at SciVerse ScienceDirect

Atmospheric Environment xxx (2012) 1e24

Contents lists available

Atmospheric Environment

journal homepage wwwelsevier comlocateatmosenv

Review

Real-time air quality forecasting part I History techniques and current status

Yang Zhang ab Marc Bocquet cd Vivien Mallet cd Christian Seigneur d Alexander Baklanov e

aDepartment of Marine Earth and Atmospheric Sciences North Carolina State University Raleigh NC 27695 USAbDepartment of Environmental Engineering and Sciences Tsinghua University Beijing Chinac INRIA Paris Rocquencourt Research Center FrancedCEREA (Atmospheric Environment Center) Joint Laboratory Eacutecole des Ponts ParisTech and EDF RampD Universiteacute Paris-Est 77455 Marne-la-Valleacutee FranceeResearch Department Danish Meteorological Institute (DMI) Copenhagen DK-2100 Denmark

h i g h l i g h t s

lt A comprehensive review for real-time air quality forecasting (RT-AQF)lt Major milestones in the history of RT-AQFlt Assessment of major RT-AQF techniqueslt A detailed review of 3-D RT-AQF modelslt Evaluation protocols and review of current forecasting skills of RT-AQF models

a r t i c l e i n f o

Article historyReceived 6 February 2012Received in revised form7 June 2012Accepted 8 June 2012

KeywordsAir quality forecastingHistoric milestoneTechniques and toolsEvaluation methods

Corresponding author Department of Marine EartCampus Box 8208 NCSU Raleigh NC 27695 USA

E-mail address yang_zhangncsuedu (Y Zhang)

1352-2310$ e see front matter 2012 Elsevier Ltdhttpdxdoiorg101016jatmosenv201206031

Please cite this article in press as Zhang YEnvironment (2012) httpdxdoiorg1010

a b s t r a c t

Real-time air quality forecasting (RT-AQF) a new discipline of the atmospheric sciences represents oneof the most far-reaching development and practical applications of science and engineering posesunprecedented scientific technical and computational challenges and generates significant opportu-nities for science dissemination and community participations This two-part review providesa comprehensive assessment of the history current status major research and outreach challenges andfuture directions of RT-AQF with a focus on the application and improvement of three-dimensional (3-D)deterministic RT-AQF models In Part I major milestones in the history of RT-AQF are reviewed Thefundamentals of RT-AQF are introduced Various RT-AQF techniques with varying degrees of sophisti-cation and skills are described comparatively Among all techniques 3-D RT-AQF models with online-coupled meteorologyechemistry and their transitions from mesoscale to unified model systems acrossscales represent a significant advancement and would greatly enhance understanding of the underlyingcomplex interplay of meteorology emission and chemistry from global to urban scales in the realatmosphere Current major 3-D global and regional RT-AQF models in the world are reviewed in terms ofmodel systems component models application scales model inputs forecast products horizontal gridresolutions and model treatments of chemistry and aerosol processes An important trend of suchmodels is their coupling with an urban model or a computational fluid dynamic model for urbanlocalscale applications at 1 km or less and with an exposure model to provide real-time public healthassessment and exposure predictions Evaluation protocols are described along with examinations ofcurrent forecasting skills and areas with large biases of major RT-AQF models

2012 Elsevier Ltd All rights reserved

h and Atmospheric Sciences

All rights reserved

et al Real-time air quality for16jatmosenv201206031

1 Introduction

11 Importance and type of applications of air quality forecasting

Air quality refers to the chemical state of the atmosphere ata given time and place Like weather air quality affects everyoneAir pollutants include gaseous and particulate species thatmay lead to non-carcinogenic andor carcinogenic adverse healtheffects Numerous studies (eg Greenbaum et al 2001 WHO

ecasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e242

2004 Georgopoulos et al 2009 Phalen and Phalen 2011 httpwwwepagovapticourse422ap7ahtml) show that acute (short-term) exposure to high levels of these species may pose serioustemporary health concerns such as eye irritation difficultybreathing pulmonary and cardio-vascular health effects andpremature death Chronic (long-term) exposure may lead to healthconcerns such as cancer premature death and damage to thebodyrsquos immune neurological reproductive and respiratorysystems People with pre-existing heart and lung diseases anddiabetics the elderly and children (so-called sensitive groups) areat an even greater risk for air pollution-related health effects Inaddition these pollutants and their derivatives can cause manyadverse effects on the environment including visibility impairmentacid deposition global climate change water quality deteriorationand plant and eco-environmental system damages (eg ESA 2004Seinfeld and Pandis 2006 Krupa et al 2006)

To protect human health and the environment the WorldHealth Organization (WHO) has issued guidelines and severalcountries and states have issued regulations (eg WHO 2010) TheUS Environmental Protection Agency (EPA) has set NationalAmbient Air Quality Standards (NAAQSs) for six air pollutants toprotect human health (US EPA 1996 Seinfeld and Pandis 2006)These pollutants are sulfur dioxide (SO2) nitrogen dioxide (NO2)carbonmonoxide (CO) ozone (O3) lead (Pb) and particulatematterwith aerodynamic diameters less than or equal to 25 mm (PM25)and 10 mm (PM10) In Europe the European Union (EU) issuesdirectives which subsequently become air quality standards orgoals in 27 member states (EU 2008) The pollutants regulated bythe EU include the six pollutants regulated in the US and benzene(C6H6) a volatile organic compound (VOC) with known carcino-genic health effects Despite significant progress in understandingemissions and fates of these pollutants as well as in reducing theirambient levels in urban areas in the past half century air pollutionis estimated to kill 3 million people a year worldwide (httpwwwworld-sciencenetothernews070814_ diseasehtm) In particularPM pollution has been directly linked to excess deaths in manycountries in the world (eg Schwartz 1991 Dockery et al 1993Kunzli et al 2000 Matanoski and Tao 2002 Millman et al 2008Jayachandran 2009 Zhang et al 2010)

To protect citizens from unhealthy air many countries have real-time air quality forecasting (RT-AQF) programs in place to forecastthe concentrations of pollutants of special health concerns such asO3 NO2 PM25 and PM10 (eg Manins 1999 US EPA 1999Pudykiewicz and Koziol 2001) Such information has been used toissue early air quality alerts that allow government and people totake precautionary measures such as temporarily shutting offmajor emission sources and car pooling or taking public trans-portation to reduce air pollution and avoid or limit their exposuresto unhealthy levels of air pollution (Wayland et al 2002) It hasbeen used to schedule and plan numerous field campaigns toeffectively track pollutant plume transport and sample pollutantconcentrations and maximize the usage of expensive instrumentedplatforms such as airplanes and other limited measurementresources (eg Lee et al 1997 Flatoslashy et al 2000) Accurate RT-AQFcan therefore offer tremendous societal and economic benefits byenabling advanced planning for individuals organizations andcommunities in order to reduce pollutant emissions and theiradverse health impacts

Driven by crucial regulations societal and economic needsscientific advancements and increasing availability of highperformance computing capacity RT-AQF has evolved fromweather forecasting and developed into a new discipline thatintegrates science and technology from several disciplinesincluding meteorology atmospheric chemistryair quality mathe-matics physics environmental statistics and computer sciences

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

engineering In light of a significant progress in the past twodecades and the unprecedented challenges of RT-AQF this workintends to offer a comprehensive review of its history currentstatus fundamentals technical approaches evaluation protocolsand improvement techniques The objectives of this review are tosummarize past achievements of RT-AQF identify major areas offuture improvement and resource investments recommendresearch priorities and provide future prospects to guide thefurther development of this new exciting yet very challengingdiscipline for the decades to come In the remaining Section 1 thehistory and current status are reviewed and the fundamentals ofRT-AQF are introduced Various RT-AQF techniques are describedcomparatively along with their strengths and limitations in Section2 Major 3-D global and regional RT-AQF models are reviewed inSection 3 Evaluation protocols and current RT-AQF skills aredescribed in Section 4

12 History and current status

A sequence of severe localized pollution episodes were repor-ted during the 1930s-1960s These include the Meuse ValleyBelgium coal and coke burning smog event with 63 excess deathsand 6000 illnesses in December 1930 (Haldane 1931) the DonoraPennsylvania US zinc smelter smog event with 20 excess deathsand 7000 illnesses in October 1948 (Battan 1966) and the LondonUK coal and chemical combustion smog events with excess deathsof 4000 in December 1952 1000 in January 1956 300e800 in 1957and 340e700 in 1962 (Brimblecombe 1987) These air pollutionepisodes and disasters led to the passage of a number of early airpollution regulations and laws such as the Air Pollution Control Actof 1955 the Clean Air Act (CAA) of 1963 and the CAA Amendmentsof 1970 and 1977 in the US as well as the CAA of 1956 and 1968 inGreat Britain

Table 1 summarizes the major milestones in the history ofRT-AQF In October 1954 heavy smog occurred in Los Angeleswhich triggered the shutdown of schools and industry for the firsttime in the US The Los Angeles County Air Pollution ControlDistrict (LACAPCD) the first regional air pollution control agency inthe US adopted a three-stage smog alert system for O3 and threeother pollutants in June 1955 to prevent the possibility of an airpollution catastrophe In the 1960s the US Weather Bureau(USWB) (ie the predecessor of the National Weather Service(NWS)) provided the first forecasts of air stagnation or pollutionpotential by using numerical weather prediction (NWP) models toforecast conditions conducive to poor air quality (eg Niemeyer1960) These forecasts were conducted strictly from a meteorolog-ical perspective without considering the emissions and chemistryof air pollutants In 1965 the US Environmental Science ServicesAdministration (ESSA) (the predecessor of the National Oceanic andAtmospheric Administration (NOAA)) was created The weatherofficials from 25 nations met in London for the First InternationalClean Air Congress in 1966 In 1970 ESSA became NOAA and theUSWB became NWS The major findings of relationships betweenmeteorological factors and air pollution potential led to the creationof the National Air Pollution Potential Forecast Program (Gross1970) On December 2 1970 the US EPA was created by Presi-dent Nixon In 1979 the NWS developed a global data assimilationsystem and its nested grid model became operational

Starting in the 1970s various RT-AQF techniques and tools weredeveloped to forecast air pollution in urban areas These techniqueswere largely based on empirical approaches and statistical modelstrained or fitted to historical air quality and meteorological data(eg McCollister and Wilson 1975 Wolff and Lioy 1978 Aron1980) Their level of sophistication increased considerably toovercome some of these limitations and address non-linearity of

recasting part I History techniques and current status Atmospheric

Table 1Selected milestones of real-time air quality forecasting (RT-AQF)

Time Milestone

October 1954 Heavy smog shutdown schools and industry in LA for the first time in the USJune 1955 A three-stage smog alert system was established in LA1960s USWB provided the first forecasts of air stagnation or pollution potential1966 The First International Clean Air Congress1970 AFR 161-22 Environmental Pollution Control required air-pollution potential forecasts and warnings the national

air pollution potential forecast program was established in the US1970se1980s Development and application of RT-AQF based on empirical approaches and statistical modelsEarly 1970s to1980s Development and application of the first-generation AQMs on urbanregional scalesLate 1970s The development and application of the first-generation AQMs on a global scale1980seearly 1990s The development and application of the second-generation AQMs on all scales1990s Application of 3-D air quality models for short-term forecasting of O3 in several countries1994e1995 The first application of an operational AQF model by the University of Cambridge UK to support the planning

of two field experiments for stratosphereMid 1990se2000s The development and application of the third-generation AQMs on all scales1997 The US EPA revised the AQI and established the AIRNow program to provide real-time air quality measured and

forecasted data to the public the first application of an operational AQF model by the Norwegian Institute for AirResearch to support a field campaign in troposphere

1999 The Meteorological Service of Canada (MSC) initiated an AQF program for eastern Canada the US EPA developedguideline for developing O3 forecasting program

Early 2000s The US Weather Research Program recommended the development and application of operational mesoscale air qualityforecasts the high-resolution advanced Weather Research and Forecasting model (WRF) was developed to serve as thebackbone of the US public weather forecasts and research WRF v10 was released for beta test in 2001 The first officialreleased version was v20 in 2004 A decision to incorporate chemistry into WRF was made at a workshop on ModelingChemistry in Cloud and Mesoscale Models held at NCAR on March 6e8 2000

Early 2000s-present The development and application of the 4th-generation air quality models on all scales2001 MSC launched the national air quality forecast in May 2001 the first reported ensemble O3 forecasting the first application

of chemical data assimilation for AQFs since their first applications to CTMs in mid 1990s2002 Under the provision of the Energy Policy Act of 2002 of USA Congress mandated NOAA to develop nationwide RT-AQF

capability for O3 NOAArsquos pilot study of predicting O3 for the New England region in 2002 using three numerical AQF modelsystems the first version of WRFChem was released WRFChem represents the first community online-coupledmeteorologyechemistry model

2003 The first RT-AQF workshop in the US organized by USWRP held during April 29-May 3 in Houston Texas NOAA and EPAformally collaborate to develop a national RT-AQF model

2004 The second New England Forecasting Program to forecast O3 and PM25 in the northeastern US during which six numericalAQF models were deployed for RT-AQF the US EPANOAArsquos NAQFC was deployed in the summer of 2004 CFD was coupledwith the 3-D AQF models for AQF over industrial plants and urban areas at horizontal grid resolutions of 1e10 m

2005 European FUMAPEX UAQIFS with improved urban meteorology air quality and population exposure models were developedand launched for operational urban AQF in 6 EU cities

2006-present Ensemble forecast based on sole sequential aggregation the development of the ensemble forecast of Analyses (EFA) approachto combine ensembles and data assimilation

2007 The first large scale inverse modeling with 4D-Var data assimilation to combine emission rate and chemical state optimizationin a complex 3-D AQF model (ie the University of Cologne mesoscale EURAD model)

Y Zhang et al Atmospheric Environment xxx (2012) 1e24 3

the photochemical system during the 1990s (eg Ryan 1995 Ryanet al 2000) The development of 3-D numerical air quality models(AQMs) on urban regional and global scales since the 1970s andtheir applications for short- and long-term RT-AQF since the mid1990s led to a significant frog leap With significant advances incomputational technologies and computer architectures RT-AQFsystems have shifted from postprocessing of forecast meteorologyand statistical methods to the use of sophisticated 3-D AQMs thataccount for meteorology emissions chemistry and removalprocesses

3-D AQMs have evolved over four generations since the 1970swith roughly one generation per decade reflecting the advance-ments in scientific understanding and numerical and computa-tional technologies Review of some AQMs can be found in theliterature (eg Seigneur 1994 2001 Peters et al 1995 Russell andDennis 2000 Zhang 2008) CTMs or AQMs have traditionally beenused to retrospectively simulate historical poor air quality scenariosin support of regulation and planning due primarily to computa-tional constraints and a lack of real-time chemical measurementsGiven their relative maturity in sciences and the advancement incomputational technology some of the 2nd 3rd and 4th genera-tions AQMs have been deployed for RT-AQF since the mid to late1990s These efforts were first begun in Germany in 1994 (egRufeger et al 1997) Japan in 1996 (eg Ohara et al 1997) Australiain 1997 (eg Manins et al 2002) and Canada in 1998 (eg

Please cite this article in press as Zhang Y et al Real-time air quality forEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

Pudykiewicz and Koziol 2001) and then expanded in the US(McHenry et al 2004 Otte et al 2005) other countries in Europe(eg Brandt et al 2001 Jakobs et al 2002) China (Han et al 2002Wang et al 2009) and other regions in Japan (Uno et al 2003) Inaddition to applications to short-term forecasts of air pollution forthe public 3-D AQMs have also been applied for chemical fore-casting during field campaigns Lee et al (1997) and Flatoslashy et al(2000) represent the first RT-AQF to support the planning of fieldexperiments for the troposphere and stratosphere respectivelyFollowing the two studies a number of RT-AQFs have been appliedbefore and during field campaigns (eg Kang et al 2005 McKeenet al 2005 2007 2009)

Another frog leap in the history of RT-AQF is the development ofreal-time data repositories such as the Aerometric InformationRetrieval Now (AIRNow) network (wwwairnowgov) In 1997 theUS EPA developed AIRNow to provide an effective platform forcommunicating real-time air quality conditions and forecasts to thepublic via the internet and other medias Differing from the tradi-tional data submission quality-control and calibration that usuallytake 3e6 months after their collection AIRNow receives real-timeO3 and PM pollution data from more than 115 US and Canadianagencies as well RT-AQFs from about 400 US cities and representsa centralized nationwide governmental repository for real-timedata AIRNow has been expanded to include real-time data fromother countries such as China Similar programs exist in the EU

ecasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e244

within the Global Monitoring for Environment and Security (GMES)Programme and the European Commission Seventh FrameworkProgramme Such efforts are also coordinated by several EuropeanCooperation in Science and Technology (COST) Actions and EUnetworks The air quality agency of the Paris region in FranceAIRPARIF makes real-time data available on its website A near-real-time observation database Base de donneacutees en temps reacuteel(BASTER) gathers all hourly measurements made by the local airquality networks every 3 h in France Germany Italy FinlandAustria and the UK (Rouiumll et al 2009) The US EPA established thePollutant Standard Index (PSI) (also known as the Air PollutionIndex (API)) in 1978 which was replaced by the Air Quality Index(AQI) to include a simple color scheme in 1997 to link air qualityconcentrations and associated health effects to a simple color-coded index that can be easily and consistently reported to thepublic (US EPA 2000 2009) Similar AQIAPI exist in more than 37countries in the world (eg Australia Canada Mexico DenmarkFrance UK Germany China India Japan Brazil Chile)

In the mid-late 1990s many countries recognized an increasingneed to implement a centralized national air quality forecastingsystem The Minister for the Environment in Australia funded theAir Pollution in Major Cities Program and developed their RT-AQFmodel in 1998 The Meteorological Service of Canada (MSC) initi-ated an RT-AQF program for eastern Canada in 1999 which wasextended to cover all of subarctic Canada in 2001 (Pudykiewiczet al 2003) In 1999 the US EPA developed a guideline for O3forecasting (US EPA 1999) which was extended to add PM25 in2003 (US EPA 2003) Under the provision of the Energy Policy Actof 2002 of the US the US Congress mandated NOAA to developa nationwide RT-AQF capability for O3 Since then NOAA and EPAcollaboratively developed a national RT-AQF model and providedO3 forecast guidance to state and local forecasters (Stockwell et al2002 Wayland et al 2002 Dabberdt et al 2004) They conductedthe first pilot study of predicting O3 for the New England region in2002 using three numerical RT-AQF models (Kang et al 2005) Thefirst workshop on RT-AQF in the US was organized by the USWRPand held in 2003 in Houston Texas during which 50 scientistsrecommended research areas and priorities for RT-AQF that helpguide the further development of the nationwide RT-AQF researchprogram (Dabberdt et al 2006) In 2004 NOAA sponsored thesecond New England Forecasting Program to forecast both O3 andPM25 in the northeastern US during which 5 RT-AQF models wereused for ensemble modeling (McKeen et al 2005 2007) The real-time National Air Quality Forecast Capability (NRT-AQFC) thatconsists of Eta-CMAQ jointly-developed by US NOAA and EPA wasdeployed for RT-AQF in the summer of 2004 it represents the firstoperational forecast implementation of CMAQ (Otte et al 2005)Region-wide efforts by universities and research organizations areprevalent in the US and many countries in the world (eg Chenet al 2008 Hogrefe et al 2007 Cai et al 2008)

Learning from initial RT-AQF applications more sophisticatedtechniques such as 4-dimensional variational method (4D-Var)Kalman-filtering and ensemble methods have been used inconjunction with 3-D RT-AQF models to improve RT-AQF resultsElbern and Schmidt (2001) conducted one of the first applicationsof chemical data assimilation (CDA) for RT-AQF using 4D-Var toassimilate O3 and NO2 observations during August 1997 overcentral Europe and showed a significant improvement in O3 fore-casts The first reported ensemble O3 forecasting was conductedwith a three-member CHIMERE ensemble using three global NWPsover Europe (Vautard et al 2001a) As one of the early real-timeensemble RT-AQFs McKeen et al (2005 2007) applied severalRT-AQF models to forecast O3 and PM25 respectively during the2004 NEAQSICARTT study and found that multimodel ensembleforecasting outperformed individual forecasting The same set of

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

RT-AQF models was applied to forecast O3 and PM25 during theTexAQS IIGoMACCS with the ensemble forecasting performingbetter than most individual members (McKeen et al 2009) whichrepresents the first PM25 ensemble bias-corrected and Kalmanfilter-corrected forecasting Doraiswamy et al (2009) conductedensemble forecast of O3 and PM25 over the state of New York usingCMAQ with WRF and MM5 and found that the ensemble forecastsoften but not always perform better than individual memberforecasts and the weighting or bias correction approaches mayimprove performance Mallet (2010) coupled an ensemble fore-casting of O3 with CDA to overcome the limitations of pureensemble forecasting and showed a 28 reduction in the root meansquare error (RMSE) Hybrid approaches using both statistical and3-D models have also been applied to improve the accuracy of theRT-AQF (eg Eben et al 2005 Guillas et al 2008 Kang et al 2008Rouiumll et al 2009) There are increasing numbers of RT-AQF appli-cations of CTMs coupled with Computational Fluid Dynamical(CFD) models over industrial plants and urban areas at horizontalgrid resolutions of 1e10 m (eg San Joseacute et al 2006 2009) Theseapplications predict chemical concentrations in the urban canopytaking into account the complex building structure

13 Fundamentals of real-time air quality forecasting

131 Major characteristics of RT-AQFFig 1 shows a diagram of an RT-AQF model system from global

to urban scales based on typical configurations available fromcurrent RT-AQF models At each scale a meteorological model andan air quality model are needed they may be coupled online oroffline While an offline meteorological model (eg MM5 Grellet al 1995 or WRF Michalakes et al 2001) provides meteoro-logical forecasts separately from a regional RT-AQF model (egCMAQ Byun and Schere (2006)) an online-coupled meteorologyand chemistry model (eg WRFChem Grell et al 2005) gener-ates both meteorological and chemical forecasts within the sametime step At a global scale the GCM and the global CTM (GCTM)(eg ECHAM5 Roeckner et al 2006 MOCAGE Rouiumll et al 2009)are initiated with climatological or reanalysis or observationaldata The GCM produces the meteorological fields needed by theGCTM An emission model is required to project real-time emis-sions based on energyfuel consumption data at all scales Anemission processor converts projected emissions into model-ready gridded emission files Forecasted meteorological informa-tion is needed for meteorology-dependent emissions (egbiogenic emissions sea-salt and erodible dust emissions VOCevaporation) The regional scale RT-AQF systems (eg CHIMERERouiumll et al 2009 EURAD Elbern et al 2010) use meteorologicalinitial and boundary conditions (ICONs and BCONs) from a GCMand chemical BCONs from a GCTM Chemical ICONs from eithera GCTM or observations are needed to initiate the first dayrsquosforecasting those for subsequent days will then use previous dayrsquosforecast The forecasting product will be post-processed andevaluated using real-time (or near real-time) observations Prod-ucts such as species concentrations and AQIs will be submitted tothe medias web sites and subscribers Some RT-AQF systemsemploy bias correction techniques to correct large systematicbiases for next-dayrsquos forecast based on previous dayrsquos forecast andobservations (eg McKeen et al 2005 2007 Kang et al 2008)The initial RT-AQFs for current day issued on a previous day maybe updated in the morning based on updated meteorologicalforecasting to improve the accuracy of the forecasting productsSome RT-AQF systems include an extended subsystem for urbanscale air quality andor human exposure and environmentalhealth forecasting (eg Tilmes et al 2002 Baklanov et al 2007San Joseacute et al 2009) The urban RT-AQF requires urban

recasting part I History techniques and current status Atmospheric

Fig 1 An automated RT-AQF system from global to urban scales The three model systems at global regional and urban scales are shown as offline-coupled meteor-ologyechemistry models They can be online-coupled systems

Y Zhang et al Atmospheric Environment xxx (2012) 1e24 5

meteorological forecasts background concentrations forecastedfrom the regional RT-AQF model and traffic emissions that arecalculated using detailed traffic information The output includesthe spatial and temporal distributions of forecasted concentra-tions The neighborhood scale human exposure or environmentalhealth forecasting requires urban meteorological forecasts fore-casted concentrations and deposition from a regional RT-AQFmodel demographic and geographic data (eg total number ofpopulation and age distribution location and time-activity ofpeople) and health data (eg mortality morbidity hospitaladmissions) (Baklanov et al 2007 and references therein) Theoutput includes the spatial and temporal distributions of fore-casted total dose and relative risks of adverse health outcome The

Please cite this article in press as Zhang Y et al Real-time air quality forEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

entire data retrieval model simulation and product processing isautomated on a day-to-day basis to ensure completion of forecastsin time

While RT-AQF is built upon weather forecasting and traditionalair quality modeling it differs from them in many aspects RT-AQFshares the same requirements for time and optimization of modeltreatments as NWP but requires simulations of additional physico-chemical processes affecting fates of air pollutants it is thus moredifficult than weather forecast (Stockwell et al 2002) While theweather forecasting is typically available for 3e10 days RT-AQF istypically available for 1e3 days at present (eg STEM-2K3Carmichael et al 2003 WRFChem McKeen et al 2005 2007MOCAGE Rouiumll et al 2009 CHIMERE Rouiumll et al 2009) due to

ecasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e246

a much greater computational demand in terms of CPU time anddisk storage required

Compared with traditional air quality modeling for research-grade and regulatory applications RT-AQF has its unique tech-nical challenges and time requirements and involves a number ofreal-time operational issues that did not exist before Backcasting isdriven by regulatory guidance and compliance aiming at repro-ducing historic pollution episodes with a high accuracy RT-AQF isdriven by societal pressures tominimize the human environmentaland economic impacts of air pollution aiming at producing warn-ings of high pollutant concentrations that will likely pose imme-diate health threats to the population This difference leads todifferences in many aspects of backcasting and RT-AQF Forexample backcasting often uses the best available model treat-ments to reproduce the species concentrations from a historicpollution episode RT-AQF typically uses simplified optimizedoptions for dynamics chemistry and physics treatments that arefast enough yet reasonably accurate to meet time requirements foroperational forecastingRT-AQF often applies special techniques(eg bias correction Kang et al 2008) to achieve accuracy andcomputational efficiency in a very short turnaround time whichare often not needed for backcasting While backcasting generallyhas no specific time window requirements RT-AQF requires a fastshort-term prediction on a day-by-day basis This time pressuredictates the implementation of a fully automated system todownload and preprocess real-time (or near real-time) datasets forRT-AQF model set up and simulations and the use of a fast set ofmodel options for an efficient deployment of RT-AQF For productevaluation while RT-AQF products can be evaluated using evalua-tion protocols for traditional AQMs it is more meaningful to usecategorical evaluation with threshold statistics (eg probability ofdetection false alarm ratio) (eg McHenry et al 2004 McKeenet al 2005 2007 2009 Kang et al 2008) because the primaryvalue of RT-AQFs is their guidance for issuing health advisories andalerts of an air pollution episode and the categorical indicesdetermine the likelihood of such an episode The end users forRT-AQF products include researchers forecasters regulators deci-sion makers in health environmental meteorological and militaryagencies polluters (eg industrialists manufacturers residentialwoodburners car users) the public the media (eg newspapersand other printed medias television radio phone pager andinternet) and the commercial sector (eg health insuranceindustry) (Dabberdt et al 2006) which involve much largercommunities than those of backcasting RT-AQF requires a specificinformation technology infrastructure that is not always neededfor backcasting (eg web-based interfaces the construction ofapplication-specific andor client-oriented datasets and a readilyaccess to forecast products)

RT-AQF can be made for the short- or long-term depending onthe objective of the applications (Dabberdt et al 2006) The short-term forecasts (1e5 days) will predict concentrations of pollutantsand can be used daily to inform the general public about thepotentially harmful conditions and to take preventive actions toreduce exposures and emissions The long-term forecasts (gt1-yr)will provide long-term variation trends of pollutants and theircorrelationswith forecasted emissions andmeteorology that can beused to guide the development of State Implementation Plans forthe designated non-attainment regions Such long-term forecastsfor greenhouse gases can also be used for global climate changemitigation This review focuses on short-term RT-AQF

132 Forecasting productsThe RT-AQF efforts since the 1990s have focused primarily on O3

(eg Cope et al 2004 McHenry et al 2004 McKeen et al 2005)and have only recently been expanded to include PM25 and PM10

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

(eg McKeen et al 2007 2009 Chuang et al 2011) Some alsoforecast other pollutants such as SO2 NOx CO VOCs air toxics (egbenzene and 13-butadiene) and dust (eg Cope et al 2004Baklanov et al 2007 Kaminski et al 2008 Kallos et al 2009Elbern et al 2010) For example in Europe NO2 is included becauseof common exceedances of the air quality standard Because therewas no exceedances of the annual NO2 standard NO2 has typicallynot been included in the US although the recent promulgation ofa 1-h average near-source NO2 standard may soon generate someinterest for NO2 forecasting in North America The forecast productsare issued in terms of spatial maps or site-specific values of hourlyconcentrations and time-averaging concentrations (eg maximum8-h average concentrations based on the US NAAQS) as well asAQI Eder et al (2010) described approaches to use national RT-AQFguidance to develop local AQI forecasts The AQI used in the US isa dimensionless six-color-coded index for reporting daily airquality to the public in a manner as easily understood as weatherforecasts (US EPA 2009) It provides a simple uniform system torelate daily forecasted levels of criteria pollutants (eg O3 PM NO2CO and SO2) to health advisories and alerts for sensitive groups andthe general public and suggests actions to reduce exposureTable A1 in the supplementary material shows the AQI for O3 andPM that was established by the US EPA (US EPA 2009 wwwairnowgov) The AQI converts a forecasted pollutant concentra-tion to a number on a scale of 0e500 A value of 100 generallycorresponds to the NAAQS established for each pollutant underthe Clean Air Act Values below 100 are considered satisfactoryValues above 100 indicates that the air is unhealthy and posesa health concern For example the 100e200 level may triggerpreventive actions such as limiting certain activities and enforcingpotential restrictions on industrial activities by state or localofficials

A European AQI the Common Air Quality Index (CAQI) wasdeveloped to compare air quality in different European cities(Elshout and Leacuteger 2007) An important feature of CAQI is that itaccounts for both urban background monitoring conditions at citybackground sites and traffic (ie near-source) pollution atneartraffic monitoring sites The background index indicates theoutdoor air quality in the city experienced by the average citizenThe obligatory background index comprises NO2 PM10 and O3with CO and SO2 as auxiliary components The traffic index indi-cates air quality in busy streets which is generally the poorest airquality in the city Citizens living working and visiting thesestreets as well as those in vehicles are all affected The obligatorytraffic index comprises NO2 and PM10 with CO as an auxiliarycomponent The two indices provide an improved assessment ofcurrent air quality over city averages because some monitoringnetworks are designed to monitor areas of poor air quality andothers provide an average city picture CAQI has 5 levels of pollu-tion using a scale of 0e25 (very low) 25e50 (low) 50e75(medium) 75e100 (high) and gt100 (very high) and the match-ing colors are green light green yellow red and dark redrespectively

2 Techniques and tools of RT-AQF and their advantageslimitations

Existing RT-AQF techniques and tools have various levels ofsophistication and forecasting skills They can be grouped into threecategories simple empirical approaches (also referred to as Criteriaor simple rules of thumb) parametric or non-parametric statisticalapproaches and more advanced physically-based approachesSome of them have been reviewed with varying levels of details inthe literature (eg Wayland et al 2002 US EPA 2003 Cope andHess 2005 Schere et al 2006 Menut and Bessagnet 2010

recasting part I History techniques and current status Atmospheric

Fig 1 An automated RT-AQF system from global to urban scales The three model systems at global regional and urban scales are shown as offline-coupled meteor-ologyechemistry models They can be online-coupled systems

Y Zhang et al Atmospheric Environment xxx (2012) 1e24 5

meteorological forecasts background concentrations forecastedfrom the regional RT-AQF model and traffic emissions that arecalculated using detailed traffic information The output includesthe spatial and temporal distributions of forecasted concentra-tions The neighborhood scale human exposure or environmentalhealth forecasting requires urban meteorological forecasts fore-casted concentrations and deposition from a regional RT-AQFmodel demographic and geographic data (eg total number ofpopulation and age distribution location and time-activity ofpeople) and health data (eg mortality morbidity hospitaladmissions) (Baklanov et al 2007 and references therein) Theoutput includes the spatial and temporal distributions of fore-casted total dose and relative risks of adverse health outcome The

Please cite this article in press as Zhang Y et al Real-time air quality forEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

entire data retrieval model simulation and product processing isautomated on a day-to-day basis to ensure completion of forecastsin time

While RT-AQF is built upon weather forecasting and traditionalair quality modeling it differs from them in many aspects RT-AQFshares the same requirements for time and optimization of modeltreatments as NWP but requires simulations of additional physico-chemical processes affecting fates of air pollutants it is thus moredifficult than weather forecast (Stockwell et al 2002) While theweather forecasting is typically available for 3e10 days RT-AQF istypically available for 1e3 days at present (eg STEM-2K3Carmichael et al 2003 WRFChem McKeen et al 2005 2007MOCAGE Rouiumll et al 2009 CHIMERE Rouiumll et al 2009) due to

ecasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e246

a much greater computational demand in terms of CPU time anddisk storage required

Compared with traditional air quality modeling for research-grade and regulatory applications RT-AQF has its unique tech-nical challenges and time requirements and involves a number ofreal-time operational issues that did not exist before Backcasting isdriven by regulatory guidance and compliance aiming at repro-ducing historic pollution episodes with a high accuracy RT-AQF isdriven by societal pressures tominimize the human environmentaland economic impacts of air pollution aiming at producing warn-ings of high pollutant concentrations that will likely pose imme-diate health threats to the population This difference leads todifferences in many aspects of backcasting and RT-AQF Forexample backcasting often uses the best available model treat-ments to reproduce the species concentrations from a historicpollution episode RT-AQF typically uses simplified optimizedoptions for dynamics chemistry and physics treatments that arefast enough yet reasonably accurate to meet time requirements foroperational forecastingRT-AQF often applies special techniques(eg bias correction Kang et al 2008) to achieve accuracy andcomputational efficiency in a very short turnaround time whichare often not needed for backcasting While backcasting generallyhas no specific time window requirements RT-AQF requires a fastshort-term prediction on a day-by-day basis This time pressuredictates the implementation of a fully automated system todownload and preprocess real-time (or near real-time) datasets forRT-AQF model set up and simulations and the use of a fast set ofmodel options for an efficient deployment of RT-AQF For productevaluation while RT-AQF products can be evaluated using evalua-tion protocols for traditional AQMs it is more meaningful to usecategorical evaluation with threshold statistics (eg probability ofdetection false alarm ratio) (eg McHenry et al 2004 McKeenet al 2005 2007 2009 Kang et al 2008) because the primaryvalue of RT-AQFs is their guidance for issuing health advisories andalerts of an air pollution episode and the categorical indicesdetermine the likelihood of such an episode The end users forRT-AQF products include researchers forecasters regulators deci-sion makers in health environmental meteorological and militaryagencies polluters (eg industrialists manufacturers residentialwoodburners car users) the public the media (eg newspapersand other printed medias television radio phone pager andinternet) and the commercial sector (eg health insuranceindustry) (Dabberdt et al 2006) which involve much largercommunities than those of backcasting RT-AQF requires a specificinformation technology infrastructure that is not always neededfor backcasting (eg web-based interfaces the construction ofapplication-specific andor client-oriented datasets and a readilyaccess to forecast products)

RT-AQF can be made for the short- or long-term depending onthe objective of the applications (Dabberdt et al 2006) The short-term forecasts (1e5 days) will predict concentrations of pollutantsand can be used daily to inform the general public about thepotentially harmful conditions and to take preventive actions toreduce exposures and emissions The long-term forecasts (gt1-yr)will provide long-term variation trends of pollutants and theircorrelationswith forecasted emissions andmeteorology that can beused to guide the development of State Implementation Plans forthe designated non-attainment regions Such long-term forecastsfor greenhouse gases can also be used for global climate changemitigation This review focuses on short-term RT-AQF

132 Forecasting productsThe RT-AQF efforts since the 1990s have focused primarily on O3

(eg Cope et al 2004 McHenry et al 2004 McKeen et al 2005)and have only recently been expanded to include PM25 and PM10

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

(eg McKeen et al 2007 2009 Chuang et al 2011) Some alsoforecast other pollutants such as SO2 NOx CO VOCs air toxics (egbenzene and 13-butadiene) and dust (eg Cope et al 2004Baklanov et al 2007 Kaminski et al 2008 Kallos et al 2009Elbern et al 2010) For example in Europe NO2 is included becauseof common exceedances of the air quality standard Because therewas no exceedances of the annual NO2 standard NO2 has typicallynot been included in the US although the recent promulgation ofa 1-h average near-source NO2 standard may soon generate someinterest for NO2 forecasting in North America The forecast productsare issued in terms of spatial maps or site-specific values of hourlyconcentrations and time-averaging concentrations (eg maximum8-h average concentrations based on the US NAAQS) as well asAQI Eder et al (2010) described approaches to use national RT-AQFguidance to develop local AQI forecasts The AQI used in the US isa dimensionless six-color-coded index for reporting daily airquality to the public in a manner as easily understood as weatherforecasts (US EPA 2009) It provides a simple uniform system torelate daily forecasted levels of criteria pollutants (eg O3 PM NO2CO and SO2) to health advisories and alerts for sensitive groups andthe general public and suggests actions to reduce exposureTable A1 in the supplementary material shows the AQI for O3 andPM that was established by the US EPA (US EPA 2009 wwwairnowgov) The AQI converts a forecasted pollutant concentra-tion to a number on a scale of 0e500 A value of 100 generallycorresponds to the NAAQS established for each pollutant underthe Clean Air Act Values below 100 are considered satisfactoryValues above 100 indicates that the air is unhealthy and posesa health concern For example the 100e200 level may triggerpreventive actions such as limiting certain activities and enforcingpotential restrictions on industrial activities by state or localofficials

A European AQI the Common Air Quality Index (CAQI) wasdeveloped to compare air quality in different European cities(Elshout and Leacuteger 2007) An important feature of CAQI is that itaccounts for both urban background monitoring conditions at citybackground sites and traffic (ie near-source) pollution atneartraffic monitoring sites The background index indicates theoutdoor air quality in the city experienced by the average citizenThe obligatory background index comprises NO2 PM10 and O3with CO and SO2 as auxiliary components The traffic index indi-cates air quality in busy streets which is generally the poorest airquality in the city Citizens living working and visiting thesestreets as well as those in vehicles are all affected The obligatorytraffic index comprises NO2 and PM10 with CO as an auxiliarycomponent The two indices provide an improved assessment ofcurrent air quality over city averages because some monitoringnetworks are designed to monitor areas of poor air quality andothers provide an average city picture CAQI has 5 levels of pollu-tion using a scale of 0e25 (very low) 25e50 (low) 50e75(medium) 75e100 (high) and gt100 (very high) and the match-ing colors are green light green yellow red and dark redrespectively

2 Techniques and tools of RT-AQF and their advantageslimitations

Existing RT-AQF techniques and tools have various levels ofsophistication and forecasting skills They can be grouped into threecategories simple empirical approaches (also referred to as Criteriaor simple rules of thumb) parametric or non-parametric statisticalapproaches and more advanced physically-based approachesSome of them have been reviewed with varying levels of details inthe literature (eg Wayland et al 2002 US EPA 2003 Cope andHess 2005 Schere et al 2006 Menut and Bessagnet 2010

recasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e246

a much greater computational demand in terms of CPU time anddisk storage required

Compared with traditional air quality modeling for research-grade and regulatory applications RT-AQF has its unique tech-nical challenges and time requirements and involves a number ofreal-time operational issues that did not exist before Backcasting isdriven by regulatory guidance and compliance aiming at repro-ducing historic pollution episodes with a high accuracy RT-AQF isdriven by societal pressures tominimize the human environmentaland economic impacts of air pollution aiming at producing warn-ings of high pollutant concentrations that will likely pose imme-diate health threats to the population This difference leads todifferences in many aspects of backcasting and RT-AQF Forexample backcasting often uses the best available model treat-ments to reproduce the species concentrations from a historicpollution episode RT-AQF typically uses simplified optimizedoptions for dynamics chemistry and physics treatments that arefast enough yet reasonably accurate to meet time requirements foroperational forecastingRT-AQF often applies special techniques(eg bias correction Kang et al 2008) to achieve accuracy andcomputational efficiency in a very short turnaround time whichare often not needed for backcasting While backcasting generallyhas no specific time window requirements RT-AQF requires a fastshort-term prediction on a day-by-day basis This time pressuredictates the implementation of a fully automated system todownload and preprocess real-time (or near real-time) datasets forRT-AQF model set up and simulations and the use of a fast set ofmodel options for an efficient deployment of RT-AQF For productevaluation while RT-AQF products can be evaluated using evalua-tion protocols for traditional AQMs it is more meaningful to usecategorical evaluation with threshold statistics (eg probability ofdetection false alarm ratio) (eg McHenry et al 2004 McKeenet al 2005 2007 2009 Kang et al 2008) because the primaryvalue of RT-AQFs is their guidance for issuing health advisories andalerts of an air pollution episode and the categorical indicesdetermine the likelihood of such an episode The end users forRT-AQF products include researchers forecasters regulators deci-sion makers in health environmental meteorological and militaryagencies polluters (eg industrialists manufacturers residentialwoodburners car users) the public the media (eg newspapersand other printed medias television radio phone pager andinternet) and the commercial sector (eg health insuranceindustry) (Dabberdt et al 2006) which involve much largercommunities than those of backcasting RT-AQF requires a specificinformation technology infrastructure that is not always neededfor backcasting (eg web-based interfaces the construction ofapplication-specific andor client-oriented datasets and a readilyaccess to forecast products)

RT-AQF can be made for the short- or long-term depending onthe objective of the applications (Dabberdt et al 2006) The short-term forecasts (1e5 days) will predict concentrations of pollutantsand can be used daily to inform the general public about thepotentially harmful conditions and to take preventive actions toreduce exposures and emissions The long-term forecasts (gt1-yr)will provide long-term variation trends of pollutants and theircorrelationswith forecasted emissions andmeteorology that can beused to guide the development of State Implementation Plans forthe designated non-attainment regions Such long-term forecastsfor greenhouse gases can also be used for global climate changemitigation This review focuses on short-term RT-AQF

132 Forecasting productsThe RT-AQF efforts since the 1990s have focused primarily on O3

(eg Cope et al 2004 McHenry et al 2004 McKeen et al 2005)and have only recently been expanded to include PM25 and PM10

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

(eg McKeen et al 2007 2009 Chuang et al 2011) Some alsoforecast other pollutants such as SO2 NOx CO VOCs air toxics (egbenzene and 13-butadiene) and dust (eg Cope et al 2004Baklanov et al 2007 Kaminski et al 2008 Kallos et al 2009Elbern et al 2010) For example in Europe NO2 is included becauseof common exceedances of the air quality standard Because therewas no exceedances of the annual NO2 standard NO2 has typicallynot been included in the US although the recent promulgation ofa 1-h average near-source NO2 standard may soon generate someinterest for NO2 forecasting in North America The forecast productsare issued in terms of spatial maps or site-specific values of hourlyconcentrations and time-averaging concentrations (eg maximum8-h average concentrations based on the US NAAQS) as well asAQI Eder et al (2010) described approaches to use national RT-AQFguidance to develop local AQI forecasts The AQI used in the US isa dimensionless six-color-coded index for reporting daily airquality to the public in a manner as easily understood as weatherforecasts (US EPA 2009) It provides a simple uniform system torelate daily forecasted levels of criteria pollutants (eg O3 PM NO2CO and SO2) to health advisories and alerts for sensitive groups andthe general public and suggests actions to reduce exposureTable A1 in the supplementary material shows the AQI for O3 andPM that was established by the US EPA (US EPA 2009 wwwairnowgov) The AQI converts a forecasted pollutant concentra-tion to a number on a scale of 0e500 A value of 100 generallycorresponds to the NAAQS established for each pollutant underthe Clean Air Act Values below 100 are considered satisfactoryValues above 100 indicates that the air is unhealthy and posesa health concern For example the 100e200 level may triggerpreventive actions such as limiting certain activities and enforcingpotential restrictions on industrial activities by state or localofficials

A European AQI the Common Air Quality Index (CAQI) wasdeveloped to compare air quality in different European cities(Elshout and Leacuteger 2007) An important feature of CAQI is that itaccounts for both urban background monitoring conditions at citybackground sites and traffic (ie near-source) pollution atneartraffic monitoring sites The background index indicates theoutdoor air quality in the city experienced by the average citizenThe obligatory background index comprises NO2 PM10 and O3with CO and SO2 as auxiliary components The traffic index indi-cates air quality in busy streets which is generally the poorest airquality in the city Citizens living working and visiting thesestreets as well as those in vehicles are all affected The obligatorytraffic index comprises NO2 and PM10 with CO as an auxiliarycomponent The two indices provide an improved assessment ofcurrent air quality over city averages because some monitoringnetworks are designed to monitor areas of poor air quality andothers provide an average city picture CAQI has 5 levels of pollu-tion using a scale of 0e25 (very low) 25e50 (low) 50e75(medium) 75e100 (high) and gt100 (very high) and the match-ing colors are green light green yellow red and dark redrespectively

2 Techniques and tools of RT-AQF and their advantageslimitations

Existing RT-AQF techniques and tools have various levels ofsophistication and forecasting skills They can be grouped into threecategories simple empirical approaches (also referred to as Criteriaor simple rules of thumb) parametric or non-parametric statisticalapproaches and more advanced physically-based approachesSome of them have been reviewed with varying levels of details inthe literature (eg Wayland et al 2002 US EPA 2003 Cope andHess 2005 Schere et al 2006 Menut and Bessagnet 2010

recasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e24 7

Kukkonen et al 2011) Table 2 summarizes major techniques forRT-AQF

21 Simple empirical approaches

The persistence method is based on the assumption that todayrsquosobserved pollutant level is tomorrowrsquos forecasted value (US EPA2003) As such it requires yesterdayrsquos air quality data Thismethod is the quickest approach among all RT-AQF techniques Itworks well under a stationary condition during which a consis-tently low or high pollution episode occurs but fails at the begin-ning or the end of the episode because it cannot handle abruptchange of weather emissions and air quality Thus it has a lowaccuracy from a long-term RT-AQF perspective and is primarilyused as a reference (or baseline) by other methods (eg McHenryet al 2004 McKeen et al 2005 2007 2009)

Climatology is based on the hypothesis that air quality is highlydependent on weather and air quality climatology is thus analo-gous to weather climatology This method is computationally veryfast It uses historical frequency of pollutant events to guide and

Table 2Major techniques and tools for RT-AQF

Strength

Simple Empirical ApproachesPersistence Computationally fast accurate during a static ambie

simple to use and requires little expertise low operClimatology Computationally fast helps guide and bound foreca

from other methods Simple to use and requires littlow operational cost

Empiricism Computationally fast an effective screening methodpollution events simple to use low operational cos

Parametric (Statistical) ModelsClassification and Regression

Trees (CART)Computationally fast works well for a given site audifferentiates between days with similar pollutant crequires modest expertise low operational cost moaccuracy

Regression Methods Computationally fast works well for a given site coused and easy to operate produces generally good frequires modest expertise moderate operational comoderatehigh accuracy

Artificial Neural Networks(ANNs)

Capacity to learn from data works well for a given shandle nonlinear and chaotic chemical system at a smodest expertise moderate operational cost modehigh accuracy computationally fast

Fuzzy Logic Method (FL) Capacity to represent inherent uncertainties of humcan handle non-linearity and chaotic chemical systemodest expertise moderatehigh accuracy moderatcost

Kalman Filter Method Provides response prediction and estimation and asuncertainties allowing time dependence of parametcomponents of the time series can be decomposed

Advanced Physically-Based ApproachesDeterministic (CTM) Models Prognostic time- and spatially-resolved concentratio

both typical and atypical scenarios and in areas thatmonitored scientific insights into pollutant formatioaccounts for the air parcel history including transpodoes not require a large quantity of measurement dmoderatehigh accuracy

CTMs with Bias CorrectionTechniques

Combines merits of deterministic models and statist(or other techniques) high accuracy

Ensemble and ProbabilisticMethods

Improved forecast skills compared to a single membhandle uncertainties in AQF provides an estimate oof an occurrence of an event on a scale from 0 to 1

Please cite this article in press as Zhang Y et al Real-time air quality forEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

bound RT-AQF It requires historical (gt2e5 yrs) air pollutant dataBecause of the averaging nature of the climatology this methodcannot account for abrupt changes in air quality due to changes inemissions (eg forest fires sudden release of air toxics) andweather patterns that did not occur historically The accuracy islow It is not a stand-alone method and is mainly used to guide RT-AQFs derived from other methods (US EPA 2003)

Empiricism (or criteria or simple rules of thumb) is based on theassumption that thresholds (ie criteria) of forecasted meteoro-logical variables can indicate future high pollutant concentrations(Wayland et al 2002) When parameters that influence pollutionare forecasted to reach a threshold high pollutant concentrationsare forecasted This method requires observed and forecastedupper-air and surface meteorological and air quality data It ismoderately accurate However it cannot forecast exact concentra-tions and does not work for pollutants that depend weakly onweather (eg CO) This method relies on a strong correlationbetween forecasted pollutant (eg O3) and meteorological vari-ables (eg temperature) which may not always hold as warmtemperatures are necessary but not sufficient for the formation of

Limitation

nt conditionsational cost

Cannot handle abrupt change of weather emissions andair quality low accuracy not a stand-alone method

sts derivedle expertise

Cannot handle abrupt change of air quality low accuracynot a stand-alone method

for hight

Cannot forecast exact concentrations does not work forpollutants that depend weakly on weather moderateaccuracy

tomaticallyoncentrationsderatehigh

Cannot accurately predict extreme concentrations limiteduse due to limited observations and large local-scalevariations of concentrations

mmonlyorecastsst

Cannot accurately predict extreme concentrations thelinear regression cannot handle non-linearity of thechemical system limited use due to limited observationsand large local-scale variations of concentrations

ite canite requiresrate

Cannot accurately predict extreme concentrations limiteduse due to limited observations and large local-scalevariations of concentrations poor generalizationperformance moderatehigh accuracy and operational cost

an knowledgem requirese operational

Limited use due to limited observations and large local-scalevariations of concentrations poor generalizationperformance needs a substantial amount of observationaldata the computational complexity due to large numberof inappropriate rules

sociateders and

Does not perform well for highly non-linear systems

ns underare notn processesrt issuesata

Biases due to imperfect and missing model treatments andinaccuracies and uncertainties in meteorological andemissions predictions and model inputs Computationallyexpensive a need for a high speed computer systemhigh-level of expertise moderatehigh operational cost

ical models Bias correction may be effective only for systematic biasesand may hinder model improvement needs computationallyexpensive and complex a need for a high speed computersystem high-level of expertise high operational cost

er canf likelihood

Observational errors are not always accounted for inherentlimitations associated with individual ensemble modelmember accuracy sensitive to weighting factorscomputationally very expensive and complex a need forsupercomputer system high-level of expertise very highoperational cost

ecasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e248

high O3 furthermore PM species respond differently to tempera-ture changes Despite these limitations this method has beencommonly used in many RT-AQF programs as a primary method orcombined with other methods to screen initial RT-AQFs anddetermine whether the situation warrants a more quantitativeRT-AQF using more sophisticated methods

22 Parametricnon-parametric statistical approaches

Parametric or non-parametric statistical methods are based onthe fact that weather and air quality variables are statisticallyrelated It uses different functions (eg regression or trained neuralnetwork systems) to forecast pollutant concentrations that dependon external conditions The commonly-used methods includeclassification and regression trees (CART) (eg Burrows et al 1995)regression method (RM) (eg Cobourn and Hubbard 1999Cobourn 2007) artificial neural networks (ANNs) (eg Prybutoket al 2000 Peacuterez and Reyes 2006) fuzzy logic (FL) method (egJorquera et al 1998 Shad et al 2009) and Kalman filter (KF)method and its variants (eg extended Kalman filters (EKF)square-root filters ensemble Kalman filter (EnKF)) (eg van derWal and Jansen 2000 Zolghadri and Cazaurang 2006) CARTuses a decision tree and RM uses a regression equation to predictconcentrations based on values of various meteorological and airquality parameters ANNs and FL are artificial intelligence-basedtechniques ANNs use simplified mathematical models of brain-like systems to enable a structure to simulate intelligent behaviorin computers FL uses a form of algebra with a range of values interms of logical variables that can have continuous values between0 and 1 (false or true respectively) to represent varying degrees oftruthfulness and falsehood (ie partially true or false) The maindifference between FL and ANNs is that FL is a mathematical tool todeal with the uncertainties in human perception and reasoning itcan thus provide some insights into the model whereas ANNs arefast computation tools with learning and adaptive capabilitiesKalman filtering is an efficient recursive computational solution fortracking in real time a time-dependent state vector with noisyevolution equation and with noisy measurements

The statistical approaches usually require a large quantity ofhistorical measured data under a variety of atmospheric conditions(eg 2e3 yrs observed O3 or PM25 concentrations) They aregenerally more suitable for the description of complex site-specificrelations between concentrations of air pollutants and potentialpredictors and often have a higher accuracy as compared todeterministic models KF provides not only the response predic-tions but also their associated uncertainties through the calculationof their variances and covariances Some methods (eg ANNs andFL) can handle non-linearity whereas some cannot (eg KF) Theyhave several common drawbacks First they are usually confined tothe area and conditions present during the measurements andcannot be generalized to other regions with different chemical andmeteorological conditions Second they cannot predict concen-trations during periods of unusual emissions andor meteorologicalconditions that deviate significantly from the historical record(Stockwell et al 2002) Third they cannot capture the contributionof distant weather-dependent sources and circumstances favorablefor the formation of secondary pollutants because of the use of anaverage relationship between local concentrations and localatmospheric conditions Fourth the forecast accuracy typicallydepends on the skill of commonly-used meteorological predictorswhich usually neglect (eg morning inversion regional pollutanttransport) or use simplified parameterizations (eg turbulenceconvection and precipitation) for some meteorological processesthat are important to the evolution of air pollutants (Ryan 1995)Fifth the nature of statistical modeling does not enable better

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

understanding of chemical and physical processes These statisticalmodels provide neither the direct linkages between precursoremissions and resultant pollution nor the interrelationships amongmultiple pollutants (ie the interactions among pollutants thatmay potentially exacerbate one pollution problem while anotherproblem is being alleviated) Explicit treatments for such linkagesand interactions in RT-AQF models are essential to the enhance-ment of understanding of the physical-chemical system theimprovement of short- and long-term RT-AQF skill and thedevelopment of integrated emission control strategies for multi-pollutants

Various statistical approaches have been applied to RT-AQFsince the late 1970s These include multiple linear regression(eg Wolff and Lioy 1978 Stadlober et al 2008 Genc et al 2010)CART (eg Burrows et al 1995) ANNs (eg Peacuterez and Reyes 2006Li and Hassan 2010) FL systems (eg Shad et al 2009 Alhanafyet al 2010) nonlinear regression (NLR) (eg Cobourn andHubbard 1999) hybrid NLR (Cobourn 2007) and KF (egChenevez and Jensen 2001 Hoi et al 2008) These techniquesshowed some RT-AQF skill The best techniques that can handlenonlinearities and interactive relationships include CART ANNs FLand NLR which are more accurate than linear regression (egComrie 1997 Chaloulakou et al 2003) Some studies (eg Dutotet al 2007) showed that ANNs can outperform a CTM such asCHIMERE To overcome their limitations ANNs may be combinedwith other methods (eg Diaz-Robles et al 2008)

23 Advanced physically-based approaches

Deterministic models of air quality (also referred to as chemicaltransport models (CTMs) or AQMs) explicitly represent all majormeteorological physical and chemical processes that lead to theformation and accumulation of air pollutants by solution of theconservation equations for the mass of various species and trans-formation relationships among chemical species and physicalstates (Wayland et al 2002) The forecasting system requiresgridded meteorological fields emissions and chemical ICONs andBCONs Compared with statistical methods this approach hasseveral strengths First it is capable of forecasting temporally- andspatially-resolved concentrations under both typical and atypicalscenarios and pollutant concentrations in areas that are notmonitored Second it is physically-based and provides scientificunderstanding of pollutant processes thus it can address issuesthat cannot be handled by other forecasting methods such as long-range transport of air pollutants intricate interplay among mete-orology emissions and chemistry and changes in air quality underdifferent meteorological and emission scenarios Third it offersa moderate to high accuracy when all influential processes areaccurately represented in the model Fourth it does not requirea large quantity of measurement data There are several drawbacksFirst it demands sound knowledge of pollution sources andprocesses governing their evolution in the atmosphere making thedevelopment of such models rather difficult and costly This crucialknowledge is often limited andor insufficient and in some casesthe processes may be too complex to be easily represented ina model Approximations and simplifications are thus often madein CTMs Such imperfect representations of some atmosphericprocesses (eg SOA formation) or missing treatments in currentmodels often lead to inaccuracies and biases in forecastedconcentrations Second the accuracy of RT-AQF depends on theaccuracy of meteorological predictions emission estimates andother model inputs such as ICONs and BCONs Biases in suchpredictions and uncertainties in model inputs can be propagatedinto RT-AQFs Third it is computationally expensive and requiresa high-speed computer system with a large memory and disk

recasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e24 9

storage Despite these constraints the use of a coupledmeteorologyechemistry model for RT-AQF represents a significantadvancement in routine operational RT-AQFs and would greatlyenhance understanding of the underlying complex interplay ofmeteorology emission and chemistry Since 1990s RT-AQFsystems based on CTMs have been developed rapidly and arecurrently in operation in many countries including AustraliaCanada Japan US France Denmark Germany Norway UK SpainBelgium Turkey the Netherlands Chile and China Progress in CTMdevelopment and computing technologies has allowed dailyRT-AQFs using simplified (eg Vautard et al 2001b) or morecomprehensive 3-D CTMs such as offline-coupled (Tilmes et al2002 Honoreacute et al 2008 Schaap et al 2008 Chen et al 2008Baldasano et al 2008 McKeen et al 2009) and online-coupledmeteorologyechemistry models (eg Grell et al 2005 de Freitaset al 2005 Baklanov et al 2008 Flemming et al 2009 Chuanget al 2011) Model evaluation demonstrates that such a modelingapproach has skills consistent with or better than current statisticalforecasting tools (McHenry et al 2004 Manders et al 2009)

The use of offline-coupled meteorology and AQMs does notpermit the simulation of meteorologyechemistry feedbacks suchas aerosol feedbacks to radiation and photolysis which areimportant and may affect the next hourrsquos air quality and meteo-rological predictions (Grell et al 2004 2005 Zhang 2008 Zhanget al 2010b Baklanov 2010) Such systems may introduce biasesin RT-AQF For example Otte et al (2005) and Eder et al (2006)reported a poor performance of their offline coupled EtaCMAQmodeling system during cloudy periods due to neglecting aerosolfeedbacks to radiation and cloud formation processes Furthermoreatmospheric information at a time scale smaller than the outputtime interval of the meteorological model (eg 1-h) is lost in theoffline-coupled model systems (Grell et al 2004 Zhang 2008)Online-coupled models are increasingly used for applications inwhich the feedbacks may be important (eg locations with highfrequencies of clouds and large aerosol loadings) the local scalewind and circulation system change quickly and the coupledmeteorologyeair quality modeling is essential for accurate modelsimulations (eg RT-AQF or simulating the impact of future climatechange on air quality) Compared with offline AQMs online modelscan provide more realistic treatments of the atmosphere particu-larly in regions with a fast local circulation or a high aerosol loadingand cloud coverage where meteorology and radiation may bemodified by the presence of chemical species through variousfeedback mechanisms Therefore an RT-AQF system that is basedon an online-coupled meteorologyechemistry model can betterrepresent the real atmosphere and thus provide more accurateRT-AQFs

Given inherent limitations of CTMs a number of methods havebeen developed to combine their use with bias correction tech-niques such as statistical models or data assimilation systemsConceptually the approach of a CTM combined with a statisticalmodel is similar to the Model Output Statistics (MOS) (Glahn andLowry 1972) or a downscaling method (Wilby and Wigley 1997)that is often used to correct biases in weather forecasting atspecific sites In this method the linear regression model is firstdeveloped between a set of variables from the CTM and anobserved variable and then used to correct forecast biases fora given site (eg Wilson and Valeacutee 2003 Honoreacute et al 2008)This approach combines the strengths of CTMs and advancedobservation-based techniques thus it can have a greater accuracythan either method alone However the bias correction may notbe effective for random errors It may hinder the identification ofproblematic areas and model improvement needs In additionthis method is computationally expensive and complex andrequires a high-level of expertise As one of the earliest efforts

Please cite this article in press as Zhang Y et al Real-time air quality forEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

exploring this method Chenevez and Jensen (2001) applieda statistical after-treatment (ie an adaptive linear regressionmodel with an optimal state estimator algorithm) of the 3-DRT-AQF to adjust the next 48-h O3 forecasts at specific siteswhere real-time O3 measurements are available They showedthat this method can give higher correlation coefficients andsmaller biasesRMSEs than 3-D model raw outputs In a post-processing step to produce reanalysis pollution maps Honoreacuteet al (2008) trained MOS using past observations and thenapplied it to correct O3 forecasts from the Prevrsquoair RT-AQF systemand reported large improvements in specific areas and for high O3levels Konovalov et al (2009) combined deterministic andstatistical approaches for PM10 forecast in Europe by makingstatistical post-processing of forecasts from a CTM (ie CHIMERE)using PM10 monitoring data They found that this approachsignificantly improved the deterministic PM10 forecasts witha maximum reduction of RMSE by 50 and the maximumincrease of the coefficient of determination (r2) of more than 85

Several data assimilation techniques have been developed andused to improveoptimize ICONs BCONs emissions and meteo-rology in conjunction with a CTM for RT-AQF (eg Elbern andSchmidt 2001 Blond and Vautard 2004 Chai et al 2007Carmichael et al 2008 Wu et al 2008) Blond and Vautard (2004)used a statistical interpolation method that combines O3 predic-tions from CHIMERE with surface O3 measurements for RT-AQF ofO3 for western Europe and showed that the RMSE was reduced byw1 ppb (w30 relative to the raw forecasts) for 1e2 days forecastsDenby et al (2008) applied two data assimilation techniques forregional PM10 forecasts using a 3-D RT-AQF model in Europestatistical interpolation (SI) based on residual kriging after linearregression of the model and EnKF After assimilation the root-mean square discrepancy between analyses and the observationsfrom validation stations was reduced from 167 to 92 mg m3 usingSI and to 135 mg m3 using EnKF and r2 increased from 021 to 066using SI and to 041 using EnKF Comparedwith a classical statisticalapproach (eg MOS) the data assimilation approach allows the useof real-time or near real-time measurements for the same day ofthe PM10 forecasts

Another approach developed to overcome limitations of a CTMis probabilistic forecast (eg Wilczak et al 2006 Delle Monacheet al 2008) There is a growing trend to shift from purely deter-ministic forecasts to probabilistic forecast Given the inherentuncertainty of predictions of a CTM probabilistic forecasts have anadvantage because they provide an estimate of likelihood of anoccurrence of an event (Pagowski and Grell 2006) The simplestprobabilistic forecasts are associated with a dichotomous (eg YesNo) predictand Multi-category probabilistic forecasts can provideseveral categories of the predictand by using a consistent set ofprobability values Forecasts based on ensembles represent thethird type of probabilistic forecast in which the full distribution ofprobabilities rather than categories (lsquolsquomaybersquorsquo versus lsquolsquoyesrsquorsquo or lsquolsquonorsquorsquo)is approximated Ensembles can handle uncertainties but requireknowledge of model errors and uncertainties in model inputs anddifferent models Also they require a powerful computer systema high-level of expertise and very high operational cost Theregional RT-AQF with ten RT-AQF models over Europe within theMACC project (httpgemsecmwfintdproductsraq) representsthe operational state of the art for the generation of an ensemble offorecasts

Because of the increasing maturity of CTMs and availability ofcomputer resources many organizations in many countries cannow afford to develop RT-AQF systems that are based on 3-D CTMsSuch physically-based RT-AQF systems represent the state of thescience and will likely become prevalent forecasting approachesand tools and will be reviewed in the remaining sections

ecasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e248

high O3 furthermore PM species respond differently to tempera-ture changes Despite these limitations this method has beencommonly used in many RT-AQF programs as a primary method orcombined with other methods to screen initial RT-AQFs anddetermine whether the situation warrants a more quantitativeRT-AQF using more sophisticated methods

22 Parametricnon-parametric statistical approaches

Parametric or non-parametric statistical methods are based onthe fact that weather and air quality variables are statisticallyrelated It uses different functions (eg regression or trained neuralnetwork systems) to forecast pollutant concentrations that dependon external conditions The commonly-used methods includeclassification and regression trees (CART) (eg Burrows et al 1995)regression method (RM) (eg Cobourn and Hubbard 1999Cobourn 2007) artificial neural networks (ANNs) (eg Prybutoket al 2000 Peacuterez and Reyes 2006) fuzzy logic (FL) method (egJorquera et al 1998 Shad et al 2009) and Kalman filter (KF)method and its variants (eg extended Kalman filters (EKF)square-root filters ensemble Kalman filter (EnKF)) (eg van derWal and Jansen 2000 Zolghadri and Cazaurang 2006) CARTuses a decision tree and RM uses a regression equation to predictconcentrations based on values of various meteorological and airquality parameters ANNs and FL are artificial intelligence-basedtechniques ANNs use simplified mathematical models of brain-like systems to enable a structure to simulate intelligent behaviorin computers FL uses a form of algebra with a range of values interms of logical variables that can have continuous values between0 and 1 (false or true respectively) to represent varying degrees oftruthfulness and falsehood (ie partially true or false) The maindifference between FL and ANNs is that FL is a mathematical tool todeal with the uncertainties in human perception and reasoning itcan thus provide some insights into the model whereas ANNs arefast computation tools with learning and adaptive capabilitiesKalman filtering is an efficient recursive computational solution fortracking in real time a time-dependent state vector with noisyevolution equation and with noisy measurements

The statistical approaches usually require a large quantity ofhistorical measured data under a variety of atmospheric conditions(eg 2e3 yrs observed O3 or PM25 concentrations) They aregenerally more suitable for the description of complex site-specificrelations between concentrations of air pollutants and potentialpredictors and often have a higher accuracy as compared todeterministic models KF provides not only the response predic-tions but also their associated uncertainties through the calculationof their variances and covariances Some methods (eg ANNs andFL) can handle non-linearity whereas some cannot (eg KF) Theyhave several common drawbacks First they are usually confined tothe area and conditions present during the measurements andcannot be generalized to other regions with different chemical andmeteorological conditions Second they cannot predict concen-trations during periods of unusual emissions andor meteorologicalconditions that deviate significantly from the historical record(Stockwell et al 2002) Third they cannot capture the contributionof distant weather-dependent sources and circumstances favorablefor the formation of secondary pollutants because of the use of anaverage relationship between local concentrations and localatmospheric conditions Fourth the forecast accuracy typicallydepends on the skill of commonly-used meteorological predictorswhich usually neglect (eg morning inversion regional pollutanttransport) or use simplified parameterizations (eg turbulenceconvection and precipitation) for some meteorological processesthat are important to the evolution of air pollutants (Ryan 1995)Fifth the nature of statistical modeling does not enable better

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

understanding of chemical and physical processes These statisticalmodels provide neither the direct linkages between precursoremissions and resultant pollution nor the interrelationships amongmultiple pollutants (ie the interactions among pollutants thatmay potentially exacerbate one pollution problem while anotherproblem is being alleviated) Explicit treatments for such linkagesand interactions in RT-AQF models are essential to the enhance-ment of understanding of the physical-chemical system theimprovement of short- and long-term RT-AQF skill and thedevelopment of integrated emission control strategies for multi-pollutants

Various statistical approaches have been applied to RT-AQFsince the late 1970s These include multiple linear regression(eg Wolff and Lioy 1978 Stadlober et al 2008 Genc et al 2010)CART (eg Burrows et al 1995) ANNs (eg Peacuterez and Reyes 2006Li and Hassan 2010) FL systems (eg Shad et al 2009 Alhanafyet al 2010) nonlinear regression (NLR) (eg Cobourn andHubbard 1999) hybrid NLR (Cobourn 2007) and KF (egChenevez and Jensen 2001 Hoi et al 2008) These techniquesshowed some RT-AQF skill The best techniques that can handlenonlinearities and interactive relationships include CART ANNs FLand NLR which are more accurate than linear regression (egComrie 1997 Chaloulakou et al 2003) Some studies (eg Dutotet al 2007) showed that ANNs can outperform a CTM such asCHIMERE To overcome their limitations ANNs may be combinedwith other methods (eg Diaz-Robles et al 2008)

23 Advanced physically-based approaches

Deterministic models of air quality (also referred to as chemicaltransport models (CTMs) or AQMs) explicitly represent all majormeteorological physical and chemical processes that lead to theformation and accumulation of air pollutants by solution of theconservation equations for the mass of various species and trans-formation relationships among chemical species and physicalstates (Wayland et al 2002) The forecasting system requiresgridded meteorological fields emissions and chemical ICONs andBCONs Compared with statistical methods this approach hasseveral strengths First it is capable of forecasting temporally- andspatially-resolved concentrations under both typical and atypicalscenarios and pollutant concentrations in areas that are notmonitored Second it is physically-based and provides scientificunderstanding of pollutant processes thus it can address issuesthat cannot be handled by other forecasting methods such as long-range transport of air pollutants intricate interplay among mete-orology emissions and chemistry and changes in air quality underdifferent meteorological and emission scenarios Third it offersa moderate to high accuracy when all influential processes areaccurately represented in the model Fourth it does not requirea large quantity of measurement data There are several drawbacksFirst it demands sound knowledge of pollution sources andprocesses governing their evolution in the atmosphere making thedevelopment of such models rather difficult and costly This crucialknowledge is often limited andor insufficient and in some casesthe processes may be too complex to be easily represented ina model Approximations and simplifications are thus often madein CTMs Such imperfect representations of some atmosphericprocesses (eg SOA formation) or missing treatments in currentmodels often lead to inaccuracies and biases in forecastedconcentrations Second the accuracy of RT-AQF depends on theaccuracy of meteorological predictions emission estimates andother model inputs such as ICONs and BCONs Biases in suchpredictions and uncertainties in model inputs can be propagatedinto RT-AQFs Third it is computationally expensive and requiresa high-speed computer system with a large memory and disk

recasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e24 9

storage Despite these constraints the use of a coupledmeteorologyechemistry model for RT-AQF represents a significantadvancement in routine operational RT-AQFs and would greatlyenhance understanding of the underlying complex interplay ofmeteorology emission and chemistry Since 1990s RT-AQFsystems based on CTMs have been developed rapidly and arecurrently in operation in many countries including AustraliaCanada Japan US France Denmark Germany Norway UK SpainBelgium Turkey the Netherlands Chile and China Progress in CTMdevelopment and computing technologies has allowed dailyRT-AQFs using simplified (eg Vautard et al 2001b) or morecomprehensive 3-D CTMs such as offline-coupled (Tilmes et al2002 Honoreacute et al 2008 Schaap et al 2008 Chen et al 2008Baldasano et al 2008 McKeen et al 2009) and online-coupledmeteorologyechemistry models (eg Grell et al 2005 de Freitaset al 2005 Baklanov et al 2008 Flemming et al 2009 Chuanget al 2011) Model evaluation demonstrates that such a modelingapproach has skills consistent with or better than current statisticalforecasting tools (McHenry et al 2004 Manders et al 2009)

The use of offline-coupled meteorology and AQMs does notpermit the simulation of meteorologyechemistry feedbacks suchas aerosol feedbacks to radiation and photolysis which areimportant and may affect the next hourrsquos air quality and meteo-rological predictions (Grell et al 2004 2005 Zhang 2008 Zhanget al 2010b Baklanov 2010) Such systems may introduce biasesin RT-AQF For example Otte et al (2005) and Eder et al (2006)reported a poor performance of their offline coupled EtaCMAQmodeling system during cloudy periods due to neglecting aerosolfeedbacks to radiation and cloud formation processes Furthermoreatmospheric information at a time scale smaller than the outputtime interval of the meteorological model (eg 1-h) is lost in theoffline-coupled model systems (Grell et al 2004 Zhang 2008)Online-coupled models are increasingly used for applications inwhich the feedbacks may be important (eg locations with highfrequencies of clouds and large aerosol loadings) the local scalewind and circulation system change quickly and the coupledmeteorologyeair quality modeling is essential for accurate modelsimulations (eg RT-AQF or simulating the impact of future climatechange on air quality) Compared with offline AQMs online modelscan provide more realistic treatments of the atmosphere particu-larly in regions with a fast local circulation or a high aerosol loadingand cloud coverage where meteorology and radiation may bemodified by the presence of chemical species through variousfeedback mechanisms Therefore an RT-AQF system that is basedon an online-coupled meteorologyechemistry model can betterrepresent the real atmosphere and thus provide more accurateRT-AQFs

Given inherent limitations of CTMs a number of methods havebeen developed to combine their use with bias correction tech-niques such as statistical models or data assimilation systemsConceptually the approach of a CTM combined with a statisticalmodel is similar to the Model Output Statistics (MOS) (Glahn andLowry 1972) or a downscaling method (Wilby and Wigley 1997)that is often used to correct biases in weather forecasting atspecific sites In this method the linear regression model is firstdeveloped between a set of variables from the CTM and anobserved variable and then used to correct forecast biases fora given site (eg Wilson and Valeacutee 2003 Honoreacute et al 2008)This approach combines the strengths of CTMs and advancedobservation-based techniques thus it can have a greater accuracythan either method alone However the bias correction may notbe effective for random errors It may hinder the identification ofproblematic areas and model improvement needs In additionthis method is computationally expensive and complex andrequires a high-level of expertise As one of the earliest efforts

Please cite this article in press as Zhang Y et al Real-time air quality forEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

exploring this method Chenevez and Jensen (2001) applieda statistical after-treatment (ie an adaptive linear regressionmodel with an optimal state estimator algorithm) of the 3-DRT-AQF to adjust the next 48-h O3 forecasts at specific siteswhere real-time O3 measurements are available They showedthat this method can give higher correlation coefficients andsmaller biasesRMSEs than 3-D model raw outputs In a post-processing step to produce reanalysis pollution maps Honoreacuteet al (2008) trained MOS using past observations and thenapplied it to correct O3 forecasts from the Prevrsquoair RT-AQF systemand reported large improvements in specific areas and for high O3levels Konovalov et al (2009) combined deterministic andstatistical approaches for PM10 forecast in Europe by makingstatistical post-processing of forecasts from a CTM (ie CHIMERE)using PM10 monitoring data They found that this approachsignificantly improved the deterministic PM10 forecasts witha maximum reduction of RMSE by 50 and the maximumincrease of the coefficient of determination (r2) of more than 85

Several data assimilation techniques have been developed andused to improveoptimize ICONs BCONs emissions and meteo-rology in conjunction with a CTM for RT-AQF (eg Elbern andSchmidt 2001 Blond and Vautard 2004 Chai et al 2007Carmichael et al 2008 Wu et al 2008) Blond and Vautard (2004)used a statistical interpolation method that combines O3 predic-tions from CHIMERE with surface O3 measurements for RT-AQF ofO3 for western Europe and showed that the RMSE was reduced byw1 ppb (w30 relative to the raw forecasts) for 1e2 days forecastsDenby et al (2008) applied two data assimilation techniques forregional PM10 forecasts using a 3-D RT-AQF model in Europestatistical interpolation (SI) based on residual kriging after linearregression of the model and EnKF After assimilation the root-mean square discrepancy between analyses and the observationsfrom validation stations was reduced from 167 to 92 mg m3 usingSI and to 135 mg m3 using EnKF and r2 increased from 021 to 066using SI and to 041 using EnKF Comparedwith a classical statisticalapproach (eg MOS) the data assimilation approach allows the useof real-time or near real-time measurements for the same day ofthe PM10 forecasts

Another approach developed to overcome limitations of a CTMis probabilistic forecast (eg Wilczak et al 2006 Delle Monacheet al 2008) There is a growing trend to shift from purely deter-ministic forecasts to probabilistic forecast Given the inherentuncertainty of predictions of a CTM probabilistic forecasts have anadvantage because they provide an estimate of likelihood of anoccurrence of an event (Pagowski and Grell 2006) The simplestprobabilistic forecasts are associated with a dichotomous (eg YesNo) predictand Multi-category probabilistic forecasts can provideseveral categories of the predictand by using a consistent set ofprobability values Forecasts based on ensembles represent thethird type of probabilistic forecast in which the full distribution ofprobabilities rather than categories (lsquolsquomaybersquorsquo versus lsquolsquoyesrsquorsquo or lsquolsquonorsquorsquo)is approximated Ensembles can handle uncertainties but requireknowledge of model errors and uncertainties in model inputs anddifferent models Also they require a powerful computer systema high-level of expertise and very high operational cost Theregional RT-AQF with ten RT-AQF models over Europe within theMACC project (httpgemsecmwfintdproductsraq) representsthe operational state of the art for the generation of an ensemble offorecasts

Because of the increasing maturity of CTMs and availability ofcomputer resources many organizations in many countries cannow afford to develop RT-AQF systems that are based on 3-D CTMsSuch physically-based RT-AQF systems represent the state of thescience and will likely become prevalent forecasting approachesand tools and will be reviewed in the remaining sections

ecasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e24 9

storage Despite these constraints the use of a coupledmeteorologyechemistry model for RT-AQF represents a significantadvancement in routine operational RT-AQFs and would greatlyenhance understanding of the underlying complex interplay ofmeteorology emission and chemistry Since 1990s RT-AQFsystems based on CTMs have been developed rapidly and arecurrently in operation in many countries including AustraliaCanada Japan US France Denmark Germany Norway UK SpainBelgium Turkey the Netherlands Chile and China Progress in CTMdevelopment and computing technologies has allowed dailyRT-AQFs using simplified (eg Vautard et al 2001b) or morecomprehensive 3-D CTMs such as offline-coupled (Tilmes et al2002 Honoreacute et al 2008 Schaap et al 2008 Chen et al 2008Baldasano et al 2008 McKeen et al 2009) and online-coupledmeteorologyechemistry models (eg Grell et al 2005 de Freitaset al 2005 Baklanov et al 2008 Flemming et al 2009 Chuanget al 2011) Model evaluation demonstrates that such a modelingapproach has skills consistent with or better than current statisticalforecasting tools (McHenry et al 2004 Manders et al 2009)

The use of offline-coupled meteorology and AQMs does notpermit the simulation of meteorologyechemistry feedbacks suchas aerosol feedbacks to radiation and photolysis which areimportant and may affect the next hourrsquos air quality and meteo-rological predictions (Grell et al 2004 2005 Zhang 2008 Zhanget al 2010b Baklanov 2010) Such systems may introduce biasesin RT-AQF For example Otte et al (2005) and Eder et al (2006)reported a poor performance of their offline coupled EtaCMAQmodeling system during cloudy periods due to neglecting aerosolfeedbacks to radiation and cloud formation processes Furthermoreatmospheric information at a time scale smaller than the outputtime interval of the meteorological model (eg 1-h) is lost in theoffline-coupled model systems (Grell et al 2004 Zhang 2008)Online-coupled models are increasingly used for applications inwhich the feedbacks may be important (eg locations with highfrequencies of clouds and large aerosol loadings) the local scalewind and circulation system change quickly and the coupledmeteorologyeair quality modeling is essential for accurate modelsimulations (eg RT-AQF or simulating the impact of future climatechange on air quality) Compared with offline AQMs online modelscan provide more realistic treatments of the atmosphere particu-larly in regions with a fast local circulation or a high aerosol loadingand cloud coverage where meteorology and radiation may bemodified by the presence of chemical species through variousfeedback mechanisms Therefore an RT-AQF system that is basedon an online-coupled meteorologyechemistry model can betterrepresent the real atmosphere and thus provide more accurateRT-AQFs

Given inherent limitations of CTMs a number of methods havebeen developed to combine their use with bias correction tech-niques such as statistical models or data assimilation systemsConceptually the approach of a CTM combined with a statisticalmodel is similar to the Model Output Statistics (MOS) (Glahn andLowry 1972) or a downscaling method (Wilby and Wigley 1997)that is often used to correct biases in weather forecasting atspecific sites In this method the linear regression model is firstdeveloped between a set of variables from the CTM and anobserved variable and then used to correct forecast biases fora given site (eg Wilson and Valeacutee 2003 Honoreacute et al 2008)This approach combines the strengths of CTMs and advancedobservation-based techniques thus it can have a greater accuracythan either method alone However the bias correction may notbe effective for random errors It may hinder the identification ofproblematic areas and model improvement needs In additionthis method is computationally expensive and complex andrequires a high-level of expertise As one of the earliest efforts

Please cite this article in press as Zhang Y et al Real-time air quality forEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

exploring this method Chenevez and Jensen (2001) applieda statistical after-treatment (ie an adaptive linear regressionmodel with an optimal state estimator algorithm) of the 3-DRT-AQF to adjust the next 48-h O3 forecasts at specific siteswhere real-time O3 measurements are available They showedthat this method can give higher correlation coefficients andsmaller biasesRMSEs than 3-D model raw outputs In a post-processing step to produce reanalysis pollution maps Honoreacuteet al (2008) trained MOS using past observations and thenapplied it to correct O3 forecasts from the Prevrsquoair RT-AQF systemand reported large improvements in specific areas and for high O3levels Konovalov et al (2009) combined deterministic andstatistical approaches for PM10 forecast in Europe by makingstatistical post-processing of forecasts from a CTM (ie CHIMERE)using PM10 monitoring data They found that this approachsignificantly improved the deterministic PM10 forecasts witha maximum reduction of RMSE by 50 and the maximumincrease of the coefficient of determination (r2) of more than 85

Several data assimilation techniques have been developed andused to improveoptimize ICONs BCONs emissions and meteo-rology in conjunction with a CTM for RT-AQF (eg Elbern andSchmidt 2001 Blond and Vautard 2004 Chai et al 2007Carmichael et al 2008 Wu et al 2008) Blond and Vautard (2004)used a statistical interpolation method that combines O3 predic-tions from CHIMERE with surface O3 measurements for RT-AQF ofO3 for western Europe and showed that the RMSE was reduced byw1 ppb (w30 relative to the raw forecasts) for 1e2 days forecastsDenby et al (2008) applied two data assimilation techniques forregional PM10 forecasts using a 3-D RT-AQF model in Europestatistical interpolation (SI) based on residual kriging after linearregression of the model and EnKF After assimilation the root-mean square discrepancy between analyses and the observationsfrom validation stations was reduced from 167 to 92 mg m3 usingSI and to 135 mg m3 using EnKF and r2 increased from 021 to 066using SI and to 041 using EnKF Comparedwith a classical statisticalapproach (eg MOS) the data assimilation approach allows the useof real-time or near real-time measurements for the same day ofthe PM10 forecasts

Another approach developed to overcome limitations of a CTMis probabilistic forecast (eg Wilczak et al 2006 Delle Monacheet al 2008) There is a growing trend to shift from purely deter-ministic forecasts to probabilistic forecast Given the inherentuncertainty of predictions of a CTM probabilistic forecasts have anadvantage because they provide an estimate of likelihood of anoccurrence of an event (Pagowski and Grell 2006) The simplestprobabilistic forecasts are associated with a dichotomous (eg YesNo) predictand Multi-category probabilistic forecasts can provideseveral categories of the predictand by using a consistent set ofprobability values Forecasts based on ensembles represent thethird type of probabilistic forecast in which the full distribution ofprobabilities rather than categories (lsquolsquomaybersquorsquo versus lsquolsquoyesrsquorsquo or lsquolsquonorsquorsquo)is approximated Ensembles can handle uncertainties but requireknowledge of model errors and uncertainties in model inputs anddifferent models Also they require a powerful computer systema high-level of expertise and very high operational cost Theregional RT-AQF with ten RT-AQF models over Europe within theMACC project (httpgemsecmwfintdproductsraq) representsthe operational state of the art for the generation of an ensemble offorecasts

Because of the increasing maturity of CTMs and availability ofcomputer resources many organizations in many countries cannow afford to develop RT-AQF systems that are based on 3-D CTMsSuch physically-based RT-AQF systems represent the state of thescience and will likely become prevalent forecasting approachesand tools and will be reviewed in the remaining sections

ecasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e2410

3 Overview of major 3-D RT-AQF models on regional andglobal scales

31 Model systems components and application scales

A number of 3-D global and regional models have beendeployed for RT-AQF Table 3 summarizes 9 global and 36 regionalRT-AQF models that are currently used in Australia North AmericaSouth America Europe and Asia in terms of component models(ie meteorological models air quality models microscalemodels)spatial scale and coupling between meteorology and chemistryThe full names of models and associated organizations are providedin a list of acronyms and symbols in an Appendix in the supple-mentary material Among them four global models (ie GEM-AQLMDzt-INCA ECHAM5 and ECMWF-IFS-CTMs) and four regionalmodels (ie WRFChem WRFChemeMADRID GEM-MACH15 andCFORS) are online-coupled models

For global models meteorological fields are produced by rean-alysis data such as National Centers for Environmental Prediction(NCEP) or general circulation models such as GEM ECMWFIFSLMDzt ECHAM5 In unified online-coupled models such asGEM-AQ and ECHAM5 a chemistry module is incorporated intoa GCM In offline or separate online-coupled models an AQM isdriven by data from meteorological reanalysis (eg GOCART) orused in conjunction with a GCM (eg MOCAGE MOZART-3 TM5)respectively Among the 9 global RT-AQFmodels ECMWFIFS-CTMsthat is being developed by ECMWF offers an integrated flexibleadvanced global forecastingmodeling tool with data assimilation ofsatellite observations In ECMWFIFS-CTMs the IFS is coupled viaa coupler software to one of the global CTMs MOCAGE MOZART-3or TM5 for global forecast and assimilation of reactive chemicalspecies The selection of multiple CTMs and their ensemble resultsprovides a range and an indication of the robustness of the fore-casts The coupled system can directly utilize the IFS 4D-Var algo-rithm to assimilate atmospheric observations (Flemming et al2009) Three out of 9 global models have been applied to regionaldomains for RT-AQF While GEM-AQ and MOCAGE were directlydownscaled to a regional domain at a finer horizontal grid resolu-tion (Neary et al 2007 Rouiumll et al 2009) the global-regionalRT-AQF model system (GR-RT-AQF) was used to provide ICONsand BCONs to drive a regional unified online-coupled model WRFChem (Takigawa et al 2007)

For regional models many of them use the most popularmeteorological models such as MM5 WRF and ECMWFIFS ManyAQMs (eg CHIMERE and PolyphemusPolair3D) can be drivenby different meteorological models The most commonly-usedregional AQMs include CMAQ WRFChem and CHIMERE WRFChem (Grell et al 2005) and WRFChemeMADRID (Zhang et al2010a) include the most coupled meteorological microphysicalchemical and radiative processes and allow the simulation ofaerosol direct semi-direct and indirect effects thus representingthe state-of-the-science online-coupled regional RT-AQF modelsSince its first release in 2002 WRFChem has been further devel-oped and improved by many researchers in the community (egFast et al 2006 Zhang et al 2010a Shrivastava et al 2010) andincreasingly applied to many regions of the world (eg Tie et al2009 Fast et al 2009 Misenis and Zhang 2010 Zhang et al2010ab 2012 Li et al 2011) Compared to WRFChem WRFChemeMADRID includes two additional gas-phase mechanisms(ie the 2005 version of carbon bond gas-phase mechanism (CB05)and the 1999 Statewide Air Pollution Research Center gas-phasemechanism (SAPRC-99)) one aerosol module (ie the model ofaerosol dynamics reaction ionization and dissolution (MADRID))one aerosol activation scheme (ie Fountoukis and Nenes 2005)and several nucleation algorithms (eg Sihto et al 2006

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

Merikanto et al 2007 Yu 2010) CB05 and SAPRC-99 are coupledwith MADRID and two existing aerosol modules (ie the ModalAerosol Dynamics Model for Europethe Secondary Organic AerosolModel (MADESORGAM) and Model for Simulating Aerosol Inter-actions and Chemistry (MOSAIC)) as well as the CMU aqueous-phase chemistry in WRFChemeMADRID (Zhang et al 2010a c2012) WRFChemeMADRID has been applied retrospectivelyover the continental US and its sub-regions (eg Zhang et al2010a 2012) and Europe (Zhang et al 2011) and also for RT-AQFat a horizontal resolution of 12-km over the southeastern USsince summer 2009 (Chuang et al 2011) Similar to the globalECMWF-IFS-CTMs the French national air quality forecasting andmonitoring system Prevrsquoair consists of three models CHIMEREMOCAGE and PolyphemusPolair3D (httpwwwPrevrsquoairorgRouiumll et al 2009) which allows ensemble RT-AQFs Since spring2003 Prevrsquoair has been in operational forecasting Case studies ofRT-AQFs with WRFChemeMADRID in the eastern US and Poly-phemus in Europe are given in Section 7 Several RT-AQF systemsare being used by regional air quality agencies (eg the Associa-tions agreacuteeacutees de surveillance de la qualiteacute de lrsquoair AASQAs) inFrance (eg AIRPARIF for the Paris region) those regional RT-AQFsystems typically use CHIMERE

Among 36 regionalmodels five are suitable for urbanlocal scaleapplications at a spatial resolution of 1 km or less These modelsinclude THOR the OPerational version of the AtmosphericNumerical pollution model for urban and regional Areas (OPANA)and four Urban Air Quality Information and Forecasting Systems(UAQIFS) models (ie UAQIFS e Norway UAQIFS-Finland UAQIFS-Italy2 and UAQIFS-Denmark) THOR includes the backgroundurban model (BUM) the Operational Street Pollution Model(OSPM) and the Danish Rimpuff and Eulerian Accidental releaseModel (DREAM) OPANA includes the microscale air qualitymodeling system (MICROSYS) to forecast air concentrations forurban areas with street level details at a 5e10 m spatial resolutionand up to 200e300m in height over themaximumbuilding heightsin one 1-km grid cell that is nested in a regional simulation domain(San Joseacute et al 2006 2009) The UAQIFS models were developed aspart of the Integrated Systems for Forecasting Urban MeteorologyAir Pollution and Population Exposure e UAQIFS (FUMAPEX-UAQIFS) project sponsored by the EU (httpfumapexdmidkBaklanov 2006) They include six separate UAQIFS that are furtherdeveloped and applied in six cities in Europe While these UAQIFSuse advanced meteorological models such as HIRLAM and RAMSthe level of sophistication in AQMs varies from the simplestdispersion model with no or very simple chemistry to the mostcomplex 3-DAQMs such as FARM and CAMx Their common featurelies in that they integrate the latest developments in urban mete-orology air quality and population exposure modeling via anoffline coupling approach to enhance the modelrsquos forecastingcapability in urban areas The enhanced modeling capabilitiesinclude one ormore areas in urban RT-AQF urbanmanagement andplanning public health assessment and exposure prediction andurban emergency preparedness One example application of anUAQIFS is given in Section 73

32 Model input output and horizontal grid resolution

Table A2 summarizes model inputs and Table 4 summarizesforecast products and horizontal grid resolutions that are used inthese models For global models anthropogenic emissions aremostly based on well known global emission inventories (eg theGlobal Emissions Inventory Activity (GEIA) and the EmissionDatabase for Global Atmospheric Research (EDGAR)) and in somecases based on emission inventories developed in internationalcollaborative projects (eg GEMS- or Megacities Emissions urban

recasting part I History techniques and current status Atmospheric

Table 3Component Models Spatial Scale and Coupling of Meteorology and Chemistry in Major RT-AQF Model Systemsa

CountryOrganization Model System MeteorologicalModel (MetM)

Air QualityModel (AQM)

MicroscaleModels

Scale MetM -AQMCoupling

References

Global ModelsUSNASA GOCART GEOS DAS GOCART None Global Offline Chin et al (2003) httpacdbextgsfcnasagov

PeopleChingocartinfohtmlCanadaYork Univ GEM-AQ GEM AQ None GlobalRegional Unified

OnlineNeary et al (2007) and Kaminski et al (2008)httpecoforecasteuindexphpidfrac142

FranceLMD LMDzt-INCA ECMWFIFSLMDzt v40with nudgedwith NCEP

INCA v3 None Global SeparateOnline

Hauglustaine et al (2004) and Folberth et al (2006)httpwwwlsceincaceafrwelcome_real_timehtml

FranceMFCNRM MOCAGE ARPEGE (global)ALADIN(regional)

MOCAGE None Global Regional Offline Rouiumll et al (2009) wwwprevairorg

USNCARGermanyMPIC

MATCH-NCAR NCEPNCAR MATCH-NCAR None Global Offline Rasch et al (1997) and Lawrence et al (2003)httpwwwmpchmainzmpgdewlawrenceforecastshtml

GermanyMPIM ECHAM5 ECHAM5 ECHAM5 None Global UnifiedOnline

Roeckner et al (2006) and Zhang et al (2010)httpwwwmpimetmpgdeensciencemodelsechamhtml

Norway NIARFLEXPART FLEXPART ECMWF NCEP FLEXPART None Global Offline Forster et al (2004) and Stohl et al (2005)httptransportnilunoflexpart

UKECMWF ECMWF-IFS-CTMs IFS MOZART-3TM5 or MOCAGE

None Global Separateonline

Flemming et al (2009) and Mangold et al (2011)httpgemsecmwfintdproductsgrgrealtime

Japan-FRCGC GR-AQF CCSRNIESFRCGCatmosphericGCM (global)WRF (regional)

CHASER (global)WRFChem (regional)

None GlobalRegional Offline forglobalregionalUnified onlinefor regional

Takigawa et al (2007)

Regional ModelsAustraliaCSIRO AAQFS LAPS CSIROrsquos CTM None Regional Offline Manins (2002) and Cope et al (2004)

wwwepavicgovauAirAAQFSUSBAMS MAQSIP-RT CMAQ BAMS-MM5 WRF MAQSIP CMAQ None Regional Offline McHenry et al (2004) and McKeen et al (2009)

wwwbaronamscomprojectsSECMEPindexhtmlUSWSU AIRPACT3 MM5 CALGRID CMAQ None Regional Offline Vaughan et al (2004) and Chen et al (2008)

wwwlarwsueduairpact-3USSUNY-Albany AQFMS SKIRONEta WRF

(NMM and ARW)CAMx CMAQ None Regional Offline Hogrefe et al (2007) Cai et al (2008) and

Doraiswamy et al (2009)asrcalbanyeduresearchaqf

USUI STEM-2K3 MM5 WRF STEM None Regional Offline Carmichael et al (2003) nascgreruiowaeduUSNOAA EPA NAQFC Eta WRF (NMM) CMAQ None Regional Offline Otte et al (2005) Yu et al (2007 2008)

Lee et al (2008) and Eder et al (2010)wwwweathergovaqairnowgov

USNOAA WRFChem WRF (ARW) WRFChem None Regional Unified Online McKeen et al (2005 2007) rucnoaagovwrfWG11_RT

USNCSU WRFChem-MADRID WRF (ARW) WRFChem None Regional UnifiedOnline

Zhang et al (2010a) and Chuang et al (2011)wwwmeasncsueduaqforecastingReal_Timehtml

CanadaEnviron Canada GEM-AURAMS Regional GEM AURAMS None Regional Offline McKeen et al (2005 2007 2009)wwwmscsmcecgccaresearchicarttaurams_ehtml

CanadaEnviron Canada GEM-CHRONOS Regional GEM CHRONOS None Regional Offline Pudykiewicz and Koziol (2001) andMcKeen et al (2005 2007 2009)wwwmsc-smcecgccaaq_smogchronos_ecfm

CanadaEnviron Canada GEM-MACH15 GEM chemistry fromAURAMS

None Regional UnifiedOnline

Talbot (2007)

FranceAIRPARIFb ESMERALDA MM5 CHIMERE None Regional Offline Vautard et al (2001ab)httpwwwesmeralda-webfr

(continued on next page)

YZhanget

alAtm

osphericEnvironm

entxxx

(2012)1e24

11

Pleasecite

thisarticle

inpress

asZhangYetalReal-tim

eair

qualityforecastingpart

IHistorytechniquesand

currentstatusA

tmospheric

Environment

(2012)httpdxdoiorg101016jatmosenv201206031

Table 3 (continued )

CountryOrganization Model System MeteorologicalModel (MetM)

Air QualityModel (AQM)

MicroscaleModels

Scale MetM -AQMCoupling

References

FranceINERIS CHIMERE MM5 WRFECWMFIFS

CHIMERE None Regional Offline Vautard et al (2001a) and Rouiumll et al (2009)wwwprevairorg

FranceCEREA POLYPHEMUS ECMWF MM5 WRF Polair3D None Regional Offline Mallet and Sportisse 2006 Mallet et al 2007Debry et al 2007 Sartelet et al 2007cereaenpcfrpolyphemus

DenmarkDMU-ATMI THOR The US NCEP Eta DEOM BUM OSPM DREAM Regional Offline Brandt et al (2001) and Tilmes et al (2002)luftdmudkAtmosphericEnvironmentThorintro_ukhtml thordmudk

DenmarkDMI DACFOS HIRLAM MOON CAMxEnviro- HIRLAM

M2UE Regional Offlineonline Chenevez and Jensen (2001) Baklanov et al (2011)and references therein

FUMAPEX UAQIFSc 1 UAQIFS-Norway2 UAQIFS-Finland3 UAQIFS-Spain4 UAQIFS-Italy15 UAQIFS-Italy26 UAQIFS-Denmark

1 HIRLAM2 HIRLAM3 RAMS4 RAMS5 LAMI6 HIRLAM

1 AirQUIS (dispersion)2 CAR-FMI (dispersion)3 CAMx (O3 only)4 FARM5 NINFA-OPPIOADAM6 DERMA-ARGOS

Some include populationexposure models someinclude urban dispersionstatistical models

Regional local Offline Baklanov (2006) and Baklanov et al (2007)

GermanyFRIUUK EURAD MM5 EURAD-IM None Regional Offline Elbern et al (2010) wwweuraduni-koelndeNorwayMET-NO EMEP-Unified ECMWFIFS Unified EMEP-CWF None Regional Offline Valdebenito and Benedictow (2010) wwwemepintSwedenSMHI MATCH ECMWFIFS HIRLAM MATCH None Regional Offline Robertson (2010) httpwwwsmhiseen

ResearchResearchdepartments Air-qualityNetherlandsTNO LOTOS-EUROS Archived analyses

or ECMWFLOTOS-EUROS None Regional Offline Schaap et al (2008) and Manders et al (2009)

httpwwwlotoseuros nlFinlandFMI SILAM v454 ECMWFIFS

HIRLAM WRFSILAM None Regional Offline Sofiev et al (2006) Sofiev and Vira (2010) and

Kukkonen et al (2011) silamfmifihttppandoramengauthgrmdsshowlongphpidfrac14168d_1 8

UKAEA WRFCMAQ WRF CMAQ None Regional offline Fraser et al (2010)wwwairqualitycoukuk_forecastingapfuk_homephp

UKMet Office NAME-III ECMWF NAME-III None Regional local Offline wwwairqualitycoukuk_forecastingapfuk_homephpSpainTUM OPANA v40 MM5 CMAQ MICROSYS Regional local Offline San Joseacute et al (2006

2009) articolmafiupmesSpainBSC-CNS CALIOPE WRF MM5 CMAQ CHIMERE

DREAMNone Regional Offline Baldasano et al (2008)

wwwbscescaliopeGreeceUA SKIRONTAPM SKIRONdust CAMx v431 None Regional Offline Kallos et al (2007 2009)

Mitsakou et al (2008) andSpyrou et al (2010) forecastuoagr

ItalyCETEMPS ForeChem MM5 CHIMERE None Regional Offline Curci (2010)pumpkinaquilainfnitforechem

ChinaIAP-CAS EMS-Beijing MM5 NAQPMS CMAQ CAMx None Regional Offline Wang et al (2009)JapanKyushu

UniversityCFORS RAMS Parameterized chemical

tracers in RAMSNone Regional Unified Online Uno et al (2003) Carmichael et al (2003) and

Hadley et al (2007)ChileMeteo Chile POLYPHEMUS MM5 Polair3D None Regional Offline Mallet et al (2007) and

Mallet and Sportisse (2006) cereaenpcfrpolyphemus

a The list of acronyms is provided in Appendixb There are several air quality systems used by regional air quality agencies (Association agreacuteeacutees de surveillance de la qualiteacute de lrsquoair (AASQAs)) and only the one for the Paris region is given here as an examplec FUMAPEX includes six UAQIFS model systems implemented in six cities in five countries as labeled 1e6 under Model System Meteorological Model (MetM) and Air Quality Model (AQM) UAQIFS-Italy1 refers to the UAQIFS

applied for the city of Turin Italy and UAQIFS-Italy2 refers to the UAQIFS applied for the city of Bologna Italy

YZhanget

alAtm

osphericEnvironm

entxxx

(2012)1e24

12Pleasecite

thisarticle

inpress

asZhangYetalReal-tim

eair

qualityforecastingpartIH

istorytechniquesandcurrent

statusAtm

osphericEnvironm

ent(2012)httpdxdoiorg101016jatm

osenv201206031

Table 4Forecasting products and horizontal resolution of major RT-AQF model systems

Organization Model System AQF Products Horizontal Grid Resolutions

Global ModelsUSNASA GOCART 24e96 h AQF of concof PM composition total

PM extinction AOD due to SO24 dust BC OC sea-salt PM

25 25 or 1 1

CanadaYork Univ GEM-AQ 5e10 days on global scale and 24e72 h on regionalscale for CO SO2 NO2 and O3

Global 15 15 or 09 09 regionala rotated variable resolution

FranceLMD LMDzt-INCA 72 h RT AQF of O3 and PM10 25 375

The Meacuteteacuteo France MOCAGE 72-h forecasts of O3 NO NO2 CO SO2 and PM10 2 2 at a global scale and 05 05 amp01 015 at a regional scale

USNCAR GermanyMPIC MATCH-NCAR 24-h forecasts of O3 NOx CO OH HNO3 PAN amp VOCs 1 1

GermanyMPIM ECHAM5 24-h forecasts of CO 28 28

Norway NIARFLEXPART 72-h RT-AQF of CO NOx and SO2 1 1

UKECMWF ECMWF-IFS-CTM 72-h RT-AQF of O3 and CO CO2 PM25 and composition AOD MOZART-3 1875 1875 TM5 3 2MOCAGE 2 2

Japan-FRCGC GR-AQF 15-h forecasts for O3 Global 28 Japan 15- and 5-kmRegional ModelsAustralian CSIRO AAQFS 24e36 h AQF of 25 species (eg O3 NO2 CO SO2

VOCs PM1 PM25 PM10 visibility and air toxics)1e5 km

USBAMS MAQSIP-RT amp CMAQ CO ethylene NOy O3 PM25 and PM25 composition 45- 15- and 5- kmUSWSU AIRPACT3 24e64 h forecasts of O3 and PM25 12-kmUSSUNY-Albany AQFMS 24-h forecasts of CO NOxNOy SO2 VOC HONO HOx O3 PM25 36- and 12-kmUSUI STEM-2K3 48-h AQF of CO ethylene NOy O3 PM25 and PM25 composition 60- 12- and 4-kmUS NOAAEPArsquos NAQFC 48-h AQF of CO ethylene NOy O3 PM25 and PM25 composition 12-kmUS NOAAESRL WRFChem 48-h AQF of CO ethylene NOy O3 PM25 and PM25 composition 12- 27- 36-kmUSNCSU WRFChem-MADRID 48-h AQF of O3 PM25 and PM10 12-kmCanadaEC GEM-AURAMS 48-h AQF of CO ethylene NOy O3 PM25 and PM25 composition 42- and 28-kmCanadaEC GEM-CHRONOS 48-h AQF of CO ethylene SO2 NOy O3 PM25 PM10 composition 21-kmCanadaEC GEM-MACH15 48-h forecasts of CO SO2 NO2 O3 PM25 and PM10 15-kmFranceAIRPARIFa 72-h forecasts of O3 NO2 PM10 PM25 Urban 3-km regional 9-kmFrance INERIS CHIMERE 72-h forecasts of O3 NO NO2 CO SO2 PM25 and PM10 01 05 50-kmFranceCEREA POLYPHEMUS 72-h RT-AQF of O3 NO2 PM25 and PM10 05

Denmark DMU-ATMI THOR 72-h forecasts of O3 NO2 NOx CO Local 1-km regional 50-kmDenmark DACFOS 48-h forecast for GEMS and 36-h for DK of O3 NO NO2

CO SO2 PM25 and PM10

02 for GEMS domain and 004 forinner domain

FUMAPEX UAQIFS 24e72-h AQF Products vary with the AQF models typicallyinclude SO2 NOx CO O3 benzene PM25 PM10 radiative tracerspecies or a subset of them Some forecasts human exposuresto ambient air and radioactive pollutants

Outer grid is 12e16 km innersingle grid isat 1e10 km most with 2e3 levels of nesting

GermanyFRIUUK EURAD 72-h AQF of O3 NO NO2 CO C6H6 SO2 PM25 PM10 125- 15- and 5- kmNorwayMET-NO EMEP-Unified 72-h AQF of O3 NO NO2 CO SO2 Rn-222 PM25 PM10 025

SwedenSMHI MATCH 72-h forecasts of O3 NO NO2 CO SO2 PM25 PM10 05

NetherlandsTNO LOTOS-EUROS 72-h RT-AQF of O3 NO2 SO2 CO PM25 PM10 025 (lon) 0125 (lat) or 05 025

FinlandFMI SILAM v454 54 to 72-h forecasts of O3 NO NO2 CO SO2 PM25 PM10 Rn120-h forecasts of pollen

02

UKAEA WRFCMAQ 48-h forecasting of O3 CO SO2 NO2 PM10 and PM25 50- 10-kmUKMet Office NAME-III 72-h forecasts of PM tracer 50- 8-kmSpainTUM OPANA v40 72-h forecasts of O3 NH3 NO2 CO SO2 PM25 PM10 Local 10-m regional 9- 3- 1-kmSpainBSC-CNS CALIOPE 24e72-h forecasts of O3 NO2 CO SO2 and PM10 12- and 4-kmGreeceUA SKIRONTAPM 48-h forecasts of O3 NO2 SO2 SO

24 and dust 025 025

ItalyCETEMPS ForeChem 72-h forecasts of O3 NO2 PM25 PM10 and dust 05 and 015

Chinese IAPCAS EMS-Beijing 72e96 h forecasts of O3 NO2 PM25 PM10 81- 27 9 and 3-kmJapan Kyushu University CFORS 8-h AQF for SO2 CO EC aerosol Al fine mode nss-SO4

222Radon dust AOD80-km

ChileMeteo Chile POLYPHEMUS 72-h forecast for O3 4-km

a There are several air quality systems used by regional air quality agencies (eg Association agreacuteeacutees de surveillance de la qualiteacute de lrsquoair (AASQAs)) and only the one for theParis region is given here as an example

Y Zhang et al Atmospheric Environment xxx (2012) 1e24 13

regional and Global Atmospheric POLlution and climate effects) orby individual researchers (eg Cooke et al 1999) Most globalmodels include either online or offline biogenic emissions Somemodels treat online or offline emissions for dust sea salt and dimethylsulfide (DMS) Only one model (ie LMDzt eINCA) treats emissionsof nitrous oxides (N2O) and methane (CH4) Chemical ICONs arebased on default settings (eg GOCART) 6-month spin up simu-lations (eg GEM-AQ) and other global model outputs (egECHAM5) Zero chemical ICONs are used in FLEXPART and GR_RT-AQF Several global models constrained concentrations of somespecies (eg O3 CH4 NO NO2 HNO3 peroxynitrous acid (HNO4)dinitrogen pentoxide (N2O5) NOy) for the upper boundary at thetop of the domain based on climatology or observations to over-come the limitation associated with the lack of stratospheric

Please cite this article in press as Zhang Y et al Real-time air quality forEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

chemistry and stratospheric-tropospheric exchange processesThese models provide 15e96 h RT-AQF at a horizontal grid reso-lution with latitude and longitude of 09 09 to 25 375with one exception for GEM-AQ which can forecast up to 10 daysOne notable feature of GEM-AQ is its rotated variable resolutiongrid with a uniform high resolution window at 01375 (w15-km)that can be placed over any region of interest on a regional scale(eg North America) (Neary et al 2007) which makes GEM-AQa promising model for multiscale air quality modeling TheRT-AQF products from global RT-AQF systems include gaseousspecies (eg CO SO2 NOx O3 OH HNO3 peroxyacytyl nitrate(PAN) and VOCs) PM species (eg PM25 PM10 PM composition)and aerosol opticalradiative properties (eg aerosol extinctioncoefficient AOD)

ecasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e2414

Regional RT-AQF models typically use regional anthropogenicemissions generated by GEMS-TNOMEGAPOLI-TNO and the Euro-pean Monitoring and Evaluation Programme Model (EMEP) overEurope the US National Emissions Inventory and Canadianinventories over North America and the Transport and ChemicalEvolution over the Pacific (TRACE-P) and the IntercontinentalChemical Transport Experiment-Phase B (INTEX-B) emissions overAsia as well as national and local emission inventories (egChinese National Environment Agency emission inventories) overindividual countries These emissions are processed intotemporally-resolved gridded emissions with emission models suchas the Sparse Matrix Operator Kernel Emissions (SMOKE) ModelingSystem and the High-Elective Resolution Modelling EmissionsSystem (HERMES) The RT-AQF models used by regional air qualityagencies such as AIRPARIF for the Paris region typically usemore accurate emission inventories developed based on localinformation than the national RT-AQF systems such as Prevrsquoairfor France Despite the needs for real-time emissions for RT-AQFonly a few RT-AQF models (eg MAQSIP-RT NRT-AQFC WRFChemeMADRID GEM-CHRONOS CHIMERE and PolyphemusPOLAIR3D) use projectedforecasted anthropogenic emissionsfor RT-AQF For example in NRT-AQFC the anthropogenic emis-sions are calculated with two approaches a static projectionapproach for emissions that are projected from historical data withpredetermined spatial and temporal variability (eg from areasources) and an ldquoonlinerdquo approach for emissions with a strongdependence on meteorology (eg from point biogenic and mobilesources) For biogenic emissions the Biogenic Emissions InventorySystem (BEIS) model is used inmostmodels Biogenic emissions areincluded as offline emissions in most models and as online emis-sions in 7 out of 36 models Online sea-salt and dust emissions areincluded in some models Inclusion of dust emissions improvessubstantially the model skills in terms of both discrete and cate-gorical statistics in regions where the influence of dust emissionscannot be neglected (eg over southern Europe by Jimeacutenez-Guerrero et al 2008) Many models do not include sources ofPM25 from wildfires

Chemical ICONs are default settings that are based on clima-tology or measurements in most models and spin up simulationsand larger scale model outputs in some models Chemical BCONsare based on climatology or continental clean conditions formost models and larger scale outputs for some models Thesemodels provide 24e72 h RT-AQF at horizontal grid resolutions of1e125 km with an exception for CFORS that provides 8-h AQF andSILAM that provides 54e72 h forecasts of Radon O3 NO NO2 COSO2 PM25 and PM10 and 120-h forecast for pollen The modelsused by regional air quality agencies such as AIRPARIF typically usea finer spatial resolution (eg 3e9 km) than the national systemPrevrsquoair (eg 50 km) All regional models except for three models(ie MAQSIP-RT THOR and NAME III) forecast both gaseous andPM species MAQSIP-RT forecasts O3 THOR forecasts CO NOx andO3 and NAME-III forecasts PM tracer AOD is forecasted only byCFORS Among 36 models three (ie UAQIFS e Norway UAQIFS-Finland and UAQIFS-Italy2) provide forecasts of human expo-sures to pollutants such as NO2 PM25 and PM10

33 Chemistry and aerosol treatments

Table A3 summarizes the gas-phase chemistry aerosol chem-istry and dynamics and cloud chemistry treated in these modelsThe gas-phase chemistry used in global RT-AQFmodels ranges fromhighly-simplified chemical mechanisms such as SO2 and DMSoxidation by prescribed OH radical and nitrate radical (NO3) inGOCART (Chin et al 2003) to a detailed chemistry such as theREactive Processes Ruling the Ozone BUdget in the Stratosphere

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

(REPROBUS) model mechanismwith 118 species and 350 reactionsused in MOCAGE in the ECMWF-IFS-CTMs system (Flemming et al2009) Two gas-phase mechanisms that are commonly used inregional AQMs are also used in two global models ie the RegionalAtmospheric Chemistry Mechanism (RACM) in MOCAGE and theCarbon BondMechanism-Version IV (CB-IV) in TM5 in ECMWF-IFS-CTMs SO4

2 elemental carbon (EC) primary organic carbon (OC)Nathorn Cl and dust are treated in GOCART ECHAM5 MOCAGE andECMWF-IFS-CTMs Other models either include prescribed SO4

2

concentrations only (eg LMDzt-INCA) or do not simulate PM (egGR-RT-AQF) None of the nine global RT-AQF models simulate NO3

and secondary organic aerosol (SOA) The particle size represen-tations are based on the modal approach with 1-mode for LMDzt-INCA and 7-mode for ECHAM5 the sectional approach with 3e12bins for some models (eg GOCART GEM-AQ) and a combinationof sectional (for sea-salt and dust) and bulk (for remaining PMspecies) approaches for ECMWF-IFS-CTMs The detailed aerosolmicrophysical processes such as nucleation coagulation andcondensation of sulfuric acid are treated only in GEM-AQ andECHAM5 Among 9 global RT-AQF models only 4 models includecloud chemistry

The most popular gas-phase chemistry mechanisms used inregional RT-AQF models include CB-IV CB05 the carbon bondmechanism e version Z (CBM-Z) SAPRC-99 the Regional AcidDeposition Model Mechanism Version 2 (RADM2) RACM and theADOM version 2 (ADOM-II) These mechanisms differ in severalaspects such as the number of species and reactions the level ofcomplexity and lumping methods for organic species and thereaction ratestoichiometry parameters Among them SAPRC-99offers the most detailed chemistry with 80 species and 214 reac-tions that can lead to high O3 formation due to high radicalconcentrations and chemical reactivity Some of these gas-phasemechanisms were compared and evaluated in either box or 3-Dmodels (eg Gross and Stockwell 2003 Faraji et al 2008Luecken et al 2008 Kim et al 2009 2011a Zhang et al 2012)While these mechanisms give similar predictions for gaseousconcentrations in rural and clean conditions the differences insimulated O3 and NOy may be large under urban conditions duemainly to different representations of organic chemistry anddifferent selection of kinetic data for inorganic chemistry Differ-ences in gas-phase mechanisms will lead to differences insecondary aerosol formation Kim et al (2011b) reported lessthan 1 mg m3 (6) differences in monthly-mean PM25 concen-trations but with up to 26 differences in PM25 composition frommodel simulations with the same aerosol module but differentgas-phase mechanisms (ie CB05 and RACM2) Zhang et al (2012)compared CBM-Z CB05 and SAPRC-99 in WRFChemeMADRIDover the continental US and found that different gas-phasemechanisms led to different predictions of concentrations of O3(up to 5 ppb) PM25 (up to 05 mg m3) secondary inorganic PM25species (up to 11 mg m3) and organic PM (up to 18 mg m3)aerosol number concentrations (up to 41) CCN (up to 58) andCDNC (up to 997) These studies illustrated the potentially largeimpact of gas-phase chemical mechanisms on overall modelpredictions

Twenty regional models treat aqueous-phase equilibrium andkinetic reactions and one model (ie STEM-2K3) only treats equi-librium partitioning of SO2 H2O2 and O3 For models with aqueous-phase chemistry the level of details varies from a highly-simplified(eg one first-order oxidation of SO2) to the most comprehensivemechanism (eg the Carnegie-Mellon University (CMU) aqueous-phase mechanism with 50 aqueous-phase species 17 aqueous-phase ionic equilibria 21 gas-phaseaqueous-phase reversiblereactions and 109 aqueous kinetic reactions) that is used in onemodel (ie WRFChemeMADRID) The CMU mechanism explicitly

recasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e24 15

treats the oxidation of dissolved S(IV) by H2O2 O3 O2 catalyzed byFe3thorn and Mn2thorn and radical species as well as the non-reactiveuptake of HNO3 hydrochloric acid (HCl) ammonia (NH3) andother trace gases

Compared with global RT-AQF models regional RT-AQF modelscontain much more detailed aerosol treatments All models exceptfor THOR and five UAQIFS models simulate SO4

2- NO3 NH4

thorn EC OCsodium and chloride UAQIFS-Spain and UAQIFS-Denmark do notsimulate PM Other models only simulate some PM species Wind-blown dust is simulated in 14 out of 36 models (eg ART-AQFSSTEM-2K3 WRFChem CHIMERE SKIRONTAPM) 21 out of 36models simulate detailed aerosol chemistry and dynamics such asthermodynamics for inorganic and organic aerosols nucleationcondensation and coagulation Three models (ie EMEP-UnifiedMATCH and LOTOS-EUROS) only simulate thermodynamics forinorganic aerosols and GEM-CHRONOS simulates an additionalprocess ie SOA formation using a fixed yield approach Ninemodels (eg ART-AQFS MAQSIP-RT THOR) do not simulate anyaerosol chemistry and dynamics

Among the three CTMs in Prevrsquoair PolyphemusPolair3Dcontains drivers for simple forward run (forecast) ensemble fore-cast and various data assimilation methods and offers the mostdetailed chemistry and aerosol treatments For example it includestwo aerosol models a Modal Aerosol Model (MAM Sartelet et al2007) and a SIze REsolved Aerosol Model (SIREAM Debry et al2007) SIREAM simulates detailed aerosol chemistry and dynamicprocesses using state-of-the-science algorithm SOA formation issimulated using either an improved version of the SORGAM model(Kim et al 2011b) which is based on an absorptive partitioningtheory and treats eight SOA classes (4 anthropogenic and 4biogenic) or the AEREPRICaltech (AEC) SOA model (Kim et al2011a) which is based on a hydrophobichydrophilic molecule-based gas-partitioning scheme Currently most SOA modulesused in most RT-AQF models do not take into account the hydro-philic behavior of organic species and only a few account for SOAformation from oxidation of gaseous precursors such as isopreneand sesquiterpene

4 Evaluation protocols and model performance of 3-DRT-AQF models

41 Evaluation protocols

Two types of evaluations are typically conducted to evaluatethe modelrsquos forecasting skills discrete evaluation that uses a set ofstatistical measures and categorical evaluation that uses a set ofcategorical indices Tables 5 and 6 summarize major statisticalmeasures and categorical indices commonly used respectivelyStatistical measures and evaluation protocol for discrete evalua-tion of RT-AQF are the same as those recommended by Seigneuret al (2000) US EPA (2001) and Yu et al (2006) for traditionalAQMs The traditional statistical measures include correlationcoefficient (r) mean bias (MB) mean absolute gross error (MAGE)RMSE mean normalized bias (MNB) mean normalized gross error(MNGE) normalized mean bias (NMB) normalized mean grosserror (NMGE) fractional bias (FB) and fractional gross error (FGE)Some problems may arise with some of those traditionalmeasures For example overpredictions are artificially given moreweight than underpredictions when MNBMNGE and NMBNMGEare used because MNB and NMB can grow disproportionally foroverpredictions but are bounded by 1 for underpredictions(Seigneur et al 2000) A set of new statistical metrics wastherefore developed by Yu et al (2006) based on the concept offactors to overcome the limitations of the traditional measuresThese new metrics include the mean normalized factor bias

Please cite this article in press as Zhang Y et al Real-time air quality forEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

(MNFB) the mean normalized gross factor error (MNGFE) thenormalized mean bias factor (NMBF) and the normalized meanerror factor (NMEF) While MNFB and MNGFE are still subjected tothe dominance by the extremely low observational values in thenormalization NMBF and NMEF provide the most robust statis-tical measures and are therefore recommended for operationalevaluation for model predictions (Yu et al 2006) In additionpeak accuracy measures (eg unpaired peak accuracy (UPA)temporally-paired peak accuracy (TPPA) spatially-paired peakaccuracy (SPPA) and pair peak accuracy (PPA)) index of agree-ment (IOA) and Brier score (BS) are often used in the discreteevaluation UPA quantifies the difference between the magnitudeof the peak observed and simulated values in the modelingdomain regardless whether this occurs at the monitoring stationand time of the peak observed value UPA tests the modelrsquos abilityto reproduce the highest observed value anywhere in the regionUPA is less useful as the two peak values may not occur at thesame location and time TPPA SPPA and PPA provide morestringent measures TPPA tests the modelrsquos ability to reproducethe highest observed value at the hour of the peak observed valueat any location within the modeling domain SPPA tests themodelrsquos ability to reproduce the highest observed value at themonitoring station of the peak observed value at any hour PPAtests the modelrsquos ability to reproduce the highest observed valueat the time and location of the peak observed value which is themost stringent measure among the four peak accuracy measuresIOA provides a measure of how well the departure of the simu-lated values from the observed mean matches with the departureof observations from the observed mean An IOA of 1 indicatesa perfect agreement BS measures the average squared deviationbetween predicted probabilities for a set of events and theiroutcomes so a lower score represents a higher accuracy It isa measure used frequently for probabilistic forecasts

The traditional categorical indices include accuracy (A) criticalsuccess index (CSI) or threat score (TS) probability of detection(POD) (or hit rate (HRate)) bias (B) false alarm ratio (FAR) falsealarm rate (FARate) Heidke skill score (HSS) Pierce skill score(PSS) (or true skill score (TSS) or Hansen and Kuipers discriminant(HKD)) skill score (SS) and economic value (EV) A indicates thepercentage of forecasts that correctly predicts an exceedance ora nonexceedance CSI indicates the percentage of the forecastedactual exceedances to all actual and forecasted exceedances PODor HRate indicates the percentage of the forecasted actualexceedances to all actual exceedances CSI is a more balancedscore because it takes into account both missed exceedances andfalse alarms whereas POD does not take into account false alarmsB judges if forecasts are underpredicted (lt1 also referred to asunderforecasting or ldquofalse negativerdquo) or overpredicted (gt1 alsoreferred to as overforecasting or ldquofalse positiverdquo) FAR measuresthe percentage of times an exceedance was forecast whenexceedance did not occur FARate indicates the proportion ofnonexceedances that were incorrectly forecasted (ie the proba-bility of false detection) it is the counterpart to the hit rate PODor HRate HSS measures the fractional improvement of the fore-cast (ie the total number of correct forecasts minus the correctrandom forecasts) over the standard random chance forecasts (iethe total number of forecasts minus the correct forecasts due tochance) Like most skill scores it is normalized by the total rangeof possible improvement over the standard which means HSS cansafely be compared on different datasets PSS (or TSS or HKD)measures skills relative to an unbiased random reference forecastby taking the differences between POD and FARate SS measuresthe relative accuracy with respect to a set of reference forecastsThe most convenient reference is the persistence forecast Anyvalid error metrics such as RMSE and NME can be used to calculate

ecasting part I History techniques and current status Atmospheric

Table 5Discrete statistical measures used in the model evaluation (compiled from Willmott 1981 Zhang et al 2006a Yu et al 2006 Dutot et al 2007)

Metrics Mathematical Expressiona Range

Correlation Coefficient r frac14(XN

ifrac141

ethMi MTHORNethOi OTHORN)(XN

ifrac141

ethMi MTHORN2XNifrac14 1

ethOi OTHORN2)1

2

1 to 1

Mean Bias (MB) MB frac14 1N

XNifrac141

ethMi OiTHORN frac14 M O N to thornN

Mean AbsoluteGross Error (MAGE)

MAGE frac14 1N

XNifrac141

jMi Oij 0 to thornN

Root MeanSquare Error (RMSE)

RMSE frac141N

XNifrac141

ethMi OiTHORN21

2

0 to thornN

Mean NormalizedBias (MNB)

MNB frac14 1N

XNifrac141

frac12ethMi OiTHORN=Oi frac141N

XNifrac141

ethMi=Oi 1THORN 1 to thornN

Mean NormalizedGross Error (MNGE)

MNGE frac14 1N

XNifrac141

frac12ethjMi OijTHORN=Oi 0 to thornN

NormalizedMean Bias (NMB)

NMB frac14XNifrac141

ethMi OiTHORNXN

ifrac14 1

Oi frac14 M

O 1 1 to thornN

NormalizedMean Error (NME)

NME frac14XNifrac141

jMi OijXN

ifrac141

Oi frac14 MAGE=O 0 to thornN

Fractional Bias (FB) FB frac14 1N

XNifrac141

Mi Oi

ethMi thorn OiTHORN=22 to thorn2

Fractional GrossError (FGE)

FGE frac14 1N

XNifrac141

jMi OijethMi thorn OiTHORN=2

0 to 2

Mean NormalizedFactor Bias (MNFB)

MNFB frac14 1N

XNifrac141

Fi where Fi frac14 Mi=Oi 10 for Mi Oi

Fi frac14 10 Oi=Mi forMi lt Oi

N to thornN

Mean NormalizedGross FactorError (MNGFE)

MNGFE frac14 1N

XNifrac141

jFij where Fi frac14 Mi=Oi 10 for Mi Oi

Fi frac14 10 Oi=Mi forMi lt Oi

0 to thornN

Normalized MeanBias Factor (NMBF)

NMBF frac14XNifrac141

Mi

XNifrac141

Oi 1 frac14 M=O 1 for M O

NMBF frac14 1XNifrac141

Oi

XNifrac141

Mi frac14 eth1 O=MTHORN for M lt O

N to thornN

Normalized Mean ErrorFactor (NMEF)

NMEF frac14XNifrac14 1

jMi OijXN

ifrac141

Oi frac14 MAGEO

for M O

NMEF frac14XNifrac14 1

jMi OijXN

ifrac141

Mi frac14 MAGEM

for M lt O

0 to thornN

Unpaired peakaccuracy (UPA)

UPA frac14Mmaxxt O

maxbxbtOmaxbxbt 1 to thornN

Temporally-paired peakaccuracy (PPA)

TPPA frac14M

maxxbt Omaxbxbt

Omaxbxbt 1 to thornN

Spatially-paired peakaccuracy (SPPA)

SPPA frac14M

maxbxt Omaxbxbt

Omaxbxbt 1 to thornN

Paired peak accuracy (PPA) PPA frac14M

maxbxbt Omaxbxbt

Omaxbxbt 1 to thornN

Index of Agreement (IOA) IOA frac14 1PN

ifrac14 1ethMi OiTHORN2PNifrac141ethjMi Oj thorn jOi OjTHORN2

0 to 1

Brier Score (BS) BS frac14 1K

XKtfrac141

ethpt otTHORN2 0 to thorn1

N is the number of samples (by time andor location)Mmaxbxbt and O

maxbxbt are time- and space-paired peak simulated and observed values with the circumflex indicating thatthe peak value occurs at the time t and location x of the peak observed value pt and ot are the probability that an event occurs according to an ensemble and the event actuallyoccurred or not at instance t (0 if it doesnrsquot happen and 1 if it happens) respectively K is the number of forecasting instances

a M frac14 eth1=NTHORNPNifrac141 MiO frac14 eth1=NTHORNPN

ifrac141 Oi Mi and Oi are values of model predictions and observations at time or location i respectively

Y Zhang et al Atmospheric Environment xxx (2012) 1e2416

SS For a perfect model A CSI POD HSS PSS and SS would be 1FAR and FARate would be zero and B would be unity

The above traditional categorical metrics provide ldquoclear cutrdquomeasures of the modelrsquos ability to predict an ldquoexceedancerdquo by usinga fixed threshold concentration and time- and space-paired

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

observationeforecast sets A main limitation is that they cannotprovide information regarding the spatial variations of pollutantconcentrations within a region of interest To overcome this limi-tation Kang et al (2007) developed three new categorical metricsfor RT-AQF evaluation the weighted success index (WSI) area hit

recasting part I History techniques and current status Atmospheric

Table 6Categorical statistical measures used in the model evaluation (compiled based on Ryan 1995 McHenry et al 2004 Kang et al 2005 2007 Pagowski and Grell 2006)

Metrics Mathematical Expressiona Range

Accuracy A frac14 bthorn cathorn bthorn cthorn d

0e1

CSI frac14 bathorn bthorn d

0e1

Critical Success Index or Threat Score (TS) WSI frac14 bthornPn1 IP

athorn bthorn d0e1

The Weighted Success Index (WSI)bIP frac14 Mi f Oi

Mi f T for Mi gt T gt Oi and Mi lt f T

IP frac14 Oi f Mi

Oi f T for Oi gt T gt Mi and Oi lt f Mi

IP frac14 0 for other conditions

Probability Of Detection Or Hit Rate POD frac14 H frac14 bbthorn d

0e1

Area Hit Ratec aH frac14 AbAbthorn Ad

0e1

Bias or Bias Ratio B frac14 athorn bbthorn d

0 to thornN

False Alarm Ratio FAR frac14 aathorn b

0e1

Area False Alarm Ratiod aFAR frac14 AaAathorn Ab

0e1

False Alarm rate FARate frac14 F frac14 aathorn c

0e1

Heidke skill score HSS frac14 2ethbc adTHORNethbthorn dTHORNethcthorn dTHORN thorn ethathorn bTHORNethathorn cTHORN N to 1

Pierce skill score (or true skill score (TSS)or Hansen and Kuipers discriminant)

PSS frac14 POD FARate frac14 bbthorn d

aathorn c

frac14 bc adethathorn cTHORN ethbthorn dTHORN 1 to 1

Skill scoreef SS frac14 Em ErefEperf Eref

0e1

Em frac14 RMSE frac141N

XNifrac141

ethMi OiTHORN21

2

or NME frac14XNifrac14 1

jMi OijXN

ifrac141

Oi frac14 MAGE=O

Eref frac14 RMSE frac141N

XNifrac14 1

ethMi KiTHORN21

2

or NME frac14XNifrac14 1

jMi MjjXN

jfrac141

Ki frac14 MAGEref=K

Eperf frac14 0

Economic value (EV) EV frac14 minethr cf THORN Feth1 cf THORNr thorn H cf eth1 rTHORN cfminethr cf THORN cf r

0e1

r frac14 CLfrac14 athorn b

d

H frac14 bbthorn d

F frac14 aathorn c

cf is the climatological frequency of an event

a Where a b c d are the number of simulated and observed data pairs at one site at a specific time in the four regions (see Table below) that represent forecast exceedancesthat did not occur forecast exceedances that did occur forecast nonexceedances that did occur and forecast nonexceedances that did not occur respectivelyContingencytable for threshold forecasts

ObservedYes No

Forecast Yes b aNo d c

b Mi and Oi are values of model prediction and observation at a location i respectively T is the threshold value used f is an empirical factor that reflects the zone in which themost points are located on the predictioneobservation scatterplots its typical value can be 1e2

c Ab is the number of exceedances that are both observed and forecast but the forecast is any exceedance that occurs in the designated area centered at themonitor locationAd is the number of observed exceedances that are not forecast within the designated area centered at the monitor location

d Aa is the number of forecast area exceedances that were not observed and Ab is the number of forecast exceedances that were observede Mi Oi and Ki are values of model prediction observation and temporal or spatial persistence value at time or location i respectively Temporal persistence value is the

previous days forecast value Spatial persistence value is the observation at the nearest location to the location i O frac14 eth1=NTHORNPNifrac141 Oi K frac14 eth1=NTHORNPN

ifrac141 Ki N is the number ofsamples (by time andor location) RMSE e Root mean square error NME e Normalized mean error MAGE e Mean Absolute Gross Error

f Em Eref and Eperf are the errors associated with model reference and perfect forecast respectively Em and Eref can be any valid error metrics eg RMSE and NME Thetemporal or spatial persistence forecast is often used as a reference forecast a perfect forecast has a zero error ie Eperf frac14 0

Y Zhang et al Atmospheric Environment xxx (2012) 1e24 17

Please cite this article in press as Zhang Y et al Real-time air quality forecasting part I History techniques and current status AtmosphericEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

Y Zhang et al Atmospheric Environment xxx (2012) 1e2418

(aH) and area false-alarm ratio (aFAR) WSI is less stringent thanCSI but a more representative measure It provides partial credit toobservationeforecast pairs that fall just outside limits establishedby CSI butwithin a designated factor that is used to judge themodelperformance (eg data pairs fall into a factor of 1e2 of theobservations)

An important practical aspect of RT-AQFs is how forecasts canbe used to mitigate the impact of air quality on economy andpeople In the calculation of EV one needs to know the climato-logical frequency (cf) of an event in addition to the values of a b cand d (see definitions in Table 6) Hits and false alarms incur a costof taking a preventive action (C) while misses are associated withlosses due to the lack of prevention (L) Correct rejections (c) incurno expense EV is defined as the reduction in mean expense rela-tive to the reduction (both defined with respect to cf) that wouldoccur if all the forecasts were correct (Pagowski and Grell 2006) Asfor weather forecasting EV provides a useful measure for theeconomic benefits of the existing RT-AQFs and the benefits ofimproving these systems to individuals businesses and society

No recommended values of performance statistics were providedfor acceptable forecasting skills by any individual countries Therecommended values for retrospective AQ modeling can providesome guidance for RT-AQF skill evaluation For example Zhang et al(2006a) recommended the use of MNBs and NMBs 15 MNGEsand NMEs 30 as indications of a satisfactory performance for O3

and PM25 Boylan and Russell (2006) recommended the use of meanfractional bias (MFB) and mean fractional error (MFE) for PM withgood performance indicted by MFBs 30 and MFEs 50 forPM25 and MFBs 60 and MFEs 75 for major PM componentsthat have concentrations 225 mg m3 The US EPA recommendedthat a so-calledweight of evidence approach inwhich a set of diverseanalyses are used to judge model performance Since no singlestatistical metrics can fully reflect RT-AQF skill discrete and cate-gorical statistics summarized in Tables 5 and 6 are recommended toprovide a comprehensive evaluation of an RT-AQF system

42 Observational datasets for model evaluation

Datasets for forecasting evaluation and improvement includereal-time or near real-time surface measurement datasets suchas the US EPArsquos AIRNOW datasets obtained by aircraft shipozonesondes and lidars during special field campaigns andsatellite data Satellite observations can complement datasetsfrom surface networks by providing derived chemical measure-ments Recent work has shown the potential of using near real-time satellite and surface data to improve RT-AQF (eg Al-Saadiet al 2005 Kondragunta et al 2008 Wang et al 2011) OverEurope the Global and Regional Earth-System MonitoringUsing Satellite and In situ Data (GEMS) and MACC projects areaimed at a near real-time data assimilation and forecast capabilityfor aerosols greenhouse gases and reactive gases The dataset canbe used to monitor the composition infer estimates of surfacefluxes and produce global short-range and medium-rangeforecasts

43 Current model forecasting skills

Most evaluations of forecasted O3 and its precursors focus onsummer and very few include winter (eg Manins et al 2002 Caiet al 2008 Doraiswamy et al 2009) Forecasting PM is moredifficult than forecasting O3 because PM consists of multiplechemical components over a broad size spectrum PM modeling isstarting to reach sufficient maturity for transition from research tooperational use Several PM models have recently been transferredinto operational models to forecast PM (Carmichael et al 2003

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

McHenry et al 2004 McKeen et al 2005 2007 Yu et al 2008Chuang et al 2011) very few studies provide detailed evaluationsof forecasted PM and its composition and precursors (eg McKeenet al 2007 2009 Yu et al 2008 Chen et al 2008 Manders et al2009) Similar to O3 forecasts most evaluations are conducted forsummer episodes and very few for winter episodes (eg Mathuret al 2008 Konovalov et al 2009) or a full year (eg Manderset al 2009) Limited evaluation was performed for coarse parti-cles such as mineral dust (eg Jimeacutenez-Guerrero et al 2008 Niuet al 2008 Menut et al 2009) that may be of major concernsfor PM10 attainment in some regions such as Asia southern Europeand the western US Jimeacutenez-Guerrero et al (2008) showed thatthe inclusion of a dust emission module substantially increases theaccuracy of both discrete and categorical statistics in the IberianPeninsula in Europewhere the influence of the Saharan dust cannotbe neglected

Table A4 summarizes the evaluation of a number of RT-AQFsystems in terms of domain and period and discrete and categor-ical performance statistics While maximum 1-h O3 averagestatistics are generally satisfactory those for maximum 8-h averageand hourly O3 sometime exceed NMB of 15 and NME of 30 forsome models (eg EtaCMAQ WRF-NMMCMAQ) over the easternor northeastern US in the O3 season during 2004e2006 indicatinga relatively poor performance The overpredictions in the low O3range (lt50 ppb) were reported in several studies (eg Yu et al2007 Chen et al 2008) Several factors may contribute to suchoverpredictions These include a poor representation of thenocturnal PBL mixing height (Gilliam et al 2006 Zhang et al2006b Eder et al 2006) an excessive downward transport ofhigh level O3 aloft and too much photolysis under high cloudconditions (Eder et al 2006) a high O3 production rate with theSAPRC-99 chemical mechanism (Arnold and Dennis 2006) themodel limitation in resolving titration of O3 by NO in urban plumes(Yu et al 2007) and the BCONs from global models (Chen et al2008) Chuang et al (2011) applied WRFChemeMADRID forRT-AQF and found that O3 overprediction in most regions in thesoutheastern US is likely caused by inaccurate emissions ofprecursors such as biogenic VOCs positive biases in 2-m temper-ature and negative biases in wind speed at 10-m O3 under-predictions in some regions could be due in part to theuncertainties in lateral BCONs

For categorical evaluation different threshold values were usedfor the same variables For example the threshold values used formaximum 1-h average O3 maximum 8-h average O3 and hourly O3are 60e125 ppb 65e85 ppb and 80 ppb respectively Some ofthese values are lower than the former US NAAQS of 120 ppb formaximum 1-h average O3 and the current US NAAQS of 75 ppb formaximum 8-h O3 respectively for the reasons stated previously(eg Hogrefe et al 2007 Yu et al 2007 Chuang et al 2011)although higher thresholds were used in some earlier applications(eg Kang et al 2005) For maximum 1-h average O3 A CSI POD Band FAR range from 15 to 998 52 to 216 69 to 89 01 to53 ppb and 03 to 941 respectively Prevrsquoair gives the lowest A of15e41 because it underestimates O3 daily maxima at high O3concentrations (Honoreacute et al 2008) All RT-AQFs give low values ofCSI and POD but higher values of FAR because they fail to forecastexceedance (ie a low value of b) but overpredict lowO3 (ie a highvalue of a) A CSI POD B and FAR for maximum 8-h average O3

range from 762 to 998 0 to 532 0 to 848 03 to 170 ppb and13 to 991 respectively Compared with maximum 1-h O3 thevalues of CSI and POD are higher and those of FAR are lower forcorresponding maximum 8-h O3 The categorical evaluation forhourly O3 was conducted by only one model (ie WRFChemeMADRID) with A CSI POD B and FAR of 992 25188 68 ppb and 972 respectively Very few evaluations were

recasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e24 19

conducted for precursors of O3 and other gaseous species (eg Caiet al 2008) and vertical profile of forecasted concentrations (egYu et al 2007) Yu et al (2007) reported that EtaCMAQ reproducedO3 vertical distributions on most of the days at low altitudes butoverpredictions occurred at altitudes gt 6 km because of a combi-nation of effects related to the specifications of meteorologicallateral BCONs as well as themodelrsquos coarse vertical resolution in theupper free troposphere

Surface PM forecasts are evaluated in terms of 24-h averagePM25 hourly PM25 and 24-h average PM10 These evaluationshowever have been very limited to date and a consistent evaluationprotocol has not yet been well established (Seigneur 2001 Zhanget al 2006a) The forecasting skill for PM25 is overall poorer thanthat for O3 For 24-h average PM25 MBs RMSEs NMBs and NMEsrange from 32 to 62 mg m3 17 to 159 mg m3 21 to 32 and37 to 81 respectively The NMB of 21 using EtaCMAQ over theeastern US during July 14eAugust 18 2004 was reported by Yuet al (2008) who attributed underpredictions to underestimatedtotal carbonaceous PM at both urban and rural sites and a signifi-cant underestimation of unspecified anthropogenic PM mass(mainly consisting of primary emitted trace elements) at rural sitesOn the other hand Hogrefe et al (2007) showed that EtaCMAQoverpredicted PM25 in the New York City due to overpredictions inorganic aerosols and crustal material Factors contributing to suchoverpredictions include underpredictions of nocturnal verticalmixing inaccurate temporal allocation of primary OM emissionsand underestimate of deposition processes Chen et al (2008)reported the NMB of 17e32 over the Pacific Northwest duringAugusteNovember 2004 The overpredictions in PM25 wereattributed to uncertainties in wildfire emission estimates and themodeling error in fire plume transport due to errors in MM5-predicted wind direction and wind speed McKeen et al (2007)evaluated RT-AQF of 6 models including WRFChem at two gridresolutions (12- and 36-km) CHRONOS AURAMS STEM-2K3 andEtaCMAQ They found that most models did not reproduce theobserved diurnal variation at urban and suburban sites particularlyduring the nighttime to early morning and the ensemble meanbased on 6 models with equal weighting gave the best possibleforecast in terms of statistics Chuang et al (2011) showed slightunderpredictions of PM25 in the O3 season over the southeasternUS by WRFChemeMADRID which were attributed to uncer-tainties in emissions such as those of biogenic VOCs and NH3overpredictions of precipitation and uncertainties in the BCONsOnly two studies evaluated hourly PM25 givingMBs RMSEs NMBsand NMEs of 33 to06 mg m3 83e113 mg m3 211 to52and 498e514 respectively For 24-h average PM10 MBs andRMSEs range from 29 to 138 mg m3 and 83 to 472 mg m3respectively Berge et al (2002) reported forecasted PM10 perfor-mance over Oslo with an MB of 209 mg m3 and an NMBof 365 at Kirkeveien due to an inaccurate emissions of PM10from the surface (ie the re-suspension of dust deposited at theroadside) on dry days and an MB of 138 mg m3 and an NMB of 24at Furuset due to errors in simulated grid-averaged wind fields

Very few evaluations were conducted for PM componentsMcKeen et al (2007) found that all the six RT-AQF models signifi-cantly underpredicted OM at the surface and overestimated SO4

2

above 2 km The overpredictions in SO42 were attributed to over-

estimate of SO2 by WRFChem and CHRONOS and the inclusion ofaqueous-phase oxidation of SO2 by AURAMS CMAQETA andSTEM-2K3 Yu et al (2007) reported that EtaCMAQ overpredictedSO4

2 due to too much in-cloud SO2 oxidation as a result of over-estimated H2O2 concentrations in the model underpredicted NH4

thorn

at the rural sites and aloft due to the exclusion of some sources ofNH3 in the real-time emission inventory underpredicted NO3

dueto overpredictions of SO4

2 and underpredicted OC due to missing

Please cite this article in press as Zhang Y et al Real-time air quality forEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

sources of primary OC in the emission inventory and missing SOAformation from the gas-phase oxidation of isoprene and sesqui-terpenes and aqueous-phase oxidation of glyoxal and methyl-glyoxal Chen et al (2008) reported that AIRPACT-3 reproduced OCwell but significantly overpredicted EC and significantly under-predicted SO4

2 NO3 and NH4

thorn due to underestimation of emis-sions of primary PM species (eg sulfate) and precursors ofsecondary PM (eg SO2 NH3 and NOx) insufficient spatial reso-lution and modelrsquos inability to capture the hourly PM variations

For categorical evaluation of PM some studies used thresholdvalues of 15e315 mg m3 for 24-h average PM25 and 30e50 mgm3

for 24-h average PM10 that are lower than the NAAQSs of 24-haverage PM25 of 35 mg m3 and 24-h average PM10 of 150 mg m3

for the same reason as mentioned previously For 24-h averagePM25 A CSI POD B and FAR range from 608 to 997 0 to 5370 to 909 07 to 19 mg m3 and 25 to 100 respectively Those for24-h average PM10 range from 167 to 100 91 to 100 150 to100 05 to 09 mg m3 and 0 to 48 respectively Categoricalevaluation of hourly PM25 was conducted only with WRFChemeMADRID with A CSI POD B and FAR of 721 205292 07 mg m3 and 591 respectively

5 Summary of Current States of RT-AQF

Emerging as a new discipline of applied sciences that integratesseveral physical sciences mathematicalstatistical tools andcomputer technology in the 1970s RT-AQF represents a uniqueapplied science affecting humanrsquos daily activities and posesunprecedented challenges intersecting many aspects of sciencesand technologies Driven by deteriorated air quality worldwideincreased societalhuman demands and rapid scientific and tech-nological advancements significant progress has been made in thepast four decades in RT-AQF A number of RT-AQF tools and modelswith varying degrees of sophistication and forecasting skillsranging from the simplest rule of thumb to the most advanced 3-Donline-coupled meteorologyechemistry models have been devel-oped Among them the deterministic physically-based RT-AQFmodels that are based on 3-D global and regional CTMs representthe state of the science enabling scientific understanding ofmechanisms for pollutant formation and development of emissioncontrol strategies Although other approaches in particular para-metric (statistical) models will continue to be used the 3-D RT-AQFmodels will likely become prevalent approaches and tools on all thescales There is a growing trend to combine these 3-D modelswith statistical models to provide more accurate RT-AQF Amongall RT-AQF models the coupled meteorologyechemistry modelrepresents a significant advancement and will greatly enhance theunderstanding of the underlying complex interplay of meteorologyemissions and chemistry in the real atmosphere Importantextensions of the RT-AQF models include their coupling with anurban model (eg traffic andor local pollutant dispersion) or a CFDmodel for urbanlocal scale applications at a spatial resolution of1 km or less and with an exposure model to provide real-timepublic health assessment and exposure predictions and urbanemergency preparedness

3-D RT-AQF models have been used for 24e72 h RT-AQFworldwide since the mid 1990s Despite their unique characteris-tics and technical requirements 3-D RT-AQF models are based onNWP models and CTM development and application efforts in thepast decades and therefore are subject to limitations of thesemodels Although no universal forecasting guidance and evaluationprotocols were developed covering all countries some guidanceand protocols were recommended by governmental agencies andthe scientific community Discrete and categorical evaluations aretwo most commonly-used evaluations Real-time or near real-time

ecasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e2420

surface air quality measurement and satellite datasets areincreasingly available for such evaluations The evaluation effortshave been primarily focused on O3 NO2 and PM (both PM10 andPM25) at surface level The forecasting skill is typically better for O3than for PM25 While their performance statistics averaged overa time period (eg monthly or daily) and the whole domain isgenerally good or satisfactory current RT-AQF models have diffi-culties forecasting variations at finer time scales (hourly and diur-nally) and individual monitoring stations In addition larger biasesexist for forecasted NOx HONO HNO3 and OH and PM compositionincluding SO4

2 NO3 and OM indicating possible error compen-

sations that lead to current levels of forecast accuracy These fore-cast inaccuracies have been attributed to a number of factors thatare highly uncertain or lack of accurate treatments in the RT-AQFmodels including mesoscale meteorological processes (eg sea-breeze circulations) and variables (eg temperature PBL height)chemical BCONs of O3 and PM25 emissions (eg SO2 NOx NH3BVOCs primary PM) physical and chemical processes (eg urbanprocesses in-cloud oxidation of SO2 SOA formation) and modelconfiguration (eg coarse grid resolution) Recent advances havebeenmade in some of those aspects Part II of this reviewwill assessand discuss these advances along with advanced computationaltechniques that can potentially improve RT-AQF Recommendationsfor research priorities and future prospects will also be provided

Acknowledgments

This work was performed under the Fellowship Award(704389J) Center for Education and ResearchEacutecole des PontsParisTech France through a visiting professorship at the Atmo-spheric Environment Center (CEREA) the Joint Laboratory Eacutecole desPonts ParisTech and EDF RampD Paris France the NSF Career Award(ATM-0348819) the EPA-Science to Achieve Results (STAR)program (R83337601) and Chinarsquos National Basic ResearchProgram (2010CB951803) Support from CEEH MEGAPOLI andCOST ES1004 is also acknowledged Thanks are due to Ming-TungChuang a former post-D researcher at NCSU for compiling anearlier version of Table A4 Thanks are due to Yelva RoustanAtmospheric Research Center (CEREA) France Christian Hogrefethe New York State Department of Environmental Conservation(NYSDEC) NY US and John McHenry Baron Advanced Meteoro-logical Systems (BAMS) US for providing some informationregarding their air quality forecasting models

Appendix A Supplementary Material

Supplementary data related to this article can be found online athttpdxdoiorg101016jatmosenv201206031

References

Alhanafy TE Zaghlool F El AS Moustafa D 2010 Neuro fuzzy modeling schemefor the prediction of air pollution J American Sci 6 (12) 605e616

Al-Saadi J Szykman J Pierce RB Kittaka C Neil D Chu DA Remer LGumley L Prins E Weinstock L MacDonald C Wayland R Dimmick FFishman J 2005 Improving national air quality forecasts with satellite aerosolobservations Bull Amer Meteor Soc 86 1249e1261

Arnold JR Dennis RL 2006 Testing CMAQ chemistry sensitivities in base caseand emissions control runs at SEARCH and SOS99 surface sites in the south-eastern US Atmos Environ 40 5027e5040

Aron RH 1980 Forecasting high level oxidant concentrations in the Los Angelesbasin J Air Pollut Control Assoc 20 1227e1228

Baklanov A Haumlnninen O Slordal LH Kukkonen J Bjergene N Fay B Finardi SHoe SC Jantunen M Karppinen A Rasmussen A Skouloudis A Sokhi RSSorensen j H Oslashdegaard V 2007 Integrated systems for forecasting urbanmeteorology air pollution and population exposure Atmos Chem Phys 7855e874 wwwatmos-chem-physnet78552007

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

Baklanov A Korsholm U Mahura A Petersen C Gross A 2008 Enviro-HIRLAMon-line coupled modelling of urban meteorology and air pollution Adv Sci Res2 41e46 wwwadv-sci-resnet2412008

Baklanov A Mahura A Sokhi R (Eds) 2011 Integrated Systems of Meso-Meteorological and Chemical Transport Models first ed Springer ISBN 978-3-642-13979-6 p 242

Baklanov A 2006 Overview of the European project FUMAPEX Atmos ChemPhys 6 2005e2015

Baklanov A 2010 Chemical weather forecasting a new concept of integratedmodeling Adv Sci Res 4 23e27

Baldasano JM Jimeacutenez-Guerrero P Jorba O Peacuterez C Loacutepez E Guumlereca PMartin F Garciacutea-Vivanco M Palomino I Querol X Pandolfi M Sanz MJDieacuteguez JJ 2008 CALIOPE an operational air quality forecasting system for theIberian Peninsula Balearic Islands and Canary Islands e First annual evaluationand ongoing developments Adv Sci Res 2 89e98 ISSN 1992-0628 May

Battan LJ 1966 The Unclean Sky Doubleday Garden City NY 141 ppBerge E Walker S-E Sorteberg A Lenkopane M Eastwood S Jablonska HI

Koltzow MOslash 2002 A real-time operational forecast model for meteorologyand air quality during peak air pollution episodes in Oslo Norway Water AirSoil Poll Focus 2 745e757

Blond N Vautard R 2004 Three-dimensional ozone analysis and their use forshort-term ozone forecasts J Geophys Res 109 D17303 httpdxdoiorg1010292004JD004515

Boylan JW Russell AG 2006 PM and light extinction model performance metricsgoals and criteria for three-dimensional air quality models Atmos Environ 40(26) 4946e4959 httpdxdoiorg101016jatmosenv200509087

Brandt J Christensen JH Frohn LM Palmgren F Berkowicz R Zlatev Z 2001Operational air pollution forecasts from European to local scale Atmos Environ35 (Suppl 1) S91eS98

Brimblecombe P 1987 The Big Smoke a History of Air Pollution in London sinceMedieval Time Routledge Revivals Methuen London UK 186 pp

Burrows WR Benjamin M Beauchamp S Lord ER McCollor D Thomson B1995 CART decision-tree statistical analysis and prediction of summer seasonmaximum surface ozone for the Vancouver Montreal and Atlantic Regions ofCanada J Appl Meteor Climatol 34 1848e1862

Byun D Schere KL 2006 Review of the governing equations computationalalgorithms and other components of the models-3 community multiscale airquality (CMAQ) modeling system Appl Mech Rev 59 51e77

Cai C Hogrefe C Katsafados P Kallos G Beauharnois M Schwab JJ Ren XBrune W Zhou X He Y Demerjian K 2008 Performance evaluation of an airquality forecast modeling system for a summer and winter season e Photo-chemical oxidants and their precursors Atmos Environ 42 8585e8599 httpdxdoiorg101026jatmosenv200808029

Carmichael GR et al 2003 Regional-scale chemical transport modeling in supportof intensive field experiments overview and analysis of the TRACE-P observa-tions J Geophys Res 108 (D21) 8823 httpdxdoiorg1010292002JD003117

Carmichael GR Sandu A Chai T Daescu DN Constantinescu EM Tang Y2008 Predicting air quality improvements through advanced methods tointegrate models and measurements J Comput Phys 227 3540e3571

Chai T Carmichael GR Tang Y Sandu A Streets DG 2007 Four-dimensionaldata assimilation experiments with international consortium for atmosphericresearch on transport and transformation ozone measurements J GeophysRes 112 D12S15 httpdxdoiorg1010292006JD007763

Chaloulakou A Saisana M Spyrellis N 2003 Comparative assessment of neuralnetworks and regression models for forecasting summertime ozone in AthensSci Total Environ 313 1e13

Chen J Vaughan J Avise J OrsquoNeill S Lamb B 2008 Enhancement and evalu-ation of the AIRPACT ozone and PM25 forecast system for the Pacific NorthwestJ Geophys Res 113 D14305 httpdxdoiorg1010292007JD009554

Chenevez J Jensen CO 2001 Operational ozone forecasts for the region of Copen-hagen by the Danish Meteorological Institute Atmos Environ 35 4567e4580

Chin M Ginoux P Lucchesi R Huebert B Weber R Anderson T Masonis SBlomquist B Bandy A Thornton D 2003 A global aerosol model forecast forthe ACE-Asia field experiment J Geophys Res 108 (D23) 8654 httpdxdoiorg1010292003JD003642

Chuang M-T Zhang Y Kang D-W 2011 Application of WRFChem-MADRID forreal-time air quality forecasting over the southeastern United States AtmosEnviron 45 (34) 6241e6250

Cobourn WG Hubbard MC 1999 An enhanced ozone forecasting model using airmass trajectory analysis Atmos Environ 33 4663e4674

Cobourn WG 2007 Accuracy and reliability of an automated air quality forecastsystem for ozone in seven Kentucky metropolitan areas Atmos Environ 415863e5875 httpdxdoiorg101016jatmosenv200703024

Comrie AC 1997 Comparing neural networks and regression models for ozoneforecasting J Air amp Waste Manag Assoc 47 653e663

Cooke WF Liousse C Cachier H Feichter J 1999 Construction of a fossil fuelemission data set for carbonaceous aerosol and implementation and radiativeimpact in the ECHAM4 model J Geophys Res 104 22137e22162

Cope ME Hess GD 2005 Air quality forecasting a review and comparison of theapproaches used internationally and in Australia Clean Air and Environ Quality39 (1) 52e60

Cope ME Hess GD Lee S Tory K Azzi M Carras J Lilley W Manins PCNelson P Ng L Puri K Wong N Walsh S Young M 2004 The Australianair quality forecasting system Part I project description and early outcomesJ Appl Meteorol 43 649e662

recasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e2414

Regional RT-AQF models typically use regional anthropogenicemissions generated by GEMS-TNOMEGAPOLI-TNO and the Euro-pean Monitoring and Evaluation Programme Model (EMEP) overEurope the US National Emissions Inventory and Canadianinventories over North America and the Transport and ChemicalEvolution over the Pacific (TRACE-P) and the IntercontinentalChemical Transport Experiment-Phase B (INTEX-B) emissions overAsia as well as national and local emission inventories (egChinese National Environment Agency emission inventories) overindividual countries These emissions are processed intotemporally-resolved gridded emissions with emission models suchas the Sparse Matrix Operator Kernel Emissions (SMOKE) ModelingSystem and the High-Elective Resolution Modelling EmissionsSystem (HERMES) The RT-AQF models used by regional air qualityagencies such as AIRPARIF for the Paris region typically usemore accurate emission inventories developed based on localinformation than the national RT-AQF systems such as Prevrsquoairfor France Despite the needs for real-time emissions for RT-AQFonly a few RT-AQF models (eg MAQSIP-RT NRT-AQFC WRFChemeMADRID GEM-CHRONOS CHIMERE and PolyphemusPOLAIR3D) use projectedforecasted anthropogenic emissionsfor RT-AQF For example in NRT-AQFC the anthropogenic emis-sions are calculated with two approaches a static projectionapproach for emissions that are projected from historical data withpredetermined spatial and temporal variability (eg from areasources) and an ldquoonlinerdquo approach for emissions with a strongdependence on meteorology (eg from point biogenic and mobilesources) For biogenic emissions the Biogenic Emissions InventorySystem (BEIS) model is used inmostmodels Biogenic emissions areincluded as offline emissions in most models and as online emis-sions in 7 out of 36 models Online sea-salt and dust emissions areincluded in some models Inclusion of dust emissions improvessubstantially the model skills in terms of both discrete and cate-gorical statistics in regions where the influence of dust emissionscannot be neglected (eg over southern Europe by Jimeacutenez-Guerrero et al 2008) Many models do not include sources ofPM25 from wildfires

Chemical ICONs are default settings that are based on clima-tology or measurements in most models and spin up simulationsand larger scale model outputs in some models Chemical BCONsare based on climatology or continental clean conditions formost models and larger scale outputs for some models Thesemodels provide 24e72 h RT-AQF at horizontal grid resolutions of1e125 km with an exception for CFORS that provides 8-h AQF andSILAM that provides 54e72 h forecasts of Radon O3 NO NO2 COSO2 PM25 and PM10 and 120-h forecast for pollen The modelsused by regional air quality agencies such as AIRPARIF typically usea finer spatial resolution (eg 3e9 km) than the national systemPrevrsquoair (eg 50 km) All regional models except for three models(ie MAQSIP-RT THOR and NAME III) forecast both gaseous andPM species MAQSIP-RT forecasts O3 THOR forecasts CO NOx andO3 and NAME-III forecasts PM tracer AOD is forecasted only byCFORS Among 36 models three (ie UAQIFS e Norway UAQIFS-Finland and UAQIFS-Italy2) provide forecasts of human expo-sures to pollutants such as NO2 PM25 and PM10

33 Chemistry and aerosol treatments

Table A3 summarizes the gas-phase chemistry aerosol chem-istry and dynamics and cloud chemistry treated in these modelsThe gas-phase chemistry used in global RT-AQFmodels ranges fromhighly-simplified chemical mechanisms such as SO2 and DMSoxidation by prescribed OH radical and nitrate radical (NO3) inGOCART (Chin et al 2003) to a detailed chemistry such as theREactive Processes Ruling the Ozone BUdget in the Stratosphere

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

(REPROBUS) model mechanismwith 118 species and 350 reactionsused in MOCAGE in the ECMWF-IFS-CTMs system (Flemming et al2009) Two gas-phase mechanisms that are commonly used inregional AQMs are also used in two global models ie the RegionalAtmospheric Chemistry Mechanism (RACM) in MOCAGE and theCarbon BondMechanism-Version IV (CB-IV) in TM5 in ECMWF-IFS-CTMs SO4

2 elemental carbon (EC) primary organic carbon (OC)Nathorn Cl and dust are treated in GOCART ECHAM5 MOCAGE andECMWF-IFS-CTMs Other models either include prescribed SO4

2

concentrations only (eg LMDzt-INCA) or do not simulate PM (egGR-RT-AQF) None of the nine global RT-AQF models simulate NO3

and secondary organic aerosol (SOA) The particle size represen-tations are based on the modal approach with 1-mode for LMDzt-INCA and 7-mode for ECHAM5 the sectional approach with 3e12bins for some models (eg GOCART GEM-AQ) and a combinationof sectional (for sea-salt and dust) and bulk (for remaining PMspecies) approaches for ECMWF-IFS-CTMs The detailed aerosolmicrophysical processes such as nucleation coagulation andcondensation of sulfuric acid are treated only in GEM-AQ andECHAM5 Among 9 global RT-AQF models only 4 models includecloud chemistry

The most popular gas-phase chemistry mechanisms used inregional RT-AQF models include CB-IV CB05 the carbon bondmechanism e version Z (CBM-Z) SAPRC-99 the Regional AcidDeposition Model Mechanism Version 2 (RADM2) RACM and theADOM version 2 (ADOM-II) These mechanisms differ in severalaspects such as the number of species and reactions the level ofcomplexity and lumping methods for organic species and thereaction ratestoichiometry parameters Among them SAPRC-99offers the most detailed chemistry with 80 species and 214 reac-tions that can lead to high O3 formation due to high radicalconcentrations and chemical reactivity Some of these gas-phasemechanisms were compared and evaluated in either box or 3-Dmodels (eg Gross and Stockwell 2003 Faraji et al 2008Luecken et al 2008 Kim et al 2009 2011a Zhang et al 2012)While these mechanisms give similar predictions for gaseousconcentrations in rural and clean conditions the differences insimulated O3 and NOy may be large under urban conditions duemainly to different representations of organic chemistry anddifferent selection of kinetic data for inorganic chemistry Differ-ences in gas-phase mechanisms will lead to differences insecondary aerosol formation Kim et al (2011b) reported lessthan 1 mg m3 (6) differences in monthly-mean PM25 concen-trations but with up to 26 differences in PM25 composition frommodel simulations with the same aerosol module but differentgas-phase mechanisms (ie CB05 and RACM2) Zhang et al (2012)compared CBM-Z CB05 and SAPRC-99 in WRFChemeMADRIDover the continental US and found that different gas-phasemechanisms led to different predictions of concentrations of O3(up to 5 ppb) PM25 (up to 05 mg m3) secondary inorganic PM25species (up to 11 mg m3) and organic PM (up to 18 mg m3)aerosol number concentrations (up to 41) CCN (up to 58) andCDNC (up to 997) These studies illustrated the potentially largeimpact of gas-phase chemical mechanisms on overall modelpredictions

Twenty regional models treat aqueous-phase equilibrium andkinetic reactions and one model (ie STEM-2K3) only treats equi-librium partitioning of SO2 H2O2 and O3 For models with aqueous-phase chemistry the level of details varies from a highly-simplified(eg one first-order oxidation of SO2) to the most comprehensivemechanism (eg the Carnegie-Mellon University (CMU) aqueous-phase mechanism with 50 aqueous-phase species 17 aqueous-phase ionic equilibria 21 gas-phaseaqueous-phase reversiblereactions and 109 aqueous kinetic reactions) that is used in onemodel (ie WRFChemeMADRID) The CMU mechanism explicitly

recasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e24 15

treats the oxidation of dissolved S(IV) by H2O2 O3 O2 catalyzed byFe3thorn and Mn2thorn and radical species as well as the non-reactiveuptake of HNO3 hydrochloric acid (HCl) ammonia (NH3) andother trace gases

Compared with global RT-AQF models regional RT-AQF modelscontain much more detailed aerosol treatments All models exceptfor THOR and five UAQIFS models simulate SO4

2- NO3 NH4

thorn EC OCsodium and chloride UAQIFS-Spain and UAQIFS-Denmark do notsimulate PM Other models only simulate some PM species Wind-blown dust is simulated in 14 out of 36 models (eg ART-AQFSSTEM-2K3 WRFChem CHIMERE SKIRONTAPM) 21 out of 36models simulate detailed aerosol chemistry and dynamics such asthermodynamics for inorganic and organic aerosols nucleationcondensation and coagulation Three models (ie EMEP-UnifiedMATCH and LOTOS-EUROS) only simulate thermodynamics forinorganic aerosols and GEM-CHRONOS simulates an additionalprocess ie SOA formation using a fixed yield approach Ninemodels (eg ART-AQFS MAQSIP-RT THOR) do not simulate anyaerosol chemistry and dynamics

Among the three CTMs in Prevrsquoair PolyphemusPolair3Dcontains drivers for simple forward run (forecast) ensemble fore-cast and various data assimilation methods and offers the mostdetailed chemistry and aerosol treatments For example it includestwo aerosol models a Modal Aerosol Model (MAM Sartelet et al2007) and a SIze REsolved Aerosol Model (SIREAM Debry et al2007) SIREAM simulates detailed aerosol chemistry and dynamicprocesses using state-of-the-science algorithm SOA formation issimulated using either an improved version of the SORGAM model(Kim et al 2011b) which is based on an absorptive partitioningtheory and treats eight SOA classes (4 anthropogenic and 4biogenic) or the AEREPRICaltech (AEC) SOA model (Kim et al2011a) which is based on a hydrophobichydrophilic molecule-based gas-partitioning scheme Currently most SOA modulesused in most RT-AQF models do not take into account the hydro-philic behavior of organic species and only a few account for SOAformation from oxidation of gaseous precursors such as isopreneand sesquiterpene

4 Evaluation protocols and model performance of 3-DRT-AQF models

41 Evaluation protocols

Two types of evaluations are typically conducted to evaluatethe modelrsquos forecasting skills discrete evaluation that uses a set ofstatistical measures and categorical evaluation that uses a set ofcategorical indices Tables 5 and 6 summarize major statisticalmeasures and categorical indices commonly used respectivelyStatistical measures and evaluation protocol for discrete evalua-tion of RT-AQF are the same as those recommended by Seigneuret al (2000) US EPA (2001) and Yu et al (2006) for traditionalAQMs The traditional statistical measures include correlationcoefficient (r) mean bias (MB) mean absolute gross error (MAGE)RMSE mean normalized bias (MNB) mean normalized gross error(MNGE) normalized mean bias (NMB) normalized mean grosserror (NMGE) fractional bias (FB) and fractional gross error (FGE)Some problems may arise with some of those traditionalmeasures For example overpredictions are artificially given moreweight than underpredictions when MNBMNGE and NMBNMGEare used because MNB and NMB can grow disproportionally foroverpredictions but are bounded by 1 for underpredictions(Seigneur et al 2000) A set of new statistical metrics wastherefore developed by Yu et al (2006) based on the concept offactors to overcome the limitations of the traditional measuresThese new metrics include the mean normalized factor bias

Please cite this article in press as Zhang Y et al Real-time air quality forEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

(MNFB) the mean normalized gross factor error (MNGFE) thenormalized mean bias factor (NMBF) and the normalized meanerror factor (NMEF) While MNFB and MNGFE are still subjected tothe dominance by the extremely low observational values in thenormalization NMBF and NMEF provide the most robust statis-tical measures and are therefore recommended for operationalevaluation for model predictions (Yu et al 2006) In additionpeak accuracy measures (eg unpaired peak accuracy (UPA)temporally-paired peak accuracy (TPPA) spatially-paired peakaccuracy (SPPA) and pair peak accuracy (PPA)) index of agree-ment (IOA) and Brier score (BS) are often used in the discreteevaluation UPA quantifies the difference between the magnitudeof the peak observed and simulated values in the modelingdomain regardless whether this occurs at the monitoring stationand time of the peak observed value UPA tests the modelrsquos abilityto reproduce the highest observed value anywhere in the regionUPA is less useful as the two peak values may not occur at thesame location and time TPPA SPPA and PPA provide morestringent measures TPPA tests the modelrsquos ability to reproducethe highest observed value at the hour of the peak observed valueat any location within the modeling domain SPPA tests themodelrsquos ability to reproduce the highest observed value at themonitoring station of the peak observed value at any hour PPAtests the modelrsquos ability to reproduce the highest observed valueat the time and location of the peak observed value which is themost stringent measure among the four peak accuracy measuresIOA provides a measure of how well the departure of the simu-lated values from the observed mean matches with the departureof observations from the observed mean An IOA of 1 indicatesa perfect agreement BS measures the average squared deviationbetween predicted probabilities for a set of events and theiroutcomes so a lower score represents a higher accuracy It isa measure used frequently for probabilistic forecasts

The traditional categorical indices include accuracy (A) criticalsuccess index (CSI) or threat score (TS) probability of detection(POD) (or hit rate (HRate)) bias (B) false alarm ratio (FAR) falsealarm rate (FARate) Heidke skill score (HSS) Pierce skill score(PSS) (or true skill score (TSS) or Hansen and Kuipers discriminant(HKD)) skill score (SS) and economic value (EV) A indicates thepercentage of forecasts that correctly predicts an exceedance ora nonexceedance CSI indicates the percentage of the forecastedactual exceedances to all actual and forecasted exceedances PODor HRate indicates the percentage of the forecasted actualexceedances to all actual exceedances CSI is a more balancedscore because it takes into account both missed exceedances andfalse alarms whereas POD does not take into account false alarmsB judges if forecasts are underpredicted (lt1 also referred to asunderforecasting or ldquofalse negativerdquo) or overpredicted (gt1 alsoreferred to as overforecasting or ldquofalse positiverdquo) FAR measuresthe percentage of times an exceedance was forecast whenexceedance did not occur FARate indicates the proportion ofnonexceedances that were incorrectly forecasted (ie the proba-bility of false detection) it is the counterpart to the hit rate PODor HRate HSS measures the fractional improvement of the fore-cast (ie the total number of correct forecasts minus the correctrandom forecasts) over the standard random chance forecasts (iethe total number of forecasts minus the correct forecasts due tochance) Like most skill scores it is normalized by the total rangeof possible improvement over the standard which means HSS cansafely be compared on different datasets PSS (or TSS or HKD)measures skills relative to an unbiased random reference forecastby taking the differences between POD and FARate SS measuresthe relative accuracy with respect to a set of reference forecastsThe most convenient reference is the persistence forecast Anyvalid error metrics such as RMSE and NME can be used to calculate

ecasting part I History techniques and current status Atmospheric

Table 5Discrete statistical measures used in the model evaluation (compiled from Willmott 1981 Zhang et al 2006a Yu et al 2006 Dutot et al 2007)

Metrics Mathematical Expressiona Range

Correlation Coefficient r frac14(XN

ifrac141

ethMi MTHORNethOi OTHORN)(XN

ifrac141

ethMi MTHORN2XNifrac14 1

ethOi OTHORN2)1

2

1 to 1

Mean Bias (MB) MB frac14 1N

XNifrac141

ethMi OiTHORN frac14 M O N to thornN

Mean AbsoluteGross Error (MAGE)

MAGE frac14 1N

XNifrac141

jMi Oij 0 to thornN

Root MeanSquare Error (RMSE)

RMSE frac141N

XNifrac141

ethMi OiTHORN21

2

0 to thornN

Mean NormalizedBias (MNB)

MNB frac14 1N

XNifrac141

frac12ethMi OiTHORN=Oi frac141N

XNifrac141

ethMi=Oi 1THORN 1 to thornN

Mean NormalizedGross Error (MNGE)

MNGE frac14 1N

XNifrac141

frac12ethjMi OijTHORN=Oi 0 to thornN

NormalizedMean Bias (NMB)

NMB frac14XNifrac141

ethMi OiTHORNXN

ifrac14 1

Oi frac14 M

O 1 1 to thornN

NormalizedMean Error (NME)

NME frac14XNifrac141

jMi OijXN

ifrac141

Oi frac14 MAGE=O 0 to thornN

Fractional Bias (FB) FB frac14 1N

XNifrac141

Mi Oi

ethMi thorn OiTHORN=22 to thorn2

Fractional GrossError (FGE)

FGE frac14 1N

XNifrac141

jMi OijethMi thorn OiTHORN=2

0 to 2

Mean NormalizedFactor Bias (MNFB)

MNFB frac14 1N

XNifrac141

Fi where Fi frac14 Mi=Oi 10 for Mi Oi

Fi frac14 10 Oi=Mi forMi lt Oi

N to thornN

Mean NormalizedGross FactorError (MNGFE)

MNGFE frac14 1N

XNifrac141

jFij where Fi frac14 Mi=Oi 10 for Mi Oi

Fi frac14 10 Oi=Mi forMi lt Oi

0 to thornN

Normalized MeanBias Factor (NMBF)

NMBF frac14XNifrac141

Mi

XNifrac141

Oi 1 frac14 M=O 1 for M O

NMBF frac14 1XNifrac141

Oi

XNifrac141

Mi frac14 eth1 O=MTHORN for M lt O

N to thornN

Normalized Mean ErrorFactor (NMEF)

NMEF frac14XNifrac14 1

jMi OijXN

ifrac141

Oi frac14 MAGEO

for M O

NMEF frac14XNifrac14 1

jMi OijXN

ifrac141

Mi frac14 MAGEM

for M lt O

0 to thornN

Unpaired peakaccuracy (UPA)

UPA frac14Mmaxxt O

maxbxbtOmaxbxbt 1 to thornN

Temporally-paired peakaccuracy (PPA)

TPPA frac14M

maxxbt Omaxbxbt

Omaxbxbt 1 to thornN

Spatially-paired peakaccuracy (SPPA)

SPPA frac14M

maxbxt Omaxbxbt

Omaxbxbt 1 to thornN

Paired peak accuracy (PPA) PPA frac14M

maxbxbt Omaxbxbt

Omaxbxbt 1 to thornN

Index of Agreement (IOA) IOA frac14 1PN

ifrac14 1ethMi OiTHORN2PNifrac141ethjMi Oj thorn jOi OjTHORN2

0 to 1

Brier Score (BS) BS frac14 1K

XKtfrac141

ethpt otTHORN2 0 to thorn1

N is the number of samples (by time andor location)Mmaxbxbt and O

maxbxbt are time- and space-paired peak simulated and observed values with the circumflex indicating thatthe peak value occurs at the time t and location x of the peak observed value pt and ot are the probability that an event occurs according to an ensemble and the event actuallyoccurred or not at instance t (0 if it doesnrsquot happen and 1 if it happens) respectively K is the number of forecasting instances

a M frac14 eth1=NTHORNPNifrac141 MiO frac14 eth1=NTHORNPN

ifrac141 Oi Mi and Oi are values of model predictions and observations at time or location i respectively

Y Zhang et al Atmospheric Environment xxx (2012) 1e2416

SS For a perfect model A CSI POD HSS PSS and SS would be 1FAR and FARate would be zero and B would be unity

The above traditional categorical metrics provide ldquoclear cutrdquomeasures of the modelrsquos ability to predict an ldquoexceedancerdquo by usinga fixed threshold concentration and time- and space-paired

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

observationeforecast sets A main limitation is that they cannotprovide information regarding the spatial variations of pollutantconcentrations within a region of interest To overcome this limi-tation Kang et al (2007) developed three new categorical metricsfor RT-AQF evaluation the weighted success index (WSI) area hit

recasting part I History techniques and current status Atmospheric

Table 6Categorical statistical measures used in the model evaluation (compiled based on Ryan 1995 McHenry et al 2004 Kang et al 2005 2007 Pagowski and Grell 2006)

Metrics Mathematical Expressiona Range

Accuracy A frac14 bthorn cathorn bthorn cthorn d

0e1

CSI frac14 bathorn bthorn d

0e1

Critical Success Index or Threat Score (TS) WSI frac14 bthornPn1 IP

athorn bthorn d0e1

The Weighted Success Index (WSI)bIP frac14 Mi f Oi

Mi f T for Mi gt T gt Oi and Mi lt f T

IP frac14 Oi f Mi

Oi f T for Oi gt T gt Mi and Oi lt f Mi

IP frac14 0 for other conditions

Probability Of Detection Or Hit Rate POD frac14 H frac14 bbthorn d

0e1

Area Hit Ratec aH frac14 AbAbthorn Ad

0e1

Bias or Bias Ratio B frac14 athorn bbthorn d

0 to thornN

False Alarm Ratio FAR frac14 aathorn b

0e1

Area False Alarm Ratiod aFAR frac14 AaAathorn Ab

0e1

False Alarm rate FARate frac14 F frac14 aathorn c

0e1

Heidke skill score HSS frac14 2ethbc adTHORNethbthorn dTHORNethcthorn dTHORN thorn ethathorn bTHORNethathorn cTHORN N to 1

Pierce skill score (or true skill score (TSS)or Hansen and Kuipers discriminant)

PSS frac14 POD FARate frac14 bbthorn d

aathorn c

frac14 bc adethathorn cTHORN ethbthorn dTHORN 1 to 1

Skill scoreef SS frac14 Em ErefEperf Eref

0e1

Em frac14 RMSE frac141N

XNifrac141

ethMi OiTHORN21

2

or NME frac14XNifrac14 1

jMi OijXN

ifrac141

Oi frac14 MAGE=O

Eref frac14 RMSE frac141N

XNifrac14 1

ethMi KiTHORN21

2

or NME frac14XNifrac14 1

jMi MjjXN

jfrac141

Ki frac14 MAGEref=K

Eperf frac14 0

Economic value (EV) EV frac14 minethr cf THORN Feth1 cf THORNr thorn H cf eth1 rTHORN cfminethr cf THORN cf r

0e1

r frac14 CLfrac14 athorn b

d

H frac14 bbthorn d

F frac14 aathorn c

cf is the climatological frequency of an event

a Where a b c d are the number of simulated and observed data pairs at one site at a specific time in the four regions (see Table below) that represent forecast exceedancesthat did not occur forecast exceedances that did occur forecast nonexceedances that did occur and forecast nonexceedances that did not occur respectivelyContingencytable for threshold forecasts

ObservedYes No

Forecast Yes b aNo d c

b Mi and Oi are values of model prediction and observation at a location i respectively T is the threshold value used f is an empirical factor that reflects the zone in which themost points are located on the predictioneobservation scatterplots its typical value can be 1e2

c Ab is the number of exceedances that are both observed and forecast but the forecast is any exceedance that occurs in the designated area centered at themonitor locationAd is the number of observed exceedances that are not forecast within the designated area centered at the monitor location

d Aa is the number of forecast area exceedances that were not observed and Ab is the number of forecast exceedances that were observede Mi Oi and Ki are values of model prediction observation and temporal or spatial persistence value at time or location i respectively Temporal persistence value is the

previous days forecast value Spatial persistence value is the observation at the nearest location to the location i O frac14 eth1=NTHORNPNifrac141 Oi K frac14 eth1=NTHORNPN

ifrac141 Ki N is the number ofsamples (by time andor location) RMSE e Root mean square error NME e Normalized mean error MAGE e Mean Absolute Gross Error

f Em Eref and Eperf are the errors associated with model reference and perfect forecast respectively Em and Eref can be any valid error metrics eg RMSE and NME Thetemporal or spatial persistence forecast is often used as a reference forecast a perfect forecast has a zero error ie Eperf frac14 0

Y Zhang et al Atmospheric Environment xxx (2012) 1e24 17

Please cite this article in press as Zhang Y et al Real-time air quality forecasting part I History techniques and current status AtmosphericEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

Y Zhang et al Atmospheric Environment xxx (2012) 1e2418

(aH) and area false-alarm ratio (aFAR) WSI is less stringent thanCSI but a more representative measure It provides partial credit toobservationeforecast pairs that fall just outside limits establishedby CSI butwithin a designated factor that is used to judge themodelperformance (eg data pairs fall into a factor of 1e2 of theobservations)

An important practical aspect of RT-AQFs is how forecasts canbe used to mitigate the impact of air quality on economy andpeople In the calculation of EV one needs to know the climato-logical frequency (cf) of an event in addition to the values of a b cand d (see definitions in Table 6) Hits and false alarms incur a costof taking a preventive action (C) while misses are associated withlosses due to the lack of prevention (L) Correct rejections (c) incurno expense EV is defined as the reduction in mean expense rela-tive to the reduction (both defined with respect to cf) that wouldoccur if all the forecasts were correct (Pagowski and Grell 2006) Asfor weather forecasting EV provides a useful measure for theeconomic benefits of the existing RT-AQFs and the benefits ofimproving these systems to individuals businesses and society

No recommended values of performance statistics were providedfor acceptable forecasting skills by any individual countries Therecommended values for retrospective AQ modeling can providesome guidance for RT-AQF skill evaluation For example Zhang et al(2006a) recommended the use of MNBs and NMBs 15 MNGEsand NMEs 30 as indications of a satisfactory performance for O3

and PM25 Boylan and Russell (2006) recommended the use of meanfractional bias (MFB) and mean fractional error (MFE) for PM withgood performance indicted by MFBs 30 and MFEs 50 forPM25 and MFBs 60 and MFEs 75 for major PM componentsthat have concentrations 225 mg m3 The US EPA recommendedthat a so-calledweight of evidence approach inwhich a set of diverseanalyses are used to judge model performance Since no singlestatistical metrics can fully reflect RT-AQF skill discrete and cate-gorical statistics summarized in Tables 5 and 6 are recommended toprovide a comprehensive evaluation of an RT-AQF system

42 Observational datasets for model evaluation

Datasets for forecasting evaluation and improvement includereal-time or near real-time surface measurement datasets suchas the US EPArsquos AIRNOW datasets obtained by aircraft shipozonesondes and lidars during special field campaigns andsatellite data Satellite observations can complement datasetsfrom surface networks by providing derived chemical measure-ments Recent work has shown the potential of using near real-time satellite and surface data to improve RT-AQF (eg Al-Saadiet al 2005 Kondragunta et al 2008 Wang et al 2011) OverEurope the Global and Regional Earth-System MonitoringUsing Satellite and In situ Data (GEMS) and MACC projects areaimed at a near real-time data assimilation and forecast capabilityfor aerosols greenhouse gases and reactive gases The dataset canbe used to monitor the composition infer estimates of surfacefluxes and produce global short-range and medium-rangeforecasts

43 Current model forecasting skills

Most evaluations of forecasted O3 and its precursors focus onsummer and very few include winter (eg Manins et al 2002 Caiet al 2008 Doraiswamy et al 2009) Forecasting PM is moredifficult than forecasting O3 because PM consists of multiplechemical components over a broad size spectrum PM modeling isstarting to reach sufficient maturity for transition from research tooperational use Several PM models have recently been transferredinto operational models to forecast PM (Carmichael et al 2003

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

McHenry et al 2004 McKeen et al 2005 2007 Yu et al 2008Chuang et al 2011) very few studies provide detailed evaluationsof forecasted PM and its composition and precursors (eg McKeenet al 2007 2009 Yu et al 2008 Chen et al 2008 Manders et al2009) Similar to O3 forecasts most evaluations are conducted forsummer episodes and very few for winter episodes (eg Mathuret al 2008 Konovalov et al 2009) or a full year (eg Manderset al 2009) Limited evaluation was performed for coarse parti-cles such as mineral dust (eg Jimeacutenez-Guerrero et al 2008 Niuet al 2008 Menut et al 2009) that may be of major concernsfor PM10 attainment in some regions such as Asia southern Europeand the western US Jimeacutenez-Guerrero et al (2008) showed thatthe inclusion of a dust emission module substantially increases theaccuracy of both discrete and categorical statistics in the IberianPeninsula in Europewhere the influence of the Saharan dust cannotbe neglected

Table A4 summarizes the evaluation of a number of RT-AQFsystems in terms of domain and period and discrete and categor-ical performance statistics While maximum 1-h O3 averagestatistics are generally satisfactory those for maximum 8-h averageand hourly O3 sometime exceed NMB of 15 and NME of 30 forsome models (eg EtaCMAQ WRF-NMMCMAQ) over the easternor northeastern US in the O3 season during 2004e2006 indicatinga relatively poor performance The overpredictions in the low O3range (lt50 ppb) were reported in several studies (eg Yu et al2007 Chen et al 2008) Several factors may contribute to suchoverpredictions These include a poor representation of thenocturnal PBL mixing height (Gilliam et al 2006 Zhang et al2006b Eder et al 2006) an excessive downward transport ofhigh level O3 aloft and too much photolysis under high cloudconditions (Eder et al 2006) a high O3 production rate with theSAPRC-99 chemical mechanism (Arnold and Dennis 2006) themodel limitation in resolving titration of O3 by NO in urban plumes(Yu et al 2007) and the BCONs from global models (Chen et al2008) Chuang et al (2011) applied WRFChemeMADRID forRT-AQF and found that O3 overprediction in most regions in thesoutheastern US is likely caused by inaccurate emissions ofprecursors such as biogenic VOCs positive biases in 2-m temper-ature and negative biases in wind speed at 10-m O3 under-predictions in some regions could be due in part to theuncertainties in lateral BCONs

For categorical evaluation different threshold values were usedfor the same variables For example the threshold values used formaximum 1-h average O3 maximum 8-h average O3 and hourly O3are 60e125 ppb 65e85 ppb and 80 ppb respectively Some ofthese values are lower than the former US NAAQS of 120 ppb formaximum 1-h average O3 and the current US NAAQS of 75 ppb formaximum 8-h O3 respectively for the reasons stated previously(eg Hogrefe et al 2007 Yu et al 2007 Chuang et al 2011)although higher thresholds were used in some earlier applications(eg Kang et al 2005) For maximum 1-h average O3 A CSI POD Band FAR range from 15 to 998 52 to 216 69 to 89 01 to53 ppb and 03 to 941 respectively Prevrsquoair gives the lowest A of15e41 because it underestimates O3 daily maxima at high O3concentrations (Honoreacute et al 2008) All RT-AQFs give low values ofCSI and POD but higher values of FAR because they fail to forecastexceedance (ie a low value of b) but overpredict lowO3 (ie a highvalue of a) A CSI POD B and FAR for maximum 8-h average O3

range from 762 to 998 0 to 532 0 to 848 03 to 170 ppb and13 to 991 respectively Compared with maximum 1-h O3 thevalues of CSI and POD are higher and those of FAR are lower forcorresponding maximum 8-h O3 The categorical evaluation forhourly O3 was conducted by only one model (ie WRFChemeMADRID) with A CSI POD B and FAR of 992 25188 68 ppb and 972 respectively Very few evaluations were

recasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e24 19

conducted for precursors of O3 and other gaseous species (eg Caiet al 2008) and vertical profile of forecasted concentrations (egYu et al 2007) Yu et al (2007) reported that EtaCMAQ reproducedO3 vertical distributions on most of the days at low altitudes butoverpredictions occurred at altitudes gt 6 km because of a combi-nation of effects related to the specifications of meteorologicallateral BCONs as well as themodelrsquos coarse vertical resolution in theupper free troposphere

Surface PM forecasts are evaluated in terms of 24-h averagePM25 hourly PM25 and 24-h average PM10 These evaluationshowever have been very limited to date and a consistent evaluationprotocol has not yet been well established (Seigneur 2001 Zhanget al 2006a) The forecasting skill for PM25 is overall poorer thanthat for O3 For 24-h average PM25 MBs RMSEs NMBs and NMEsrange from 32 to 62 mg m3 17 to 159 mg m3 21 to 32 and37 to 81 respectively The NMB of 21 using EtaCMAQ over theeastern US during July 14eAugust 18 2004 was reported by Yuet al (2008) who attributed underpredictions to underestimatedtotal carbonaceous PM at both urban and rural sites and a signifi-cant underestimation of unspecified anthropogenic PM mass(mainly consisting of primary emitted trace elements) at rural sitesOn the other hand Hogrefe et al (2007) showed that EtaCMAQoverpredicted PM25 in the New York City due to overpredictions inorganic aerosols and crustal material Factors contributing to suchoverpredictions include underpredictions of nocturnal verticalmixing inaccurate temporal allocation of primary OM emissionsand underestimate of deposition processes Chen et al (2008)reported the NMB of 17e32 over the Pacific Northwest duringAugusteNovember 2004 The overpredictions in PM25 wereattributed to uncertainties in wildfire emission estimates and themodeling error in fire plume transport due to errors in MM5-predicted wind direction and wind speed McKeen et al (2007)evaluated RT-AQF of 6 models including WRFChem at two gridresolutions (12- and 36-km) CHRONOS AURAMS STEM-2K3 andEtaCMAQ They found that most models did not reproduce theobserved diurnal variation at urban and suburban sites particularlyduring the nighttime to early morning and the ensemble meanbased on 6 models with equal weighting gave the best possibleforecast in terms of statistics Chuang et al (2011) showed slightunderpredictions of PM25 in the O3 season over the southeasternUS by WRFChemeMADRID which were attributed to uncer-tainties in emissions such as those of biogenic VOCs and NH3overpredictions of precipitation and uncertainties in the BCONsOnly two studies evaluated hourly PM25 givingMBs RMSEs NMBsand NMEs of 33 to06 mg m3 83e113 mg m3 211 to52and 498e514 respectively For 24-h average PM10 MBs andRMSEs range from 29 to 138 mg m3 and 83 to 472 mg m3respectively Berge et al (2002) reported forecasted PM10 perfor-mance over Oslo with an MB of 209 mg m3 and an NMBof 365 at Kirkeveien due to an inaccurate emissions of PM10from the surface (ie the re-suspension of dust deposited at theroadside) on dry days and an MB of 138 mg m3 and an NMB of 24at Furuset due to errors in simulated grid-averaged wind fields

Very few evaluations were conducted for PM componentsMcKeen et al (2007) found that all the six RT-AQF models signifi-cantly underpredicted OM at the surface and overestimated SO4

2

above 2 km The overpredictions in SO42 were attributed to over-

estimate of SO2 by WRFChem and CHRONOS and the inclusion ofaqueous-phase oxidation of SO2 by AURAMS CMAQETA andSTEM-2K3 Yu et al (2007) reported that EtaCMAQ overpredictedSO4

2 due to too much in-cloud SO2 oxidation as a result of over-estimated H2O2 concentrations in the model underpredicted NH4

thorn

at the rural sites and aloft due to the exclusion of some sources ofNH3 in the real-time emission inventory underpredicted NO3

dueto overpredictions of SO4

2 and underpredicted OC due to missing

Please cite this article in press as Zhang Y et al Real-time air quality forEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

sources of primary OC in the emission inventory and missing SOAformation from the gas-phase oxidation of isoprene and sesqui-terpenes and aqueous-phase oxidation of glyoxal and methyl-glyoxal Chen et al (2008) reported that AIRPACT-3 reproduced OCwell but significantly overpredicted EC and significantly under-predicted SO4

2 NO3 and NH4

thorn due to underestimation of emis-sions of primary PM species (eg sulfate) and precursors ofsecondary PM (eg SO2 NH3 and NOx) insufficient spatial reso-lution and modelrsquos inability to capture the hourly PM variations

For categorical evaluation of PM some studies used thresholdvalues of 15e315 mg m3 for 24-h average PM25 and 30e50 mgm3

for 24-h average PM10 that are lower than the NAAQSs of 24-haverage PM25 of 35 mg m3 and 24-h average PM10 of 150 mg m3

for the same reason as mentioned previously For 24-h averagePM25 A CSI POD B and FAR range from 608 to 997 0 to 5370 to 909 07 to 19 mg m3 and 25 to 100 respectively Those for24-h average PM10 range from 167 to 100 91 to 100 150 to100 05 to 09 mg m3 and 0 to 48 respectively Categoricalevaluation of hourly PM25 was conducted only with WRFChemeMADRID with A CSI POD B and FAR of 721 205292 07 mg m3 and 591 respectively

5 Summary of Current States of RT-AQF

Emerging as a new discipline of applied sciences that integratesseveral physical sciences mathematicalstatistical tools andcomputer technology in the 1970s RT-AQF represents a uniqueapplied science affecting humanrsquos daily activities and posesunprecedented challenges intersecting many aspects of sciencesand technologies Driven by deteriorated air quality worldwideincreased societalhuman demands and rapid scientific and tech-nological advancements significant progress has been made in thepast four decades in RT-AQF A number of RT-AQF tools and modelswith varying degrees of sophistication and forecasting skillsranging from the simplest rule of thumb to the most advanced 3-Donline-coupled meteorologyechemistry models have been devel-oped Among them the deterministic physically-based RT-AQFmodels that are based on 3-D global and regional CTMs representthe state of the science enabling scientific understanding ofmechanisms for pollutant formation and development of emissioncontrol strategies Although other approaches in particular para-metric (statistical) models will continue to be used the 3-D RT-AQFmodels will likely become prevalent approaches and tools on all thescales There is a growing trend to combine these 3-D modelswith statistical models to provide more accurate RT-AQF Amongall RT-AQF models the coupled meteorologyechemistry modelrepresents a significant advancement and will greatly enhance theunderstanding of the underlying complex interplay of meteorologyemissions and chemistry in the real atmosphere Importantextensions of the RT-AQF models include their coupling with anurban model (eg traffic andor local pollutant dispersion) or a CFDmodel for urbanlocal scale applications at a spatial resolution of1 km or less and with an exposure model to provide real-timepublic health assessment and exposure predictions and urbanemergency preparedness

3-D RT-AQF models have been used for 24e72 h RT-AQFworldwide since the mid 1990s Despite their unique characteris-tics and technical requirements 3-D RT-AQF models are based onNWP models and CTM development and application efforts in thepast decades and therefore are subject to limitations of thesemodels Although no universal forecasting guidance and evaluationprotocols were developed covering all countries some guidanceand protocols were recommended by governmental agencies andthe scientific community Discrete and categorical evaluations aretwo most commonly-used evaluations Real-time or near real-time

ecasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e2420

surface air quality measurement and satellite datasets areincreasingly available for such evaluations The evaluation effortshave been primarily focused on O3 NO2 and PM (both PM10 andPM25) at surface level The forecasting skill is typically better for O3than for PM25 While their performance statistics averaged overa time period (eg monthly or daily) and the whole domain isgenerally good or satisfactory current RT-AQF models have diffi-culties forecasting variations at finer time scales (hourly and diur-nally) and individual monitoring stations In addition larger biasesexist for forecasted NOx HONO HNO3 and OH and PM compositionincluding SO4

2 NO3 and OM indicating possible error compen-

sations that lead to current levels of forecast accuracy These fore-cast inaccuracies have been attributed to a number of factors thatare highly uncertain or lack of accurate treatments in the RT-AQFmodels including mesoscale meteorological processes (eg sea-breeze circulations) and variables (eg temperature PBL height)chemical BCONs of O3 and PM25 emissions (eg SO2 NOx NH3BVOCs primary PM) physical and chemical processes (eg urbanprocesses in-cloud oxidation of SO2 SOA formation) and modelconfiguration (eg coarse grid resolution) Recent advances havebeenmade in some of those aspects Part II of this reviewwill assessand discuss these advances along with advanced computationaltechniques that can potentially improve RT-AQF Recommendationsfor research priorities and future prospects will also be provided

Acknowledgments

This work was performed under the Fellowship Award(704389J) Center for Education and ResearchEacutecole des PontsParisTech France through a visiting professorship at the Atmo-spheric Environment Center (CEREA) the Joint Laboratory Eacutecole desPonts ParisTech and EDF RampD Paris France the NSF Career Award(ATM-0348819) the EPA-Science to Achieve Results (STAR)program (R83337601) and Chinarsquos National Basic ResearchProgram (2010CB951803) Support from CEEH MEGAPOLI andCOST ES1004 is also acknowledged Thanks are due to Ming-TungChuang a former post-D researcher at NCSU for compiling anearlier version of Table A4 Thanks are due to Yelva RoustanAtmospheric Research Center (CEREA) France Christian Hogrefethe New York State Department of Environmental Conservation(NYSDEC) NY US and John McHenry Baron Advanced Meteoro-logical Systems (BAMS) US for providing some informationregarding their air quality forecasting models

Appendix A Supplementary Material

Supplementary data related to this article can be found online athttpdxdoiorg101016jatmosenv201206031

References

Alhanafy TE Zaghlool F El AS Moustafa D 2010 Neuro fuzzy modeling schemefor the prediction of air pollution J American Sci 6 (12) 605e616

Al-Saadi J Szykman J Pierce RB Kittaka C Neil D Chu DA Remer LGumley L Prins E Weinstock L MacDonald C Wayland R Dimmick FFishman J 2005 Improving national air quality forecasts with satellite aerosolobservations Bull Amer Meteor Soc 86 1249e1261

Arnold JR Dennis RL 2006 Testing CMAQ chemistry sensitivities in base caseand emissions control runs at SEARCH and SOS99 surface sites in the south-eastern US Atmos Environ 40 5027e5040

Aron RH 1980 Forecasting high level oxidant concentrations in the Los Angelesbasin J Air Pollut Control Assoc 20 1227e1228

Baklanov A Haumlnninen O Slordal LH Kukkonen J Bjergene N Fay B Finardi SHoe SC Jantunen M Karppinen A Rasmussen A Skouloudis A Sokhi RSSorensen j H Oslashdegaard V 2007 Integrated systems for forecasting urbanmeteorology air pollution and population exposure Atmos Chem Phys 7855e874 wwwatmos-chem-physnet78552007

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

Baklanov A Korsholm U Mahura A Petersen C Gross A 2008 Enviro-HIRLAMon-line coupled modelling of urban meteorology and air pollution Adv Sci Res2 41e46 wwwadv-sci-resnet2412008

Baklanov A Mahura A Sokhi R (Eds) 2011 Integrated Systems of Meso-Meteorological and Chemical Transport Models first ed Springer ISBN 978-3-642-13979-6 p 242

Baklanov A 2006 Overview of the European project FUMAPEX Atmos ChemPhys 6 2005e2015

Baklanov A 2010 Chemical weather forecasting a new concept of integratedmodeling Adv Sci Res 4 23e27

Baldasano JM Jimeacutenez-Guerrero P Jorba O Peacuterez C Loacutepez E Guumlereca PMartin F Garciacutea-Vivanco M Palomino I Querol X Pandolfi M Sanz MJDieacuteguez JJ 2008 CALIOPE an operational air quality forecasting system for theIberian Peninsula Balearic Islands and Canary Islands e First annual evaluationand ongoing developments Adv Sci Res 2 89e98 ISSN 1992-0628 May

Battan LJ 1966 The Unclean Sky Doubleday Garden City NY 141 ppBerge E Walker S-E Sorteberg A Lenkopane M Eastwood S Jablonska HI

Koltzow MOslash 2002 A real-time operational forecast model for meteorologyand air quality during peak air pollution episodes in Oslo Norway Water AirSoil Poll Focus 2 745e757

Blond N Vautard R 2004 Three-dimensional ozone analysis and their use forshort-term ozone forecasts J Geophys Res 109 D17303 httpdxdoiorg1010292004JD004515

Boylan JW Russell AG 2006 PM and light extinction model performance metricsgoals and criteria for three-dimensional air quality models Atmos Environ 40(26) 4946e4959 httpdxdoiorg101016jatmosenv200509087

Brandt J Christensen JH Frohn LM Palmgren F Berkowicz R Zlatev Z 2001Operational air pollution forecasts from European to local scale Atmos Environ35 (Suppl 1) S91eS98

Brimblecombe P 1987 The Big Smoke a History of Air Pollution in London sinceMedieval Time Routledge Revivals Methuen London UK 186 pp

Burrows WR Benjamin M Beauchamp S Lord ER McCollor D Thomson B1995 CART decision-tree statistical analysis and prediction of summer seasonmaximum surface ozone for the Vancouver Montreal and Atlantic Regions ofCanada J Appl Meteor Climatol 34 1848e1862

Byun D Schere KL 2006 Review of the governing equations computationalalgorithms and other components of the models-3 community multiscale airquality (CMAQ) modeling system Appl Mech Rev 59 51e77

Cai C Hogrefe C Katsafados P Kallos G Beauharnois M Schwab JJ Ren XBrune W Zhou X He Y Demerjian K 2008 Performance evaluation of an airquality forecast modeling system for a summer and winter season e Photo-chemical oxidants and their precursors Atmos Environ 42 8585e8599 httpdxdoiorg101026jatmosenv200808029

Carmichael GR et al 2003 Regional-scale chemical transport modeling in supportof intensive field experiments overview and analysis of the TRACE-P observa-tions J Geophys Res 108 (D21) 8823 httpdxdoiorg1010292002JD003117

Carmichael GR Sandu A Chai T Daescu DN Constantinescu EM Tang Y2008 Predicting air quality improvements through advanced methods tointegrate models and measurements J Comput Phys 227 3540e3571

Chai T Carmichael GR Tang Y Sandu A Streets DG 2007 Four-dimensionaldata assimilation experiments with international consortium for atmosphericresearch on transport and transformation ozone measurements J GeophysRes 112 D12S15 httpdxdoiorg1010292006JD007763

Chaloulakou A Saisana M Spyrellis N 2003 Comparative assessment of neuralnetworks and regression models for forecasting summertime ozone in AthensSci Total Environ 313 1e13

Chen J Vaughan J Avise J OrsquoNeill S Lamb B 2008 Enhancement and evalu-ation of the AIRPACT ozone and PM25 forecast system for the Pacific NorthwestJ Geophys Res 113 D14305 httpdxdoiorg1010292007JD009554

Chenevez J Jensen CO 2001 Operational ozone forecasts for the region of Copen-hagen by the Danish Meteorological Institute Atmos Environ 35 4567e4580

Chin M Ginoux P Lucchesi R Huebert B Weber R Anderson T Masonis SBlomquist B Bandy A Thornton D 2003 A global aerosol model forecast forthe ACE-Asia field experiment J Geophys Res 108 (D23) 8654 httpdxdoiorg1010292003JD003642

Chuang M-T Zhang Y Kang D-W 2011 Application of WRFChem-MADRID forreal-time air quality forecasting over the southeastern United States AtmosEnviron 45 (34) 6241e6250

Cobourn WG Hubbard MC 1999 An enhanced ozone forecasting model using airmass trajectory analysis Atmos Environ 33 4663e4674

Cobourn WG 2007 Accuracy and reliability of an automated air quality forecastsystem for ozone in seven Kentucky metropolitan areas Atmos Environ 415863e5875 httpdxdoiorg101016jatmosenv200703024

Comrie AC 1997 Comparing neural networks and regression models for ozoneforecasting J Air amp Waste Manag Assoc 47 653e663

Cooke WF Liousse C Cachier H Feichter J 1999 Construction of a fossil fuelemission data set for carbonaceous aerosol and implementation and radiativeimpact in the ECHAM4 model J Geophys Res 104 22137e22162

Cope ME Hess GD 2005 Air quality forecasting a review and comparison of theapproaches used internationally and in Australia Clean Air and Environ Quality39 (1) 52e60

Cope ME Hess GD Lee S Tory K Azzi M Carras J Lilley W Manins PCNelson P Ng L Puri K Wong N Walsh S Young M 2004 The Australianair quality forecasting system Part I project description and early outcomesJ Appl Meteorol 43 649e662

recasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e24 15

treats the oxidation of dissolved S(IV) by H2O2 O3 O2 catalyzed byFe3thorn and Mn2thorn and radical species as well as the non-reactiveuptake of HNO3 hydrochloric acid (HCl) ammonia (NH3) andother trace gases

Compared with global RT-AQF models regional RT-AQF modelscontain much more detailed aerosol treatments All models exceptfor THOR and five UAQIFS models simulate SO4

2- NO3 NH4

thorn EC OCsodium and chloride UAQIFS-Spain and UAQIFS-Denmark do notsimulate PM Other models only simulate some PM species Wind-blown dust is simulated in 14 out of 36 models (eg ART-AQFSSTEM-2K3 WRFChem CHIMERE SKIRONTAPM) 21 out of 36models simulate detailed aerosol chemistry and dynamics such asthermodynamics for inorganic and organic aerosols nucleationcondensation and coagulation Three models (ie EMEP-UnifiedMATCH and LOTOS-EUROS) only simulate thermodynamics forinorganic aerosols and GEM-CHRONOS simulates an additionalprocess ie SOA formation using a fixed yield approach Ninemodels (eg ART-AQFS MAQSIP-RT THOR) do not simulate anyaerosol chemistry and dynamics

Among the three CTMs in Prevrsquoair PolyphemusPolair3Dcontains drivers for simple forward run (forecast) ensemble fore-cast and various data assimilation methods and offers the mostdetailed chemistry and aerosol treatments For example it includestwo aerosol models a Modal Aerosol Model (MAM Sartelet et al2007) and a SIze REsolved Aerosol Model (SIREAM Debry et al2007) SIREAM simulates detailed aerosol chemistry and dynamicprocesses using state-of-the-science algorithm SOA formation issimulated using either an improved version of the SORGAM model(Kim et al 2011b) which is based on an absorptive partitioningtheory and treats eight SOA classes (4 anthropogenic and 4biogenic) or the AEREPRICaltech (AEC) SOA model (Kim et al2011a) which is based on a hydrophobichydrophilic molecule-based gas-partitioning scheme Currently most SOA modulesused in most RT-AQF models do not take into account the hydro-philic behavior of organic species and only a few account for SOAformation from oxidation of gaseous precursors such as isopreneand sesquiterpene

4 Evaluation protocols and model performance of 3-DRT-AQF models

41 Evaluation protocols

Two types of evaluations are typically conducted to evaluatethe modelrsquos forecasting skills discrete evaluation that uses a set ofstatistical measures and categorical evaluation that uses a set ofcategorical indices Tables 5 and 6 summarize major statisticalmeasures and categorical indices commonly used respectivelyStatistical measures and evaluation protocol for discrete evalua-tion of RT-AQF are the same as those recommended by Seigneuret al (2000) US EPA (2001) and Yu et al (2006) for traditionalAQMs The traditional statistical measures include correlationcoefficient (r) mean bias (MB) mean absolute gross error (MAGE)RMSE mean normalized bias (MNB) mean normalized gross error(MNGE) normalized mean bias (NMB) normalized mean grosserror (NMGE) fractional bias (FB) and fractional gross error (FGE)Some problems may arise with some of those traditionalmeasures For example overpredictions are artificially given moreweight than underpredictions when MNBMNGE and NMBNMGEare used because MNB and NMB can grow disproportionally foroverpredictions but are bounded by 1 for underpredictions(Seigneur et al 2000) A set of new statistical metrics wastherefore developed by Yu et al (2006) based on the concept offactors to overcome the limitations of the traditional measuresThese new metrics include the mean normalized factor bias

Please cite this article in press as Zhang Y et al Real-time air quality forEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

(MNFB) the mean normalized gross factor error (MNGFE) thenormalized mean bias factor (NMBF) and the normalized meanerror factor (NMEF) While MNFB and MNGFE are still subjected tothe dominance by the extremely low observational values in thenormalization NMBF and NMEF provide the most robust statis-tical measures and are therefore recommended for operationalevaluation for model predictions (Yu et al 2006) In additionpeak accuracy measures (eg unpaired peak accuracy (UPA)temporally-paired peak accuracy (TPPA) spatially-paired peakaccuracy (SPPA) and pair peak accuracy (PPA)) index of agree-ment (IOA) and Brier score (BS) are often used in the discreteevaluation UPA quantifies the difference between the magnitudeof the peak observed and simulated values in the modelingdomain regardless whether this occurs at the monitoring stationand time of the peak observed value UPA tests the modelrsquos abilityto reproduce the highest observed value anywhere in the regionUPA is less useful as the two peak values may not occur at thesame location and time TPPA SPPA and PPA provide morestringent measures TPPA tests the modelrsquos ability to reproducethe highest observed value at the hour of the peak observed valueat any location within the modeling domain SPPA tests themodelrsquos ability to reproduce the highest observed value at themonitoring station of the peak observed value at any hour PPAtests the modelrsquos ability to reproduce the highest observed valueat the time and location of the peak observed value which is themost stringent measure among the four peak accuracy measuresIOA provides a measure of how well the departure of the simu-lated values from the observed mean matches with the departureof observations from the observed mean An IOA of 1 indicatesa perfect agreement BS measures the average squared deviationbetween predicted probabilities for a set of events and theiroutcomes so a lower score represents a higher accuracy It isa measure used frequently for probabilistic forecasts

The traditional categorical indices include accuracy (A) criticalsuccess index (CSI) or threat score (TS) probability of detection(POD) (or hit rate (HRate)) bias (B) false alarm ratio (FAR) falsealarm rate (FARate) Heidke skill score (HSS) Pierce skill score(PSS) (or true skill score (TSS) or Hansen and Kuipers discriminant(HKD)) skill score (SS) and economic value (EV) A indicates thepercentage of forecasts that correctly predicts an exceedance ora nonexceedance CSI indicates the percentage of the forecastedactual exceedances to all actual and forecasted exceedances PODor HRate indicates the percentage of the forecasted actualexceedances to all actual exceedances CSI is a more balancedscore because it takes into account both missed exceedances andfalse alarms whereas POD does not take into account false alarmsB judges if forecasts are underpredicted (lt1 also referred to asunderforecasting or ldquofalse negativerdquo) or overpredicted (gt1 alsoreferred to as overforecasting or ldquofalse positiverdquo) FAR measuresthe percentage of times an exceedance was forecast whenexceedance did not occur FARate indicates the proportion ofnonexceedances that were incorrectly forecasted (ie the proba-bility of false detection) it is the counterpart to the hit rate PODor HRate HSS measures the fractional improvement of the fore-cast (ie the total number of correct forecasts minus the correctrandom forecasts) over the standard random chance forecasts (iethe total number of forecasts minus the correct forecasts due tochance) Like most skill scores it is normalized by the total rangeof possible improvement over the standard which means HSS cansafely be compared on different datasets PSS (or TSS or HKD)measures skills relative to an unbiased random reference forecastby taking the differences between POD and FARate SS measuresthe relative accuracy with respect to a set of reference forecastsThe most convenient reference is the persistence forecast Anyvalid error metrics such as RMSE and NME can be used to calculate

ecasting part I History techniques and current status Atmospheric

Table 5Discrete statistical measures used in the model evaluation (compiled from Willmott 1981 Zhang et al 2006a Yu et al 2006 Dutot et al 2007)

Metrics Mathematical Expressiona Range

Correlation Coefficient r frac14(XN

ifrac141

ethMi MTHORNethOi OTHORN)(XN

ifrac141

ethMi MTHORN2XNifrac14 1

ethOi OTHORN2)1

2

1 to 1

Mean Bias (MB) MB frac14 1N

XNifrac141

ethMi OiTHORN frac14 M O N to thornN

Mean AbsoluteGross Error (MAGE)

MAGE frac14 1N

XNifrac141

jMi Oij 0 to thornN

Root MeanSquare Error (RMSE)

RMSE frac141N

XNifrac141

ethMi OiTHORN21

2

0 to thornN

Mean NormalizedBias (MNB)

MNB frac14 1N

XNifrac141

frac12ethMi OiTHORN=Oi frac141N

XNifrac141

ethMi=Oi 1THORN 1 to thornN

Mean NormalizedGross Error (MNGE)

MNGE frac14 1N

XNifrac141

frac12ethjMi OijTHORN=Oi 0 to thornN

NormalizedMean Bias (NMB)

NMB frac14XNifrac141

ethMi OiTHORNXN

ifrac14 1

Oi frac14 M

O 1 1 to thornN

NormalizedMean Error (NME)

NME frac14XNifrac141

jMi OijXN

ifrac141

Oi frac14 MAGE=O 0 to thornN

Fractional Bias (FB) FB frac14 1N

XNifrac141

Mi Oi

ethMi thorn OiTHORN=22 to thorn2

Fractional GrossError (FGE)

FGE frac14 1N

XNifrac141

jMi OijethMi thorn OiTHORN=2

0 to 2

Mean NormalizedFactor Bias (MNFB)

MNFB frac14 1N

XNifrac141

Fi where Fi frac14 Mi=Oi 10 for Mi Oi

Fi frac14 10 Oi=Mi forMi lt Oi

N to thornN

Mean NormalizedGross FactorError (MNGFE)

MNGFE frac14 1N

XNifrac141

jFij where Fi frac14 Mi=Oi 10 for Mi Oi

Fi frac14 10 Oi=Mi forMi lt Oi

0 to thornN

Normalized MeanBias Factor (NMBF)

NMBF frac14XNifrac141

Mi

XNifrac141

Oi 1 frac14 M=O 1 for M O

NMBF frac14 1XNifrac141

Oi

XNifrac141

Mi frac14 eth1 O=MTHORN for M lt O

N to thornN

Normalized Mean ErrorFactor (NMEF)

NMEF frac14XNifrac14 1

jMi OijXN

ifrac141

Oi frac14 MAGEO

for M O

NMEF frac14XNifrac14 1

jMi OijXN

ifrac141

Mi frac14 MAGEM

for M lt O

0 to thornN

Unpaired peakaccuracy (UPA)

UPA frac14Mmaxxt O

maxbxbtOmaxbxbt 1 to thornN

Temporally-paired peakaccuracy (PPA)

TPPA frac14M

maxxbt Omaxbxbt

Omaxbxbt 1 to thornN

Spatially-paired peakaccuracy (SPPA)

SPPA frac14M

maxbxt Omaxbxbt

Omaxbxbt 1 to thornN

Paired peak accuracy (PPA) PPA frac14M

maxbxbt Omaxbxbt

Omaxbxbt 1 to thornN

Index of Agreement (IOA) IOA frac14 1PN

ifrac14 1ethMi OiTHORN2PNifrac141ethjMi Oj thorn jOi OjTHORN2

0 to 1

Brier Score (BS) BS frac14 1K

XKtfrac141

ethpt otTHORN2 0 to thorn1

N is the number of samples (by time andor location)Mmaxbxbt and O

maxbxbt are time- and space-paired peak simulated and observed values with the circumflex indicating thatthe peak value occurs at the time t and location x of the peak observed value pt and ot are the probability that an event occurs according to an ensemble and the event actuallyoccurred or not at instance t (0 if it doesnrsquot happen and 1 if it happens) respectively K is the number of forecasting instances

a M frac14 eth1=NTHORNPNifrac141 MiO frac14 eth1=NTHORNPN

ifrac141 Oi Mi and Oi are values of model predictions and observations at time or location i respectively

Y Zhang et al Atmospheric Environment xxx (2012) 1e2416

SS For a perfect model A CSI POD HSS PSS and SS would be 1FAR and FARate would be zero and B would be unity

The above traditional categorical metrics provide ldquoclear cutrdquomeasures of the modelrsquos ability to predict an ldquoexceedancerdquo by usinga fixed threshold concentration and time- and space-paired

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

observationeforecast sets A main limitation is that they cannotprovide information regarding the spatial variations of pollutantconcentrations within a region of interest To overcome this limi-tation Kang et al (2007) developed three new categorical metricsfor RT-AQF evaluation the weighted success index (WSI) area hit

recasting part I History techniques and current status Atmospheric

Table 6Categorical statistical measures used in the model evaluation (compiled based on Ryan 1995 McHenry et al 2004 Kang et al 2005 2007 Pagowski and Grell 2006)

Metrics Mathematical Expressiona Range

Accuracy A frac14 bthorn cathorn bthorn cthorn d

0e1

CSI frac14 bathorn bthorn d

0e1

Critical Success Index or Threat Score (TS) WSI frac14 bthornPn1 IP

athorn bthorn d0e1

The Weighted Success Index (WSI)bIP frac14 Mi f Oi

Mi f T for Mi gt T gt Oi and Mi lt f T

IP frac14 Oi f Mi

Oi f T for Oi gt T gt Mi and Oi lt f Mi

IP frac14 0 for other conditions

Probability Of Detection Or Hit Rate POD frac14 H frac14 bbthorn d

0e1

Area Hit Ratec aH frac14 AbAbthorn Ad

0e1

Bias or Bias Ratio B frac14 athorn bbthorn d

0 to thornN

False Alarm Ratio FAR frac14 aathorn b

0e1

Area False Alarm Ratiod aFAR frac14 AaAathorn Ab

0e1

False Alarm rate FARate frac14 F frac14 aathorn c

0e1

Heidke skill score HSS frac14 2ethbc adTHORNethbthorn dTHORNethcthorn dTHORN thorn ethathorn bTHORNethathorn cTHORN N to 1

Pierce skill score (or true skill score (TSS)or Hansen and Kuipers discriminant)

PSS frac14 POD FARate frac14 bbthorn d

aathorn c

frac14 bc adethathorn cTHORN ethbthorn dTHORN 1 to 1

Skill scoreef SS frac14 Em ErefEperf Eref

0e1

Em frac14 RMSE frac141N

XNifrac141

ethMi OiTHORN21

2

or NME frac14XNifrac14 1

jMi OijXN

ifrac141

Oi frac14 MAGE=O

Eref frac14 RMSE frac141N

XNifrac14 1

ethMi KiTHORN21

2

or NME frac14XNifrac14 1

jMi MjjXN

jfrac141

Ki frac14 MAGEref=K

Eperf frac14 0

Economic value (EV) EV frac14 minethr cf THORN Feth1 cf THORNr thorn H cf eth1 rTHORN cfminethr cf THORN cf r

0e1

r frac14 CLfrac14 athorn b

d

H frac14 bbthorn d

F frac14 aathorn c

cf is the climatological frequency of an event

a Where a b c d are the number of simulated and observed data pairs at one site at a specific time in the four regions (see Table below) that represent forecast exceedancesthat did not occur forecast exceedances that did occur forecast nonexceedances that did occur and forecast nonexceedances that did not occur respectivelyContingencytable for threshold forecasts

ObservedYes No

Forecast Yes b aNo d c

b Mi and Oi are values of model prediction and observation at a location i respectively T is the threshold value used f is an empirical factor that reflects the zone in which themost points are located on the predictioneobservation scatterplots its typical value can be 1e2

c Ab is the number of exceedances that are both observed and forecast but the forecast is any exceedance that occurs in the designated area centered at themonitor locationAd is the number of observed exceedances that are not forecast within the designated area centered at the monitor location

d Aa is the number of forecast area exceedances that were not observed and Ab is the number of forecast exceedances that were observede Mi Oi and Ki are values of model prediction observation and temporal or spatial persistence value at time or location i respectively Temporal persistence value is the

previous days forecast value Spatial persistence value is the observation at the nearest location to the location i O frac14 eth1=NTHORNPNifrac141 Oi K frac14 eth1=NTHORNPN

ifrac141 Ki N is the number ofsamples (by time andor location) RMSE e Root mean square error NME e Normalized mean error MAGE e Mean Absolute Gross Error

f Em Eref and Eperf are the errors associated with model reference and perfect forecast respectively Em and Eref can be any valid error metrics eg RMSE and NME Thetemporal or spatial persistence forecast is often used as a reference forecast a perfect forecast has a zero error ie Eperf frac14 0

Y Zhang et al Atmospheric Environment xxx (2012) 1e24 17

Please cite this article in press as Zhang Y et al Real-time air quality forecasting part I History techniques and current status AtmosphericEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

Y Zhang et al Atmospheric Environment xxx (2012) 1e2418

(aH) and area false-alarm ratio (aFAR) WSI is less stringent thanCSI but a more representative measure It provides partial credit toobservationeforecast pairs that fall just outside limits establishedby CSI butwithin a designated factor that is used to judge themodelperformance (eg data pairs fall into a factor of 1e2 of theobservations)

An important practical aspect of RT-AQFs is how forecasts canbe used to mitigate the impact of air quality on economy andpeople In the calculation of EV one needs to know the climato-logical frequency (cf) of an event in addition to the values of a b cand d (see definitions in Table 6) Hits and false alarms incur a costof taking a preventive action (C) while misses are associated withlosses due to the lack of prevention (L) Correct rejections (c) incurno expense EV is defined as the reduction in mean expense rela-tive to the reduction (both defined with respect to cf) that wouldoccur if all the forecasts were correct (Pagowski and Grell 2006) Asfor weather forecasting EV provides a useful measure for theeconomic benefits of the existing RT-AQFs and the benefits ofimproving these systems to individuals businesses and society

No recommended values of performance statistics were providedfor acceptable forecasting skills by any individual countries Therecommended values for retrospective AQ modeling can providesome guidance for RT-AQF skill evaluation For example Zhang et al(2006a) recommended the use of MNBs and NMBs 15 MNGEsand NMEs 30 as indications of a satisfactory performance for O3

and PM25 Boylan and Russell (2006) recommended the use of meanfractional bias (MFB) and mean fractional error (MFE) for PM withgood performance indicted by MFBs 30 and MFEs 50 forPM25 and MFBs 60 and MFEs 75 for major PM componentsthat have concentrations 225 mg m3 The US EPA recommendedthat a so-calledweight of evidence approach inwhich a set of diverseanalyses are used to judge model performance Since no singlestatistical metrics can fully reflect RT-AQF skill discrete and cate-gorical statistics summarized in Tables 5 and 6 are recommended toprovide a comprehensive evaluation of an RT-AQF system

42 Observational datasets for model evaluation

Datasets for forecasting evaluation and improvement includereal-time or near real-time surface measurement datasets suchas the US EPArsquos AIRNOW datasets obtained by aircraft shipozonesondes and lidars during special field campaigns andsatellite data Satellite observations can complement datasetsfrom surface networks by providing derived chemical measure-ments Recent work has shown the potential of using near real-time satellite and surface data to improve RT-AQF (eg Al-Saadiet al 2005 Kondragunta et al 2008 Wang et al 2011) OverEurope the Global and Regional Earth-System MonitoringUsing Satellite and In situ Data (GEMS) and MACC projects areaimed at a near real-time data assimilation and forecast capabilityfor aerosols greenhouse gases and reactive gases The dataset canbe used to monitor the composition infer estimates of surfacefluxes and produce global short-range and medium-rangeforecasts

43 Current model forecasting skills

Most evaluations of forecasted O3 and its precursors focus onsummer and very few include winter (eg Manins et al 2002 Caiet al 2008 Doraiswamy et al 2009) Forecasting PM is moredifficult than forecasting O3 because PM consists of multiplechemical components over a broad size spectrum PM modeling isstarting to reach sufficient maturity for transition from research tooperational use Several PM models have recently been transferredinto operational models to forecast PM (Carmichael et al 2003

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

McHenry et al 2004 McKeen et al 2005 2007 Yu et al 2008Chuang et al 2011) very few studies provide detailed evaluationsof forecasted PM and its composition and precursors (eg McKeenet al 2007 2009 Yu et al 2008 Chen et al 2008 Manders et al2009) Similar to O3 forecasts most evaluations are conducted forsummer episodes and very few for winter episodes (eg Mathuret al 2008 Konovalov et al 2009) or a full year (eg Manderset al 2009) Limited evaluation was performed for coarse parti-cles such as mineral dust (eg Jimeacutenez-Guerrero et al 2008 Niuet al 2008 Menut et al 2009) that may be of major concernsfor PM10 attainment in some regions such as Asia southern Europeand the western US Jimeacutenez-Guerrero et al (2008) showed thatthe inclusion of a dust emission module substantially increases theaccuracy of both discrete and categorical statistics in the IberianPeninsula in Europewhere the influence of the Saharan dust cannotbe neglected

Table A4 summarizes the evaluation of a number of RT-AQFsystems in terms of domain and period and discrete and categor-ical performance statistics While maximum 1-h O3 averagestatistics are generally satisfactory those for maximum 8-h averageand hourly O3 sometime exceed NMB of 15 and NME of 30 forsome models (eg EtaCMAQ WRF-NMMCMAQ) over the easternor northeastern US in the O3 season during 2004e2006 indicatinga relatively poor performance The overpredictions in the low O3range (lt50 ppb) were reported in several studies (eg Yu et al2007 Chen et al 2008) Several factors may contribute to suchoverpredictions These include a poor representation of thenocturnal PBL mixing height (Gilliam et al 2006 Zhang et al2006b Eder et al 2006) an excessive downward transport ofhigh level O3 aloft and too much photolysis under high cloudconditions (Eder et al 2006) a high O3 production rate with theSAPRC-99 chemical mechanism (Arnold and Dennis 2006) themodel limitation in resolving titration of O3 by NO in urban plumes(Yu et al 2007) and the BCONs from global models (Chen et al2008) Chuang et al (2011) applied WRFChemeMADRID forRT-AQF and found that O3 overprediction in most regions in thesoutheastern US is likely caused by inaccurate emissions ofprecursors such as biogenic VOCs positive biases in 2-m temper-ature and negative biases in wind speed at 10-m O3 under-predictions in some regions could be due in part to theuncertainties in lateral BCONs

For categorical evaluation different threshold values were usedfor the same variables For example the threshold values used formaximum 1-h average O3 maximum 8-h average O3 and hourly O3are 60e125 ppb 65e85 ppb and 80 ppb respectively Some ofthese values are lower than the former US NAAQS of 120 ppb formaximum 1-h average O3 and the current US NAAQS of 75 ppb formaximum 8-h O3 respectively for the reasons stated previously(eg Hogrefe et al 2007 Yu et al 2007 Chuang et al 2011)although higher thresholds were used in some earlier applications(eg Kang et al 2005) For maximum 1-h average O3 A CSI POD Band FAR range from 15 to 998 52 to 216 69 to 89 01 to53 ppb and 03 to 941 respectively Prevrsquoair gives the lowest A of15e41 because it underestimates O3 daily maxima at high O3concentrations (Honoreacute et al 2008) All RT-AQFs give low values ofCSI and POD but higher values of FAR because they fail to forecastexceedance (ie a low value of b) but overpredict lowO3 (ie a highvalue of a) A CSI POD B and FAR for maximum 8-h average O3

range from 762 to 998 0 to 532 0 to 848 03 to 170 ppb and13 to 991 respectively Compared with maximum 1-h O3 thevalues of CSI and POD are higher and those of FAR are lower forcorresponding maximum 8-h O3 The categorical evaluation forhourly O3 was conducted by only one model (ie WRFChemeMADRID) with A CSI POD B and FAR of 992 25188 68 ppb and 972 respectively Very few evaluations were

recasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e24 19

conducted for precursors of O3 and other gaseous species (eg Caiet al 2008) and vertical profile of forecasted concentrations (egYu et al 2007) Yu et al (2007) reported that EtaCMAQ reproducedO3 vertical distributions on most of the days at low altitudes butoverpredictions occurred at altitudes gt 6 km because of a combi-nation of effects related to the specifications of meteorologicallateral BCONs as well as themodelrsquos coarse vertical resolution in theupper free troposphere

Surface PM forecasts are evaluated in terms of 24-h averagePM25 hourly PM25 and 24-h average PM10 These evaluationshowever have been very limited to date and a consistent evaluationprotocol has not yet been well established (Seigneur 2001 Zhanget al 2006a) The forecasting skill for PM25 is overall poorer thanthat for O3 For 24-h average PM25 MBs RMSEs NMBs and NMEsrange from 32 to 62 mg m3 17 to 159 mg m3 21 to 32 and37 to 81 respectively The NMB of 21 using EtaCMAQ over theeastern US during July 14eAugust 18 2004 was reported by Yuet al (2008) who attributed underpredictions to underestimatedtotal carbonaceous PM at both urban and rural sites and a signifi-cant underestimation of unspecified anthropogenic PM mass(mainly consisting of primary emitted trace elements) at rural sitesOn the other hand Hogrefe et al (2007) showed that EtaCMAQoverpredicted PM25 in the New York City due to overpredictions inorganic aerosols and crustal material Factors contributing to suchoverpredictions include underpredictions of nocturnal verticalmixing inaccurate temporal allocation of primary OM emissionsand underestimate of deposition processes Chen et al (2008)reported the NMB of 17e32 over the Pacific Northwest duringAugusteNovember 2004 The overpredictions in PM25 wereattributed to uncertainties in wildfire emission estimates and themodeling error in fire plume transport due to errors in MM5-predicted wind direction and wind speed McKeen et al (2007)evaluated RT-AQF of 6 models including WRFChem at two gridresolutions (12- and 36-km) CHRONOS AURAMS STEM-2K3 andEtaCMAQ They found that most models did not reproduce theobserved diurnal variation at urban and suburban sites particularlyduring the nighttime to early morning and the ensemble meanbased on 6 models with equal weighting gave the best possibleforecast in terms of statistics Chuang et al (2011) showed slightunderpredictions of PM25 in the O3 season over the southeasternUS by WRFChemeMADRID which were attributed to uncer-tainties in emissions such as those of biogenic VOCs and NH3overpredictions of precipitation and uncertainties in the BCONsOnly two studies evaluated hourly PM25 givingMBs RMSEs NMBsand NMEs of 33 to06 mg m3 83e113 mg m3 211 to52and 498e514 respectively For 24-h average PM10 MBs andRMSEs range from 29 to 138 mg m3 and 83 to 472 mg m3respectively Berge et al (2002) reported forecasted PM10 perfor-mance over Oslo with an MB of 209 mg m3 and an NMBof 365 at Kirkeveien due to an inaccurate emissions of PM10from the surface (ie the re-suspension of dust deposited at theroadside) on dry days and an MB of 138 mg m3 and an NMB of 24at Furuset due to errors in simulated grid-averaged wind fields

Very few evaluations were conducted for PM componentsMcKeen et al (2007) found that all the six RT-AQF models signifi-cantly underpredicted OM at the surface and overestimated SO4

2

above 2 km The overpredictions in SO42 were attributed to over-

estimate of SO2 by WRFChem and CHRONOS and the inclusion ofaqueous-phase oxidation of SO2 by AURAMS CMAQETA andSTEM-2K3 Yu et al (2007) reported that EtaCMAQ overpredictedSO4

2 due to too much in-cloud SO2 oxidation as a result of over-estimated H2O2 concentrations in the model underpredicted NH4

thorn

at the rural sites and aloft due to the exclusion of some sources ofNH3 in the real-time emission inventory underpredicted NO3

dueto overpredictions of SO4

2 and underpredicted OC due to missing

Please cite this article in press as Zhang Y et al Real-time air quality forEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

sources of primary OC in the emission inventory and missing SOAformation from the gas-phase oxidation of isoprene and sesqui-terpenes and aqueous-phase oxidation of glyoxal and methyl-glyoxal Chen et al (2008) reported that AIRPACT-3 reproduced OCwell but significantly overpredicted EC and significantly under-predicted SO4

2 NO3 and NH4

thorn due to underestimation of emis-sions of primary PM species (eg sulfate) and precursors ofsecondary PM (eg SO2 NH3 and NOx) insufficient spatial reso-lution and modelrsquos inability to capture the hourly PM variations

For categorical evaluation of PM some studies used thresholdvalues of 15e315 mg m3 for 24-h average PM25 and 30e50 mgm3

for 24-h average PM10 that are lower than the NAAQSs of 24-haverage PM25 of 35 mg m3 and 24-h average PM10 of 150 mg m3

for the same reason as mentioned previously For 24-h averagePM25 A CSI POD B and FAR range from 608 to 997 0 to 5370 to 909 07 to 19 mg m3 and 25 to 100 respectively Those for24-h average PM10 range from 167 to 100 91 to 100 150 to100 05 to 09 mg m3 and 0 to 48 respectively Categoricalevaluation of hourly PM25 was conducted only with WRFChemeMADRID with A CSI POD B and FAR of 721 205292 07 mg m3 and 591 respectively

5 Summary of Current States of RT-AQF

Emerging as a new discipline of applied sciences that integratesseveral physical sciences mathematicalstatistical tools andcomputer technology in the 1970s RT-AQF represents a uniqueapplied science affecting humanrsquos daily activities and posesunprecedented challenges intersecting many aspects of sciencesand technologies Driven by deteriorated air quality worldwideincreased societalhuman demands and rapid scientific and tech-nological advancements significant progress has been made in thepast four decades in RT-AQF A number of RT-AQF tools and modelswith varying degrees of sophistication and forecasting skillsranging from the simplest rule of thumb to the most advanced 3-Donline-coupled meteorologyechemistry models have been devel-oped Among them the deterministic physically-based RT-AQFmodels that are based on 3-D global and regional CTMs representthe state of the science enabling scientific understanding ofmechanisms for pollutant formation and development of emissioncontrol strategies Although other approaches in particular para-metric (statistical) models will continue to be used the 3-D RT-AQFmodels will likely become prevalent approaches and tools on all thescales There is a growing trend to combine these 3-D modelswith statistical models to provide more accurate RT-AQF Amongall RT-AQF models the coupled meteorologyechemistry modelrepresents a significant advancement and will greatly enhance theunderstanding of the underlying complex interplay of meteorologyemissions and chemistry in the real atmosphere Importantextensions of the RT-AQF models include their coupling with anurban model (eg traffic andor local pollutant dispersion) or a CFDmodel for urbanlocal scale applications at a spatial resolution of1 km or less and with an exposure model to provide real-timepublic health assessment and exposure predictions and urbanemergency preparedness

3-D RT-AQF models have been used for 24e72 h RT-AQFworldwide since the mid 1990s Despite their unique characteris-tics and technical requirements 3-D RT-AQF models are based onNWP models and CTM development and application efforts in thepast decades and therefore are subject to limitations of thesemodels Although no universal forecasting guidance and evaluationprotocols were developed covering all countries some guidanceand protocols were recommended by governmental agencies andthe scientific community Discrete and categorical evaluations aretwo most commonly-used evaluations Real-time or near real-time

ecasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e2420

surface air quality measurement and satellite datasets areincreasingly available for such evaluations The evaluation effortshave been primarily focused on O3 NO2 and PM (both PM10 andPM25) at surface level The forecasting skill is typically better for O3than for PM25 While their performance statistics averaged overa time period (eg monthly or daily) and the whole domain isgenerally good or satisfactory current RT-AQF models have diffi-culties forecasting variations at finer time scales (hourly and diur-nally) and individual monitoring stations In addition larger biasesexist for forecasted NOx HONO HNO3 and OH and PM compositionincluding SO4

2 NO3 and OM indicating possible error compen-

sations that lead to current levels of forecast accuracy These fore-cast inaccuracies have been attributed to a number of factors thatare highly uncertain or lack of accurate treatments in the RT-AQFmodels including mesoscale meteorological processes (eg sea-breeze circulations) and variables (eg temperature PBL height)chemical BCONs of O3 and PM25 emissions (eg SO2 NOx NH3BVOCs primary PM) physical and chemical processes (eg urbanprocesses in-cloud oxidation of SO2 SOA formation) and modelconfiguration (eg coarse grid resolution) Recent advances havebeenmade in some of those aspects Part II of this reviewwill assessand discuss these advances along with advanced computationaltechniques that can potentially improve RT-AQF Recommendationsfor research priorities and future prospects will also be provided

Acknowledgments

This work was performed under the Fellowship Award(704389J) Center for Education and ResearchEacutecole des PontsParisTech France through a visiting professorship at the Atmo-spheric Environment Center (CEREA) the Joint Laboratory Eacutecole desPonts ParisTech and EDF RampD Paris France the NSF Career Award(ATM-0348819) the EPA-Science to Achieve Results (STAR)program (R83337601) and Chinarsquos National Basic ResearchProgram (2010CB951803) Support from CEEH MEGAPOLI andCOST ES1004 is also acknowledged Thanks are due to Ming-TungChuang a former post-D researcher at NCSU for compiling anearlier version of Table A4 Thanks are due to Yelva RoustanAtmospheric Research Center (CEREA) France Christian Hogrefethe New York State Department of Environmental Conservation(NYSDEC) NY US and John McHenry Baron Advanced Meteoro-logical Systems (BAMS) US for providing some informationregarding their air quality forecasting models

Appendix A Supplementary Material

Supplementary data related to this article can be found online athttpdxdoiorg101016jatmosenv201206031

References

Alhanafy TE Zaghlool F El AS Moustafa D 2010 Neuro fuzzy modeling schemefor the prediction of air pollution J American Sci 6 (12) 605e616

Al-Saadi J Szykman J Pierce RB Kittaka C Neil D Chu DA Remer LGumley L Prins E Weinstock L MacDonald C Wayland R Dimmick FFishman J 2005 Improving national air quality forecasts with satellite aerosolobservations Bull Amer Meteor Soc 86 1249e1261

Arnold JR Dennis RL 2006 Testing CMAQ chemistry sensitivities in base caseand emissions control runs at SEARCH and SOS99 surface sites in the south-eastern US Atmos Environ 40 5027e5040

Aron RH 1980 Forecasting high level oxidant concentrations in the Los Angelesbasin J Air Pollut Control Assoc 20 1227e1228

Baklanov A Haumlnninen O Slordal LH Kukkonen J Bjergene N Fay B Finardi SHoe SC Jantunen M Karppinen A Rasmussen A Skouloudis A Sokhi RSSorensen j H Oslashdegaard V 2007 Integrated systems for forecasting urbanmeteorology air pollution and population exposure Atmos Chem Phys 7855e874 wwwatmos-chem-physnet78552007

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

Baklanov A Korsholm U Mahura A Petersen C Gross A 2008 Enviro-HIRLAMon-line coupled modelling of urban meteorology and air pollution Adv Sci Res2 41e46 wwwadv-sci-resnet2412008

Baklanov A Mahura A Sokhi R (Eds) 2011 Integrated Systems of Meso-Meteorological and Chemical Transport Models first ed Springer ISBN 978-3-642-13979-6 p 242

Baklanov A 2006 Overview of the European project FUMAPEX Atmos ChemPhys 6 2005e2015

Baklanov A 2010 Chemical weather forecasting a new concept of integratedmodeling Adv Sci Res 4 23e27

Baldasano JM Jimeacutenez-Guerrero P Jorba O Peacuterez C Loacutepez E Guumlereca PMartin F Garciacutea-Vivanco M Palomino I Querol X Pandolfi M Sanz MJDieacuteguez JJ 2008 CALIOPE an operational air quality forecasting system for theIberian Peninsula Balearic Islands and Canary Islands e First annual evaluationand ongoing developments Adv Sci Res 2 89e98 ISSN 1992-0628 May

Battan LJ 1966 The Unclean Sky Doubleday Garden City NY 141 ppBerge E Walker S-E Sorteberg A Lenkopane M Eastwood S Jablonska HI

Koltzow MOslash 2002 A real-time operational forecast model for meteorologyand air quality during peak air pollution episodes in Oslo Norway Water AirSoil Poll Focus 2 745e757

Blond N Vautard R 2004 Three-dimensional ozone analysis and their use forshort-term ozone forecasts J Geophys Res 109 D17303 httpdxdoiorg1010292004JD004515

Boylan JW Russell AG 2006 PM and light extinction model performance metricsgoals and criteria for three-dimensional air quality models Atmos Environ 40(26) 4946e4959 httpdxdoiorg101016jatmosenv200509087

Brandt J Christensen JH Frohn LM Palmgren F Berkowicz R Zlatev Z 2001Operational air pollution forecasts from European to local scale Atmos Environ35 (Suppl 1) S91eS98

Brimblecombe P 1987 The Big Smoke a History of Air Pollution in London sinceMedieval Time Routledge Revivals Methuen London UK 186 pp

Burrows WR Benjamin M Beauchamp S Lord ER McCollor D Thomson B1995 CART decision-tree statistical analysis and prediction of summer seasonmaximum surface ozone for the Vancouver Montreal and Atlantic Regions ofCanada J Appl Meteor Climatol 34 1848e1862

Byun D Schere KL 2006 Review of the governing equations computationalalgorithms and other components of the models-3 community multiscale airquality (CMAQ) modeling system Appl Mech Rev 59 51e77

Cai C Hogrefe C Katsafados P Kallos G Beauharnois M Schwab JJ Ren XBrune W Zhou X He Y Demerjian K 2008 Performance evaluation of an airquality forecast modeling system for a summer and winter season e Photo-chemical oxidants and their precursors Atmos Environ 42 8585e8599 httpdxdoiorg101026jatmosenv200808029

Carmichael GR et al 2003 Regional-scale chemical transport modeling in supportof intensive field experiments overview and analysis of the TRACE-P observa-tions J Geophys Res 108 (D21) 8823 httpdxdoiorg1010292002JD003117

Carmichael GR Sandu A Chai T Daescu DN Constantinescu EM Tang Y2008 Predicting air quality improvements through advanced methods tointegrate models and measurements J Comput Phys 227 3540e3571

Chai T Carmichael GR Tang Y Sandu A Streets DG 2007 Four-dimensionaldata assimilation experiments with international consortium for atmosphericresearch on transport and transformation ozone measurements J GeophysRes 112 D12S15 httpdxdoiorg1010292006JD007763

Chaloulakou A Saisana M Spyrellis N 2003 Comparative assessment of neuralnetworks and regression models for forecasting summertime ozone in AthensSci Total Environ 313 1e13

Chen J Vaughan J Avise J OrsquoNeill S Lamb B 2008 Enhancement and evalu-ation of the AIRPACT ozone and PM25 forecast system for the Pacific NorthwestJ Geophys Res 113 D14305 httpdxdoiorg1010292007JD009554

Chenevez J Jensen CO 2001 Operational ozone forecasts for the region of Copen-hagen by the Danish Meteorological Institute Atmos Environ 35 4567e4580

Chin M Ginoux P Lucchesi R Huebert B Weber R Anderson T Masonis SBlomquist B Bandy A Thornton D 2003 A global aerosol model forecast forthe ACE-Asia field experiment J Geophys Res 108 (D23) 8654 httpdxdoiorg1010292003JD003642

Chuang M-T Zhang Y Kang D-W 2011 Application of WRFChem-MADRID forreal-time air quality forecasting over the southeastern United States AtmosEnviron 45 (34) 6241e6250

Cobourn WG Hubbard MC 1999 An enhanced ozone forecasting model using airmass trajectory analysis Atmos Environ 33 4663e4674

Cobourn WG 2007 Accuracy and reliability of an automated air quality forecastsystem for ozone in seven Kentucky metropolitan areas Atmos Environ 415863e5875 httpdxdoiorg101016jatmosenv200703024

Comrie AC 1997 Comparing neural networks and regression models for ozoneforecasting J Air amp Waste Manag Assoc 47 653e663

Cooke WF Liousse C Cachier H Feichter J 1999 Construction of a fossil fuelemission data set for carbonaceous aerosol and implementation and radiativeimpact in the ECHAM4 model J Geophys Res 104 22137e22162

Cope ME Hess GD 2005 Air quality forecasting a review and comparison of theapproaches used internationally and in Australia Clean Air and Environ Quality39 (1) 52e60

Cope ME Hess GD Lee S Tory K Azzi M Carras J Lilley W Manins PCNelson P Ng L Puri K Wong N Walsh S Young M 2004 The Australianair quality forecasting system Part I project description and early outcomesJ Appl Meteorol 43 649e662

recasting part I History techniques and current status Atmospheric

Table 5Discrete statistical measures used in the model evaluation (compiled from Willmott 1981 Zhang et al 2006a Yu et al 2006 Dutot et al 2007)

Metrics Mathematical Expressiona Range

Correlation Coefficient r frac14(XN

ifrac141

ethMi MTHORNethOi OTHORN)(XN

ifrac141

ethMi MTHORN2XNifrac14 1

ethOi OTHORN2)1

2

1 to 1

Mean Bias (MB) MB frac14 1N

XNifrac141

ethMi OiTHORN frac14 M O N to thornN

Mean AbsoluteGross Error (MAGE)

MAGE frac14 1N

XNifrac141

jMi Oij 0 to thornN

Root MeanSquare Error (RMSE)

RMSE frac141N

XNifrac141

ethMi OiTHORN21

2

0 to thornN

Mean NormalizedBias (MNB)

MNB frac14 1N

XNifrac141

frac12ethMi OiTHORN=Oi frac141N

XNifrac141

ethMi=Oi 1THORN 1 to thornN

Mean NormalizedGross Error (MNGE)

MNGE frac14 1N

XNifrac141

frac12ethjMi OijTHORN=Oi 0 to thornN

NormalizedMean Bias (NMB)

NMB frac14XNifrac141

ethMi OiTHORNXN

ifrac14 1

Oi frac14 M

O 1 1 to thornN

NormalizedMean Error (NME)

NME frac14XNifrac141

jMi OijXN

ifrac141

Oi frac14 MAGE=O 0 to thornN

Fractional Bias (FB) FB frac14 1N

XNifrac141

Mi Oi

ethMi thorn OiTHORN=22 to thorn2

Fractional GrossError (FGE)

FGE frac14 1N

XNifrac141

jMi OijethMi thorn OiTHORN=2

0 to 2

Mean NormalizedFactor Bias (MNFB)

MNFB frac14 1N

XNifrac141

Fi where Fi frac14 Mi=Oi 10 for Mi Oi

Fi frac14 10 Oi=Mi forMi lt Oi

N to thornN

Mean NormalizedGross FactorError (MNGFE)

MNGFE frac14 1N

XNifrac141

jFij where Fi frac14 Mi=Oi 10 for Mi Oi

Fi frac14 10 Oi=Mi forMi lt Oi

0 to thornN

Normalized MeanBias Factor (NMBF)

NMBF frac14XNifrac141

Mi

XNifrac141

Oi 1 frac14 M=O 1 for M O

NMBF frac14 1XNifrac141

Oi

XNifrac141

Mi frac14 eth1 O=MTHORN for M lt O

N to thornN

Normalized Mean ErrorFactor (NMEF)

NMEF frac14XNifrac14 1

jMi OijXN

ifrac141

Oi frac14 MAGEO

for M O

NMEF frac14XNifrac14 1

jMi OijXN

ifrac141

Mi frac14 MAGEM

for M lt O

0 to thornN

Unpaired peakaccuracy (UPA)

UPA frac14Mmaxxt O

maxbxbtOmaxbxbt 1 to thornN

Temporally-paired peakaccuracy (PPA)

TPPA frac14M

maxxbt Omaxbxbt

Omaxbxbt 1 to thornN

Spatially-paired peakaccuracy (SPPA)

SPPA frac14M

maxbxt Omaxbxbt

Omaxbxbt 1 to thornN

Paired peak accuracy (PPA) PPA frac14M

maxbxbt Omaxbxbt

Omaxbxbt 1 to thornN

Index of Agreement (IOA) IOA frac14 1PN

ifrac14 1ethMi OiTHORN2PNifrac141ethjMi Oj thorn jOi OjTHORN2

0 to 1

Brier Score (BS) BS frac14 1K

XKtfrac141

ethpt otTHORN2 0 to thorn1

N is the number of samples (by time andor location)Mmaxbxbt and O

maxbxbt are time- and space-paired peak simulated and observed values with the circumflex indicating thatthe peak value occurs at the time t and location x of the peak observed value pt and ot are the probability that an event occurs according to an ensemble and the event actuallyoccurred or not at instance t (0 if it doesnrsquot happen and 1 if it happens) respectively K is the number of forecasting instances

a M frac14 eth1=NTHORNPNifrac141 MiO frac14 eth1=NTHORNPN

ifrac141 Oi Mi and Oi are values of model predictions and observations at time or location i respectively

Y Zhang et al Atmospheric Environment xxx (2012) 1e2416

SS For a perfect model A CSI POD HSS PSS and SS would be 1FAR and FARate would be zero and B would be unity

The above traditional categorical metrics provide ldquoclear cutrdquomeasures of the modelrsquos ability to predict an ldquoexceedancerdquo by usinga fixed threshold concentration and time- and space-paired

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

observationeforecast sets A main limitation is that they cannotprovide information regarding the spatial variations of pollutantconcentrations within a region of interest To overcome this limi-tation Kang et al (2007) developed three new categorical metricsfor RT-AQF evaluation the weighted success index (WSI) area hit

recasting part I History techniques and current status Atmospheric

Table 6Categorical statistical measures used in the model evaluation (compiled based on Ryan 1995 McHenry et al 2004 Kang et al 2005 2007 Pagowski and Grell 2006)

Metrics Mathematical Expressiona Range

Accuracy A frac14 bthorn cathorn bthorn cthorn d

0e1

CSI frac14 bathorn bthorn d

0e1

Critical Success Index or Threat Score (TS) WSI frac14 bthornPn1 IP

athorn bthorn d0e1

The Weighted Success Index (WSI)bIP frac14 Mi f Oi

Mi f T for Mi gt T gt Oi and Mi lt f T

IP frac14 Oi f Mi

Oi f T for Oi gt T gt Mi and Oi lt f Mi

IP frac14 0 for other conditions

Probability Of Detection Or Hit Rate POD frac14 H frac14 bbthorn d

0e1

Area Hit Ratec aH frac14 AbAbthorn Ad

0e1

Bias or Bias Ratio B frac14 athorn bbthorn d

0 to thornN

False Alarm Ratio FAR frac14 aathorn b

0e1

Area False Alarm Ratiod aFAR frac14 AaAathorn Ab

0e1

False Alarm rate FARate frac14 F frac14 aathorn c

0e1

Heidke skill score HSS frac14 2ethbc adTHORNethbthorn dTHORNethcthorn dTHORN thorn ethathorn bTHORNethathorn cTHORN N to 1

Pierce skill score (or true skill score (TSS)or Hansen and Kuipers discriminant)

PSS frac14 POD FARate frac14 bbthorn d

aathorn c

frac14 bc adethathorn cTHORN ethbthorn dTHORN 1 to 1

Skill scoreef SS frac14 Em ErefEperf Eref

0e1

Em frac14 RMSE frac141N

XNifrac141

ethMi OiTHORN21

2

or NME frac14XNifrac14 1

jMi OijXN

ifrac141

Oi frac14 MAGE=O

Eref frac14 RMSE frac141N

XNifrac14 1

ethMi KiTHORN21

2

or NME frac14XNifrac14 1

jMi MjjXN

jfrac141

Ki frac14 MAGEref=K

Eperf frac14 0

Economic value (EV) EV frac14 minethr cf THORN Feth1 cf THORNr thorn H cf eth1 rTHORN cfminethr cf THORN cf r

0e1

r frac14 CLfrac14 athorn b

d

H frac14 bbthorn d

F frac14 aathorn c

cf is the climatological frequency of an event

a Where a b c d are the number of simulated and observed data pairs at one site at a specific time in the four regions (see Table below) that represent forecast exceedancesthat did not occur forecast exceedances that did occur forecast nonexceedances that did occur and forecast nonexceedances that did not occur respectivelyContingencytable for threshold forecasts

ObservedYes No

Forecast Yes b aNo d c

b Mi and Oi are values of model prediction and observation at a location i respectively T is the threshold value used f is an empirical factor that reflects the zone in which themost points are located on the predictioneobservation scatterplots its typical value can be 1e2

c Ab is the number of exceedances that are both observed and forecast but the forecast is any exceedance that occurs in the designated area centered at themonitor locationAd is the number of observed exceedances that are not forecast within the designated area centered at the monitor location

d Aa is the number of forecast area exceedances that were not observed and Ab is the number of forecast exceedances that were observede Mi Oi and Ki are values of model prediction observation and temporal or spatial persistence value at time or location i respectively Temporal persistence value is the

previous days forecast value Spatial persistence value is the observation at the nearest location to the location i O frac14 eth1=NTHORNPNifrac141 Oi K frac14 eth1=NTHORNPN

ifrac141 Ki N is the number ofsamples (by time andor location) RMSE e Root mean square error NME e Normalized mean error MAGE e Mean Absolute Gross Error

f Em Eref and Eperf are the errors associated with model reference and perfect forecast respectively Em and Eref can be any valid error metrics eg RMSE and NME Thetemporal or spatial persistence forecast is often used as a reference forecast a perfect forecast has a zero error ie Eperf frac14 0

Y Zhang et al Atmospheric Environment xxx (2012) 1e24 17

Please cite this article in press as Zhang Y et al Real-time air quality forecasting part I History techniques and current status AtmosphericEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

Y Zhang et al Atmospheric Environment xxx (2012) 1e2418

(aH) and area false-alarm ratio (aFAR) WSI is less stringent thanCSI but a more representative measure It provides partial credit toobservationeforecast pairs that fall just outside limits establishedby CSI butwithin a designated factor that is used to judge themodelperformance (eg data pairs fall into a factor of 1e2 of theobservations)

An important practical aspect of RT-AQFs is how forecasts canbe used to mitigate the impact of air quality on economy andpeople In the calculation of EV one needs to know the climato-logical frequency (cf) of an event in addition to the values of a b cand d (see definitions in Table 6) Hits and false alarms incur a costof taking a preventive action (C) while misses are associated withlosses due to the lack of prevention (L) Correct rejections (c) incurno expense EV is defined as the reduction in mean expense rela-tive to the reduction (both defined with respect to cf) that wouldoccur if all the forecasts were correct (Pagowski and Grell 2006) Asfor weather forecasting EV provides a useful measure for theeconomic benefits of the existing RT-AQFs and the benefits ofimproving these systems to individuals businesses and society

No recommended values of performance statistics were providedfor acceptable forecasting skills by any individual countries Therecommended values for retrospective AQ modeling can providesome guidance for RT-AQF skill evaluation For example Zhang et al(2006a) recommended the use of MNBs and NMBs 15 MNGEsand NMEs 30 as indications of a satisfactory performance for O3

and PM25 Boylan and Russell (2006) recommended the use of meanfractional bias (MFB) and mean fractional error (MFE) for PM withgood performance indicted by MFBs 30 and MFEs 50 forPM25 and MFBs 60 and MFEs 75 for major PM componentsthat have concentrations 225 mg m3 The US EPA recommendedthat a so-calledweight of evidence approach inwhich a set of diverseanalyses are used to judge model performance Since no singlestatistical metrics can fully reflect RT-AQF skill discrete and cate-gorical statistics summarized in Tables 5 and 6 are recommended toprovide a comprehensive evaluation of an RT-AQF system

42 Observational datasets for model evaluation

Datasets for forecasting evaluation and improvement includereal-time or near real-time surface measurement datasets suchas the US EPArsquos AIRNOW datasets obtained by aircraft shipozonesondes and lidars during special field campaigns andsatellite data Satellite observations can complement datasetsfrom surface networks by providing derived chemical measure-ments Recent work has shown the potential of using near real-time satellite and surface data to improve RT-AQF (eg Al-Saadiet al 2005 Kondragunta et al 2008 Wang et al 2011) OverEurope the Global and Regional Earth-System MonitoringUsing Satellite and In situ Data (GEMS) and MACC projects areaimed at a near real-time data assimilation and forecast capabilityfor aerosols greenhouse gases and reactive gases The dataset canbe used to monitor the composition infer estimates of surfacefluxes and produce global short-range and medium-rangeforecasts

43 Current model forecasting skills

Most evaluations of forecasted O3 and its precursors focus onsummer and very few include winter (eg Manins et al 2002 Caiet al 2008 Doraiswamy et al 2009) Forecasting PM is moredifficult than forecasting O3 because PM consists of multiplechemical components over a broad size spectrum PM modeling isstarting to reach sufficient maturity for transition from research tooperational use Several PM models have recently been transferredinto operational models to forecast PM (Carmichael et al 2003

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

McHenry et al 2004 McKeen et al 2005 2007 Yu et al 2008Chuang et al 2011) very few studies provide detailed evaluationsof forecasted PM and its composition and precursors (eg McKeenet al 2007 2009 Yu et al 2008 Chen et al 2008 Manders et al2009) Similar to O3 forecasts most evaluations are conducted forsummer episodes and very few for winter episodes (eg Mathuret al 2008 Konovalov et al 2009) or a full year (eg Manderset al 2009) Limited evaluation was performed for coarse parti-cles such as mineral dust (eg Jimeacutenez-Guerrero et al 2008 Niuet al 2008 Menut et al 2009) that may be of major concernsfor PM10 attainment in some regions such as Asia southern Europeand the western US Jimeacutenez-Guerrero et al (2008) showed thatthe inclusion of a dust emission module substantially increases theaccuracy of both discrete and categorical statistics in the IberianPeninsula in Europewhere the influence of the Saharan dust cannotbe neglected

Table A4 summarizes the evaluation of a number of RT-AQFsystems in terms of domain and period and discrete and categor-ical performance statistics While maximum 1-h O3 averagestatistics are generally satisfactory those for maximum 8-h averageand hourly O3 sometime exceed NMB of 15 and NME of 30 forsome models (eg EtaCMAQ WRF-NMMCMAQ) over the easternor northeastern US in the O3 season during 2004e2006 indicatinga relatively poor performance The overpredictions in the low O3range (lt50 ppb) were reported in several studies (eg Yu et al2007 Chen et al 2008) Several factors may contribute to suchoverpredictions These include a poor representation of thenocturnal PBL mixing height (Gilliam et al 2006 Zhang et al2006b Eder et al 2006) an excessive downward transport ofhigh level O3 aloft and too much photolysis under high cloudconditions (Eder et al 2006) a high O3 production rate with theSAPRC-99 chemical mechanism (Arnold and Dennis 2006) themodel limitation in resolving titration of O3 by NO in urban plumes(Yu et al 2007) and the BCONs from global models (Chen et al2008) Chuang et al (2011) applied WRFChemeMADRID forRT-AQF and found that O3 overprediction in most regions in thesoutheastern US is likely caused by inaccurate emissions ofprecursors such as biogenic VOCs positive biases in 2-m temper-ature and negative biases in wind speed at 10-m O3 under-predictions in some regions could be due in part to theuncertainties in lateral BCONs

For categorical evaluation different threshold values were usedfor the same variables For example the threshold values used formaximum 1-h average O3 maximum 8-h average O3 and hourly O3are 60e125 ppb 65e85 ppb and 80 ppb respectively Some ofthese values are lower than the former US NAAQS of 120 ppb formaximum 1-h average O3 and the current US NAAQS of 75 ppb formaximum 8-h O3 respectively for the reasons stated previously(eg Hogrefe et al 2007 Yu et al 2007 Chuang et al 2011)although higher thresholds were used in some earlier applications(eg Kang et al 2005) For maximum 1-h average O3 A CSI POD Band FAR range from 15 to 998 52 to 216 69 to 89 01 to53 ppb and 03 to 941 respectively Prevrsquoair gives the lowest A of15e41 because it underestimates O3 daily maxima at high O3concentrations (Honoreacute et al 2008) All RT-AQFs give low values ofCSI and POD but higher values of FAR because they fail to forecastexceedance (ie a low value of b) but overpredict lowO3 (ie a highvalue of a) A CSI POD B and FAR for maximum 8-h average O3

range from 762 to 998 0 to 532 0 to 848 03 to 170 ppb and13 to 991 respectively Compared with maximum 1-h O3 thevalues of CSI and POD are higher and those of FAR are lower forcorresponding maximum 8-h O3 The categorical evaluation forhourly O3 was conducted by only one model (ie WRFChemeMADRID) with A CSI POD B and FAR of 992 25188 68 ppb and 972 respectively Very few evaluations were

recasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e24 19

conducted for precursors of O3 and other gaseous species (eg Caiet al 2008) and vertical profile of forecasted concentrations (egYu et al 2007) Yu et al (2007) reported that EtaCMAQ reproducedO3 vertical distributions on most of the days at low altitudes butoverpredictions occurred at altitudes gt 6 km because of a combi-nation of effects related to the specifications of meteorologicallateral BCONs as well as themodelrsquos coarse vertical resolution in theupper free troposphere

Surface PM forecasts are evaluated in terms of 24-h averagePM25 hourly PM25 and 24-h average PM10 These evaluationshowever have been very limited to date and a consistent evaluationprotocol has not yet been well established (Seigneur 2001 Zhanget al 2006a) The forecasting skill for PM25 is overall poorer thanthat for O3 For 24-h average PM25 MBs RMSEs NMBs and NMEsrange from 32 to 62 mg m3 17 to 159 mg m3 21 to 32 and37 to 81 respectively The NMB of 21 using EtaCMAQ over theeastern US during July 14eAugust 18 2004 was reported by Yuet al (2008) who attributed underpredictions to underestimatedtotal carbonaceous PM at both urban and rural sites and a signifi-cant underestimation of unspecified anthropogenic PM mass(mainly consisting of primary emitted trace elements) at rural sitesOn the other hand Hogrefe et al (2007) showed that EtaCMAQoverpredicted PM25 in the New York City due to overpredictions inorganic aerosols and crustal material Factors contributing to suchoverpredictions include underpredictions of nocturnal verticalmixing inaccurate temporal allocation of primary OM emissionsand underestimate of deposition processes Chen et al (2008)reported the NMB of 17e32 over the Pacific Northwest duringAugusteNovember 2004 The overpredictions in PM25 wereattributed to uncertainties in wildfire emission estimates and themodeling error in fire plume transport due to errors in MM5-predicted wind direction and wind speed McKeen et al (2007)evaluated RT-AQF of 6 models including WRFChem at two gridresolutions (12- and 36-km) CHRONOS AURAMS STEM-2K3 andEtaCMAQ They found that most models did not reproduce theobserved diurnal variation at urban and suburban sites particularlyduring the nighttime to early morning and the ensemble meanbased on 6 models with equal weighting gave the best possibleforecast in terms of statistics Chuang et al (2011) showed slightunderpredictions of PM25 in the O3 season over the southeasternUS by WRFChemeMADRID which were attributed to uncer-tainties in emissions such as those of biogenic VOCs and NH3overpredictions of precipitation and uncertainties in the BCONsOnly two studies evaluated hourly PM25 givingMBs RMSEs NMBsand NMEs of 33 to06 mg m3 83e113 mg m3 211 to52and 498e514 respectively For 24-h average PM10 MBs andRMSEs range from 29 to 138 mg m3 and 83 to 472 mg m3respectively Berge et al (2002) reported forecasted PM10 perfor-mance over Oslo with an MB of 209 mg m3 and an NMBof 365 at Kirkeveien due to an inaccurate emissions of PM10from the surface (ie the re-suspension of dust deposited at theroadside) on dry days and an MB of 138 mg m3 and an NMB of 24at Furuset due to errors in simulated grid-averaged wind fields

Very few evaluations were conducted for PM componentsMcKeen et al (2007) found that all the six RT-AQF models signifi-cantly underpredicted OM at the surface and overestimated SO4

2

above 2 km The overpredictions in SO42 were attributed to over-

estimate of SO2 by WRFChem and CHRONOS and the inclusion ofaqueous-phase oxidation of SO2 by AURAMS CMAQETA andSTEM-2K3 Yu et al (2007) reported that EtaCMAQ overpredictedSO4

2 due to too much in-cloud SO2 oxidation as a result of over-estimated H2O2 concentrations in the model underpredicted NH4

thorn

at the rural sites and aloft due to the exclusion of some sources ofNH3 in the real-time emission inventory underpredicted NO3

dueto overpredictions of SO4

2 and underpredicted OC due to missing

Please cite this article in press as Zhang Y et al Real-time air quality forEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

sources of primary OC in the emission inventory and missing SOAformation from the gas-phase oxidation of isoprene and sesqui-terpenes and aqueous-phase oxidation of glyoxal and methyl-glyoxal Chen et al (2008) reported that AIRPACT-3 reproduced OCwell but significantly overpredicted EC and significantly under-predicted SO4

2 NO3 and NH4

thorn due to underestimation of emis-sions of primary PM species (eg sulfate) and precursors ofsecondary PM (eg SO2 NH3 and NOx) insufficient spatial reso-lution and modelrsquos inability to capture the hourly PM variations

For categorical evaluation of PM some studies used thresholdvalues of 15e315 mg m3 for 24-h average PM25 and 30e50 mgm3

for 24-h average PM10 that are lower than the NAAQSs of 24-haverage PM25 of 35 mg m3 and 24-h average PM10 of 150 mg m3

for the same reason as mentioned previously For 24-h averagePM25 A CSI POD B and FAR range from 608 to 997 0 to 5370 to 909 07 to 19 mg m3 and 25 to 100 respectively Those for24-h average PM10 range from 167 to 100 91 to 100 150 to100 05 to 09 mg m3 and 0 to 48 respectively Categoricalevaluation of hourly PM25 was conducted only with WRFChemeMADRID with A CSI POD B and FAR of 721 205292 07 mg m3 and 591 respectively

5 Summary of Current States of RT-AQF

Emerging as a new discipline of applied sciences that integratesseveral physical sciences mathematicalstatistical tools andcomputer technology in the 1970s RT-AQF represents a uniqueapplied science affecting humanrsquos daily activities and posesunprecedented challenges intersecting many aspects of sciencesand technologies Driven by deteriorated air quality worldwideincreased societalhuman demands and rapid scientific and tech-nological advancements significant progress has been made in thepast four decades in RT-AQF A number of RT-AQF tools and modelswith varying degrees of sophistication and forecasting skillsranging from the simplest rule of thumb to the most advanced 3-Donline-coupled meteorologyechemistry models have been devel-oped Among them the deterministic physically-based RT-AQFmodels that are based on 3-D global and regional CTMs representthe state of the science enabling scientific understanding ofmechanisms for pollutant formation and development of emissioncontrol strategies Although other approaches in particular para-metric (statistical) models will continue to be used the 3-D RT-AQFmodels will likely become prevalent approaches and tools on all thescales There is a growing trend to combine these 3-D modelswith statistical models to provide more accurate RT-AQF Amongall RT-AQF models the coupled meteorologyechemistry modelrepresents a significant advancement and will greatly enhance theunderstanding of the underlying complex interplay of meteorologyemissions and chemistry in the real atmosphere Importantextensions of the RT-AQF models include their coupling with anurban model (eg traffic andor local pollutant dispersion) or a CFDmodel for urbanlocal scale applications at a spatial resolution of1 km or less and with an exposure model to provide real-timepublic health assessment and exposure predictions and urbanemergency preparedness

3-D RT-AQF models have been used for 24e72 h RT-AQFworldwide since the mid 1990s Despite their unique characteris-tics and technical requirements 3-D RT-AQF models are based onNWP models and CTM development and application efforts in thepast decades and therefore are subject to limitations of thesemodels Although no universal forecasting guidance and evaluationprotocols were developed covering all countries some guidanceand protocols were recommended by governmental agencies andthe scientific community Discrete and categorical evaluations aretwo most commonly-used evaluations Real-time or near real-time

ecasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e2420

surface air quality measurement and satellite datasets areincreasingly available for such evaluations The evaluation effortshave been primarily focused on O3 NO2 and PM (both PM10 andPM25) at surface level The forecasting skill is typically better for O3than for PM25 While their performance statistics averaged overa time period (eg monthly or daily) and the whole domain isgenerally good or satisfactory current RT-AQF models have diffi-culties forecasting variations at finer time scales (hourly and diur-nally) and individual monitoring stations In addition larger biasesexist for forecasted NOx HONO HNO3 and OH and PM compositionincluding SO4

2 NO3 and OM indicating possible error compen-

sations that lead to current levels of forecast accuracy These fore-cast inaccuracies have been attributed to a number of factors thatare highly uncertain or lack of accurate treatments in the RT-AQFmodels including mesoscale meteorological processes (eg sea-breeze circulations) and variables (eg temperature PBL height)chemical BCONs of O3 and PM25 emissions (eg SO2 NOx NH3BVOCs primary PM) physical and chemical processes (eg urbanprocesses in-cloud oxidation of SO2 SOA formation) and modelconfiguration (eg coarse grid resolution) Recent advances havebeenmade in some of those aspects Part II of this reviewwill assessand discuss these advances along with advanced computationaltechniques that can potentially improve RT-AQF Recommendationsfor research priorities and future prospects will also be provided

Acknowledgments

This work was performed under the Fellowship Award(704389J) Center for Education and ResearchEacutecole des PontsParisTech France through a visiting professorship at the Atmo-spheric Environment Center (CEREA) the Joint Laboratory Eacutecole desPonts ParisTech and EDF RampD Paris France the NSF Career Award(ATM-0348819) the EPA-Science to Achieve Results (STAR)program (R83337601) and Chinarsquos National Basic ResearchProgram (2010CB951803) Support from CEEH MEGAPOLI andCOST ES1004 is also acknowledged Thanks are due to Ming-TungChuang a former post-D researcher at NCSU for compiling anearlier version of Table A4 Thanks are due to Yelva RoustanAtmospheric Research Center (CEREA) France Christian Hogrefethe New York State Department of Environmental Conservation(NYSDEC) NY US and John McHenry Baron Advanced Meteoro-logical Systems (BAMS) US for providing some informationregarding their air quality forecasting models

Appendix A Supplementary Material

Supplementary data related to this article can be found online athttpdxdoiorg101016jatmosenv201206031

References

Alhanafy TE Zaghlool F El AS Moustafa D 2010 Neuro fuzzy modeling schemefor the prediction of air pollution J American Sci 6 (12) 605e616

Al-Saadi J Szykman J Pierce RB Kittaka C Neil D Chu DA Remer LGumley L Prins E Weinstock L MacDonald C Wayland R Dimmick FFishman J 2005 Improving national air quality forecasts with satellite aerosolobservations Bull Amer Meteor Soc 86 1249e1261

Arnold JR Dennis RL 2006 Testing CMAQ chemistry sensitivities in base caseand emissions control runs at SEARCH and SOS99 surface sites in the south-eastern US Atmos Environ 40 5027e5040

Aron RH 1980 Forecasting high level oxidant concentrations in the Los Angelesbasin J Air Pollut Control Assoc 20 1227e1228

Baklanov A Haumlnninen O Slordal LH Kukkonen J Bjergene N Fay B Finardi SHoe SC Jantunen M Karppinen A Rasmussen A Skouloudis A Sokhi RSSorensen j H Oslashdegaard V 2007 Integrated systems for forecasting urbanmeteorology air pollution and population exposure Atmos Chem Phys 7855e874 wwwatmos-chem-physnet78552007

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

Baklanov A Korsholm U Mahura A Petersen C Gross A 2008 Enviro-HIRLAMon-line coupled modelling of urban meteorology and air pollution Adv Sci Res2 41e46 wwwadv-sci-resnet2412008

Baklanov A Mahura A Sokhi R (Eds) 2011 Integrated Systems of Meso-Meteorological and Chemical Transport Models first ed Springer ISBN 978-3-642-13979-6 p 242

Baklanov A 2006 Overview of the European project FUMAPEX Atmos ChemPhys 6 2005e2015

Baklanov A 2010 Chemical weather forecasting a new concept of integratedmodeling Adv Sci Res 4 23e27

Baldasano JM Jimeacutenez-Guerrero P Jorba O Peacuterez C Loacutepez E Guumlereca PMartin F Garciacutea-Vivanco M Palomino I Querol X Pandolfi M Sanz MJDieacuteguez JJ 2008 CALIOPE an operational air quality forecasting system for theIberian Peninsula Balearic Islands and Canary Islands e First annual evaluationand ongoing developments Adv Sci Res 2 89e98 ISSN 1992-0628 May

Battan LJ 1966 The Unclean Sky Doubleday Garden City NY 141 ppBerge E Walker S-E Sorteberg A Lenkopane M Eastwood S Jablonska HI

Koltzow MOslash 2002 A real-time operational forecast model for meteorologyand air quality during peak air pollution episodes in Oslo Norway Water AirSoil Poll Focus 2 745e757

Blond N Vautard R 2004 Three-dimensional ozone analysis and their use forshort-term ozone forecasts J Geophys Res 109 D17303 httpdxdoiorg1010292004JD004515

Boylan JW Russell AG 2006 PM and light extinction model performance metricsgoals and criteria for three-dimensional air quality models Atmos Environ 40(26) 4946e4959 httpdxdoiorg101016jatmosenv200509087

Brandt J Christensen JH Frohn LM Palmgren F Berkowicz R Zlatev Z 2001Operational air pollution forecasts from European to local scale Atmos Environ35 (Suppl 1) S91eS98

Brimblecombe P 1987 The Big Smoke a History of Air Pollution in London sinceMedieval Time Routledge Revivals Methuen London UK 186 pp

Burrows WR Benjamin M Beauchamp S Lord ER McCollor D Thomson B1995 CART decision-tree statistical analysis and prediction of summer seasonmaximum surface ozone for the Vancouver Montreal and Atlantic Regions ofCanada J Appl Meteor Climatol 34 1848e1862

Byun D Schere KL 2006 Review of the governing equations computationalalgorithms and other components of the models-3 community multiscale airquality (CMAQ) modeling system Appl Mech Rev 59 51e77

Cai C Hogrefe C Katsafados P Kallos G Beauharnois M Schwab JJ Ren XBrune W Zhou X He Y Demerjian K 2008 Performance evaluation of an airquality forecast modeling system for a summer and winter season e Photo-chemical oxidants and their precursors Atmos Environ 42 8585e8599 httpdxdoiorg101026jatmosenv200808029

Carmichael GR et al 2003 Regional-scale chemical transport modeling in supportof intensive field experiments overview and analysis of the TRACE-P observa-tions J Geophys Res 108 (D21) 8823 httpdxdoiorg1010292002JD003117

Carmichael GR Sandu A Chai T Daescu DN Constantinescu EM Tang Y2008 Predicting air quality improvements through advanced methods tointegrate models and measurements J Comput Phys 227 3540e3571

Chai T Carmichael GR Tang Y Sandu A Streets DG 2007 Four-dimensionaldata assimilation experiments with international consortium for atmosphericresearch on transport and transformation ozone measurements J GeophysRes 112 D12S15 httpdxdoiorg1010292006JD007763

Chaloulakou A Saisana M Spyrellis N 2003 Comparative assessment of neuralnetworks and regression models for forecasting summertime ozone in AthensSci Total Environ 313 1e13

Chen J Vaughan J Avise J OrsquoNeill S Lamb B 2008 Enhancement and evalu-ation of the AIRPACT ozone and PM25 forecast system for the Pacific NorthwestJ Geophys Res 113 D14305 httpdxdoiorg1010292007JD009554

Chenevez J Jensen CO 2001 Operational ozone forecasts for the region of Copen-hagen by the Danish Meteorological Institute Atmos Environ 35 4567e4580

Chin M Ginoux P Lucchesi R Huebert B Weber R Anderson T Masonis SBlomquist B Bandy A Thornton D 2003 A global aerosol model forecast forthe ACE-Asia field experiment J Geophys Res 108 (D23) 8654 httpdxdoiorg1010292003JD003642

Chuang M-T Zhang Y Kang D-W 2011 Application of WRFChem-MADRID forreal-time air quality forecasting over the southeastern United States AtmosEnviron 45 (34) 6241e6250

Cobourn WG Hubbard MC 1999 An enhanced ozone forecasting model using airmass trajectory analysis Atmos Environ 33 4663e4674

Cobourn WG 2007 Accuracy and reliability of an automated air quality forecastsystem for ozone in seven Kentucky metropolitan areas Atmos Environ 415863e5875 httpdxdoiorg101016jatmosenv200703024

Comrie AC 1997 Comparing neural networks and regression models for ozoneforecasting J Air amp Waste Manag Assoc 47 653e663

Cooke WF Liousse C Cachier H Feichter J 1999 Construction of a fossil fuelemission data set for carbonaceous aerosol and implementation and radiativeimpact in the ECHAM4 model J Geophys Res 104 22137e22162

Cope ME Hess GD 2005 Air quality forecasting a review and comparison of theapproaches used internationally and in Australia Clean Air and Environ Quality39 (1) 52e60

Cope ME Hess GD Lee S Tory K Azzi M Carras J Lilley W Manins PCNelson P Ng L Puri K Wong N Walsh S Young M 2004 The Australianair quality forecasting system Part I project description and early outcomesJ Appl Meteorol 43 649e662

recasting part I History techniques and current status Atmospheric

Table 6Categorical statistical measures used in the model evaluation (compiled based on Ryan 1995 McHenry et al 2004 Kang et al 2005 2007 Pagowski and Grell 2006)

Metrics Mathematical Expressiona Range

Accuracy A frac14 bthorn cathorn bthorn cthorn d

0e1

CSI frac14 bathorn bthorn d

0e1

Critical Success Index or Threat Score (TS) WSI frac14 bthornPn1 IP

athorn bthorn d0e1

The Weighted Success Index (WSI)bIP frac14 Mi f Oi

Mi f T for Mi gt T gt Oi and Mi lt f T

IP frac14 Oi f Mi

Oi f T for Oi gt T gt Mi and Oi lt f Mi

IP frac14 0 for other conditions

Probability Of Detection Or Hit Rate POD frac14 H frac14 bbthorn d

0e1

Area Hit Ratec aH frac14 AbAbthorn Ad

0e1

Bias or Bias Ratio B frac14 athorn bbthorn d

0 to thornN

False Alarm Ratio FAR frac14 aathorn b

0e1

Area False Alarm Ratiod aFAR frac14 AaAathorn Ab

0e1

False Alarm rate FARate frac14 F frac14 aathorn c

0e1

Heidke skill score HSS frac14 2ethbc adTHORNethbthorn dTHORNethcthorn dTHORN thorn ethathorn bTHORNethathorn cTHORN N to 1

Pierce skill score (or true skill score (TSS)or Hansen and Kuipers discriminant)

PSS frac14 POD FARate frac14 bbthorn d

aathorn c

frac14 bc adethathorn cTHORN ethbthorn dTHORN 1 to 1

Skill scoreef SS frac14 Em ErefEperf Eref

0e1

Em frac14 RMSE frac141N

XNifrac141

ethMi OiTHORN21

2

or NME frac14XNifrac14 1

jMi OijXN

ifrac141

Oi frac14 MAGE=O

Eref frac14 RMSE frac141N

XNifrac14 1

ethMi KiTHORN21

2

or NME frac14XNifrac14 1

jMi MjjXN

jfrac141

Ki frac14 MAGEref=K

Eperf frac14 0

Economic value (EV) EV frac14 minethr cf THORN Feth1 cf THORNr thorn H cf eth1 rTHORN cfminethr cf THORN cf r

0e1

r frac14 CLfrac14 athorn b

d

H frac14 bbthorn d

F frac14 aathorn c

cf is the climatological frequency of an event

a Where a b c d are the number of simulated and observed data pairs at one site at a specific time in the four regions (see Table below) that represent forecast exceedancesthat did not occur forecast exceedances that did occur forecast nonexceedances that did occur and forecast nonexceedances that did not occur respectivelyContingencytable for threshold forecasts

ObservedYes No

Forecast Yes b aNo d c

b Mi and Oi are values of model prediction and observation at a location i respectively T is the threshold value used f is an empirical factor that reflects the zone in which themost points are located on the predictioneobservation scatterplots its typical value can be 1e2

c Ab is the number of exceedances that are both observed and forecast but the forecast is any exceedance that occurs in the designated area centered at themonitor locationAd is the number of observed exceedances that are not forecast within the designated area centered at the monitor location

d Aa is the number of forecast area exceedances that were not observed and Ab is the number of forecast exceedances that were observede Mi Oi and Ki are values of model prediction observation and temporal or spatial persistence value at time or location i respectively Temporal persistence value is the

previous days forecast value Spatial persistence value is the observation at the nearest location to the location i O frac14 eth1=NTHORNPNifrac141 Oi K frac14 eth1=NTHORNPN

ifrac141 Ki N is the number ofsamples (by time andor location) RMSE e Root mean square error NME e Normalized mean error MAGE e Mean Absolute Gross Error

f Em Eref and Eperf are the errors associated with model reference and perfect forecast respectively Em and Eref can be any valid error metrics eg RMSE and NME Thetemporal or spatial persistence forecast is often used as a reference forecast a perfect forecast has a zero error ie Eperf frac14 0

Y Zhang et al Atmospheric Environment xxx (2012) 1e24 17

Please cite this article in press as Zhang Y et al Real-time air quality forecasting part I History techniques and current status AtmosphericEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

Y Zhang et al Atmospheric Environment xxx (2012) 1e2418

(aH) and area false-alarm ratio (aFAR) WSI is less stringent thanCSI but a more representative measure It provides partial credit toobservationeforecast pairs that fall just outside limits establishedby CSI butwithin a designated factor that is used to judge themodelperformance (eg data pairs fall into a factor of 1e2 of theobservations)

An important practical aspect of RT-AQFs is how forecasts canbe used to mitigate the impact of air quality on economy andpeople In the calculation of EV one needs to know the climato-logical frequency (cf) of an event in addition to the values of a b cand d (see definitions in Table 6) Hits and false alarms incur a costof taking a preventive action (C) while misses are associated withlosses due to the lack of prevention (L) Correct rejections (c) incurno expense EV is defined as the reduction in mean expense rela-tive to the reduction (both defined with respect to cf) that wouldoccur if all the forecasts were correct (Pagowski and Grell 2006) Asfor weather forecasting EV provides a useful measure for theeconomic benefits of the existing RT-AQFs and the benefits ofimproving these systems to individuals businesses and society

No recommended values of performance statistics were providedfor acceptable forecasting skills by any individual countries Therecommended values for retrospective AQ modeling can providesome guidance for RT-AQF skill evaluation For example Zhang et al(2006a) recommended the use of MNBs and NMBs 15 MNGEsand NMEs 30 as indications of a satisfactory performance for O3

and PM25 Boylan and Russell (2006) recommended the use of meanfractional bias (MFB) and mean fractional error (MFE) for PM withgood performance indicted by MFBs 30 and MFEs 50 forPM25 and MFBs 60 and MFEs 75 for major PM componentsthat have concentrations 225 mg m3 The US EPA recommendedthat a so-calledweight of evidence approach inwhich a set of diverseanalyses are used to judge model performance Since no singlestatistical metrics can fully reflect RT-AQF skill discrete and cate-gorical statistics summarized in Tables 5 and 6 are recommended toprovide a comprehensive evaluation of an RT-AQF system

42 Observational datasets for model evaluation

Datasets for forecasting evaluation and improvement includereal-time or near real-time surface measurement datasets suchas the US EPArsquos AIRNOW datasets obtained by aircraft shipozonesondes and lidars during special field campaigns andsatellite data Satellite observations can complement datasetsfrom surface networks by providing derived chemical measure-ments Recent work has shown the potential of using near real-time satellite and surface data to improve RT-AQF (eg Al-Saadiet al 2005 Kondragunta et al 2008 Wang et al 2011) OverEurope the Global and Regional Earth-System MonitoringUsing Satellite and In situ Data (GEMS) and MACC projects areaimed at a near real-time data assimilation and forecast capabilityfor aerosols greenhouse gases and reactive gases The dataset canbe used to monitor the composition infer estimates of surfacefluxes and produce global short-range and medium-rangeforecasts

43 Current model forecasting skills

Most evaluations of forecasted O3 and its precursors focus onsummer and very few include winter (eg Manins et al 2002 Caiet al 2008 Doraiswamy et al 2009) Forecasting PM is moredifficult than forecasting O3 because PM consists of multiplechemical components over a broad size spectrum PM modeling isstarting to reach sufficient maturity for transition from research tooperational use Several PM models have recently been transferredinto operational models to forecast PM (Carmichael et al 2003

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

McHenry et al 2004 McKeen et al 2005 2007 Yu et al 2008Chuang et al 2011) very few studies provide detailed evaluationsof forecasted PM and its composition and precursors (eg McKeenet al 2007 2009 Yu et al 2008 Chen et al 2008 Manders et al2009) Similar to O3 forecasts most evaluations are conducted forsummer episodes and very few for winter episodes (eg Mathuret al 2008 Konovalov et al 2009) or a full year (eg Manderset al 2009) Limited evaluation was performed for coarse parti-cles such as mineral dust (eg Jimeacutenez-Guerrero et al 2008 Niuet al 2008 Menut et al 2009) that may be of major concernsfor PM10 attainment in some regions such as Asia southern Europeand the western US Jimeacutenez-Guerrero et al (2008) showed thatthe inclusion of a dust emission module substantially increases theaccuracy of both discrete and categorical statistics in the IberianPeninsula in Europewhere the influence of the Saharan dust cannotbe neglected

Table A4 summarizes the evaluation of a number of RT-AQFsystems in terms of domain and period and discrete and categor-ical performance statistics While maximum 1-h O3 averagestatistics are generally satisfactory those for maximum 8-h averageand hourly O3 sometime exceed NMB of 15 and NME of 30 forsome models (eg EtaCMAQ WRF-NMMCMAQ) over the easternor northeastern US in the O3 season during 2004e2006 indicatinga relatively poor performance The overpredictions in the low O3range (lt50 ppb) were reported in several studies (eg Yu et al2007 Chen et al 2008) Several factors may contribute to suchoverpredictions These include a poor representation of thenocturnal PBL mixing height (Gilliam et al 2006 Zhang et al2006b Eder et al 2006) an excessive downward transport ofhigh level O3 aloft and too much photolysis under high cloudconditions (Eder et al 2006) a high O3 production rate with theSAPRC-99 chemical mechanism (Arnold and Dennis 2006) themodel limitation in resolving titration of O3 by NO in urban plumes(Yu et al 2007) and the BCONs from global models (Chen et al2008) Chuang et al (2011) applied WRFChemeMADRID forRT-AQF and found that O3 overprediction in most regions in thesoutheastern US is likely caused by inaccurate emissions ofprecursors such as biogenic VOCs positive biases in 2-m temper-ature and negative biases in wind speed at 10-m O3 under-predictions in some regions could be due in part to theuncertainties in lateral BCONs

For categorical evaluation different threshold values were usedfor the same variables For example the threshold values used formaximum 1-h average O3 maximum 8-h average O3 and hourly O3are 60e125 ppb 65e85 ppb and 80 ppb respectively Some ofthese values are lower than the former US NAAQS of 120 ppb formaximum 1-h average O3 and the current US NAAQS of 75 ppb formaximum 8-h O3 respectively for the reasons stated previously(eg Hogrefe et al 2007 Yu et al 2007 Chuang et al 2011)although higher thresholds were used in some earlier applications(eg Kang et al 2005) For maximum 1-h average O3 A CSI POD Band FAR range from 15 to 998 52 to 216 69 to 89 01 to53 ppb and 03 to 941 respectively Prevrsquoair gives the lowest A of15e41 because it underestimates O3 daily maxima at high O3concentrations (Honoreacute et al 2008) All RT-AQFs give low values ofCSI and POD but higher values of FAR because they fail to forecastexceedance (ie a low value of b) but overpredict lowO3 (ie a highvalue of a) A CSI POD B and FAR for maximum 8-h average O3

range from 762 to 998 0 to 532 0 to 848 03 to 170 ppb and13 to 991 respectively Compared with maximum 1-h O3 thevalues of CSI and POD are higher and those of FAR are lower forcorresponding maximum 8-h O3 The categorical evaluation forhourly O3 was conducted by only one model (ie WRFChemeMADRID) with A CSI POD B and FAR of 992 25188 68 ppb and 972 respectively Very few evaluations were

recasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e24 19

conducted for precursors of O3 and other gaseous species (eg Caiet al 2008) and vertical profile of forecasted concentrations (egYu et al 2007) Yu et al (2007) reported that EtaCMAQ reproducedO3 vertical distributions on most of the days at low altitudes butoverpredictions occurred at altitudes gt 6 km because of a combi-nation of effects related to the specifications of meteorologicallateral BCONs as well as themodelrsquos coarse vertical resolution in theupper free troposphere

Surface PM forecasts are evaluated in terms of 24-h averagePM25 hourly PM25 and 24-h average PM10 These evaluationshowever have been very limited to date and a consistent evaluationprotocol has not yet been well established (Seigneur 2001 Zhanget al 2006a) The forecasting skill for PM25 is overall poorer thanthat for O3 For 24-h average PM25 MBs RMSEs NMBs and NMEsrange from 32 to 62 mg m3 17 to 159 mg m3 21 to 32 and37 to 81 respectively The NMB of 21 using EtaCMAQ over theeastern US during July 14eAugust 18 2004 was reported by Yuet al (2008) who attributed underpredictions to underestimatedtotal carbonaceous PM at both urban and rural sites and a signifi-cant underestimation of unspecified anthropogenic PM mass(mainly consisting of primary emitted trace elements) at rural sitesOn the other hand Hogrefe et al (2007) showed that EtaCMAQoverpredicted PM25 in the New York City due to overpredictions inorganic aerosols and crustal material Factors contributing to suchoverpredictions include underpredictions of nocturnal verticalmixing inaccurate temporal allocation of primary OM emissionsand underestimate of deposition processes Chen et al (2008)reported the NMB of 17e32 over the Pacific Northwest duringAugusteNovember 2004 The overpredictions in PM25 wereattributed to uncertainties in wildfire emission estimates and themodeling error in fire plume transport due to errors in MM5-predicted wind direction and wind speed McKeen et al (2007)evaluated RT-AQF of 6 models including WRFChem at two gridresolutions (12- and 36-km) CHRONOS AURAMS STEM-2K3 andEtaCMAQ They found that most models did not reproduce theobserved diurnal variation at urban and suburban sites particularlyduring the nighttime to early morning and the ensemble meanbased on 6 models with equal weighting gave the best possibleforecast in terms of statistics Chuang et al (2011) showed slightunderpredictions of PM25 in the O3 season over the southeasternUS by WRFChemeMADRID which were attributed to uncer-tainties in emissions such as those of biogenic VOCs and NH3overpredictions of precipitation and uncertainties in the BCONsOnly two studies evaluated hourly PM25 givingMBs RMSEs NMBsand NMEs of 33 to06 mg m3 83e113 mg m3 211 to52and 498e514 respectively For 24-h average PM10 MBs andRMSEs range from 29 to 138 mg m3 and 83 to 472 mg m3respectively Berge et al (2002) reported forecasted PM10 perfor-mance over Oslo with an MB of 209 mg m3 and an NMBof 365 at Kirkeveien due to an inaccurate emissions of PM10from the surface (ie the re-suspension of dust deposited at theroadside) on dry days and an MB of 138 mg m3 and an NMB of 24at Furuset due to errors in simulated grid-averaged wind fields

Very few evaluations were conducted for PM componentsMcKeen et al (2007) found that all the six RT-AQF models signifi-cantly underpredicted OM at the surface and overestimated SO4

2

above 2 km The overpredictions in SO42 were attributed to over-

estimate of SO2 by WRFChem and CHRONOS and the inclusion ofaqueous-phase oxidation of SO2 by AURAMS CMAQETA andSTEM-2K3 Yu et al (2007) reported that EtaCMAQ overpredictedSO4

2 due to too much in-cloud SO2 oxidation as a result of over-estimated H2O2 concentrations in the model underpredicted NH4

thorn

at the rural sites and aloft due to the exclusion of some sources ofNH3 in the real-time emission inventory underpredicted NO3

dueto overpredictions of SO4

2 and underpredicted OC due to missing

Please cite this article in press as Zhang Y et al Real-time air quality forEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

sources of primary OC in the emission inventory and missing SOAformation from the gas-phase oxidation of isoprene and sesqui-terpenes and aqueous-phase oxidation of glyoxal and methyl-glyoxal Chen et al (2008) reported that AIRPACT-3 reproduced OCwell but significantly overpredicted EC and significantly under-predicted SO4

2 NO3 and NH4

thorn due to underestimation of emis-sions of primary PM species (eg sulfate) and precursors ofsecondary PM (eg SO2 NH3 and NOx) insufficient spatial reso-lution and modelrsquos inability to capture the hourly PM variations

For categorical evaluation of PM some studies used thresholdvalues of 15e315 mg m3 for 24-h average PM25 and 30e50 mgm3

for 24-h average PM10 that are lower than the NAAQSs of 24-haverage PM25 of 35 mg m3 and 24-h average PM10 of 150 mg m3

for the same reason as mentioned previously For 24-h averagePM25 A CSI POD B and FAR range from 608 to 997 0 to 5370 to 909 07 to 19 mg m3 and 25 to 100 respectively Those for24-h average PM10 range from 167 to 100 91 to 100 150 to100 05 to 09 mg m3 and 0 to 48 respectively Categoricalevaluation of hourly PM25 was conducted only with WRFChemeMADRID with A CSI POD B and FAR of 721 205292 07 mg m3 and 591 respectively

5 Summary of Current States of RT-AQF

Emerging as a new discipline of applied sciences that integratesseveral physical sciences mathematicalstatistical tools andcomputer technology in the 1970s RT-AQF represents a uniqueapplied science affecting humanrsquos daily activities and posesunprecedented challenges intersecting many aspects of sciencesand technologies Driven by deteriorated air quality worldwideincreased societalhuman demands and rapid scientific and tech-nological advancements significant progress has been made in thepast four decades in RT-AQF A number of RT-AQF tools and modelswith varying degrees of sophistication and forecasting skillsranging from the simplest rule of thumb to the most advanced 3-Donline-coupled meteorologyechemistry models have been devel-oped Among them the deterministic physically-based RT-AQFmodels that are based on 3-D global and regional CTMs representthe state of the science enabling scientific understanding ofmechanisms for pollutant formation and development of emissioncontrol strategies Although other approaches in particular para-metric (statistical) models will continue to be used the 3-D RT-AQFmodels will likely become prevalent approaches and tools on all thescales There is a growing trend to combine these 3-D modelswith statistical models to provide more accurate RT-AQF Amongall RT-AQF models the coupled meteorologyechemistry modelrepresents a significant advancement and will greatly enhance theunderstanding of the underlying complex interplay of meteorologyemissions and chemistry in the real atmosphere Importantextensions of the RT-AQF models include their coupling with anurban model (eg traffic andor local pollutant dispersion) or a CFDmodel for urbanlocal scale applications at a spatial resolution of1 km or less and with an exposure model to provide real-timepublic health assessment and exposure predictions and urbanemergency preparedness

3-D RT-AQF models have been used for 24e72 h RT-AQFworldwide since the mid 1990s Despite their unique characteris-tics and technical requirements 3-D RT-AQF models are based onNWP models and CTM development and application efforts in thepast decades and therefore are subject to limitations of thesemodels Although no universal forecasting guidance and evaluationprotocols were developed covering all countries some guidanceand protocols were recommended by governmental agencies andthe scientific community Discrete and categorical evaluations aretwo most commonly-used evaluations Real-time or near real-time

ecasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e2420

surface air quality measurement and satellite datasets areincreasingly available for such evaluations The evaluation effortshave been primarily focused on O3 NO2 and PM (both PM10 andPM25) at surface level The forecasting skill is typically better for O3than for PM25 While their performance statistics averaged overa time period (eg monthly or daily) and the whole domain isgenerally good or satisfactory current RT-AQF models have diffi-culties forecasting variations at finer time scales (hourly and diur-nally) and individual monitoring stations In addition larger biasesexist for forecasted NOx HONO HNO3 and OH and PM compositionincluding SO4

2 NO3 and OM indicating possible error compen-

sations that lead to current levels of forecast accuracy These fore-cast inaccuracies have been attributed to a number of factors thatare highly uncertain or lack of accurate treatments in the RT-AQFmodels including mesoscale meteorological processes (eg sea-breeze circulations) and variables (eg temperature PBL height)chemical BCONs of O3 and PM25 emissions (eg SO2 NOx NH3BVOCs primary PM) physical and chemical processes (eg urbanprocesses in-cloud oxidation of SO2 SOA formation) and modelconfiguration (eg coarse grid resolution) Recent advances havebeenmade in some of those aspects Part II of this reviewwill assessand discuss these advances along with advanced computationaltechniques that can potentially improve RT-AQF Recommendationsfor research priorities and future prospects will also be provided

Acknowledgments

This work was performed under the Fellowship Award(704389J) Center for Education and ResearchEacutecole des PontsParisTech France through a visiting professorship at the Atmo-spheric Environment Center (CEREA) the Joint Laboratory Eacutecole desPonts ParisTech and EDF RampD Paris France the NSF Career Award(ATM-0348819) the EPA-Science to Achieve Results (STAR)program (R83337601) and Chinarsquos National Basic ResearchProgram (2010CB951803) Support from CEEH MEGAPOLI andCOST ES1004 is also acknowledged Thanks are due to Ming-TungChuang a former post-D researcher at NCSU for compiling anearlier version of Table A4 Thanks are due to Yelva RoustanAtmospheric Research Center (CEREA) France Christian Hogrefethe New York State Department of Environmental Conservation(NYSDEC) NY US and John McHenry Baron Advanced Meteoro-logical Systems (BAMS) US for providing some informationregarding their air quality forecasting models

Appendix A Supplementary Material

Supplementary data related to this article can be found online athttpdxdoiorg101016jatmosenv201206031

References

Alhanafy TE Zaghlool F El AS Moustafa D 2010 Neuro fuzzy modeling schemefor the prediction of air pollution J American Sci 6 (12) 605e616

Al-Saadi J Szykman J Pierce RB Kittaka C Neil D Chu DA Remer LGumley L Prins E Weinstock L MacDonald C Wayland R Dimmick FFishman J 2005 Improving national air quality forecasts with satellite aerosolobservations Bull Amer Meteor Soc 86 1249e1261

Arnold JR Dennis RL 2006 Testing CMAQ chemistry sensitivities in base caseand emissions control runs at SEARCH and SOS99 surface sites in the south-eastern US Atmos Environ 40 5027e5040

Aron RH 1980 Forecasting high level oxidant concentrations in the Los Angelesbasin J Air Pollut Control Assoc 20 1227e1228

Baklanov A Haumlnninen O Slordal LH Kukkonen J Bjergene N Fay B Finardi SHoe SC Jantunen M Karppinen A Rasmussen A Skouloudis A Sokhi RSSorensen j H Oslashdegaard V 2007 Integrated systems for forecasting urbanmeteorology air pollution and population exposure Atmos Chem Phys 7855e874 wwwatmos-chem-physnet78552007

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

Baklanov A Korsholm U Mahura A Petersen C Gross A 2008 Enviro-HIRLAMon-line coupled modelling of urban meteorology and air pollution Adv Sci Res2 41e46 wwwadv-sci-resnet2412008

Baklanov A Mahura A Sokhi R (Eds) 2011 Integrated Systems of Meso-Meteorological and Chemical Transport Models first ed Springer ISBN 978-3-642-13979-6 p 242

Baklanov A 2006 Overview of the European project FUMAPEX Atmos ChemPhys 6 2005e2015

Baklanov A 2010 Chemical weather forecasting a new concept of integratedmodeling Adv Sci Res 4 23e27

Baldasano JM Jimeacutenez-Guerrero P Jorba O Peacuterez C Loacutepez E Guumlereca PMartin F Garciacutea-Vivanco M Palomino I Querol X Pandolfi M Sanz MJDieacuteguez JJ 2008 CALIOPE an operational air quality forecasting system for theIberian Peninsula Balearic Islands and Canary Islands e First annual evaluationand ongoing developments Adv Sci Res 2 89e98 ISSN 1992-0628 May

Battan LJ 1966 The Unclean Sky Doubleday Garden City NY 141 ppBerge E Walker S-E Sorteberg A Lenkopane M Eastwood S Jablonska HI

Koltzow MOslash 2002 A real-time operational forecast model for meteorologyand air quality during peak air pollution episodes in Oslo Norway Water AirSoil Poll Focus 2 745e757

Blond N Vautard R 2004 Three-dimensional ozone analysis and their use forshort-term ozone forecasts J Geophys Res 109 D17303 httpdxdoiorg1010292004JD004515

Boylan JW Russell AG 2006 PM and light extinction model performance metricsgoals and criteria for three-dimensional air quality models Atmos Environ 40(26) 4946e4959 httpdxdoiorg101016jatmosenv200509087

Brandt J Christensen JH Frohn LM Palmgren F Berkowicz R Zlatev Z 2001Operational air pollution forecasts from European to local scale Atmos Environ35 (Suppl 1) S91eS98

Brimblecombe P 1987 The Big Smoke a History of Air Pollution in London sinceMedieval Time Routledge Revivals Methuen London UK 186 pp

Burrows WR Benjamin M Beauchamp S Lord ER McCollor D Thomson B1995 CART decision-tree statistical analysis and prediction of summer seasonmaximum surface ozone for the Vancouver Montreal and Atlantic Regions ofCanada J Appl Meteor Climatol 34 1848e1862

Byun D Schere KL 2006 Review of the governing equations computationalalgorithms and other components of the models-3 community multiscale airquality (CMAQ) modeling system Appl Mech Rev 59 51e77

Cai C Hogrefe C Katsafados P Kallos G Beauharnois M Schwab JJ Ren XBrune W Zhou X He Y Demerjian K 2008 Performance evaluation of an airquality forecast modeling system for a summer and winter season e Photo-chemical oxidants and their precursors Atmos Environ 42 8585e8599 httpdxdoiorg101026jatmosenv200808029

Carmichael GR et al 2003 Regional-scale chemical transport modeling in supportof intensive field experiments overview and analysis of the TRACE-P observa-tions J Geophys Res 108 (D21) 8823 httpdxdoiorg1010292002JD003117

Carmichael GR Sandu A Chai T Daescu DN Constantinescu EM Tang Y2008 Predicting air quality improvements through advanced methods tointegrate models and measurements J Comput Phys 227 3540e3571

Chai T Carmichael GR Tang Y Sandu A Streets DG 2007 Four-dimensionaldata assimilation experiments with international consortium for atmosphericresearch on transport and transformation ozone measurements J GeophysRes 112 D12S15 httpdxdoiorg1010292006JD007763

Chaloulakou A Saisana M Spyrellis N 2003 Comparative assessment of neuralnetworks and regression models for forecasting summertime ozone in AthensSci Total Environ 313 1e13

Chen J Vaughan J Avise J OrsquoNeill S Lamb B 2008 Enhancement and evalu-ation of the AIRPACT ozone and PM25 forecast system for the Pacific NorthwestJ Geophys Res 113 D14305 httpdxdoiorg1010292007JD009554

Chenevez J Jensen CO 2001 Operational ozone forecasts for the region of Copen-hagen by the Danish Meteorological Institute Atmos Environ 35 4567e4580

Chin M Ginoux P Lucchesi R Huebert B Weber R Anderson T Masonis SBlomquist B Bandy A Thornton D 2003 A global aerosol model forecast forthe ACE-Asia field experiment J Geophys Res 108 (D23) 8654 httpdxdoiorg1010292003JD003642

Chuang M-T Zhang Y Kang D-W 2011 Application of WRFChem-MADRID forreal-time air quality forecasting over the southeastern United States AtmosEnviron 45 (34) 6241e6250

Cobourn WG Hubbard MC 1999 An enhanced ozone forecasting model using airmass trajectory analysis Atmos Environ 33 4663e4674

Cobourn WG 2007 Accuracy and reliability of an automated air quality forecastsystem for ozone in seven Kentucky metropolitan areas Atmos Environ 415863e5875 httpdxdoiorg101016jatmosenv200703024

Comrie AC 1997 Comparing neural networks and regression models for ozoneforecasting J Air amp Waste Manag Assoc 47 653e663

Cooke WF Liousse C Cachier H Feichter J 1999 Construction of a fossil fuelemission data set for carbonaceous aerosol and implementation and radiativeimpact in the ECHAM4 model J Geophys Res 104 22137e22162

Cope ME Hess GD 2005 Air quality forecasting a review and comparison of theapproaches used internationally and in Australia Clean Air and Environ Quality39 (1) 52e60

Cope ME Hess GD Lee S Tory K Azzi M Carras J Lilley W Manins PCNelson P Ng L Puri K Wong N Walsh S Young M 2004 The Australianair quality forecasting system Part I project description and early outcomesJ Appl Meteorol 43 649e662

recasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e2418

(aH) and area false-alarm ratio (aFAR) WSI is less stringent thanCSI but a more representative measure It provides partial credit toobservationeforecast pairs that fall just outside limits establishedby CSI butwithin a designated factor that is used to judge themodelperformance (eg data pairs fall into a factor of 1e2 of theobservations)

An important practical aspect of RT-AQFs is how forecasts canbe used to mitigate the impact of air quality on economy andpeople In the calculation of EV one needs to know the climato-logical frequency (cf) of an event in addition to the values of a b cand d (see definitions in Table 6) Hits and false alarms incur a costof taking a preventive action (C) while misses are associated withlosses due to the lack of prevention (L) Correct rejections (c) incurno expense EV is defined as the reduction in mean expense rela-tive to the reduction (both defined with respect to cf) that wouldoccur if all the forecasts were correct (Pagowski and Grell 2006) Asfor weather forecasting EV provides a useful measure for theeconomic benefits of the existing RT-AQFs and the benefits ofimproving these systems to individuals businesses and society

No recommended values of performance statistics were providedfor acceptable forecasting skills by any individual countries Therecommended values for retrospective AQ modeling can providesome guidance for RT-AQF skill evaluation For example Zhang et al(2006a) recommended the use of MNBs and NMBs 15 MNGEsand NMEs 30 as indications of a satisfactory performance for O3

and PM25 Boylan and Russell (2006) recommended the use of meanfractional bias (MFB) and mean fractional error (MFE) for PM withgood performance indicted by MFBs 30 and MFEs 50 forPM25 and MFBs 60 and MFEs 75 for major PM componentsthat have concentrations 225 mg m3 The US EPA recommendedthat a so-calledweight of evidence approach inwhich a set of diverseanalyses are used to judge model performance Since no singlestatistical metrics can fully reflect RT-AQF skill discrete and cate-gorical statistics summarized in Tables 5 and 6 are recommended toprovide a comprehensive evaluation of an RT-AQF system

42 Observational datasets for model evaluation

Datasets for forecasting evaluation and improvement includereal-time or near real-time surface measurement datasets suchas the US EPArsquos AIRNOW datasets obtained by aircraft shipozonesondes and lidars during special field campaigns andsatellite data Satellite observations can complement datasetsfrom surface networks by providing derived chemical measure-ments Recent work has shown the potential of using near real-time satellite and surface data to improve RT-AQF (eg Al-Saadiet al 2005 Kondragunta et al 2008 Wang et al 2011) OverEurope the Global and Regional Earth-System MonitoringUsing Satellite and In situ Data (GEMS) and MACC projects areaimed at a near real-time data assimilation and forecast capabilityfor aerosols greenhouse gases and reactive gases The dataset canbe used to monitor the composition infer estimates of surfacefluxes and produce global short-range and medium-rangeforecasts

43 Current model forecasting skills

Most evaluations of forecasted O3 and its precursors focus onsummer and very few include winter (eg Manins et al 2002 Caiet al 2008 Doraiswamy et al 2009) Forecasting PM is moredifficult than forecasting O3 because PM consists of multiplechemical components over a broad size spectrum PM modeling isstarting to reach sufficient maturity for transition from research tooperational use Several PM models have recently been transferredinto operational models to forecast PM (Carmichael et al 2003

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

McHenry et al 2004 McKeen et al 2005 2007 Yu et al 2008Chuang et al 2011) very few studies provide detailed evaluationsof forecasted PM and its composition and precursors (eg McKeenet al 2007 2009 Yu et al 2008 Chen et al 2008 Manders et al2009) Similar to O3 forecasts most evaluations are conducted forsummer episodes and very few for winter episodes (eg Mathuret al 2008 Konovalov et al 2009) or a full year (eg Manderset al 2009) Limited evaluation was performed for coarse parti-cles such as mineral dust (eg Jimeacutenez-Guerrero et al 2008 Niuet al 2008 Menut et al 2009) that may be of major concernsfor PM10 attainment in some regions such as Asia southern Europeand the western US Jimeacutenez-Guerrero et al (2008) showed thatthe inclusion of a dust emission module substantially increases theaccuracy of both discrete and categorical statistics in the IberianPeninsula in Europewhere the influence of the Saharan dust cannotbe neglected

Table A4 summarizes the evaluation of a number of RT-AQFsystems in terms of domain and period and discrete and categor-ical performance statistics While maximum 1-h O3 averagestatistics are generally satisfactory those for maximum 8-h averageand hourly O3 sometime exceed NMB of 15 and NME of 30 forsome models (eg EtaCMAQ WRF-NMMCMAQ) over the easternor northeastern US in the O3 season during 2004e2006 indicatinga relatively poor performance The overpredictions in the low O3range (lt50 ppb) were reported in several studies (eg Yu et al2007 Chen et al 2008) Several factors may contribute to suchoverpredictions These include a poor representation of thenocturnal PBL mixing height (Gilliam et al 2006 Zhang et al2006b Eder et al 2006) an excessive downward transport ofhigh level O3 aloft and too much photolysis under high cloudconditions (Eder et al 2006) a high O3 production rate with theSAPRC-99 chemical mechanism (Arnold and Dennis 2006) themodel limitation in resolving titration of O3 by NO in urban plumes(Yu et al 2007) and the BCONs from global models (Chen et al2008) Chuang et al (2011) applied WRFChemeMADRID forRT-AQF and found that O3 overprediction in most regions in thesoutheastern US is likely caused by inaccurate emissions ofprecursors such as biogenic VOCs positive biases in 2-m temper-ature and negative biases in wind speed at 10-m O3 under-predictions in some regions could be due in part to theuncertainties in lateral BCONs

For categorical evaluation different threshold values were usedfor the same variables For example the threshold values used formaximum 1-h average O3 maximum 8-h average O3 and hourly O3are 60e125 ppb 65e85 ppb and 80 ppb respectively Some ofthese values are lower than the former US NAAQS of 120 ppb formaximum 1-h average O3 and the current US NAAQS of 75 ppb formaximum 8-h O3 respectively for the reasons stated previously(eg Hogrefe et al 2007 Yu et al 2007 Chuang et al 2011)although higher thresholds were used in some earlier applications(eg Kang et al 2005) For maximum 1-h average O3 A CSI POD Band FAR range from 15 to 998 52 to 216 69 to 89 01 to53 ppb and 03 to 941 respectively Prevrsquoair gives the lowest A of15e41 because it underestimates O3 daily maxima at high O3concentrations (Honoreacute et al 2008) All RT-AQFs give low values ofCSI and POD but higher values of FAR because they fail to forecastexceedance (ie a low value of b) but overpredict lowO3 (ie a highvalue of a) A CSI POD B and FAR for maximum 8-h average O3

range from 762 to 998 0 to 532 0 to 848 03 to 170 ppb and13 to 991 respectively Compared with maximum 1-h O3 thevalues of CSI and POD are higher and those of FAR are lower forcorresponding maximum 8-h O3 The categorical evaluation forhourly O3 was conducted by only one model (ie WRFChemeMADRID) with A CSI POD B and FAR of 992 25188 68 ppb and 972 respectively Very few evaluations were

recasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e24 19

conducted for precursors of O3 and other gaseous species (eg Caiet al 2008) and vertical profile of forecasted concentrations (egYu et al 2007) Yu et al (2007) reported that EtaCMAQ reproducedO3 vertical distributions on most of the days at low altitudes butoverpredictions occurred at altitudes gt 6 km because of a combi-nation of effects related to the specifications of meteorologicallateral BCONs as well as themodelrsquos coarse vertical resolution in theupper free troposphere

Surface PM forecasts are evaluated in terms of 24-h averagePM25 hourly PM25 and 24-h average PM10 These evaluationshowever have been very limited to date and a consistent evaluationprotocol has not yet been well established (Seigneur 2001 Zhanget al 2006a) The forecasting skill for PM25 is overall poorer thanthat for O3 For 24-h average PM25 MBs RMSEs NMBs and NMEsrange from 32 to 62 mg m3 17 to 159 mg m3 21 to 32 and37 to 81 respectively The NMB of 21 using EtaCMAQ over theeastern US during July 14eAugust 18 2004 was reported by Yuet al (2008) who attributed underpredictions to underestimatedtotal carbonaceous PM at both urban and rural sites and a signifi-cant underestimation of unspecified anthropogenic PM mass(mainly consisting of primary emitted trace elements) at rural sitesOn the other hand Hogrefe et al (2007) showed that EtaCMAQoverpredicted PM25 in the New York City due to overpredictions inorganic aerosols and crustal material Factors contributing to suchoverpredictions include underpredictions of nocturnal verticalmixing inaccurate temporal allocation of primary OM emissionsand underestimate of deposition processes Chen et al (2008)reported the NMB of 17e32 over the Pacific Northwest duringAugusteNovember 2004 The overpredictions in PM25 wereattributed to uncertainties in wildfire emission estimates and themodeling error in fire plume transport due to errors in MM5-predicted wind direction and wind speed McKeen et al (2007)evaluated RT-AQF of 6 models including WRFChem at two gridresolutions (12- and 36-km) CHRONOS AURAMS STEM-2K3 andEtaCMAQ They found that most models did not reproduce theobserved diurnal variation at urban and suburban sites particularlyduring the nighttime to early morning and the ensemble meanbased on 6 models with equal weighting gave the best possibleforecast in terms of statistics Chuang et al (2011) showed slightunderpredictions of PM25 in the O3 season over the southeasternUS by WRFChemeMADRID which were attributed to uncer-tainties in emissions such as those of biogenic VOCs and NH3overpredictions of precipitation and uncertainties in the BCONsOnly two studies evaluated hourly PM25 givingMBs RMSEs NMBsand NMEs of 33 to06 mg m3 83e113 mg m3 211 to52and 498e514 respectively For 24-h average PM10 MBs andRMSEs range from 29 to 138 mg m3 and 83 to 472 mg m3respectively Berge et al (2002) reported forecasted PM10 perfor-mance over Oslo with an MB of 209 mg m3 and an NMBof 365 at Kirkeveien due to an inaccurate emissions of PM10from the surface (ie the re-suspension of dust deposited at theroadside) on dry days and an MB of 138 mg m3 and an NMB of 24at Furuset due to errors in simulated grid-averaged wind fields

Very few evaluations were conducted for PM componentsMcKeen et al (2007) found that all the six RT-AQF models signifi-cantly underpredicted OM at the surface and overestimated SO4

2

above 2 km The overpredictions in SO42 were attributed to over-

estimate of SO2 by WRFChem and CHRONOS and the inclusion ofaqueous-phase oxidation of SO2 by AURAMS CMAQETA andSTEM-2K3 Yu et al (2007) reported that EtaCMAQ overpredictedSO4

2 due to too much in-cloud SO2 oxidation as a result of over-estimated H2O2 concentrations in the model underpredicted NH4

thorn

at the rural sites and aloft due to the exclusion of some sources ofNH3 in the real-time emission inventory underpredicted NO3

dueto overpredictions of SO4

2 and underpredicted OC due to missing

Please cite this article in press as Zhang Y et al Real-time air quality forEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

sources of primary OC in the emission inventory and missing SOAformation from the gas-phase oxidation of isoprene and sesqui-terpenes and aqueous-phase oxidation of glyoxal and methyl-glyoxal Chen et al (2008) reported that AIRPACT-3 reproduced OCwell but significantly overpredicted EC and significantly under-predicted SO4

2 NO3 and NH4

thorn due to underestimation of emis-sions of primary PM species (eg sulfate) and precursors ofsecondary PM (eg SO2 NH3 and NOx) insufficient spatial reso-lution and modelrsquos inability to capture the hourly PM variations

For categorical evaluation of PM some studies used thresholdvalues of 15e315 mg m3 for 24-h average PM25 and 30e50 mgm3

for 24-h average PM10 that are lower than the NAAQSs of 24-haverage PM25 of 35 mg m3 and 24-h average PM10 of 150 mg m3

for the same reason as mentioned previously For 24-h averagePM25 A CSI POD B and FAR range from 608 to 997 0 to 5370 to 909 07 to 19 mg m3 and 25 to 100 respectively Those for24-h average PM10 range from 167 to 100 91 to 100 150 to100 05 to 09 mg m3 and 0 to 48 respectively Categoricalevaluation of hourly PM25 was conducted only with WRFChemeMADRID with A CSI POD B and FAR of 721 205292 07 mg m3 and 591 respectively

5 Summary of Current States of RT-AQF

Emerging as a new discipline of applied sciences that integratesseveral physical sciences mathematicalstatistical tools andcomputer technology in the 1970s RT-AQF represents a uniqueapplied science affecting humanrsquos daily activities and posesunprecedented challenges intersecting many aspects of sciencesand technologies Driven by deteriorated air quality worldwideincreased societalhuman demands and rapid scientific and tech-nological advancements significant progress has been made in thepast four decades in RT-AQF A number of RT-AQF tools and modelswith varying degrees of sophistication and forecasting skillsranging from the simplest rule of thumb to the most advanced 3-Donline-coupled meteorologyechemistry models have been devel-oped Among them the deterministic physically-based RT-AQFmodels that are based on 3-D global and regional CTMs representthe state of the science enabling scientific understanding ofmechanisms for pollutant formation and development of emissioncontrol strategies Although other approaches in particular para-metric (statistical) models will continue to be used the 3-D RT-AQFmodels will likely become prevalent approaches and tools on all thescales There is a growing trend to combine these 3-D modelswith statistical models to provide more accurate RT-AQF Amongall RT-AQF models the coupled meteorologyechemistry modelrepresents a significant advancement and will greatly enhance theunderstanding of the underlying complex interplay of meteorologyemissions and chemistry in the real atmosphere Importantextensions of the RT-AQF models include their coupling with anurban model (eg traffic andor local pollutant dispersion) or a CFDmodel for urbanlocal scale applications at a spatial resolution of1 km or less and with an exposure model to provide real-timepublic health assessment and exposure predictions and urbanemergency preparedness

3-D RT-AQF models have been used for 24e72 h RT-AQFworldwide since the mid 1990s Despite their unique characteris-tics and technical requirements 3-D RT-AQF models are based onNWP models and CTM development and application efforts in thepast decades and therefore are subject to limitations of thesemodels Although no universal forecasting guidance and evaluationprotocols were developed covering all countries some guidanceand protocols were recommended by governmental agencies andthe scientific community Discrete and categorical evaluations aretwo most commonly-used evaluations Real-time or near real-time

ecasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e2420

surface air quality measurement and satellite datasets areincreasingly available for such evaluations The evaluation effortshave been primarily focused on O3 NO2 and PM (both PM10 andPM25) at surface level The forecasting skill is typically better for O3than for PM25 While their performance statistics averaged overa time period (eg monthly or daily) and the whole domain isgenerally good or satisfactory current RT-AQF models have diffi-culties forecasting variations at finer time scales (hourly and diur-nally) and individual monitoring stations In addition larger biasesexist for forecasted NOx HONO HNO3 and OH and PM compositionincluding SO4

2 NO3 and OM indicating possible error compen-

sations that lead to current levels of forecast accuracy These fore-cast inaccuracies have been attributed to a number of factors thatare highly uncertain or lack of accurate treatments in the RT-AQFmodels including mesoscale meteorological processes (eg sea-breeze circulations) and variables (eg temperature PBL height)chemical BCONs of O3 and PM25 emissions (eg SO2 NOx NH3BVOCs primary PM) physical and chemical processes (eg urbanprocesses in-cloud oxidation of SO2 SOA formation) and modelconfiguration (eg coarse grid resolution) Recent advances havebeenmade in some of those aspects Part II of this reviewwill assessand discuss these advances along with advanced computationaltechniques that can potentially improve RT-AQF Recommendationsfor research priorities and future prospects will also be provided

Acknowledgments

This work was performed under the Fellowship Award(704389J) Center for Education and ResearchEacutecole des PontsParisTech France through a visiting professorship at the Atmo-spheric Environment Center (CEREA) the Joint Laboratory Eacutecole desPonts ParisTech and EDF RampD Paris France the NSF Career Award(ATM-0348819) the EPA-Science to Achieve Results (STAR)program (R83337601) and Chinarsquos National Basic ResearchProgram (2010CB951803) Support from CEEH MEGAPOLI andCOST ES1004 is also acknowledged Thanks are due to Ming-TungChuang a former post-D researcher at NCSU for compiling anearlier version of Table A4 Thanks are due to Yelva RoustanAtmospheric Research Center (CEREA) France Christian Hogrefethe New York State Department of Environmental Conservation(NYSDEC) NY US and John McHenry Baron Advanced Meteoro-logical Systems (BAMS) US for providing some informationregarding their air quality forecasting models

Appendix A Supplementary Material

Supplementary data related to this article can be found online athttpdxdoiorg101016jatmosenv201206031

References

Alhanafy TE Zaghlool F El AS Moustafa D 2010 Neuro fuzzy modeling schemefor the prediction of air pollution J American Sci 6 (12) 605e616

Al-Saadi J Szykman J Pierce RB Kittaka C Neil D Chu DA Remer LGumley L Prins E Weinstock L MacDonald C Wayland R Dimmick FFishman J 2005 Improving national air quality forecasts with satellite aerosolobservations Bull Amer Meteor Soc 86 1249e1261

Arnold JR Dennis RL 2006 Testing CMAQ chemistry sensitivities in base caseand emissions control runs at SEARCH and SOS99 surface sites in the south-eastern US Atmos Environ 40 5027e5040

Aron RH 1980 Forecasting high level oxidant concentrations in the Los Angelesbasin J Air Pollut Control Assoc 20 1227e1228

Baklanov A Haumlnninen O Slordal LH Kukkonen J Bjergene N Fay B Finardi SHoe SC Jantunen M Karppinen A Rasmussen A Skouloudis A Sokhi RSSorensen j H Oslashdegaard V 2007 Integrated systems for forecasting urbanmeteorology air pollution and population exposure Atmos Chem Phys 7855e874 wwwatmos-chem-physnet78552007

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

Baklanov A Korsholm U Mahura A Petersen C Gross A 2008 Enviro-HIRLAMon-line coupled modelling of urban meteorology and air pollution Adv Sci Res2 41e46 wwwadv-sci-resnet2412008

Baklanov A Mahura A Sokhi R (Eds) 2011 Integrated Systems of Meso-Meteorological and Chemical Transport Models first ed Springer ISBN 978-3-642-13979-6 p 242

Baklanov A 2006 Overview of the European project FUMAPEX Atmos ChemPhys 6 2005e2015

Baklanov A 2010 Chemical weather forecasting a new concept of integratedmodeling Adv Sci Res 4 23e27

Baldasano JM Jimeacutenez-Guerrero P Jorba O Peacuterez C Loacutepez E Guumlereca PMartin F Garciacutea-Vivanco M Palomino I Querol X Pandolfi M Sanz MJDieacuteguez JJ 2008 CALIOPE an operational air quality forecasting system for theIberian Peninsula Balearic Islands and Canary Islands e First annual evaluationand ongoing developments Adv Sci Res 2 89e98 ISSN 1992-0628 May

Battan LJ 1966 The Unclean Sky Doubleday Garden City NY 141 ppBerge E Walker S-E Sorteberg A Lenkopane M Eastwood S Jablonska HI

Koltzow MOslash 2002 A real-time operational forecast model for meteorologyand air quality during peak air pollution episodes in Oslo Norway Water AirSoil Poll Focus 2 745e757

Blond N Vautard R 2004 Three-dimensional ozone analysis and their use forshort-term ozone forecasts J Geophys Res 109 D17303 httpdxdoiorg1010292004JD004515

Boylan JW Russell AG 2006 PM and light extinction model performance metricsgoals and criteria for three-dimensional air quality models Atmos Environ 40(26) 4946e4959 httpdxdoiorg101016jatmosenv200509087

Brandt J Christensen JH Frohn LM Palmgren F Berkowicz R Zlatev Z 2001Operational air pollution forecasts from European to local scale Atmos Environ35 (Suppl 1) S91eS98

Brimblecombe P 1987 The Big Smoke a History of Air Pollution in London sinceMedieval Time Routledge Revivals Methuen London UK 186 pp

Burrows WR Benjamin M Beauchamp S Lord ER McCollor D Thomson B1995 CART decision-tree statistical analysis and prediction of summer seasonmaximum surface ozone for the Vancouver Montreal and Atlantic Regions ofCanada J Appl Meteor Climatol 34 1848e1862

Byun D Schere KL 2006 Review of the governing equations computationalalgorithms and other components of the models-3 community multiscale airquality (CMAQ) modeling system Appl Mech Rev 59 51e77

Cai C Hogrefe C Katsafados P Kallos G Beauharnois M Schwab JJ Ren XBrune W Zhou X He Y Demerjian K 2008 Performance evaluation of an airquality forecast modeling system for a summer and winter season e Photo-chemical oxidants and their precursors Atmos Environ 42 8585e8599 httpdxdoiorg101026jatmosenv200808029

Carmichael GR et al 2003 Regional-scale chemical transport modeling in supportof intensive field experiments overview and analysis of the TRACE-P observa-tions J Geophys Res 108 (D21) 8823 httpdxdoiorg1010292002JD003117

Carmichael GR Sandu A Chai T Daescu DN Constantinescu EM Tang Y2008 Predicting air quality improvements through advanced methods tointegrate models and measurements J Comput Phys 227 3540e3571

Chai T Carmichael GR Tang Y Sandu A Streets DG 2007 Four-dimensionaldata assimilation experiments with international consortium for atmosphericresearch on transport and transformation ozone measurements J GeophysRes 112 D12S15 httpdxdoiorg1010292006JD007763

Chaloulakou A Saisana M Spyrellis N 2003 Comparative assessment of neuralnetworks and regression models for forecasting summertime ozone in AthensSci Total Environ 313 1e13

Chen J Vaughan J Avise J OrsquoNeill S Lamb B 2008 Enhancement and evalu-ation of the AIRPACT ozone and PM25 forecast system for the Pacific NorthwestJ Geophys Res 113 D14305 httpdxdoiorg1010292007JD009554

Chenevez J Jensen CO 2001 Operational ozone forecasts for the region of Copen-hagen by the Danish Meteorological Institute Atmos Environ 35 4567e4580

Chin M Ginoux P Lucchesi R Huebert B Weber R Anderson T Masonis SBlomquist B Bandy A Thornton D 2003 A global aerosol model forecast forthe ACE-Asia field experiment J Geophys Res 108 (D23) 8654 httpdxdoiorg1010292003JD003642

Chuang M-T Zhang Y Kang D-W 2011 Application of WRFChem-MADRID forreal-time air quality forecasting over the southeastern United States AtmosEnviron 45 (34) 6241e6250

Cobourn WG Hubbard MC 1999 An enhanced ozone forecasting model using airmass trajectory analysis Atmos Environ 33 4663e4674

Cobourn WG 2007 Accuracy and reliability of an automated air quality forecastsystem for ozone in seven Kentucky metropolitan areas Atmos Environ 415863e5875 httpdxdoiorg101016jatmosenv200703024

Comrie AC 1997 Comparing neural networks and regression models for ozoneforecasting J Air amp Waste Manag Assoc 47 653e663

Cooke WF Liousse C Cachier H Feichter J 1999 Construction of a fossil fuelemission data set for carbonaceous aerosol and implementation and radiativeimpact in the ECHAM4 model J Geophys Res 104 22137e22162

Cope ME Hess GD 2005 Air quality forecasting a review and comparison of theapproaches used internationally and in Australia Clean Air and Environ Quality39 (1) 52e60

Cope ME Hess GD Lee S Tory K Azzi M Carras J Lilley W Manins PCNelson P Ng L Puri K Wong N Walsh S Young M 2004 The Australianair quality forecasting system Part I project description and early outcomesJ Appl Meteorol 43 649e662

recasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e24 19

conducted for precursors of O3 and other gaseous species (eg Caiet al 2008) and vertical profile of forecasted concentrations (egYu et al 2007) Yu et al (2007) reported that EtaCMAQ reproducedO3 vertical distributions on most of the days at low altitudes butoverpredictions occurred at altitudes gt 6 km because of a combi-nation of effects related to the specifications of meteorologicallateral BCONs as well as themodelrsquos coarse vertical resolution in theupper free troposphere

Surface PM forecasts are evaluated in terms of 24-h averagePM25 hourly PM25 and 24-h average PM10 These evaluationshowever have been very limited to date and a consistent evaluationprotocol has not yet been well established (Seigneur 2001 Zhanget al 2006a) The forecasting skill for PM25 is overall poorer thanthat for O3 For 24-h average PM25 MBs RMSEs NMBs and NMEsrange from 32 to 62 mg m3 17 to 159 mg m3 21 to 32 and37 to 81 respectively The NMB of 21 using EtaCMAQ over theeastern US during July 14eAugust 18 2004 was reported by Yuet al (2008) who attributed underpredictions to underestimatedtotal carbonaceous PM at both urban and rural sites and a signifi-cant underestimation of unspecified anthropogenic PM mass(mainly consisting of primary emitted trace elements) at rural sitesOn the other hand Hogrefe et al (2007) showed that EtaCMAQoverpredicted PM25 in the New York City due to overpredictions inorganic aerosols and crustal material Factors contributing to suchoverpredictions include underpredictions of nocturnal verticalmixing inaccurate temporal allocation of primary OM emissionsand underestimate of deposition processes Chen et al (2008)reported the NMB of 17e32 over the Pacific Northwest duringAugusteNovember 2004 The overpredictions in PM25 wereattributed to uncertainties in wildfire emission estimates and themodeling error in fire plume transport due to errors in MM5-predicted wind direction and wind speed McKeen et al (2007)evaluated RT-AQF of 6 models including WRFChem at two gridresolutions (12- and 36-km) CHRONOS AURAMS STEM-2K3 andEtaCMAQ They found that most models did not reproduce theobserved diurnal variation at urban and suburban sites particularlyduring the nighttime to early morning and the ensemble meanbased on 6 models with equal weighting gave the best possibleforecast in terms of statistics Chuang et al (2011) showed slightunderpredictions of PM25 in the O3 season over the southeasternUS by WRFChemeMADRID which were attributed to uncer-tainties in emissions such as those of biogenic VOCs and NH3overpredictions of precipitation and uncertainties in the BCONsOnly two studies evaluated hourly PM25 givingMBs RMSEs NMBsand NMEs of 33 to06 mg m3 83e113 mg m3 211 to52and 498e514 respectively For 24-h average PM10 MBs andRMSEs range from 29 to 138 mg m3 and 83 to 472 mg m3respectively Berge et al (2002) reported forecasted PM10 perfor-mance over Oslo with an MB of 209 mg m3 and an NMBof 365 at Kirkeveien due to an inaccurate emissions of PM10from the surface (ie the re-suspension of dust deposited at theroadside) on dry days and an MB of 138 mg m3 and an NMB of 24at Furuset due to errors in simulated grid-averaged wind fields

Very few evaluations were conducted for PM componentsMcKeen et al (2007) found that all the six RT-AQF models signifi-cantly underpredicted OM at the surface and overestimated SO4

2

above 2 km The overpredictions in SO42 were attributed to over-

estimate of SO2 by WRFChem and CHRONOS and the inclusion ofaqueous-phase oxidation of SO2 by AURAMS CMAQETA andSTEM-2K3 Yu et al (2007) reported that EtaCMAQ overpredictedSO4

2 due to too much in-cloud SO2 oxidation as a result of over-estimated H2O2 concentrations in the model underpredicted NH4

thorn

at the rural sites and aloft due to the exclusion of some sources ofNH3 in the real-time emission inventory underpredicted NO3

dueto overpredictions of SO4

2 and underpredicted OC due to missing

Please cite this article in press as Zhang Y et al Real-time air quality forEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

sources of primary OC in the emission inventory and missing SOAformation from the gas-phase oxidation of isoprene and sesqui-terpenes and aqueous-phase oxidation of glyoxal and methyl-glyoxal Chen et al (2008) reported that AIRPACT-3 reproduced OCwell but significantly overpredicted EC and significantly under-predicted SO4

2 NO3 and NH4

thorn due to underestimation of emis-sions of primary PM species (eg sulfate) and precursors ofsecondary PM (eg SO2 NH3 and NOx) insufficient spatial reso-lution and modelrsquos inability to capture the hourly PM variations

For categorical evaluation of PM some studies used thresholdvalues of 15e315 mg m3 for 24-h average PM25 and 30e50 mgm3

for 24-h average PM10 that are lower than the NAAQSs of 24-haverage PM25 of 35 mg m3 and 24-h average PM10 of 150 mg m3

for the same reason as mentioned previously For 24-h averagePM25 A CSI POD B and FAR range from 608 to 997 0 to 5370 to 909 07 to 19 mg m3 and 25 to 100 respectively Those for24-h average PM10 range from 167 to 100 91 to 100 150 to100 05 to 09 mg m3 and 0 to 48 respectively Categoricalevaluation of hourly PM25 was conducted only with WRFChemeMADRID with A CSI POD B and FAR of 721 205292 07 mg m3 and 591 respectively

5 Summary of Current States of RT-AQF

Emerging as a new discipline of applied sciences that integratesseveral physical sciences mathematicalstatistical tools andcomputer technology in the 1970s RT-AQF represents a uniqueapplied science affecting humanrsquos daily activities and posesunprecedented challenges intersecting many aspects of sciencesand technologies Driven by deteriorated air quality worldwideincreased societalhuman demands and rapid scientific and tech-nological advancements significant progress has been made in thepast four decades in RT-AQF A number of RT-AQF tools and modelswith varying degrees of sophistication and forecasting skillsranging from the simplest rule of thumb to the most advanced 3-Donline-coupled meteorologyechemistry models have been devel-oped Among them the deterministic physically-based RT-AQFmodels that are based on 3-D global and regional CTMs representthe state of the science enabling scientific understanding ofmechanisms for pollutant formation and development of emissioncontrol strategies Although other approaches in particular para-metric (statistical) models will continue to be used the 3-D RT-AQFmodels will likely become prevalent approaches and tools on all thescales There is a growing trend to combine these 3-D modelswith statistical models to provide more accurate RT-AQF Amongall RT-AQF models the coupled meteorologyechemistry modelrepresents a significant advancement and will greatly enhance theunderstanding of the underlying complex interplay of meteorologyemissions and chemistry in the real atmosphere Importantextensions of the RT-AQF models include their coupling with anurban model (eg traffic andor local pollutant dispersion) or a CFDmodel for urbanlocal scale applications at a spatial resolution of1 km or less and with an exposure model to provide real-timepublic health assessment and exposure predictions and urbanemergency preparedness

3-D RT-AQF models have been used for 24e72 h RT-AQFworldwide since the mid 1990s Despite their unique characteris-tics and technical requirements 3-D RT-AQF models are based onNWP models and CTM development and application efforts in thepast decades and therefore are subject to limitations of thesemodels Although no universal forecasting guidance and evaluationprotocols were developed covering all countries some guidanceand protocols were recommended by governmental agencies andthe scientific community Discrete and categorical evaluations aretwo most commonly-used evaluations Real-time or near real-time

ecasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e2420

surface air quality measurement and satellite datasets areincreasingly available for such evaluations The evaluation effortshave been primarily focused on O3 NO2 and PM (both PM10 andPM25) at surface level The forecasting skill is typically better for O3than for PM25 While their performance statistics averaged overa time period (eg monthly or daily) and the whole domain isgenerally good or satisfactory current RT-AQF models have diffi-culties forecasting variations at finer time scales (hourly and diur-nally) and individual monitoring stations In addition larger biasesexist for forecasted NOx HONO HNO3 and OH and PM compositionincluding SO4

2 NO3 and OM indicating possible error compen-

sations that lead to current levels of forecast accuracy These fore-cast inaccuracies have been attributed to a number of factors thatare highly uncertain or lack of accurate treatments in the RT-AQFmodels including mesoscale meteorological processes (eg sea-breeze circulations) and variables (eg temperature PBL height)chemical BCONs of O3 and PM25 emissions (eg SO2 NOx NH3BVOCs primary PM) physical and chemical processes (eg urbanprocesses in-cloud oxidation of SO2 SOA formation) and modelconfiguration (eg coarse grid resolution) Recent advances havebeenmade in some of those aspects Part II of this reviewwill assessand discuss these advances along with advanced computationaltechniques that can potentially improve RT-AQF Recommendationsfor research priorities and future prospects will also be provided

Acknowledgments

This work was performed under the Fellowship Award(704389J) Center for Education and ResearchEacutecole des PontsParisTech France through a visiting professorship at the Atmo-spheric Environment Center (CEREA) the Joint Laboratory Eacutecole desPonts ParisTech and EDF RampD Paris France the NSF Career Award(ATM-0348819) the EPA-Science to Achieve Results (STAR)program (R83337601) and Chinarsquos National Basic ResearchProgram (2010CB951803) Support from CEEH MEGAPOLI andCOST ES1004 is also acknowledged Thanks are due to Ming-TungChuang a former post-D researcher at NCSU for compiling anearlier version of Table A4 Thanks are due to Yelva RoustanAtmospheric Research Center (CEREA) France Christian Hogrefethe New York State Department of Environmental Conservation(NYSDEC) NY US and John McHenry Baron Advanced Meteoro-logical Systems (BAMS) US for providing some informationregarding their air quality forecasting models

Appendix A Supplementary Material

Supplementary data related to this article can be found online athttpdxdoiorg101016jatmosenv201206031

References

Alhanafy TE Zaghlool F El AS Moustafa D 2010 Neuro fuzzy modeling schemefor the prediction of air pollution J American Sci 6 (12) 605e616

Al-Saadi J Szykman J Pierce RB Kittaka C Neil D Chu DA Remer LGumley L Prins E Weinstock L MacDonald C Wayland R Dimmick FFishman J 2005 Improving national air quality forecasts with satellite aerosolobservations Bull Amer Meteor Soc 86 1249e1261

Arnold JR Dennis RL 2006 Testing CMAQ chemistry sensitivities in base caseand emissions control runs at SEARCH and SOS99 surface sites in the south-eastern US Atmos Environ 40 5027e5040

Aron RH 1980 Forecasting high level oxidant concentrations in the Los Angelesbasin J Air Pollut Control Assoc 20 1227e1228

Baklanov A Haumlnninen O Slordal LH Kukkonen J Bjergene N Fay B Finardi SHoe SC Jantunen M Karppinen A Rasmussen A Skouloudis A Sokhi RSSorensen j H Oslashdegaard V 2007 Integrated systems for forecasting urbanmeteorology air pollution and population exposure Atmos Chem Phys 7855e874 wwwatmos-chem-physnet78552007

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

Baklanov A Korsholm U Mahura A Petersen C Gross A 2008 Enviro-HIRLAMon-line coupled modelling of urban meteorology and air pollution Adv Sci Res2 41e46 wwwadv-sci-resnet2412008

Baklanov A Mahura A Sokhi R (Eds) 2011 Integrated Systems of Meso-Meteorological and Chemical Transport Models first ed Springer ISBN 978-3-642-13979-6 p 242

Baklanov A 2006 Overview of the European project FUMAPEX Atmos ChemPhys 6 2005e2015

Baklanov A 2010 Chemical weather forecasting a new concept of integratedmodeling Adv Sci Res 4 23e27

Baldasano JM Jimeacutenez-Guerrero P Jorba O Peacuterez C Loacutepez E Guumlereca PMartin F Garciacutea-Vivanco M Palomino I Querol X Pandolfi M Sanz MJDieacuteguez JJ 2008 CALIOPE an operational air quality forecasting system for theIberian Peninsula Balearic Islands and Canary Islands e First annual evaluationand ongoing developments Adv Sci Res 2 89e98 ISSN 1992-0628 May

Battan LJ 1966 The Unclean Sky Doubleday Garden City NY 141 ppBerge E Walker S-E Sorteberg A Lenkopane M Eastwood S Jablonska HI

Koltzow MOslash 2002 A real-time operational forecast model for meteorologyand air quality during peak air pollution episodes in Oslo Norway Water AirSoil Poll Focus 2 745e757

Blond N Vautard R 2004 Three-dimensional ozone analysis and their use forshort-term ozone forecasts J Geophys Res 109 D17303 httpdxdoiorg1010292004JD004515

Boylan JW Russell AG 2006 PM and light extinction model performance metricsgoals and criteria for three-dimensional air quality models Atmos Environ 40(26) 4946e4959 httpdxdoiorg101016jatmosenv200509087

Brandt J Christensen JH Frohn LM Palmgren F Berkowicz R Zlatev Z 2001Operational air pollution forecasts from European to local scale Atmos Environ35 (Suppl 1) S91eS98

Brimblecombe P 1987 The Big Smoke a History of Air Pollution in London sinceMedieval Time Routledge Revivals Methuen London UK 186 pp

Burrows WR Benjamin M Beauchamp S Lord ER McCollor D Thomson B1995 CART decision-tree statistical analysis and prediction of summer seasonmaximum surface ozone for the Vancouver Montreal and Atlantic Regions ofCanada J Appl Meteor Climatol 34 1848e1862

Byun D Schere KL 2006 Review of the governing equations computationalalgorithms and other components of the models-3 community multiscale airquality (CMAQ) modeling system Appl Mech Rev 59 51e77

Cai C Hogrefe C Katsafados P Kallos G Beauharnois M Schwab JJ Ren XBrune W Zhou X He Y Demerjian K 2008 Performance evaluation of an airquality forecast modeling system for a summer and winter season e Photo-chemical oxidants and their precursors Atmos Environ 42 8585e8599 httpdxdoiorg101026jatmosenv200808029

Carmichael GR et al 2003 Regional-scale chemical transport modeling in supportof intensive field experiments overview and analysis of the TRACE-P observa-tions J Geophys Res 108 (D21) 8823 httpdxdoiorg1010292002JD003117

Carmichael GR Sandu A Chai T Daescu DN Constantinescu EM Tang Y2008 Predicting air quality improvements through advanced methods tointegrate models and measurements J Comput Phys 227 3540e3571

Chai T Carmichael GR Tang Y Sandu A Streets DG 2007 Four-dimensionaldata assimilation experiments with international consortium for atmosphericresearch on transport and transformation ozone measurements J GeophysRes 112 D12S15 httpdxdoiorg1010292006JD007763

Chaloulakou A Saisana M Spyrellis N 2003 Comparative assessment of neuralnetworks and regression models for forecasting summertime ozone in AthensSci Total Environ 313 1e13

Chen J Vaughan J Avise J OrsquoNeill S Lamb B 2008 Enhancement and evalu-ation of the AIRPACT ozone and PM25 forecast system for the Pacific NorthwestJ Geophys Res 113 D14305 httpdxdoiorg1010292007JD009554

Chenevez J Jensen CO 2001 Operational ozone forecasts for the region of Copen-hagen by the Danish Meteorological Institute Atmos Environ 35 4567e4580

Chin M Ginoux P Lucchesi R Huebert B Weber R Anderson T Masonis SBlomquist B Bandy A Thornton D 2003 A global aerosol model forecast forthe ACE-Asia field experiment J Geophys Res 108 (D23) 8654 httpdxdoiorg1010292003JD003642

Chuang M-T Zhang Y Kang D-W 2011 Application of WRFChem-MADRID forreal-time air quality forecasting over the southeastern United States AtmosEnviron 45 (34) 6241e6250

Cobourn WG Hubbard MC 1999 An enhanced ozone forecasting model using airmass trajectory analysis Atmos Environ 33 4663e4674

Cobourn WG 2007 Accuracy and reliability of an automated air quality forecastsystem for ozone in seven Kentucky metropolitan areas Atmos Environ 415863e5875 httpdxdoiorg101016jatmosenv200703024

Comrie AC 1997 Comparing neural networks and regression models for ozoneforecasting J Air amp Waste Manag Assoc 47 653e663

Cooke WF Liousse C Cachier H Feichter J 1999 Construction of a fossil fuelemission data set for carbonaceous aerosol and implementation and radiativeimpact in the ECHAM4 model J Geophys Res 104 22137e22162

Cope ME Hess GD 2005 Air quality forecasting a review and comparison of theapproaches used internationally and in Australia Clean Air and Environ Quality39 (1) 52e60

Cope ME Hess GD Lee S Tory K Azzi M Carras J Lilley W Manins PCNelson P Ng L Puri K Wong N Walsh S Young M 2004 The Australianair quality forecasting system Part I project description and early outcomesJ Appl Meteorol 43 649e662

recasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e2420

surface air quality measurement and satellite datasets areincreasingly available for such evaluations The evaluation effortshave been primarily focused on O3 NO2 and PM (both PM10 andPM25) at surface level The forecasting skill is typically better for O3than for PM25 While their performance statistics averaged overa time period (eg monthly or daily) and the whole domain isgenerally good or satisfactory current RT-AQF models have diffi-culties forecasting variations at finer time scales (hourly and diur-nally) and individual monitoring stations In addition larger biasesexist for forecasted NOx HONO HNO3 and OH and PM compositionincluding SO4

2 NO3 and OM indicating possible error compen-

sations that lead to current levels of forecast accuracy These fore-cast inaccuracies have been attributed to a number of factors thatare highly uncertain or lack of accurate treatments in the RT-AQFmodels including mesoscale meteorological processes (eg sea-breeze circulations) and variables (eg temperature PBL height)chemical BCONs of O3 and PM25 emissions (eg SO2 NOx NH3BVOCs primary PM) physical and chemical processes (eg urbanprocesses in-cloud oxidation of SO2 SOA formation) and modelconfiguration (eg coarse grid resolution) Recent advances havebeenmade in some of those aspects Part II of this reviewwill assessand discuss these advances along with advanced computationaltechniques that can potentially improve RT-AQF Recommendationsfor research priorities and future prospects will also be provided

Acknowledgments

This work was performed under the Fellowship Award(704389J) Center for Education and ResearchEacutecole des PontsParisTech France through a visiting professorship at the Atmo-spheric Environment Center (CEREA) the Joint Laboratory Eacutecole desPonts ParisTech and EDF RampD Paris France the NSF Career Award(ATM-0348819) the EPA-Science to Achieve Results (STAR)program (R83337601) and Chinarsquos National Basic ResearchProgram (2010CB951803) Support from CEEH MEGAPOLI andCOST ES1004 is also acknowledged Thanks are due to Ming-TungChuang a former post-D researcher at NCSU for compiling anearlier version of Table A4 Thanks are due to Yelva RoustanAtmospheric Research Center (CEREA) France Christian Hogrefethe New York State Department of Environmental Conservation(NYSDEC) NY US and John McHenry Baron Advanced Meteoro-logical Systems (BAMS) US for providing some informationregarding their air quality forecasting models

Appendix A Supplementary Material

Supplementary data related to this article can be found online athttpdxdoiorg101016jatmosenv201206031

References

Alhanafy TE Zaghlool F El AS Moustafa D 2010 Neuro fuzzy modeling schemefor the prediction of air pollution J American Sci 6 (12) 605e616

Al-Saadi J Szykman J Pierce RB Kittaka C Neil D Chu DA Remer LGumley L Prins E Weinstock L MacDonald C Wayland R Dimmick FFishman J 2005 Improving national air quality forecasts with satellite aerosolobservations Bull Amer Meteor Soc 86 1249e1261

Arnold JR Dennis RL 2006 Testing CMAQ chemistry sensitivities in base caseand emissions control runs at SEARCH and SOS99 surface sites in the south-eastern US Atmos Environ 40 5027e5040

Aron RH 1980 Forecasting high level oxidant concentrations in the Los Angelesbasin J Air Pollut Control Assoc 20 1227e1228

Baklanov A Haumlnninen O Slordal LH Kukkonen J Bjergene N Fay B Finardi SHoe SC Jantunen M Karppinen A Rasmussen A Skouloudis A Sokhi RSSorensen j H Oslashdegaard V 2007 Integrated systems for forecasting urbanmeteorology air pollution and population exposure Atmos Chem Phys 7855e874 wwwatmos-chem-physnet78552007

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

Baklanov A Korsholm U Mahura A Petersen C Gross A 2008 Enviro-HIRLAMon-line coupled modelling of urban meteorology and air pollution Adv Sci Res2 41e46 wwwadv-sci-resnet2412008

Baklanov A Mahura A Sokhi R (Eds) 2011 Integrated Systems of Meso-Meteorological and Chemical Transport Models first ed Springer ISBN 978-3-642-13979-6 p 242

Baklanov A 2006 Overview of the European project FUMAPEX Atmos ChemPhys 6 2005e2015

Baklanov A 2010 Chemical weather forecasting a new concept of integratedmodeling Adv Sci Res 4 23e27

Baldasano JM Jimeacutenez-Guerrero P Jorba O Peacuterez C Loacutepez E Guumlereca PMartin F Garciacutea-Vivanco M Palomino I Querol X Pandolfi M Sanz MJDieacuteguez JJ 2008 CALIOPE an operational air quality forecasting system for theIberian Peninsula Balearic Islands and Canary Islands e First annual evaluationand ongoing developments Adv Sci Res 2 89e98 ISSN 1992-0628 May

Battan LJ 1966 The Unclean Sky Doubleday Garden City NY 141 ppBerge E Walker S-E Sorteberg A Lenkopane M Eastwood S Jablonska HI

Koltzow MOslash 2002 A real-time operational forecast model for meteorologyand air quality during peak air pollution episodes in Oslo Norway Water AirSoil Poll Focus 2 745e757

Blond N Vautard R 2004 Three-dimensional ozone analysis and their use forshort-term ozone forecasts J Geophys Res 109 D17303 httpdxdoiorg1010292004JD004515

Boylan JW Russell AG 2006 PM and light extinction model performance metricsgoals and criteria for three-dimensional air quality models Atmos Environ 40(26) 4946e4959 httpdxdoiorg101016jatmosenv200509087

Brandt J Christensen JH Frohn LM Palmgren F Berkowicz R Zlatev Z 2001Operational air pollution forecasts from European to local scale Atmos Environ35 (Suppl 1) S91eS98

Brimblecombe P 1987 The Big Smoke a History of Air Pollution in London sinceMedieval Time Routledge Revivals Methuen London UK 186 pp

Burrows WR Benjamin M Beauchamp S Lord ER McCollor D Thomson B1995 CART decision-tree statistical analysis and prediction of summer seasonmaximum surface ozone for the Vancouver Montreal and Atlantic Regions ofCanada J Appl Meteor Climatol 34 1848e1862

Byun D Schere KL 2006 Review of the governing equations computationalalgorithms and other components of the models-3 community multiscale airquality (CMAQ) modeling system Appl Mech Rev 59 51e77

Cai C Hogrefe C Katsafados P Kallos G Beauharnois M Schwab JJ Ren XBrune W Zhou X He Y Demerjian K 2008 Performance evaluation of an airquality forecast modeling system for a summer and winter season e Photo-chemical oxidants and their precursors Atmos Environ 42 8585e8599 httpdxdoiorg101026jatmosenv200808029

Carmichael GR et al 2003 Regional-scale chemical transport modeling in supportof intensive field experiments overview and analysis of the TRACE-P observa-tions J Geophys Res 108 (D21) 8823 httpdxdoiorg1010292002JD003117

Carmichael GR Sandu A Chai T Daescu DN Constantinescu EM Tang Y2008 Predicting air quality improvements through advanced methods tointegrate models and measurements J Comput Phys 227 3540e3571

Chai T Carmichael GR Tang Y Sandu A Streets DG 2007 Four-dimensionaldata assimilation experiments with international consortium for atmosphericresearch on transport and transformation ozone measurements J GeophysRes 112 D12S15 httpdxdoiorg1010292006JD007763

Chaloulakou A Saisana M Spyrellis N 2003 Comparative assessment of neuralnetworks and regression models for forecasting summertime ozone in AthensSci Total Environ 313 1e13

Chen J Vaughan J Avise J OrsquoNeill S Lamb B 2008 Enhancement and evalu-ation of the AIRPACT ozone and PM25 forecast system for the Pacific NorthwestJ Geophys Res 113 D14305 httpdxdoiorg1010292007JD009554

Chenevez J Jensen CO 2001 Operational ozone forecasts for the region of Copen-hagen by the Danish Meteorological Institute Atmos Environ 35 4567e4580

Chin M Ginoux P Lucchesi R Huebert B Weber R Anderson T Masonis SBlomquist B Bandy A Thornton D 2003 A global aerosol model forecast forthe ACE-Asia field experiment J Geophys Res 108 (D23) 8654 httpdxdoiorg1010292003JD003642

Chuang M-T Zhang Y Kang D-W 2011 Application of WRFChem-MADRID forreal-time air quality forecasting over the southeastern United States AtmosEnviron 45 (34) 6241e6250

Cobourn WG Hubbard MC 1999 An enhanced ozone forecasting model using airmass trajectory analysis Atmos Environ 33 4663e4674

Cobourn WG 2007 Accuracy and reliability of an automated air quality forecastsystem for ozone in seven Kentucky metropolitan areas Atmos Environ 415863e5875 httpdxdoiorg101016jatmosenv200703024

Comrie AC 1997 Comparing neural networks and regression models for ozoneforecasting J Air amp Waste Manag Assoc 47 653e663

Cooke WF Liousse C Cachier H Feichter J 1999 Construction of a fossil fuelemission data set for carbonaceous aerosol and implementation and radiativeimpact in the ECHAM4 model J Geophys Res 104 22137e22162

Cope ME Hess GD 2005 Air quality forecasting a review and comparison of theapproaches used internationally and in Australia Clean Air and Environ Quality39 (1) 52e60

Cope ME Hess GD Lee S Tory K Azzi M Carras J Lilley W Manins PCNelson P Ng L Puri K Wong N Walsh S Young M 2004 The Australianair quality forecasting system Part I project description and early outcomesJ Appl Meteorol 43 649e662

recasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e24 21

Curci G 2010 An air quality forecasting tool over Italy In Proceedings of 31stTechnical Meeting on Air Pollution Modelling and its Application Torino ItalySeptember

Dabberdt WF Carroll MA Baumgardner D Carmichael G Cohen R Dye TEllis J Grell G Grimmond S Hannaa S Irwin J Lamb B Madronich SMcQueen J Meagher J Odman T Pleim J Schmid HP Westphal DL 2004Meteorological research needs for improved air quality forecasting report ofthe 11th prospectus development team of the U S Weather Research ProgramBull Amer Meteor Soc 85 563e586

Dabberdt WF Carroll MA Appleby W Baumgardner D Carmichael GDavidson P Doran JC Dye TS Grimmond S Middeleton P Neff WZhang Y 2006 USWRP workshop on air quality forecasting Bull AmerMeteor Soc 86 215e221

de Freitas ED Martins LD Silva Dias PL Faacutetima Andrade M 2005 A simplephotochemical module implemented in RAMS for tropospheric ozoneconcentration forecast in the metropolitan area of Satildeo Paulo Brazil couplingand validation Atmos Environ 39 6352e6361

Debry Eacute Fahey K Sartelet K Sportisse B Tombette M 2007 Technical notea new size REsolved Aerosol Model SIREAM Atmos Chem Phys 7 1537e1547

Delle Monache L Wilczak J McKeen S Grell G Pagowski M Peckham SStull R McHenry J McQueen J 2008 A Kalman filter bias correction methodapplied to deterministic ensemble averaged and probabilistic forecasts ofsurface ozone Tellus Ser B 60 238e249 httpdxdoiorg101111j1600-0889200700332x

Denby B Schaap M Segers A Builtjes P Horaacutelek J 2008 Comparison of twodata assimilation methods for assessing PM10 exceedances on the Europeanscale Atmos Environ 42 7122e7134

Diaz-Robles LA Ortega JC Fu JS Reed GD Chow JC Watson JG Moncada-Herrera JA 2008 A hybrid ARIMA and artificial neural networks model toforecast particulate matter in urban areas the case of Temuco Chile AtmosEnviron 42 8331e8340

Dockery D Pope C Xu XP Spengler JD Ware JH Fay ME Ferris BGSpeizer FE 1993 An association between air-pollution and mortality in 6United-states cities N Engl J Med 329 (24) 1753e1759

Doraiswamy P Hogrefe C Hao W Colle B Beauharnois M Demerjian KKu J-Y Sistla G 2009 Preliminary Experiences with the Multi-Model AirQuality Forecasting System for New York State Paper presented at the 8thAnnual Community Modeling and Analysis System (CMAS) ConferenceChapel Hill NC

Dutot A-L Rynkiewicz J Steiner FE Rude J 2007 A 24-h forecast ofozone peaks and exceedance levels using neural classifiers and weatherpredictions Environ Model Software 22 1261e1269 httpdxdoiorg101016jenvsoft200608002

Eben K Jurus

ˇ

P Resler J Belda M Pelikaacuten E Kruumlger BC Keder J 2005 Anensemble Kalman filter for short-term forecasting of tropospheric ozoneconcentrations Q J R Meteorol Soc 131 3313e3322

Eder BK Kang D Mathur R Yu S Schere K 2006 An operational evaluation ofthe EtaeCMAQ air quality forecast model Atmos Environ 40 4894e4905

Eder B Kang D Rao ST Mathur R Yu S Otte T Schere K Waylard RJackson S Davidson P McQueen J Bridgers G 2010 Using national airquality forecast guidance to develop local air quality index forecasts Bull AmerMeteor Soc 91 313e326

Elbern H Schmidt H 2001 Ozone episode analysis by four dimensional varia-tional chemistry data assimilation J Geophys Res 106 3569e3590

Elbern H Friese E Strunk A 2010 EURAD-IM Products Quality and BackgroundInformation Report Prepared for MACC Report D_R-ENS_132 March

Elshout S v d Leacuteger K 2007 Comparing Urban Air Quality across Borders ea Review of Existing Air Quality Indices and the Proposal of a Common Alter-native Final Report on the CITEAIR Project of the INTERREG IIIC programmeprepared by DCMR Environmental Protection Agency Rijnmond TheNetherlands June

Ecological Society of America (ESA) 2004 Impacts of Atmospheric Pollution onAquatic Ecosystems In Issues in Ecology vol 12 ISSN 1092-8987 TheEcological Society of America Washington DC 20006 also Available at httpwwwesaorgsbisbi_issues 26 pp

European Union (EU) 2008 Directive 200850EC of the European Parliament andof the Council of 21 May 2008 on ambient air quality and cleaner air for Europe1162008 L OJEU 152 1e44

Faraji M Kimura Y McDonald-Buller E Allen D 2008 Comparison of the carbonbond and SAPRC photochemical mechanisms under conditions relevant tosoutheast Texas Atmos Environ 42 5821e5836

Fast JD Gustafson Jr WI Easter RC Zaveri RA Barnard JC Chapman EGGrell GA Peckham SE 2006 Evolution of ozone particulates and aerosoldirect radiative forcing in the vicinity of Houston using a fully coupledmeteorology-chemistry-aerosol model J Geophys Res 111 D21305 httpdxdoiorg1010292005JD006721

Fast J et al 2009 Evaluating simulated primary anthropogenic and biomassburning organic aerosols during MILAGRO implications for assessing treat-ments of secondary organic aerosols Atmos Chem Phys 9 6191e6215 httpdxdoiorg105194acp-9-6191-2009

Flatoslashy F Hov O Schlager H 2000 Chemical forecasts used for measurementflight planning during POLINAT 2 Geophys Res Lett 27 951e954

Flemming J Inness A Flentje H Huijnen V Moinat P Schultz MG Stein O2009 Coupling global chemistry transport models to ECMWFrsquos integratedforecast system Geosci Model Dev 2 253e265

Please cite this article in press as Zhang Y et al Real-time air quality forEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

Folberth GA Hauglustaine DA Lathire J Rocheton F 2006 Interactive chem-istry in the Laboratoire de Meacuteteacuteorologie Dynamique general circulation modelmodel description and impact analysis of biogenic hydrocarbons on tropo-spheric chemistry Atmos Chem Phys 6 2273e2319

Forster C et al 2004 Lagrangian transport model forecasts and a transportclimatology for the Intercontinental Transport and Chemical Transformation2002 (ITCT 2K2) measurement campaign J Geophys Res 109 D07S92 httpdxdoiorg1010292003JD003589

Fountoukis C Nenes A 2005 Continued development of a cloud droplet forma-tion parameterization for global climate models J Geophys Res 110 D11212httpdxdoiorg1010292004JD005591

Fraser A Davies T Lingard J 2010 Air Quality Forecasting with WRF and CMAQNational Air Quality Forecasting Seminar July 14 AEA Harwell UK

Genc DD Yesilyurt C Tuncel G 2010 Air pollution forecasting in Ankara Turkeyusing air pollution index and its relation to assimilative capacity of the atmo-sphere Environ Monit Assess 166 11e27 httpdxdoiorg101007s10661-009-0981-y

Georgopoulos et al 2009 Air quality modeling needs for exposure assessmentfrom the source-to-outcome perspective Environ Manage October Issue26-35

Gilliam RC Hogrefe C Rao ST 2006 New methods for evaluating meteoro-logical models used in air quality applications Atmos Environ 40 5073e5086

Glahn HR Lowry DA 1972 The use of model output statistics (MOS) in objectiveweather forecasting J Appl Meteorol 11 1203e1211

Greenbaum DS Bachmann JD Krewski D Samet JM White R Wyzga RE2001 Particulate air pollution standards and morbidity and mortality casestudy Am J Epidemiol 154 (12) Supplement S78eS90

Grell G Dudhia J Stauffer D 1995 A Description of the Fifth-Generation PennStateNCAR Mesoscale Model (MM5) NCAR Technical Note NCARTN-398thornSTR

Grell GA Knoche R Peckham SE McKeen SA 2004 Online versus offline airquality modeling on cloud-resolving scales Geophys Res Lett 31 L16117httpdxdoiorg1010292004GL020175

Grell GA Peckham SE Schmitz R McKeen SA Frost G Skamarock WCEder B 2005 Fully coupled ldquoonlinerdquo chemistry within the WRF model AtmosEnviron 39 6957e6975

Gross A Stockwell WR 2003 Comparison of the EMEP RADM2 and RACMmechanisms J Atmos Chem 44 151e170

Gross E 1970 The National Air Pollution Potential Forecast Program EESA TechMemo WBTM NMC 47 ETAC TN 70-9 National Meteorological Center SuitlandMD November

Guillas S Bao J Choi Y Wang Y 2008 Statistical correction and downscaling ofchemical transport model ozone forecasts over Atlanta Atmos Environ 421338e1348 httpdxdoiorg101016jatmosenv200710027

Hadley OL Ramanathan V Carmichael GR Tang Y Corrigan CE Roberts GCMauger GS 2007 Trans-Pacific transport of black carbon and fine aerosols(D lt 25 mm) into North America J Geophys Res 112 D05309 httpdxdoiorg1010292006JD007632

Haldane J 1931 Atmospheric pollution and fogs Br Med J 366e367Han Z-W Du S-Y Lei X-E Ju L-X Wang Q-G 2002 Numerical model system

of urban air pollution prediction and its application China Environ Sci 22 (3)202e206 (in Chinese)

Hauglustaine DA Hourdin F Jourdain L Filiberti MA Walters S Lamarque JFHolland EA 2004 Interactive chemistry in the Laboratoire de MeacuteteacuteorologieDynamique general circulation model description and background tropo-spheric chemistry evaluation J Geophys Res 109 D04314 httpdxdoiorg1010292003JD003957

Hogrefe C Hao W Civerolo K Ku J-Y Sistla G Gaza RS Sedefian L Schere KGilliland A Mathur R 2007 Daily simulation of ozone and fine particulatesover New York State findings and challenges J Appl Meteorol Climatol 46 (7)961e979 httpdxdoiorg101175JAM25201

Hoi KI Yuen KV Mok KM 2008 Kalman filter based prediction system forwintertime PM10 concentrations in Macau Global NEST J 10 (2) 140e150

Honoreacute C Rouiumll L Vautard R Beekmann M Bessagnet B Dufour AElichegaray C Flaud J-M Malherbe L Meleux F Menut L Martin DPeuch A Peuch V-H Poisson N 2008 Predictability of European airquality assessment of 3 years of operational forecasts and analysis by thePREVrsquoAIR system J Geophys Res 113 D04301 httpdxdoiorg1010292007JD008761

Jakobs HJ Tilmes S Heidegger A Nester K Smiatek G 2002 Short-term ozoneforecasting with a network model system during summer 1999 J Atmos Chem42 23e40

Jayachandran S 2009 Air quality and early-life mortality evidence fromIndonesiarsquos wildfires J Hum Resour 44 (4) 916e954

Jimeacutenez-Guerrero P Peacuterez C Jorba O Baldasano JM 2008 Contribution ofSaharan dust in an integrated air quality system and its on-line assessmentGeophys Res Lett 35 L03814 httpdxdoiorg1010292007GL031580

Jorquera H Perez R Cipriano A Espejo A Letelier MV Acuna G 1998Forecasting ozone daily maximum levels at Santiago Chile Atmos Environ 323415e3424

Kallos G Astitha M Katsafados P Spyrou C 2007 Long-range transport ofanthropogenically and naturally produced particulate matter in the Mediter-ranean and North Atlantic current state of knowledge J Appl MeteorolClimatol 46 1230e1251

Kallos G Spyrou C Astitha M Mitsakou C Solomos S Kushta J Pytharoulis IKatsafados P Mavromatidis E Papantoniou N 2009 Ten-year operational

ecasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e2422

dust forecasting e Recent model development and future plans IOP Conf SerEarth Environ Sci 7 012012 httpdxdoiorg1010881755-130771012012

Kaminski JW et al 2008 GEM-AQ an on-line global multiscale chemical weathermodelling system model description and evaluation of gas phase chemistryprocesses Atmos Chem Phys 8 3255e3281

Kang D Eder BK Stein AF Grell GA Peckham SE McHenry J 2005 The NewEngland air quality forecasting pilot program development of an evaluationprotocol and performance benchmark J Air amp Waste Manage Assoc 551782e1796

Kang D Mathur R Schere K Yu S Eder B 2007 New categorical metrics for airquality model evaluation J Appl Meteorol Climatol 46 549e555 httpdxdoiorg101175JAM24791

Kang D Mathur R Rao ST Yu S 2008 Bias adjustment techniques forimproving ozone air quality forecasts J Geophys Res 113 D23308 httpdxdoiorg1010292008JD010151

Kim Y Sartelet K Seigneur C 2009 Comparison of two gas-phase chemicalkinetic Mechanisms of ozone formation over Europe J Atmos Chem 6289e119

Kim Y Sartelet K Seigneur C 2011a Formation of secondary aerosols impact ofthe gas-phase chemical mechanism Atmos Chem Phys 11 583e598 httpdxdoiorg105194acp-11-583-2011

Kim Y Couvidat F Sartelet K Seigneur C 2011b Comparison of different gas-phase mechanisms and aerosol modules for simulating particulate matterformation J Air Waste Manag Assoc 61 1218e1226 httpdxdoiorg101080104732892011603999

Kondragunta S Lee P McQueen J Kittaka C Prados AI Ciren P Laszlo IPierce RB Hoff R Szykman JJ 2008 Air quality forecast verification usingsatellite data J Appl Meteor Climatol 47 425e442

Konovalov IB Beekmann M Meleux F Dutot A Foret G 2009 Combiningdeterministic and statistical approaches for PM10 forecasting in Europe AtmosEnviron 43 6425e6434 httpdxdoiorg101016jatmosenv200906039

Krupa SV Grunhage L Jager HJ Nosal M Manning WJ Legge AHHanewald K 2006 Ambient ozone (O3) and adverse crop response a unifiedview of cause and effect Environ Pollut 87 119e126

Kukkonen J et al 2011 Operational regional-scale chemical weather forecastingmodels in Europe Atmos Chem Phys Discuss 11 5985e6162

Kunzli N Kaiser R Medina S Studnicka M Filliger P Chanel O Herry MHorak F Puybonnieux-Texier V Quenel P Schneider J Seethaler RVergnaud JC Sommer H 2000 Public-health impact of outdoor and traffic-related air pollution a European assessment Lancet 356 (9232) 795e801

Lawrence MG et al 2003 Global chemical weather forecasts for field campaignplanning predictions and observations of large-scale features during MINOSCONTRACE and INDOEX Atmos Chem Phys 3 267e289

Lee AM Carver GD Chipperfield MP Pyle JA 1997 Three dimensionalchemical forecasting a methodology J Geophys Res 102 3905e3919

Lee P Kang D McQueen J Tsisulko M Hart M DiMego G Seaman NDavidson P 2008 Impact of domain size on modeled ozone forecast for theNortheastern United states J Appl Meteorol Climatol 47 443e461 httpdxdoiorg1011752007JAMC14081

Li M Hassan R 2010 Urban air pollution forecasting using artificial intelligence-based tools In Villanyi Vanda (Ed) Chpt 9 Air Pollution In Tech ISBN978-953-307-143-5 pp 195e219

Li G Zavala M Lei W Tsimpidi AP Karydis VA Pandis SN Canagaratna MRMolina LT 2011 Simulations of organic aerosol concentrations in Mexico Cityusing the WRF-CHEM model during the MCMA-2006MILAGRO campaignAtmos Chem Phys 11 3789e3809

Luecken DJ Phillips S Sarwar G Jang C 2008 Effects of using the CB05 vsSAPRC-99 vs CB4 chemical mechanism on model predictions ozone andgas-phase precursor concentrations Atmos Environ 42 5805e5820

Mallet V Sportisse B 2006 Ensemble-based air quality forecasts a multimodelapproach applied to ozone J Geophys Res 111 D18302 httpdxdoiorg1010292005JD006675

Mallet V Queacutelo D Sportisse B Ahmed de Biasi M Debry Eacute Korsakissok IWu L Roustan Y Sartelet K Tombette M Foudhil H 2007 Technical notethe air quality modeling system Polyphemus Atmos Chem Phys 7 (20)5479e5487

Mallet V 2010 Ensemble forecast of analyses coupling data assimilation andsequential aggregation J Geophys Res 115 D24303 httpdxdoiorg1010292010JD014259

Manders AMM Schaap M Hoogerbrugge R 2009 Testing the capability of thechemistry transport model LOTOS-EUROS to forecast PM10 levels in theNetherlands Atmos Environ 43 4050e4059 httpdxdoiorg101016jatmosenv200905006

Mangold A De Backer H De Paepe B Dewitte S Chiapello I Derimian YKacenelenbogen M Leacuteon J-F Huneeus N Schulz M Ceburnis DOrsquoDowd C Flentje H Kinne S Benedetti A Morcrette JJ Boucher O 2011Aerosol analysis and forecast in the European Centre for medium-rangeweather forecasts integrated forecast system 3 Evaluation by means of casestudies J Geophys Res 116 D03302

Manins PC Cope ME Hess GD Nelson PF Puri K Wong N Young M 2002The Australian air quality forecasting system prognostic air quality forecastingin Australia Clean Air and Environ Quality 36 (2) 43e48

Manins P October 1999 Air Quality Forecasting for Australiarsquos Major Cities 1stProgress Report SB1407 25 CSIRO Atmospheric Research AspendaleAustralia httpwwwdarcsiroaupublicationsManins_1999apdf

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

Matanoski GM Tao X 2002 Case-cohort Study of Styrene Exposure and IschemicHeart Disease Research ReportHealth Effects Institute 108 pp 1e29

Mathur R Yu S Kang D Schere KL 2008 Assessment of the wintertimeperformance of particulate matter forecasts with the Eta-Community Multi-scale Air Quality modeling system J Geophys Res 113 D02303 httpdxdoiorg1010292007JD008580

McCollister G Wilson K 1975 Linear stochastic models for forecasting dailymaxima and hourly concentrations of air pollutants Atmos Environ 9417e423

McHenry JN Ryan WF Seaman NL Coats Jr CJ Pudykiewics JArunachalam S Vukovich JM 2004 A real-time Eulerian photochemicalmodel forecast system overview and initial ozone forecast performance in theNortheast US corridor Bull Amer Meteor Soc 85 525e548

McKeen S Wilczak J Grell G Djalova I Peckham S Hsie E-Y Gong WBouchet V Meacutenard S Moffet R McHenry J McQueen J Tang YCarmichael GR Pagowski M Chan A Dye t Frost G Lee P Mathur R2005 Assessment of an ensemble of seven real-time ozone forecasts overeastern North America during the summer of 2004 J Geophys Res 110D21307 httpdxdoiorg1010292005JD005858

McKeen S Chung SH Wilczak J Grell G Djalalova I Peckham S Gong WBouchet V Moffet R Tang Y Carmichael GR Mathur R Yu S 2007Evaluation of several PM25 forecast models using data collected during theICARTTNEAQS 2004 field study J Geophys Res 112 D10S20 httpdxdoiorg1010292006JD007608

McKeen S et al 2009 An evaluation of real-time air quality forecasts and theirurban emissions over eastern Texas during the summer of 2006 Second TexasAir Quality Study field study J Geophys Res 114 D00F11 httpdxdoiorg1010292008JD011697

Menut L Bessagnet B 2010 Atmospheric composition forecasting in Europe AnnGeophys 28 61e74 wwwann-geophysnet28612010

Menut L Chiapello I Moulin C 2009 Previsibility of mineral dust concentra-tions the CHIMERE-DUST forecast during the first AMMA experiment dryseason J Geophys Res 114 D07202 httpdxdoiorg1010292008JD010523

Merikanto J Napari I Vehkamaumlki H Anttila T Kulmala M 2007 New param-eterization of sulfuric acid-ammonia-water ternary nucleation rates at tropo-spheric conditions J Geophys Res 112 D15207 httpdxdoiorg1010292006JD007977

Michalakes J Chen S Dudhia J Hart L Klemp J Middlecoff J Skamarock W2001 Development of a next-generation regional weather forecast modelDevel In Zwieflhofer Walter Kreitz Norbert (Eds) Teracomputing Proceed-ings of the Ninth ECMWF Workshop on the Use of High PerformanceComputing in Meteorology World Scientific Singapore pp 269e276

Millman A Tang DL Perera FP 2008 Air pollution threatens the health ofchildren in China Pediatrics 122 620e628

Misenis C Zhang Y 2010 An examination of WRFChem physical parameteri-zations nesting options and grid resolutions Atmos Res 97 315e334

Mitsakou C Kallos G Papantoniou N Spyrou C Solomos S Astitha MHousiadas C 2008 Saharan dust levels in Greece and received inhalationdoses Atmos Chem Phys 8 7181e7192

Neary L Kaminski JW Lupu A McConnell JC 2007 Developments and Resultsfrom a Global Multiscale Air Quality Model (GEM-AQ) In Air PollutionModeling and Its Application XVII vol 5 httpdxdoiorg101007978-0-387-68854-1_44 pp 403e410

Niemeyer LE 1960 Forecasting air pollution potential Mon Wee Rev 88 (3)88e96

Niu T Gong SL Zhu GF Liu HL Hu XQ Zhao CH Wang YQ 2008 Dataassimilation of dust aerosol observations for the CUACEdust forecastingsystem Atmos Chem Phys 8 3473e3482

Ohara T Fujita A Kizu T Okamoto S 1997 Advanced air quality forecastingsystem for Chiba Prefecture Proceedings of the 22nd NATOCCMS InternationalTechnical Meeting on Air Pollution Modelling and its Application Clermont-Ferrand France 2e6 June

Otte TL Pouliot G Pleim JE Young JO Schere KL Wong DC Lee PCSTsidulko M McQueen JT Davidson P Mathur R Chuang H-Y DiMego GSeaman NL 2005 NCEP Notes linking the Eta model with the communitymultiscale air quality (CMAQ) modeling system to build a national air qualityforecasting system Weather Forecast 20 367e384

Pagowski M Grell GA 2006 Ensemble-based ozone forecasts skill andeconomic value J Geophys Res 111 D23S30 httpdxdoiorg1010292006JD007124

Peacuterez P Reyes J 2006 An integrated neural network model for PM10 forecastingAtmos Environ 40 2845e2851 httpdxdoiorg101016jatmosenv200601010

Peters LK Berkowitz CM Carmichael GR Easter RC Fairweather G Ghan SJHales JM Ruby Leung L Pennell WR Potra FA Saylor RD Tsang TT1995 The current state and future direction of Eulerian models in simulatingthe tropospheric chemistry and transport of trace species a review AtmosEnviron 29 189e222

Phalen RF Phalen RN 2011 Introduction to Air Pollution Science ISBN 100763780448 Jones amp Bartlett Learning

Prybutok VR Yi J Mitchell D 2000 Comparison of neural network models withARIMA and regression models for prediction of Houstonrsquos daily maximumozone concentrations Eur J Oper Res 122 31e40

Pudykiewicz JA Koziol AS 2001 The application of Eulerian models for airquality prediction and the evaluation of emission control strategies in CanadaInt J Environ Pollut 16 425e438

recasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e24 23

Pudykiewicz JA et al 2003 Operational Air Quality Forecasting in CanadaNumerical Model Guidance for Ground-level Ozone and Particulate MatterPreprints Fifth Conf on Atmospheric Chemistry Gases Aerosols and CloudsLong Beach CA Amer Meteor Soc the 83rd Annual Meeting CD-ROM 32

Rasch PJ Mahowald NM Eaton BE 1997 Representations of transportconvection and the hydrologic cycle in chemical transport models Implica-tions for the modeling of short-lived and soluble species J Geophys Res 10228127e28138

Robertson L 2010 MATCH Products Quality and Background Information ReportPrepared for MACC Report D_R-ENS_152 March

Roeckner E Brokopf R Esch M Giorgetta M Hagemann S Kornblueh LManzini E Schlese U Schulzweida U 2006 Sensitivity of simulated climateto horizontal and vertical resolution in the ECHAM5 Atmosphere ModelJ Climate 19 3771e3791

Rouiumll L Honoreacute C Vautard R Beekmann M Bssagnet B Malherbe LMeleux F Dufour A Elichegaray C Flaud J-M Menut L Martin DPeuch A Peuch V-H Poisson N 2009 Prevrsquoair an operational forecastingand mapping system for air quality in Europe Bull Amer Meteor Soc 9073e83 httpdxdoiorg1011752008BAMS23901

Rufeger W Mieth P Lux T 1997 Applying the DYMOS System in ConurbationsProc Modsim 97 4 TAS Australia International Congress on Modelling andSimulation Hobart pp 1797e1801

Russell A Dennis R 2000 NARSTO critical review of photochemical models andmodeling Atmos Environ 34 2283e2324

Ryan WF Petty CA Luebehusen ED 2000 Air quality forecasts in the mid-Atlantic region current practice and benchmark skill Weather Forecast 1546e60

Ryan WF 1995 Forecasting ozone episodes in the Baltimore metropolitan areaAtmos Environ 29 (17) 2387e2398

San Joseacute R Peacuterez JL Gonzaacutelez RM 2006 CFD and Mesoscale Air QualityApplication in Urban Environment MADRID Case Study Proceedings of the 4thWSEAS International Conference on Environment Ecosystems and Develop-ment (EED rsquo06) Venice Italy November 20e22 2006

San Joseacute R Peacuterez JL Morant JL Gonzaacutelez Barras RM 2009 The use of Modernthird-generation air quality models (MM5-EMIMO-CMAQ) for real-time oper-ational air quality impact assessment of industrial plants Water Air Soil PolluFocus 9 (1e2) 27e37 httpdxdoiorg101007s11267-008-9196-4

Sartelet KN Debry E Fahey K Roustan Y Tombette M Sportisse B 2007Simulation of aerosols and gas-phase species over Europe with the Polyphemussystem Part I-Model-to-data comparison for 2001 Atmos Environ 416116e6131 httpdxdoiorg101016jatmosenv200704024

Schaap M Timmermans RMA Sauter FJ Roemer M Velders GJMBoersen GAC Beck JP Builtjes PJH 2008 The LOTOS-EUROS model descrip-tion validation and latest developments Int J Environ Pollut 32 (2) 270e289

Schere K Bouchet V Grell G McHenry J McKeen S 2006 The Emergence ofNumerical Air Quality Forecasting Models and Their Applications presentationat The AMS Forum on Managing Our Physical and Natural Resources February2 Atlanta GA

Schwartz J 1991 Particulate air pollution and daily mortality in Detroit EnvironRes 56 204e213

Seigneur C et al 2000 Guidance for the performance evaluation of three-dimensional air quality modeling systems for particulate matter and visibilityJ Air Waste Manage Assoc 50 588e599

Seigneur C 1994 The status of mesoscale air quality models In Solomon P (Ed)Planning and Managing Air Quality Modeling and Measurement Studiesa Perspective Through SJVAQSAUSPEX Lewis Publishers Boca Raton FloridaUSA pp 403e433 (Chapter 3-2)

Seigneur C 2001 Current status of air quality modeling for particulate matter J AirWaste Manage Assoc 51 1508e1821

Seinfeld JH Pandis SN 2006 Atmospheric Chemistry and Physics From AirPollution to Climate Change John Wiley amp Sons Inc ISBN 978-0-471-72018-8

Shad R Mesgari MS Abkar A Shad A 2009 Predicting air pollution using fuzzygenetic linear membership kriging in GIS Comput Environ Urban 33 472e481

Shrivastava M Fast J Easter R Gustafson Jr WI Zaveri RA Jimenez JLSaide P Hodzic A 2010 Modeling organic aerosols in a megacity comparisonof simple and complex representations of the volatility basis set approachAtmos Chem Phys Discuss 10 30205e30277

Sihto S Kulmala M Kerminen V Dal Maso M Petaumljauml T Riipinen IKorhonen H Arnold F Janson R Boy M 2006 Atmospheric sulphuric acidand aerosol formation Implications from atmospheric measurements fornucleation and early growth mechanisms Atmos Chem Phys 6 4079e4091httpdxdoiorg105194acp-6-4079-2006

Sofiev M Vira J 2010 SILAM Products Quality and Background InformationReport Prepared for MACC Report D_R-ENS_192 March

Sofiev M Siljamo P Ranta H Rantio-Lehtimaumlki A 2006 Towards numericalforecasting of long-range air transport of birch pollen theoretical consider-ations and a feasibility study Int J Biometeorol 50 392e402

Spyrou C Mitsakou C Kallos G Louka P Vlastou G 2010 An improved limitedarea model for describing the dust cycle in the atmosphere J Geophys Res 115D17211 httpdxdoiorg1010292009JD013682

Stadlober E Houmlrmann S Pfeiler B 2008 Quality and performance of a PM10 dailyforecasting model Atmos Environ 42 1098e1109 httpdxdoiorg101016jatmosenv200710073

Stockwell WR et al 2002 The scientific basis of NOAArsquos air quality forecastprogram Environ Manage December 20e27

Please cite this article in press as Zhang Y et al Real-time air quality forEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

Stohl A Forster C Frank A Seibert P Wotawa G 2005 Technical note theLagrangian particle dispersion model FLEXPART version 62 Atmos Chem Phys5 2461e2474

Takigawa M Niwano M Akimoto H Takahashi M 2007 Development of a one-way nested global-regional air quality forecasting model SOLA 3 81e84 httpdxdoiorg102151sola2007e021

Talbot D 2007 Development of a New Canadian Operational Air Quality Fore-cast Model GEM-MACH paper presented at the International TechnicalMeeting on Air Pollution Modelling and its Application Aveiro-Portugal24e28 September

Tie X Geng FH Peng L Gao W Zhao CS 2009 Measurement and modeling ofO3 variability in Shanghai China application of the WRF-Chem model AtmosEnviron 43 4289e4302

Tilmes S et al 2002 Comparison of five Eulerian air pollution forecasting systemsfor the summer of 1999 using the German ozone monitoring data J AtmosChem 42 91e121

US (EPA) 1996 Air Quality Criteria for Particulate Matter Research Triangle ParkNC Office of Research and Development Office of Health and EnvironmentalAssessment EPA report no EPA600P-95001aF

US EPA 1999 Guideline for Developing an Ozone Forecasting Program USEnvironmental Protection Agency Rep EPA-454R-99e009 88 pp Availablefrom Office of Air Quality Planning and Standards EPA Research Triangle ParkNC 27711

US EPA 2000 Air Quality Index a Guide to Air Quality and Your Health EPA-454R-00-005 US Environmental Protection Agency Office of Air and RadiationWashington DC httpwwwepagovairnowaqibroch

US EPA 2001 Guidance for Demonstrating Attainment of Air Quality Goals forPM25 and Regional Haze Draft 21 January 2 2001 The US EnvironmentalProtection Agency Office of Air and RadiationOffice of Air Quality Planning andStandards Research Triangle Park NC 27711

US EPA 2003 Guideline for Developing an Air Quality (Ozone and PM25) Fore-casting Program US Environmental Protection Agency Rep EPA-456R-03-002126 pp Available from Office of Air Quality Planning and Standards EPAResearch Triangle Park NC 27711 Also Available Online at httpwwwepagovairnowaq_forecasting_guidance-1016pdf

US EPA 2009 Technical Assistance Document for Reporting of Daily Air QualitydAirQuality Index (AQI) EPA-454B-09-001 US Environmental Protection AgencyOffice of Air Quality Planning and Standards Research Triangle Park NC

Uno I et al 2003 Regional chemical weather forecasting system CFORS modeldescriptions and analysis of surface observations at Japanese island stationsduring the ACE-Asia experiment J Geophys Res 108 8668 httpdxdoiorg1010292002JD002845

Valdebenito B Benedictow AacuteM AC 2010 EMEP Products Quality and Back-ground Information Report Prepared for MACC Report D_R-ENS_122 March

van der Wal JT Jansen LHJM 2000 Analysis of spatial and temporal variationsof PM10 concentrations in the Netherlands using Kalman filtering AtmosEnviron 34 3675e3687

Vaughan J et al 2004 A numerical daily air quality forecast system for the PacificNorthwest Bull Amer Meteor Soc 85 549e561

Vautard R Blond N Schmidt H Derognat C Beekmann M 2001a Multi-modelensemble ozone forecasts over Europe analysis of uncertainty In Brandt J(Ed) Mesoscale Transport of Air Pollution OA15 EGS XXXVI General AssemblyNice European Geophysical Society Katlenburg-Lindau Germany France 26March

Vautard R Beekmann M Roux J Gombert D 2001b Validation of a determin-istic forecasting system for the ozone concentrations over the Paris area AtmosEnviron 35 2449e2461

Wang Z-F Wu Q-Z Gbaguidi A Yan P-Z Zhang W Wang W Tang X 2009Ensemble air quality multi-model forecast system for Beijing (EMS-Beijing)model description and preliminary application J of NUIST Natural ScienceEdition 1 (1) 19e26 (in Chinese)

Wang X Mallet V Berroir J-P Herlin I 2011 Assimilation of OMI NO2 retrievalsinto a regional chemistry-transport model for improving air quality forecastsover Europe Atmos Environ 45 485e492

Wayland RA White JE Dickerson PG Dye TS 28-36 December 2002Communicating real-time and forecasted air quality to the public EnvironManage

Wilby RL Wigley TML 1997 Downscaling general circulation model outputa review of methods and limitations Prog Phys Geog 21 530e548

Wilczak J McKeen S Djalalova I Grell G Peckam S Gong W Bouchet VMoffet R McHenry J Mc-Queen J Lee P Tang Y Carmichael GR 2006Bias-corrected ensemble and probabilistic forecasts of surface ozone overeastern North America during the summer of 2004 J Geophys Res 111D23S28 httpdxdoiorg1010292006JD007598

Willmott CJ 1981 On the validation of models Phys Geogr 2 184e194Wilson LJ Valeacutee M 2003 The Canadian Updateable Model Output

Statistics (UMOS) system validation against perfect prog Weather Forecast 18288e302

Wolff GT Lioy PJ 1978 An empirical model for forecasting maximum dailyozone levels in the northeastern United States J Air Pollut Control Assoc 281034e1038

World Health Organization (WHO) 2004 Health Aspects of Air Pollution Resultsfrom the WHO Project ldquoSystematic Review of Health Aspects of Air Pollution inEurope rdquo June E83080 World Health Organization Regional Office for EuropeScherfigsvej 8 DK-2100 Copenhagen Oslash Denmark

ecasting part I History techniques and current status Atmospheric

Y Zhang et al Atmospheric Environment xxx (2012) 1e2424

World Health Organization (WHO) 2010 WHO Guidelines for Indoor Air QualitySelected Pollutants World Health Organization egional Office for EuropeScherfigsvej 8 DK-2100 Copenhagen Oslash Denmark ISBN 9789289002134

Wu L Mallet V Bocquet M Sportisse B 2008 A comparison study of dataassimilation algorithms for ozone forecasts J Geophys Res 113 D20310 httpdxdoiorg1010292008JD009991

Yu SC Eder B Dennis R Chu S-H Schwartz S 2006 New unbiased symmetricmetrics for evaluation of air quality models Atmos Sci Lett 7 26e34

Yu SC Mathur R Kang D Schere K Pleim J Otte TL 2007 A detailed eval-uation of the Eta-CMAQ forecast model performance for O3 its relatedprecursors and meteorological parameters during the 2004 ICARTT studyJ Geophys Res 112 D12S14 httpdxdoiorg1010292006JD007715

Yu S Mathur R Schere K Kang D Pleim J Young J Tong D Pouliot GMcKeen SA Rao ST 2008 Evaluation of real-time PM25 forecasts andprocess analysis for PM25 formation over the eastern United States using theEta-CMAQ forecast model during the 2004 ICARTT study J Geophys Res 113D06204 httpdxdoiorg1010292007JD009226

Yu F 2010 Ion-mediated nucleation in the atmosphere key controlling parame-ters implications and look-up table J Geophys Res 115 D03206 httpdxdoiorg1010292009JD012630

Zhang Y Liu P Pun B Seigneur C 2006a A comprehensive performance eval-uation of MM5-CMAQ for the summer 1999 southern oxidants study episodepart-I Evaluation protocols databases and meteorological predictions AtmosEnviron 40 4825e4838

Zhang Y Liu P Queen A Misenis C Pun B Seigneur C Wu S-Y 2006bA comprehensive performance evaluation of MM5-CMAQ for the summer 1999southern oxidants study episode Part-II Gas and aerosol predictions AtmosEnviron 40 4839e4855

Zhang K Wan H Wang B Zhang M Feichter J Liu X 2010 Troposphericaerosol size distributions simulated by three online global aerosol models using

Please cite this article in press as Zhang Y et al Real-time air quality foEnvironment (2012) httpdxdoiorg101016jatmosenv201206031

the M7 microphysics module Atmos Chem Phys 10 6409e6434 httpdxdoiorg105194acp-10-6409-2010

Zhang Y Pan Y Wang K Fast JD Grell GA 2010a WRFChem-MADRIDincorporation of an aerosol module into WRFChem and its initial applicationto the TexAQS2000 episode J Geophys Res 115 D18202 httpdxdoiorg1010292009JD013443

Zhang Y Liu P Liu X-H Pun B Seigneur C Jacobson MZ Wang W-X 2010bFine scale modeling of wintertime aerosol mass number and size distributionsin Central California J Geophys Res 115 D15207 httpdxdoiorg1010292009JD012950

Zhang Y Wen X-Y Jang CJ 2010c Simulating climate-chemistry-aerosol-cloud-radiation feedbacks in continental US using online-coupled WRFChemAtmos Environ 44 (29) 3568e3582

Zhang Y Chen Y-S Zhu S Wu S-Y Sartelet K Tran P Seigneur C 2011Application of WRFChem-MADRID in Europe Model Evaluation and Aerosol-Meteorology Interactions Presentation at the EGU General Assembly 2011Vienna Austria 3e8 April

Zhang Y Karamchandani P Glotfelty T Streets DG Grell G Nenes A Yu F-QBennartz R 2012 Development and Initial Application of the Global-Through-Urban Weather Research and Forecasting Model with Chemistry (GU-WRFChem) J Geophys Res in review

Zhang Y Chen Y-C Sarwar G Schere K 2012 Impact of gas-phasemechanisms on WRFChem predictions mechanism implementation andcomparative evaluation J Geophys Res 117 D1 httpdxdoiorg1010292011JD015775

Zhang Y 2008 Online coupled meteorology and chemistry models historycurrent status and outlook Atmos Chem Phys 8 2895e2932

Zolghadri A Cazaurang F 2006 Adaptive nonlinear state-space modelling for theprediction of daily mean PM10 concentrations Environ Model Software 21885e894 httpdxdoiorg101016jenvsoft200504008

recasting part I History techniques and current status Atmospheric

• Real-time air quality forecasting part I History techniques and current status
• • 1 Introduction
  • • 11 Importance and type of applications of air quality forecasting
    • 12 History and current status
    • 13 Fundamentals of real-time air quality forecasting
    • • 131 Major characteristics of RT-AQF
      • 132 Forecasting products
      • • 2 Techniques and tools of RT-AQF and their advantageslimitations
        • • 21 Simple empirical approaches
          • 22 Parametricnon-parametric statistical approaches
          • 23 Advanced physically-based approaches
          • • 3 Overview of major 3-D RT-AQF models on regional and global scales
            • • 31 Model systems components and application scales
              • 32 Model input output and horizontal grid resolution
              • 33 Chemistry and aerosol treatments
              • • 4 Evaluation protocols and model performance of 3-D RT-AQF models
                • • 41 Evaluation protocols
                  • 42 Observational datasets for model evaluation
                  • 43 Current model forecasting skills
                  • • 5 Summary of Current States of RT-AQF
                    • Acknowledgments
                    • Appendix A Supplementary Material
                    • References