thesis ann

Upload: taridan

Post on 29-Oct-2015

29 views

Category:

Documents


0 download

TRANSCRIPT

  • A Hybrid Computational Model for

    Building Systems Control

    Seongju ChangSchool of Architecture

    Carnegie Mellon University

    Ph.D. Committee

    Prof. Ardeshir Mahdavi, Ph.D.(Chair)Prof. Volker Hartkopf, Ph.D.Prof. Sebastian Thrun, Ph.D.Prof. Bernd Bruegge, Ph.D.

  • i

    Table of contents

    Table of contents..................................................................................................................................................iList of Figures.....................................................................................................................................................iiList of Tables......................................................................................................................................................viAcknowledgment................................................................................................................................................vii

    Copyright Declaration......................................................................................................................................viii

    Abstract...............................................................................................................................................................ix

    1 Introduction........................................................................................................................1

    1.1 Motivation............................................................................................................................................1

    1.2 Thesis goals & hypotheses...................................................................................................................3

    1.3 Thesis outline.......................................................................................................................................4

    2 Background.........................................................................................................................6

    2.1 Building process and its control...........................................................................................................6

    2.1.1 Overview..............................................................................................................................6

    2.1.2 Daylight and electrical lighting process ..............................................................................7

    2.1.3 New perspective on building control process ......................................................................9

    2.2 Models for building process control ..................................................................................................10

    2.2.1 Simulation model...............................................................................................................10

    2.2.2 Machine learning model ....................................................................................................16

    3 HISSTO: a hybrid system for environmental control ..................................................21

    3.1 Problem identification........................................................................................................................21

    3.1.1 Issues in building control models ......................................................................................21

    3.2 Requirement elicitation......................................................................................................................24

    3.2.1 Functional requirements ....................................................................................................24

    3.2.2 Non-Functional requirements ............................................................................................30

    3.3 System analysis..................................................................................................................................32

    3.3.1 Analysis of hybridization...................................................................................................32

    3.3.2 System analysis object model ............................................................................................343.3.3 System sequence model .....................................................................................................35

    3.3.4 User interface mock-up......................................................................................................37

    3.4 System design ....................................................................................................................................39

    3.4.1 System architecture............................................................................................................39

    3.4.2 System component design .................................................................................................40

    3.4.3 Concurrency identification ................................................................................................52

  • ii

    3.4.4 Hardware/software mapping..............................................................................................52

    3.4.5 Data management ..............................................................................................................54

    3.4.6 Global resources management ...........................................................................................55

    3.4.7 Software control implementation ......................................................................................56

    3.4.8 Boundary conditions ..........................................................................................................60

    3.4.9 Adaptation to the system changes......................................................................................62

    3.4.10 Design priorities.................................................................................................................63

    3.4.11 Design trade-off .................................................................................................................63

    3.5 Object design .....................................................................................................................................643.5.1 Object design of the Predictor ...........................................................................................643.5.2 Object design of the Tester ................................................................................................833.5.3 Object design of the Commander ......................................................................................883.5.4 Object design of the Sensor ...............................................................................................893.5.5 Object design of the Actuator ............................................................................................903.5.6 Object design of the daylight responsive lighting control .................................................90

    4 System test & evaluation .................................................................................................96

    4.1 Test bay & control target systems......................................................................................................96

    4.2 Data acquisition for system test .........................................................................................................99

    4.2.1 Analysis of the measurement...........................................................................................100

    4.2.2 Data preparation for testing the predictors ......................................................................102

    4.3 Performance test of the predictors ...................................................................................................104

    4.3.1 Prediction performance of the LUMINA.........................................................................104

    4.3.2 Prediction performances of the hybrid predictors............................................................106

    4.4 Test of predictive control scenarios .................................................................................................121

    4.4.1 Test of simulation-based control......................................................................................121

    4.4.2 Test of hybrid predictor-based control .............................................................................124

    4.4.3 Test of the convergence time to reach a control action decision .....................................126

    4.5 Overall evaluation of HISSTO ........................................................................................................128

    5 Conclusion.......................................................................................................................130

    5.1 Contributions ...................................................................................................................................130

    5.2 Future Research ...............................................................................................................................131

    6 References .......................................................................................................................133

    Appendix A: HISSTOs test instrumentation

    Appendix B: HISSTOs control performance evaluation

  • iii

    List of Figures

    Figure 2.1: Daylight and electrical light process in a space ................................................................................9

    Figure 2.2:Feed-forward (a) and feed-back (b) control process........................................................................12Figure 2.3: Feed-forward control process example ...........................................................................................12

    Figure 2.4:: Use of BDI and GAT for simulation-assisted building control (Mahadavi, 1997)........................13Figure 3.1:Use case diagram of HISSTO: UML notation .................................................................................30

    Figure 3.2: Structure of a virtual building and its relationship with other entities............................................32

    Figure 3.3:Typical hybridization schemes between a simulator and a machine learner ...................................34

    Figure 3.4: System analysis class diagram of HISSTO: UML notation............................................................35

    Figure 3.5:Sequence diagram illustrating interactions among the key objects in HISSTO: UML notation .....36Figure 3.6: HISSTOs Monitoring Window on the web....................................................................................37

    Figure 3.7: HISSTOs Control Window on the web..........................................................................................38

    Figure 3.8:Component diagram describing HISSTOs system architecture: UML notation ............................40

    Figure 3.9:Top level logical model describing HISSTOs subsystem breakdown: UML notation ...................41

    Figure 3.10:Logical model of HISSTOs SystemManager component: UML notation....................................42

    Figure 3.11:Logical model of HISSTOs EnvironmentDescriptor component: UML notation ........................43

    Figure 3.12:Logical model of HISSTOs MonitoringManager component: UML notation .............................44

    Figure 3.13:Logical model of HISSTOs EventManager component: UML notation ......................................45

    Figure 3.14:Sequence diagram showing HISSTOs event handling process: UML notation............................46

    Figure 3.15:Logical model of HISSTOs PredictionManager component: UML notation ...............................47

    Figure 3.16:Logical model of HISSTOs ControlManager component: UML notation ...................................48

    Figure 3.17:Logical model of HISSTOs DatabaseManager component: UML notation .................................49

    Figure 3.18:Logical model of HISSTOs CommunicationManager component: UML notation......................50

    Figure 3.19:Logical model of HISSTOs UserInterfaceManager component: UML notation..........................51

    Figure 3.20:Deployment diagram describing HISSTOs hardware topology: UML notation...........................54

    Figure 3.21: Physical connectivity for HISSTOs system instrumentation .......................................................54

    Figure 3.22: Access privilege management in HISSTO: UML notation...........................................................55

    Figure 3.23: HISSTOs event notification process based on the Observer design pattern: UML notation.......57

    Figure 3.24: HISSTOs alternative control strategies based on the Strategy design pattern: UML notation ....58

    Figure 3.25:State transition diagram describing HISSTOs control mode changes: UML notation .................59

    Figure 3.26: HISSTOs handling of control command based on the Command design pattern: UML notation60

    Figure 3.27: Construction of a View based on the Builder design pattern: UML notation ...............................62

  • iv

    List of Figures

    Figure 3.28: Viewer information necessary for LUMINAs glare calculation ..................................................65

    Figure 3.29:Iconic representation scheme of the neural networks designed for HISSTO.................................71

    Figure 3.30: Generic architecture of the neural networks design in HISSTO...................................................73

    Figure 3.31: Different neural network training schemes based on the progression of time ..............................75

    Figure 3.32:Collaboration diagram describing SM12CNNs training process: UML notation.........................79

    Figure 3.33:Collaboration diagram describing SM12CNNs prediction process: UML notation.....................80

    Figure 3.34:Exemplary collaborative process between WFNN(2) and BRIDGENN in WFC1NN..................81Figure 3.35: Training and operation processes of the different hybrid predictors in HISSTO..........................83

    Figure 3.36: Exemplary preference functions for HISSTOs performance indicators ......................................87

    Figure 3.37:Activity diagram of HISSTOs daylight responsive lighting control process: UML notation........92

    Figure 4.1: Intelligent Workplace at Carnegie Mellon University as the system test bed ................................96

    Figure 4.2: HISSTO system test bay at the Intelligent Workplace....................................................................97

    Figure 4.3: Illustrative photos showing exterior and interior of the test bay ....................................................97

    Figure 4.4: Plan and elevation view of the test bay with louver, luminaires, and indoor sensors.....................98

    Figure 4.5: Daylight station on the top of the IW measuring sky condition parameters...................................99

    Figure 4.6:Sky illuminance and irradiance measurements on a typical clear day (3/13/98)...........................100Figure 4.7:Sky illuminance measurement showing daily and seasonal variations .........................................100

    Figure 4.8:Indoor illuminance profile on all reference points influenced by the louver angle changes .........101

    Figure 4.9:Indoor illuminance profiles depending on the sensor positions and louver angles (3/13/98)........102Figure 4.10:LUMINAs indoor illuminance prediction performance in the test bay ......................................105

    Figure 4.11:Relative error of LUMINAs indoor illuminance prediction .......................................................106

    Figure 4.12:Histogram showing LUMINAs relative indoor illuminance prediction error distribution .........106

    Figure 4.13:Learning curves of the neural networks in HISSTO....................................................................107

    Figure 4.14:Different hybrid predictors indoor illuminance prediction accuracy expressed in RMS of relative

    errors (Test case 1with D1 and D2 data sets in Table 4.1)................................................................109Figure 4.15:Different hybrid predictors indoor illuminance prediction accuracy expressed in RMS of relative

    errors (Test case 2 with D4 and D5 data sets in Table 4.2)...............................................................110Figure 4.16: Different hybrid predictors performance in predicting both reference parameters and performance

    indicators represented as the averaged relative distances from the simulators prediction...............111

    Figure 4.17: Average relative distance and standard deviation of hybrid predictors reference parameters and

    preference indicators prediction from the simulations (with D1 and D2 data sets in Table 4.1)......113Figure 4.18: MS1NNs indoor illuminance prediction accuracy with the selected training sensor position....114

  • v

    List of Figures

    Figure 4.19:WFC1NNs indoor illuminance prediction performance (a) and RMS of the predicted performance indicators relative deviations from the simulations (b) with different calibration sensor points....115

    Figure 4.20: Indoor illuminance prediction decay curves of the hybrid predictors in HISSTO......................116

    Figure 4.21: Decay curves of the hybrid predictors predictions for the reference parameters and performance

    indicators expressed in terms of their relative distances from the simulation output......................117

    Figure 4.22: Change adaptation performance of each hybrid predictor in indoor illuminance prediction......118

    Figure 4.23: Change adaptation profile calculated as the closeness of each hybrid predictors reference

    parameters and performance indicators prediction to the simulation output ...................................119

    Figure 4.24:Overall evaluation of the hybrid predictors in a graphical form..................................................120

    Figure 4.25: Mapping between ranked louver positions and control quality index ........................................122

    Figure 4.26: Time requirements comparison needed for a single control time window .................................127

  • vi

    List of Tables

    Table 2.1: Factors influencing thermal and lighting process in a building..........................................................7

    Table 3.1: Examples of the challenges in energy simulation validation process ..............................................22

    Table 3.2:Visionary scenarios of HISSTO ........................................................................................................25

    Table 3.3:HISSTOs use cases...........................................................................................................................29

    Table 3.4: Design priorities of a typical building control system......................................................................63

    Table 3.5: Input and output parameters for LUMINA.......................................................................................66

    Table 3.6: Input and output parameters for the neural networks designed in HISSTO.....................................69

    Table 3.7:Selected input (I) and output (O) parameters for each neural network designed for HISSTO .........70Table 3.8:Design characteristics of the neural networks to construct HISSTOs predictors.............................72

    Table 3.9:Diagrammatic description of each hybrid predictors prediction process .........................................82

    Table 3.10: Various sensors used in HISSTO....................................................................................................89

    Table 3.11: Louver operation schedule example ...............................................................................................94

    Table 4.1:Test data sets filtered by the validation criteria (see A.2.1.1) .........................................................103

    Table 4.2: Test data sets within % deviation range from the Perez model prediction .............................103

    Table 4.3: Neural network training data sets generated by the weather file based simulations ......................103

    Table 4.4:Example of typical simulation validation methods .........................................................................104

    Table 4.5:Overall evaluation of the hybrid predictors designed in HISSTO...................................................120

    Table 4.6:Overall evaluation of the daylight control test ................................................................................122

    Table 4.7:Percentage of instances with a specific control quality index .........................................................123

    Table 4.8:Initial state for the inputs to the predictor........................................................................................123

    Table 4.9:10 best control actions suggested by the pure simulator .................................................................123

    Table 4.10:Order of the best louver positions suggested by the simulator......................................................124

    Table 4.11:10 best control actions suggested by the simulator + luminaire matrix predictor .........................124

    Table 4.12:Order of the best louver positions suggested by the MSC0NN predictor .....................................125

    Table 4.13:10 best control actions suggested by the MSC0NN predictor.......................................................125

    Table 4.14:Order of the best louver positions suggested by the WFC1NN predictor .....................................125

    Table 4.15:10 best control actions suggested by the WFC1NN predictor.......................................................126

    Table 4.16:Final systems control states (with utility values) suggested by the tested predictors....................126Table 4.17: Comparison of the required time for HISSTOs different controllers..........................................128

    Table 4.18:Assessment of the research outcome including HISSTO implementation....................................129

    20

  • vii

    Acknowledgment

    It has been a long journey through which I learned that life has many hidden layers to be discovered. Ifeel sincere gratitude to Professor Ardeshir Mahdavi who brought a great range of expertise and interest

    to my work. He has been an ever-present source of guidance and encouragement throughout my doc-

    toral program residency.

    I should like to thank Professor Volker Hartkopf for his visionary inspirations and support, letting me

    perform necessary experiments at the Intelligent Workplace. Surely, an essential part of contribution

    came from Professor Sebastian Thrun through his continuous enthusiasm, commenting and answering

    my machine learning related questions with unsurpassed clarity. Professor Bernd Bruegge has been pas-

    sionate and resourceful in bringing object-orientation and component technology into my thesis systemdevelopment effort. I treasure the time spent with him and his ambitious students at the software engi-

    neering laboratory.

    It would certainly have been much difficult if I had not have my colleagues around, lending me their

    shoulders to lean on, providing invaluable moments of friendship. Among those deserve being named, I

    want to thank Vineeta Pal for sparing her time to discuss the issues involved in her daylight simulation

    program with me. It is she who has been always willing to be victimized by my run-time humors even

    when I broke her no-more-than-ten-minutes rule on purpose. I also thank Jamieson Shulte at the

    School of Computer Science for his help in the instrumentation of my thesis experiments as well as the

    neural network implementation. To those named persons, I must add Rohini Brame, Paul Mathew, Rob-

    ert Ries, and Jay Shankavaram for their friendly help and knowledge exchanges with unexpected occa-

    sional laughters.

    I owe many thanks to my family members, especially to my mother whose unconditional support and

    trust made this thesis finally see the outside world. I dedicate this thesis to her with love and respect.

  • viii

    Copyright Declaration

    I hereby declare that I am the sole author of this thesis.

    I authorize Carnegie Mellon University, Pittsburgh, Pennsylvania to lend this thesis to other institutions

    or individuals for the purpose of scholarly research.

    I authorize Carnegie Mellon University, Pittsburgh, Pennsylvania to reproduce this thesis by photocopy-

    ing or by other means, in total or in part, at the request of other institutions or individuals for the pur-

    pose of scholarly research.

    Copyright 1999 by Seongju Changc

  • ix

    Abstract

    The increased complexity in building systems integration provides a new challenge for building opera-

    tion process. A wide distribution of the direct digital control approach has inspired researchers and engi-

    neers to pay more attention to the computational solutions in building control. The enhanced knowledge

    about building systems behaviors as well as the increased computing power make attractive the poten-

    tial use of a computational building model as the agent to carry on building control task. On the other

    hand, some simulation programs are computationally too intensive to be effective for real-time control

    purpose. A machine learner can address this problem. However, it often requires large amount of data

    for training or retraining, which makes the technique inevitably sensor-dependent. This thesis intends to

    demonstrate that analytical approaches (i.e. simulation) and inductive learning methods (i.e. neural net-works) can cooperate to facilitate building systems control. In this process, the role of a simulator isaugmented as the source of the system knowledge by which a supervised learner, implemented in neural

    networks, is trained for faster predictions. A machine learner, trained either to replace a simulator or to

    predict the simulation error, serves as the pivotal component for a better predictor through a hybridiza-

    tion process.

    This thesis has computationally implemented and tested seamless control scenarios based on the differ-

    ent hybridization schemes in order to identify desirable processes for such hybridization. The hybrid

    constructs designed to explore diverse synergistic effects have been tested in daylight responsive light-

    ing control domain. Multiple hybrid controllers are designed to meet four control goals: enriching the

    informational repertoire of systems control operations for lighting (by inclusion of performance indica-tors for glare and solar gain), reducing the number of sensing units necessary for capturing the states ofbuildings visual performance indicators in real time, enhancing the accuracy of predictions necessary

    for the identification of the best control option, and maximizing the searches in the lighting system con-

    trol state space within a limited time.

    HISSTO (Hybrid Intelligence for System State Transition Operation), the resulting pilot hybrid controlsystem is capable of regulating target lighting systems effectively by covering a wider portion of the

    control state space with increased agility and precision even in the environments subject to dynamicchanges. HISSTO demonstrates that the various modes of user-system interactions can be facilitated by

    using component-based object-oriented software engineering techniques along with internet-activatedcontrol and monitoring interfaces. HISSTOs software design pattern driven modular system architec-

    ture also allows it to be easily extended even to the other building control domains.

  • 11 Introduction

    1.1 Motivation

    Traditionally, building control systems have operated based on a homeostatic short-term feed back

    mechanism. For example, thermostatic control of HVAC components involves typical operations

    (on/off, change in volume and/or temperature of heating/cooling media, etc.) that are essentiallyguided by temperature sensing in a space. More recently, building control systems have become

    increasingly sophisticated. One of the approaches has been to utilize various methods and tools

    (including neural networks) to accurately capture a buildings dynamic characteristics so as to pro-vide a more reliable basis for the control of its behavior (Curtis 1996, Mistry and Nair 1993). In thisscenario, control options can be improved ("optimized"), as their past impact on the buildingsdynamic behavior is reflected in the collected information by the sensing system. Capturing build-

    ings dynamic behavior toward enhanced control strategies can be supported using advanced com-

    putational performance simulation routines. Generate-and-test as well as bi-directional inference

    methods can be used to derive preferable control schemes and required values for control variables

    based on parametric and iterative simulations (Mahdavi 1997b).

    A building used to be considered as a static and a passive physical construct. This view has been

    significantly challenged by the growing complexity of various building systems and components as

    well as the dynamic interactions among them. For example, because of the thermal inertia associ-

    ated with a building, the effect of external environmental changes on building interior conditions

    takes place over a time scale on the order of hours. These dynamics make control action take place

    on a time scale: on the order of seconds for actions involving local control loops, to a time scale on

    the order of minutes for changes in zone conditions (Kelly et al. 1984). Such dynamism makes abuilding systems control task a complex and sophisticated process. Most of the building processes

    can be quite accurately predicted using a set of well-established models, yet some of them still

    remain to be constantly monitored. This research starts from the idea that a well organized hybrid

    construct out of both simulation and machine learning technique can support the control of a

    dynamic building process such as the visual environment in a building.

    Developing a hybrid computational model for building control is the major focus of this thesis.There have been many hybridization efforts in control domain, but few of them, if any, addressed

  • 2the issue of combining model with machine learning techniques, especially for the visual environ-

    ment control in a building. At the same time, a well-defined evaluation scheme for different control

    system designs and implementation options is also important. Normally, the evaluation criteria of

    the visual environment in a building has been limited to very few quantitative indicators such as

    illuminance and uniformity. Enriching the performance evaluation criteria is, therefore, highly

    desired as long as the control system can deal with resulting complexity in both generation and

    assessment of control options. Identifying potential synergistic effects between simulation and

    machine learning for an enhanced building systems control can be attempted through a hybridiza-

    tion process. Hybridization of different models or components is not an entirely new approach,

    since there have been such attempts as neuro-fuzzy systems and neural network assisted PID con-

    trollers (Curtis 1996). For example, PID control algorithm using a feed-back loop with three math-ematical terms: proportional, integral, and derivative works with well-behaved linear, cumulative,

    and time-variant processes. But under non-linear conditions, this algorithm can yield unstable con-

    trol outputs unless it is tuned correctly. A neural network could serve for this tuning purpose. The

    major motivation for the hybridization between simulation and machine learning in this thesis liesin the importance of prior knowledge in learning process. It has been argued that prior knowledge

    can help significantly in learning (Mitchell 1997) and simulation can provide such knowledge to amachine learner. In this scheme, the role of building simulations can be augmented as the source of

    the system knowledge. Simulation can assist the identification of control-sensitive variables. The

    mapping between possible states of a system and their outcomes useful for the training or retraining

    of a machine learner can also be provided by simulations. Both data dependency and sensor depen-

    dency in neural network implementation can be dramatically reduced in this process. Beyond

    enhancing the effectiveness of dynamic control systems, the additional benefits of the suggested

    approach include:

    Prediction of the effects of the changes to building hardware and its control systems

    Beta-testing of building control system hardware based on the simulated actions in the context

    of a virtual building

    Pre-training of machine learning systems simulation data on building behaviors prior to their

    field utilization

    Retraining of machine learning systems to account for the effects of abrupt modifications to

    building configuration (i.e. retrofit) using simulation data

    Reduction of the number of sensing units necessary for capturing building's real-time opera-

    tional status

  • 3On the other hand, a machine learner can be utilized for on-line simulation calibration. This tuning

    process can improve the performance of simulation predictions. A machine learner can also copy

    the simulators knowledge so that it can reduce the time and computational load significantly in

    predicting the outcome of control actions after the training is completed. This benefit could lead to

    a major improvement in building control performance as long as the proper design and implemen-tation of the hybridization are achieved. By combining a simulation model and a machine learning

    technique such as neural networks, some of the critical issues in model-based control can be

    resolved. The expected benefits that simulation model can have when it is combined with machine

    learners are as follows:

    Calibration of simulation outcomes by compensating for the simulators prediction error

    Reducing computational load and time in predictions by copying and replacing a simulator

    In order to realize this idea, a conventional building automation system must be supplemented with

    a multi-aspect virtual model of the building that runs parallel to the building's actual operation.

    While the real building can only react to the actual contextual conditions (i.e. sky luminance distri-bution pattern, occupancy) and building control actions, the simulation-based virtual model allowsfor additional operations: i) the virtual model can move backward in time so as to analyze the build-ings past behavior and/or to calibrate the model toward improved predictive potency; ii) the virtualmodel can move forward in time so as to predict the buildings responses to the various alternative

    control scenarios (Mahdavi et al. 1999).

    Interactive and modular building control software development is the other driving force of

    this thesis. The primary motivation for the thesis system development effort is to be able to dynam-

    ically change the values of control parameters on-line while the system allows implementation-

    independence and flexible future extension.

    1.2 Thesis goals & hypotheses

    The scientific goal of this thesis is to verify the benefits of the hybridization between simulation

    and machine learning technique in building control domain. The engineering goal is to identify a

    prototypical building control software architecture allowing flexible adaptation to dynamic envi-

    ronmental and user preference changes. To summarize, the major goals of this thesis are to identify:

    TG1) Hybridization architecture between a simulator and a machine learner for high-performancebuilding control operations with increased speed and accuracy.

  • 4TG2) Well organized building control software architecture for enhanced accessibility, changeadaptability, and interoperability while being successfully integrated with the developed hybrid

    computational models

    The fulfillment of the scientific thesis goal (TG1) will be tested against the following hypotheses:

    TH1) Hybrid prediction model can increase the speed of convergence to a control decision.

    TH2) Hybrid prediction model can increase the accuracy of predictions for more precise control.

    TH3) Hybrid prediction model can decrease the sensor dependency of control operations whileenriching the performance evaluation criteria involved in control action decision process.

    TH4) Hybrid prediction model can provide better chance for the control system to adapt to buildingconfiguration or environmental changes.

    TH5) System flexibility can be maximized by using design patterns and component-based object-oriented programing methodology.

    The engineering thesis goal (TG2) will be tested against the systems use cases and non-functionalrequirements described in Chapter 3, system requirement elicitation section.

    1.3 Thesis outline

    This thesis involves a system development effort and the experiments for the developed system val-

    idation. As is stated, demonstrating the performance enhancements of the hybrid prediction models

    in building control domain is critical. For this reason, multiple hybridization schemes in the various

    forms of computational models are designed, implemented, and tested. These prediction models are

    combined with a testing algorithm which performs control options evaluation task to produce the

    most appropriate control decision. Different controllers based on the different hybrid prediction

    models are then evaluated in the visual environment control operation of an actual building. The

    performance of each controller is evaluated against measurement. Both the predictors and the con-

    trollers using those predictors are tested according to the defined evaluation criteria. The prototypi-

    cal control software architecture designed for HISSTO, the thesis control system, is engineered to

    provide all necessary functionalities, namely use cases, within the target control domain. The suc-

    cess of this prototype control software is evaluated based on those use cases described in the system

    requirement elicitation section. This thesis is progressed by the following steps:

  • 5Step 1) Outline of contextual knowledge: Chapter 2

    Step 2) Identification of the requirements for the prototype system development: Chapter 3

    Step 3) Identification and resolution of the prototype system design issues: Chapter 3

    Step 4) Detail design of the hybrid prediction models and other system components: Chapter 3

    Step 5) Instrumentation, planning, execution, and evaluation of the prototype system: Chapter 4

    Step 6) Validation of the thesis hypotheses and the system use cases: Chapter 4

    Step 7) Summary of the achievement and future work: Chapter 5

  • 62 Background

    2.1 Building process and its control

    2.1.1 Overview

    A building has diversified physical processes interacting with each other. The factors being

    impacted by such phenomena as lighting and thermal process in a building include both human

    comfort and energy implication. Daylight and electrical lighting can be complimentary in terms of

    satisfying the required indoor illuminance level. Electrical lighting increases energy consumption

    whereas daylight is basically free of charge. Both of them increase internal cooling load in summer

    and decrease heating load in winter. Building envelope components such as walls, windows, and

    shades interact with outside weather condition and influence indoor thermal condition. Occupants

    and equipment also have impact on thermal environment by introducing heat and moisture inside.

    A HVAC (Heating Ventilation & Air Conditioning) system controls temperature, humidity, and airflow which, in turn, influences PMV(Predicted Mean Vote) as the index of thermal comfort. AHVAC system consumes certain amount of energy based on the system type, operation mode and

    state. There have been significant amount of modeling efforts to describe each physical process in a

    building as well as the inter-dependencies among various such processes.

    From the control point of view, a building is a system which has multi-variant dynamic subsystems

    showing various linear or mostly non-linear behaviors. Both environmental and occupancy changes

    in a building increase the complexity of control operations. Occupants not only impose control

    goals, but also influence various building processes as is shown in Table 2.1. The building factors

    and the system factors are to be controlled to satisfy occupants requirements under the different

    states of the natural factors. The building factors are normally determined during design and con-

    struction phase of a building life cycle while the states of the system factors are practically decided

    in the buildings operation phase. Both the natural and the occupancy factors are dynamic but nor-

    mally not controllable, therefore, the building and the system factors need to be interactively

    adjusted to these factors.

    The control goal of the system factors is to minimize cost while maximizing occupants comfort

    (i.e. an adequate indoor illuminance level and an acceptable PMV (Predicted Mean Vote)). Build-

  • 7ing control normally focuses on the manipulations of the target systems to make up for the gap

    between the desired state and the current one. Because of the complex inter-dependency of all fac-

    tors involved in a building process, determining even a single control target system state requires a

    proper understanding of the interactions among all related factors (Table 2.1).

    2.1.2 Daylight and electrical lighting process

    Visual environment in a building is determined by numerous factors. Few performance factors have

    been used for selection of lighting equipment and its location throughout 1950s, 1960s, and early

    1970s (Flynn et al. 1992). Such performance criteria as illuminance level and energy consumptionare the typical example of those few factors. Figure 2.1 shows how light interacts with a built envi-

    ronment. For electrical lighting calculation, only direct and indirect internal components are to be

    considered. The sun is a single source of daylight. There are three components to be considered in

    daylight calculation:

    Sky component: The rays from the sun are partially absorbed and modified by the atmo-

    sphere. Incoming sunlight is also diffused by the clouds before reaching a building envelope.

    Modeling of this modified sunlight by the atmosphere and clouds is normally based on the tur-

    bidity applied to the entire sky. Angular relationship between the sun and the target space is

    another factor creating daily and seasonal variations of the indoor daylight profile. Solar azi-

    muth, solar altitude, and building orientation are, therefore, critical factors influencing this

    Table 2.1: Factors influencing thermal and lighting process in a building

    Classification Thermal process Lighting process

    Natural factors solar insolation, wind, outdoor tem-

    perature, humidity

    solar irradiance, sky condition, ground sur-

    face reflectance

    Occupancy factors number of occupants, activity type,

    occupant schedule, clothing

    number of occupants, occupants location,

    activity type, occupancy schedule

    Building factors building geometry, orientation, over-

    hang, glazing area, glazing material,

    wall construction, wall material, inte-

    rior finish, ceiling & floor materials,

    room volume, surface area

    site latitude & longitude, building geometry,

    orientation, overhang, shades, glazing area,

    glazing reflectance, glazing transmittance,

    interior finish, interior surface area, interior

    surface reflectance, interior surface transmit-

    tance

    System factors external shade, internal shade,

    HVAC system characteristics

    light redirection louver, blinds, luminaires

  • 8phenomenon.

    Externally reflected component: The sky component sometimes can directly hit the indoor

    reference point but part of it is either diffused or reflected by the external objects such asneighboring buildings, ground, or overhang before being introduced to the indoor space.

    Internally reflected component: Along with the externally reflected component, the inter-

    nally reflected light component is also important part of indoor illuminance calculation. Sur-

    face properties such as transmittance, absorption, reflectance as well as specularity are

    important to calculate the amount of light bouncing back from a surface. For more precise illu-

    minance calculation, the radiosity method is used for diffuse surfaces whereas the ray tracing is

    applied to specular surfaces.

    Average illuminance can be derived to represent the illuminance of a certain area having multiple

    illuminance reference points. Uniformity factor is another index to evaluate how evenly illumi-

    nance values are distributed throughout a space. Dimming is basically a control scheme to compen-

    sate daylight with electrical lighting so that both human comfort and energy saving can be

    achieved. Glare is another phenomenon which has impact on visual comfort. Glare is caused by

    unbalanced perception of light on the human retina when there is a significant luminance difference

    in his/her visual field. Glare can come from either the daylight source or an electrical light source.

    Especially the glare on a computer monitor is becoming an important factor to be considered as the

    typical office work evolves toward a computer-supported information processing task. Controlling

    glare is a non-trivial task because glare is also dependent on such factors as the occupants position

    and the mode of activities in a space which are normally dynamic in their nature (Li 1997).

    Solar radiation brings not just light but heat into the space. This phenomenon has energy implica-tion especially during summer and winter because it increases cooling load in summer and

    decreases heating load in winter (Reynolds 1992). In daylight responsive lighting control, deter-mining how much daylight should be introduced to the inside is, therefore, a very delicate issue

    (Mahdavi et al. 1995). There have been many new daylight modification systems introduced to theconstruction of a building. Device such as light redirection louver is adding more sophistication to

    our visual environment control capability by turning a traditionally static building facade into a

    dynamic one. By having this type of device, flexibility in controlling our visual environment can

    increase significantly even after a building is designed and constructed.

  • 9Figure 2.1: Daylight and electrical light process in a space

    2.1.3 New perspective on building control process

    New building systems, components, and assemblies have been introduced along with increasing

    attention to the innovative concepts for energy conservation and environmental quality. The goal of

    building operation is to maximize occupants satisfaction while minimizing operational cost. Fur-

    thermore, the integration of building systems for an enhanced performance is becoming more and

    more important to increase user satisfaction, efficiently. In this context, a more intelligent and reli-

    able building control approach is indispensable. DDC (Direct Digital Control) has greatly influ-enced the overall building control practice by innovating both control hardware and software

    (Brandt and Shavit 1984). Conventional PID (Proportional, Integral, and derivative) controller isstill being widely used, but new approaches using advanced process models or AI (Artificial Intel-ligence) techniques are now available to control various building processes including thermal,visual, and air flow environment (Jeannette et al. 1998). In recent times, internet-based buildingmonitoring and control systems have been introduced which enable location-independent building

    operation with proper set of software considering security and various levels of access privilege (Soet al. 1998). Neural network, as one of the machine learning techniques, has been used for thermalload prediction and HVAC system control and is sometimes combined with PID controller to

    enhance its adaptation capability. Four major issues are identified for this research considering thenew trends in building control domain:

    Extending the number of control criteria beyond the limited traditional ones

    Enhancing prediction capability of a building control model

    Internal reflection

    Direct component

    Luminaire

    External reflection

  • 10

    Agile and precise adaptation to the environmental and system changes

    Providing a well-established interaction scheme between users and building control system

    2.2 Models for building process control

    2.2.1 Simulation model

    2.2.1.1 Simulator to support building process control

    Simulation is the transparent model of a system constructed through examining pertinent objectsand processes in the target system. Simulation, as a virtual system, allows one to make system

    parameter changes and to study their effect without relying on the actual system which is some-

    times inaccessible or costly. A model can be gained in a mathematical form describing interactions

    among its components. Some of the ways to build a process model are discussed as follows:

    By acquiring a cross-correlation function out of input and output data and use it to derive one

    of multiple candidate model forms.

    By establishing a model having a core of formulations describing already explained mecha-

    nisms in the process followed by numeric manipulations to decide the unknown coefficients.

    By using regression method to determine the magnitudes coupled with hypothesized relations.

    By applying numeric method to fit unknown coefficients to the statistically defined dynamic

    relations.

    The increasing need for more energy-sensitive and adaptive systems for building control has been

    encouraging the use of more precise and delicate computational models of various building sys-

    tems. Those computational models are being used to supplement or sometimes even replace costly

    equipment testing to develop more energy-saving building systems and components. Simulations in

    real-time helps to build the control logic for complicated building systems to evaluate the appar-

    ently explosive number of potential options. The possible outcome of this process could be an

    option minimizing energy consumption and guaranteeing stable environmental control with reliable

    equipment at the same time.

    In building control domain, on-line simulation models can be used for predicting actual systems

    dynamic behavior by having input data from the sensors or other sources. Simulation, as the virtual

    building, can provide an useful tool for generating and evaluating multiple control alternatives

    based on the given set of preference criteria without being restricted by the constraints that actual

  • 11

    building and its systems have.

    Contextual data such as the fluctuation of indoor occupancy, exceptional environmental conditions

    (i.e. lighting), and the level of equipment use can be provided to the simulator to increase its predic-tive potential. The following techniques are often used to achieve different goals by using simula-

    tion as a model for control:

    Sensitivity analysis: Identification of control-sensitive variables is important. A series of

    experiments could be done to achieve this goal by systematically increasing or decreasing a

    subset of the independent variables, while others remain constant, to observe the impact of

    such operations.

    Goal seeking: In goal seeking, the desired target output of a system should be defined first to

    identify set of inputs which possibly yield that target output. Special techniques are necessary

    to get to the desired input because there can be multiple input sets generating the same target

    output. For example, to satisfy target indoor illuminance level, one can either change the lumi-

    naire dimming level or simply manipulate the blind. When a target output is defined, both sen-

    sitivity analysis and what-if analysis can help to guide which input set better be modified to

    produce the target output.

    What-if analysis: What-if analysis is to identify the consequence of a certain assumed control

    action. By setting simulation input values the virtual action, the outputs of the simulation can

    be analyzed to see the outcome of that assumed action taken as assumption. This process is

    also useful for system fault detection or for identifying an adaptation strategy when a building

    component system or the building configuration is changed.

    The way of using a system model in an actual control practice is another issue to be explored. Typ-

    ical feed-forward control (model-based) process as opposed to the feed-back control (sensor-based)process is illustrated in Figure 2.2. Usually, a system model identifies necessary set of inputs to the

    system by predicting the system outputs with different input sets and comparing them with the tar-

    get output. This goal-seeking process needs a reasonably precise system model which is an

    essential part of feed-forward (open-loop) control architecture.

  • 12

    Figure 2.2: Feed-forward (a) and feed-back (b) control process

    Figure 2.3 shows an example of how an analytical system model such as a simulator can be used for

    the control purpose. Loop a) is the model calibration cycle by which the simulator is calibrated basedon the detected model error. Loop b) is the goal seeking loop by which a sufficiently calibrated sim-ulator is used to identify a proper set of system inputs for generating the given target output.

    Figure 2.3: Feed-forward control process example

    A critical task toward the realization of a simulation-assisted building control system lies in the

    development of a strategy to create a well-defined set of control options that, together with the pro-

    jected contextual conditions, serve as the basis for comparative and/or parametric simulation runs.The results of these simulations enable the control system to anticipate the impact of a control

    action choice on the selected performance indicators. While there may be numerous methods to

    derive at a structured set of such control options, two principal approaches are briefly described

    below (Mahdavi and Berberidou 1994):

    system model system

    system model system

    sensor

    r(t) u(t) y(t)

    u(t) y(t)r(t) c(t)+

    -

    a) Feed-forward control process

    b) Feed-back control process

    r(t): model input, c(t): calibrated model input, u(t): system input, and y(t): system output

    system model

    target output

    model output

    system output

    -model error

    -control error

    system

    input

    -deviation

    a) calibration loop

    b) goal seeking loop

  • 13

    The Generate-and-Test Method (GAT): This method involves the rule-based generation of afinite number of discrete control options. Such options may involve, for example, various posi-

    tions of a movable external shading device, positions of operable windows, or various on/off

    timing schemes for intermittent heating/cooling. These schemes are then evaluated and ranked

    (possibly in view of multiple criteria involving power consumption, cost, emissions, visual andthermal comfort, etc.) based on the results of multiple simulation runs.

    The Bi-directional Inference Method (BDI): This method involves the explicit definition ofcontrol variables and performance indicators. An example of a control variable would be the

    position of a moveable external shading and/or light-redirection device, or the deviation of

    heating/cooling set-point temperature from the space target temperature. Examples of the per-

    formance indicators are the average illuminance on a desk surface, illuminance distribution

    uniformity, annual building energy demand, the average cumulative deviation of the main-

    tained space temperature from the set-point temperature, or the average cumulative PPD (Pre-dicted Percentage of Dissatisfied) in a space. Starting from an initial operational state, the bi-directional inference mechanism facilitates the derivation of required changes in the control

    variables based on desired changes in the performance indicators. This derivation can be

    accomplished via the investigative projection technique.

    Figure 2.4 illustrates the use of BDI and GAT for simulation-assisted building control.

    Figure 2.4: : Use of BDI and GAT for simulation-assisted building control (Mahadavi 1997b)

    target perf.attributes

    preferenceprocessing

    virtual building control variableattribute

    sensing control

    rule-based optiongenerator

    controlschemes matrix

    parametricsimulation

    control schemeselection

    Generate-And-Test (GAT)

    Bi-Directional Inference (BDI)

  • 14

    2.2.1.2 LUMINA as a lighting simulator

    To be able to use a simulator as the prediction tool for both daylight and electrical lighting, it must

    have:

    building model

    sky model

    system model (light redirection louver, shade, luminaire etc.)

    occupancy model

    daylight process model

    electrical light process model

    Lighting process model can be coupled with thermal process model to assess the impact of heat

    gain from solar radiation and luminaires. As a lighting simulation program, LUMINA (Pal andMahdavi 1999) is capable of modeling both daylight and electrical lighting process even in thespaces having a complex geometry. LUMINA uses both radiosity and ray tracing as its core algo-

    rithms. Ray tracing and radiosity based simulation techniques have overcome many of the limita-

    tions a simplified table-based hand calculation method has.

    The major concerns of the ray tracing or the radiosity based methods are computational storage andtime. The computational storage is somewhat more of an issue in the case of radiosity (especiallyfor diffuse surfaces) while the computation time is critical in the case of ray tracing (especially forspecular surfaces). Most global illumination simulation techniques may use a combination of radi-osity and ray tracing method, making a trade-off between computational storage and computational

    time (Sillion 1994).

    Since most architectural spaces include diffuse surfaces with occasional specular surfaces, a pre-

    dominantly radiosity-based method which enters into a recursive ray tracing mode whenever a

    specular surface is encountered is preferred for the development of LUMINA. In a radiosity based

    algorithm, the most computationally expensive part of the calculation, i.e. the calculation of form

    factors, needs to be performed just once and these results can be used repeatedly for multiple itera-tions. A two-pass radiosity algorithm using extended form factors which include ray traced specu-

    lar transfers is given by Tellier and Bouatouch 1994. This is extended to include transmitting

    surfaces and is implemented in the following sub-processes (Pal 1999):

  • 15

    1. Read internal surfaces from a file. Surface information includes vertex coordinates (x, y, z) foreach vertex followed by reflectance, transmittance, diffuseness of reflectance, specularity of reflec-

    tance, phong exponent for reflectance, diffuseness of transmittance, specularity of transmittance,

    and a phong exponent for transmittance.

    2. Read external surfaces from a file. The surface information is the same as for internal surfaces.

    3. Divide internal and external surfaces into patches by overlaying a grid and determining which

    grid elements lie inside the surface. A smaller grid is used for internal surfaces and louvers and

    overhangs. A bigger grid is used for surrounding building surfaces.

    4. Read windows and lights. Repeat the process except that lights don't reflect or transmit but have

    an intensity distribution which is read from IES formatted luminaire files.

    5. Divide windows and lights into smaller elements by binary subdivision.

    6. Calculate the visibility matrix for internal surface elements.

    7. Calculate the visibility matrix for external surface elements.

    8. Calculate the extended form factor matrix for external surface elements. The extended form fac-

    tors use ray-tracing to determine the specular component of the extended form factor.

    9. Similarly, calculate the extended form factors for internal surface elements.

    10. For each time step of the simulation period:

    a) Build the sky luminance distribution using SOLARIS and Perez sky algorithm.

    b) Compute the direct component of illuminance on the external surface elements.

    c) Compute the radiosity solution for the external surfaces using progressive radiosity.

    d) Compute the direct component of daylight and also the initial direct lighting from the electriclights for the internal surface elements. Visibility of window or light patch is taken account. The

    sky is divided into patches and if a sky patch is obstructed by an external surface element, the lumi-

  • 16

    nance of the sky patch is replaced by the luminance of the obstruction. If the obstruction is specular,

    backward ray tracing is used to compute the specular component of light transfer from the external

    element to the internal element. The environment is partitioned into exterior and interior spaces.

    The radiosity solution for the exterior space is calculated first and then its contribution to the inte-

    rior space is computed. This is more efficient than treating them as one space, since it is easier to

    solve two smaller systems of equations rather than one large one. This is based on the assumption

    that the interior space has an insignificant contribution to the exterior space.

    e) Compute the radiosity solution for the internal surfaces using progressive radiosity.

    Running LUMINA takes about a minute for a normal sized room (1,000 patches). Most of the timeneeded is taken up by the visibility and extended form factor calculations. Light redirection louver

    is considered as a dynamic external object having multiple surfaces modifying incoming daylightbefore it is introduced into the interior space. LUMINA uses Perez model for generating sky condi-

    tions necessary for the simulation. Earlier daylight simulation techniques relied on simplified sky

    models such as the uniform sky, the CIE standard overcast sky, or the clear sky. These have limited

    applicability for predominantly intermediate skies. With the more advanced sky luminance distri-

    bution models available today, the applicability and accuracy of daylight prediction tools are

    expected to increase. However, most of these models have either been validated for the data from

    which they are empirically derived or have differential performance results when applied to differ-

    ent climatic conditions (Littlefair 1994). All weather sky model has been developed by Perez et al.(1993) on the basis of sky luminance data collected for predominantly clear skies, this model uses aclearness index and a sky brightness factor to compute the sky luminance distribution (Pal andMahdavi 1999).

    2.2.2 Machine learning model

    While simulation involves the explicit description of real systems and their processes, a typical

    machine learning technique relies on inductive reasoning based on the given data sets. On-going

    researches show that there are many alternative statistical methods (i.e. high-dimensional non-par-ticipant regression methods) available for dealing with data mining tasks with a weaker model andlarge data sets (Banks et al. 1999) Machine learning technique is capable of on-line learning fromthe real system operation. A machine learner can be tuned for optimal system operation even in an

    uncertain and dynamic environment. Most of machine learning modules need training and cannot

    handle untrained conditions beyond their capability of generalization. Depending on the source of

    knowledge and the way a machine learner captures it, a machine learning method can be classified

    as one of the following three categories:

  • 17

    Supervised learning: A supervised learner is given both inputs and corresponding outputs so that

    it can predict right condition-outcome mapping after training. Supervised learning is capable of

    function approximation and system identification. Supervised learning requires the correct output

    signal for each input vector to be specified. Neural network is one of the algorithms for implement-

    ing this type of learning.

    Reinforcement learning: Reinforcement learning is suitable when there is no condition-out-

    come example to be trained with. Instead, the learner only gets reward (either positive or negative)as a state is changed into another by an action. Reinforcement learning progresses by receiving an

    implicit scalar evaluation of an action and is capable of policy identification for choosing actions to

    achieve its goals (Mitchell 1997). This learning scheme achieves its goal through identifying aseries of actions that maximize rewards.

    Unsupervised learning: Unsupervised learning clusters and characterizes input values without any

    target output or reward guiding its learning process. An unsupervised learner self-adjusts its param-eters and structure to capture the regularities of input vectors without receiving explicit information

    from the external environment. This self-organizing learning scheme is useful especially when

    there is no prior knowledge either explicit or implicit except for the data set which only includes

    inputs. Clustering and discovery are two major application domains of unsupervised learning tech-nique. Clustering attempts to classify objects based on feature descriptions. Discovery learning ismore likely to be used in theoretical environment. For example, a program named BACON, as a

    model of data-driven scientific discovery, discovered a number of scientific laws such as ideal gas

    theory, Ohms law, and Joules law.

    The basic motivation of using machine learning technique is to capture the pattern of dynamic system

    behaviors (system identification) and to project its knowledge into the near future to predict possiblestates of a system or a controlled process (parameter estimation). Unlike rule-based systems, machinelearning techniques are useful in the control of unidentified complex and dynamic systems by learning

    directly from the multiple instances of the system input-output mappings called patterns. Machine

    learning module for a system control has strength especially in the following situations (Narendra1990):

    When a controller is theoretically definable, but too complex for implementation

    When the controlled process has a poor transient response for initial conditions in a domain of

    interest. In this case, the identification of the system over a long period of time followed by the use

  • 18

    of nonlinear controller is needed

    When stabilization of a controlled process is needed for the efficient operation of a system at sev-

    eral equilibrium states

    When a control should be carried out on the basis of output patterns and the state space is parti-

    tioned into disjoint regions which are equivalent for purposes of control, therefore, the optimalinput corresponding to each region is required

    One the advantages of having a machine learner as the major control agent is its adaptation capability todynamic conditions. When noticeable deviations are expected due to, for example, an incomplete

    model, adaptive control systems is supposed to compensate through backing-up the model based on the

    available information. Adaptive control techniques are normally designed to deal with uncertainty in the

    controlled process. A process model for control can be refined by a certain additional program which

    observes the behaviors of the control process model in operation. Adaptive system, therefore, is useful

    in building a controller capable of responsively dealing with a complicated and dynamic environment.

    Adaptive control allows a control system to observe its own behavior and adjust accordingly, whichrequires both learning and prediction capabilities.

    Neural network is one of the most popular supervised learning techniques. It has been engineered as a

    computational system modeled after the learning abilities and parallelism of biological nervous sys-

    tems. The applications of this technique include speech recognizers, explosives detectors, airline seat

    allocation systems, and loan risk assessors. Neural networks are suitable for non-linear modeling for

    dynamic system control. They can successfully be integrated with the overall control system. Dynamic

    systems are generally history-dependent and cannot be modeled by static input-output maps. Therefore,

    neural network architectures and training techniques for building dynamic models are required.

    Neural network learning methods provide a robust approach to approximating real-valued, discrete-val-

    ued, and vector-valued target functions. Learning to interpret complex real-world sensor data is one of

    the best fits of neural network. Neural network is robust to noise in training data (Mitchell 1997). Train-ing time of neural network depends on the number of weights in the network, the number of training

    examples considered, and the settings of various learning algorithm parameters. Neural network con-

    sists of input and output layer and hidden layers in between. Each layer contains nodes and links and

    those weights associated with the links are to be adjusted to fit the input-output relationships. The back-propogation algorithm is the most common network learning method. Backpropogation searches the

    space of possible hypotheses using gradient descent to iteratively reduce the error in the network fit to

    training examples. Gradient descent converges to a local minima in the training error with respect to the

  • 19

    network weights. Gradient descent is useful for searching continuously parameterized hypothesis

    spaces where the training error is a differentiable function of hypothesis parameters. The application of

    neural network typically comprises two phases. A learning phase and an operation phase. Learning is

    the process through which neural networks are acquiring the knowledge embedded in the training data.

    A trained network subsequently represents a static knowledge base which can be recalled during its

    operation phase. Following is the backpropagation algorithm used for capturing knowledge embedded

    in training examples (cp. Mitchell 1997):

    1. Create a feed-forward network with input units, hidden units, and output units

    2. Initialize all network weights to small random numbers (i.e. between -0.0.5 and 0.05).3. Until the termination condition is met, Do

    For each in training examples, Do

    propagate the input forward through the network:

    3-1. Input the instance to the network and compute the output of every unit in the network.

    propagate the errors backward through the network:

    3-2. For each network output, calculate its error term

    (2.3)

    3-3. For each hidden unit number, calculate its error term

    (2.4)

    3-4. Update each network weight

    for output unit weights (2.5)

    for hidden unit weights (2.6)

    Each training example is a pair of the form , where is the vector of network input values, and

    is the vector of target network output values. is the learning rate. is the output computed by unit ,

    is the target output for unit . The input from unit into unit is denoted , the weight from unit

    to unit is denoted , and the output error is represented as .

    The capabilities of neural networks in control domain can be summarized as follows (Barto 1990):

    Copying existing controller: Neural network can be trained from already existing controller, sys-

    nin nhidden nout

    x t,

    x ou u

    k

    k ok 1 ok( ) tk ok( )

    h

    h oh 1 oh( ) khkk outputs

    j i

    j i wjiEd

    tj oj( )oj 1 oj( )xji= =

    j i jxji=

    x t, x t oj j

    tj j i j xji i

    j wji Ed

  • 20

    tem model, or a human expert who has knowledge about effective control rules.

    Adaptive prediction: The input to the neural network consists of delayed values of the signals, and

    the target output consists of the current values of the signals. The network tries to match the current

    signal values by adjusting a function of their past values. When the input to the network bypassesthe delay units, the output of the network is a prediction of the values the signal will have in the

    future.

    System identification: Training information can be obtained by observing the input-output behav-

    ior of a system where the network receives the same input as the system, and the system output is

    the target neural network output. If one selects a class of model to design an effective controller, a

    control law can be expressed in terms of the parameter estimates produced by the system identifica-

    tion procedure.

    Identification of system inverse: In this method, the input to a network is the output of the system,

    and the target output of the neural network is the system input. If the neural network is a system

    inverse, by feeding desired system output to the neural network, it can predict what the system

    input should be. For error learning, the input to the neural network should be actual system output

    and a feedback controller needs to be used during training.

    Differentiating a model: This method is for training a controller relying on backpropogation

    between layered neural networks such as forward model (system model) and controller. The propa-gation of the errors between actual and target system outputs back through the system model gener-

    ates the error in control signal useful to train another neural network for controller.

  • 21

    3 HISSTO: a hybrid system for environmental control

    3.1 Problem identification

    3.1.1 Issues in building control models

    With the advances in computation and DDC (Direct Digital Control), the model-based building controlbecomes an attractive option. A model may not always precisely capture the actual system behavior

    due, in part, to the difficulty of acquiring exact descriptions of building system properties, such as mate-

    rials and construction features. Some simulation programs are computationally too intensive to be effec-

    tive for real-time control purpose. An example is a lighting simulator that uses ray tracing in its

    modeling process. Such simulation programs cannot be used for control purposes unless the control

    state search space is dramatically reduced. Machine learning can address this problem. However, it

    often requires large amount of data for training. For instance, before a neural network is trained, or if it

    encounters unexpected conditions, it is not able to predict accurately (Curtiss 1996). The need forretraining makes it difficult for a machine learner to respond quickly to the system retrofit and/or sea-

    sonal weather pattern changes. Its sensor-dependency represents an additional difficulty, especially

    when placing and/or maintaining sensors are costly or otherwise not desirable. Conventional lighting

    control systems do not necessarily allow multiple control modes such as manual mode, scheduled

    mode, and predictive mode. Limited accessibility due to the location-dependency of building systems

    control also needs to be addressed.

    3.1.1.1 Challenges in simulation

    Accuracy of prediction

    Simulation is a generic analytical model for a specific process and can not always describe actual

    world. In many cases, validation of a simulator based on empirical data is expensive and time-consum-

    ing. It is normally pursued at a national or an international level, although many individual research

    groups undertake small monitoring schemes which focus on specific areas of building performance

    (Clark 1984). It usually takes much time to build and validate a simulation. Simulation can be used for

  • 22

    model-based control, training and testing of other control models (i.e. machine learning), and by gener-ating the pairs of control action and predicted outputs so that the most promising control option can be

    identified. It also can be used to train or test other control models (i.e. machine learner). The feed-for-ward (model-based) control scheme requires high resolution system behavior description of the systemmodel. Ultimately, it is only by comparing model predictions with the corresponding results from actual

    building in operation that a models usefulness as a predictive agent must be ascertained. Unfortunately,

    this task is difficult due to the lack of comprehensive data relating to the performance of real buildings

    and the shortcomings inherent in even the most sophisticated technique, which make it difficult to

    model reality exactly. Commonly encountered problems in validating an energy simulation program are

    shown in the Table 3.1 (Clark 1984).

    Increased computational load

    Some simulation programs are too computation-intensive for real-time or quasi real-time control pur-

    pose. A typical example is any simulator using FEM (Finite Element Method) or ray tracing as the pri-mary modeling algorithm. Normally, model predictions and the control decision based on them should

    be done within a certain time frame. For this reason, those simulation programs demanding heavy com-

    putation are not suitable for the control purpose unless the control state search space is dramatically

    reduced to limit the number of simulation sessions necessary for testing different control options. This

    problem has been significantly eased by the dramatic increase of available computing power, yet a care-

    ful consideration needs to be made when developing of a building product or process model for the con-

    trol operation demanding heavy computation.

    Table 3.1: Exemplary challenges in energy simulation validation process

    Complex process Monitoring air movement and specifying the results in a form meaningful to modelers

    Data validation On-line validation, organization and management of large data sets in a manner which al-

    lows estimation of the data reliability

    Occupants behavior Ill-understood actions of occupants with regard to window opening, blind operation, and

    control manipulation

    Design/construction

    discrepancy

    Accurate determination of thermo-physical properties as required by the models

    Cost of development Cost associated with the data logging system, sensor installation, and quality technical staff

    Climatic data Availability of simultaneously recorded climatic data relating to the test site (This is espe-cially problematic with many historical performance data sets.)

    Instrumentation Necessity to ensure, by constant checking, that all instruments remain as calibrated for the

    duration of the monitoring period

  • 23

    3.1.1.2 Challenges in machine learning

    Data dependency

    Machine Learning technique often requires large amount of data for training. For example, neural net-

    work training process is normally data-intensive and time consuming even though a trained neural net-

    work quickly produces outputs for the given inputs. Also, in case of building configuration or system

    changes, a machine learning module needs retraining for adaptation. This makes the neural network

    slow in responding not just during the initial training phase but also when it needs a update throughretraining. This problem leads to the following requirements:

    1) Provision of seamless transition from the training mode to the operation mode2) Creation of an algorithm whereby the training occurs with minimum memory requirement3) Inclusion of a substitute algorithm to generate values for the controller output while the neural net-work is trained

    Sensor dependency

    Typical neural networks require the inputs/target outputs to be measured, which makes it inevitably sen-

    sor dependent. This sometimes causes problem, especially when placing or maintaining sensors is diffi-

    cult or costly. There are certain cases no sensor could be placed because it simply doesnt exist. For

    example, glare in an office environment can not easily be measured because it again is the function of

    the viewers location and view angle which are dynamic in their nature. This sensor dependency can

    easily limit the boundary of the adoptable performance criteria by which a desired control option is

    selected. In fact, lots of control variables in building operation can only be calculated instead of being

    measured, and this poses problems in using neural networks for control purpose.

  • 24

    3.2 Requirement elicitation

    HISSTO is a pilot control system applied to an intelligent daylight responsive lighting control task. In

    general, human comfort and cost minimization are the primary driving forces behind any building sys-

    tems control efforts. For a successful daylight responsive lighting control, HISSTO must minimize the

    lighting and thermal energy consumption and maximize occupants visual comfort in the target control

    zone. Visual comfort in a space is not sufficiently captured just by identifying indoor illuminance distri-bution, therefore, additional performance indicators need to be introduced. HISSTO makes use of the

    simulator-assisted machine learner training as well as the machine learner-supported simulator calibra-

    tion and tuning. HISSTO needs to capture seasonal changes of the sky condition while adapting itself to

    the changes of building configuration (e.g. spatial retrofit). There are also some non-functional require-ments (Bruegge and Dutoit, 1999) for developing HISSTO. Predicting and evaluating the values of theperformance indicators for all candidate control options should be done within a certain time frame (i.e.10 to 15 minutes). It is not feasible to use a large number of sensors in the target space beyond a certainperiod of data collection. Therefore, the sensors required during HISSTOs operation stage need to be

    confined to a small number (i.e. one or two).

    3.2.1 Functional requirements

    3.2.1.1 Visionary scenarios of HISSTO

    A scenario, as an instance of a use case, is a concrete, focused, informal description of a single feature

    of the system from the viewpoint of a single actor (Bruegge and Tutoit 1999). A scenario could be oneof four types (Carroll 1995). As-is scenarios reflect current situation regarding the current systemunderstood by observing and describing user actions. Visionary scenarios focus on future system to

    refine ideas on the system needs to be developed. Evaluation scenarios describe user tasks against

    which the system is to be evaluated. Training scenarios teach users how to use the system. Since devel-

    oping HISSTO is a greenfield engineering, a new system development effort, which starts form the

    scratch, elucidating visionary scenarios is the best choice to identify HISSTOs potential requirements.

    Table 3.2 shows an exemplary set of HISSTOs visionary scenarios describing possible system usages

    of the different users.

  • 25

    Table 3.2: Visionary scenarios of HISSTO

    ID Scenario name Actorinstance Flow of events

    VS1 managerOnTravel

    David:

    Facility-Manager

    1. David is out of Pittsburgh participating in the COMDEX show inAtlanta.2. David opens up the web-browser on his laptop computer, connects toHISSTO web interface and be authorized.3. David presses monitoring button and select ControlZone 14. All sensor values and the values of control performance indicators forControlZone1 are displayed on the screen.5. David presses stop button and closes the web-browser.

    VS2 repeatedLight-ening

    Kate:

    Occupant

    1. While working at her office, Kate notices that on-going lightening dis-tracts ambient light level too often due to the predictive control mode inoperation.2. Kate opens up the web-browser on her desktop computer, connects toHISSTO web interface and be authorized.3. Kate selects ControlZone 1 and change control mode from predictive tomanual.4. Kate sets a comfortable ambient light level by manipulating luminaireslide bars on the screen.5. Kate presses exit button and closes the web-browser.

    VS3 teamAOnPicnic Mark:

    Occupantin team A

    1. Team A members decide to have a brief outdoor picnic.2. Mark, the team leader, opens up the web-browser on his desktop com-puter, connects to HISSTO web interface and be authorized.3. Mark selects ControlZone 1 and changes control mode from predictiveto scheduled.4. HISSTO sets ambient light level and louver angle based on the non-occupancy schedule.5. Mark presses exit button and closes the web-browser.

    VS4 workOnWeek-end

    Kate:

    Occupant

    1. Kate comes to her workspace on Sunday to finish the remaining work.2. Kate opens up the web-browser on her desktop computer, connects toHISSTO web interface and be authorized.3. Kate selects ControlZone 1 and change control mode from scheduled topredictive.4. Kate presses exit button and closes the web-browser.

    VS5 partition-Change

    David:

    Facility-Manager

    1. David gets a phone call from Mark, the team A leader, notifying that thepartitions in their workplace have been reconfigured.2. David checks out the new partition configuration, connects to HISSTOserver interface and be authorized.3. David selects ControlZone 1 and changes building configuration fileaccording to the new configuration.4. David presses exit button and terminates the HISSTO server interface.5. HISSTO updates simulator and machine learner for change adaptation.

    VS6 workspaceRe-assignment

    Julia:

    Occupant

    1. Julia moves to Kates workspace after the recent work group change.2. Julia opens up the web-browser on her desktop computer, connects toHISSTO web interface and be authorized.3. Julia s