coupled breeding for ensemble multiweek prediction
DESCRIPTION
Coupled Breeding for Ensemble Multiweek Prediction. www.cawcr.gov.au. Harry Hendon, Patricia Okely, Debra Hudson, Yonghong Yin, Oscar Alves, Griff Young, Andrew Marshall plus others. Outline Motivation for coupled breeding approach A step toward extending seasonal system to multiweek - PowerPoint PPT PresentationTRANSCRIPT
The Centre for Australian Weather and Climate ResearchA partnership between CSIRO and the Bureau of Meteorology
Coupled Breeding for Ensemble Multiweek Prediction
Harry Hendon, Patricia Okely, Debra Hudson, Yonghong Yin, Oscar Alves, Griff Young, Andrew Marshall
plus others
www.cawcr.gov.au
Outline• Motivation for coupled breeding approach
A step toward extending seasonal system to multiweek
lagged ensemble is underdispersive
need a consistent set of atmos/ocean/land perturbations
On the path to coupled ensemble-based assimilation
• Examples of forecast benefits for multiweek leads
• Some (limited) analysis of statistics of bred perturbations
focus on first 10 days comparing bred/lagged
primarily focused on atmos perturbations (not discounting importance of coupling)
2
Developing an intraseasonal forecasting capability from seasonal system
POAMA 1.5b (ABOM1)2007-2011
AGCM T47-L17 OGCM 2-0.5
Basic ocean data assimilation (T only) in offline OGCM
atmos/land initialization (ALI); strongly nudge AGCM to ERA in AMIP run
hindcasts 10 member lagged ensemble (successive 6 hour earlier;
0.25-2.5 days)
realtime: once per day
POAMA M24 (ABOM2) 2013-
Ensemble-based ocean data assimilation; T and S (PEODAS)
ALI atmos/land initialization
Coupled breeding to generate atmos-ocean perturbations
Multi-model (3 versions)
33 member burst ensemble every 5 days (twice per week operationally)
Some useful intraseasonal forecasts
Lagged ensemble cumbersome
Inconsistent realtime/hindcasts
Reliability issues (under-dispersive)
Improved Intraseasonal forecasting capability
Improved reliability and products
Consistency hindcasts/realtime
POAMA: Predictive Ocean Atmosphere Model for Australia
3
POAMA Ensemble Ocean Data Assimilation System Yin et al. 2011
runs in OGCM forced by ERA surface fluxes and SST relaxed strongly to Reynolds OI
Ensemble OI (Oke et al. 2005)
Cross-Covariances from ensemble spread
(3D multi-variate-time evolving)
Assimilation only into central member
ASSIMASSIM
Observations T&S
Compress Ensemble
Nudge to central analysis
Synthetic Perturbed wind forcing
(Alves and Robert 2005)
+ ocean perturbation
Coupled Breeding builds on Ensemble Ocean Assimilation
provides ensemble of ocean states, but not for atmos
Ensemble of OGCM integrations
1 day
4
Coupled Model forecasts
1 dayCentral unperturbed analyses:
PEODAS (ocean) and ALI (atmos)
Coupled Breeding Initialisation System
Member perturbations rescaledseparate norms for ocean and atmosthen centred to the central analyse
s
First Step Towards Coupled Assimilation...
Based on the PEODAS and ALI infrastructures:
Atmos: zonal mean rmsd surface zonal wind=analysis uncertainty (ERA-NCEP)
Ocean: 3-d T/S rmsd = analysis uncertainty (PEODAS)
rescale threshold met ~everyday in midlat atmos; every 4-5 days in tropics and oceans
Output an ensemble of atmos/ocean/land perturbations
5
Coupled breeding
Impact of ensemble generation (and multimodel)
POAMA-1.5
POAMA-2 intraseasonal
Burst ensemble
Time-lagged ensemble
6 hour lagged Atmos IC
Ocean IC
Ensemble spread
NRMSE of ensemble mean
SHEM 500 hPa Geopotential hghts
0 10 20 30d
0 10 20 30d
6
Improved forecast reliability
POAMA-1.5
POAMA-2 (seas)
POAMA-2 (intra)
Weeks 1 and 2
(all forecast start months 1980-2006)
Weeks 3 and 4
Probability of rainfall in upper tercile
All grid points over Australia
7
MJO Forecast skill 1982-2011 (1st each mnth)RMM1 and RMM2
Bred spread
Bred RMSE
ABOM2
Lagged ABOM1
Bivariate RMSE/Spread
Coupled breeding significant improvement over lagged, but still under-dispersed in Tropics
Improvement of ensemble mean over individual members
Courtesy D. Waliser
8
Some analyses of the statistics of the perturbations
Compare lagged to bred, plus a sensitivity exp using jumbled
Jumbled: make a new set of perturbations by randomly sampling the bred perturbations from all other years
should elucidate day-to-day “flow dependence”
•How flow dependent are the bred perturbations? (highly)
•Are “flow of the day” any better than jumbled? (not really)
•How coupled are they, or does coupling matter? (can’t fully answer
yet but apparently not important for longer leads)
•How optimal are the bred perturbations? (certainly better than
lagged but still under-dispersed in Tropics)
9
Spread and RMSE (9 member ensemble)DJF Southern Hemisphere Z500
Slight benefit of “flow of the day” perturbations
Black = control bredRed = laggedYellow = old lagged: p15bBlue = jumbled bred
spread
rmse
10
Spatial correlation of Initial Perturbations (90S-90N)
Perturbations defined wrt to central member
Bred, Jumbled, Lagged (6hourly)
Mean of absolute correlation of perturbations from member 1 with other 9 members
1st Dec, 1st Jan, and 1st Feb 1982-2011
6 hour lagged mean abs(r)=0.55
Bred 0.18
Jumbled 0.14
0.0
0.7
11
Examining flow-of-day sensitivity: U850 spread composites for ENSO
Initial time
After 10 days
U850 spread/anomaly along equator
warm
cold
warm-cold
12
Control vs Jumbled results
Control at initialisation
Jumbled at initialisation
After 10 days
Factor 10 smaller scale
Implication: coupled ocean-atmos perturbations not critical for reliable long lead prediction of ENSO; might not be true for short lead prediction of MJO
13
Association of spread with westerly anomalies is general throughout tropics
Correlation U850 spread (from breeding) with obs U850 anomaly at initialisation
high spread goes with westerly anomalies in tropics> convective regimes
14
Comparing Bred vs Lagged (6 hourly: 0.25-2.5 days)
Mean variance of perturbations MSLP initial time
DJF 1982-2011Bred
Lagged
varianceZonal mean amplitude
Variance normalized by zonal mean at each lat.
15
Bred Lagged
day 4
day 11
16
Bred lagged
day 1
Perturbation Growth rate
day 5
Wavenumber
0 20
)(
)(ln1
)(tICamp
tFampt
17
Summary
• Coupled-breeding ensemble generation has led to increased multiweek skill and reliability in POAMA-2 (ABOM2)
But breeding is still underdispersive in Tropics
•Despite simplicity, lagged ICs are far from optimal for multiweek, esp due to slow growth in Tropics
Lagged initial conditions 6hr apart are too similar and not good sample of analysis uncertainty
Infer> need to be 2-3 days apart but pay accuracy penalty for multiweek
•Next steps Further analysis of perturbations (coupling, MJO-dependence, etc)Refine breeding cycle to target increased tropical spreadImplement weakly coupled assimilation (show tomorrow)Fully coupled assimilation (cross covariance ocean-atmos)
18
Define mean perturbation amplitude for initial conditions and forecasts over the M ensemble members at each time as
]))()((1[)( 2 tAtAM
sqrttICamp centrali
Assuming exponential growth , define growth rate:
]))()((1[)( 2 tFtFM
sqrttFamp centrali
)(
)(ln1
)(tICamp
tFampt
Decompose ICamp and Famp as functions of zonal wavenumber
perturbation growth rates as function of zonal scale
)exp()()( tICamptFamp
19