Caio A. S. Coelho e-mail: [email protected] Supervisors: D. B. Stephenson, F. J. Doblas-Reyes (*) Thanks to CAG, S. Pezzulli and M. Balmaseda.

Caio A. S. Coelho

e-mail: [email protected]

Supervisors: D. B. Stephenson, F. J. Doblas-Reyes (*)

Thanks to CAG, S. Pezzulli and M. Balmaseda (*)

Department of Meteorology, University of Reading and ECMWF (*)

Forecast Assimilation of DEMETER Coupled Model Seasonal Predictions

Plan of talk1. Issues2. Conceptual framework (“Forecast Assimilation”)3. DEMETER4. Examples of application: 0-d, 1-d, 2-d.

5. Conclusions

•

1. Issues

• Why do forecasts need it?• Which are the best ways

to calibrate?• How to get good probability

estimates?

Calibration

Combination • Why to combine?• Should model predictions be

selected?• How best to combine?

2. Conceptual framework

)y(p

)x(p)x|y(p)y|x(p

i

iiiii

Data Assimilation “Forecast Assimilation”

)x(p

)y(p)y|x(p)x|y(p

f

fffff

3. Multi-model ensemble approach

DEMETER DEMETER Development of a European Multi-Model Ensemble

System forSeasonal to Interannual Prediction

Solution: Multi-model Ensemble

Errors: Model formulationInitial conditions

http://www.ecmwf.int/research/demeter

DEMETER Multi-model ensemble system

7 coupled global circulation models

Hindcast period: 1980-2001 (1959-2001)

9 member ensembles

ERA-40 initial conditions

SST and wind perturbations

4 start dates per year

(Feb, May, Aug and Nov)

6 month hindcasts

Model Country

ECMWF International

LODYC France

CNRM France

CERFACS France

INGV Italy

MPI Germany

UKMO U.K.

4. Examples of application

•

• Niño-3.4 index (0-d)• Equatorial Pacific SST (1-d)• South American rainfall (2-d)

Example 1: Niño-3.4 forecasts

Well-calibrated: Most observations in the 95% prediction interval (P.I.)

95% P.I.valueJulyY

valueDecemberY

),Y(N~Y|Y

5t

t

2t05t1o5tt

ECMWF coupled model ensemble forecasts

Observations not within the 95% prediction interval! Coupled model forecasts need calibration

m=9DEMETER: 5-month

lead

2X

2ttt

2ttt sˆ;Xˆ);,(N~X

Prior:

Univariate X and Y

),(N~Y 2t0t0t

)V,Y(N~Y|X tttt

'

2X

t m

m

m

sV

),(N~X|Y 2tttt

t

t

2

2t0

t02

t

t

t

2

2t0

2t

X

V

V

11

)X(p

)Y(p)Y|X(p)X|Y(p

t

ttttt

Posterior:

Likelihood:

Bayes’ theorem:

Modelling the likelihood p(X|Y)

y

Combined forecasts

Note: most observations within the 95% prediction interval!

All forecasts

Forecast MAE

(C)

MAESS (%)

BS BSS

(%)

Uncert

(C)

Climatol. 1.16 0 0.25 0 1.19

Empirical 0.53 55 0.05 79 0.61

Coupled 0.57 51 0.18 29 0.33

Combined 0.31 74 0.04 81 0.32

MAESS = [1- MAE/MAE(clim.)]*100%

Empirical Coupled

Combined

BSS = [1- BS/BS(clim.)]*100%

)C,Y(N~Y b

1TT

111T

obba

)SGCG(CGL

C)LGI()CGSG(D

)]YY(GX[LYY

)S],YY[G(N~Y|X o

Prior:

Likelihood:

Posterior:

1YYXYSSG

YGXGYo T

YYXX GGSSS

)D,Y(N~X|Y a

Multivariate X and Y

bias

qq:D

qn:Y

pn:X

qq:C q1:Yb

pp:S qn:Ya

Matrices

Example 2: Equatorial Pacific SST

Forecast Brier Score (BS)

BSS

(%)

Climatol p=0.5 0.25 0

Multi-model 0.19 24

FA 58-01 0.17 31

)0YPr(p tt

SST anomalies: Y (°C)Forecast probabilities: p

DEMETER: 7 coupled models; 6-month lead

BSS = [1- BS/BS(clim.)]*100%

Y 0Y

Brier Score as a function of longitude

Forecast assimilation reduces (i.e. improves) the Brier score in the eastern and western equatorial Pacific

1BS0)op(n

1BS

n

1k

2kk

Brier Score decomposition

1BS0)op(n

1BS

n

1k

2kk

)o1(o)oo(Nn

1)op(N

n

1BS

l

1i

2ii

l

1i

2iii

iNk

ki

i1i oN

1)p|o(po

n

1kko

n

1o

reliability resolution uncertainty

Forecast assimilation improves reliability in the western Pacific

Reliability as a function of longitudeReliability as a function of longitude

Resolution as a function of longitude

Forecast assimilation improves resolution in the eastern Pacific

Why South America?

El Niño (DJF)

La Niña (DJF)

Source: Climate Prediction Center (http://www.cpc.ncep.noaa.gov)

Seasonal climate potentially predictable

DEMETERMulti-model

Correlation: DJF rainfall

Why South American rainfall?

Agriculture

Electricity: More than 90% produced by hydropower stations

e.g. Itaipu (Brazil/Paraguay):• World largest hydropower plant• Installed power: 12600 MW • 18 generation units (700 MW each)• ~25% electricity consumed in Brazil• ~95% electricity consumed in Paraguay

Itaipu

Example 3: South American rainfall anomalies

Obs Multi-modelForecastAssimilation

(mm/day)

DEMETER: 3 coupled models

(ECMWF, CNRM, UKMO)

1-month lead

Start: Nov DJF

ENSO composites: 1959-2001

• 16 El Nino years

• 13 La Nina years

r=0.51

r=0.28

r=0.97

r=0.82

South American DJF rainfall anomaliesObs Multi-model Forecast

Assimilation

(mm/day)

r=-0.09

r=0.32

r=0.59

r=0.56

South American DJF rainfall anomaliesObs Multi-model Forecast

Assimilation

(mm/day)

r=0.04

r=0.08

r=0.32

r=0.38

Brier Skill Score for S. American rainfall

Forecast assimilation improves the Brier Skill Score (BSS) in the tropics

limcBS

BS1BSS

)0YPr(p tt

Reliability component of the BSS

Forecast assimilation improves reliability over many regions

limc

reliabreliab BS

BSBSS

Resolution component of the BSS

Forecast assimilation improves resolution in the tropics

limc

resolresol BS

BSBSS

• unified framework for the calibration and combination of predictions – “forecast assimilation”

• improves the skill of probability forecasts• Example 1: Niño-3.4

improved mean forecast value and prediction uncertainty estimate

• Example 2: Equatorial Pacific SST improved reliability (west) and resolution

(east)• Example 3: South American rainfall

improved reliability and resolution in the tropics improved reliability over subtropical and central regions

5. Conclusions:

•Coelho C.A.S. “Forecast Calibration and Combination: Bayesian Assimilation of Seasonal Climate Predictions”. PhD Thesis. University of Reading (to be submitted) • Coelho C.A.S., D. B. Stephenson, F. J. Doblas-Reyes and M. Balmaseda: “From Multi-model Ensemble Predictions to Well-calibrated Probability Forecasts: Seasonal Rainfall Forecasts over South America 1959-2001”. CLIVAR Exchanges (submitted).• Stephenson, D. B., Coelho, C. A. S., Doblas-Reyes, F.J. and Balmaseda, M.“Forecast Assimilation: A Unified Framework for the Combination of Multi-Model Weather and Climate Predictions.” Tellus A - DEMETER special issue (in press).• Coelho C.A.S., S. Pezzulli, M. Balmaseda, F. J. Doblas-Reyes and D. B. Stephenson, 2004: “Forecast Calibration and Combination: A Simple Bayesian Approach for ENSO”. Journal of Climate. Vol. 17, No. 7, 1504-1516.

• Coelho C.A.S., S. Pezzulli, M. Balmaseda, F. J. Doblas-Reyes and D. B. Stephenson, 2003: “Skill of Coupled Model Seasonal Forecasts: A Bayesian Assessment of ECMWF ENSO Forecasts”. ECMWF Technical Memorandum No. 426, 16pp. Available at http://www.met.rdg.ac.uk/~swr01cac

More information …

Reliability diagram (Multi-model)

(pi)

(oi)

o

Reliability diagram (FA 58-01)

o

(pi)

(oi)

Operational Seasonal forecasts for S. America• Coupled models

U.S.A: http://iri.columbia.edu

• Atmospheric models forced by persisted/forecast SSTs

Brazil: http://www.cptec.inpe.br

Europe: http://www.ecmwf.int

U.K: http://www.metoffice.com

Mean Anomaly Correlation Coefficient

Momentum measure of skewness

n

1i

3

y

i1 s

yy

n

1b

Measure of asymmetry of the distribution