Correcting monthly precipitation in 8 RCMs over Europe

Blaž Kurnik (European Environment Agency)Andrej Ceglar, Lucka Kajfez – Bogataj (University of Ljubljana)

Outline

• Regional climate models and observation - observation from E-OBS - RCMs from ENSEMBLES project

• Techniques for correcting precipitation prior use in impact models – bias corrections

• Validation of the methodology with results

The question

Can we use precipitation fields from RCMs directly in impact models?

Climate models

Clim

ate

mod

elIm

pact

mod

els

Ensembles of Climate models -simplified

RCM1 RCM2RCM3

RCM4

RCM5RCM6RCM7

GCM

RCMs used in the study RCM

GCM*

SMHIRCA3

MPIREMO

KNMIRACMO

ETHZCLM

DMIHIRLAM

CNRMALADIN

BCMMETNO

ECHAM5MPI

HadCM3QUK - MET

ARPEGECNRM* Only 1 scenario - A1B - which is version of A1 SRES scenario

Outputs from RCMsMonthly precipitation PDFs at different locations

Correction of the climate model data – workflow

Observations

SM1

DM2

ETH

MPI

CNR

DM1

SM2

KNM

25 k

m x

1 d

ayEu

rope

, bet

wee

n 19

61 -

1990

Bias correction

Correction of the climate model data

• Adjusting of the distribution function at every grid cell

• Long time series (> 40 years) of observation data are needed - correction and validation of the model (20 +20 years)

• Corrections are needed for each model separately

Precipitation correction the climate model data – transfer function

𝑐𝑑𝑓 (𝑥 ,𝛼 , 𝛽 )=∫0

𝑥 𝑒− 𝑥 ′𝛽 𝑥 ′𝛼− 1

Γ (𝛼)𝛽𝛼 𝑑𝑥 ′+cdf (0)

cdfobs(y) = cdfsim(x)

𝑦= 𝑓 (𝑥 )=𝑐𝑑𝑓 𝑜𝑏𝑠− 1 [𝑐𝑑𝑓 𝑠𝑖𝑚(𝑥 )]

Piani et al, 2010

Cumulative distribution

Probability for dry event

Fulfilling criteria

Corrected precipitation Modelled precipitation

Bias corrected data – ensemble mean of annual/July precipitation

Observed Simulated Corrected

Observed Simulated Corrected

Annual1991 - 2010

July1991 - 2010

Kurnik et al, 2011, submitted to IJC

RMSE of simulated and corrected

simulated corrected

RMSE cor=1𝑁 √∑i=1

N

(RRcor− RRobs)2

RMSE ¿=1𝑁 √∑i=1

N

(RR¿−RR obs)2

Failed correction – number of models

RMSEsim < RMSEcor

1.5 % area all models failed4.5 % area > 6/8 models failedDM1 90% cases cor(RMSE) < sim(RMSE)ETH 75% cases cor(RMSE) < sim(RMSE)

Brier Score – zero precipitation

simulated corrected

BS cor=1𝑁 √∑i=1

N

(PROB (RR=0 )cor−OBS (RR=0))2BS¿=1𝑁 √∑i=1

N

(PROB (RR=0 )¿−OBS (RR=0))2

BS 0: the best probabilistic predictionBS 1: the worst probabilistic prediction

Brier Score – heavy precipitation (RR> 200mm)

simulated corrected

BS cor=1𝑁 √∑i=1

N

(PROB (R R>200 )cor−OBS (RR>20 0))2BS¿=1𝑁 √∑i=1

N

(PROB (RR>20 0 )¿−OBS (RR>20 0))2

BS 0: the best probabilistic predictionBS 1: the worst probabilistic prediction

Brier skill score– extremesKurnik et al, 2011, submitted to IJC

Dry event RR > 200 mm

BSS=1- BScor / BSsim

BSS < 0: no improvementsBSS > 0: corrections improve predictions

Conclusions

• Various RCMs have been corrected, using same approach

• Bias correction is necessary, prior use of data in impact models – significant improvements

• Bias correction needs to be relatively “robust” • Dry months need to be studied carefully• Selection of validation technics is important (RMSE,

BS, BSS)