caio a. s. coelho centro de previs ã o de tempo e estudos clim á ticos (cptec)
DESCRIPTION
An intercomparison of multivariate regression methods for the calibration and combination of seasonal climate predictions. Caio A. S. Coelho Centro de Previs ã o de Tempo e Estudos Clim á ticos (CPTEC) Instituto Nacional de Pesquisas Espaciais (INPE) [email protected]. - PowerPoint PPT PresentationTRANSCRIPT
Caio A. S. CoelhoCentro de Previsão de Tempo e Estudos Climáticos (CPTEC)
Instituto Nacional de Pesquisas Espaciais (INPE)[email protected]
11th International Meeting on Statistical Climatology Edinburgh, 12-16 July 2010
PLAN OF TALK• Motivation• Datasets and regression methods• Skill assessment• Summary
An intercomparison of multivariate regression methods for the calibration and
combination of seasonal climate predictions
Thanks to David Stephenson
Framework for calibration and combination of climate predictions
)y|x(p ii
Data Assimilation “Forecast Assimilation”)x|y(p ff
Stephenson et al. (2005)Tellus A, 57(3), 253-264
Multi-model ensemble approach
ENSEMBLESENSEMBLE-based Predictions of Climate Changes
and their Impacts
Solution: Multi-model Ensemble
Errors: Model formulationInitial conditions
http://www.ecmwf.int/research/EU_projects/ENSEMBLES/
Hindcast period: 1960-2005 (46 years)
ENSEMBLES multi-model seasonal predictions
Coupled model Country ECMWF International Meteo-France France INGV Italy IFM-Kiel Germany UK Met Office U.K.
1-month leadprecip. predictionsfor DJFover S. America (i.e. issued in Nov)
9 ens memb. eachtotal: 45 members
http://www.ecmwf.int/research/EU_projects/ENSEMBLES/
Multivariate regression model for thecalibration and combination of climate predictions
Y|X ~ N (L (X - Xo),D)
TYZ
1XXYXYY
o
1XXYX
SSSSD
LXYXL
SSL
Y: DJF precipitation X: DJF precipitation predictions
pn:Xqn:Y
qq:D
Use PCs of X: Principal component (PC) regression
Use Maximum Covariance Analysis (MCA) modes of YT X: MCA regression
Multivariate regression model for thecalibration and combination of climate predictions
Y|X ~ N (L (X - Xo),D)
YY
o
XXYX
SDLXYXL
SSL
1
Y: DJF precipitation X: DJF precipitation predictions
Use PCs of X: Ridge principal component regression
pn:Xqn:Y
qq:D 1
pXXYX )IS(SL
Taking advantage of forecast skill over the Pacificto improve forecasts over land
Source: Franco Molteni (ECMWF)
Y
X
Cross validated skill assessment
Ridge PC regr. PC regressionMCA regression
Correlation maps: DJF precip. anomalies
Hindcast period: 1960-2005 (46 years)
6 PCs 3 MCAs All PCs
Ridge PC shows improved skill in central South AmericaPC and MCA regression show improved skill in SE South America
PC regression
Hindcast period: 1960-2005 (46 years)
6 PCs 3 MCAs
Gerrity score for DJF tercile precip. categories
Ridge PC regr. MCA regression
All PCs
PC regression
Hindcast period: 1960-2005 (46 years)
6 PCs 3 MCAs
ROC skill score for DJF positive anomalies
Ridge PC regr. MCA regression
All PCs
Summary• Multivariate regression is a powerful tool for the calibration and
combination of multi-model ensemble predictions
• Ridge principal component regression allows incorporation of full forecast variability in the calibration and combination procedure (advantage to PC and MCA regression that require truncation)
South American austral summer predictions:
• Principal component regression requires retaining more modes to achieve similar level of skill to MCA regression
• Over Central South America ridge principal component regression shows improved skill compared to PC and MCA regression
• Over SE South America PC and MCA regression show improved skill compared to ridge principal component regression