CoupledCoupled SeasonalSeasonal ForecastingForecasting atat CCCmaCCCma

Bill Merryfield, George Boer, Greg Flato, Slava Kharin, Badal Pal, John Scinocca

Canadian Centre for Climate Modelling and Analysis

Funded by the Canadian Foundation for Climate and Atmospheric Sciences Dynamical Seasonal Forecasting in Canada

Canadian Clivar Network 2001-2007 • Historical Forecast Project (HFP) - 6-run ensembles, 2 models (GCM2, SEF) - persisted SSTs - Current operational system (SEF → GEM in 2004) • HFP2 - 10-run ensembles, 4 models (GCM2, GCM3, SEF, GEM) - persisted SSTs - Operational Fall 2007 • Begin development of coupled forecast capability at CCCma - prognostic SSTs , enabling longer leads GOAPP • Coupled Historical Forecast Project (CHFP ) Dynamical Seasonal Forecasting in Canada

Canadian Clivar Network 2001-2007 • Historical Forecast Project (HFP) - 6-run ensembles, 2 models (GCM2, SEF) - persisted SSTs → 3 month outlook - Current operational system (SEF GEM in 2004) lead 0 mo • HFP2 - 10-run ensembles, 4 models (GCM2, GCM3, SEF, GEM) - persisted SSTs 3 month outlook - Operational Fall 2007 lead 0, 1 mo • Initial development of coupled forecast capability at CCCma - prognostic SSTs , enabling longer leads GOAPP • Coupled Historical Forecast Project (CHFP ) Observed SST autocorrelation time scale

Months before autocorrelation < 0.5 : Feb May

2 3 4 5 6 7 2 3 4 5 6 7 Aug Nov

2 3 4 5 6 7 2 3 4 5 6 7 months months Goddard & Mason ( Clim Dyn 2002) CHFP Role within GOAPP

I.2.1 Indep Coupled Assimilation I.1.3 I.1.4 Multivar Ocean Ocn Reanalysis I.2.1 Assimilation & Forecasting Joint Coupled Assimilation

II.2.1 II.2.2 II.2.3 Coupled Model CHFP Forecast Initialization Calibration CHFP Role within GOAPP

I.2.1 Indep Coupled Assimilation I.1.3 I.1.4 Multivar Ocean Ocn Reanalysis I.2.1 Assimilation & Forecasting Joint Coupled Future Assimilation Operational System?

II.2.1 II.2.2 II.2.3 Coupled Model CHFP Forecast Initialization Calibration CHFP Prototype

• CGCM3.1 (T63) • Ocean initialization via SST “nudging” to obs τττ=10d • Initialize 1 Sep, 1 Dec, 1 Mar, 1 Jun 1972/3-2001/2 • 10-run ensembles: 1 Sep nudging assimilation run

Aug 25 26 27 28 29 30 31 Forecast 1 12 mos AGCM Forecast 2 Forecast 3

OGCM …

mix & match Forecast 10 lead 0 lead 1 lead 2 forecast runs Forecast methodology/verification

• Bias correction from 30y forecast

• Verify by computing Mean sq fcst error - anomaly correlation σσσ2 - mean-square skill score: MSSS= 1 – __fcst →→→ > 0 if more skillful than σσσ2 clim climatological forecast Mean sq error of - % correct of categorical forecasts climatological fcst - Brier skill score, etc.

c August Prototype forecast results Initialization Anomaly correlation MSSS

October Lead 1

December Lead 3

February Lead 5 November Prototype forecast results Initialization Anomaly correlation MSSS

January Lead 1

March Lead 3

May Lead 5 Prototype forecast results NINO3.4 “plume” Prototype forecast results NINO3.4 “plume” Init Nov 1992

ensemble ensemble obs mean

C) members ° ° ° °

assim

NINO3.4 anomaly anomaly ( NINO3.4 persistence

0 2 4 6 8 10 Lead Surface temperature forecast for Canada Lead 1 Proto-CHFP MSSS HFP2 MSSS

OND forecast

0.0475 0.0028

JFM forecast

0.2802 0.0093 Surface temperature forecast for Canada Lead 1 Proto-CHFP MSSS HFP2 MSSS

OND forecast

0.0475 CHFP Wins!0.0028

JFM forecast

0.2802 0.0093 Surface temperature forecast for Canada Lead 0 Proto-CHFP MSSS HFP2 MSSS

SON forecast

0.0207 0.0511

DJF forecast

-0.2716 0.0982 Why proto-CHFP loses to HFP2 at lead 0

• HFP2 initialized by →close to “truth” • Proto-CHFP atmospheric IC have incorrect “weather” σσσ2 σσσ (T ran -Tobs ) p(T) T σσσ2 σσσ 2 = (T ran )+ (T obs ) σσσ2 = 2 T _ _ T ran Tclim Tobs T σσσ2 MSSS = 1- __T clim σσσ2 = 0 T 2σσσ2 MSSS = 1- __T ran σσσ2 = -1 T Surface temperature forecast for Canada TAS Canada, cangrd, 1972−2001

0.6 IC=AUG IC=DEC 0.5 0.4 0.3 0.2 0.1 0.0 −0.1 −0.2 −0.3 Corr MSSS CHFP HFP2 −0.4 JJA DJF MJJ JAS NDJ JFM AMJ FMA ASO SON OND MAM Surface temperature forecast for Canada TAS Canada, cangrd, 1972−2001

0.6 IC=AUG IC=DEC 0.5 0.4 0.3 0.2 0.1 0.0 −0.1 −0.2 −0.3 Corr MSSS CHFP HFP2 −0.4 JJA DJF MJJ JAS NDJ JFM AMJ FMA ASO SON OND MAM Proto-CHFP vs HFP2

• Proto-CHFP outperforms HFP2 @ lead 1, even though - 10 run ××× 1 model ensemble, vs 10 run ××× 4 models - simple ocean initialization - forecast NINO3.4 beaten by persistence until lead ≈ 7 - CGCM3.1 ENSO SST variability weak, concentrated in central rather then E Pacific

c Standard deviation of anomalous monthly SST

Observations 1950-99 CGCM3.1

AGCM3 + OGCM4 CGCM4 Regression of annual mean SST vs NINO3.4

Observations 1950-99 CGCM3.1

AGCM3 + OGCM4 CGCM4 What’s next • Newer CGCM version • Initialize AGCM with obs as in HFP, HFP2 • Ocean - 2D Var (Tang et al. JGR 2004) - Semi-prognostic method (Greatbatch et al. Ocean Modelling 2006) - Spectral nudging (Thompson et al. Ocean Modelling 2006)

• Multiple input reanalyses: NCEP SODA GFDL +Theme I? What’s next • Newer CGCM version • Initialize AGCM with obs as in HFP, HFP2 • Ocean data assimilation - 2D Var (Tang et al. JGR 2004) - Semi-prognostic method (Greatbatch et al. Ocean Modelling 2006) - Spectral nudging (Thompson et al. Ocean Modelling 2006)

• Multiple input reanalyses: NCEP SODA GFDL +Theme I? • Task Force of Seasonal Prediction → TFSP Experiment : multi-model ensemble of retrospective coupled forecasts

TAS glb, era40, 1972−2001

0.6 IC=AUG IC=DEC 0.5

0.4

0.3

0.2

0.1

0.0

−0.1

−0.2 Corr MSSS CHFP HFP2 −0.3 JJA MJJ DJF JAS JFM NDJ AMJ FMA ASO SON OND MAM Prototype SST forecast vs persistence November Coupled forecast MSSS Persistence MSSS Initialization January Lead 1

March Lead 3

May Lead 5