Seasonal Forecasting From DEMETER to ENSEMBLES Francisco J. Doblas-Reyes ECMWF
DEMETER • Coupled ocean-atmosphere systems without assimilation of sub-surface ocean observations. • Multi-model (ECMWF, GloSea, Météo-France, IfM-Kiel, CERFACS, INGV, LODYC) ensemble re-forecasts. • Re-forecast period 1959-2001, seasonal (6 months, February, May, August and November start date), 9-member ensembles, ERA40 initialization in most cases.
ENSEMBLES • Model uncertainty is a major source of forecast error. Three approaches to deal with model uncertainty are being investigated in ENSEMBLES: multi-model (ECMWF, GloSea, DePreSys, Météo-France, IfM-Kiel, CERFACS, INGV), stochastic physics (ECMWF) and perturbed parameters (DePreSys). • Hindcasts in two streams: • Stream 1: hindcast period 1991-2001, seasonal (7 months, May and November start date), annual (14 months, November start date), 9-member ensembles, ERA40 initialization in most cases; DePreSys (IC and PP ensembles) 10-year runs in every instance. • Stream 2: As in Stream 1 but over 1960-2005, with 4 start dates for seasonal hindcasts, at least 1 for annual and at least one 3-member decadal hindcast every 5 years; DePreSys 10-year runs once a year and 30-year runs every 5 years.
Assume a multi-model ensemble system with coupled initialized GCMs Lead time = 7 Ensemble climate forecast systems Model 1 Model 2 Model 3Model 4Model 5Model 6 K models x M ensemble members M*K-member ensemble Feb 87 May 87 Aug 87 Nov 87 Feb 88 ...
Assume a multi-model ensemble system with coupled initialized GCMs Lead time = 4 Ensemble climate forecast systems Model 1 Model 2 Model 3Model 4Model 5Model 6 Feb 87 May 87 Aug 87 Nov 87 Feb 88 ...
Differences in climatological pdfs Reference pdf Model pdf Actual occurrences t=2 Different variabilities Mean bias t=1 t=3 Temperature Threshold Systematic error in ensemble forecasts Main systematic errors in dynamical climate forecasts: • Differences between the model climatological pdf (computed for a lead time from all start dates and ensemble members) and the reference climatological pdf (for the corresponding times of the reference dataset): systematic errors in mean and variability. • Conditional biases in the forecast pdf: errors in conditional probabilities implying that probability forecasts are not trustworthy. This type of systematic error is best assessed using the reliability diagram. Forecast PDF
Systematic error in seasonal forecasts ENSEMBLES Stream 2 T2m mean bias wrt ERA40/OPS, 1960-2005 First month May Months 2-4 JJA Months 5-7 SON ECMWF Météo-France
Attributes diagram Attributes diagrams for 1-month lead seasonal (JJA) precipitation above the upper tercile over the tropical band for the ENSEMBLES Stream 1 multi-model (left, 45 members), stochastic physics (centre, 9 members) and perturbed parameters (right, 9 members) hindcasts started in May over the period 1991-2001 verified against GPCP. The Brier and ROC skill scores, along with 95% confidence intervals (in brackets) computed using a bootstrap method, are shown on top of each panel. Stochastic physics Perturbed parameters Multi-model 0.129 (0.082,0.178) 0.441 (0.378,0.502) 0.059 (0.005,0.105) 0.391 (0.322,0.453) 0.050 (0.002,0.105) 0.381 (0.329,0.443)
Direct model output Reliability diagrams for 1-month lead seasonal (JJA) precipitation above the upper tercile over the tropical band for the ENSEMBLES Stream 1 multi-model (left, 45 members), stochastic physics (centre, 9 members) and perturbed parameters (right, 9 members) hindcasts over the period 1991-2001 verified against GPCP. Direct model output (no bias correction) and threshold (upper tercile) computed from the reference climatology Stochastic physics Perturbed parameters Multi-model
ROCSS for anomalies above the upper tercile Anomaly correlation coefficient Ratio between spread and RMSE Stream 1 seasonal hindcasts Scores for southern South America precipitation from Stochastic Physics, Perturbed Parameters (both with 9-member ensembles) and Multi-model (5 models, 45 members). Sample values are shown with black dots along with 95% confidence intervals obtained using a bootstrap method (verified against GPCP over 1991-2001).
Stream 2 seasonal hindcasts Brier skill score for several regions (Northern Hemisphere, Tropics, Southern Hemisphere), events (anomalies above/below the upper/lower tercile),lead times (2-4, 5-7 months), start dates (Feb, May, Aug and Nov) and variables (near-surface temperature, precipitation, Z500, T850 and MSLP) computed over the period 1960-2005. The inset numbers indicate the number of cases where a system is superior. 3 64 237 176
DEMETER vs ENSEMBLES Brier skill score for Niño3 SST re-forecasts in DEMETER, ENSEMBLES and DEMETER+ENSEMBLES using all start dates over the period 1980-2001. x’<lower tercile x’>upper tercile Forecast period 2-4 months Forecast period 4-6 months
DEMETER+ENSEMBLES ENSEMBLES DEMETER vs ENSEMBLES Brier skill score for re-forecasts of near-surface temperature and precipitation for different land regions over the period 1980-2001. DEMETER
EUROBRISA at IC3 • The Catalan Institute for Climate Sciences (IC3, Barcelona, Spain) will start this winter a new group on seasonal and interannual climate forecasting. • The main goals are to develop a capability to perform research on climate forecasting and to work on methods that provide useful climate information; the target regions are the Mediterranean area, South America and Africa. • A solid link to EUROBRISA is expected at IC3. Two members of the group will work on seasonal forecasting calibration and combination with a focus on the Mediterranean region and the calibration of forecasts using non-stationary series (in the presence of trends and low-frequency variability).
Summary • Substantial systematic error, including lack of reliability, is still a fundamental problem in dynamical seasonal and interannual forecasting and forces a posteriori corrections to obtain useful predictions. • Comprehensive assessments of the forecast quality measures (including estimates of their standard error) are indispensable in forecast system comparisons. • Perturbed-parameter ensembles are competitive with multi-model ensembles. • The ENSEMBLES multi-model is marginally better than the DEMETER multi-model; much is still to be gained from robust calibration and combination.