Presentation on theme: "ECMWF long range forecast systems"— Presentation transcript:
1 ECMWF long range forecast systems Dr. Tim StockdaleEuropean Centre for Medium-Range Weather Forecasts
2 Outline Overview of System 4 Some recent research results EUROSIP multi-model forecastsForecasts for JJA 2013Seasonal prediction at ECMWFHistory/approachStrengths and weaknesses of S4History of EUROSIP, including theory of multi-modelInc DEMETER results as motivationAlso theoretical basis for mm averagingPractical considerations – real-time risks, confidence etcNCEP joiningMention Paco’s assessmentDifficulty of 14 years for assessmentShow some results from LauraNew suite and productsMention that NCEP processing suite now operationalNew multi-model suite still in esuite – want to be sure everything is complete, and manage handover for MSExample plots, esp inc Nino pdfs.SummaryOperational system, plus benefits for research, some already, more for future. Plus meeting together, discussing, sharing expertise etc.
3 System 4 seasonal forecast model IFS (atmosphere)TL255L91 Cy36r4, 0.7 deg grid for physics (operational in Dec 2010)Full stratosphere, enhanced stratospheric physicsSingular vectors from EPS system to perturb atmosphere initial conditionsOcean currents coupled to atmosphere boundary layer calculationsNEMO (ocean)Global ocean model, 1x1 resolution, 0.3 meridional near equatorNEMOVAR (3D-Var) analyses, newly developed.CouplingFully coupled, no flux adjustmentsSea-ice based on sampling previous five years
4 Reduced mean state errors JJA T850 bias (left), DJF U50 – stratosphere – right.S4 bias (top), S3 (bottom) – on same scale!!Note for T850, much reduced error in structure of T850 field over N America And surrounding oceansThese plots are typical of very many fields – almost all aspects of climate improved. Almost all!
5 Tropospheric scores One month lead Four month lead Spatially averaged grid-point temporal ACCOne month leadFour month lead
6 S4 extended hindcast set Scores are smoother and systematically higher with 51 member hindcasts
7 S4 extended hindcast set Gain over S3 is now stronger and more robust
15 QBOA big reduction in vertical diffusion, and a further tuning of non-orographic GWD, has given a big additional improvement in the QBO compared to S4.Period and downward penetration match observationsSemi-annual oscillation still poorly represented
18 NH winter forecastsEven with 101 members, ensemble mean signal not always well defined
19 NH winter forecastsNew version has weaker signal, more noise
20 NH winter forecastsForecast skill is above perfect model predictability limit
21 EUROSIP A European multi-model seasonal forecast system Operational since 2005Data archive and real-time forecast productsInitial partners: ECMWF, Met Office, Météo-FranceNCEP an Associate Partner; forecasts included since 2012Products released at 12Z on the 15th of each monthAim is a high quality operational systemData policy issues are always a factor in EuropeInitially, all models ran at ECMWF, comprehensive data archive; now, some models run remotely with reduced data volumesAlso more diversity in forecast/re-forecast strategy
22 Recent changes: variance scaling Robust implementationLimit to maximum scaling (1.4)Weakened upscaling for very large anomaliesImproves every individual modelImproves consistency between modelsImproves accuracy of multi-model ensemble mean
25 Calibrated p.d.f. ENSO forecasts have good past performance data We can calibrate forecast spread based on past performanceWe can also allow varying weights for modelsWe have to be very careful not to overfit data at any point.Represent forecast with a p.d.f.This is the natural output of our calibration procedureEasier visual interpretation by userCalibration and combination in general caseIdeally apply similar techniques to all forecast values (T2m maps etc)More difficult because less information on past (higher noise levels)Hope to get there eventually ….. .
27 P.d.f. interpretation P.d.f. based on past errors Bayesian p.d.f. The risk of a real-time forecast having a new category of error is not accounted for. E.g. Tambora volcanic eruption.We plot 2% and 98%ile. Would not go beyond this in tails.Risk of change in bias in real-time forecast relative to re-forecast.Bayesian p.d.f.Explicitly models uncertainty coming from errors in forecasting systemTwo different systems will calculate different pdf’s – both are correctValidationRank histograms show pdf’s are remarkably accurate (cross-validated)Verifying different periods shows relative bias of different periods can distort pdf – sampling issue in our validation data.