The Ensemble-MOS of Deutscher Wetterdienst 15th EMS/12th ECAM Session ECAM1 07 September 2015 Reinhold Hess, Jenny Glashoff, and Bernhard K. Reichert Deutscher Wetterdienst
The Ensemble-MOS of Deutscher Wetterdienst Outline Motivation: AutoWARN, operational MOS Systems at DWD Ensemble MOS: basic idea and current setup Application to COSMO-DE-EPS and ECMWF-EPS ModelMIX: MOS of MOS Reinhold Hess, FEZE-B 2
Ensemble-MOS and ModelMIX Basis of the Decision Support System AutoWARN for the Weather Warning Service at DWD Generation of Automated Warning Proposals for Forecasters Ensemble-MOS: MOS for ensembles: designed for the calibration of probabilistic forecasts ModelMIX: Consistent combination of individual MOS Systems for ICON, IFS/ECMWF, COSMO-DE-EPS, IFS-EPS ICON-EPS (DWD) B.K. Reichert et al. Reinhold Hess, FEZE-B 3
MOS Systems at DWD MOS (Model Output Statistics): stepwise multiple regression of predictors (known values as model forecasts and latest observations) to predictands (unknown values as observations and derived elements at valid times) Operational Systems at DWD MOS-MIX: ICON-MOS, ECMWF-MOS global, up to 240h, at synoptic stations, based on ICON and IFS/ECMWF ICON-WarnMOS, ECMWF-WarnMOS, WarnMOS-MIX provides 27 warning criteria on 1x1 km grid for Germany AUTO-TAF spezialised forecasts for airports CellMOS nowcasting thunderstorms on advecting cells (Lagrange) Ensemble-MOS, ModelMIX (in development, based on WarnMOS) calibration of ensemble forecasts (COSMO-DE-EPS, ECMWF-EPS, ICON-EPS) Reinhold Hess, FEZE-B 4
Ensemble-MOS basic idea: exploit probabilistic information of ensembles for event probabilities statistical optimisation of ensemble forecasts using synoptical observations optimisation and interpretation of deterministic variables (e.g. visibility) calibration of event probabilities (e.g. prob(rr/1h>15mm) ) forecast of forecast errors (e.g. absolute error of 2m temperature) apply for COSMO-DE-EPS and ECMWF-EPS (ICON-EPS, planned) Reinhold Hess, FEZE-B 5
Ensemble-MOS efforts: ensemble products as model predictors: mean and stddev (quantiles, etc.) surrounding of stations (mean and stddev of surrounding) linear regression for deterministic forecast elements (e.g. 2m temperature) logistic regression for probabilistic forecast elements (e.g. prob(rr/1h>15mm) ) forecast of forecast errors and forecast uncertainty maximize number of extreme and rare events (wind gusts, heavy precipitation) use of long time series, e.g. 4 years for COSMO-DE-EPS multi station approach (9 climatological clusters in Germany) multi time equations (up to +/- 9 hours depending on rareness) gauge adjusted radar data additionally to precipitation observations Reinhold Hess, FEZE-B 6
Gauge adjusted radar products as predictands (T. Hirsch) Probabilities of Precipitation Prob(RR>15 mm/ 1h) Prob(RR>40 mm/12h) standard: synoptic observations idea: use radar-data (RW, 1x1 km) surrounding of stations (r=8 km und 40 km) relative frequencies of threshold exceedances in surrounding improved statistical sample higher representativity more extreme cases 1-hourly estimation of precipitation (gauge adjusted at stations) Reinhold Hess, FEZE-B 7
Application to COSMO-DE-EPS MOS: stepwise multiple linear and logistic regression provide set of predictors, e.g. model forecasts, observations, derived predictors (e.g. sqrt RR, Rel_Div_10m, CAPE index, etc.) select predictor with highest statistical correlation to predictand select further predictors correlated to residuum, as long as statistically significant example: 2m temperature based on 3 UTC issue of COSMO-DE-EPS 99903 Issue=02:00z +001:00 TTT Season: spr _MS: medium scale: 28 km _LS: large scale: 54 km Name Lin Reg 1 Co Wgt ----------------------------------- T_G_MS 0.05 5 TD_2M_LS -0.02 2 TTT(-1)Obs 0.88 78 Td(+0)StF 0.08 7 ----------------------------------- Const. = -5.2 RMSE = 6.28 Co: coefficient of regression Wgt: normalised weigth of predictor in equation forecast time: 1h 1 equation for each predictand (about 160), cluster (9), forecast time (21), season (4), issue of EPS(8) Reinhold Hess, FEZE-B 8
Application to COSMO-DE-EPS Calibration of wind gusts > 14m/s Impact of logistic regression reliability diagram for 3-hourly forecasts COSMO-DE-EPS (not calibrated, grey) shows low resolution and significant overforecasting for high probabilities. MOS with linear regression (blue, green) shows underforecasting for high probabilites. Ensemble MOS with logistic regression (red) is correcting, however not yet perfectly. Still overforecasting for small probabilities (problem found). C. Primo Reinhold Hess, FEZE-B 9
Application to ECMWF-EPS EnsembleMOS for ECMWF-EPS (J. Glashoff) Verification of 2013 2m temperature errors : MOS Forecasts of MOS Errors : MOS Errors Frankfurt TIGGE/THORPEX data 50 ensembles, 1 high resolution run 2m temperature, mean wind, cloud coverage, 24h precipitation observations as predictands ensemble products, mean, stddev as predictors training sample 2002-2012 free forecasts for 2013 C. Primo Reinhold Hess, FEZE-B 10
Application to ECMWF-EPS Forecast of Forecast Errors (of MOS forecast) compute regression of any forecast element use absolute error of residuum as predictand compute regression of this predictand example: absolute error of 2m temperature based on 12 UTC issue of ECMWF-EPS (TIGGE/THORPEX data, 4 variables) 99906 Issue=12:00z +024:00 TTT Season: win Name Lin Reg 1 Co Wgt ----------------------------------- temp 1.04 88 Cos_Dag 0.07 3 temp_dev 0.04 3 Sin_3*Dag 0.01 3 wind -0.16 3 ----------------------------------- Const. = -2.5 RMSE = 10.90 T2m temp: ensemble mean of temperature temp_dev: standard deviation of ensemble 99906 Issue=12:00z +024:00 E_TTT Season: win Name Lin Reg 1 Co Wgt ----------------------------------- temp_dev 0.05 37 wind -0.19 29 temp -0.02 18 Sin_3*Dag -0.01 16 ----------------------------------- Const. = 9.3 RMSE = 7.10 abs. err. of T2m stddev is increased and calibrated (stddev = absolute error/0.8 for Gaussian distribution) Reinhold Hess, FEZE-B 11
Application to ECMWF-EPS EnsembleMOS for ECMWF-EPS (J. Glashoff) Verification of 2013 24h precipitation errors Frankfurt C. Primo Dublin C. Primo Reinhold Hess, FEZE-B 12
ModelMix MOS of MOS combination of MOS systems ICON-WarnMOS ECMWF-WarnMOS COSMO-DE-EPS-EnsembleMOS ECMWF-EPS-EnsembleMOS ICON-EPS-EnsembleMOS statistically optimal combinations consistent probabilitstic products for warning criteria at locations of stations and on 1km-grid basis for the automated warning proposals for AutoWARN Reinhold Hess, FEZE-B 13
ModelMIX: Thunderstorm Probability for thunderstorm +16h oper. WarnMOS (GME + ECMWF) COSMO-DE-EPS-MOS ModelMIX Thunder T. Hirsch Combination of MOS forecasts signal for thunderstorm is enhanced Reinhold Hess, FEZE-B 14
ModelMIX normalised weights for COSMO-DE-EPS-EmsebleMOS, GME-WarnMOS und ECMWF-WarnMOS (all issues, seasons, stations) consistent transition from COSMO-DE-EPS to GME/ICON and ECMWF Reinhold Hess, FEZE-B 15
Application to ECMWF-EPS EnsembleMOS for ECMWF-EPS Verification of 2013 24h precipitation errors Frankfurt C. Primo Dublin C. Primo Reinhold Hess, FEZE-B 16
Thank you for attention References: Knüpffer, K., 1996. Methododical and predictability aspects of MOS systems, in: 13th Conf. on Probability and Statistics in Atmosph. Sciences, Amer. Meteorol. Soc., San Francisco, CA. pp. 190 197. Kriesche, B., R. Hess, B. K. Reichert, and V. Schmidt (2015). A probabilistic approach to the prediction of area weather events, applied to precipitation. Spat. Stat. 12, 15 30. Peralta, C., Ben Bouallègue, Z., Theis, S.E., Gebhardt, C. and M. Buchhold (2012): Accounting for initial condition uncertainties in COSMO-DE-EPS. J. Geophys. Res., 117 (D7) Reinhold Hess, FEZE-B 17