ModelMIX Optimal combination of NWP Model Forecasts for AutoWARN Tamas Hirsch, Reinhold Hess, Sebastian Trepte, Cristina Primo, Jenny Glashoff, Bernhard Reichert, Dirk Heizenreder Deutscher Wetterdienst (DWD), Offenbach, Germany Research and Development (FEZE) World Weather Open Science Conference 2014 16 21 August 2014, Montreal, Canada
AutoWARN process AutoWARN - An automated decision support system for the weather warning service with the possibility of manual modification by the forecaster AutoWARN 2.0 planned for 2015 Observations Synop, MREP, etc. ASG ASE AutoWARN Status Editor PVW Automated Generation and Delivery of Warning Products NowCastMIX Integrated Nowcasting Product (1km) Automatic warn proposals Manually modifiable warn status Customer specified warning products ModelMIX MOS-based NWP (ensemble) post-processing (1km) 2
Requirements for ModelMIX Integrating as much valuable forecast information as possible Using several different NWP models, including ensemble models to take into account meteorological uncertainties Reducing model forecast errors Post-processing of direct model output needed Deriving parameters not forecast by NWP models E.g. probability of thunderstorms Best combined results Combination of all forecast information into a single best forecast product 3
Realization by the MOS approach MOS => Model Output Statistics Statistical connection between predictands (parameters of interest) and predictors (parameters with known values: raw model forecasts, obs, etc.) Raw model forecasts long historical data sets Observations long historical data sets Development Regression minimise RMSE MOS coefficients Operation Raw model forecasts current values Observations current values Post-processed forecasts 4
MOS systems at DWD Much experience with MOS at DWD Standard MOS, MOSMIX, Applied to weather warnings: WarnMOS (2010) Probability of warning events Extension to ensemble forecast: EnsembleMOS (R.Hess) Predictors: statistical properties of the whole ensemble distribution of model forecasts ensemble mean std. deviation percentiles First application to COSMO-DE-EPS (2013) 5
ModelMIX Process Post-Processing separately by WarnMOS / EnsembleMOS Combination for each model run, time step, parameter, season, etc. by MOSMIX ECMWF Global model (IFS) GME => ICON DWD current / future global model COSMO-DE-EPS DWD conv. scale Ens.Predict.System ECMWF-EPS ECMWF Ensemble Prediction System ICON-EPS DWD future global Ens.Predict.System ECMWF-WarnMOS GME-WarnMOS => ICON-WarnMOS COSMO-DE-EPS-WarnMOS ECMWF-EPS-WarnMOS ICON-EPS-WarnMOS ModelMIX Statist. optimal combination of Probability (warning events) with unified spatial and temp. resolution INPUT Raw model forecasts Post-processed model forecasts OUTPUT Best combined probability product 6
ModelMIX DWD warning events Priority of DWD warning event groups for ModelMIX Warning levels Wind gust 14-18 m/s 18-25 m/s 25-29 m/s 29-33 m/s 33-39 m/s > 39 m/s 1. Wind gust 2. Snowfall 3. Persistent rain 4. Freezing rain 5. Blowing snow (from snowfall) 6. Heavy rain (convective, short period) 7. Thunderstorm 8. Frost (T< 0 C) 9. Thaw (snow melt with rain) 10. Fog 7
ModelMIX Examples: SNOWFALL Probability of snowfall > 0.1 mm/6h GME + ECMWF (WarnMOS) ModelMIX 02 08 UTC 12 Feb 2014 Synop observations (WW) Area of snowfall at 02 UTC Tamas Hirsch, DWD WWOSC 2014 8
ModelMIX Examples: SNOWFALL Probability of snowfall > 0.1 mm/6h GME + ECMWF (WarnMOS) ModelMIX 02 08 UTC 12 Feb 2014 Synop observations (WW) Area of snowfall at 04 UTC Tamas Hirsch, DWD WWOSC 2014 9
ModelMIX Examples: SNOWFALL Probability of snowfall > 0.1 mm/6h GME + ECMWF (WarnMOS) ModelMIX 02 08 UTC 12 Feb 2014 Synop observations (WW) Area of snowfall at 06 UTC Tamas Hirsch, DWD WWOSC 2014 10
ModelMIX Examples: SNOWFALL Probability of snowfall > 0.1 mm/6h GME + ECMWF (WarnMOS) Mostly low probabilities ModelMIX COSMO-DE-EPS also included 02 08 UTC 12 Feb 2014 Clear signal for the right area Synop observations (WW) Area of snowfall at 08 UTC Tamas Hirsch, DWD WWOSC 2014 11
ModelMIX Examples: THUNDERSTORM Probability of thunderstorms at +16h GME + ECMWF (WarnMOS) COSMO-DE-EPS-WMOS ModelMIX Lightning measurements By combining all models Clear signal of possible thunderstorms near the right location even at +16h Tamas Hirsch, DWD WWOSC 2014 12
ModelMIX Verification Probability of wind gusts > 14 m/s +3h Reliability diagram Near diagonal (ideal) Significant improvement on raw model forecast Observed rel. frequency GME/ECMWF-WarnMOS COSMO-DE-EPS-WarnMOS Calibration of global/regional models COSMO-DE-EPS (raw) WarnMOS components combined into ModelMIX Underforecasting for low probabilities Forecast probability Overforecasting for high probabilities 13
ModelMIX - Improvements Improved MOS (regression) Improved combination of probabilities ECMWF Global model (IFS) GME => ICON (2014/2015) DWD current / future global model COSMO-DE-EPS DWD conv. scale Ens.Predict.System ECMWF-EPS (2014) ECMWF Ensemble Prediction System ICON-EPS ( ) DWD future global Ens.Predict.System ECMWF-WarnMOS GME-WarnMOS => ICON-WarnMOS COSMO-DE-EPS-WarnMOS ECMWF-EPS-WarnMOS ICON-EPS-WarnMOS ModelMIX Statist. optimal combination of Probability (warning events) with unified spatial and temp. resolution GFS EFS? NCEP Glob. Ensemble Forecast System MOGREPS? Met Office Ensemble Prediction System CMC EPS? Canadian Met.Center Ens. Pred. System 14
ModelMIX New observational data sets Essential for MOS development and also for appropriate verification of results NOW: observations mostly from surface stations (synop, ) NEW: use of high resolution observational datasets Radar estimated accumulated precipitation 20120911 1850 UTC Direct calculation of probabilities Improved representativity More extreme cases might be identified Better training data set for development / verification 15
ModelMIX Summary Post-processed, combined NWP model forecasts MOS approach, WarnMOS, EnsembleMOS NWP Models used: global, regional, deterministic, ensemble Examples of resulting probability fields for warning events Improvements in progress or planned Using new and improved NWP models Involving new high resolution observational data sets Improved MOS and combination of probabilities Verification Improvement on raw (uncalibrated) NWP model forecasts Currently for stations, to be extended to grid and to more parameters Presentation on the use of ModelMIX forecasts in project AutoWARN G. Schröder, B. Reichert, D. Heizenreder: Automated weather warning proposals based on post-processed numerical weather forecasts (19 August 2014, 17:40-18:00) 16
Thank you very much for your attention! Tamas Hirsch, DWD WWOSC 2014 17