160 likes | 310 Views
Vanessa Stauch Offenbach, September 2009. Postprocessing of temperature and wind for COSMO-7 and COSMO-2. COSMO General Meeting. calibration with Kalman Filter. >> recursive estimation of forecast error (prediction – correction) >> requires online observations
E N D
Vanessa Stauch Offenbach, September 2009 Postprocessing of temperature and wind for COSMO-7 and COSMO-2 COSMO General Meeting
calibration with Kalman Filter >> recursive estimation of forecast error (prediction – correction) >> requires online observations >> can be used quasi-instantaneously (no large historical database) >> cannot predict fast changes (assumption of persistent error) >> suitable for a subset of parameters (normally distributed errors)
error : ^ error model: with with states evolution: ^ prediction: calibration with Kalman Filter
COSMO models COSMO-LEPS COSMO-LEPS 10km, +132 hours COSMO-7 6.6km, +72 hours COSMO-2 2.2km, +24 hours COSMO-2 COSMO-7
Kalman Filter @ MeteoSwiss operational: T2m, TD2m for COSMO-LEPS mean COSMO-7 COSMO-2 IFS in preparation: FF10m, TW2m, RH2m for COSMO-LEPS mean COSMO-7 COSMO-2 IFS
Swiss met. measurement network 62 stations
T2m predictions COSMO-7 COSMO-7 COSMO-7 COSMO-7 COSMO-2 COSMO-2 KFC7 KFC2 performance?
benefit COSMO-2 vs COSMO-7? C2 vs C7 C2-KF vs C7-KF =
stations for wind speed calibration Schaffhausen (SHA) Chasseral (CHA) Üetliberg (UEB) Oron (ORO) Piz Martegnas (PMA) Gütsch (GUE) SMN station WKA Evionnaz (EVI)
SHA CHA/WiCro UEB WiFel WiGue ORO PMA EVI/WiCol height differences
represenativeness of met. station wind turbine Gütsch model prediction representative for (mean) grid box local point observation (specific conditions)
SHA CHA UEB PMA ORO GUE EVI COSMO-7 vs COSMO-2 rRMSE (%) für 1-24h, period 01.09.08 – 31.03.09
SHA CHA UEB PMA ORO GUE EVI effect on MOS-postprocessing rRMSE (%) für 1-24h, period 01.09.08 – 31.03.09
SHA CHA UEB PMA ORO GUE EVI effect on KF-postprocessing rRMSE (%) für 1-24h, Zeitraum 01.09.08 – 31.03.09
summary >> statistical postprocessing profits from a better NWP input model >> „dynamical downscaling“ does not replace statistical adaptation to local observations (in particular if results being verified against those)