170 likes | 311 Views
Cost Efficient Use of COSMO-LEPS Reforecasts. Felix Fundel, André Walser , Mark Liniger, Christof Appenzeller. Motivation: Reforecast Bias 1971-2000. Temperature. Precipitation. too wet. too warm. Motivation: Reforecast Bias 1971-2000. Temperature. Precipitation.
E N D
Cost Efficient Use of COSMO-LEPS Reforecasts Felix Fundel, André Walser, Mark Liniger, Christof Appenzeller
Motivation: Reforecast Bias 1971-2000 Temperature Precipitation too wet too warm
Motivation: Reforecast Bias 1971-2000 Temperature Precipitation • Calibrating with reforecast reduces the forecast bias • Ensemble forecasts become more reliable • Shown for up to date global and regional EPS
Outline • Status of the reforecast project • New results • Outlook and open issues
Status of the reforecast project • Research-Program has finished • Work in progress to make calibrated COSMO-LEPS products operational • Large reforecast data set available, also used in other research projects
Reforecasts Data Set Intention: statistically independent data for a long period • reforecasts over a period of 30 years (1971-2000) • deterministic run of COSMO-LEPS v 4.0 convection and physics randomly switched • ERA40 reanalysis as initial/boundary fields • 90h lead time, every 3rd day 1971-01-01 1971-01-03 1971-01-06 LT: 90h Available at ECMWF mars archive: . . . retrieve, time=12:00:00, date=2008-03-10, stream=enfh, step=0/to/90, levtype=sfc, expver=2, type=cf, hdate=1971-03-10/to/2000-03-10 class=co, param=61.2 2000-12-27 2000-12-29 2000-12-31
Gridded observations for Switzerland • 24h total precipitation • domain Switzerland • COSMO-LEPS grid (SYMAP, C. Frei) • >450 Observations
Calibration Calibration method: - CDF mapping - each grid point calibrated individually - using gridded observations over Switzerland winter precip. summer precip. Brier Skill Score day 1 day 2 day 3 day 4 day 1 day 2 day 3 day 4
However… • Reforecasts are very expensive • cost for 2 years reforecasts = cost for 1 forecast member (current setup, reforecast every 3rd day, 90h lead-time) What‘s the skill of COSMO-LEPS with 16-n forecast members, calibrated with 2xn years of reforecasts?
Isocost analysis - Precipitation RPS Summer-Fall 2008 RPS Winter-Spring 2007/2008 Best setup: 16 member & 30y reforecasts Best isocost setup: 11 member & 10y reforecasts
Isocost analysis - Precipitation 2AFC sore Summer-Fall 2008 2AFC score Winter-Spring 2007/2008
Isocost analysis – T2m Based on 61 SwissMetNet stations Calibration using reforecasts and observations from 1981-2000 No seasonal or lead-time stratification RPS Winter-Spring Dec 07-Nov 08 2AFC Winter-Spring Dec 07-Nov 08
Isocost analysis – T2m 16 member DMO DMO calibr. w. 1y refcst DMO calibr. w. 20y refcst Calibrating T2m improves ensemble spread Most efficient around noon Still not enough to correct under-dispersiveness
Conclusion • Forecast skill can be improved without additional (CPU) cost improved reliability, small loss in resolution • Optimal COSMO-LEPS setup ~12 members / 8 reforecast years (for 24h precipitation and temperature) But: • Some applications require large ensembles • Corresponding observations needed for calibrated values
Outlook & open issues • Adapt reforecast suite for 7 km COSMO-LEPS suite, but......how should we initialize the soil of these reforecasts? • How perform alternative (and cheaper) post-processing methods vs. our calibrated reforecasts? CONSENS • Should the reforecast project become part of the operational COSMO-LEPS suite of COSMO?