Experiences in Calibration of EPS forecasts
Juha KILPINEN
Finnish Meteorological Institute, Finland
Abstract
While evaluating the usefulness of ECMWF EPS (Ensemble Prediction System) forecasts several sources of uncertainty will be encountered. In this work the most difficult surface parameters like mslp, 10m wind speed and 2m temperature are considered.
EPS wind forecast are verified and compared with deterministic forecasts. A recursive calibration method is used calibrate the EPS wind speed forecasts and recursive error dressing method is used to produce probability forecasts from deterministic wind speed forecasts. Standard verification methods like Brier Skill Score and ROC area are used to compare the results of these methods. For comparison the verification is made both against original wind speed observations and corrected wind speed observations. The correction of wind speed observations takes into account the height of the instrument (using neutral stratification) and the reduction effect of surrounding obstacles (like buildings, towers, hills etc.).
The results indicate that EPS mslp forecasts are useful guidance without any calibration. However, other surface parameters suffer from same shortcomings as deterministic forecasts. There are systematic errors and noise due to imperfect parameterizations, not detailed enough orography etc. Calibration of EPS wind forecasts makes their value as guidance for e.g. wind warnings much higher.
Error dressing of deterministic wind forecasts provides useful probabilistic guidance for shorter lead times while calibrated EPS wind forecast are better in longer lead times. The results with original and corrected wind speed observation provide diverse results. For some stations correction also produces better forecasts but for some stations the results are less good. After further examination it was found out that for some stations the surroundings have changed but recalibration of correction coefficients has not been carried out.