Analysis of the Spread–Skill Relations Using the ECMWF Ensemble Prediction System over Europe

Abstract

The Ensemble Prediction System (EPS) of the European Centre for Medium-Range Weather Forecasts (ECMWF) was used to analyze various aspects of the ensemble-spread forecast-skill relation. It was shown that synoptic-scale upper-air spread measures can be used as first estimators of local forecast skill, although the relation was weaker than expected. The synoptic-scale spread measures were calculated based on upper-air fields (Z500 and T850) over western Europe for the period June 1997 to December 2000. The spread?skill relations for the operational ECMWF EPS were tested using several different spread definitions including a neural network-based measure. It was shown that spreads based on upper-air root-mean-square (rms) measures showed a strong seasonal cycle unlike anomaly correlation (AC)-based measures. The deseasonalized spread?skill correlations for the upper-air fields were found to be useful even for longer lead times (168?240 h). Roughly 68%?83% of small or large spread was linked to the corresponding high or low skill. A comparison with a perfect model approach showed the potential for improving the ECMWF EPS spread?skill relations by up to 25?30 correlation percentage points for long lead times. Local forecasts issued by operational forecasters for the Swiss Alpine region, as well as station precipitation forecasts for Geneva were used to test the limits of the synoptic-scale upper-air spread as an estimator of local surface skill. A weak relation was found for all upper-air spread measures used. Although the probabilistic EPS direct model precipitation forecast for Geneva exhibited a considerable bias, the spread?skill relation was recovered at least up to 144 h. A neural network downscaling technique was able to correct the precipitation forecast bias, but did not increase the synoptic-scale spread surface-skill relation.