Deterministic evaluation of ensemble streamflow predictions in Sweden

Abstract

A system for ensemble streamflow prediction, ESP, has been operational at SMHI since July 2004, based on 50 meteorological ensemble forecasts from ECMWF. Hydrological ensemble forecasts are produced daily for 51 basins in Sweden. All ensemble members, as well as statistics (minimum, 25% quartile, median, 75% quartile and maximum), are stored in a database. This paper presents an evaluation of the first 18 months of ESP median forecasts from this system, and in particular their performance in comparison with today's categorical forecast. The evaluation was made in terms of three statistical measures: bias B, root mean square error RMSE and absolute peak flow error PE. For ESP forecasts the bias ranged between -20% and 80% with a systematic overestimation for Sweden as a whole. A comparison between bias in input precipitation and ESP output, respectively, revealed only a weak relationship, but streamflow overestimation is likely related mainly to model properties. The results from the streamflow forecast comparison showed that the ESP median in deterministic terms performs overall as well as the presently used categorical forecast. Further, ESP has the advantage of providing at least a qualitative measure of the uncertainty in the forecasts, with probability forecasts being the ultimate goal.