Combining Postprocessed Ensemble Weather Forecasts and Multiple Hydrological Models for Ensemble Streamflow Predictions

Abstract

Ensemble streamflow prediction (ESP), which is generally achieved by combining ensemble weather forecast (EWF) and hydrological model, has a wide application. However, the EWF is biased and underdispersive, and therefore cannot be directly used to build ESP. The skillful forecast lead time of EWF in ESP needs to be determined, and the uncertainty of hydrological models is also nonnegligible. In this study, raw meteorological forecasts are corrected by the generator-based postprocessing method (GPP), the skillful forecast lead time of EWF is determined by comparison with a historical resampling method, and hydrological model uncertainty is investigated using Bayesian model average. The results indicate that GPP can significantly reduce bias and improve dispersion. With a superior postprocessing method, the skillful forecast lead times are 9 and 14 lead days for precipitation and temperature, respectively. With the synthetic effects of precipitation and temperature, the ESP has a skillful forecast lead time for around 10 lead days in terms of both deterministic and probabilistic metrics. However, the skillful lead time may be shortened to 5 days for flood season streamflow predictions. In addition, the hydrological model is an important source of uncertainty in ESPs, especially when evaluating ESPs in terms of probabilistic metrics. The ESP based on a combination of multiple hydrological models outperforms that based on a single model. Overall, this study indicates that the combination of postprocessed EWFs and multiple hydrological models is an effective approach for ESPs.