Short-Range Ensemble Forecast Post-processing

Short-term hydrological ensemble forecasts do not usually account for the uncertainty in the initial conditions. Consequently, raw forecasts are often biased and under-dispersed and must be post-processed. Both precipitation and streamflow forecasts for short lead-time depart from the Gaussian distribution, and this important characteristic limits the choice of possible post-processing approaches. Post-processing is performed by calibrating a statistical model using a training dataset containing past forecasts and the corresponding observations. This chapter covers the most common post-processing approaches for short-term hydrological forecasts. They are divided into four categories: analog methods, regressions, kernel dressing, and Bayesian Model Averaging. The vast majority of post-processing methods can be categorized as regression-based. A selection of the most commonly encountered ones in hydrology is presented: quantile regression, nonhomogeneous regression, and logistic regression. Any post-processing approach brings benefits and drawbacks, which are discussed at the end of this chapter. However, according to the few existing comparative studies, no single method is appropriate for all forecasting situation. Therefore, the reader should make his or her own mind regarding which one to choose, according to his or her own specific needs and limitations.