Convolutional neural network-based statistical post-processing of ensemble precipitation forecasts

Abstract

Raw forecasts from numerical weather prediction models suffer from systematic bias and cannot be directly used in applications such as hydrological forecasting. Statistical post-processing methods can be used to remove the bias and achieve reliable ensemble forecasts. However, traditional post-processing methods generally use local raw forecasts as the only predictor, which limits their ability to extract spatial information from raw forecasts. In this paper, we develop a convolutional neural network (CNN)-based post-processing method for precipitation forecasts to make use of spatial information and atmospheric circulation variables as auxiliary predictors. The results show that the proposed CNN-based post-processing model outperforms traditional methods in forecast accuracy and reliability, especially for heavy rain. The improvements of the CNN-based model relative to a state-of-the-art joint probability model can reach 5% and 8% in terms of Brier skill score for heavy rain and rainstorms at the lead time of 1 day. The improvements can be attributed to the use of auxiliary predictors such as total column water forecasts in the proposed model. Moreover, the CNN-based model transcends a fully connected network (FCN)-based post-processing model, which illustrates the benefits of using CNNs to extract spatial information. The results illustrate the proposed CNN-based post-processing model is able to utilize spatial information and various auxiliary predictors to improve forecast skill and quantify forecast uncertainty, which is important for further applications such as hydrological forecasting.