Statistical Postprocessing of Wind Speed Forecasts Using Convolutional Neural Networks

Abstract Current statistical postprocessing methods for probabilistic weather forecasting are not capable of using full spatial patterns from the numerical prediction (NWP) model. In this paper, we incorporate wind speed information by convolutional neural networks (CNNs) and obtain forecasts in Netherlands 48 h ahead, based on KNMI’s deterministic HARMONIE-AROME NWP The CNNs shown to have higher Brier skill scores medium speeds, as well a better continuous ranked probability score (CRPS) logarithmic score, than fully connected quantile regression forests. As secondary result, compared three different density estimation [quantized softmax (QS), kernel mixture networks, fitting truncated normal distribution], found QS method be best.