Ensemble weather forecast post-processing with a flexible probabilistic neural network approach

Abstract

Ensemble forecast post-processing is a necessary step in producing accurate probabilistic forecasts. Many post-processing methods operate by estimating the parameters of a predetermined probability distribution; others operate on a per-lead-time or per-station basis. All of the aforementioned factors either limit the expressive power of the methods in question or require additional models, one for each lead time and station. We propose a novel, neural network-based method that produces forecasts for all lead times jointly and requires a single model for all stations. We incorporate normalizing spline flows as flexible parametric distribution estimators, which enables us to model complex forecast distributions. Furthermore, we demonstrate the effectiveness of our method in the context of the EUPPBench benchmark, where we conduct 2-m temperature forecast post-processing for stations in a subregion of Europe. We show that our novel method exhibits state-of-the-art performance on the benchmark, improving upon other well-performing entries. Additionally, by providing a detailed comparison of three variants of our novel post-processing method, we elucidate the reasons why our method outperforms per-lead-time-based approaches and approaches with distributional assumptions.