Anet1: post-processing of ensemble weather forecasts using neural networks

Abstract

Ensemble forecast post-processing plays a crucial role in generating more accurate probabilistic weather forecasts. Traditional methods estimate parameters of a parametric distribution separately for each location or lead time while assuming the target distribution of the post-processed weather variable. We propose a novel, neural network-based approach, denoted as ANET1, that produces forecasts jointly for all locations and lead times. Our model post-processes individual ensemble members and uses their latent encodings to estimate the parameters of a predictive normal distribution. To evaluate our method, we conduct temperature forecast post-processing for stations in a sub-region of western Europe using the EUPPBench benchmark. Our results demonstrate that ANET1 showcases state-of-the-art performance, improving upon existing methods in challenging mountainous regions. Compared to the two best methods, EMOS and DVQR, ANET1 exhibits better continuous ranked probability score and quantile loss, resulting in tangible improvements in the calibration of the forecast.