Abstract Accurate weather forecasts rely on computationally intensive high‐resolution models solving partial differential equations (PDEs) explicitly on a fine grid. Traditionally, downscaling is used to achieve a reduction of the computational cost by first solving a lower‐resolution weather model, providing the boundary conditions for a regional high‐resolution weather model. In recent years, the use of machine learning has been investigated to accelerate weather prediction further through learning the mapping from low‐resolution inputs to high‐resolution predictions—known as super‐resolution methods. In this work, we benchmark recent super‐resolution approaches used previously for downscaling wind predictions as well as new transformer‐based architecture on a novel dataset, with a large variety of input variables, consisting of European Centre for Medium‐Range Weather Forecasts Integrated Forecasting System (ECMWF‐IFS) data and predictions from a high‐resolution regional weather model over Switzerland, and test the generalization performance on weather forecasts from additional locations across Europe. We investigate the relevance of each input variable when downscaling wind further and propose Windflow‐SRnet, a transformer‐based neural network. Using the high‐resolution model as label, the proposed (and best‐performing) machine‐learning model resulted in an average wind‐speed prediction error of on a held‐out test set, a 35% reduction compared with the error for downscaling using linear interpolation.
Ericson et al. (Fri,) studied this question.