Interpretable weather forecasting for worldwide stations with a unified deep model

Abstract

Automatic weather stations are essential for fine-grained weather forecasting; they can be built almost anywhere around the world and are much cheaper than radars and satellites. However, these scattered stations only provide partial observations governed by the continuous space–time global weather system, thus introducing thorny challenges to worldwide forecasting. Here we present the Corrformer model with a novel multi-correlation mechanism, which unifies spatial cross-correlation and temporal auto-correlation into a learned multi-scale tree structure to capture worldwide spatiotemporal correlations. Corrformer reduces the canonical double quadratic complexity of spatiotemporal modelling to linear in spatial modelling and log-linear in temporal modelling, achieving collaborative forecasts for tens of thousands of stations within a unified deep model. Our model can generate interpretable predictions based on inferred propagation directions of weather processes, facilitating a fully data-driven artificial intelligence paradigm for discovering insights for meteorological science. Corrformer yields state-of-the-art forecasts on global, regional and citywide datasets with high confidence and provided skilful weather services for the 2022 Winter Olympics. Worldwide weather station forecasting is challenging because of high computational costs and the difficulty of modelling spatiotemporal correlations from partial observations. Wu et al. propose a transformer-based method that can reconstruct such complex correlations from scattered weather stations, leading to efficient and interpretable state-of-the-art forecasts.