Spatiotemporal weather forecasting via multi-scale graph neural networks and latent diffusion models.

Abstract

Accurate weather prediction is crucial in agriculture, disaster prevention, and public safety. Traditional numerical models have high computational costs and struggle with atmospheric nonlinearity and chaos, while existing deep learning methods face limitations in handling spatial heterogeneity and non-Euclidean data. This paper introduces the STGLDWeather method. It combines multi-scale spatiotemporal graph neural networks (MS-ST-GNN) and latent diffusion models (LDM) to capture multi-scale spatiotemporal dependencies in weather data and model the temporal evolution of weather conditions in latent space. Experiments on real weather datasets show that STGLDWeather significantly outperforms existing state-of-the-art baselines in prediction accuracy and computational efficiency, particularly excelling in temperature, geopotential height, and wind speed forecasts.