A variable-resolution unstructured spatio-temporal graph neural network: an application to very short-range weather forecasting in Guangdong, China

Abstract The increase in extreme weather events has raised the demand for very short-range forecasts in Guangdong, the most populous province in China. Although global AI-based weather forecasting models provide strong medium-range guidance, they remain suboptimal for regional very short-range predictions. To address this, we developed a spatio-temporal graph neural network (STGNN) for Guangdong, featuring an unstructured, variable-resolution graph refined over the Pearl River Delta. The core of the network is 1-D dilated causal temporal convolutions and spatial convolutions on combined static and dynamic graphs. A spatial mean constraint and Laplacian-residual regularization were incorporated into the loss function to enhance physical consistency. Trained on historical reanalysis dataset, the model generates hourly forecasts for multiple near-surface atmospheric variables for the next six hours. Verification against an independent test-set and observations, using multiple metrics, demonstrates high predictive skill and strong spatio-temporal coherence for 2 m temperature and sea-level pressure. For 10 m winds and 2 m humidity, model errors grow larger with lead time. Case studies of an extreme heatwave event in Guangzhou and the post-landfall evolution of Super Typhoon Saola effectively capture the evolution characteristics and spatial patterns, while underestimates the peak temperatures, maximum winds and minimum sea-level pressures. These results support the use of variable-resolution STGNN as a practical approach for very short-range regional forecasting in Guangdong, China.