A Structured Graph Neural Network for Improving the Numerical Weather Prediction of Rainfall

Abstract The current challenges of numerical weather prediction (NWP) of rainfall mainly stem from the complex and multiscale nature of rainfall. In recent years, as observation capability improved worldwide, there has been an increased feasibility to use data‐driven models to enhance forecasting performance with rainfall observation. Compared to traditional statistical and machine learning models, deep‐learning models show considerable promise in capturing the spatial‐temporal features of weather processes from multiple predictors, but the convolution‐based feature extractor is suboptimal due to the linear nature of convolution kernels. In this study, a multilevel forecasting model is proposed to forecast each rainfall level, in which each submodel adopts a graph neural network for feature extraction. Spatial and temporal propagation functions based on grid structure are designed to explicitly represent feature fusion and propagation of multiple predictors across multiple scales. On model training, a weight setting strategy that balances the impact of samples with different rainfall values on the total training loss is proposed, and a soft classification label is designed to convert observed rainfall into the probability of rainfall above each threshold. The proposed model was trained and validated on NWP data provided by European Center for Medium‐Range Weather Forecast, and results show significant improvement over the NWP in terms of threat score (TS) and Heidke Skill Score (HSS) scores. Analysis of the forecast results for two typical rainfall processes also illustrates that the proposed method can predict rainfall with more reasonable location and intensity.