Fusing Multi-Scale Numerical Model Forecasts to Improve Short-Term Intense Rainfall Forecast with a Deep Learning Rain Network

Severe precipitation ranks among the most impactful meteorological disasters, exerting profound social and economic consequences. Forecasting heavy rainfall on an hourly scale, driven by complex multi-scale thermodynamic processes, remains a formidable challenge in numerical weather prediction (NWP). This study introduces a deep-learning-based fusion rain network (FRNet) designed to improve intense rainfall forecasting by effectively integrating the strengths of multi-scale model rain features. The model’s performance in 3-h rainfall forecasts within a 24-h lead time over North China was comprehensively validated. Statistical assessments, based on independent tests conducted for 2021 and 2022, demonstrated FRNet's robust generalization capability. The results of the evaluations reveal that the deep learning model significantly enhances forecast accuracy across various precipitation thresholds—ranging from light to torrential rain—across all lead times within the 24-h period, when compared with the original model forecasts and the traditional multi-model similarity ensemble correction method. Notably, improvements become more pronounced with increasing precipitation intensity. Case studies of typical intense rainfall events have illustrated that FRNet effectively synthesizes the strengths of global models in predicting rainband location and movement, as well as mesoscale numerical models in forecasting intensity and fine-scale structures. This integration results in a more comprehensive and precise adjustment of precipitation forecasts, offering superior guidance for intense rainfall events compared to any single numerical model. Structure-amplitude-location (SAL) verification shows that FRNet significantly adjusts the structure and location of heavy precipitation but shows overestimation at the same time. Ablation experiments confirm that the performance gains of FRNet are primarily driven by the inclusion of high-resolution global and mesoscale model forecasts, complemented by geographic and temporal information. This study underscores the potential of deep learning techniques to integrate more multi-scale physical and observational features from different sources, paving the way for future advancements in intense rainfall forecasting.