Improving the Intense Rainfall Forecast with a Fusion Rain Network

Severe precipitation is one of the most significant meteorological disasters, causing substantial social and economic impacts. The fine-scale prediction of short-term intense rainfall remains one of the most challenging aspects of numerical weather forecast. This lecture addresses the issue of the long-tail distribution of precipitation samples and developes a deep learning correction model which based on multi-scale numerical forecasts (Fusion Rain Network, FRNet). Then the ability of FRNet to extract and integrate advantages from numerical forecasts and the improvement on the skill of heavy rainfall forecast was comprehensively validated. Two independent parallel tests in year 2021 and 2022 of the 24h rain forecasts in North China were conducted, which shows a good generalization capabilities of FRNet. Statistical evaluations reveal that the deep learning model significantly improves the forecast scores for various thresholds, ranging throughout light rain, moderate rain, heavy rain to torrential rain, across all lead times within the 24h period, and the greater improvements are observeds with the increase of precipitation intensity. Case studies of typical intense rain events show that FRNet can learn the strengths of global models in forecasting the location and movement of rain bands and the advantages of mesoscale numerical models in predicting precipitation intensity and fine structures. Consequently, it provides a more comprehensive adjustment of precipitation forecasts, offering greater guidance for intense rainfall than any single numerical model. The machine learning ablation experiments further confirm that the above improvement by FRNet are primarily attributed to the high-resolution global model and mesoscale model forecasts, subsidiary the feature of geographic and temporal information. This study provides valuable insights for future applications of deep learning techniques to integrate more physical and observational data, thereby enhancing the forecast skill of intense rainfall forecasts.