TG-net: A Physically Interpretable Deep Learning Forecasting Model for Thunderstorm Gusts

Thunderstorm gusts are a common and hazardous type of severe convective weather, characterized by a small spatial scale, short duration, and significant destructive power. They often lead to severe disasters, highlighting the critical importance of their accurate forecasting. Previous studies have explored the environmental factors and spatiotemporal distribution characteristics of thunderstorm gusts, highlighting the need for improved forecasting methods. In recent years, artificial intelligence techniques have shown promise in enhancing the accuracy of thunderstorm gust forecasting, with various machine learning algorithms and models having been developed. This paper proposes a multi-scale feature fusion module called Thunderstorm Gusts Block (TG-Block) and a deep learning model named Thunderstorm Gusts net (TG-net) based on the Attention U-net and TG-TransUnet models, and employs interpretable methods such as Integrated Gradient, Deep Learning Importance Features, and Shapley Additive exPlanations to validate the model’s practical relevance and reliability. The analysis of feature importance underscores the model’s ability to capture key thermodynamic and multiscale weather characteristic information for thunderstorm gust nowcasting. It is, however, worth emphasizing that these conclusions are only based on a limited number of thunderstorm gust examples, and the evaluation results may be affected by specific weather types and sample sizes. Nonetheless, TG-net has been put into real-time operation at the Institute of Urban Meteorology, and we will continue to rigorously validate its performance and make any necessary optimizations and enhancements based on feedback to ensure the robustness and stability of the model.