Probabilistic Nowcasting of Heavy Rainfall with a Spatio-Temporal Graph Transformer
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Short-term heavy rainfall, as a highly sudden and destructive weather phenomenon, poses a significant challenge for accurate forecasting in the meteorological field. Addressing the shortcomings of existing forecasting methods in spatiotemporal feature fusion and uncertainty quantification, this paper proposes a novel probabilistic forecasting method based on a Spatiotemporal Graph Transformer. This method innovatively combines Graph Attention Networks and the Transformer architecture: the graph attention mechanism adaptively learns spatial dependencies between meteorological stations, while the Transformer encoder captures the temporal evolution patterns of meteorological elements, achieving end-to-end deep integration of spatiotemporal features. Experiments based on observational data from 156 meteorological stations in a southwestern province of China from 2018 to 2024 show that this method achieves a TS score of 0.341, which is 45.7 % and 17.2 % higher than traditional XGBoost and ConvLSTM models, respectively. The AUC indicator reaches 0.867, and the Brier Score is 0.048. Ablation experiments further validate the effectiveness of each core component, with the graph attention mechanism contributing an 11.4 % performance improvement and Transformer temporal modeling contributing an 18.4 % improvement. The research demonstrates that the proposed Spatio-temporal Graph Transformer model can effectively enhance the accuracy and reliability of short-term heavy rainfall forecasts, providing a new technical approach for the operational application of intelligent meteorological forecasting.
Fields of Study & Subjects
Hydrological Forecasting Using AI
Traffic Prediction and Management Techniques
Meteorological Phenomena and Simulations
Publication details
- Journal / source
- International Conference on Electrical Automation and Artificial Intelligence (ICEAAI)
- Pages
- 1071–1075
- Publication date
- 16 January 2026
- Language
- EN