In recent years, affected by the combined effects of monsoon circulation and topography, short-term extreme precipitation events have occurred frequently in North China with high disaster severity. A typical case is the "7·29" extreme heavy rainfall in the Beijing-Tianjin-Hebei region in 2023, which exposes the shortcomings of existing regional forecasting models in capturing precipitation processes characterized by "strong suddenness and high spatial heterogeneity". To address the challenge of short-term extreme precipitation forecasting in North China, this study proposes a deep learning framework integrating satellite and ground-based observations. Its core is to leverage the advantages of multi-source data to match the precipitation characteristics of North China, thereby achieving accurate capture of macro-micro temporal and spatial dynamics.
The innovation of the model architecture is designed around the needs of short-term extreme precipitation forecasting in North China. Its cross-attention multi-source fusion module, targeting the complex water vapor transport paths in North China, enables adaptive feature interaction between the GNSS-PWV macro water vapor field, radar mesoscale dynamic field, and ground precipitation field, thus accurately integrating the associated signals of "water vapor-dynamics-precipitation". The terrain-aware spatiotemporal attention mechanism takes Vision Transformer (ViT) as the core and incorporates digital elevation model information of North China. Experiments based on the 2021-2025 heavy rainfall dataset in North China show that the model adapts to the precipitation characteristics of North China in the 0-3 hour short-term forecasting task. Under the heavy precipitation threshold, the Probability of Detection (POD) is 18%-25% higher than that of the benchmark model.
 

Short-term extreme precipitation forecasting; Multi-source data fusion; Beidou water vapor retrieval; Terrain awareness; Spatiotemporal attention mechanism