Precipitation nowcasting model based on multi-source data fusion and efficient additive attention

Short-term precipitation nowcasting serves as an indispensable technical support for the early warning and risk management of meteorological hazards. To improve the nowcasting accuracy by leveraging complementary information from multi-source data, this study proposes a precipitation nowcasting model based on multi-source data fusion and efficient additive attention (EAA-MNFP). The model adopts a hierarchical feature processing framework, while a video Transformer-based grid feature encoder and a spatiotemporal graph convolution-based station feature encoder are designed to independently extract spatiotemporal sequence features from radar echo data and automatic weather station data. Subsequently, an internal feature encoder integrated with efficient additive attention mechanism is applied to encode sequential features, followed by the cross feature encoder utilized cross attention mechanism to align and exchange information between the grid-based and station-based feature representations. Finally, the predicted precipitation results are generated and obtained. Quantitative evaluation on real meteorological datasets reveals the effectiveness of the fusion strategy in short-term precipitation nowcasting tasks, providing a feasible technical approach for enhancing the performance of precipitation nowcasting models.