A Dynamic Quality Weight Adjustment-Based Multi-source Fusion Method for Precipitation Forecasting

Accurate precipitation forecasting plays a pivotal role in weather forecasting and disaster management. This study introduces a novel approach, the Dynamic Quality Weighting Multi-source Fusion Method, designed for precipitation forecasting by leveraging a Transformer-based architecture. The proposed model integrates multi-source precipitation data, including radar echo and satellite infrared imagery, to effectively capture spatiotemporal dependencies and cross-modal interactions. In order to enhance the robustness to noise and missing data, a dynamic quality weighting mechanism is added to the fusion module part, which dynamically adjusts the contribution of data of each source according to the forecasting confidence of precipitation data from different sources. Experiments were conducted on the MeteoNet public dataset, focusing on forecasting radar echo maps for the subsequent six hours using observations from the preceding six hours. The results demonstrate that the proposed method surpasses state-of-the-art models, including ConvLSTM, PredRNN, and SimVP, in terms of both forecasting accuracy and the structural consistency of precipitation patterns. This study underscores the potential of dynamic quality weighting and Transformer-based architectures to advance precipitation forecasting and multi-source data fusion methodologies.