Abstract
Accurate summer precipitation forecast over the Yangtze River Basin (YRB) is crucial for flood early warning and effective regional water resource management. The Climate‐Weather Research and Forecasting (CWRF) model is a useful tool for providing across‐season monthly precipitation prediction during flooding season in the YRB, but it still exhibits notable limitations in accurately capturing precipitation variations both spatially and temporally. To alleviate these limitations, a novel deep learning framework utilising a spatio‐temporal graph neural network (GNN) is proposed to improve the prediction skill of CWRF. The result suggests that the proposed GNN model outperforms the existing neural network‐based models across multiple evaluation metrics on precipitation prediction over the YRB. Specifically, GNN achieves the most substantial reduction in root‐mean‐square error across the YRB. It also outperforms other models in terms of temporal correlation coefficient, with overall monthly increases (exceeding 0.1 in June) compared to the raw CWRF output. In terms of anomaly correlation coefficient, GNN shows obvious enhancements, indicating better year‐to‐year consistency. These findings provide valuable guidance for enhancing precipitation forecasts, particularly in regions characterised by complex topography, irregular spatial domains, and diverse climate conditions.