Middle‐ and Long‐Term Runoff Forecast Model for Water Resource and Climate Security Based on Self‐Attention Mechanism

ABSTRACT Reliable middle‐ and long‐term streamflow forecasts are critical for ensuring water resources management and climate security. This study establishes a novel runoff forecasting model based on the self‐attention (SA) mechanism and variational mode decomposition‐gated recurrent unit (VMD‐GRU) framework to improve monthly streamflow prediction accuracy. The maximal information coefficient (MIC) method is adopted to screen key drivers of runoff variability. The proposed model integrates VMD to decompose the runoff sequence into intrinsic components and applies GRU coupled with SA to predict each component. The whale optimization algorithm (WOA) is adopted for the optimization of VMD‐SA‐GRU model hyperparameters, and then monthly streamflow forecast results are obtained by superimposing the results of each component. Using 40 years of runoff data from the South‐to‐North Water Diversion Project in China, the VMD‐SA‐GRU model is evaluated against GRU and VMD‐GRU benchmarks. Results demonstrate that the VMD‐SA‐GRU model leverages the strengths of its constituent algorithms, significantly improving forecast accuracy. Compared to the VMD‐GRU model, the VMD‐SA‐GRU framework enhances the Nash‐Sutcliff efficiency coefficient (NSE) by 6%–35%, reduces the root mean square error (RMSE) by 15%–30%, and decreases the mean absolute error (MAE) by 15%–33%. This robust forecasting framework provides a reliable tool for sustainable water resource management and addressing climate‐related challenges.