Hierarchical Temporal-Scale Framework for Real-Time Streamflow Prediction in Reservoir-Regulated Basins

Reservoir construction has profoundly altered natural runoff evolution in river basins. Dynamic conflicts among multi-objective operational strategies—such as flood control, water supply, and ecological compensation—across varying temporal scales exacerbate uncertainties in runoff prediction, primarily due to the complex interplay between hydrological rhythm variations and anthropogenic regulation. To address these challenges, this study proposes a hierarchical multi-scale coupling framework. Long short-term memory (LSTM) networks are employed to extract implicit operational patterns from long-term reservoir records at monthly and weekly scales, while short-term decision dynamics are captured through deviations from these established long-term rules. The proposed framework is validated in the Dongjiang River Basin, a key water source for the Guangdong–Hong Kong–Macao Greater Bay Area. Compared to single-scale models, the hierarchical approach improves prediction accuracy with an average Nash–Sutcliffe Efficiency (NSE) increase of 9.4% and reductions in the Root Mean Square Error (RMSE) and the Mean Absolute Error (MAE) of 13.2% and 9.6%, respectively. When coupled with a hydrological model, the framework enhances simulation accuracy in reservoir-regulated basins by up to 37.8%. By integrating multi-source decision variables, the framework captures the feedback mechanisms between natural flow variability and human interventions across temporal scales, providing a transferable strategy to reconcile operational conflicts with ecological flow requirements. Its flexibility supports optimized water allocation in regulated river basins, contributing to enhanced water security for downstream urban agglomerations.