Abstract
We propose a hierarchical temporal scale framework for real-time streamflow prediction in reservoir-regulated basins to ensure effective water resources management. As an important anthropogenic interference in the hydrologic cycle, reservoir operation behavior remains challenging to properly represent in hydrologic models, thus limiting the capability of predicting streamflow under the interactions between hydrologic variability and operational preferences. We employ a data-driven model (LSTM) for streamflow prediction in reservoir-regulated basins. Given the difficulty in predicting streamflow processes caused by varying operational objectives of different reservoirs across different time scales within the basin, we simulate the monthly storage and release patterns of reservoirs using historical daily operation data and then capture the deviations between daily scales and these patterns to model the actual reservoir operation rules. Finally, we predict the watershed streamflow based on the reservoir release volume combined with hydrometeorological data. We enhance model performance and interpretability using the Optuna method and Shapley additive explanation (SHAP). The Dongjiang River Basin (DRB) serves as the study area. Results indicate that the framework excellently captures the operational patterns of the three major reservoirs in the basin and significantly improves the daily streamflow prediction accuracy. Model interpretability results show that the contribution of main stem reservoir releases to downstream streamflow is greater than that of tributary reservoir releases.