Abstract Coordinated optimization of cascade reservoirs is critical for maximizing a river basin's economic, social, and ecological benefits. However, conventional hydropower scheduling lacks adaptability to complex future scenarios, constrained by seasonal hydrological variability and uncertain inflows. While AI algorithms offer new avenues for reservoir operation, bridging the gap between historical runoff‐based prediction and efficient future scheduling remains a key challenge. This study proposes an Optimization–Learning–Simulation (OLS) framework—integrating model‐based, data‐driven, and physics‐informed approaches—to enhance dynamic adaptability and long‐term robustness of cascade reservoir operation under hydrological uncertainty. “Physics‐informed” in the context of reservoir operation refers to the fulfillment of water balance and boundary conditions throughout a complete operational cycle. This means that in the optimization process, the operation of the reservoir must not only aim to maximize scheduling benefits but also ensure basic physical consistency, avoiding violations of the reservoir's physical operational conditions and management requirements. The framework combines optimization (INSGA‐III for multi‐objective optimization, VIKOR for balanced decisions), learning (physics‐informed LSTM with reinforcement learning, PIRLSTM), and simulation (SARIMA for runoff forecasting, Bootstrap/Cholesky for uncertain inflows) to synergize adaptability and robustness. Validated on four mega cascade reservoirs in the lower Jinsha River, the OLS framework learns optimized rules across extreme dry, normal, and extreme wet years. It achieves Nash–Sutcliffe Efficiency (NSE) up to 0.96 and Water Balance Index (WBI) near 1.00, highlighting the framework's superior performance in maintaining physical consistency and predictive accuracy; Simulation results under future runoff scenarios further demonstrate that the OLS framework guarantees scheduling schemes that rigorously comply with reservoir operation boundaries and the fundamental constraint of water balance. Moreover, it effectively identifies and dynamically adjusts scheduling strategies in response to complex hydrological events, including alternating extreme wet–dry sequences and sudden shifts between drought and flood. By integrating artificial intelligence algorithms with hydrological and physical principles, this study offers a novel and practically applicable approach for the coordinated optimization of cascade reservoir operations under complex and uncertain future conditions, bridging theoretical innovation with engineering practice.
Zhu et al. (2026) studied this question.