Deep Reinforcement Learning Aided Differential Evolution for High-Dimensional Reservoir Optimization: The Case of Jinsha River - Three Gorges Cascade

This paper proposes the Neural Evolution-Enhanced Differential Evolution (NEDE), a novel meta-heuristic algorithm designed to optimize the operations of complex, high-dimensional reservoir clusters—a critical task for efficient hydropower utilization and global carbon reduction. Integrating Differential Evolution (DE) with Deep Reinforcement Learning (DRL) and the concept of "evolutionary paths," NEDE features a novel neural network-based architecture comprising a population encoder, policy selector, and parameter controller. By leveraging the adaptive capabilities of DRL to dynamically adjust DE parameters and mutation strategies, the algorithm effectively addresses the challenges of hydraulic coupling and hydrological uncertainty inherent in high-dimensional systems. The neural network is trained within a reinforcement learning framework utilizing fitness rewards and entropy regularization. Comparative analyses on the IEEE CEC 2020 benchmark functions and the real-world downstream Jinsha River - Three Gorges cascade system demonstrate that NEDE delivers superior solution accuracy and faster convergence than conventional algorithms. Specifically, compared to LSHADE, NEDE increases average and maximum power generation by 0.4% and 0.9% in wet years, 1.0% and 0.9% in normal years, and 1.0% and 1.2% in dry years, respectively, validating its robustness and effectiveness in dynamic reservoir scheduling.