Multi-objective Reservoir Flood Control Optimization uing a Real-time Adaptive Hybrid CNN-MCTS Framework with Spatiotemporal Feature Extraction

This study presents a real-time adaptive framework for multi-objective reservoir flood control optimization, that integrates Convolutional Neural Networks (CNN) with Monte Carlo Tree Search (MCTS). The proposed method addresses key challenges in operational decision-making under hydrological uncertainty, including timely outflow adjustment, storage efficiency, and downstream flood risk mitigation. A novel three-dimensional spatiotemporal feature tensor is constructed to represent dynamic interactions among inflow, outflow, and reservoir water levels. A dual-network architecture is employed: a policy network predicts feasible outflow actions, while a value network evaluates their performance across multiple objectives. These networks guide the MCTS to explore and refine control strategies under real-world operational constraints. The framework is validated at China’s Chitan Hydropower Station, achieving over 95% accuracy in outflow prediction (± 500 m 3 /s tolerance) and objective evaluation (± 0.15 error), with 10% higher peak flow reduction and 40% lower flood storage usage compared to historical operations. This framework method demonstrates robust performance across diverse flood scenarios (0.2–5% frequencies) and real-time feasibility (< 60 s per decision step), The results indicate that the proposed approach is a scalable, transferable module for intelligent reservoir scheduling systems. Highlights. A real-time reservoir flood control framework is developed using hybrid CNN-MCTS. 3D spatiotemporal tensors enable CNN to extract dynamic hydrological features. The method outperforms historical rules with faster decisions and better flood mitigation.