Optimizing the operation of sediment pass-through tunnel of Nanhua Reservoir with machine learning and dynamic programming approach

Sediment accumulation in reservoirs poses a persistent challenge to sustainable water management, particularly during flood events when opportunities for sediment routing emerge. An innovative approach is introduced to optimize operations of the sediment pass-through tunnel at Nanhua Reservoir, Taiwan, integrating surrogate modeling with dynamic programming (DP). The SRH-2D model is utilized to simulate sediment transport dynamics. Its high computational demand, however, restricts real-time use. An XGBoost-based surrogate model is thus developed to efficiently predict sediment pass-through under diverse flood conditions. This model is paired with a DP framework to identify optimal tunnel operation strategies, maximizing sediment discharge while ensuring flood control and water supply reliability. Historical flood data from 2008 to 2024 are evaluated to validate the method. Results show sediment pass-through increases significantly—reaching 0.146 MCM during Typhoon Gaemi—compared to 0.076 MCM in baseline operations, with negligible downstream flood impacts. Sensitivity analyses highlight inflow rates and sediment concentration as key drivers of routing efficiency. The proposed methodology provides a practical, computationally efficient solution for real-time reservoir management.