An ecohydrological framework for balancing hydropower and aquatic health using machine learning and multi-objective optimization

Jinglin Zeng, Yulei Xie, Hang Wan, Ran Li, Yanpeng Cai, Zhifeng Yang,
An ecohydrological framework for balancing hydropower and aquatic health using machine learning and multi-objective optimization,
Journal of Hydrology,
Volume 663, Part B,
2025,
134307,
ISSN 0022-1694,
https://doi.org/10.1016/j.jhydrol.2025.134307.
(https://www.sciencedirect.com/science/article/pii/S0022169425016476)
Abstract: Dam construction and operation alter flow regimes and dissolved gas saturation, threatening downstream fish habitats and aquatic ecosystem safety. Efforts to mitigate these impacts while balancing reservoir operation objectives have drawn significant attention. Although previous studies have considered ecological flow and supersaturated total dissolved gas (TDG) impacts, comprehensive analyses of power generation-ecological protection trade-offs and intelligent methods for impact prediction and scheduling optimization are still lacking. Hence, a multi-objective optimization model was established in the paper, combining machine learning for supersaturated TDG level prediction and a genetic algorithm for reservoir operation, aiming to maximize power generation and minimize supersaturated TDG levels. The results showed that XGBoost (Extreme Gradient Boosting) model outperformed other machine learning models in predicting supersaturated TDG levels, with a mean absolute error of 1.3923, a mean square error of 3.4899, and an R2 value of 0.8845. SHapley Additive exPlanations (SHAP) analysis revealed that total flow was the key factor affecting the prediction accuracy of the XGBoost model for supersaturated TDG levels, followed by flood discharge, with average SHAP values of 0.093 and 0.033. The multi-objective optimization results showed that the maximum supersaturated TDG levels during the optimized flood discharge process remained below 125 %, with a duration of only 8 h, while fish exposure time were shorter than their median lethal times. Furthermore, full-load power generation was achieved, highlighting a synergistic enhancement of both ecological and economic benefits in dam operations. The results showed that the established method combining machine learning predictions with genetic algorithm optimization could effectively address multi-objective ecological scheduling problems. This study also recognized challenges in real-time scheduling under limited data contexts and the need for stakeholder input and policy integration. While not fully addressed here, the proposed framework establishes a foundation for extending to hydrological forecasting, adaptive scheduling, and policy-relevant decision support.
Keywords: Machine learning; Supersaturated total dissolved gas; Multi-objective optimization; SHAP analysis; Ecological scheduling