Deriving reservoir operating rule based on multiple evolutionary search, data mining and machine learning methods

Reservoir operating rules are essential guidelines for effective reservoir management and the realization of integrated benefits. Although optimization-based reservoir operation methods have advanced rapidly in recent years, a significant gap between theory and practice has hindered their direct application in real-world operations. To address this issue, this study proposes an integrated methodology that combines evolutionary search, data mining, and machine learning to derive reservoir operating rules. The proposed framework consists of three components: (1) the formulation of a maximum hydroenergy production optimization model solved using Genetic Algorithms (GA), Differential Evolution (DE), and Pattern Search (PS); (2) the identification of key decision factors through Grey Relational Analysis and Kendall’s tau rank correlation coefficient; (3) the derivation and performance evaluation of operating rules using traditional methods (MLR, SGM), classical machine learning methods (ANN, SVM), and emerging machine learning methods (GRU, CatBoost). This methodology is applied to the Ma’erdang Reservoir in the source region of the Yellow River. Characteristic reservoir param and ten-day runoff data from the Jungong station (1980 ~ 2023) are utilized for multi-scenario simulation and comparative analysis. The results indicate that: 1) Without consideration of guaranteed output, all three evolutionary search algorithms yield an average annual energy production of 7.48 billion kWh, and with consideration of guaranteed output, the long-term average generation decreases to 7.46, 7.45, and 7.31 billion kWh for Genetic Algorithm (GA), Differential Evolution (DE), and Pattern Search (PS), respectively. 2) The factor selection results indicate that, in both without consideration of guaranteed output and with consideration of guaranteed output scenarios, six key factors were consistently identified, covering aspects such as the month of the time period, inflow, outflow, and forebay water level. This method reduced the dimensionality of multi-dimensional decision factors by 60% while maintaining over 95% of the decision relevance. 3) Dividing the hydrological time series into a training set (1980 ~ 2014) and a test set (2015 ~ 2023) allowed comparison of the fitting and prediction performance of machine learning methods (ANN, SVM, GRU, CatBoost) and traditional methods (MLR, SGM). Taking the multi-year average energy production obtained from direct optimization as the benchmark, operating rules derived using emerging machine learning models (GRU, CatBoost) achieved up to 99% of the benchmark, those derived from classical machine learning models (ANN, SVM) reached 96%, and those from traditional methods (MLR, SGM) attained 91%.