Cascade reservoirs operation rules derivation using integrated decision-making method and combinatorial evolution network

Abstract

Implicit reservoir operation rules map known reservoir conditions to discharge flow or water level decisions to guide reservoir operation. Existing data-driven operation rule extraction models are influenced by both features and hyperparameters, and lack subjective-objective coordination, hindering the formulation of reasonable water energy utilization strategies. To improve the comprehensive benefits of cascade reservoirs, this study proposes a novel multi-objective cascade reservoir operation rule derivation framework. First, a multi-objective optimal operation model for cascade reservoirs (MOOMCR) considering power generation, environment, and navigation is constructed and solved using a Speed-constrained Multi-objective Particle Swarm Optimization algorithm (SMPSO). The relationships among the requirements were explained using a four-dimensional Pareto front visualization analysis. Then, an Integrated Decision-making Method based on Multilevel Subspace Domination (IDMSD) is proposed to circumvent the subjective defects of single decision-making models and reliably select the optimal scheme with the most significant comprehensive benefits from all typical years. Finally, a feature-hyperparameter combination optimization technique based on the differential evolutionary algorithm (FHCC-DE/best/1) is proposed, which exhibits its strength in the identification of key influencing factors on the optimal discharge flow from the IDMSD. The method was embedded in a deep learning model to construct a Combinatorial Evolution Network (CEN) to extract the implicit reservoir operation rules from the optimal IDMSD schemes for the MOOMCR. The results show that the RMSE of the CEN simulating the optimal discharge flow is improved by 5.98%, 28.03%, and 27.89% for the XLD reservoir and by 11.01%, 1.66%, and 60.98% for the XJB reservoir, respectively, over the best results of all benchmark models. Meanwhile, the operation simulation results confirmed that the CEN outperforms conventional reservoir scheduling, with advanced multi-objective comprehensive effects. This suggests that multimethod integrated decision-making and deep learning evolution optimization can help improve the accuracy, generalization, and reliability of implicit reservoir operation rules.