Reinforcement learning-assisted walrus optimization algorithm for multi-equipment optimization in industrial cooling water systems

Abstract

In industrial refrigeration plant cooling water systems, chillers, cooling water pumps, and cooling towers are strongly coupled, making independent equipment regulation insufficient for global optimality. Existing chiller plant control studies often insufficiently characterize cross time thermal coupling caused by cooling water transport delay, have limited coordinated decision making, and give inadequate attention to condenser thermal safety. To address these issues, this study proposes a two step look ahead dynamic optimization method based on a reinforcement learning assisted Walrus Optimizer. Centered on chiller heat balance, the method constructs a composite objective function that incorporates cross time energy coupling and condenser thermal safety into a unified optimization process. To handle strongly coupled constraints in complex coordinated optimization, a Q learning adaptive operator selection mechanism is embedded in WO, enabling search strategies to adjust with population evolution and balance global exploration with local exploitation. To exploit thermal state continuity and maintain thermal safety, a cross time hierarchical elite archive is designed to preserve, inject, and warm start high quality solutions across adjacent control steps, improving efficiency, energy saving performance, and stability. Tests on nine benchmark functions show that RL-WO outperforms WO, GA, PSO, GWO, WOA, and DE in mean value and standard deviation. Field validation with historical operation data shows average energy saving rates of 2.58%, 4.67%, 8.14%, and 10.51% under high, medium high, medium low, and low load conditions. Condenser outlet temperatures remain below the safety threshold, and computation time satisfies engineering constraints, verifying effectiveness and safety.