Combined cooling in CSP: Reinforcement learning applied to a two-stage optimization

Abstract

To improve the sustainability and economic viability of concentrated solar thermal energy systems, this study proposes a cooling system that combines wet and dry cooling technologies. The effectiveness of this novel cooling solution relies on optimal operation strategies that balance water and electricity use under constrained water availability. To address this challenge, a two-stage optimization framework is proposed. In the first stage, a sequence of multi-step optimization problems is solved to generate Pareto fronts representing trade-offs between resource consumption. In the second stage, these fronts are traversed through a path-search algorithm that minimizes the cumulative cooling cost over the optimization horizon. This work focuses on the first stage by comparing two methodologies for obtaining Pareto fronts: (i) an exhaustive search across the entire operating range and (ii) a reinforcement learning agent. The results show that the latter significantly reduces computational time, which is particularly advantageous given the long operational lifetime of such plants, while maintaining a comparable solution quality.