Reinforcement Learning and Machine Learning Controllers for Enhancing Water Quality and Process Efficiency in Electrochemical Desalination

Abstract

Effective control of electrochemical desalination is limited by the intricate relationship between operating parameters, performance, and feedwater quality dynamics. This complexity cannot be adequately represented in conventional mathematical models or controllers due to their simplified assumptions and inability to account for numerous influencing parameters. Artificial intelligence (AI) techniques present a promising solution for modeling complex nonlinear relationships and adaptively responding to changing conditions. This study investigated two advanced AI controllers: the artificial neural network-based model predictive controller (MPC) and reinforcement learning controller (RLC). Both maintained water quality near the drinkable concentration (17 mM). The MPC ensured continuous efficiency and effluent water quality (purity), while the RLC exhibited deviations with unseen dynamic variations. Despite its longer response time, the MPC generalization accuracy (70%) far surpassed the RLC (30%). By highlighting the distinct strengths and limitations of both controllers, this study underscores the potential of AI control strategies for complex dynamic desalination operations.