Autonomous water quality management in an electrochemical desalination process

Abstract

This research explores advanced control strategies to enhance water quality in membrane capacitive deionization (MCDI) systems, employing a validated modified Donnan model. Three types of artificial neural network (ANN) controllers were developed and evaluated, namely, ANN-proportional-integral-derivative, ANN-Integral, and Multiple Parallel ANN-Integral (MPAI) Controllers. Among these, the MPAI Controller demonstrated the best performance and was selected for further optimization. It was then compared with an offline reinforcement learning controller using the Conservative Q-Learning (CQL) algorithm. To optimize the CQL Controller, various reward functions were tested, including quadratic penalty, exponential penalty, and a Gaussian reward function, with the Gaussian function ultimately selected for its effectiveness, achieving a reward at approximately one. Both control strategies maintained the effluent concentration at approximately 17 mM, despite variations in inlet concentration and fouling dynamics, with absolute errors under 0.4 mM. Notably, the MPAI Controller showed the highest precision, with an error margin approaching nearly zero compared to the CQL Controller. This study underscores the potential of AI-driven controllers in enhancing the efficiency and reliability of MCDI systems, contributing to advancements in water treatment technologies.