Development of AI-based process controller of sour water treatment unit using deep reinforcement learning

Abstract

BackgroundDue to the variability in the feedstock conditions and the nonlinearity of the sour water stripping process, determining the optimal operating conditions for Sour Water Treatment Unit (SWTU) is a huge challenge. MethodsIn this study, we propose an AI-Based Process Controller (AIPC) for optimizing the SWTU, combining deep reinforcement learning (DRL) and expert knowledge. A surrogate model of an industrial SWTU digital twin was developed to serve as the environment for DRL. A reward function was designed and compared with others for evaluation. A method for seamless switching was devised to guarantee uninterrupted device operation by preventing any interference from the policy network. Significant FindingsIn contrast to the alternative control schemes, the AIPC not only demonstrates superior performance in mitigating overshooting and enhancing setpoint tracking precision but achieves a reduction in stripping steam usage. The proposed method has great potential in the field of real-time optimization.