A reinforcement learning-based transformed inverse model strategy for nonlinear process control

Abstract

Process control is evolving with a rapid inclusion of artificial intelligence (AI) based methods, which have the potential to significantly boost controller performances. This development is valuable to control engineers, especially for the control of nonlinear processes plagued by higher offsets and larger delays. With this motivation, we develop in this work a new strategy utilizing AI for process control by assimilating reinforcement learning (RL) with a novel approach for inverse model control. This approach involves expressing a dynamic process model in the controlled variable domain to obtain a computationally efficient, transformed inverse model (TIM). In the developed control strategy, the TIM provides baseline control, which is subsequently improved by RL’s deep deterministic policy gradient method in the resultant TIM-RL controller. This controller is examined in three different in-silico case studies of varying nonlinearity and complexity. Relative to nonlinear model predictive control, the TIM-RL controller is observed to have superior performance with more than 30% reduction in integral absolute error, 45% lower settling time, and about an order of magnitude smaller latency.