Reinforcement Learning based Multivariate Nonlinear Model Predictive Controller Validation on a Lab Scale Batch Reactor

Researchers proposed different approaches for the development of accurate and energy-efficient batch reactor control because they are still crucial in the chemical process industries. In this article, the authors propose a novel cost function aimed at minimizing tracking errors for the nonlinear model predictive controller (NMPC) and improving overall system performance. In the first phase, a batch reactor dynamic model was developed from the input and output data collected using an open-loop experiment by simultaneously varying the coolant flow rate and heater current. In the second phase, a recurrent neural network (RNN) model was developed for a multi-input single-output (MISO) reactor system that follows the nonlinear characteristics of reactor system dynamics. The developed RNN-based model was then used to implement NMPC using sigmoidal weights to enhance computational efficiency. The NMPC was also implemented using a reinforcement learning (RL) approach for adaptive control optimization. The simulation results of both models were validated through real-time implementation using the NVIDIA Jetson Orin 8GB environment platform. The results show the effectiveness of the proposed cost in a controlled environment for intelligent and energy-efficient batch processing in the NMPC system.