Neural network based model predictive control for a steel pickling process

Abstract

A neural network based model predictive control strategy (NNMPC) is developed for a multivariable nonlinear system in this research. A steel pickling process which is a fundamental industry in Thailand and has long existed and served the country's steel demand is chosen as the case study. The hydrochloric acid concentrations in the acid baths are the controlled variables. The baths exhibit common features in an industrial systems including nonlinear dynamics and interaction among variables. In the modeling, multiple-input single-output multilayer feedforward neural network models are developed using input-output data sets obtaining from mathematical model simulation. The Levenberg-Marquardt algorithm is used to train the neural network process models. In the control algorithm, the neural network models are used to predict the future process response in a model predictive control (MPC) algorithm for searching the optimal control actions via the Successive quadratic programming (SQP). The proposed algorithm is tested to control the steel pickling process in simulation for several cases such as set point tracking, disturbance, model mismatch and presence of noise. The results of NNMPC show better performance in control of the system over conventional PI controller in most cases. In addition, implementation of inverse neural network (InvNN) controller and Dual Mode (DM) strategy to a steel pickling process is investigated. Simulation results have demonstrated that DM control strategy give good control results for the steel pickling process and remove the offset when compared to the NNDIC and conventional PI controller.