Disturbance-Encoding-Based Neural Hammerstein–Wiener Model for Industrial Process Predictive Control

Abstract

The control reliability of model predictive control is largely determined by the accuracy of the process model. The Hammerstein–Wiener (HW) model is an important nonlinear process modeling technique that has obtained great success in some process industries. Disturbances result in model mismatch and steady-state deviation, but little effort has been devoted to the coupling effects and inertial information in measured disturbances. In addition, few studies try to construct a disturbance observer (DO) to alleviate unmeasured disturbances. The present work proposes prompt disturbance rejection. First, a spatial–temporal long short-term memory-based measurable disturbance encoder is devised to analyze time-series information from measured disturbances and their coupling effects. The encoder can further clarify the status of inertial interference components and the disturbance intensity. Second, a DO is designed to estimate and attenuate unmeasured disturbances. Third, to create the new HW network, which is improved by integrating the disturbance encoder and observer differential, neural networks are used as nonlinear parts. Finally, a model predictive controller based on this improved model is constructed for real-time industrial process control. Simulation comparison experiments have demonstrated the superiority of the proposed methods. Real industry application in the country’s largest lead-zinc froth flotation plant in China validated the proposed model’s effectiveness in controlling chemical reagents.