AI-BASED ANALYSIS AND DIAGNOSIS OF NMPC CONTROLLER PERFORMANCE

The purpose of this study is to analyze and diagnose the performance of model predictive controllers using a data-based approach. Random Forest (RF) and Multilayer Perceptron (MLP) Neural Network, both in cascade, are compared. Poor controller tuning and modeling errors are considered as sources of anomalies, together with valve stiction. The study was conducted in a simulated Van de Vusse reactor controlled by a nonlinear predictive controller (NMPC) with the presence of an extended, constrained Kalman filter (CEKF). The results indicate that both MLP and RF models had successful overall performance. However, the RF-based system shows superior capability in diagnosing the causes of control performance deterioration.