Monitoring model predictive control systems

Despite the widespread application of model predictive control systems (MPC) in the process industries over the past two decades, only recently has attention been focused on the important problem of monitoring MPC system performance. The primary objective of this research is to develop practical and effective approaches for MPC monitoring. Ideally, MPC monitoring techniques should address the following questions: (1) Is the MPC system performing "normally"? (2) If the performance is abnormal, is the cause due to: (a) Unusual disturbances? (b) Sensor problems? (c) An inaccurate process model (for current operating conditions)? Thus, the monitoring technique should be able to detect changes in control system performance and diagnose the cause of the performance degradation.; In order to achieve these goals, a novel approach to the MPC monitoring problem based on pattern classification of time-series data is proposed. First, a simulated database of closed-loop responses is created using the process model in the MPC system. This database contains several types of generic disturbances, sensor problems, and plant model changes. Then, two main classification approaches are utilized to perform controller monitoring: (1) neural network based pattern classifiers classify the MPC performance into different classes---normal vs. abnormal, whether or not an unusual disturbance is present, and whether a significant plant change has occurred; (2) a novel principal component analysis (PCA) based on the PCA, distance, and variance similarity factors compares a window of current MPC operating data to the simulated responses. The PCA approach indirectly classifies MPC performance by indicating which simulated responses are most similar to the current operation.; A case study is performed on a closed-loop MPC simulation of the Wood-Berry distillation column model to evaluate the capabilities of the proposed monitoring approaches. Additional techniques developed to complement these approaches are utilized to monitor and diagnose a simulated and actual refinery process controlled by a commercial MPC system, AspenTech DMCplus(TM).; To augment the proposed MPC monitoring approaches, analytical closed-loop models for simplified, unconstrained, two-stage MPC are developed to investigate the closed-loop stability thresholds with respect to modeling error at the earlier controller design stage.