Data-Driven Method For Model Predictive Control Performance Monitoring

Model predictive control (MPC) performance monitoring is of great importance for both academic research and practical application to assure the stable and effective operation. It represents a pivotal question in data-driven MPC performance monitoring that a monitoring index which can characterize the current condition of MPC is defined and an intelligent approach which can systematically integrates both the assessment and diagnosis procedure is proposed via making full use of highly accessible field data. Centering on this, this paper focuses on integrated method of MPC performance assessment and diagnosis in hope of providing feasible solution.This paper first introduces the research objective and meaning of data-driven MPC performance monitoring, then reviews the present research status. Based on this, extensive researches are conducted on both MPC performance assessment and diagnosis parts.For the performance assessment part, a PCA-based performance assessment method is proposed. Due to the feature of normal distributions of the process variables, a unified Mahalanobis distance based monitoring index is introduced with its benchmark deduced. This index can capture the deviation of the process variables in both principal subspace and residual subspace so as to improve the assessment performance. Later on, concrete steps of performance assessment are designed. A case study of the Wood-Berry distillation column process is used to demonstrate the superiority of the proposed method over other conventional ones in assessment performance.For the performance diagnosis part, a SVM-based performance assessment method is proposed. Considering four common MPC degradation factors, namely noise variance change, model mismatch, control variables constraint saturation, and manipulated variables constraint saturation, MPC performance degradation patterns are classified into corresponding four groups. Performance signatures are extracted from both principle component and residual subspace, and classifier is constructed to identify four common performance degradation patterns via support vector machine. The effectiveness of the proposed method is demonstrated on Wood-Berry distillation column process.To verify the effectiveness of the proposed method in practical application, two-tank liquid level system based MPC monitoring experiment is first setup. Then the proposed MPC performance assessment and diagnosis method are demonstrated respectively. Results show that compared with other assessment method, the proposed method is superior in that it enjoys both lower false negative rate and false positive rate and that when a poor performance is detected, root cause of MPC performance degradation can be identified.