(325a) Closed-Loop Subspace Projection Based State-Space Model-Plant Mismatch Detection and Isolation for MIMO MPC Performance Monitoring

Model predictive control (MPC) system determines the best moves of manipulated variables by solving an online optimization problem so as to achieve the optimal targets including the production throughput, energy consumption and economic profit. However, the performance of industrial multi-input multi-output (MIMO) MPC applications often degrade dramatically after a period of operation due to various factors such as model-plant mismatch, poor controller tuning, changes of noise disturbances, sensor/actuator faults, inappropriate control design, and changes of constraint sets. Among the aforementioned factors causing MPC performance deterioration, the model-plant mismatch is a very significant one because the process model is needed in MPC systems for enabling the horizon based predictions of all controlled variables. The unreliable plant models can result in poor predictions on the process outputs, which in turn affect the optimized move sequences of system inputs. Consequently, it is necessary and useful to detect different kinds of model-plant mismatches and resolve the model quality issues rapidly for sustaining MPC performance. Though re-identification of plant models can improve model quality and prediction accuracy, it is very time consuming and economically expensive in industrial practice. Therefore, the automatic detection and isolation of the model-plant mismatches are highly desirable to monitor and improve MPC performance.

In this study, a new closed-loop MPC performance monitoring approach is proposed to detect model-plant mismatches within discrete-time state-space formulations through subspace projections and statistical hypothesis testing. This approach makes use of closed-loop operating data without the intrusive step tests, and a set of quadratic indices based on different kinds of subspace projections are developed for identifying the significant mismatches on various state-space model matrices. Moreover, the model residuals are utilized for mismatch isolation and the feedback effects under closed-loop operation are eliminated through various orthogonal projection of the future disturbances onto the past inputs and outputs. A monitoring framework consisting of a series of quadratic indices and the corresponding confidence limits is developed to identify model-plant mismatches precisely.

The presented method is applied to a simulated multivariable MPC system in paper machine process and different test cases with model-plant mismatches are designed to examine the monitoring capability of the proposed approach. The computational results demonstrate that the proposed method can accurately detect and isolate the mismatches on system matrices. Therefore, this approach may offer an effective way for closed-loop MPC performance monitoring and especially model-plant mismatch detection.