Detection of model-plant mismatch in MPC applications

In model predictive control of processes, the process model plays an important role. The performance of the controller depends on the quality of the model and hence on the model-plant mismatch. Although model-plant mismatch is inevitable, it is highly desirable to minimize it. For processes with large number of inputs and outputs, re-identification of the model is a costly exercise as keeping a large number of inputs in a perturbed or excited state for a long time means loss of normal production time. Hence, it would be highly desirable to detect the precise location of the mismatch so that only a few inputs would have to be perturbed and only the degraded portion of the model updated. In this work, a methodology is proposed for the detection of mismatch from closed-loop operating data. The proposed methodology is based on the analysis of partial correlations between the model residuals and the manipulated variables. Since partial correlation analysis aids in spotting spurious as well as hidden correlations, the proposed technique is able to accurately resolve mismatch due to model inaccuracies and disturbances. The efficacy of the proposed mismatch detection technique is demonstrated on two simulation case studies and its application to data from an industrial process.