Abstract
Model Predictive Control systems may suffer from performance degradation mainly for two reasons: (i) external unmeasured disturbances are not estimated correctly, (ii) the (linear) dynamic model used by the MPC does not match (any longer) the actual process response. In this work we present the application of a method to detect when performance is not optimal, to diagnose the source of performance degradation and to propose appropriate corrections. In the simplest situation (i), optimal performance can be restored by recomputing the estimator parameters; in the other case (ii), re-identification becomes necessary. The method is based on analysis of the prediction error, i.e. the difference between the actual measured output and the corresponding model prediction, and uses three main tools: a statistical (whiteness) test on the prediction error sequence, a subspace identification method to detect the order of the input-to-prediction error system, and a nonlinear optimization algorithm to recompute optimal estimator parameters. We illustrate the effectiveness of the method on a large-scale rigorously simulated industrial process.
