Model Mismatch Assessment Of Multivariate Model Predictive Control

The performance of model predictive controller may decline and even invalid because the work point of processes may drift to another one and the actual processes are always non-linear and several other unknown external disturbances will also influent the processes. If the performance of model predictive controller is not improved in time, the model predictive controller will not increase the economic efficiency as much as it can. As a model-based control algorithm, the performance of controller will be hard to be improved only by re-tuning the parameters of controller if there is serious mismatch between model and plant. Model mismatch is a common problem in model predictive control, as well as one of the most important reasons of bad performance. As a result, it is important and urgent to assess model quality and provide guidance to the maintenance of model predictive controller.In this paper, the model plant mismatch problem is classified and defined. An existed method of model plant mismatches assessment formulated in terms of discrete time state space model is introduced. Then a method of model plant mismatch assessment in terms of non-parameter model is proposed. The relationship between model plant mismatch and model output residues is analyzed and a pair of time series sequences which contains mismatch information is found. A new method based on whiteness test of a time series sequence using statistical inference is presented and it is promoted to on line control performance monitoring.The mismatch of each channel needs to be evaluated in multivariate process. To mine model mismatch of each channel of multivariate model predictive controllers much correctly, another new scheme based on partial correlation analysis is demonstrated. This method can depart the coupling between process variables and isolate model mismatch information of each channel. A new model mismatch index is defined to express the mismatch quantity exactly. Numerical simulation examples have shown that it can detect model mismatch of each channel effectively. What's more, this new method can deal with non stationary disturbance as well.