An Automated Method For Model-Plant Mismatch Detection And Correction In Process Plants Employing Model Predictive Control (MPC)

A model-predictive controller (MPC) uses the process model to predict future outputs
of the system. Hence, its performance is directly related to the quality of the model. The
difference between the model and the actual plant is termed model-plant mismatch (MPM).
Since MPM has significant effect on MPC performance, the model has to be corrected and
updated whenever high MPM is detected. Re-identification of the process model with large
number of inputs and outputs is costly due to potential production losses and high manpower
efforts. Therefore, detection of the location of the mismatch is needed so that only that
channel is re-identified.
Detection methods using partial correlation analysis as well as other methods have
been developed, but these are qualitative methods that does not indicate the extent of the
mismatch clearly and whether or not corrective action is necessary. The proposed
methodology of this project uses a quantitative variable (e/u) which is the model errors
divided by the manipulated variables, to identify changes in the plant gain and hence the
mismatch. Taguchi experiments were carried out to identity the most contributing gains to the
overall process, and then focus on these major contributors to find the threshold limits of
mismatch by trial and error. When the mismatch indicated by the variable (e/u) exceeds the
threshold limit, auto-correction of the model gain of the controller is made to match with the
new plant gain.
The proposed method was assessed in simulations using MA TLAB and Simulink on
the Wood and Berry distillation column case study and was successfully validated. Testing
for various mismatch scenarios for both two major contributors to the process, the algorithm
was able to bring the output back to the desired set-point in a very short time.