A PLS approach to identifying predictive ARX models

MPC (Model Predictive Control) based on linear models is an extensively used methodology in the industrial field as a control solution for MIMO processes. The identification of ARX models for multivariable systems from input-output data often requires the use of LVMs (Latent Variable Methods) such as PCR (Principal Components Regression) or PLS (Partial Least Squares) due to the so called ldquocurse of dimensionalityrdquo. LVMs however, do not take into consideration the prediction horizon in which the model will be used in MPC. PLS-PH (Partial Least Squares Prediction Horizon) is presented in this paper as a modification to PLS aiming to provide a model which performs better within a given prediction horizon. The advantage of using PLS-PH is shown in a simulation example.