ENHANCING THE PREDICTION STAGE OF A MODEL PREDICTIVE CONTROL SYSTEMS THROUGH META-CLASSIFIERS

Abstract

A Model Predictive Control (MPC) is a system which allows to control a production plant. Thanks to this type of system, it is possible to make a production close to zero defects. To achieve its main goal, this kind of systems is composed of several phases. One of the most critical one is the phase that predicts the plant situation for a given time. Currently, the majority of the research in this field is related to linear MPCs, although the process may not be. Previously, we presented several experiments that prove that the forecast phase, usually represented by a single mathematical function, can be represented by machine-learning models. Nevertheless, the employment of standalone classifiers raises some limitations. In this paper, we extend our previous research and we propose a general method to foresee all the defects, building a meta-classifier using the combination of different methods and removing the need of selecting the best algorithm for each objective or dataset. Finally, we compare the obtained results, showing that the new approach obtains better results, in terms of accuracy and error rates.