Caracterização de agrupamentos de termos na seleção de estruturade modelos polinomiais NARX

Abstract

The modelling system is present in almost every areas of science. However, it is not easy model systems using physical laws that describe the process dynamics. So, the systems identification is a very feasible alternative. By measurements of input and output are found models which describe the process dynamics. Although the easy parametrization of these models, the determination of a method for structure selectionof models that best adjust the system data has not yet been completely discussed. This problem worsens in applications which require a non-linear model. Hence, choosing model structure is essential to avoid overparametrization problems. As possible solution, this work proposes a new methodology for selection of NARX polynomial models structures, named Modelling via Monte Carlo simulations with Constraints. Where, using the concept of clusters of terms, the model structure is a variablebased on approximations of static characteristic of the system. It is applied a procedure of random generation of parameters, and selection of the best models achieved. After testing several approaches for characteristic static, the eective clusters are selected. Along the text, eorts were directed to the following objectives: (I) interpret theproblem of structure selection of models and propose a new methodology, (II) obtain through simulations of experimental and simulated systems, situations in which the algorithm implemented is applicable and (III) to compare the results with other classicalmethods. The results showed that the use of the new method it is possible to distinguish between under and overparametrization structures of NARX polynomial models.