Utilização de auto-consistência como ferramenta auxiliar na seleção de estrutura de modelos Narx Polinomiais

Abstract

The structure selection of models used to represent dynamics described by data is a crucial problem in system identication. The ERR index (error reduction ratio), which is based on the prediction error minimization (one step ahead), despite being a criterion widely used in structure selection, can choose incorrect or redundant terms in non ideal identication conditions, that is, when the available data are not suitable (oversampled or noisy) or when the input signal is relatively slow. On the other hand, the SRR criterion (simulation error reduction ratio), dierently from the ERR, may be eective in non ideal identication conditions. Moreover, SRR yields more compact models that are, therefore, more robust. However, such criterion, which is based on the simulation error minimization (free-run prediction), requires a signicantly large computational eort. Thus, in this work, a criterion based on the prediction error of two steps ahead minimization (ERR2) is proposed to be applied on the cases in which the input signal is relatively slow. To accomplish that, were investigated ve cases studies: three with simulated data and two cases with experimental data from real systems. The results presented here show that the use of self-consistency between the criteria ERR and ERR2 can assist in the selection of structure of polynomial NARX models.