NARX model identification with error filtering

Model identification of polynomial NARX models involves a lengthy and computationally intensive procedure for selecting the model structure among a possibly large set of candidate regressors. If the model structure is under-parameterized to reduce the burden of the model selection phase, unsatisfactory results are generally obtained. This inaccuracy problem can be somewhat circumvented by focusing the identification process on the obtainment of an accurate local model over a specific frequency range. Such frequency tailoring is achieved in the nonlinear modeling framework by direct error filtering, as opposed to the data pre-filtering practice adopted in the linear context. This work discusses the application of error filtering to classical NARX model identification methods. A simulation example is provided to show the performance of the proposed approach in deriving accurate local models, despite an underparameterized model structure. It is also shown that a proper error filtering may increase the model accuracy in simulation with respect to available identification techniques.