On the use of minimal parametrisations in multivariable ARMAX identification (Q1267250)

The paper is devoted to the problem of determining the order and the structure of minimally parameterized ARMAX models for linear multivariable systems. A new minimal parameterization for multivariable ARMAX predictors is given. A complete prediction error identification procedure is proposed. Also an illustration of possible non-identifiability conditions associated with the non-minimal parameterization is presented. It is also shown that the determination of a suitable order and structure for a model of a real process does not involve critical choices. A description of a complete and easily implementable prediction method procedure is provided and an example of the procedure is shown, using real data.