Identification of Multivariable, Linear, Dynamic Models:  Comparing Regression and Subspace Techniques

Dynamic models identified as ARX or FIR models using two regression techniques, PLS and CCR, are compared with state-space models identified using a predictive error method and two subspace algorithms, CVA and N4SID. The objective functions for PLS and CCR are shown to be related. A comprehensive simulation study of several processes with different characteristics and noise properties is used to compare the identification methods. The results indicate that, if the time delay structure is known or estimated accurately, the identified subspace models tend to be more accurate than the models identified using regression. The state-space models identified using the CVA algorithm are especially accurate.