Local Model Networks for the Identification of Nonlinear State Space Models

Abstract

A novel algorithm for the identification of nonlinear state space models is proposed. The local model state space network (LMSSN) uses local model networks for the approximation of state and output equations of a nonlinear state space model. Thereby, the LMSSN is trained with an adapted version of the local linear model tree (LOLIMOT) algorithm. The combination of nonlinear state space models with the LOLIMOT algorithm is utilized in this form for the first time. Especially the rescaling of the state trajectory, the possibility to perform splits within the state dimensions, and the local model error estimation constitute novel ideas compared to previous works. It is shown that the proposed method performs superior to other dynamics realizations and comparable to other state space approaches on a hysteresis benchmark.