An extended self-organizing map for nonlinear system identification

Local model networks (LMN) are employed to represent a nonlinear dynamical system with a set of locally valid sub-models across the operating range. A new extended self-organizing map network (ESOM) is developed for the identification of the LMN. The ESOM is a multi-layered network that integrates the basic elements of traditional self-organizing maps and a feedforward network into a connectionist structure which distributes the learning tasks. A novel two-phase learning algorithm is introduced for constructing the ESOM from plant input-output data, with which the structure is determined through the self-organizing and the parameters are obtained with the linear least squares optimization method. The predictive performance of the model derived from the ESOM is evaluated in three case studies. Simulation results demonstrate the effectiveness of the proposed scheme in comparison with other methods.