System identification using a new dynamic neural network

Abstract

A dynamic neural network model is proposed for system identification. In this network each neuron is governed by high-order dynamics. Both the order and range of dynamics can be adjusted by chaining values or the parameters associated with each neuron. With this mechanism, the network can be adapted to capture precisely the temporal characteristics of the system to be modeled. This obviates a priori grouping of system's past inputs and outputs which is required when static neural networks are used. The network learning rules are derived based on the gradient descent technique. Advantages of the proposed model over several other dynamic models are also discussed. Finally, simulation results are given to demonstrate the validity of the modal.