Adaptive Neural Networks for Nonlinear Dynamic Systems Identification

Abstract

A new scheme for adaptive neural networks for nonlinear dynamic system identification is proposed in this paper. The network of structure multi-layer perceptron with external recurrence is trained offline at first to get the initial network parameters. The parameters of the network are classified into short-term memory part and long-term memory part. The short-term memory part includes the parameters which are linear to the network output. In the implementation, the network is validated in each sampling time using a set of new measurement data. Training procedure will be executed if the model error exceeds a specified value and the short-term memory part will be adjusted. The application in modelling of room thermal behaviour demonstrates the performance of the proposed scheme.