A neural network structure with parameter expansion for adaptive modeling of dynamic systems

Abstract

A new neural network structure for adaptive modeling of dynamic system is presented in this paper. Based on multi-layer perceptron (MLP), the network possesses parameter expansion and external recurrence. Parameter expansion is obtained by using tapped delay lines (TDLs) to the outputs of the hidden layer. This increases the number of parameters between the hidden layer and the output layer. Furthermore, external recurrence is obtained by connecting the output and the input of the network. Proper learning algorithm is derived to accommodate the aforementioned modifications. Afterwards, the network is integrated in an adaptive scheme so that it can model systems with changing property or operating condition. The application in modeling of a water tank system demonstrates the ability of the proposed scheme.