Identification of nonlinear dynamic systems using higher order diagonal recurrent neural network

Abstract

A new neural network architecture, called a higher order diagonal recurrent neural network (HDRNN), is presented. The architecture of an HDRNN is a modified model of the diagonal recurrent neural network (DRNN) with one hidden layer which is composed of self-recurrent neurons and additional multiplication inputs between conventional inputs and self-recurrent neurons. The authors derive a generalised dynamic backpropagation algorithm and show that the proposed HDRNN not only gives more accurate identification results, but also requires a shorter training time to obtain the desired accuracy.