Modeling Non-Linear Process Dynamics Using Partially Recurrent Neural Networks

Abstract

The representation of non-linear process dynamics can cause problems for feedforward networks, and recurrent networks hold promise in this area. Two partially recurrent neural network structures with local output feedback are investigated, along with four training algorithms, for their ability to model non-linear processes. The training algorithms studied included two well-known backpropagation techniques and are all shown to be derivatives of the more general recursive prediction error (RPE) method. Networks were trained and tested on data from a simulated, non-linear chemical process. The performance of the partially recurrent networks is compared to a standard time delay neural network model. The results show that partially recurrent networks can model as well as time delay networks. Issues relating to the implementation of partially recurrent networks for non-linear modelling are discussed.