Comparisons of artificial neural networks for on-line identification of a nonlinear multivariate electromechanical process

Abstract

This paper aims at assessing the performance of three different artificial neural network (ANN) structures for identifying typical power-system dynamics. The first ANN, chosen as reference, is the popular multi-layer perceptron (MLP) equipped with taped-delay lines. The second pertains to the family of feedforward neural networks with first-order filters added locally to the neurons, while the third is recurrent in the usual sense, with an architecture that mimics a nonlinear discrete state-space system. In contrast with the MLP, the two latter ANNs theoretically allow system dynamics to be identified without having to feed past inputs and outputs explicitly. Based on realistic data obtained by simulating a line fault with a sample hydro-generator connected to an infinite bus, it is shown that all ANNs can successfully identify a three-input four-output model of the underlying electromechanical process. However, their performance varies widely, according to their numerical complexity, convergence characteristics and accuracy in predicting the system behavior for new inputs not seen during training.