Applications of artificial neural networks to the identification of dynamical systems

Abstract

This paper presents two types of artificial neural network (ANN) for application to the identification of dynamical systems. The first pertains to the family of feedforward neural networks with temporal recurrent elements added to the neurons. This structure allows memory neuron networks to identify systems without having to feed past inputs and outputs explicitly. The second ANN is recurrent Its architecture looks like a discrete state-space system with a sigmoidal function in the recurrent loop. The main attraction of this three-layer recurrent neural network is its simplicity of use and the faster speed of convergence of the learning phase. The two ANNs have been tested on two dynamical systems, a fifth-order discrete theoretical plant with many multiplications between internal states in order to introduce nonlinearities, and a nonlinear transfer function from the terminal voltage to the magnetizing flux in a power transformer The challenge the ANNs is to catch the ferroresonance phenomenon as seen from the primary of the set-up transformer after a fault. The performance of these ANNs is discussed in light of various aspects of their utilisation. The comparison is based on important points such as difficulties of use, their speed and ability to converge, and their ability to generalize the behaviour of the system to inputs not available for training.