SYSTEMS IDENTIFICATION USING NEURAL NETWORKS AND EXTREME LEARNING MACHINE

Systems identification is the procedure through which mathematical models are obtained from observed data of a given process. An effective way to perform system identification is through neural networks. However, traditional methods of training feedforward networks require several iterative cycles to adjust weights, which makes this procedure very slow. In this paper we present the algorithm Extreme Learning Machine, used for training single-hidden layer feedforward neural network. This algorithm allows the neural network to achieve the best learning performance, at an extremely high speed. Experimental data from tests performed on a buck converter are used. The best network topology is determined by cross validation. The Extreme Learning Machine and Levenberg-Marquardt algorithms are applied to a single-hidden layer feedforward neural network to perform the buck converter identification. The results of simulations indicate that the algorithm Extreme Learning Machine allows to obtain smaller validation errors in very reduced periods of time.