Practical considerations in the application of neural networks to the identification of harmonic loads

Abstract The application of artificial neural networks to the identification of harmonic loads has been demonstrated before with great success. In this paper, practical aspects of such an application are discussed. While it is known that the selection of features, the choice of the architecture (number of hidden layers) and topology (number of hidden units in each hidden layer) of the neural network are heuristic decisions involving engineering judgment, this paper shows one such way and claims success in that implementation. A supervised learning procedure is adopted and the familiar back-propagation technique is used in training the network. The need for preprocessing raw data from a power system before being input to a neural network is examined. Finally, the performance of the neural network is evaluated with the aid of a benchmark case.