Applying Artificial Neural Networks and Nonlinear Optimization Techniques to Fault Location in Transmission Lines—Statistical Analysis

Abstract

This study presents applications of artificial neural networks and nonlinear optimization techniques for fault location in transmission lines using simulated data in an electromagnetic transient program and actual data occurring in transmission lines. The localization is performed by a modular structure of 4 neural networks and by the minimization of objective functions descriptive of the problem, defined according to the parameters of the line and the type of short circuit, submitted to the methods Quasi-Newton, Ellipsoidal, and Real Polarized Genetic Algorithm. The results obtained are compared statistically with those of a classical analytical method. The analysis of the variance of location errors presented by the methods revealed, with 5% significance, statistical evidence that allowed the conclusion that the type of method used affects fault location indication. In simulated scenarios, minor errors were obtained with the neural network and larger with the analytical method. For field oscillographic, the largest errors were in the neural network; there is no evidence to reject the equality between the results of the analytical method and the nonlinear optimization techniques. The Tukey test identified no differences between the nonlinear optimization methods applied to the proposed objective functions, but the low computational cost associated with the Quasi-newton method highlights it. The nonlinear optimization methods used for the localization function proved to be promising for application in companies that operate electrical systems, providing localization errors similar to those presented by the classical analytical method.