ELEMENTARY PERCEPTRON AS A TOOL FOR THE TRANSIENTS ANALYZING

Abstract

The use of elemental perceptron as the simplest artificial feedforward neural network is proposed to evaluate transient processes in electrical networks. Signals with random amplitude, phase, frequency and attenuation were used to test the neural network algorithm as well as the superposition of an aperiodic component, also having a random amplitude and a time-constant. Each signal from the sample was thus determined by six independent random parameters, varying in different ranges. Based on the results of numerical modeling it is shown that such signals are typical for oscillograms of current at short circuits on power lines. It is shown that at the frequency of digitization of signals of 600 Hz in measuring organs on the time interval during industrial frequency it is possible to assess the parameters of a transition process with the accuracy not lower than several percents. The accuracy of the definition for each parameter depending on the range of their variation is analyzed. Transition process parameters that have the greatest impact on neural network training and testing errors are highlighted. Estimates of the possible running speed of the proposed neural network algorithm are made.