Deep Learning to Assess Voltage Dips Validity

Abstract

In this work, a method based on Deep Learning algorithms is applied to assess the validity of voltage dips. The aim is to clean voltage dips statistics from voltage drops due to the measurement transformers saturation, by analyzing their waveforms. In the proposed solution, the signal processing problem addressed is transformed in an image recognition task. The algorithm adopted is based on a convolutional neural network properly trained on a set of images, extracted from the voltage waveforms monitored in the distribution network by the Italian research monitoring system, QuEEN, managed by RSE. The algorithm answer is Boolean: true or false voltage dip, tertium non datur. The obtained results are compared with those usually achieved by a different criterion, based on the detection of a second harmonic component in the measured voltages, already active in both the QuEEN and the Italian national monitoring system. The performances of the two methods in detecting real events with overlapping saturation are compared referring, in particular, to those events for which the traditional method cannot provide an answer (undefined cases). Their pros and cons are then discussed.