Failure Prediction in Automatic Reclosers Using Machine Learning Approaches

Abstract

Industry 4.0 opened new frontiers of the Smart Grid (SG) area to understand the behavior of the power grid assets. The asset sensor data stored has been supporting the analysis of the life cycle. Also, it helps to improve predictive maintenance activities. On the other hand, Machine Learning approaches have been achieving the current state-of-the-art in many problems of the SG area. In this context, this paper proposes a preliminary study using ML methods to predict failure in Automatic Reclosers (ARs). In this work, we compare five ML methods: Naive Bayes (NB), Deep Neural Network (DNN), Decision Tree (DT), Gradient Boosted Tree (GBT), and Random Forest (RF). A dataset was generated with temperature information and historical data of the ARs sensors. The failures were labeled using a rule-based approach proposed by the AR specialists. Among the ML methods, RF obtained the best result with 82.47\% F1-Score in the test set, indicating a promising result for the failure prediction area.