A machine learning approach for condition monitoring of high voltage insulators in polluted environments

Abstract

Este artículo propone un nuevo enfoque para monitorizar el estado de los aisladores eléctricos basado en la combinación de técnica de submuestreo aleatorio con un algoritmo de refuerzo adaptativo (RUSBoost), con el objetivo de estimar los indicadores clave del estado a partir de los datos meteorológicos y ambientales.

This paper proposes a new approach for insulator condition monitoring based on the combination of the random under sampling technique with an adaptative boosting algorithm (RUSBoost) and aiming to estimate key condition indicators from the meteorological and environmental data. The research was conducted at a 245 kV test station located in a severely polluted area in France, where one glass insulator string and two mirroring strings, but composed by full and half silicone-coated (bottom surface only) glass insulators, were monitored in real operational conditions during two consecutive years. The definition of the condition indicators was carried out through the characterization of the leakage current obtained in laboratory tests, subjecting the glass insulator string to different artificial pollution levels until flashover. Afterwards, the performance of the new proposed RUSBoost approach was evaluated and compared with AdaBoost, Bagging, Random Subspace Ensemble with k-nearest neighbors (KNN) and support vector machines (SVM) algorithms. The results show the effectiveness of RUSBoost in addressing the estimation of the highly imbalanced insulator condition indicators and its advantage over other methods by achieving a macro-averaged F-score of 0.757 for the non-coated string and a F-score of 0.768 for the half-coated string and 0.792 for the full coated string.