Ensemble learning methods using the Hodrick–Prescott filter for fault forecasting in insulators of the electrical power grids

Abstract

Electrical power grid insulators installed outdoors are exposed to environmental conditions, such as the accumulation of contaminants on their surface. The contaminants increase the surface conductivity of the insulators, increasing leakage current until there is a flashover. Evaluating the increase in leakage current in relation to the contamination level is one way to determine the insulation condition. This paper evaluates a time series of leakage current from a high-voltage laboratory experiment using porcelain pin-type insulators. Time series forecasting is performed with a collection of machine learning models known as ensemble learning approaches, which include blending, bootstrap aggregation (bagging), sequential learning (boosting), random subspace, and stacked generalization. According to this paper’s findings, applying these ensemble learning approaches is useful for enhancing the performance of the machine learning models in forecasting the occurrence of breakdowns in the electrical power system. The Hodrick–Prescott filter reduces the root mean square error performance metric (to be minimized) by 2.69 times using the ensemble random subspace approach. According to the results of this paper, the proposed method is stable, with low variance when a statistical analysis is performed, being superior to the long short-term memory neural network.