Machine Learning Tool for Fault Prediction in Electric Grid Transformers

Abstract

This work aims for the development of a tool for forecasting faults within the transformers connectedto the electric grid in order to support the maintenance team. With this purpose, the fault databaseis treated and converted to a list of monthly time series, one for each transformer, new features arecalculated based on the information on the database, followed by the fault risk estimation for thenext month in the series for each transformer and the ranking of the transformers. The riskestimation is done using the Pointwise Learning to Rank (LTR) methodology, in which the items on alist are ranked based on metric calculated using classifiers or regressors. A significant gain inperformance is demonstrated due to the feature engineering process.