Power Transformer Failure Prediction: Classification in Imbalanced Time Series

This paper describes a study on applying data mining techniques to power transformer failure prediction. The data set used consisted not only on DGA (dissolved gas analysis) tests, but also in other tests done to the insulating oil of the transformer. This data set presented several challenges, such as highly imbalanced classes (common in failure prediction and maintenance problems), and the temporal nature of the observations. In order to overcome these challenges, several techniques were applied to make predictions and better understand the data set. Pre-processing and time lag incorporation in the data set is discussed. For prediction, support vector machines (SVM) for 2-class classification and 1-class novelty detection, decision trees and random forests, as well as a long short term memory (LSTM) neural network for classification were applied to the data set. As the prediction performance was low (high false-positive rate), it was conducted a test to ascertain if the amount of data collected was sufficient. Results indicate that the frequency of data collection was not adequate, hinting that the degradation period was shorter than the periodicity of data collection.