Power Transformer Failure Prediction: Classification in Imbalanced Time Series

Luís Guimarães,Vera L Miguéis,Eduardo E Oliveira,José Borges

doi:10.24840/2183-6493_003.002_0004

Abstract

This paper describes a study on applying data mining techniques to power transformer failure prediction. The data set used consisted not only on DGA tests, but also in other tests done to the transformer’s insulating oil. This dataset presented several challenges, such as highly imbalanced classes (common in failure prediction problems), and the temporal nature of the observations.To overcome these challenges, several techniques were applied for prediction and better understand the dataset. Pre-processing and temporality incorporation in the dataset is discussed. For prediction, a 1-class and 2-class SVM, decision trees and random forests, as well as a LSTM neural network were applied to the dataset.As the prediction performance was low (high false-positive rate), we 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.

Highlights

Asset management plays a central role in ensuring that operational and investment costs are minimized, and ensuring the quality of products or services
Note that increasing the time lag leads to a decrease in the number of available instances, as there is a need to eliminate missing values which will originate in the first instances of each transformer
When it comes to decision trees or random forest, it is unclear which is the best criteria (Gini or entropy), but it seems evident that limiting the number of leaves has a positive effect on performance, especially in recall, while maximum depth should not be controlled

Summary

Introduction

Asset management plays a central role in ensuring that operational and investment costs are minimized, and ensuring the quality of products or services (i.e., low amount of defects, no energy shortages, low delivery times, etc.). Being able to predict failure in physical assets/equipment can be a powerful tool to aid asset management, as it can help determine the best time for maintenance actions, minimizing the amount of such actions while improving the availability of the equipment. This paradigm of using failure predictions to determine maintenance actions is called predictive (or condition-based) maintenance. We use data that is already collected for safety inspection and non-predictive maintenance This data was collected in order to evaluate the need for a change of isolating oil due to continuous or sudden degradation.

Results

Discussion

Conclusion