Fault Prediction for Power Transformer Using Optical Spectrum of Transformer Oil and Data Mining Analysis

Abstract

Periodic preventive maintenance of power transformer should be conducted for its health monitoring and early fault detection. Transformer oil is a vital element where its contents and properties need to be monitored during the service life of a power transformer. This paper presents an optical spectroscopy measurement from 200 nm to 3300 nm to characterize the transformer oil, which were sampled from the main tanks and `on-load tap changer' of power transformers. The correlation of the optical characteristics in the range of 2120 nm to 2220 nm to the Dissolved Gas Analysis results and Duval Triangle interpretation demonstrates that the low energy electrical discharges, high energy electrical discharges as well as the thermal faults rated at temperatures above 700°C in power transformers can be accurately predicted. For faster and accurate analysis of fault prediction, a data mining analytics tool was constructed using Rapid Miner server to analyze and verify the predictions for a total of 108 oil samples. For the optimization, continuous iterations were performed to determine the best absorbance-wavelength combination that can improve the accuracy of the prediction. The performance of the optical spectroscopy technique integrated with data analytic tool was analyzed and it was found that the technique contributes to a high accuracy of 98.1% in fault prediction. It is a cost-effective and quicker complementing approach to carry out pre-screening of the transformer oil in order to know the condition of the power transformers based on the transformer oil's optical characteristics.

Keywords

• Data mining
• fault prediction
• optical spectroscopy
• power transformers
• transformer oil