Modern machine learning techniques for power transformer condition assessment

Ma, H., Saha, T. K., Ekanayake, C. and Allen, D. (2012). Modern machine learning techniques for power transformer condition assessment. In: CIGRE Session 44, Paris, France, (1-8). 26-31 August 2012.

Attached Files (Some files may be inaccessible until you login with your UQ eSpace credentials)
Name Description MIMEType Size Downloads
Author Ma, H.
Saha, T. K.
Ekanayake, C.
Allen, D.
Title of paper Modern machine learning techniques for power transformer condition assessment
Conference name CIGRE Session 44
Conference location Paris, France
Conference dates 26-31 August 2012
Journal name 44th International Conference on Large High Voltage Electric Systems 2012
Place of Publication Paris, France
Publisher Council on Large Electric Systems (CIGRE)
Publication Year 2012
Sub-type Fully published paper
Start page 1
End page 8
Total pages 8
Collection year 2013
Language eng
Formatted Abstract/Summary
A power transformer is a critical asset in the power system. Its health condition needs to be regularly monitored and effectively assessed to avoid any catastrophic failure. In recent years a number of techniques such as Dissolved Gas Analysis (DGA), Polarization and Depolarization Current (PDC) measurement and dielectric Frequency Domain Spectroscopy (FDS) have been widely adopted by utilities for transformer condition assessment. However, there are still considerable issues and challenges, which need to be addressed for accurately interpreting measurement data obtained from these techniques and for making explicit conclusion of a transformer’s condition.

Over the past several years the authors of this paper have investigated modern machine learning techniques and their applications for condition assessment of transformers. As a consequence a number of intelligent algorithms for interpreting measured data and evaluating transformers condition have been developed. These algorithms exploit the statistical dependency between the historic datasets and the conditions of the corresponding transformers and subsequently build up an approximation model, which can be used to provide evaluation on the transformer of interest. The major benefit of machine learning algorithm lies in its capability of utilizing the information contained in historical data to investigate properties about future data. This paper details the key concepts and methodologies of modern machine learning algorithms and presents two intelligent algorithms, namely support vector machine (SVM) and self-organizing map (SOM) for transformer insulation diagnosis.

Case studies are provided in this paper to demonstrate the applicability of machine learning algorithms in transformer condition assessment. In DGA data interpretation, the machine learning algorithms are capable of indicating the fault conditions in a transformer by providing probability of each fault condition. In PDC data interpretation, the machine learning algorithms make use of historical PDC data collected from other transformers as reference to acquire the knowledge of underlying statistical relationship between the measured polarization and depolarization currents and the condition of transformers insulation; such knowledge is then utilized to evaluate the condition of transformer of interest. In FDS data interpretation, the dielectric response of composite duct insulation with different geometric combination at different temperatures and moisture contents is computed. A fitting algorithm is then applied to estimate the moisture contents in field transformers from their FDS measurements. The case studies also aim to reveal some correlations between DGA, PDC, and FDS measurements on a same group of transformers.
Keyword Condition monitoring
Dissolved gas analysis
Frequency domain spectroscopy
Machine learning
Polarization/depolarization currents
Power transformer
Q-Index Code EX
Q-Index Status Confirmed Code
Institutional Status UQ
Additional Notes 44th International Council on Large Electric Systems (CIGRE) Session.

Version Filter Type
Citation counts: Scopus Citation Count Cited 0 times in Scopus Article
Google Scholar Search Google Scholar
Created: Wed, 19 Dec 2012, 19:44:36 EST by Dr Hui Ma on behalf of School of Information Technol and Elec Engineering