Implementation of Self-Organizing Map and Logistic Regression in Dissolved Gas Analysis of Transformer oils

Abstract

The occurrence of incipient faults in a transformer is attributed to several types of thermal, electrical, chemical and mechanical stresses which deteriorate the transformer insulation and cause ageing. Thus it is of utmost importance to carry out periodic maintenance of this electrical equipment as well as to develop a method that would provide an early stage diagnosis of the transformer insulation abnormalities. Dissolved Gas Analysis (DGA) is widely considered to be a powerful approach to detect the incipient faults in oil-immersed transformers. But shortcomings in the conventional methods based on DGA are nowadays addressed by various intelligent techniques with enhanced accuracy of fault detection. This paper introduces the implementation of machine learning and proposes a method which utilizes Self-Organizing Map (SOM) and Logistic Regression (LR) to detect and predict the faults occurring in a transformer based on DGA. The DGA dataset used in this paper consists of a variety of cases with six types of faults. The proposed model presents the interpretation of this dataset using clustering analysis by SOM and the model performance parameters are found to be superior to other machine learning algorithms as well as traditional methods. Unlike other fault diagnostic methods which mostly implement single classifiers, this technique uses clustering of the DGA data followed by the application of a classification algorithm which demonstrates high accuracy and reliability of fault identification.