The purpose of this paper is to explore three different machine learning models to diagnose fault types inside power transformers. Power transformers are a key and expensive part of transmission and distribution systems. The traditional way of diagnosing faults has been to analyze the cooling oil for the presence of dissolved gases. From these gasses, methods of diagnoses have been developed such as Duval’s triangle and the IEEE key gas concentration method. While these methods can be accurate in some cases, they do not consider the change in gas levels over time. This paper aims to develop models that use seven raw gas values over a period of five years to output a correct fault type for 12kV and above power transformers. Previous works have applied machine learning algorithms to a single sample of Dissolved Gas Analysis (DGA), however this paper aims to include multiple years of data. The rate of change of combustible gasses can be factored in when using multiple years of data, which is the key difference between this paper and previous work. Of the methods considered, it was found that Artificial Neural Networks (ANNs) had the best accuracy. An accuracy rate of 89% was achieved when using four hidden layers training on three years of data. K-Nearest Neighbors (K-NNs) and Support Vector Machine (SVM) also achieved adequate classification results with the best cases being 88% and 86% respectively. Previous work had an accuracy rate of 91% on a single sample of DGA classified using a single conventional method. An accuracy of 89% on multiple years of data is notable, because of the increased complexity of the training data and more in-depth method of classification.

Artificial Neural Network Dissolved Gas Analysis K-Nearest Neighbors Machine Learning Power Transformer Support Vector Machine