Investigation on machine learning algorithms to support transformer dissolved gas analysis fault identification

Dissolved gas analysis (DGA) is a powerful tool to monitor the condition of a power transformer. Several interpretation methods have been proposed, one of the most reliable of which is the graphical Duval triangle method (DTM). The method consists of several triangles, which still requires expertise for fault identification. The use of computer-based technology has been implemented in recent years to support transformer fault identification. However, no study has done thorough investigation on the use of suitable machine learning algorithm for the ML-based implementation of this matter. This study examines six commonly used machine learning algorithms to support DGA fault identification of power transformer: decision tree, support vector machine, random forest (RF), neural network, Naive Bayes, and AdaBoost. Three DGA fault identification methods for mineral oil insulated transformer were studied, namely DTM1, DTM4, and DTM5. The training and testing datasets were generated for each DGA method, and trained to each ML algorithm. The tenfold cross validation was used to evaluate the results using five criteria, namely classification accuracy, area under curve, F1, Precision, and Recall. RF models demonstrated the best performance in classifying fault codes of most DGA methods. A validation was carried out using the validation dataset, comparing the selected RF-based models to the graphical DGA fault identification. The combination method was also implemented in the developed model. The results show that the proposed model is reliable, and especially useful to be used for fault identification of a large number of transformer populations.

Automated summary

Dissolved gas analysis is a powerful tool for monitoring the condition of power transformers. The graphical Duval triangle method is one of the most reliable fault identification methods. This study compares six commonly used machine learning algorithms and validates the proposed model using a validation dataset, demonstrating its reliability and usefulness in fault identification for a large number of transformers.