Muhammad Akmal A. Putra, Rahman Azis Prasojo, Suwarno
Previous studies have extensively discussed the use of machine learning (ML) algorithms for transformer fault identification. However, the unavailability of data and the resulting imbalance can reduce the accuracy of the developed models. This study aims to develop a ML-based system for transformer fault identification using a Multi-Method approach with imbalanced Dissolved Gas Analysis (DGA) data modified by various oversampling techniques, this technique aims to balance class distribution by increasing the number of samples in the minority class. Thus, the developed ML model can achieve good performance even when using an imbalanced dataset. The methods used for DGA identification in the Multi-Method approach include Roger Ratio, IEC Ratio, Duval Triangle, and Duval Pentagon. The dataset consists of 343 imbalanced data points, which are then processed using oversampling techniques such as SMOTE, ADASYN, SMOTE-Tomek Link, and SMOTE-ENN. Neural Network, Random Forest, and Naïve Bayes algorithms are employed as ML models in this study. The best combination is achieved with the application of a Neural Network on a dataset modified using SMOTE-ENN, yielding a F1 score of 0.994 and a Precision of 0.994. The results of this study indicate that data balance and the number of misclassified classes can significantly impact the accuracy of the developed model. The developed system model can assist utilities in performing maintenance and making decisions based on DGA results for power transformers. © 2024 IEEE.
Institut Teknologi Bandung, School of Electrical Engineering and Informatics, Bandung, Indonesia; Politeknik Negeri Malang, Department of Electrical Engineering, Malang, Indonesia