Muhammad Akmal A. Putra, Muhammad Rafi Sudrajat, Suwarno, Rahman Azis Prasojo
Early fault detection is critical to enhancing transformer reliability and preventing system failures that could result in substantial losses. One primary approach to diagnosing transformer faults is using Dissolved Gas Analysis (DGA). However, the limitations of raw DGA data present a significant challenge in developing accurate and reliable machine learning (ML) models. This study proposes the use of the Gaussian Process Classifier (GPC) method to address data availability limitations in ML modeling, with data validation from IEC TC-10 and other real-world cases. The GPC method not only achieves higher prediction accuracy for smaller data but also provides uncertainty levels for its predictions. This feature offers a key advantage over other ML models, which typically do not provide true uncertainty quantification. Through comparative analysis with models such as Random Forest, Support Vector Machine, and Neural Networks, the study demonstrates that GPC outperforms other methods in scenarios with limited data and offers additional insights through uncertainty quantification. This capability makes GPC a more reliable solution for transformer fault diagnosis, thereby enhancing decision-making confidence in the field, where quick decisions are paramount. The findings of this study are expected to significantly contribute to the development of more reliable prediction systems while paving the way for broader applications of the GPC approach in machine learning-based diagnostics. © 2025 IEEE.
Institut Teknologi Bandung, School of Electrical Engineering and Informatics, Bandung, Indonesia; Australian National University, School of Computing, College of System and Society, Canberra, Australia; Politeknik Negeri Malang, Department of Electrical Engineering, Malang, Indonesia