Constellation, le dépôt institutionnel de l'Université du Québec à Chicoutimi

Résumé

Power transformers represent one of the most abundant and expensive components in the electric power industry. Dissolved gas analysis (DGA) of transformer is the most widely accepted diagnostic tool across the globe to understand insulation incipient failures. Nevertheless, DGA fault gas interpretation is a remarkable challenge for transformer owners and utility engineers. Several computational techniques have been adopted for DGA fault classification along with offline methods. However, limited data availability, high ambiguity in DGA interpretation, suitability, and model accuracy are critical challenges in the DGA fault classification using computational techniques. In this work, highly diverse and large DGA data samples of in-service transformer fleets from five different utilities have been used to develop an efficient fault classification methodology. A total of 4580 DGA samples and IEC TC 10 database are used for training and testing, respectively, for various machine learning algorithms. Discussions on performance indicators and evaluation of several algorithms to verify the most suitable class algorithms are also the focus of this work. Furthermore, a best-performing model is identified based on various performance indicators. The hyperparameters of the best model are further tuned to achieve a most precise fault classification. It is inferred that non-parametric methods and non-linear SVM are best suitable for transformer DGA fault classification. Importantly, the rankings in the present study suggest that transformer DGA fault prediction is better with ensemble learning methods.