Résumé
Power transformers are the critical and costly resources in the electric power system including more than 50% of total investment and requiring periodic diagnostic and maintenance. Actually, different condition monitoring and diagnostic techniques (e.g., oil characteristics, dissolved gas analysis, winding measurements) are currently used to check the power transformer components' health (e.g., winding, bushing, tank, core). However, these methods are generally offline requiring a holding time for maintenance and a total stop of the system. These condition-monitoring techniques are realized after a failure of the power transformer or during routine diagnostics in order to identify significant parameter changes that are indicative of a developing fault. In this paper, an online data mining approach is used to detect faults in power transformers based on the real-time analysis of electrical and thermal measurements without the need to stop the system. A power distribution transformer dataset is used to train machine learning algorithms, including transformer oil level gauge, transformer oil temperature, oil temperature indicator alarm, and phase-lines voltage. Predictions are used to generate predictive alerts in case of efficiency and health degradation of the power transformer. The results obtained show the performance of the proposed method to detect the abnormal operation of the power transformer and identify the need for maintenance before the development of the issue to catastrophic failures. The temperature and phases voltages are used to train machine learning models in order to detect the anomaly. The detected faults are compared to a set of alerts history showing the usefulness of the proposed technique.