Transformer health index calculation using evidence theory and considering online monitoring data and machine learning

In this paper, methods for calculating a reliability score for distribution power transformers are investigated. The analysed algorithms employ a worst-case approach, a simple numeric score, a non-linear score, a combination of belief functions and, as an outlook, an unsupervised machine learning solution. The first-mentioned algorithms are described in technical brochure 761 of CIGRE. The other approach is a generalisation of Bayesian probability theory that allows belief and plausibility functions to represent uncertainty.

In the following, the condition indicators for the calculations are offline values from real transformers, but relevant online monitoring parameters are identified and discussed. The primary goal is to incorporate more input from these systems and use the investigated algorithms to get a prioritised list of the transformer fleet. This list can help to select transformers for repair, change or retrofit. A case study involving 56 power transformers with real data was conducted, and condition indices were assigned. The transformers vary in age, operational status, protection system history, oil sample results, and retrofit status.

In the present work, the transformer is divided into four subcomponents. That means the active part, the insulating oil, the bushings, and the on-load tap-changer (OLTC). The categorisation scheme comprises six categories for each condition indicator, ranging from good to bad, and the thresholds are guided by CIGRE, IEEE, IEC, or own specifications. Each subcomponent is assigned a score, and in combination, these scores yield an overall health index for the transformer.

The given results are plausible, and the overall indices based on mass functions provide a finergrained classification. In addition, the ability to model uncertainty is a significant advantage.