Bayesian Duval Triangle Method for Fault Prediction and Assessment of Oil Immersed Transformers

Abstract

Dissolved gas analysis of transformer insulating oil is considered the best indicator of a transformer’s overall condition and is most widely used. In this study, a Bayesian network was developed to predict failures of electrical transformers. The Duval triangle method was used to develop the Bayesian model. The proposed prediction model represents a transformer fault prediction, possible faulty behaviors produced by this transformer (symptoms), along with results of possible dissolved gas analysis. The model essentially captures how possible faults of a transformer can manifest themselves by symptoms (gas proportions). Using our model, it is possible to produce a list of the most likely faults and a list of the most informative gas analysis. Also, the proposed approach helps to eliminate the uncertainty that could exist, regarding the fault nature due to gases trapped in the transformer, or faults that result in more simultaneous gas percentages. The model accurately provides transformer fault diagnosis and prediction ability by calculating the probability of released gases. Furthermore, it predicts failures based on their relationships in the Bayesian network. Finally, we show how the approach works for five distinct electrical transformers of a power plant, by describing the advantages of having available a Bayesian network model based on the Duval triangle method for the fault prediction tasks.