Optimal scheduling of power transformers preventive maintenance with Bayesian statistical learning and influence diagrams

Abstract

Transformer preventive maintenance scheduling—as an integral part of the power utility company’s asset–management process—if optimized properly affords considerable savings in maintenance costs and results in the increased reliability of transformer stations. However, situation is exacerbated by the considerable uncertainty involved in the assessment of the transformer health status, not necessarily associated with its age, as well as in the projected asset replacement and maintenance costs. There is also uncertainty inherent in the transformer tests as such. This paper proposes a novel transformer optimal preventive maintenance scheduling approach, with the use of the Bayesian statistical learning and a generalization of the Bayesian networks, for decision-making based on the expected utility hypothesis. Proposed approach, in essence, manages risk associated with taking different preventive actions by leveraging uncertain information regarding the transformer health and the expected costs of these actions. In this way, each prospective decision has a monetary value attached to it, which depends on the particular circumstances defined by the power utility company. Time-related effects are introduced into the decision-making process by the Bayesian updating of the prior belief concerning the transformer health with new information. Solving a decision problem amounts to determining an (optimal) strategy that maximizes the expected utility for the decision maker. On top of that, Bayesian approach, uniquely, offers explicit account for the full uncertainty associated with every step of the decision-making process. The proposed approach, and its advantages, will be demonstrated on a fleet of power transformers. It will be shown that model can predict with a 90% accuracy and a high confidence the transformer health index class from a small dataset. It will also be shown that it can predict with high expected profits which transformers in the dataset would benefit the most from the early intervention (recuperating around 25% of the projected costs and retaining over 90% of the projected savings).