In electrical distribution systems, unforeseen transformer faults have caused service interruptions, deteriorated supply quality and economical damage on unprecedented scale over the past decades. Although predictive maintenance has adopted machine learning techniques, most existing studies have concentrated on classification accuracy without taking into proper account probability calibration, economic consequences modeling, or time resilience. This paper introduces an economic risk and machine learning model for the prediction of rare events in transformer burnout within distribution grids. One years of operational data are employed to train models using historical data and to test them on a future year to mimic real operating environments. A number of imbalanced learning strategies and ensemble algorithms are utilized and evaluated using precision-recall metrics, which account for the rarity of transformer failures. The study involves the use of probability calibration to make decisions more reliable and to maximize the benefits of risk management in terms of monetary value. Furthermore, the predictive results are attributed to engineering-relevant stress factors, such as outage exposure, network density, and environmental conditions, using explainable analysis. In addition to predictive performance, the expected failures are converted into expected monetary risk is provided by accounting for replacement costs, outage value, and preventive maintenance efficacy. A life-cycle economic model, including discounting, is applied to determine the optimal inspection portfolio to maximize net present value and support reliability-centered maintenance decision making. Sensitivity analysis and Monte Carlo simulations are used to indicate the robustness of the economic results in the face of parameter uncertainty.

Keywords

Cost-Sensitive Learning, Economic Risk Modeling, Transformer Failure Prediction, Distribution Network Reliability and Smart Grid Asset Management.