Intelligent Remaining Useful Life Prediction of Power Transformers Using Machine Learning

Abstract

Background: Power transformers play an integral role in power generation and transmission. Transformer failures can lead to service interruptions, reduce system reliability, and incur significant failure-related costs. Efficient transformer maintenance requires accurate condition assessment and effective maintenance decision-making. Objective: The purpose of this paper is to develop a machine-learning-based decision-support framework for power transformer maintenance that integrates health index estimation, Remaining Useful Life (RUL) prediction, and Life Cycle Cost (LCC) evaluation. Methods: The framework draws on multimodal monitoring data from 166 oil-immersed power transformers collected over a ten-year period. Six machine learning models (XGBoost, Random Forest, Support Vector Regression, Convolutional Neural Networks, Ridge Regression, and K-Nearest Neighbors) were evaluated using various feature sets. Results: Among the evaluated models, XGBoost achieved the best performance under the basic feature configuration, with an RMSE of 0.285 years, an MAE of 0.199 years, and an R² of 0.956. The robustness evaluation showed that the model retained satisfactory predictive performance under moderate feature reduction. Based on the Life Cycle Cost assumptions applied in this study, the predictive maintenance scenario was estimated to reduce the total cost of the transformer fleet by approximately 47%. Conclusion: The proposed approach combines accurate RUL prediction with technical and economic analyses to provide a practical decision-making tool for transformer asset management.