Application of Machine Learning Methods and Hybrid Modeling for Predicting the Remaining Useful Life of Equipment

The article examines the problem of predicting power transformers remaining useful life based on the analysis of dissolved gas concentrations in transformer oil. The study is based on a time-series analysis of hydrogen, carbon monoxide, ethylene, and acetylene concentration measurements over an extended operational period. The study involved data preprocessing and time series aggregation to identify patterns and factors influencing the remaining useful life of the equipment. Various data aggregation strategies were applied, including classical averaging and automated extraction of informative features. Different machine learning methods were employed to predict the remaining useful life of transformers, including ensemble methods such as gradient boosting, decision trees, and multilayer perceptrons. The implementation additionally included a feature classification stage aimed at identifying the most significant factors, improving regression forecasting accuracy. A comparative analysis of different approaches was conducted using established prediction quality metrics: mean absolute error, coefficient of determination, root mean square error, mean absolute percentage error, and classification evaluation metric. The research yielded a hybrid approach that combines classification and regression methods, facilitating a more accurate estimation of the equipment's residual service life. The proposed method can be integrated into transformer condition monitoring systems.