Application of Machine Learning in the Predictive Maintenance of Distribution Substations Aimed at Reducing Failures

Cueva M.D.F.; Vila F.S.; Camayo J.M.F. · 2026 16th International Conference on Power, Energy, and Electrical Engineering, CPEEE 2026 (2026) · DOI: 10.1109/cpeee69412.2026.11521194

Resumen

Reliable operation of distribution substations is critical to ensuring continuity of power supply in industrial and residential sectors. This article presents a predictive maintenance framework based on machine learning (ML) aimed at reducing unscheduled failures and optimizing asset management in distribution substations. The methodology integrates Nigel Cross' design model and the Quality Function Deployment (QFD) tool to translate operational needs into technical specifications, supported by SCADA and IoT platforms. Supervised (Random Forest, XGBoost, CNN) and unsupervised (K-Means, autoencoders) models were trained with historical and online monitoring data, considering dissolved gas analysis, temperature, humidity, vibration and thermographic images. The results show that CNN and XGBoost achieved the best performances, exceeding 95 % accuracy and an F 1-score of 0.95, while the Weibull-ML hybrid model estimated the residual useful life of transformers with a mean absolute error of 6.8 %. The implementation of the scheme improved reliability indices (SAIDI by 18.4 % and SAIFI by 20.9 %), reduced corrective costs by 20 %, and extended transformer useful life by 15 %. It is concluded that the combination of robust predictive modeling, digital infrastructure, and explainable artificial intelligence techniques constitutes a replicable and scalable approach for intelligent substation management, in compliance with IEEE/IEC international standards.

Instituciones: Universidad Continental