Predictive Maintenance of Medium Voltage Cable Networks using Machine Learning

The increasing demands of modernizing power infrastructure and managing aging assets present significant challenges for Distribution System Operators (DSOs). There is a critical need for advanced reliability models capable of accounting for diverse operating conditions and accurately predicting individual component health to optimize maintenance strategies. Medium-voltage (MV) cable networks, which consist of both relatively younger Cross-Linked Polyethylene (XLPE) and older Paper-Insulated Lead-Covered (PILC) cables, are critical yet underrepresented components in existing reliability analysis studies. This research emphasizes the transformative potential of machine learning (ML) in their reliability assessment, focusing on data-centric and collaborative model development that helps DSOs improve asset management decisions. Due to the accessibility and relevance of realistic asset and failure data, this thesis focuses on Denmark. After a detailed description of the technical design, aging mechanisms, and maintenance practices specific to Danish MV cable systems, it critically reviews previously applied methods, research, and standards. Building on the review insights, the work proposes a data-centric methodology for facilitating data-driven reliability prediction, detailing required data collection, harmonization, management, and feature engineering efforts. Furthermore, the work proposes a benchmark design to compare various model dimensions and formulates further opportunities for advancements, such as data enrichment, federated learning, and regulatory implications. Lastly, this work provides recommendations for industry and research communities, outlining avenues for future work paving the way for more data-driven, collaborative asset management.