Machine Learning–Based Failure-Risk Profiling and Preventive Maintenance Prioritisation for 11 kV Distribution Feeders Using Operational Data

Abstract

Conventional reliability assessment provides valuable information on historical feeder performance but offers limited insight into the operating conditions under which future failures are most likely to occur. This paper presents a machine learning–based framework for failure-risk profiling and preventive maintenance prioritisation of the Mokwa and Jebba–Bacita 11 kV distribution feeders using historical operational data from their respective injection substations. A Random Forest classifier was developed using strictly causal pre-fault features comprising instantaneous feeder load, rolling load statistics, recent outage history, and temporal operating characteristics to estimate hour-by-hour failure probability without target leakage. The resulting probabilistic outputs were subsequently translated into practical maintenance decision indicators, namely the Failure Risk Reduction Index (FRRI), Preventive Maintenance Priority Score (PMPS), and High-Risk Exposure Index (HREI), to support evidence-based maintenance planning. The predictive model achieved ROC–AUC values of approximately 0.65–0.70 for the Mokwa feeder and 0.58–0.63 for the Jebba–Bacita feeder, indicating moderate predictive capability given the constraints of operational utility data. The analysis further reveals that the Mokwa feeder exhibits stronger load-dependent failure behaviour, with elevated failure probabilities observed within the 12.5–14 MW operating range and across much of the daily operating cycle, whereas Jebba–Bacita demonstrates comparatively weaker load sensitivity and lower overall operational risk. The derived maintenance indices indicate that a relatively small proportion of operating conditions contributes disproportionately to feeder failures, enabling maintenance resources to be prioritised toward high-risk operating states. The proposed framework demonstrates how machine learning can complement conventional reliability assessment by transforming historical operational records into practical maintenance intelligence that supports proactive asset management in resource-constrained electricity distribution networks.