Objective-Driven Efficient Sparse Metering for Machine Learning-Based Fault Diagnosis and Self-Healing Restoration in Active Distribution Networks

Abstract Machine learning is a leading research approach to the diagnostic stage of fault location, isolation and service restoration (FLISR), but most schemes assume dense or synchronized metering and few link diagnostic accuracy to reliability. We present an objective-driven framework in which a few low-cost, unsynchronized feeder meters support a learning pipeline for fault classification, section localization, isolation, and self-healing restoration on the IEEE 33-bus system in OpenDSS/MATLAB. The two diagnostic tasks differ markedly in their sensing requirements: classification saturates with three to four meters, whereas localization keeps improving up to sixteen. The difference extends to model choice: the learner most accurate on clean data is not the one that minimizes interruption duration. A greedy wrapper run separately per task, under an event-held-out protocol, beats coverage, filter and random placement at all three budgets tested. Minimizing a lifetime cost that prices meters and unserved energy selects seven meters; there faults are localized to within one hop with 97.3% accuracy and FLISR cuts SAIDI, the system average interruption duration index, by 82.2% under IEEE Std 1366, with a nine-month payback. A data-driven isolability measure explains why localization errors fall on adjacent sections, and a closed-form relation maps accuracy to SAIDI.