Sparse-Positive Outage Anomaly Ranking in Power Distribution Telemetry Using Positive-Unlabeled Similarity and Graph Propagation

Few outages have confirmed labels in power distribution data, which makes anomaly ranking difficult. Semi-Supervised Local Outlier Factor with Positive-Unlabeled Learning (SSLOF-PU) combines local density with similarity to confirmed outages. The method shares similarity information among close observations and uses the resulting scores to rank anomalies. Calibration sets alarm thresholds without changing the ranking. We evaluated the method on one utility dataset and seven public datasets, using 20 paired runs per dataset. SSLOF-PU had the highest mean average precision (MAP) among the compared approaches on all eight datasets. Its numerical gains over the strongest comparator ranged from 1.4% to 6.1%. None of the paired differences remained statistically significant after Holm–Bonferroni correction. On the utility dataset, MAP was 0.52 ± 0.04, with a 95% confidence interval of 0.501 to 0.539. Examining the top 50 of 10,000 observations achieved an alarm hit rate of 0.700 and a recall of 0.350. Removing graph propagation decreases MAP to 0.39; deleting learned similarity bandwidths reduced it to 0.45. Increasing the calibration sample from 500 to 5000 reduced the standard deviation of the alarm threshold from 0.048 to 0.014. These results present the method’s potential for ranking observations under limited inspection budgets.