Machine Learning for Plug-in-Time Classification of Charging Faults and Abnormal Termination Events in Public Electric-Vehicle Charging Sessions

Public charging operators typically identify abnormal charging outcomes only after a session terminates, limiting proactive maintenance and reducing service reliability. This study develops a plug-in-time machine-learning pipeline that uses only information available when a charging session begins. The analysis uses 441,077 public charging sessions from 92 charging posts in Jiaxing, China, with 44 calendar, weather, tariff, location, user-history, and charging-post-history features. Eight classifiers are evaluated primarily with rolling-origin chronological validation. HistGradientBoosting achieved a mean PR-AUC of 0.598 and ROC-AUC of 0.819 on the four 2021 test quarters. A stricter fold-local repeated-cross-validation audit, in which validation outcomes never update user or post histories, ranked XGBoost first with a mean PR-AUC of 0.583. Charging-post recent abnormal history was the dominant predictor, but a single-feature post-history baseline achieved a pooled PR-AUC of 0.511 compared with 0.604 for the full HistGradientBoosting pipeline. Flagging the highest-risk 1% of sessions achieved 98.4% precision and 5.3% recall; quarter-specific precision ranged from 95.7% to 99.5%. Lagged prior-quarter isotonic recalibration reduced expected calibration error to 0.016–0.021 in quarters F2–F4. These results support capacity-constrained monitoring and maintenance triage rather than exhaustive fault detection.