Multivariate Anomaly Detection of Three-Phase Electrical Power Systems Using Isolation Forest

The increasing availability of electrical monitoring data provides opportunities for data-driven identification of atypical operating conditions in three-phase electrical systems. This study investigates multivariate anomaly detection using Isolation Forest on an open-source dataset comprising three-phase voltage, current, active power, reactive power, power factor, and frequency measurements. Following data integration and quality assessment, 33,211 complete observations (99.88%) across 18 electrical variables were analyzed. An Isolation Forest model with 300 estimators and contamination set to ‘auto’ identified 7,118 observations (21.43%) as candidate anomalies under its automatic decision threshold, rather than confirmed electrical faults. Continuous anomaly scores and percentile-based lower-tail rankings were subsequently used for characterization. The lowest anomaly scores were primarily associated with phase-B and phase-C current, active power, and reactive power variables, accompanied by reduced phase-C power factor. Frequency exhibited negligible monotonic association with anomaly scores (Spearman’s ρ = 0.0034, p = 0.535), and its exclusion minimally affected anomaly rankings (ρ = 0.9826). Feature-group ablation demonstrated high overall ranking stability (ρ = 0.940–0.983), indicating that observation rankings remained relatively consistent despite the exclusion of individual electrical feature groups. Temporal analysis further revealed a pronounced concentration of low-score observations during 18:00–21:00. These findings demonstrate that combining anomaly ranking, electrical characterization, feature-group sensitivity analysis, and temporal analysis provides an interpretable unsupervised framework for screening and prioritizing atypical multivariate operating conditions in three-phase electrical systems.