An Integrated Machine Learning Framework for EV Charging Behavior Characterization and Anomaly Detection in Public Charging Infrastructure

The rapid expansion of electric vehicle (EV) adoption has increased the demand for efficient charging infrastructure and data-driven approaches for understanding charging behavior. Analyzing charging patterns and identifying abnormal charging sessions are essential for improving charging network reliability, infrastructure utilization, and operational efficiency. This study proposes a comprehensive machine learning framework for EV charging behavior analysis and anomaly detection using real-world charging session data collected from six charging bays. Four charging behavior indicators, namely energy consumption (Usage), charging duration (Duration), average charging output power (Average Output), and Energy Consumption Ratio (ECR), were extracted through a feature engineering process. K-Means clustering was employed to identify distinct user behavior groups, while Principal Component Analysis (PCA) was utilized to visualize cluster separability. Isolation Forest was subsequently applied to detect anomalous charging sessions and investigate abnormal charging behavior patterns. Statistical validation was conducted using Analysis of Variance (ANOVA), and Pearson correlation analysis was performed to examine relationships among charging features and anomaly occurrence. The results identified four distinct charging behavior clusters representing moderate users, regular users, inefficient users, and high-power users. Clustering validation achieved a silhouette score of 0.6086, while PCA retained 89.8% of the total variance using two principal components. An anomaly detection analysis revealed that inefficient charging behavior exhibited the highest anomaly occurrence, whereas regular users demonstrated highly consistent charging patterns. Analysis indicated that average charging output power and ECR were the most influential variables contributing to anomaly identification. ANOVA results confirmed statistically significant differences among all identified clusters (p < 0.001), while correlation analysis demonstrated a strong positive relationship between charging power and charging efficiency (r = 0.95). The anomaly detection framework achieved accuracy, precision, recall, and F1-score of 80.0%. The proposed framework provides a comprehensive approach for EV charging behavior characterization, anomaly detection, and charging infrastructure assessment. The findings can support charging network operators in improving charging efficiency, identifying abnormal charging activities, and enabling data-driven management of EV charging systems.