A Two-Stage Framework for Spatio-Temporal Clustering and Forecasting of Electric Vehicle Charging

The increasing adoption of electric vehicles (EVs) presents both challenges and opportunities for power systems, particularly in understanding and forecasting charging behavior. This paper presents a two-stage framework to segment and forecast electric-vehicle (EV) charging behavior using spatio-temporal features extracted from one year of 15-minute telematics for 264 EVs in Saskatchewan, Canada. The feature set combines temporal usage (power/energy levels and variability) with spatial mobility (home-charging ratio, location diversity, travel distance and angular dispersion). We compare three complementary clustering families including Fuzzy C-Means (soft membership), Gaussian Mixtures (probabilistic partitions), and Hierarchical clustering (multiscale structure) and using algorithm-appropriate selection criteria (Xie–Beni, BIC, elbow/Calinski–Harabasz) to ensure a fair assessment. The framework identifies five behaviorally meaningful groups ranging from home-dominant commuters to high-mobility, DC fast-charging users. Hierarchical clustering provides the most balanced and interpretable segmentation (lowest Davies–Bouldin; highest Calinski–Harabasz), while GMM attains the highest Silhouette (0.585) with coarser partitions. In the supervised stage, ensemble classifiers forecast user-segment membership from features, with XGBoost achieving 98.1% accuracy. The outputs translate directly into planning inputs: expected residential feeder peaks for home-centric segments, DC fast-charging demand for high-mobility users, and demand-response potential for managed charging. The results indicate that spatio-temporal features coupled with hierarchical segmentation and gradient boosting provide actionable and scalable inputs for siting, tariff design, and grid operations.