Boosting Classification of Electric Vehicles from Charging Patterns for Smart Mobility and Sustainability

While electric vehicles (EVs) offer an eco-friendly transportation option, uncoordinated residential charging may lead to peak demand, asset stress, and inefficient capacity utilization, which adversely affects smart grid operations. Addressing these challenges requires timely and accurate identification of EV charging behavior at fine temporal resolution. Accordingly, this paper addresses the practical task of classifying EVs directly from statistics over 1-minute charging profiles using a two-stage pipeline suitable for real-time deployment. In the first stage, segmentation is performed by applying K-means clustering with hysteresis to the average real power signal, allowing the classifier to be trained on charging-only intervals. In the second stage, additional features are extracted to summarize short-horizon dynamics including level, slope, and stability over two rolling windows. We then train three gradient-boosting tree classifiers and analyze their performance using different metrics. The best performing model is LGBM with an accuracy of 0.9951, macro-F1 of 0.9936, MCC of 0.9945, top-3 accuracy and AUC (macro/micro) of approximately 1. The inference latency is approximately 0.044 ms/sample and the training time is approximately 3.13s. Treating segmentation as preprocessing sharpens class signatures, and causal short-horizon statistics disambiguate overlapping plateaus to produce strong, interpretable decision boundaries—enabling type-aware power budgeting, safer concurrent charging, and standardized interoperability tests using only commodity smart meter signals.