Deep Embedded Clustering Based Profiling of Electric Vehicle Charging Behavior

The rapid growth of electric vehicles (EVs) has made the randomness and uncertainty of their charging demand a threat to grid stability. Accurately identifying user-group charging preferences and demand patterns, then tailoring charging services and dispatch strategies, is essential for orderly charging and system security. To this end, this paper proposes a deep embedded clustering framework-Dynamic Auto-Encoder with TreeContrastive Learning (DynAE-TreeCL), for profiling EV charging behavior. A dynamic auto-encoder first extracts latent features from raw charging data to obtain an initial grouping. LightGBM then supervises these preliminary clusters to generate pseudolabels, after which a feature-projection network with a contrastive loss sharpens representation discrimination. Tests on real-world EV charging data demonstrate that DynAE-TreeCL decisively outperforms conventional clustering methods and standalone deep clustering models, delivering fine-grained segmentation of distinct user cohorts.