Unsupervised Multi-Source Behavioral Fusion for Identifying High-Value Electric Vehicle Users in Demand Response

Accurately identifying electric vehicle (EV) users with high demand response (DR) potential is critical for grid stability but remains challenging due to behavioral heterogeneity, data sparsity, and the subjectivity of expert-dependent methods. In particular, the absence of behavior labels and the low temporal frequency of EV charging events limit the effectiveness of conventional rule-based and clustering approaches. To address these issues, we propose a novel unsupervised framework that integrates deep behavioral modeling with multi-source indicator fusion. Our approach begins by developing a behavior recognition model robust to sparse data, effectively characterizing user charging patterns. Subsequently, a multi-dimensional potential feature system is established. A key innovation lies in our unsupervised weighting mechanism, which automatically learns the importance of each indicator by assessing inter-indicator correlations, thereby eliminating subjective bias. Finally, the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) is employed to rank users comprehensively based on their fused potential scores. Case studies on a large-scale real-world EV charging dataset demonstrate that the proposed method can effectively distinguish high-potential users from low-potential ones. The results indicate clear separability across multiple behavioral and willingness-related dimensions. This provides a practical and data-driven basis for targeted DR incentive design and user-side flexible resource planning.