Synthetic data generation for electric vehicle charging

The rapid adoption of electric vehicles (EVs) creates new challenges for energy systems due to growing residential charging demand. Effective grid management requires detailed charging data, yet publicly available session-level datasets remain scarce. This study develops a conditional diffusion model to generate synthetic residential EV charging sessions conditioned on season, weekday group, and location group. The proposed model is benchmarked against an extended Conditional Tabular Generative Adversarial Network (CTGAN) and TabDDPM, a general-purpose tabular diffusion benchmark. Model performance is evaluated across four residential location groups and reported as mean ± standard deviation. Holm–Bonferroni-adjusted exploratory paired comparisons between the proposed diffusion model and CTGAN were consistent with improved diffusion performance for six of eight yearly distributional metrics, while connection-time differences were better interpreted descriptively. For aggregate load-profile reconstruction, the proposed diffusion model and CTGAN showed complementary strengths under the simplified uniform-energy reconstruction: diffusion better reproduced the mean daily load curve, whereas CTGAN achieved lower peak and variability related errors. A sensitivity analysis using front-loaded fixed-power charging confirmed the diffusion model’s advantage in reconstructing the mean daily load curve, while model rankings for peak and variability-related errors depended on the charging-power reconstruction assumption. TabDDPM showed weaker overall fidelity under the same protocol. The findings indicate that conditional synthetic charging sessions can support indicative scenario construction and downstream energy-system analysis, while detailed charging dynamics and dependency preservation require further validation.