A Time- and Space-Consistent Generative Model for User-Level Electric Vehicle Charging Sessions

User-level electric vehicle (EV) charging records are valuable for vehicle-charger–grid studies, but raw longitudinal records are difficult to share because they may reveal mobility routines, charging preferences, station-use habits, and commercially sensitive network operation. Public EV charging datasets more often provide session-, station-, transaction-, or aggregate-level information. These data support load forecasting and station-utilization analysis, but they are less suited to studies that require traceable user-level charging-session sequences. This paper proposes a temporal–spatial load-consistent model for generating synthetic user-level EV charging sessions. The model first constructs uncalibrated simulated charging-session sets from vehicle attributes, user preferences, state of charge (SOC), charging power, and location choice. These charging sessions are aggregated into a temporal–spatial load matrix and calibrated by an entropy-regularized bi-marginal projection so that the generated charging sessions reduce deviations from the targets with minimum distortion of the pre-calibration structure. The calibrated matrix is then mapped back to individual charging sessions through a charging-session-level feedback operator, followed by physically constrained session allocation. A Shanghai electric-taxi case study compares temporal and spatial consistency with an MC baseline and a statistical generator and further evaluates behavioral fidelity, sensitivity to target uncertainty, and computational efficiency. Relative to the MC baseline, temporal RMSE decreased from 2423.79 kWh to 1082.08 kWh, and spatial WMAPE decreased from 0.07028 to 0.02351, while the largest energy-weighted Jensen–Shannon divergence, observed for final SOC, was 0.00652. The resulting synthetic charging-session set provides traceable inputs for charging-load simulation, infrastructure planning, distribution-network assessment, and vehicle–grid integration studies under data-sharing constraints.