Spatio-temporal forecasting of EV charging demand in urban communities considering differentiated user characteristics

• Chebyshev polynomials enhance spatio-temporal forecasting precision. • Iterative calibration framework adapts traffic models to historical patterns. • Multivariate data screening improves model interpretability and validation. As the penetration rate of electric vehicles (EVs) increases, large-scale, uncontrolled charging of EVs causes significant fluctuations in load, posing a significant challenge to the accuracy of charging demand forecasting. To further investigate the user characteristics associated with EV chargers, this paper proposes a spatio-temporal forecasting method for urban EV charging demand, considering the differentiated user characteristics. First, a temporal graph convolutional network is employed to extract and learn from historical traffic datasets, providing short-term spatio-temporal traffic forecasts. Second, the maximum entropy estimation method obtains the corresponding origin–destination (O-D) matrix for traffic forecasts. Then, a traffic model is constructed based on the transportation network (TN), while a dynamic traffic assignment (DTA) model incorporating iterative calibration was developed to capture 24-hour traffic flow patterns. Furthermore, a precise traffic flow measurement strategy is proposed to comprehensively capture user-specific charging behaviors, systematically addressing the intrinsic heterogeneity of charging characteristics. These steps collectively form a weekday travel simulation model. Finally, based on user characteristics, two types of charging methods are modeled using the M/M/C/K queue model and Monte Carlo simulation to forecast spatio-temporal charging demand. Simulation experiments based on traffic data from a city in East China further validate the unique advantages of the proposed method. Comparative analysis reveals significant performance improvements: Total Demand Prediction Deviation (TDPD) is reduced by 11.02%, Mean Absolute Percentage Error (MAPE) decreases by 39.26%, and R-squared (R 2 ) increases by 13.27%, conclusively validating the method’s computational precision and predictive reliability.