Forecasting Electric Vehicle Charging Behavior in Workplace Charging Infrastructure with Limited Privacy-Restricted Real Data

Workplace Charging Stations (CSs) are well-suited to improve grid stability by scheduling the charging process over the parking duration and thereby reducing the peak load. Therefore, the energy demand and parking duration of single charging sessions must be known as well as the future occupancy of the CS. Since user IDs are often unknown for privacy reasons, this paper investigates how these parameters can be predicted for future charging events. The charging behavior is examined for its characteristic features, such as location, arrival, and departure times. First, calendar, weather, lag and CS-specific features are implemented and used to train nine different machine-learning algorithms. For the observed data, the Random Forest Regressor yields the best results for parking duration and energy demand. For parking duration, a 33.7% improvement in Mean Absolute Percentage Error (MAPE) over the baseline (the mean parking duration) can be achieved. The MAPE of the parking duration forecast is 71.0% and for the energy demand, it is 84.0% which leads to the conclusion that without the knowledge of user IDs predicting the charging behavior of users is possible only to a limited extent.