Short-Term Utilization Forecasting of Electric Vehicle Charging Infrastructures

To operate electric vehicle (EV) fleets in a safe and efficient manner, many companies have been deploying charging infrastructures (CIs) at their premises. Forecasting of different system parameters of a CI, such as how many charging points will be occupied during the day, can help create accurate charge plans. In this paper, we deal with the applicability of continuous Nowcasting, i.e., frequently executed short-term forecasts, to predict the next few data points based on the past and current situation in a CI. Specifically, we forecast the number of charging EVs over a rolling two-hour horizon using XGBoost and LSTM. In the experiments, we apply different weighting schemes to emphasize the relevance of the most recent observations combined with different multi-horizon forecasting strategies. Experimental results using a real-world dataset show that a linear weighting schema combined with a direct forecasting strategy using XGBoost achieves the lowest RMSE value of 0.906 for the 15 min forecasting horizon when predicting the number of active charging stations. For the 2 h horizon, the best RMSE of 2.545 is achieved with XGBoost using the strategy Direct, but with an exponential weighting strategy. We then illustrate how short-term predictions can be used to improve the operational efficiency of an example CI by dynamically adjusting power limits based on the latest prediction results.