Data-Driven, Short-Term Prediction of Charging Station Occupation

Abstract

Enhancing electric vehicle infrastructure by forecasting the availability of charging stations can boost the attractiveness of electric vehicles. The transportation sector plays a crucial role in battling climate change. The majority of available prediction algorithms either achieve poor accuracy or predict the availability at certain points in time in the future. Both of these situations are not ideal and may potentially hinder the model’s applicability to real-world situations. This paper provides a new model for estimating the charging duration of charging events in real time, which may be used to estimate the waiting time of users at fully occupied charging stations. First, the prediction is made using the random forest regressor (RF), and then the prediction is enhanced utilizing the findings of the RF model and real-time information of the currently occurring charging events. We compare the proposed method with the RF model, which is the approach’s foundational model, and the best-performing prediction model of the light gradient boosting machine (LightGBM). Here, we make use of historical information of charging events gathered from 2079 charging stations across Germany’s 4602 fast-charging connectors. To reduce data bias, we specifically simulate prediction requests for 30% of the charging events with various characteristics that were not trained with the model. Overall, the suggested method performs better than both the RF and the LightGBM. In addition, the model’s structure is adaptable and can incorporate real-time information on charging events.

Keywords

• electric vehicles
• charging infrastructure
• random forest (RF)
• ensemble learning