Temporal Locality in the Prediction of Hourly EV Station Availability

This paper presents a feature-engineered, data-driven approach for predicting the hourly availability of electric vehicle (EV) charging stations though machine learning (ML) models. We leverage historical usage data from over 400 stations across multiple European countries to train several models to predict hourly availability from one day in advance, up to seven days in advance. We demonstrate that combining short-term behavioural patterns with long-term features greatly enhances the model’s predictive strength. Specifically, we encode station usage patterns in terms of both their global availability behavior and temporally localized availability trends. Our experiments compare several classification models, identifying XGBoost as the most effective, achieving up to 84% accuracy when predicting one day in advance. By extending the local features to different day offsets, we illustrate that the model remains robust up to seven days ahead, though accuracy gradually declines with larger forecast windows. Notably, these results are attained without relying on costly or hard-to-acquire external data, such as weather or traffic information. This straightforward yet powerful approach can facilitate optimized charger utilization, pricing strategies, and user satisfaction. Moreover, it provides a scalable, interpretable solution that can also support demand response programs, ultimately contributing to more resilient and efficient EV charging infrastructures.