Survival analysis of electric vehicle charging behavior and the temporal evolution of feature effects

Abstract

Abstract This study proposes a survival-based modeling framework that combines behavioral features with interpretable machine learning to understand and predict user churn in electric vehicle charging services. Using a dataset of 1,074 users and 107,531 charging sessions from Central European countries, we modeled time-to-churn while handling censored observations. The best-performing model, a Stacked Weibull survival model based on gradient boosting, achieved a concordance index of 0.826 ± 0.041 and Integrated Brier Score of 0.078 ± 0.008 (5-fold cross-validation), with strong calibration relative to Kaplan-Meier survival estimates. Interpretability analyses identified sustained session frequency, positive engagement trends, and temporal regularity in charging behavior as key predictors of reduced churn risk. These findings highlight the potential of survival modeling integrated with behavioral analytics to predict churn risk and inform retention strategies in electric vehicle charging networks.