Predicting dwell time of logistics electric vehicles in urban last-mile delivery: A SHAP-based ensemble approach

Logistics Electric Vehicles (LEVs) are increasingly essential for sustainable urban freight and last-mile delivery, driven by the global push toward low-emission transportation and smart mobility ecosystems. However, optimizing LEV operations remains challenging due to the complex interplay of energy constraints, charging behavior, and urban logistics dynamics. This study aims to predict the dwell time (i.e., stop duration) of LEVs using an interpretable machine learning (ML) technique to support efficient fleet scheduling and energy planning. This study utilizes a real-world dataset of 1,065 LEV stops collected over one month in Shanghai, comprising operational, temporal, and energy-related variables. A stacked ensemble model integrating XGBoost, LightGBM, and CatBoost is developed to achieve high predictive accuracy, while SHAP analysis is employed to interpret the influence of key features. The proposed model achieves an R2 of 0.993, significantly outperforming individual learners, and reveals complex non-linear relationships among operational, temporal, and energy-related variables. SHAP analysis reveals that end state-of-charge (end_soc) and start_soc emerge as dominant drivers of dwell time, followed by trip speed, distance, time_of_day, and charging status indicators. These findings highlight the critical role of energy conditions and time windows in shaping dwell time. The study provides actionable insights for logistics firms, such as improved route optimization, charging station placement, and shift planning. It also offers policy guidance for urban planners and regulators in designing smart grid-compatible infrastructure, incentive schemes, and public–private data collaborations to enhance LEV ecosystem performance.