Multi-scale modeling of electric vehicle fatal crash risk: uncovering spatial heterogeneity and infrastructure-land...

Understanding the spatial heterogeneity of electric vehicle (EV) crash risks is critical for safeguarding electrified mobility systems, yet existing methods cannot simultaneously capture spatial dynamics while maintaining model interpretability. By accounting for spatially varying macro-level spatial non-stationarity and complex interaction, a Multi-Scale Adaptive Bandwidth Geographically Weighted Random Forest (MSAB-GWRF) method is developed to uncover the spatial heterogeneity and infrastructure-land use coupling mechanisms of EV fatal crash risk at the regional level. Other than traffic and weather factors, this study investigates the effects of charging infrastructure density, the Social Vulnerability Index (SVI), and land-use characteristics. Meanwhile, a cost-based EV crash risk indicator is formulated by incorporating crash-associated economic costs. Based on monthly state-level observations across the U.S. from 2020 to 2022 (N = 1836), the modelling results indicate that the MSAB-GWRF method outperforms the fixed-bandwidth GWRF, improving average R² by 18.2%, particularly in the high-risk regions. More importantly, SHAP-based explanations further reveal pronounced spatial non-stationarity in the drivers of EV fatal crash risk. Land-use and EV-related indicators dominate in most states, outweighing conventional traffic exposure, while social vulnerability dominates in the South and Appalachia. SHAP analysis further identifies a capacity-dependent threshold effect. In high-risk regions, charging infrastructure and land-use intensity act as risk amplifiers once regional exposure exceeds critical levels, whereas low-risk regions absorb comparable expansion without safety deterioration. These findings indicate that uniform safety policies are insufficient, calling for regionally differentiated strategies coordinating charging deployment, land-use planning, and equity-oriented interventions.