A Digital-Twin-Oriented Framework for Candidate-Locker Demand and Road-Risk Prediction

Last-mile planning benefits from two complementary predictive signals before routing: expected package demand at candidate consolidation nodes and area-level road risk. This study develops a retrospective, digital-twin-oriented framework for a Los Angeles case study. It aligns delivery, weather, map, and collision records in a provenance-aware analytical state, preserves candidate-node demand and grid-day collision occurrence as separately defined and validated prediction tasks, and exposes both outputs through a common planning interface. Chronological evaluation across statistical, machine learning, neural, rule-based, and spatial-frequency methods identifies complementary strengths: the LSTM provides the strongest package-unit demand accuracy, while frequency-based road-risk models lead discrimination, classification, and probability performance. A held-out counterfactual planning experiment further shows that using both predictive signals improves service reliability and lowers average route-risk intensity while making the associated fleet and distance trade-offs explicit. The principal contribution is an auditable dual predictive architecture that connects heterogeneous urban data, task-specific validation, and operational planning without collapsing distinct demand and safety targets into a single model.