Bayesian Network-Driven Demand Prediction and Multi-Trip Two-Echelon Routing for Fleet-Constrained Metropolitan Logistics

Urban logistics in metropolitan areas faces mounting pressure to deliver faster while controlling operational costs under strict fleet size constraints. Traditional vehicle routing models assume unlimited vehicle availability, overlooking realistic fleet utilization and spatial-temporal demand imbalances. This paper introduces the fleet-constrained metropolitan logistics problem (FCMLP), a novel framework integrating trunk linehaul scheduling, two-echelon routing, multi-trip operations, and anticipatory fleet positioning. We model the FCMLP as a Markov Decision Process capturing the stochastic and dynamic nature of metropolitan delivery flows. Our solution framework combines interpretable Bayesian Network-based demand forecasting for transparent proactive vehicle relocation decisions, parameterized cost-function approximation for dynamic order-to-linehaul assignment, and Adaptive Large Neighborhood Search for multi-trip vehicle routing. Computational experiments on synthetic instances and real-world data from a major e-commerce platform in Jakarta demonstrate 20–26% total cost reduction. Multi-trip operations alone reduce fleet size by 23%, while interpretable predictive relocation further improves performance by 7% through a 20% reduction in emergency deployments. The framework’s interpretability enhances operator trust and facilitates practical adoption, offering logistics platforms a path to improve vehicle utilization through operational efficiency and transparent predictive intelligence without expanding fleet size.