Learning-enhanced rolling-horizon optimisation for port-centric freight systems

This paper presents a centralised rolling-horizon framework for dynamic freight matching in port-centric supply chains. Rather than solving one large static assignment, the framework advances through the planning horizon in short steps, splitting the dynamic many-to-many problem over time into a sequence of small per-step assignments. Within each step, a Graph Neural Network makes the real-time match or wait decisions. The Hungarian algorithm then solves that step’s assignment optimally. Vicinity-based consolidation groups nearby movements, with full and empty containers handled jointly. Many-to-many consolidation emerges as trucks are assigned further demands at later steps, and because each step solves only a small assignment, the framework stays tractable and supports scalable coordination under intermodal and operational constraints. Using throughput data from Sydney’s Port Botany, eight scenarios are simulated across fleet compositions and optimisation priorities. Results show clear trade-offs: small-truck-dominant fleets maximise timeliness and demand coverage but raise cost and emissions, whereas large-truck strategies lower cost and carbon intensity at the expense of service levels. Export flows consistently achieve the highest timeliness.