Improving Zone-to-Zone Freight OD Estimation in Florida via GNN and LLM Integration

Luohan Zhou

College: College of Business and Public Management

Major: BS.ACCOUNTING

Faculty Research Mentor: Liu, Dan

Abstract:

Accurate estimation of freight origin–destination (OD) flows is essential for transportation planning, infrastructure optimization, and regional freight policy analysis. Graph Neural Networks (GNNs) have recently emerged as powerful tools for modeling network-structured transportation systems due to their ability to capture spatial dependencies and topological interactions. However, conventional GNN-based approaches primarily rely on numerical and structural features, limiting their capacity to incorporate heterogeneous semantic attributes such as freight type classifications, trade characteristics, and domain-specific knowledge. Moreover, their purely data-driven learning paradigm may constrain generalization and interpretability in complex, multi-source freight environments.To address these limitations, this study proposes a hybrid framework integrating Graph Neural Networks (GNN) with Large Language Models (LLM) for enhanced zone-to-zone freight OD estimation in Florida’s road transportation network. The GNN component models spatial flow propagation by learning graph-based representations derived from tonnage, shipment volume, and roadway attributes. The LLM component complements this structure-aware learning by enriching semantic feature representations, structuring heterogeneous freight information, and supporting knowledge-guided integration across multi-source datasets.By combining spatial graph learning with language-informed semantic augmentation, the proposed framework improves robustness, interpretability, and estimation accuracy compared to standalone GNN models. The resulting system generates refined zone-to-zone freight OD estimates incorporating tonnage, volume, and freight type dimensions under a unified multi-source modeling architecture. This research demonstrates the potential of AI-driven graph–language integration for scalable and semantically enriched freight flow estimation.