Operator Learning with Branch–Trunk Factorization for Macroscopic Short-Term Speed Forecasting

Logistics has become an integral part of economic activity, with new formats such as front warehouses and hourly delivery demanding real-time visibility and rapid response. Minute-level road speed prediction is essential for platoon control, routing, and signal optimization, yet remains challenging due to heterogeneous and noisy data sources, highly coupled spatio-temporal interactions, and frequent distribution shifts. This paper proposes a Deep Operator Network–based framework that links logistics demand with traffic states. Warehouse and customer data are projected onto a five-kilometer subnetwork, generating six scenarios and about 1.2 million link–time samples. The proposed model decouples historical speeds in the branch from exogenous states in the trunk, allowing boundary changes to be incorporated as functional inputs rather than requiring retraining. Experiments demonstrate that the proposed method outperforms both classical regression and deep learning baselines, while ablation analyses verify robustness and interpretability. These findings establish operator learning as a promising direction for adaptive logistics forecasting.