Research on Dynamic Path Planning and Causal Inference for Last-Mile Logistics Based on Spatiotemporal Graph Neural Networks

Abstract

With China’s annual express delivery volume surpassing 120 billion pieces, the last mile of logistics is increasingly constrained by static path planning, delayed dynamic responses, and insufficient integration of heterogeneous data sources. This study targets small and medium-sized distribution stations and proposes a dynamic path planning method based on spatiotemporal graph neural networks combined with causal reasoning. The delivery area is modeled as a spatiotemporal graph that fuses 12 categories of heterogeneous information, including geographic coordinates, real-time traffic conditions, rider load, and weather factors, to capture both structural and temporal dependencies. On this basis, a causal reasoning module is introduced to distinguish sporadic disturbances from routine patterns through backdoor adjustment and intervention models, thereby refining graph attention weights and improving the robustness of routing decisions. Furthermore, a federated learning framework is employed to support encrypted model training and incremental updates across multiple stations, enhancing cross-station generalization while preserving data privacy. Pilot experiments conducted at 10 small and medium-sized stations in Jinhua, Zhejiang, demonstrate that the proposed approach reduces delivery timeout rates from 11.2% to 0.7%, decreases riders’ average daily mileage by 18 kilometers, shortens system response time from 5 minutes to 8 seconds, and improves resource utilization by 35%. These results indicate that the method forms an intelligent perception–reasoning–decision closed loop and provides a scalable technical solution for efficient, privacy-aware last-mile logistics management.