STAR: Spatial–Temporal Attention Reasoning model for dynamic logistics network routing in Cyber–Physical Internet

Abstract

The Cyber–Physical Internet (CPI) provides a protocol framework that enables logistics transportation to attain levels of reachability, reliability, and efficiency comparable to those of data transmission over the Internet. Within this framework, the routing problem involves determining optimal transportation routes in the logistics network by maintaining routing tables, with the objective of minimizing cost while satisfying constraints such as transportation mode and Estimated Time of Arrival (ETA). However, the inherent spatial and temporal dynamics of real-world logistics networks present substantial challenges to optimal route decision-making. On one hand, unforeseen events such as geopolitical conflicts and natural disasters render certain network nodes unavailable, altering the network topology and introducing spatial dynamics. On the other hand, the transportation time and cost of the same routes fluctuate due to changing supply–demand relationships and varying congestion levels, resulting in temporal dynamics. To tackle these uncertainties, we propose the Spatial–Temporal Attention Reasoning (STAR) model based on Reinforcement Learning (RL), which dynamically updates routing tables by leveraging the current topology and state of logistics networks. STAR uniquely combines a Topology-Aware Graph Convolutional Network (TAGCN), a Temporal-Correlated Recurrent Neural Network (TCRNN), and a Hierarchical Reward (HR) module to comprehensively capture spatial–temporal dynamics of logistics networks, thereby facilitating the adaptive decision-making of the most cost-effective routes that comply with transportation mode and ETA requirements. Numerical experiments based on real Modular-integrated Construction (MiC) cases in the Greater Bay Area (GBA) demonstrate the effectiveness of STAR in optimizing routing decisions within dynamic logistics networks.