Source-linked AI summary
Resilient Routing: Risk-Aware Dynamic Routing in Smart Logistics via Spatiotemporal Graph Learning
Zhiming Xue, Sichen Zhao, Yalun Qi, Xianling Zeng, Zihan Yu
TL;DR
Static routing struggles with congestion and fluctuating demand, while existing work often separates traffic prediction from route optimization. RADR combines GPS-based graph construction, GCN-GRU risk prediction, and dynamic routing penalties. In reported experiments, the framework trades a small distance increase for lower congestion-risk exposure and integrates prediction directly into routing decisions.
Problem
Static routing cannot adequately accommodate congestion and fluctuating demand, while much existing work separates traffic prediction from route optimization.
Method
RADR constructs a logistics graph from GPS trajectories, predicts congestion risk with a hybrid GCN-GRU ST-GNN, and integrates predictions into dynamic routing decisions.
Results
2.1% increase in physical driving distance accompanied a 17.6% reduction in congestion risk exposure in the reported case study.
Takeaways & Limitations
The framework demonstrates a data-driven way to couple risk prediction with routing while balancing delivery distance and supply-chain resilience.
Abstract
from arXiv · showhide
With the rapid development of the e-commerce industry, the logistics network is experiencing unprecedented pressure. The traditional static routing strategy most time cannot tolerate the traffic congestion and fluctuating retail demand. In this paper, we propose a Risk-Aware Dynamic Routing(RADR) framework which integrates Spatiotemporal Graph Neural Networks (ST-GNN) with combinatorial optimization. We first construct a logistics topology graph by using the discrete GPS data using spatial clustering methods. Subsequently, a hybrid deep learning model combining Graph Convolutional Network (GCN) and Gated Recurrent Unit (GRU) is adopted to extract spatial correlations and temporal dependencies for predicting future congestion risks. These prediction results are then integrated into a dynamic edge weight mechanism to perform path planning. We evaluated the framework on the Smart Logistics Dataset 2024, which contains real-world Internet of Things(IoT) sensor data. The experimental results show that the RADR algorithm significantly enhances the resilience of the supply chain. Particularly in the case study of high congestion scenarios, our method reduces the potential congestion risk exposure by 19.3% while only increasing the transportation distance by 2.1%. This empirical evidence confirms that the proposed data-driven approach can effectively balance delivery efficiency and operational safety.
I. INTRODUCTION
The paper addresses the limits of static routing and isolated traffic prediction in volatile last-mile logistics. It proposes an integrated data-driven framework combining spatiotemporal prediction with dynamic routing decisions.
- Motivation: Static Dijkstra and A* routing can direct vehicles into congestion because they assume constant network topology and edge weights.This can produce delivery delays and increased operational costs.
- Research context: LSTM-based traffic models capture temporal patterns but neglect spatial dependencies across the logistics network.GNNs model topological structure, while existing approaches often lack end-to-end integration with downstream decisions.
- Contribution: The proposed framework closes the loop by using real-time spatiotemporal forecasts to drive dynamic routing decisions.It moves beyond treating prediction and optimization as isolated tasks.
- Contribution: K-Means discretizes GPS trajectories, a GCN-GRU model captures spatial and temporal traffic patterns, and risk penalties guide path planning.The framework is designed to balance distance minimization with supply-chain resilience.
- Research context: Spatiotemporal graph models extend earlier statistical and recurrent traffic methods by jointly representing road-network topology and temporal evolution.Earlier approaches such as ARIMA and LSTM are described as limited for complex nonlinear or spatially structured traffic patterns.
B. Logistics Routing Optimization
Vehicle routing must handle combinatorial complexity while adapting to changing demand and traffic. The paper frames an end-to-end alternative to approaches that separate traffic prediction from route optimization.
- Routing problem: The vehicle routing problem determines delivery routes under constraints such as travel cost, time windows, and vehicle capacity.Its combinatorial and NP-hard nature has motivated extensive optimization research.
- Existing methods: Heuristic and metaheuristic routing methods perform well in static or moderately sized instances but adapt less effectively to dynamic conditions.The passage identifies parameter tuning, changing demand, real-time traffic, and large network changes as challenges.
- Learning-based routing: Learning-based GNN routing methods represent logistics systems as graphs to adapt routing strategies across changing operational conditions.Prior work integrates GNNs with reinforcement learning for varying graph topologies.
- Research gap: Most existing literature treats traffic prediction and route optimization as separate tasks, with predictions fed to solvers afterward.The paper identifies this separation as the central research gap.
- Proposed direction: RADR directly integrates ST-GNN risk prediction into dynamic decisions to support proactive rather than reactive routing.This couples forecasting with route optimization within one framework.
III. METHODOLOGY
RADR converts noisy logistics trajectories into a directed graph and couples its structure to risk-aware routing. The methodology uses clustering, trajectory-derived connectivity, pruning, and graph optimization.
- Method overview: The framework aims to transform noisy IoT logistics data into reliable routing decisions by coupling risk prediction with graph-based optimization.This establishes the overall methodological objective.
- Graph construction: K-Means clusters all GPS points into N = 10 functional zones, with each cluster centroid represented as a graph node.The discretization makes continuous trajectories compatible with graph neural networks.
- Graph construction: The choice of N balances spatial granularity and computational efficiency.Smaller values can over-smooth traffic patterns, whereas larger values increase graph sparsity and model instability.
- Graph topology construction: Connectivity is derived from temporally consecutive GPS records within each truck trajectory rather than geometric proximity alone.Transition counts Cij between zones are row-normalized into a directed adjacency matrix.
- Graph topology construction: Edges with Aij < τ are pruned using τ = 0.01 to suppress spurious connections from noise or rare movements.The resulting directed graph captures asymmetric mobility patterns, and normalized transition weights serve as GCN edge weights.
B. Spatiotemporal Risk Prediction Model
The risk-prediction model combines graph convolution with recurrent processing to capture spatial congestion diffusion and temporal evolution. Its output supports one-step-ahead congestion-risk prediction for routing.
- Model motivation: Traffic congestion can spread across neighboring hubs, so the model must represent spatial leakage alongside temporal dynamics.This motivates a hybrid spatiotemporal graph architecture rather than an isolated time-series model.
- Architecture: The ST-GNN architecture is designed to capture complex spatiotemporal dependencies in the road network.Its workflow consists of spatial graph convolution followed by temporal recurrent modeling.
- Spatial modeling via graph convolution: At each time step, a GCN aggregates node-level traffic features from neighboring nodes to model spatial dependencies.The graph operation incorporates localized traffic conditions and captures spatial congestion diffusion.
- Temporal modeling via GRU: The GCN representations are passed to a GRU that models temporal dependencies over a historical observation window of T = 10.The GRU produces a one-step-ahead congestion-risk prediction and captures congestion buildup and dissipation.
C. Risk-Aware Dynamic Routing
RADR incorporates predicted congestion risk into dynamic edge costs while retaining a standard shortest-path solver. The resulting optimization explicitly trades travel distance against congestion exposure to support more reliable routing.
- Risk-aware cost formulation: RADR embeds predicted congestion risk into a dynamic cost function rather than modifying the routing algorithm itself.The framework uses a standard Dijkstra shortest-path solver with risk-aware edge weighting.
- Risk-aware cost formulation: Wdyn(u, v) = dist(u, v) × (1 + λ · Riskavg), increasing effective costs for edges with higher predicted congestion risk.Riskavg is the predicted average edge risk, while λ controls risk sensitivity.
- Path optimization: The selected route minimizes cumulative dynamic cost across feasible paths from source node u to destination node v.The formulation allows routing to trade marginal travel-distance increases for reductions in congestion risk.
- Route-level risk evaluation: The path-level Risk Score sums predicted average congestion risks across route edges to quantify total exposure.Lower Risk Scores indicate safer routes with reduced likelihood of congestion-induced delays.
- Framework outcome: RADR tightly couples spatiotemporal risk prediction with classical graph optimization for proactive and resilient routing decisions.
A. Dataset & Settings
The study uses a year-long logistics dataset containing timestamped operational and traffic records. GPS trajectories are clustered into a directed logistics graph, normalized into temporal snapshots, and split chronologically for evaluation.
- Dataset: The Smart Logistics Supply Chain Dataset (2024) contains real-world logistics and supply-chain operational records collected throughout 2024.
- Dataset: The data includes GPS trajectories, shipment identifiers, delivery events, traffic indicators, and delay labels reflecting dynamic operational variability.
- Dataset: The dataset spans January–December 2024 and contains tens of thousands of records for spatiotemporal routing analysis.
- Graph construction: K-Means clusters continuous GPS trajectories into N = 10 logistics nodes, with directed edges based on Truck ID transition frequencies.Node features are aggregated over fixed temporal intervals into graph snapshots.
- Graph construction: Dense, non-uniform GPS coverage motivates clustering to preserve dominant traffic flows while reducing noise and model complexity.
- Temporal settings: The data is divided chronologically, using 80% of earlier time steps for training and 20% for testing without temporal information leakage.A rolling historical window of length T = 10 is used for model input.
B. Performance Analysis
The ST-GNN shows stable training and tracks observed congestion variability, while ablation results indicate that both spatial and temporal modeling contribute to lower prediction error.
- Model performance: The training loss drops sharply during the first 15 epochs and then plateaus, indicating stable early convergence.The reported dynamics suggest the model captured spatiotemporal patterns without oscillation.
- Model performance: The congestion predictions closely follow ground-truth values, including sudden sharp spikes in traffic conditions.The comparison is presented on the right side of Fig. 3.
- Ablation study: The ablation study tests whether the graph component is necessary compared with a time-series-only model.Two variants remove either spatial or temporal modeling from the full framework.
- Ablation study: MSE increased to 0.048 for the GRU-only variant, while the GCN-only model reached MSE 0.050.These variants respectively omit spatial modeling and temporal trend modeling.
- Ablation study: The full RADR framework achieves the lowest error, with MSE 0.038.The result supports simultaneously modeling spatial topology and temporal evolution for risk prediction.
D. Case Study
The case study evaluates routing during a peak congestion window, where RADR avoids a high-risk link through a strategic detour. This trades a 2.1% distance increase for a 17.6% reduction in congestion risk exposure.
- At Time Step 98, the experiment routes a shipment from Node 0 to Node 5 during a peak congestion window.
- The static baseline selects the shortest physical path 0 →5 despite the model flagging that link as high risk.
- RADR instead calculates a strategic detour through Node 9 and Node 6 in response to the predicted risk score.
- 2.1% increase in physical driving distance accompanies a 17.6% reduction in congestion risk exposure.The case study presents this as the trade-off produced by the detour.
V. CONCLUSION
The conclusion presents RADR as a pipeline combining clustering, spatiotemporal modeling, and dynamic optimization to support resilient logistics decisions. It also identifies broader unstructured-data integration and fleet-level coordination as future directions.
- RADR combines K-Means clustering, ST-GNN modeling, and dynamic optimization in one pipeline for logistics decisions.
- The framework is described as predicting traffic while selecting routes that navigate around identified congestion risks.
- Future work targets unstructured logistics data, including driver logs, maintenance reports, and weather forecasts, through LLM-based feature extraction.
- The authors also plan to extend single-path optimization toward multi-agent reinforcement learning for fleet-level coordination.