Source-linked AI summary
Neighbor-embedded Graph Neural Network-based Crowd Delivery Traffic Management in Smart City
Kishu Gupta, Deepika Saxena, Ashutosh Kumar Singh, Chung-Nan Lee
TL;DR
Existing smart-city traffic management approaches often address congestion prediction or time optimization without integrating crowd-delivery optimization and intelligent vehicle selection. NeCDM combines neighbor-embedded GNN traffic prediction with real-time vehicle selection, reporting accurate traffic-flow prediction and reduced sustainability-critical parameters.
Problem
Existing approaches inadequately integrate congestion prediction with crowd-delivery optimization, vehicle reduction, and intelligent vehicle selection for sustainable smart-city transportation.
Method
NeCDM combines a neighbor-embedded Graph Neural Network for delivery-station traffic prediction with a decision-making unit for selecting suitable vehicles for crowd delivery requests.
Results
NeCDM accurately predicts traffic flow across delivery stations and supports vehicle selection intended to reduce active delivery vehicles and sustainability-critical parameters.
Takeaways & Limitations
Crowd delivery integrated with traffic prediction and vehicle selection is presented as a framework for improving congestion mitigation and delivery efficiency in smart-city environments.
Takeaways & Limitations
The paper identifies complex, volatile, continuously growing traffic data and interdependencies among vehicles, routes, and delivery operations as constraints on existing traffic-management effectiveness.
Abstract
from arXiv · showhide
The significant upsurge in vehicle traffic presents a considerable challenge in the pursuit of smart mobilization and transportation (SMT) worldwide. Current approaches primarily focus on vehicular traffic management through congestion prediction but fall short in addressing essential objectives such as traffic reduction and appropriate vehicle selection to alleviate congestion in smart cities ($SmCt$). To address these concerns, this work introduces a novel \textit{Neighbor-Embedded Graph Neural Network-based Crowd Delivery Traffic Management} (NeCDM) Model, comprising two key components: the Traffic Congestion Prediction Unit (TCPu) and the Traffic Observation and Management Unit (TOMu). The TCPu utilizes Graph Neural Network (GNN) optimization to accurately predict traffic flow levels at various delivery stations within $SmCt$ ecosystems. Additionally, the TOMu facilitates the intelligent selection of the most suitable delivery vehicles for fulfilling crowd delivery requests ($CDR$). This work emphasizes the potential of crowd delivery as a feasible solution for achieving SMT goals while adhering to smart city parameters ($\mathcal{SCP}$s), such as reduced carbon emissions, shorter travel times, and minimized travel distances. The proposed model achieves notable improvements in computational efficiency, including reductions of up to 4.03\% in L1 loss ($£$), 16.66\% in L2 loss ($£_{rmse}$), and 7.64\% in computation time.
I. INTRODUCTION
The paper motivates crowd delivery as a way to reduce active delivery vehicles and improve sustainability-critical parameters, then proposes NeCDM to combine congestion prediction with adaptive vehicle management.
- Rising on-road vehicle volumes strain smart-city traffic management and motivate Smart Mobility and Transportation frameworks.
- Crowd delivery consolidates proximate deliveries into one route, reducing active vehicles and potentially improving carbon emissions, travel distance, duration, and cost.
- Existing frameworks often separate congestion prediction from time-based optimization and insufficiently integrate vehicle reduction, crowd delivery, and real-time vehicle selection.
- NeCDM combines a neighbor-embedded GNN-based TCPu for traffic-flow prediction with a TOMu for adaptive vehicle choice, rerouting, and traffic countermeasures.
- The integrated units continuously monitor congestion and select appropriate vehicles for crowd delivery requests while curbing the number of on-road vehicles.
II. RELATED WORK
Prior work advances traffic forecasting through deep, recurrent, convolutional, attention-based, and graph-based methods, but leaves gaps in closed-loop management and integrated resource optimization.
- Traffic forecasting research spans hybrid spatio-temporal deep learning, graph and multi-graph neural architectures, and resource-aware deployment frameworks.
- Convolutional, recurrent, attention, and dynamic graph methods improve representation of periodic patterns, temporal components, and spatial correlations.
- Related routing studies address cooperative multi-vehicle routing and multi-depot vehicle routing under congestion and emission constraints.
- Three gaps remain: forecasting is often disconnected from closed-loop management, spatial graphs may be static or partly dynamic, and resource optimization is rarely integrated with control.
- The NeCDM system model includes users, delivery stations, delivery vehicles, a smart city, and TOMu, with vehicle choice affected by fuel type, mileage, distance, trip time, and load capacity.
- Its assumptions constrain vehicle types to Electric, CNG, Petrol, and Diesel and crowd delivery requests to one through ten drop stations.
C. Problem Statement and Design Goals
The problem is to solve a multi-objective, constrained smart-transportation task that predicts congestion and selects delivery vehicles for crowd delivery requests across multiple stations.
- The problem combines numerous delivery stations, crowd delivery requests, appropriate vehicle selection, and sustainability-critical parameters within a constrained smart-city transportation setting.
- The congestion-prediction goal is to estimate long-term traffic before finalizing a route and vehicle while incorporating real-time traffic conditions.
- NeCDM uses TCPu for congestion forecasting across delivery stations and TOMu for real-time vehicle selection and traffic-management decisions.
- The vehicle-selection goal is to process requests from different users and identify an appropriate delivery vehicle for each crowd delivery request.
- A smart city is operationalized as a living space using intelligent transportation, data analytics, and predictive modeling for automated decisions that reduce congestion and support transport reliability.
- A crowd delivery request specifies a source station and one or more destination stations fulfilled by a single delivery vehicle.
A. Neighbor Dynamic Graph Network Architecture
The traffic graph models delivery stations as nodes and routes as edges, with node and edge features updated dynamically to represent evolving traffic. Neighbor information is aggregated with self-features through layered message passing for congestion prediction.
- Delivery stations form graph vertices, while routes form undirected, unweighted, homogeneous edges connecting stations.
- Node features encode local traffic attributes, while edge features describe travel time, flow rate, and distance between stations.
- The adjacency matrix and edge weights are dynamically updated with time to encode current traffic states.
- A continuous weight function uses observed travel time and a scaling constant to represent traffic conditions for graph learning.
- Layered message passing iteratively aggregates and updates information from neighboring nodes using aggregation, update, and ReLU functions.The dynamic graph adapts adjacency and edge weights in each time window.
- The TCPu adapts its graph representation to real-time congestion correlations and travel times for prediction under varying temporal conditions.
B. Traffic Congestion Prediction
The TCPu predicts congestion for crowd delivery requests by constructing traffic graphs, normalizing traffic data, and training a neighbor-embedded GNN. Its estimated flow is then passed to vehicle-management decisions.
- The TCPu predicts congestion for multiple crowd delivery requests submitted by users across source and delivery stations.Each request includes a source station and one or more delivery stations.
- Traffic graphs represent requests across stations, and geodesic distance between source and delivery stations is computed from latitude and longitude.The distance calculation uses the approximate Earth radius R = 6373 km.
- Missing values are replaced with −1, after which min–max normalization rescales high-variance traffic samples to [0, 1].
- The training dataset is divided into training, validation, and testing segments, while neighboring-node features and each node’s self-features support traffic prediction.
- The neighbor-embedded graph representation is used to train the prediction function and generate the estimated traffic congestion flow.
- The NeCDM prediction function uses Adam optimization with ReLU activation and minimizes L1 loss, with L2 error used to evaluate loss more clearly.The estimated flow ϱ is passed to TOMu for delivery-vehicle selection.
C. Delivery Vehicle Selection
The TOMu selects a delivery vehicle by combining predicted traffic flow with vehicle, route, distance, and delivery-demand constraints. A weighted feasibility score aggregates these selection parameters.
- The TOMu selects a delivery vehicle using predicted flow, flow limit, distance limit, drop count, and course of travel.
- Flow limit evaluates whether a vehicle’s speed and operating conditions satisfy admissibility criteria based on available fuel or energy.
- Distance limit checks whether the required travel distance can be covered by the selected vehicle without additional service.
- A large number of delivery drop points favors a heavy-load vehicle with greater power and payload capacity.
- Course of travel distinguishes outside-city routes, associated with lower energy needs, from congested local roads within the smart city.
- The model aggregates influential selection parameters into a feasibility score, with weights that can prioritize vehicle type, distance, or traffic conditions.
D. Operational Design and Complexity
NeCDM operationally combines ride-request and traffic-detail stores, graph construction, congestion prediction, performance evaluation, and vehicle selection. Its complexity includes request, graph, station, and neighbor-embedded GNN processing costs.
- Operational Design: The operational workflow predicts congestion, evaluates performance, retrains until the desired performance is reached, and then deploys the model for estimation.
- Operational Design: After congestion prediction, NeCDM applies vehicle-selection equations to identify an appropriate vehicle for each crowd delivery request.
- Operational Design: NeCDM uses Ride Request Details and Temporal and Spatial Traffic Details as its two input data stores.
- Operational Design: Users submit crowd delivery requests containing source and multiple delivery stations, with vehicle type determined by each request’s requirements.
- Operational Design: Traffic graph networks are created from source and destination details, and the model processes each crowd delivery request and its travel stations.
- Complexity: The stated complexity is O(N) for request processing, O(V + E) for traffic-graph generation, and O(n) for station-level processing and objective-cost computation.
A. Experimental Setup
The experiments use a specified computing environment and evaluate NeCDM on an extended Beijing trajectory dataset split into training, validation, and testing subsets.
- The experiments run on a 40-core Intel Xeon Silver 4114 server with 128 GB RAM, Ubuntu 16.04, Python 3.9, and PyTorch.
- The evaluation uses 70,124 extended Beijing trajectory samples collected for smart-city traffic analysis.
- The dataset is randomly divided into 80% training, 10% validation, and 10% testing subsets.
- Spatial inputs include GPS coordinates and station identifiers for 113 delivery stations.
- Temporal inputs contain hourly traffic observations from 02-02-2021 23:00 through 23-03-2023 21:00.
C. Results
NeCDM is evaluated through optimization, traffic-flow prediction, route outcomes, efficiency analysis, and comparisons with GNN baselines. Results describe rapid loss convergence, accurate prediction across delivery stations, and improved performance with more training data and neighborhood modeling.
- Model Optimization: NeCDM loss decreases sharply through 20 steps, then gradually converges to a minimal value around 200 steps.Figure 3 reports training and validation L1 loss over iterations.
- Model Optimization: The loss range for traffic-flow prediction over epochs [0, 100] is [280.9755, 29.5213].Table IV presents traffic prediction loss and execution-time optimization for a crowd delivery request.
- Traffic Flow Prediction: Traffic-flow plots compare actual and predicted values across delivery stations, and the model accurately predicts flow despite an interval of unavailable data.
- Crowd Delivery Route: The optimized process produces delivery routes across stations using neighbor-embedded GNN traffic-flow prediction.
- Efficiency Analysis: Both L1 (£) and L2 (£rmse) losses decrease as the training-set size increases, consistent with incremental learning on larger datasets.The efficiency analysis evaluates varying traffic-data allocations and reports improved prediction accuracy with larger training sets.
- Loss Comparison: NeCDM outperforms GNN and GNN-NeiEx by achieving greater loss reduction in fewer steps and more accurate congestion predictions across delivery stations.GNN-NeiEx excludes edge information, while NeCDM explicitly models nearest-neighbor interactions.
2) Traffic Flow Comparison:
NeCDM provides more accurate traffic-flow predictions across delivery stations by embedding neighbor information in graph learning, with reported gains in loss and computation time over comparable approaches.
- NeCDM predicts delivery-station traffic flow more accurately than the GNN baseline, especially when traffic data is unavailable.The authors attribute this to evaluating neighbor importance while learning traffic patterns.
- Up to 4.03% improvement is reported in mean absolute loss (£) over comparable approaches.The comparison includes PCNN, LSTM, GNN, and GNN-NeiEx.
- Up to 16.66% improvement is reported in root mean square error (£rmse) over comparable approaches.The metric is part of the computational comparison presented in Table VI.
- Up to 7.64% improvement is reported in overall computation time over comparable approaches.The reported comparison covers PCNN, LSTM, GNN, and GNN-NeiEx.
- The model’s neighbor-aware design is presented as the main reason for more realistic traffic-flow prediction.Neighbor features are incorporated into the learning process to account for their impact on traffic-flow prediction.