Neighbor-embedded Graph Neural Network-based Crowd Delivery Traffic Management in Smart City

Cities are breathing harder than ever. The sheer volume of vehicles on the road has turned daily commutes and delivery routes into a tangled web of stop-and-go frustration, choking the air and wasting time. In the vision of a "smart city," technology is supposed to untangle this mess, not just by predicting where the traffic will be, but by actively managing how goods move through the grid. The core challenge lies in two difficult tasks: knowing exactly how busy a street will be before a truck even leaves the depot, and then choosing the right vehicle for the job. A delivery van might be perfect for a short trip through a quiet neighborhood but a terrible choice for a long haul across a congested highway, especially if it runs on electricity with a limited range. Getting these two decisions right—forecasting the flow and picking the machine—is the key to reducing the number of vehicles on the road, which in turn lowers carbon emissions and shortens travel times for everyone.

A team of researchers has tackled this problem with a new system designed to act as a central nervous system for urban deliveries. They call their creation the Neighbor-Embedded Graph Neural Network-based Crowd Delivery Traffic Management model. While the name is technical, the concept is straightforward: it is a two-part engine that first predicts traffic jams and then uses that prediction to select the best vehicle for a specific delivery route. The system is built to handle "crowd delivery," a strategy where multiple nearby packages are consolidated into a single trip rather than sending out a separate vehicle for each one. This approach aims to shrink the total number of active vehicles on the street, directly addressing the goal of making cities more sustainable.

The first part of this system, the Traffic Congestion Prediction Unit, acts as a crystal ball for road conditions. Instead of looking at traffic data in isolation, it treats the city's delivery stations as points on a map that are all connected to one another. The system learns how traffic at one station influences the traffic at its neighbors, much like how a ripple in a pond spreads outward. By analyzing historical data on vehicle counts, speeds, and travel times, it builds a dynamic picture of how congestion flows through the network. This allows it to forecast traffic levels at specific delivery stops with high precision, even when data is missing or conditions are volatile. The researchers found that by paying close attention to these neighbor relationships, the system could predict traffic flow significantly better than older methods that treated each location as a separate island.

Once the system knows what the traffic will look like, the second part, the Traffic Observation and Management Unit, takes over to make the actual decisions. This unit receives the traffic forecast and matches it against the specific needs of a delivery request. It considers a variety of factors: how far the vehicle needs to travel, how many stops it must make, the type of road it will use, and the vehicle's own capabilities, such as its fuel type and cargo capacity. The system then calculates a score to determine which vehicle is the best fit. For example, it might decide that a heavy-duty diesel truck is necessary for a long route with many stops, while a smaller electric vehicle is the right choice for a short, local run. This intelligent matching ensures that the vehicle selected is not just available, but actually suited to the job, preventing situations where a small van gets stuck in a long-haul trip or a large truck wastes energy on a tiny errand.

To test if this approach actually works, the researchers ran extensive evaluations using an extended real-world dataset of Beijing trajectory data, a city with complex and heavy traffic patterns. They fed the system thousands of delivery requests and watched how it performed compared to existing methods. The results were clear: the new model made fewer mistakes in its predictions and required less time to compute its answers. Specifically, the system reduced the error in its traffic predictions by a noticeable margin compared to standard models, and it cut down the time needed to process the data. In the evaluations, the model successfully identified the most efficient routes and vehicle combinations, proving that it could handle the chaotic, changing nature of city traffic without getting overwhelmed.

The study also highlighted what happens when you ignore the connections between different parts of the city. When the researchers tested versions of the system that did not look at how neighboring stations influenced each other, the predictions became less accurate. This confirmed that traffic is not just a collection of isolated events but a connected flow where a jam in one area inevitably affects the next. By embedding these relationships directly into the learning process, the new model captures the reality of how traffic moves. The researchers concluded that this integrated approach—combining accurate forecasting with smart vehicle selection—offers a viable path toward reducing the number of delivery vehicles on the road. While the work utilized historical trajectory data for evaluation, the results suggest that such a system could significantly improve the efficiency of urban logistics, helping cities move toward a future with less congestion and a cleaner environment.