Graph Neural Network Model Predicts City Traffic Flow and Congestion With High Accuracy

Traffic engineers have long struggled with a stubborn problem: the models that forecast how vehicles will move through a city are often not accurate enough to be useful for managing it. A new study published in Computational Urban Science proposes a way to close that gap by combining two types of neural networks into a single prediction system, then layering a clustering algorithm on top to judge when roads tip into congestion. The reported results are striking. On a standard benchmark, the model achieved an average mean square error of 1.176 and a precision of 97.99%, figures the author reports as better than every comparison model tested, including several well-known graph-based architectures.

The work was carried out by Dong Hong and appears as an open access original paper published on 26 October 2026 in Volume 6 of the journal. Its starting point is a familiar one for anyone stuck in a rush hour jam. As urbanization accelerates and the number of motor vehicles grows, congestion is becoming increasingly serious, and smart city transportation systems that collect real-time data and analyze it intelligently have emerged as a response. Yet the author argues that existing systems suffer from low accuracy in predicting both traffic flow and congestion, and that many prior efforts have left key questions unanswered.

The technical core of the new approach is a spatio-temporal graph convolutional network, or STGCN-style architecture built from a graph convolutional network paired with a gated recurrent unit. Graph convolutional networks are designed to process graph-structured data, making them a natural fit for road networks, where roads are nodes and their relationships are edges. The gated recurrent unit, an improved recurrent neural network structure, handles the temporal side, extracting features from past traffic flow to capture how conditions evolve over time. Together, the two components let the model represent both where roads sit in the network and how their traffic patterns change through the day.

One of the study’s claimed innovations lies in how the graph itself is built. Conventional adjacency matrices rely on static physical connectivity, assigning weights only to roads that are topologically connected. The proposed spatio-temporal correlation graph instead measures functional correlations from historical traffic patterns using cosine similarity and the Earth Mover’s Distance, a method for quantifying differences between probability distributions. The rationale is that physical adjacency does not necessarily imply functional traffic correlation: two geographically disconnected roads serving similar zones, such as parallel corridors feeding the same business district, may show strongly synchronized flow, while a main arterial and a local access road sitting side by side may remain dynamically decoupled. A sparsity parameter eta set to 0.01 keeps only the strongest connections, and a sensitivity analysis across candidate values from 0.001 to 0.1 found the model achieved its optimal trade-off at 0.01, with results remaining stable between 0.005 and 0.05.

Data preprocessing played a substantial role. The study used a roughly 50 GB dataset of taxi GPS trajectory points collected between 2024 and 2025 through a ride-hailing platform in a Chinese urban area, totaling about 2 billion records sampled once per second, with road network information drawn from OpenStreetMap for the Wuhou District of Chengdu. The pipeline cleaned incomplete and abnormal records, compressed trajectories with the Douglas-Peucker algorithm, matched points to road segments using a distance-based geometric method, and defined road flow from vehicles moving at least 10 meters with more than two trajectory points within 15-minute intervals. Missing values were filled using an improved locality-sensitive hashing algorithm that retrieves historically similar traffic days, an approach the author credits with enhancing the integrity of the flow data.

For training and quantitative evaluation, the study turned to the publicly available Los-loop benchmark, covering traffic speed data at 5-minute intervals from 207 sensors on the Los Angeles highway network over March 1 to 7, 2012. The data were split strictly in temporal order, with the first five days for training, the sixth for validation and the seventh for testing, and no random shuffling, to prevent data leakage. The prediction horizon was deliberately conservative: a single step, meaning the model predicted flow only five minutes ahead. The author is candid that this setting inherently yields higher accuracy because of the strong short-term autocorrelation in highway traffic data, and that multi-step forecasting was reserved for future work.

Against that benchmark, the model’s numbers were consistently the best in its comparison set. Its average mean square error of 1.176 compared with 2.33 for a CNN-LSTM model, 3.67 for a GA-SVM model and 4.12 for a temporal convolutional network, while its average mean absolute error of 0.102 beat the others by a wide margin. Accuracy reached 98.88% and the coefficient of determination reached 0.979, with the author attributing the high fit partly to the single-step horizon and the regularity of aggregated loop-detector data. A further comparison with six representative graph-based models, including STGCN, DCRNN, Graph WaveNet, ASTGCN, AGCRN and MTGNN, under identical experimental conditions found the strongest competitor, MTGNN, reached 94.85% accuracy and 93.96% precision, still below the proposed model’s figures.

The second half of the system addresses congestion. Building on the flow predictions, the author applied an improved fuzzy C-means clustering algorithm that incorporates the entropy weight method, which objectively quantifies each traffic indicator’s information entropy rather than assuming equal weights. Indicators such as average speed, flow density and traffic saturation feed the clustering, which outputs six operating states, from very smooth to extremely severe congestion, mapped onto the Highway Capacity Manual’s level-of-service categories A through F, with an option to aggregate them into binary smooth or congested decisions. Ground-truth labels were generated from fixed HCM thresholds rather than the data-driven clustering process, a design intended to prevent data leakage. On the Los-loop test day, the congestion model achieved an F1 value of 98.56% and a recall rate of 97.88%, against 92.19% and 89.41% for CNN-LSTM, 88.54% and 85.23% for GA-SVM, and 82.69% and 81.33% for the temporal convolutional network. Per-state accuracies ranged from 97.6% to 98.9% across four representative congestion states.

An ablation study reinforced the claim that each component matters. Replacing the dynamic correlation graph with a conventional static adjacency matrix caused noticeable performance drops, swapping the gated recurrent unit for a vanilla recurrent network degraded temporal modeling, removing the graph convolutional module entirely produced the largest decline, and reverting to standard fuzzy C-means with equal indicator weights reduced both F1 and recall. The full model retained its peak values of 98.56% F1 and 97.88% recall. The author frames these results as evidence that the entropy-based weighting mechanism enhances the stability of congestion discrimination and that the dynamic functional correlation graph contributes significantly to capturing spatial dependencies that static topology alone cannot reveal.

Limitations and caveats remain. The congestion experiments covered 20 randomly selected road sections over roughly 250 five-minute time steps within a single test day, and the headline accuracy figures rest on a one-step-ahead prediction task on a highway benchmark rather than a sprawling, stochastic urban arterial network. The author identifies the main limitation of the research as the lack of a visualization tool for integrating the prediction models with actual traffic control systems, noting that converting predictions into decision support for traffic managers remains an unsolved problem. Future work is planned around a real-time visualization platform incorporating digital twin technology to simulate urban traffic dynamics. For now, the study offers a quantitative, reproducibly tested template for pairing graph-based spatial modeling with entropy-weighted clustering, and a demonstration that data-driven road relationships, not just physical maps, can sharpen the picture of how a city moves.

Subject of Research: A spatio-temporal graph convolutional network combined with entropy-weighted fuzzy clustering for urban traffic flow and congestion prediction

Article Title: Smart city traffic flow and congestion prediction model based on spatio-temporal graph neural network

Article References: Hong, D. (2026). Smart city traffic flow and congestion prediction model based on spatio-temporal graph neural network. Computational Urban Science, 6(1), Article 68. https://doi.org/10.1007/s43762-026-00303-3

Image Credits: AI Generated

DOI: 10.1007/s43762-026-00303-3

Keywords: traffic flow prediction, congestion prediction, graph neural network, spatio-temporal graph convolutional network, gated recurrent unit, fuzzy C-means clustering, entropy weight method, smart city transportation, Los-loop benchmark, taxi GPS trajectory data, Highway Capacity Manual, machine learning