|
Spatiotemporal Traffic Prediction Using Graph Neural Networks, Recurrent Units, and the Spatial Structure of the Network
|
Fateme Noori Alamouti *   |
| K.N Toosi university of technology |
|
|
Abstract: (341 Views) |
Optimal urban traffic management requires accurate and timely prediction of traffic conditions across the spatial–temporal scales, as congestion patterns within the urban road networks are governed by complex spatial interactions among the road segments as well as short- and long-term temporal dynamics. Therefore, the development of frameworks capable of concurrent capturing of the spatial and temporal dependencies plays a critical role in traffic forecasting and in supporting urban decision-making processes. In this study, a hybrid spatiotemporal framework that is based on Graph Neural Networks (GNN) and Gated Recurrent Units (GRU) is proposed for traffic state classification. The dataset consists of the road network of four urban districts of Tehran and the traffic conditions of 5,037 road segments recorded in 15-minute intervals over a three-month period. The road network is modeled as a directed graph and utilized as input to the graph neural network. To enhance the spatial modeling, in addition to the first-order adjacency matrix, the second-order neighborhood relationships and the key structural network measures—including betweenness centrality, closeness centrality, and weighted degree—are explicitly incorporated into the graph learning process.
The GNN extracts the spatial dependencies and structural interactions among road segments, and its output, together with the temporal sequences of the traffic states, is fed into the GRU to perform predictions over 15-, 30-, and 60-minute time horizons. The experimental results demonstrate that the proposed model (Enhanced GNN-GRU) consistently outperforms the baseline models, including GRU, LSTM, and the conventional GNN-GRU, across all of the prediction horizons. Specifically, the Enhanced GNN-GRU achieves an accuracy (ACC) of 0.76 and an F1-score of 0.67 for the 15-minute horizon, indicating a significant improvement over the baseline GNN-GRU model. This performance gain remains stable across longer forecasting horizons. The findings suggest that the explicit integration of the multi-order spatial dependencies and the structural network features substantially enhances the spatial representation learning, improves discrimination among different congestion levels, and increases prediction robustness over the extended time horizons, and it also highlights the critical role of road network topology in the urban traffic modeling and forecasting.
|
|
| Keywords: Traffic prediction, spatiotemporal models, network spatial structure, graph neural networks, recurrent units |
|
|
Full-Text [PDF 1010 kb]
(27 Downloads)
|
Type of Study: Research |
Subject:
GIS
Received: 2026/01/11 | Accepted: 2026/06/16 | ePublished ahead of print: 2026/06/17 | Published: 2026/06/30
|
|
|
|
|
|
|
| Send email to the article author |
|
|
