Multi-Head Spatiotemporal Attention Graph Convolutional Network for Traffic Prediction.

Area of Science:

• Computer Science
• Transportation Engineering
• Artificial Intelligence

Background:

• Intelligent transportation systems (ITSs) are crucial for traffic management and infrastructure planning.
• Accurate traffic prediction is challenging due to complex road networks and data dynamics.
• Existing methods struggle with non-Euclidean road structures and topological constraints.

Purpose of the Study:

• To develop a novel traffic forecasting model for enhanced accuracy in ITSs.
• To effectively capture spatio-temporal dependencies and dynamic variations in traffic data.
• To address the limitations of current traffic prediction methodologies.

Main Methods:

• A hybrid model combining Graph Convolutional Network (GCN), Gated Recurrent Unit (GRU), and Multi-Head Attention (MHA).
• Simultaneous incorporation of spatial, temporal, and topological traffic data features.
• Utilizing deep learning architectures for complex pattern recognition in traffic flow.

Main Results:

• Achieved 91.8% accuracy for 15-minute traffic prediction on the Los Angeles highway traffic (Los-loop) dataset.
• Obtained an R2 score of 85% for 15- and 30-minute predictions on the Shenzhen City (SZ-taxi) dataset.
• Demonstrated superior performance in learning global spatial variations and dynamic temporal sequences.

Conclusions:

• The proposed model offers state-of-the-art traffic forecasting capabilities.
• Effectively integrates spatio-temporal and topological information for robust predictions.
• Paves the way for improved traffic management and urban planning through accurate forecasting.