TEA-GCN: Transformer-Enhanced Adaptive Graph Convolutional Network for Traffic Flow Forecasting.

Area of Science:

• Transportation Science
• Artificial Intelligence
• Data Science

Background:

• Accurate traffic flow forecasting is essential for efficient urban traffic management and resource optimization.
• Existing spatial-temporal graph models struggle with dynamic spatial correlations due to fixed network structures.
• Capturing complex spatial-temporal dependencies in traffic data is key for precise predictions.

Purpose of the Study:

• To develop an advanced model for traffic flow forecasting that addresses the limitations of traditional methods.
• To enhance the capture of dynamic spatial correlations and complex temporal dependencies in traffic data.
• To improve the accuracy and reliability of urban traffic condition predictions.

Main Methods:

• Proposed a Transformer-Enhanced Adaptive Graph Convolutional Network (TEA-GCN).
• Implemented an adaptive graph convolutional module for dynamic road dependency learning.
• Incorporated a local-global temporal attention module for capturing diverse temporal dependencies.

Main Results:

• The TEA-GCN model demonstrated superior performance in traffic flow prediction.
• Experimental results validated the model's effectiveness on two public traffic datasets.
• The proposed method outperformed several state-of-the-art traffic flow prediction techniques.

Conclusions:

• The TEA-GCN effectively captures dynamic spatial-temporal dependencies in traffic data.
• The adaptive graph convolutional and temporal attention modules contribute to improved forecasting accuracy.
• This approach offers a significant advancement in urban traffic flow prediction.