A Dual-Stream Cross AGFormer-GPT Network for Traffic Flow Prediction Based on Large-Scale Road Sensor Data.

Area of Science:

• Artificial Intelligence
• Transportation Engineering
• Data Science

Background:

• Accurate traffic flow prediction is crucial for traffic management and route optimization.
• Large-scale historical traffic data presents challenges due to high non-linearity.
• Existing models struggle to effectively capture complex spatial and temporal dependencies.

Purpose of the Study:

• To propose a novel network architecture for improved traffic flow prediction.
• To integrate diverse traffic data streams (occupancy, speed) for enhanced accuracy.
• To leverage the strengths of adaptive graph neural networks and large language models.

Main Methods:

• Developed a dual-stream cross AGFormer-GPT network incorporating prompt engineering.
• Utilized traffic occupancy and speed as prompts, integrated via cross-attention.
• Employed a dual-stream cross structure to mine spatial and temporal correlations.

Main Results:

• The proposed model demonstrated improved traffic prediction accuracy.
• Achieved approximately 1.2% enhancement in prediction accuracy across different road networks.
• Experimental validation conducted on two PeMS road network datasets.

Conclusions:

• The dual-stream cross AGFormer-GPT network effectively combines adaptive graph neural networks and large language models.
• The model successfully captures complex spatial and temporal traffic dynamics.
• This approach offers a significant advancement in traffic flow prediction accuracy and reliability.