Unifying spatiotemporal and frequential attention for traffic prediction

Area of Science:

• Intelligent Transportation Systems
• Data Science
• Signal Processing

Background:

• Urban traffic flow forecasting is crucial for intelligent transportation systems.
• Existing methods primarily focus on spatial and temporal dependencies, neglecting spectral characteristics in traffic data.
• There is a need for advanced models that incorporate frequency domain analysis for improved traffic prediction.

Purpose of the Study:

• To develop an innovative traffic prediction model, the Space-Time-Frequency Attention Network (STFAN).
• To integrate attention mechanisms for capturing correlations across space, time, and frequency domains.
• To enhance the accuracy of urban traffic flow forecasting by leveraging spectral characteristics.

Main Methods:

• Utilized deep learning for spatial correlation analysis in traffic flow.
• Applied spectral analysis to fuse time series data with periodic correlations in time and frequency domains.
• Developed the STFAN model employing attention mechanisms to project current traffic features across dimensions to future states.

Main Results:

• The STFAN model demonstrated superior predictive accuracy compared to baseline models on PeMS04 and PeMS08 datasets.
• The model showed particular effectiveness in mid- and long-term traffic flow forecasting.
• Ablation studies confirmed the significant influence of frequency domain characteristics on future traffic conditions.

Conclusions:

• The proposed STFAN model offers a significant advancement in urban traffic flow forecasting.
• Integrating spectral analysis and attention mechanisms provides a comprehensive approach to learning traffic dynamics.
• The findings highlight the practical effectiveness of considering frequency domain features for accurate traffic prediction.