Spatial-temporal combination and multi-head flow-attention network for traffic flow prediction.

Area of Science:

• Artificial Intelligence
• Transportation Engineering
• Data Science

Background:

• Traffic flow prediction is crucial for effective traffic management.
• Existing methods struggle with complex spatial-temporal correlations and attention mechanism inefficiencies.
• Nonlinear spatial-temporal data presents significant modeling challenges.

Purpose of the Study:

• To propose a novel network for modeling spatial-temporal correlations in road networks.
• To address limitations in capturing complex correlations and quadratic complexity in attention mechanisms.
• To enhance the accuracy and efficiency of traffic flow prediction.

Main Methods:

• Developed a spatial-temporal combination and multi-head flow-attention network (STCMFA).
• Introduced a temporal sequence multi-head flow attention (TS-MFA) with source competition and sink allocation mechanisms.
• Integrated GRU for enhanced temporal modeling and GCN for spatial-temporal correlation capture, alongside residual mechanisms and feature aggregation.

Main Results:

• The proposed STCMFA model demonstrated excellent performance on four real-world traffic datasets.
• Significantly outperformed existing baseline methods in traffic flow prediction tasks.
• Effectively captured complex spatial-temporal correlations, overcoming limitations of previous approaches.

Conclusions:

• The STCMFA network offers a superior approach to traffic flow prediction.
• The novel attention mechanisms and integrated deep learning components enhance predictive accuracy.
• This research contributes to advancing intelligent transportation systems through improved predictive modeling.