A multi-feature spatial-temporal fusion network for traffic flow prediction.

Area of Science:

• Artificial Intelligence
• Transportation Engineering
• Data Science

Background:

• Traffic congestion is a major urban problem, necessitating accurate traffic flow prediction.
• Existing deep learning models struggle with real-world data discontinuity and irregular distributions.
• There is a need for models that leverage multi-feature fusion over continuous sequence dependencies.

Purpose of the Study:

• To develop a robust traffic flow prediction model that handles data discontinuity and irregular distributions.
• To improve the accuracy and interpretability of traffic flow predictions by utilizing multiple traffic features.
• To address the limitations of current deep learning models in practical traffic management scenarios.

Main Methods:

• Proposed an Adaptive Traffic Features Extraction Mechanism (ATFEM) to select key influence factors and construct joint temporal and global spatial feature matrices.
• Introduced a Multi-feature Spatial-Temporal Fusion Network (MFSTN) incorporating a temporal transformer encoder and graph attention network.
• Developed a scaled spatial-temporal fusion module for automatic optimal weight learning and adaptation to inconsistent dimensions.

Main Results:

• The proposed model demonstrated superior performance compared to various baseline methods in traffic flow prediction.
• Achieved accurate traffic flow predictions even with a high data missing rate.
• The multi-layer perceptron component enhanced the interpretability of the prediction outcomes.

Conclusions:

• The novel ATFEM and MFSTN approach effectively captures complex spatial-temporal dependencies in traffic data.
• The model offers a significant advancement in traffic flow prediction, particularly in challenging real-world conditions with incomplete data.
• This research provides a more interpretable and accurate solution for intelligent transportation systems.