STFDSGCN: Spatio-Temporal Fusion Graph Neural Network Based on Dynamic Sparse Graph Convolution GRU for Traffic Flow...

Area of Science:

• Traffic flow forecasting
• Graph neural networks
• Spatio-temporal data analysis

Background:

• Multivariate heterogeneity in traffic flow presents significant forecasting challenges.
• Existing models often struggle with dynamic spatio-temporal patterns and unforeseen events.

Purpose of the Study:

• To develop an advanced model for accurate traffic flow forecasting.
• To address limitations in capturing multivariate heterogeneity and dynamic spatial structures.

Main Methods:

• Proposing a Spatio-Temporal Fusion Graph Neural Network (STFDSGCN).
• Incorporating a dynamic sparse graph convolution gated recurrent unit (DSGCN-GRU) with adaptive sparse graph convolution.
• Utilizing a spatio-temporal attention fusion scheme with a gating mechanism.

Main Results:

• The STFDSGCN model demonstrated superior performance on real-world datasets.
• Achieved improvements in Mean Absolute Error (MAE) by 4.01%, Root Mean Square Error (RMSE) by 1.33%, and Mean Absolute Percentage Error (MAPE) by 1.03% compared to baseline methods.
• Effectively captured heterogeneous, local, and dynamic spatial features.

Conclusions:

• The STFDSGCN model offers a robust approach for traffic flow forecasting, especially in complex and dynamic environments.
• The integration of DSGCN-GRU and spatio-temporal attention enhances the model's ability to handle long-term patterns and traffic emergencies.
• This method provides a unified representation of multi-scale spatio-temporal traffic dynamics.