DSTFGCN: A dynamic spatial-temporal fusion graph convolution network for traffic flow forecasting.

Area of Science:

• Intelligent Transportation Systems
• Machine Learning for Transportation
• Graph Neural Networks

Background:

• Traffic flow prediction is crucial for intelligent transportation systems.
• Existing methods struggle with complex spatial-temporal dependencies.
• Recurrent neural networks and adaptive adjacency matrices have limitations.

Purpose of the Study:

• To propose a novel network for accurate traffic flow prediction.
• To overcome limitations in modeling spatial-temporal dependencies.
• To enhance the effectiveness of intelligent transportation systems.

Main Methods:

• Developed a dynamic spatial-temporal fusion graph convolution network (DSTFGCN).
• Incorporated gated dilated causal convolution for local temporal dependencies.
• Utilized node-independent temporal graph convolution for global temporal dependencies.
• Introduced a dynamic graph convolution block for spatial dependencies.

Main Results:

• DSTFGCN effectively captures both local and global spatial-temporal dependencies.
• The proposed dynamic graph convolution block constructs adaptive graphs.
• Experiments on six real-world datasets demonstrate superior performance over mainstream methods.
• Achieved higher accuracy in traffic flow prediction.

Conclusions:

• DSTFGCN offers a significant advancement in traffic flow prediction.
• The model's ability to fuse dynamic spatial and temporal information is key.
• This research contributes to more effective intelligent transportation systems.