ADSTGCN: A Dynamic Adaptive Deeper Spatio-Temporal Graph Convolutional Network for Multi-Step Traffic Forecasting.

Area of Science:

• Artificial Intelligence
• Transportation Engineering
• Data Science

Background:

• Multi-step traffic forecasting is complex due to dynamic traffic conditions.
• Graph Convolutional Networks (GCNs) extract spatial traffic data but often suffer from shallow architectures and over-smoothing.
• Existing models lack flexibility due to fixed node structures.

Purpose of the Study:

• To develop a novel traffic forecasting model that addresses the limitations of existing GCNs.
• To mitigate the over-smoothing phenomenon in deeper GCNs for traffic prediction.
• To enhance the adaptability and flexibility of traffic forecasting models for real-world networks.

Main Methods:

• Proposed Dynamic Adaptive Deeper Spatio-Temporal Graph Convolutional Networks (ADSTGCN).
• Implemented dynamic hidden layer connections and adaptive weight adjustments to combat over-smoothing.
• Introduced a parameter-sharing adaptive matrix for learning spatial dependencies and network structure adaptation.

Main Results:

• ADSTGCN effectively addresses the over-smoothing issue in deeper GCNs.
• The model demonstrates improved performance in multi-step traffic forecasting.
• Evaluations on highway and urban road networks show promising results.

Conclusions:

• ADSTGCN offers a more robust and flexible approach to traffic forecasting.
• The adaptive mechanisms enhance the model's ability to capture dynamic traffic patterns.
• The proposed method shows significant potential for real-world traffic prediction applications.