Multiscale Spatiotemporal Graph Convolutional Networks With Dynamic Delay Awareness for Traffic Forecasting.

Area of Science:

• Artificial Intelligence
• Computer Science
• Transportation Engineering

Background:

• Traffic forecasting presents significant spatiotemporal challenges.
• Graph convolutional networks (GCNs) have improved forecasting by modeling network topology.
• Existing GCN methods suffer from information redundancy and inconsistent timescale correlations.

Purpose of the Study:

• To develop a novel dynamic delay-aware multiscale spatiotemporal graph convolutional network (DDAMGCN) for enhanced traffic forecasting.
• To address limitations in GCN-based traffic forecasting, including information redundancy and inconsistent temporal correlations.

Main Methods:

• Designed a dynamic delay-aware module to identify key nodes and model important delays, focusing on critical information and reducing computational load.
• Developed a multiscale spatiotemporal graph convolution module for fine-grained modeling of spatiotemporal correlations across different timescales.
• Evaluated the DDAMGCN model on eight real-world traffic datasets.

Main Results:

• The proposed DDAMGCN method demonstrated superior performance compared to state-of-the-art baseline methods.
• The dynamic delay-aware module effectively reduced information redundancy and computational costs.
• The multiscale module accurately captured spatiotemporal correlations at various timescales.

Conclusions:

• The DDAMGCN model offers a significant advancement in traffic forecasting accuracy and efficiency.
• The method's ability to handle dynamic delays and multiscale correlations makes it robust for complex traffic networks.