PT-TDGCN: Pre-Trained Trend-Aware Dynamic Graph Convolutional Network for Traffic Flow Prediction.

Area of Science:

• Intelligent Transportation Systems
• Machine Learning
• Graph Neural Networks

Background:

• Accurate traffic flow prediction is essential for efficient intelligent transportation systems.
• Existing models struggle with complex spatiotemporal dynamics and time-varying spatial relationships.
• Short input windows limit the ability to capture long-term traffic patterns.

Purpose of the Study:

• To develop an advanced deep learning framework for enhanced traffic flow prediction.
• To address limitations of existing methods in modeling dynamic spatial dependencies and multi-scale temporal patterns.
• To improve the accuracy and robustness of traffic flow forecasting.

Main Methods:

• Proposed a two-stage framework: Pre-trained Trend-aware Dynamic Graph Convolutional Network (PT-TDGCN).
• Utilized a Transformer-based masked autoencoder for learning segment-level temporal representations during pre-training.
• Integrated dynamic graph learning via tensor decomposition, convolutional trend-aware attention, and spatial graph convolution with fusion projection for prediction.

Main Results:

• PT-TDGCN demonstrated superior predictive accuracy and robustness across four real-world traffic datasets.
• Consistently outperformed 14 established baseline models in traffic flow prediction tasks.
• The proposed methods effectively captured time-varying spatial relations and long-term traffic dynamics.

Conclusions:

• The PT-TDGCN framework offers a significant advancement in traffic flow prediction.
• The integration of pre-training, dynamic graph learning, and trend-aware attention enhances model performance.
• This approach provides a more robust and accurate solution for intelligent transportation systems.