Traffic Flow Prediction Based on Hypergraph Spatiotemporal Interaction Network.

Area of Science:

• Intelligent Transportation Systems
• Data Science
• Network Science

Background:

• Existing traffic flow prediction models struggle with complex spatiotemporal dependencies in road networks.
• Accurate short-term traffic prediction is crucial for efficient intelligent transportation systems.

Purpose of the Study:

• To propose a novel traffic flow prediction model, the Hypergraph Spatio-Temporal Interaction Network (HGSTIN), to enhance accuracy and stability.
• To effectively model complex spatiotemporal dependencies in traffic flow data.

Main Methods:

• Constructed a multi-dimensional traffic pattern input tensor integrating proximity, intra-day, and intra-week temporal scales.
• Employed a Transformer architecture with a Dynamic Tanh (DyT) mechanism for temporal modeling.
• Combined neighborhood and DTW-based semantic hypergraphs with spatial self-attention and hypergraph neural networks for spatial modeling.
• Integrated an adaptive feature fusion module and a temporal gradient consistency loss function.

Main Results:

• The HGSTIN model achieved average improvements of 5.15% in MAE, 1.76% in RMSE, and 3.88% in MAPE over the second-best baseline on PEMS04 and PEMS08 datasets.
• Demonstrated superior performance in multi-step prediction scenarios with minimal degradation.
• Ablation studies validated the effectiveness of individual model components.

Conclusions:

• HGSTIN effectively captures dynamic spatiotemporal characteristics of complex traffic scenarios.
• The proposed model provides high-precision prediction support for intelligent transportation systems.
• HGSTIN offers a robust and accurate solution for short-term traffic flow forecasting.