Multi-scale spatio-temporal graph neural network for urban traffic flow prediction

Area of Science:

• Urban planning and transportation science
• Artificial intelligence and machine learning
• Complex systems analysis

Background:

• Urban traffic flow exhibits complex, non-linear spatiotemporal patterns due to internal and external factors.
• Accurate traffic flow prediction is challenging but crucial for efficient urban mobility.
• Existing models struggle to capture the intricate dynamics of traffic flow propagation.

Purpose of the Study:

• To propose a novel deep learning model for enhanced urban traffic flow prediction.
• To address the challenges posed by complex non-linear spatiotemporal patterns in traffic data.
• To develop a model capable of integrating multi-timescale traffic flow features.

Main Methods:

• Developed a Spatio-Temporal Graph neural network with Multi-timeScale (STGMS).
• Implemented a multi-timescale feature decomposition strategy to separate traffic flow signals and residuals.
• Designed a unified spatio-temporal feature encoding module for integrated feature representation.
• Trained the model to map multi-timescale spatiotemporal features to future traffic flow.

Main Results:

• STGMS demonstrated superior performance compared to eleven baseline models on four real-world datasets.
• Achieved average improvement rates of 17.69% in Mean Absolute Error (MAE).
• Achieved average improvement rates of 15.65% in Root Mean Square Error (RMSE) and 10.30% in Mean Absolute Percentage Error (MAPE).

Conclusions:

• The proposed STGMS model effectively captures complex urban traffic flow dynamics.
• STGMS offers significant improvements in traffic flow prediction accuracy.
• The multi-timescale approach is key to enhancing the performance of traffic prediction models.