Transformer-based short-term traffic forecasting model considering traffic spatiotemporal correlation.

Area of Science:

• Intelligent Transportation Systems
• Machine Learning
• Traffic Engineering

Background:

• Accurate traffic forecasting is vital for optimizing transportation networks and managing traffic flow.
• Existing models often fail to capture complex spatiotemporal dependencies in traffic data due to nonlinearity and high dimensionality.

Purpose of the Study:

• To propose a novel short-term traffic forecasting model, Trafficformer, leveraging the Transformer framework.
• To enhance the accuracy of traffic speed prediction by effectively modeling spatiotemporal patterns.

Main Methods:

• Feature extraction from historical traffic data using a multilayer perceptron.
• Spatial interaction enhancement via Transformer-based encoding and road network topology integration.
• Noise reduction and irrelevant interaction filtering using a spatial mask for improved prediction accuracy.

Main Results:

• Trafficformer demonstrated superior prediction accuracy compared to six baseline methods on the Seattle Loop Detector dataset.
• The model effectively identified key road network sections, indicating robust performance.
• Evaluated using Mean Absolute Error, Mean Absolute Percentage Error, and Root Mean Square Error.

Conclusions:

• Trafficformer offers significant improvements in short-term traffic forecasting accuracy.
• The model shows great potential for intelligent traffic control optimization and refined traffic resource allocation.
• Effective modeling of spatiotemporal traffic dynamics is crucial for advanced transportation management.