Graph Convolutional Network: Traffic Speed Prediction Fused with Traffic Flow Data.

Area of Science:

• Intelligent Transportation Systems
• Machine Learning
• Traffic Engineering

Background:

• Traffic speed prediction is crucial for intelligent transportation systems.
• Graph Convolutional Networks (GCNs) show promise but struggle with inaccurate historical speed data and static spatial dependencies.
• Existing methods overlook the impact of dynamic traffic on spatial relationships.

Purpose of the Study:

• To develop a novel graph convolutional network model, FSTGCN, for improved traffic speed prediction.
• To address limitations of existing GCNs by incorporating more accurate traffic flow data and dynamic spatial correlations.
• To enhance the accuracy and reliability of traffic speed forecasting.

Main Methods:

• Proposed FSTGCN model with a full convolutional structure to avoid repeated iterations.
• Fused historical traffic flow data, which has a more exact mapping relationship with traffic speed, to reduce prediction error.
• Designed a dynamic adjacency matrix based on traffic flow covariance to capture dynamic spatial correlations.

Main Results:

• FSTGCN demonstrated superior performance compared to state-of-the-art methods in traffic speed prediction.
• The integration of traffic flow data significantly reduced prediction errors.
• The dynamic adjacency matrix effectively captured real-time spatial dependencies in the traffic network.

Conclusions:

• FSTGCN offers a significant advancement in traffic speed prediction accuracy.
• The model's ability to integrate dynamic traffic flow and spatial information provides a more robust forecasting approach.
• This research contributes to the development of more efficient and intelligent transportation systems.