Global-Local Spatial-Temporal Residual Correlation Network for Urban Traffic Status Prediction

Area of Science:

• Artificial Intelligence
• Machine Learning
• Urban Computing

Background:

• Existing Spatial-Temporal Residual Network (ST-ResNet) models effectively extract spatial and temporal features for urban traffic prediction.
• However, ST-ResNet primarily focuses on local spatial characteristics, neglecting crucial global spatial information.

Purpose of the Study:

• To propose a novel Global-Local Spatial-Temporal Residual Correlation Network (GL-STRCN) model.
• To enhance the accuracy of urban traffic status prediction by integrating global and local spatial features.

Main Methods:

• Utilized Pearson's correlation coefficient to identify highly correlated data series.
• Incorporated 2D convolution and residual operations to capture both global and local spatial features.
• Developed a novel temporal feature extraction component using Long Short-Term Memory (LSTM) or Gated Recurrent Unit (GRU).
• Aggregated spatial and temporal features using a weighted approach for final prediction.

Main Results:

• The GL-STRCN model demonstrated superior prediction performance on TaxiCD and PEMS-BAY datasets.
• Outperformed baseline models including CNN, ST-ResNet, GL-TCN, and DGLSTNet in urban traffic status prediction tasks.

Conclusions:

• The proposed GL-STRCN model effectively captures comprehensive spatial-temporal dependencies for improved urban traffic prediction.
• GL-STRCN offers a significant advancement over existing methods, paving the way for more accurate traffic management systems.