An Adaptive Spatio-Temporal Traffic Flow Prediction Using Self-Attention and Multi-Graph Networks.

Area of Science:

• Intelligent Transportation Systems (ITSs)
• Traffic Flow Prediction
• Spatio-temporal Data Analysis

Background:

• Current traffic flow prediction models struggle with complex spatio-temporal patterns.
• Existing methods often use single models for temporal dependencies, ignoring varying time influences.
• Limited ability to model complex spatial relationships due to static or dynamic graphs.

Purpose of the Study:

• To introduce the Adaptive Spatio-Temporal Attention-Based Multi-Model (ASTAM) for enhanced traffic flow prediction.
• To address limitations in modeling temporal and spatial dependencies in traffic data.
• To improve the accuracy and robustness of traffic flow prediction systems.

Main Methods:

• ASTAM utilizes multi-temporal gated convolution with multi-scale temporal inputs for non-linear temporal correlations.
• Employs static and dynamic parallel multi-graphs to model complex spatial dependencies.
• Incorporates a spatio-temporal self-attention mechanism for dynamic, long-term variations.

Main Results:

• ASTAM outperformed 13 baseline approaches on four real-world traffic datasets.
• Achieved average performance improvements: 5.0% reduction in Mean Absolute Error (MAE), 13.28% in Root Mean Square Error (RMSE), and 6.46% in Mean Absolute Percentage Error (MAPE).

Conclusions:

• The proposed ASTAM architecture effectively captures complex spatio-temporal dependencies in traffic flow.
• ASTAM offers a significant advancement in traffic flow prediction accuracy and performance.
• Demonstrates the potential of adaptive multi-model approaches in intelligent transportation systems.