Traffic Flow Prediction Method Based on Spatiotemporal Features and Multi-graph Fusion

To address traffic congestion, accurate traffic flow forecasting is essential. This paper presents a novel approach integrating spatiotemporal characteristics and multi-graph fusion for traffic prediction. The proposed framework employs a Transformer-based architecture to capture temporal periodicity in traffic data. Concurrently, three graph representations are utilized: A<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">s</inf> (static road network graph) for fixed spatial dependencies, A<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">d</inf>(dynamic graph) for dynamic spatial relationships, and A<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">t</inf> (similarity graph) for potential spatial correlation in traffic data respectively. These components are combined through a spatiotemporal graph convolution module to generate comprehensive spatiotemporal features. The final predictive outputs are produced via a regression layer. Validation experiments on PEMS04 and PEMS08 datasets demonstrate that this method outperforms existing approaches across multiple evaluation metrics, confirming its effectiveness and advantages in real-world traffic prediction scenarios.