Decomposition dynamic multi-graph convolutional recurrent network for traffic forecasting

Abstract

Accurate traffic flow prediction is crucial for urban traffic management. Traffic data is typically collected from sensors deployed along roadways, which often record both valid and erroneous data. However, most existing studies assume that the collected data is perfectly accurate, overlooking the existence of erroneous data. Meanwhile, graph neural networks are widely applied in traffic forecasting due to their ability to effectively capture correlations between nodes in a network. However, existing methods often rely solely on either static or dynamic graph structures, which may not accurately reflect the complex spatial relationships between nodes. To address these issues, we propose a decomposition dynamic multi-graph convolutional recurrent network (DDMGCRN). DDMGCRN utilizes a residual decomposition mechanism to separate erroneous data from valid data, thereby mitigating its impact. Additionally, DDMGCRN introduces sensor-specific spatial identity embeddings and timestamp embeddings to construct dynamic graphs. It further integrates static graphs for multi-graph fusion, facilitating more effective spatial feature extraction. Furthermore, to address the limitations of RNN-based models in capturing global temporal dependencies, DDMGCRN incorporates a global temporal attention module. Experimental results on four real-world datasets show that DDMGCRN outperforms all baseline models on the PEMS08 dataset, achieving a mean absolute error (MAE) of 14.13, which improves performance by approximately 4.85% compared to the best baseline model. The source code is available at https://github.com/hulongfei123/DDMGCRN.