Multi-Source Information Fusion Graph Convolution Network for traffic flow prediction

As a fundamental technology in the field of intelligent transportation systems, traffic flow prediction has a wide range of applications. The utilization of Graph Convolutional Network (GCN) models is notable for capturing the complex spatial–temporal dependencies in traffic data, leading to a significant improvement in prediction accuracy. However, most existing graph construction methods overlook joint impact of auxiliary information such as weather and traffic speed on the road topology. Moreover, the research on interactions within time series at coarse temporal resolutions remains insufficiently explored, giving rise to unsatisfactory long-term prediction performance. In this study, we present a novel framework, namely Multi-Source Information Fusion Graph Convolution Network (MIFGCN), for spatial–temporal traffic flow prediction. Our key innovation lies in creating a dynamic graph that integrates weather, traffic speed, and global spatial information, effectively simulating significant traffic fluctuations caused by subtle ancillary information in the road network. Simultaneously, it captures the evolving hidden adjacency relationships between nodes over time. Furthermore, by combining with an attention-based temporal interaction module, MIFGCN learns multiscale temporal correlations at course temporal resolutions, enhancing the ability for long-term prediction. Experiments conducted on four real-world traffic datasets demonstrate that MIFGCN outperforms various state-of-the-art baselines, especially achieving a 10.50% average improvement on the PeMS08 dataset.