Multi-Period Diffusion Generative Graph Recurrent Transformer Network for Traffic Flow Prediction in Vehicular Networks

Real-time and accurate traffic flow prediction enhances the efficiency of vehicular networks and improves traffic management. As vehicular networks become essential in intelligent transportation systems (ITS), precise predictions optimize performance and reduce congestion. However, graph convolutional network (GCN)-based methods rely on preprocessed data, losing critical features and reducing accuracy in complex environments. To address this challenge, we propose a novel Multi-period Diffusion generative Graph Recurrent Transformer Network (MD-GRTN) for traffic flow prediction. MD-GRTN’s innovation lies in its enhanced performance under noisy conditions. It consists of three main components: the Multi-period Diffusion Attention Fusion (MDAF) module, the Multi-Graph Recurrent Convolution (MGRC) module, and the Spatial-Temporal Transformer (STFormer) module. The MDAF module combines the Multi-period Diffusion (MD) module and the Multi-head Attention Fusion (MAF) module to enhance historical trend features of multi-period traffic flows. The MGRC module integrates a multi-graph fusion module, a graph convolutional network, and a gated recurrent unit to strengthen spatial features influenced by various factors. Lastly, the STFormer module, comprising spatial and temporal transformer modules, further enhances the global dynamic spatial-temporal features of traffic flow. The key advantage of MD-GRTN lies in its ability to effectively capture multi-period temporal dependencies, global spatial correlations, and latent features in noisy traffic data, significantly improving prediction robustness under real-world conditions. Extensive experiments on the PEMS (03–08) datasets demonstrate that MD-GRTN outperforms state-of-the-art models, achieving average RMSE reductions of 4.7% when compared with STFGCN and PDFormer.