Aiming at the problems that the multi-scale features and long-term dependencies of traffic flow are not easy to model synchronously and the higher-order hidden dynamic spatiotemporal features are challenging to capture, this paper proposes a multi-scale dual-dynamic spatiotemporal graph convolution network (MSD \(^{\varvec{2}}\) GCN) for traffic flow prediction. The dual hypergraph is employed to efficiently capture the spatiotemporal dependencies of the edges of the traffic flow graphs. The dynamic interactive convolution (DIC) between the dual-dynamic GCNs propagation features, the spatiotemporal dynamic hypergraph convolutional network (STDHCN), and the spatiotemporal dynamic graph convolutional network (STDGCN) are used to model the dynamic properties of the traffic flow graph’s edges and nodes, respectively. The Multi-scale Axial Attention Module (MSCAM) utilizes bar convolution to incorporate multi-scale features into the axial attention computation. It establishes a double cross-attention between the two traffic flow spatial axial attentions to better capture different spatial flow scale features and improve the model’s long-term prediction capability. The experimental findings demonstrate that the MSD \(^{\varvec{2}}\) GCN model outperforms the majority of the existing baseline models in terms of prediction performance.
