Research on traffic flow prediction based on multi-scale spatio-temporal diffusion graph convolutional network

Accurate traffic flow prediction is indispensable for addressing urban traffic issues. Aiming at the problems that complex temporal and spatial features of traffic flow dynamics are challenging to model, and that it is difficult to model traffic flow effectively at different scales by self-attention mechanism. A traffic flow prediction method utilizing a Multi-scale Spatio-temporal Diffusion Graph Convolutional Network (MSTDGCN) is proposed in this paper. MSTDGCN constructs a Node Adaptive Learning-Diffusion Graph Convolutional Network to learn dynamic correlations among road nodes at various time points and catch the dynamic spatio-temporal traits of traffic flow via multiple rounds of node information propagation and aggregation. Additionally, MSTDGCN constructs Multi-Scale Spatio-Temporal Attention, which applies the self-attention mechanism and dynamic deep convolutional modules in parallel to learn from long sequences of traffic flow data across various scales, thereby capturing multi-scale spatio-temporal features of traffic flow. As shown in the experimental results, the prediction accuracy of the MSTDGCN model proposed in this paper consistently outperforms that of the currently popular baseline model, showing excellent prediction performance.