Current traffic flow prediction methods based on spatio-temporal feature extraction has problems of insufficient capture of global spatial correlation and dynamic long-term temporal dependency， where spatial correlation mining relies on the quality of graph structure heavily. Therefore， a Multi-Graph Diffusion Attention Network （MGDAN） was proposed， consisting of a Multi-Graph Diffusion Attention Module （MGDAM） and a temporal attention module. Firstly， adaptive spatio-temporal embedding generator was used to construct dynamic spatio-temporal information. Secondly， a Maximal Information Coefficient （MIC） matrix and an adaptive matrix were utilized to explore fine-grained spatial information， and a global spatial attention mechanism was employed to capture dynamic spatial correlation. Finally， the temporal attention module was used to extract nonlinear temporal correlation， and the integration of the three modules was carried out to realize effective extraction of spatio-temporal correlation. Experimental results demonstrate that， on PEMS08 dataset， the Mean Absolute Error （MAE） of MGDAN model within one hour has 19.34% and 5.74% reductions compared to those of Spatio-Temporal AutoEncoder （ST_AE） and Spatial-Temporal IDentity （STID） models， respectively. At the same time， MGDAN model outperforms 9 baseline models in overall prediction performance， and can conduct medium- and long-term traffic flow prediction accurately， providing theoretical basis for urban traffic dispersion.
