Abstract: Traditional traffic flow prediction models use static adjacency matrices for spatiotemporal modeling, ignoring the potential spatial dependencies between road network nodes and the periodic similarity presented by Traffic flow data. Therefore, a dynamic spatio-temporal graph network model ADSTGN based on multi-head attention mechanism is proposed. Firstly, the traffic flow data are processed into three channel periodic time series, and the temporal features of sequence data are modeled by temporal and positional encoding to ensure sequence order. Secondly, a dynamic graph generator module is proposed in terms of spatial dimension, which uses similarity spatial features and distance features to represent road network structure information, and combines adaptive adjacency matrix to capture hidden spatial dependencies between road network nodes. Finally, a method of using one-dimensional convolution to integrate local context for multi head attention mechanism is proposed in terms of time dimension, which fully captures the time correlation under three channel periodic data and filters historical redundant information. Experiments are conducted on four publicly available datasets on highways(PEMS03, PEMS04, PEMS07 and PEMS08), and the results show that the ADSTGN model has better performance indicators compared to existing baseline models, demonstrating the effectiveness of the model in traffic flow prediction.