High precision traffic flow prediction plays an important role in traffic management and intelligent travel in large cities. Mining the temporal and spatial correlation of traffic flow dynamics is the key to improve the prediction accuracy. In view of the insufficient consideration of the high similarity of traffic flow in different time scales and the similarity of traffic flow changes between non adjacent nodes in similar functional areas, a dynamic graph convolution neural network considering spatio-temporal similarity(STS-DGCN) is constructed. The time similarity of traffic flow data is expressed by the data input tensor under multiple time scales such as adjacent periods, days and weeks, and the time-space similarity of traffic flow data is expressed by the adjacency matrix with multi-attribute characteristics such as distance measurement, similarity measurement, adaptive embedding and dynamic correlation between road network nodes. Then, based on these adjacency matrices, a dynamic graph reflecting the temporal and spatial dynamic changes of road network nodes is constructed, the corresponding spatio-temporal feature mining algorithm is designed. The results show that the prediction result of the model is better than the more advanced baseline model and has higher prediction accuracy.