In recent years,Spatial-Temporal Graph Convolutional Network (STGCN) has been introduced into traffic flow prediction, which has good spatial-temporal traffic data modeling ability and has achieved advanced performance, but there are still two problems: (1) Traffic flow data have strong temporal and spatial correlation; (2) Static pre-defined graphs are difficult to capture the spatio-temporal dependence of dynamic changes in traffic flow over time. To solve the above problems, a new Spatial-Temporal Decomposed Framework (STDF) is proposed, which uses residual connection, forgetting gate and update gate to organically connect time module and space module to decompose and predict input information in multiple dimensions. In addition, by instantiating STDF, a new traffic prediction model based on input traffic signal decomposition Decomposed Dynamic Spatial-Temporal Graph Convolutional Network (DDSTGCN) is proposed. It captures the spatiotemporal dependencies of traffic and designs a dynamic graph learning module that takes into account the dynamic nature of spatial dependencies. Finally, two real traffic flow data are used to compare with the existing traffic flow prediction algorithms. The experimental results show that the proposed method has good performance in the accuracy of traffic flow prediction and can effectively complete the traffic flow prediction in the real scenario.