Abstract: Accurate traffic forecasting can effectively solve the problems of traffic congestion and environmental pollution, but the existing methods cannot fully characterize the features of traffic data. To solve the above problems, a sequential to sequence graph convolution gated recurrent unit (Seq2Seq-GCGRU) model is proposed to extract the temporal and spatial characteristics of traffic speed. The model consisted of three parts, which were used to model the weekly, daily and near-term information of traffic speed with time shifting, and a new seq2seq training method was proposed to overcome the defect that the inherent method was not suitable for time series. The experimental results show that the proposed algorithm has higher prediction accuracy compared with other common traffic flow prediction models. The root mean square error (RMSE) and mean absolute error (MAE) are reduced by at least 25% and 24% respectively.