Abstract: Traffic flow prediction is the core component of the intelligent transportation system. In view of the low accuracy of the current traffic flow prediction methods, a new traffic flow prediction model GA2T is proposed. By building a Transformer architecture with fused encoder and decoder to model traffic data temporally, and using graph attention networks to model traffic data spatially, the complex spatial-temporal dependencies of traffic flow are captured. The experimental results on two real traffic datasets METR-LA and PEMS-BAY show that compared with the best baseline model DCRNN, GA2T reduces the three evaluation metrics (MAE, MAPE, RMSE) by 0.25, 0.38, 0.89 and 0.14, 0.34, 0.44. This proves the effectiveness and feasibility of GA2T, and verifies its advances in peer works.