For the characteristics of non-stationarity, spatial correlation and temporal dependence of short-term traffic flow time series, this paper proposes a combined prediction model of traffic flow based on improved variational mode decomposition ( VMD), graph attention (GAT) network and gated recurrent unit (GRU) network to improve its prediction accuracy and convergence speed. First, the variable mode decomposition algorithm improved by mutual information entropy (MI) is used to decompose the traffic flow time series into a series of amplitude modulation and frequency modulation signal sub-sequences, which reduces the non-stationarity of the time series signal and improves the prediction accuracy of the model. Then, they are sent to the graph attention network to capture the traffic flow of adjacent nodes of the road network to different degrees on the traffic flow of the central prediction node, so as to realize the spatial correlation modeling and further improve the prediction accuracy of the combined model. Next, the traffic flow component sub-sequences are sent to the gated recurrent unit network separately to capture the temporal dependence of the traffic flow sequence, and use the improved RMSPRop optimization algorithm to iteratively search for optimization, which not only improves the convergence speed of the optimization algorithm, but also improves the prediction accuracy of the model. Finally, the prediction values of each component subsequences are combined as the final output of the prediction model. The experiment used traffic data from the RTMC system, the results show that compared with LSTM, GCN and GAT baseline models, the mean absolute error (MAE) is reduced by 9. 35, 4. 12 and 4. 09, respectively, and the mean absolute percentage error ( MAPE) is reduced by 16. 42%, 7. 32%, and 8. 1%, respectively. The convergence speed of the optimization algorithm and the prediction accuracy of the combined model are effectively improved.