Abstract: Objectives: Aiming at the problem that the pre-defined static graph structure cannot contain global spatiotemporal information and non-stationarity in traffic flow prediction, a traffic flow prediction model combining variational mode decomposition (VMD) and adaptive graph convolutional gated recurrent network (AGCGRN) is proposed. Methods: First, the model first uses the VMD algorithm to decompose the original traffic flow sequence data into a series of amplitude modulated and frequency modulated intrinsic mode function components, reducing the non-stationary nature of the sequence. Then, an adaptive adjacency matrix is introduced to establish an adaptive graph convolutional network, and the adaptive graph convolution is embedded in the entry-level control loop unit. An AGCGRN is designed to achieve adaptive learning of the spatiotemporal characteristics of traffic flow at each node. Finally, the AGCGRN model is used to predict each component, and the predicted values of each component are reconstructed to obtain the final predicted values. Results: Experimental validation was conducted using the PeMS04 dataset and compared with seven models: HA, SVR, LSTM, GRU, STGCN, ASTGCN, and STSGCN. (1) In the one-step prediction experiment, the proposed VMD-AGCGRN model had the highest prediction accuracy among all models, and compared to the suboptimal STSGCN model, it reduced the evaluation indexes of MAE, RMSE and MAPE by 33.266%, 36.166%, and 33.443%, respectively. (2) In multi-step prediction experiments, the proposed VMD-AGCGRN model consistently outperformed other comparative models in predicting results at different times, and its growth was relatively stable. Compared with the suboptimal model, the MAE, RMSE and MAPE of the predicted results at 15 minutes, 30 minutes, and 60 minutes decreased by 30.659%/34.106%/31.397%, 28.597%/32.809%/28.474%, and 24.846%/30.765%/20.589%, respectively. (3) In order to evaluate the effectiveness of various components in the proposed model, each ablation model was compared with the proposed model. Compared with the VMD-GCGRN model without adaptive graph structure, the proposed model reduced the MAE, RMSE, and MAPE evaluation indicators by 15.110%, 13.930%, and 32.358%, respectively. Compared with the AGCGRN model without VMD module, the proposed model reduced the MAE, RMSE, and MAPE evaluation indexes by 26.794%, 31.553%, and 32.802%, respectively. Overall, the proposed model has higher prediction accuracy compared to all ablation models, indicating that the introduction of VMD and adaptive graph structure has a positive contribution to predicting traffic flow. Conclusions: The VMD-AGCGRN model can effectively extract the spatiotemporal characteristics of traffic flow and reduce the impact of non-stationary factors, improve the accuracy of traffic flow prediction, and has certain feasibility and superiority. In traffic flow prediction, the graph convolutional network optimized by adaptive graph structure learning can better learn the spatial correlation characteristics of the traffic network, and the introduction of VMD can improve the prediction accuracy of traffic flow.