Objectives Accurate traffic flow prediction plays a crucial role in enhancing traffic efficiency and safety, making it a prominent research focus in intelligent transportation systems. However, traditional static graph structures struggle to capture the global spatiotemporal dynamics of traffic flow, while short-term fluctuations in traffic introduce non-stationarity, further complicating accurate predictions.

Methods To address these challenges, this paper proposes a traffic flow prediction model that integrates variational mode decomposition (VMD) with an adaptive graph convolutional gated recurrent network. First, considering the non-stationarity of traffic flow, VMD is employed to decompose the time series into stable intrinsic mode function (IMF) components. Then, the model leverages an adaptive adjacency matrix and graph convolutional gated units to dynamically learn the complex spatiotemporal interactions among the IMF components, enabling a more precise characterization of traffic flow trends. Finally, the reconstructed model predicts the IMF components, yielding the overall traffic flow forecast.

Results The experiments conducted on the PeMS04 dataset demonstrate that the proposed model significantly outperforms baseline models in metrics such as mean absolute error, root mean square error, and mean absolute percentage error, achieving over 33% improvement. Compared to the ablation models, the performance of the proposed model improves over 15%.

Conclusions The results highlight the substantial advantages and the promising potential of the proposed model for applications in traffic flow prediction tasks.