As a core task of intelligent transportation systems, traffic flow prediction is characterised by high spatio-temporal complexity. Due to the limitations of existing methods in modelling complex spatio-temporal dependencies, particularly regarding medium- to long-term prediction accuracy and generalisation capabilities, this paper proposes a combination prediction model based on a multi-level dynamic GCN-Transformer framework (DGTFormer), to enhance the accuracy of short- and long-term traffic flow predictions. DGTFormer adopts a dual-stream architecture to achieve spatio-temporal decoupling modelling. The spatio-temporal dynamic graph convolutional network processes dynamic changes in the road network structure, and the temporal transformer encoder processes temporal information related to traffic flow. A spatio-temporal gated fusion mechanism is introduced to deeply couple spatial and temporal information. Experimental results on three real-world traffic datasets (PeMSD4, PeMSD7 and PeMSD8) demonstrate that DGTFormer significantly outperforms mainstream baseline models on multiple key evaluation metrics. Compared with advanced baseline methods, DGTFormer achieves performance improvements of up to 8.61% and 9.15% in RMSE and MAE, respectively. Furthermore, the coefficient of determination, R², remains stable at an excellent level above 0.9 across different prediction time steps, which fully validates that the DGTFormer model possesses superior predictive performance and generalisation capabilities.
LIAN et al. (2026) studied this question.