Urban traffic flow prediction accuracy was improved by proposing a multi-view diffusion graph neural network (MVDGNN). Static topology graphs and dynamic information graphs were integrated to construct traffic networks from multiple views. A diffusion graph convolutional network (DGCN) effectively modeled spatial dependencies, while an adaptive weighted fusion mechanism dynamically adjusted the contribution of static and dynamic features. In the temporal dimension, a multi-head attention mechanism and a bidirectional gated recurrent unit (Bi-GRU) captured temporal features. A multi-layer perceptron (MLP) produced predictions for multiple future time steps. The model was validated on two real-world datasets, PEMS04 and PEMS08. Results show that MVDGNN reduced mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE) by 3.01%-24.73%, 1.65%-36.04%, and 1.21%-47.50%, respectively, compared with baseline models. Ablation experiments confirm the collaborative effect of multi-view graph structures and spatio-temporal modules. The integrated model performs significantly better than single-view graph structures and offers new insights into urban traffic flow prediction.