Despite the rapid development of new energy vehicles, fuel vehicles remain essential for long-distance transport. Therefore, monitoring and predicting urban road vehicle emissions is a key measure to regulate urban pollution status. However, the existing research generally lacks causal explanation and is not robust to noise, and how to improve the long-term prediction ability of the model is a difficult problem. To solve the above problems, In this paper, we propose a diffusion variational graph neural network combined with a selective state space Mamba bidirectional temporal scan module. SDE is used to enhance the interpretability of the relationship between nodes and the robustness of the model, and bidirectional temporal scanning and spatiotemporal feature fusion are used to improve the prediction accuracy. Specifically, this paper first uses VG-AE and SDE to reason about the dynamic adjacency relationship between node-s, and then uses dynamic graph convolution, spatiotemporal feature fusion, and bidirectional temporal scanning modules for prediction. In this paper, experiments are carried out on the emission dataset of Xi ‘an City in China and the Tdrive transportation dataset, with ablation experiments performed on both datasets to demonstrate the effectiveness of the model and related modules.