Abstract: Accurate prediction of gas concentration is crucial for preventing gas disasters. The prediction accuracy is influenced by both the temporal variation patterns of gas concentration and the spatiotemporal distribution characteristics of gas diffusion. Existing model-driven prediction methods struggle to handle long-term and large-scale gas concentration prediction tasks, while data-driven prediction methods do not consider the impact of dynamic spatial features, resulting in poor generalization performance. To capture the spatiotemporal dependency of gas concentration changes and improve the prediction accuracy, a Temporal-Dynamic Graph Convolutional Transformer with Multi-Scale Mechanism (TDMformer) was proposed to construct a gas concentration prediction model. Based on the ITransformer framework, a temporal-variable attention mechanism was designed to model the temporal and variable features simultaneously. A dynamic graph convolutional network was integrated to describe the topology of underground gas sensor networks and capture the spatial dependency of gas concentration data. A multi-scale gated Tanh unit was introduced to enhance the multi-scale feature extraction capability. The experimental results showed that, compared with Graph-WaveNet, GRU, Transformer, AGCRN, DSformer, STAEformer, and FourierGNN, the root mean square error of the TDMformer model decreased by 24.87%, 26.37%, 21.69%, 19.57%, 11.90%, 10.84%, and 9.20%, respectively. The mean absolute error decreased by 17.09%, 25.58%, 26.89%, 14.56%, 11.10%, 5.75%, and 4.53%, respectively. The coefficient of determination increased by 5.94%, 6.51%, 4.79%, 4.12%, 2.21%, 2.08%, and 1.76%, respectively, verifying that this model had higher prediction accuracy and better data fitting performance.