Abstract: Gas concentration series exhibit non-stationarity, multi-scale fluctuation characteristics, and long-range temporal dependencies, and their evolution is influenced by the coupling of multiple environmental factors. Existing gas concentration prediction models mainly focus on single time series modeling or shallow feature combinations, making it difficult to simultaneously capture temporal dependency representation and cross-variable association modeling. To address this issue, a dynamic prediction model of gas concentration based on deep feature fusion and temporal dependency modeling was proposed. First, Variational Mode Decomposition (VMD) was introduced to adaptively decompose the original gas concentration series into several Intrinsic Mode Functions (IMFs) and a residual component. Then, graph nodes for variables were constructed by combining the VMD decomposition results with multi-source environmental parameters, and an adjacency matrix was established based on the correlations between different environmental parameters and the gas concentration series, providing structural priors for cross-variable association modeling. Next, a Temporal Convolutional Network (TCN) was used to extract short-term fluctuation features and long-term dependency information from multivariate sequences composed of IMFs, residual components, and multi-source environmental parameters. Finally, a Multi-Head Graph Attention (MGA) with adjacency mask constraints and scaled dot-product attention was employed to model dynamic coupling relationships among variables and to achieve heterogeneous feature fusion from multiple sources. The experimental results showed that, compared with mainstream prediction models, the proposed model achieved the best performance in Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (R²), with values of 0.028 6, 0.021 5, and 0.954, respectively, and outperformed the comparison models in overall accuracy, local fluctuation characterization, and adaptability to complex scenarios.