Accurate prediction of coal mine gas emission is of great significance for ensuring safe production in coal mines. However, traditional prediction methods often fail to fully capture the complex nonlinear relationships between gas emissions and multidimensional features. To address the issues of data insufficiency and improve model accuracy, a gas emission prediction method based on generative adversarial networks (GAN) and a dual-parallel spatiotemporal convolutional attention architecture was proposed. First, synthetic data with a distribution similar to that of the source data were generated through GAN to augment the training dataset and enhance the robustness of the model. Subsequently, a spatiotemporal convolutional network was employed to extract spatial dependency features from the data, and a temporal attention mechanism was integrated to focus on features at critical time steps, thereby further strengthening the model's capability in modeling spatiotemporal nonlinear relationships. Through a comparison of 30 mainstream machine learning algorithms, XGBoost was selected as the core model and combined with the spatiotemporal convolutional attention feature extraction module. Experimental results indicate that the proposed model effectively improves prediction accuracy and stability across different mining areas and under multi-source heterogeneous data conditions. Compared with traditional regression models and classical deep learning methods, the model demonstrates superior performance in terms of the coefficient of determination (R²) and mean absolute error (MAE), among other metrics.