Abstract: Gas emission is influenced by the coupling of multiple factors, and its data exhibit complex nonlinearity, strong temporal correlation, and dynamic variations. To address these characteristics, this study introduces variational mode decomposition (VMD), grey wolf optimization (GWO), and bidirectional long short-term memory (BiLSTM) network, and proposes a dynamic prediction model following the "feature extraction-parameter optimization-regression prediction" framework to accurately forecast gas emission. First, VMD decomposes the gas emission sequence to capture features across different frequency components, thereby reducing the complexity and volatility of the data. Then, by leveraging the global search capability of GWO and the temporal modeling ability of BiLSTM, a GWO-BiLSTM prediction model is established for each sub-mode. Finally, the prediction results of all sub-modes are reconstructed to obtain the final gas emission prediction. Experimental studies are conducted using data collected from a mine in Shanxi Province. The results show that the proposed model achieves significantly lower errors than LSTM, BiLSTM, and GWO-BiLSTM, with a coefficient of determination (R²) of 0.989 1. This provides an efficient and reliable solution for gas emission prediction under complex multi-factor coupling.