Abstract: Accurate gas concentration prediction is crucial for preventing dynamic disasters in deep coal mines. To address the limitations of current prediction methods that rely heavily on historical data and dynamic-static parameters, the frequency spectrum acoustic method employed in Russian coal mines was introduced. Indicators closely related to gas concentration were screened from artificial acoustic signal indicators of the frequency spectrum acoustic method using Grey Relational Analysis (GRA) and Hierarchical Cluster Analysis (HCA). The iterative annealing strategy of the Simulated Annealing Algorithm (SAA) was applied to determine the number of modal components (K) and the penalty coefficient (α) in Variational Mode Decomposition (VMD). The optimized VMD was then used to decompose noisy gas concentration signals into several relatively stable intrinsic mode functions with different frequencies. The optimal smoothing factor (σ) of the Generalized Regression Neural Network (GRNN) was identified through the stochastic perturbation strategy of SAA. The refined GRNN model was utilized to effectively predict each modal component and reconstruct the prediction results. The findings demonstrate that the "decomposition-prediction-reconstruction" mechanism effectively suppresses noise interference and significantly reduces nonlinear complexity, thereby enhancing prediction accuracy. Compared with four alternative models, the VMD-SAA-GRNN model based on the frequency spectrum acoustic method exhibits superior generalization capability and higher precision in dynamic gas concentration prediction, providing a valuable reference for gas control in deep tunneling faces.