Abstract: Coal mine gas concentration series exhibit nonlinear, nonstationary, and multiscale fluctuation characteristics. Existing prediction models have limited capacity to jointly handle these characteristics, resulting in insufficient prediction accuracy. To address this problem, a gas concentration prediction model based on an Improved Sparrow Search Algorithm (ISSA) and Long Short-Term Memory (LSTM), integrating STL and Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), was proposed. STL was used to extract the trend, seasonal, and irregular components of the gas concentration series. CEEMDAN was then employed to decompose the strongly nonlinear irregular component, thereby eliminating nonstationary interference from the original gas concentration data. ISSA incorporating opposition-based learning and a Lévy flight mechanism was used to optimize the key hyperparameters of LSTM, and an ISSA-LSTM model was constructed. The ISSA-LSTM model was used to predict the components decomposed by STL and CEEMDAN, and the predicted values were reconstructed to obtain the gas concentration prediction results. Experiments were conducted using gas concentration data measured in a coal mine. The results showed that the ISSA-LSTM model outperformed LSTM models optimized using genetic algorithms, particle swarm optimization, and other algorithms in terms of the number of convergence iterations, convergence success rate, and prediction error. In three-fold cross-validation across different time periods, the ISSA-LSTM model integrating STL and CEEMDAN achieved coefficients of determination (R²) greater than 0.998 and Mean Absolute Percentage Errors (MAPEs) below 0.004. For the 1- and 7-h-ahead prediction tasks, the MAPEs were 0.018 7 and 0.026 9, respectively. In cross-dataset validation, the R² was 0.955 9 and the MAPE was 0.012 17, demonstrating the strong generalization performance of the proposed model.