Traditional methane extraction concentration prediction methods have limited capability in capturing complex temporal features and responding to dynamic operating conditions. To address these limitations, a multi-parameter dynamic prediction model for methane extraction concentration based on the long short-term memory (LSTM) neural network was developed. The LSTM-based prediction model was implemented in a Python environment. Monitoring data collected from a key monitoring point on the main pipeline of a methane extraction system in a mine located in Jiaozuo mining area were used as the sample dataset. Data preprocessing, including missing value imputation, outlier correction, and normalization, was performed to improve data quality. Model hyperparameters, including the time step, number of neurons, and batch size, were optimized using an early stopping strategy during the training process. The results show that the optimized LSTM model exhibits high prediction accuracy and good trend-fitting performance. The mean absolute error (MAE) is 0.585%, the root mean square error (RMSE) is 0.759%, and the mean absolute percentage error (MAPE) is 2.90%. The proposed model effectively predicts the temporal variation of methane extraction concentration with low prediction error and stable overall performance. The findings provide a theoretical basis for intelligent regulation of coal mine methane extraction systems and offer technical support for efficient methane extraction.