Abstract: To address issues such as multivariate nonlinear coupling, inadequate long-term dependency modeling, and serious lagged response of the model for the task of mine gas concentration time series prediction, this paper proposed a hybrid prediction model, xLSTM-Informer, which fused the Extended Long Short-Term Memory (xLSTM) network with the Informer architecture. The xLSTM was used as a pre-processor to extract fluctuation patterns and inter-variable coupling information within short time windows through multi-layer residual memory units, converting them into a structured time series representation. The processed temporal representation was then input into the Informer backbone architecture to further extract global dependencies and stable trends over an extended time window, thereby enhancing the temporal continuity of predictions while preserving detailed responses. Based on multi-source environmental parameter data collected from an underground tube bundle monitoring system, a feature importance analysis was conducted to select O₂ concentration, temperature, and wind speed as input variables for constructing the input feature system. The model utilized xLSTM to extract deep temporal features, and the ProbSparse attention mechanism within the Informer effectively captured global dependencies in the temporal features, thus improving the model’s ability to predict non-stationary gas concentration. To evaluate the performance advantages of the xLSTM-Informer model, it was compared with the xLSTM, Transformer, and classic Informer models. The results showed that the xLSTM-Informer model achieved optimal performance on the metrics of Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and Coefficient of Determination R². Its R² reached 0.954, representing improvements of 21.4%, 17.8%, and 19.4% over the comparison models, respectively. To further validate the effectiveness and adaptability of the xLSTM-Informer model, measured sensor data from a fully mechanized mining face in a mine were used for a case study. The model was also compared with four other hybrid models: LSTM-Transformer, RNN-Informer, LSTM-Informer, and a Bidirectional LSTM-Informer (BiLSTM-Informer). The results indicated that the xLSTM-Informer model outperformed the comparison models in its response to gas concentration trends and key turning points, demonstrating high goodness of fit and temporal synchronization.