Abstract: Current research on gas concentration prediction in working faces of coal mines often suffers from limited feature dimensions and small dataset sizes, making it difficult to extract long-term fluctuation patterns from large-scale time-series data. To address this issue, this study proposes a Principal Component Analysis (PCA)-Transformer-based prediction algorithm for gas concentration in working faces. Firstly, raw gas concentration-related data was cleaned and normalized using min-max scaling. Then, PCA was applied to reduce the dimensionality of seven influencing factors (methane concentration at the upper corner, return airflow methane concentration, oxygen concentration, carbon monoxide concentration, temperature, net flow rate, and wind speed), effectively eliminating weakly correlated features. Finally, the processed training set was fed into a Transformer model, where the encoder and decoder extracted intrinsic patterns and features of gas concentration variations. Using monitoring data from working face 224 of a high-gas mine in Tongchuan as a sample, the PCA-Transformer model was compared with Long Short-Term Memory (LSTM), PCA-Long Short-Term Memory (PCA-LSTM), and Transformer models. The results show that: ① The PCA-Transformer model achieves a Mean Absolute Error (MAE) of 0.020 3, Mean Squared Error (MSE) of 0.047 2, and a runtime of 86 seconds, meeting the accuracy and timeliness requirements for gas concentration prediction in coal production. ② Compared to LSTM, PCA-LSTM, and Transformer models, the PCA-Transformer model better fits gas concentration trends, effectively identifies peak and trough sequences, and requires the least computational time, demonstrating its superior performance.