In order to prevent gas disasters during coal mining, this paper proposes an intelligent prediction model SBLPformer, which combines Transformer, BiLSTM and time series decomposition algorithm to predict gas concentration in working face. The model mainly includes three parts: data sequence decomposition, data interaction and attention mechanism. Using codec architecture, the multi-processing flow modeling method is proposed for different pattern information in data. Firstly, the data is decomposed, the time series is converted into frequency domain information by Fourier transform, and three types of pattern information are aggregated according to the parity and zero value classification of non-zero frequency, so as to realize the detailed decomposition of complex patterns. Secondly, the data interaction module is used to extract and interact information through convolution and pooling in the feature dimension to coordinate the time dependence of multiple modes. Moreover, the segmented multi-head attention mechanism is used to extract high-level abstract features through two-dimensional convolution to effectively capture the time dependence between segment sequences and solve the long-term dependence problem. Finally, make full use of BiLSTM’s advantage in dealing with nonlinear relations, decode and predict the data by BiLSTM. The experimental results show that SBLPformer has high prediction accuracy and provides a new method for the prediction of gas concentration.