To overcome the limitations of existing gas concentration prediction models in capturing multi-scale time-frequency features and modeling long-range temporal dependencies, time-frequency and stabilized attention network (TFSA-Net), a novel deep learning framework for forecasting working-face gas concentrations was proposed. The architecture first leveraged wavelet transform-based convolution to extract multi-scale time-frequency features, thereby enhancing the model’s ability to represent dynamic patterns across diverse temporal scales. Next, a stabilized attention-enhanced gated recurrent unit (SA-GRU) was incorporated to capture sequential dependencies effectively, with layer normalization and multi-head attention mechanisms jointly improving training stability and enabling more efficient exploitation of long-range contextual relationships. Finally, an adaptive spectral module was introduced to transform features into the frequency domain via the fast Fourier transform (FFT), where an adaptive high-frequency masking strategy selectively strengthened salient frequency components while attenuating noise interference. The experimental results show that TFSA-Net consistently outperforms state-of-the-art models such as Nlinear, LDlinear, long short-term memory (LSTM), bidirectional long short-term memory (BiLSTM), gate recurrent unit (GRU), bidirectional gated recurrent unit (BiGRU), multi-head attention (MHA)-BiGRU, Crossformer, and Informer in terms of mean absolute error (MAE) and root mean square error (RMSE) metrics, with reductions ranging from 6.09% to 43.47% and 5.42% to 36.45%, respectively. Compared to the best-performing comparative model MHA-BiGRU, its error metrics are reduced by approximately 0.6% and 5.42%, while the R² score improves by about 0.54%. Real-world engineering validation demonstrates that TFSA-Net delivers superior prediction accuracy and generalization performance, providing reliable and robust technical support for intelligent early-warning systems in coal-mine gas hazard prevention.