To overcome the limitations of existing gas concentration prediction models in capturing multi-scale time–frequency features and modeling long-range temporal dependencies, this study proposes TFSA-Net, a novel deep learning framework for forecasting working-face gas concentrations. The architecture first leverages 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) is 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 is introduced to transform features into the frequency domain via the fast Fourier transform (FFT), where an adaptive high-frequency masking strategy selectively strengthens salient frequency components while attenuating noise interference. Experimental results show that TFSA-Net consistently outperforms state-of-the-art models—including Nlinear, LDlinear, LSTM, BiLSTM, GRU, BiGRU, MHA-BiGRU, Crossformer, and Informer—by achieving error reductions of 6.09%–43.47% in MAE and 5.42%–36.45% in RMSE. Notably, when compared to the strongest baseline, MHA-BiGRU, TFSA-Net further reduces MAE and RMSE by 0.6% and 5.42%, respectively, while increasing R2 by 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.