With the transfer of coal mining to deeper areas in China, the threat of gas disasters is becoming increasingly serious and more complex. Gas concentration prediction in mining face is an important means to reduce or even avoid the losses caused by gas hazards. In order to solve the low accuracy in long-term gas concentration prediction caused by the difficulty in mining time-dependent features and strong noise interference, we propose a method by combining the complete ensemble empirical mode decomposition with adaptive noise analysis (CEEMDAN) and gated recurrent unit (GRU) model to predict gas concentration in the long-term. Firstly, gas concentration data is pre-processed using \(3\sigma\) and the multiple Lagrangian interpolation method to identify the abnormal monitoring data and supplement missing data. In response to the strong noise interference and severe signal distortion in on-site monitoring data of gas concentration. An adaptive noise complete set empirical mode decomposition method is adopted to separate the high-frequency noise components in gas signals. Wavelet decomposition method is then used to remove the high-frequency noise component by optimally selecting the wavelet coefficient thresholds. On this basis, a gated recurrent unit is used to autonomously analyze the fluctuation specificities of the gas concentration signal after denoising to deeply mine the temporal dependence of gas concentration fluctuations and improving long-term prediction accuracy. The on-site verification shows the proposed method can effectively remove strong noise of the fully mechanized mining face and achieve high accuracy of gas concentration long-term prediction.