瓦斯浓度预测对于保障矿井安全生产至关重要。瓦斯浓度数据具有样本量少、存在时间依赖性等特点，传统机器学习等方法通常效果不佳。提出一种时间卷积改进时序生成对抗网络(TCN-TimeGAN)，基于生成对抗网络特性改善瓦斯数据小样本过拟合的问题，利用TimeGAN网络捕捉瓦斯序列的时间特征，基于TCN网络扩大感受野以便读取长时间维度特征。在损失函数设计中，利用Wasserstein距离衡量瓦斯数据分布，并给鉴别网络损失函数添加自适应权重的梯度惩罚项，以解决数据不对性、梯度消失等问题，提高训练稳定性和预测准确度。在进行模型训练时，首先对瓦斯时间序列进行归一化、数据缺失值处理等操作，将处理结果作为模型的输入序列，输入到嵌入网络和恢复网络进行训练，以降低重构损失。随后，将输入序列输入监督网络中进行训练，以减小监督损失；最后进行联合训练，其总损失为生成网络损失和鉴别网络损失之和。实验表明，改进模型生成的数据能更全面地覆盖原始数据分布，利用改进模型生成的数据进行预测所得到预测结果的均方误差(MSE)、均方根误差(RMSE)和平均绝对误差(MAE)等指标都远小于对比模型，在所有时间段都能保持稳定准确的预测。

The prediction of gas concentration is of great importantance to ensure the safety of mine production. The gas concentration data has the characteristics of small sample size and time dependence, and traditional machine learning methods are not effective. A time convolution improved time series Generative adversarial network (TCN-TimeGAN) is proposed in this paper. Based on the characteristics of generative adversarial network (GAN), the problem of over-fitting of small samples of gas data is improved, and the receptive field is enlarged based on TCN network to read long-term dimension features. In the design of loss function, Wasserstein distance is used to measure the distribution of gas data, and the gradient penalty term of adaptive weight is added to the identification network loss function, so as to solve the problems of data irregularity and gradient disappearance, and improve training stability and prediction accuracy. When conducting model training, the first step is to normalize the gas time series and process missing data values. The processing results are used as input sequences of the embedding network and recovery network to reduce reconstruction loss. Subsequently, the input sequences are also input into the supervised network to reduce supervision loss. Finally, joint training is conducted, and the total loss is the sum of the generated network loss and the discriminative network loss. The experiment results show that the data generated by the proposed model can cover the original data distribution more comprehensively, and the mean square error (MSE), root mean square error (RMSE) and mean absolute error (MAE) of the results predicted from the generated data by the improved model are much smaller than those of the comparison model, and the prediction can be stable and accurate in all time periods.