Abstract: The superior performance of deep learning in predicting time series processes largely benefits from the large number of training samples. However, in the actual process, sample data is usually difficult to collect and cannot be accurately modeled. In order to solve the problem of small samples in time series prediction, this paper proposes a data augmentation network (ATCLSTM-TimeGAN) based on the attention mechanism and integrating temporal convolutional network and long short-term memory network. By incorporating a Soft-Attention mechanism into the Time-series Generative Adversarial Network (TimeGAN), the problem of dynamic information loss is addressed. Because the input to the generator is a series of random vector, the temporal convolution structure is combined with the Self-Attention mechanism to make the distribution of value in the normalized range interval and the distribution of real data form a corresponding relationship and achieve better data generation performance. In order to verify the authenticity and usefulness of the generated data, this paper compares the distribution differences of the data generated by different data augmentation methods and the predictive performance of the synthesized data when used for prediction. In the actual process dataset, the usefulness of the synthetic data for prediction is further verified. The experimental results show that compared to other data enhancement methods, ATCLSTM-TimeGAN can better cover the distribution of the original data and effectively reduce the prediction error under small samples.