Abstract: Data-driven time series prediction models usually require a large amount of training data, and the accuracy of modeling will decline when the amount of data is insufficient. Aiming at the problem of few shot in time series prediction, this paper proposed a method to enhance time series data based on improving VAE, which aims to generate virtual data different from the original data but similar in distribution. The multi-head self-attention mechanism was introduced into the coding network to mine the deep features of the original data and provide comprehensive feature information for the decoding network to generate data. The residual connection is introduced to avoid the problem of disappearing gradient. Due to the trend and periodicity of time series data, the trend component and seasonal component were introduced into the decoding network to accurately represent the time characteristics of original data, and to endow the generation process of data with time interpretability. In order to verify the effectiveness of the proposed method, the experimental comparison was made with the current commonly used time series data enhancement methods. The experimental results show that the proposed method has good performance in virtual sample generation and time series regression prediction.