TFGAN: A time-series Generation Adversarial Networks based on Time-Frequency Consistency

TimeGAN, a classical time series generation model, has been widely concerned for its high quality synthesis of synthetic data with time series statistical and local static features. However, the performance of the model is limited by the failure to integrate frequency domain information and the degree of training of the generator is greater than that of the discriminator during training. In order to solve these problems, the TFGAN is proposed in this paper, which introduces time-frequency consistency constraints and designs a new discriminator framework (TFC) combined with low-frequency signal analysis, aiming to further improve the accuracy and generalization of the original model. The TFGAN is based on the external framework of TimeGAN and effectively captures the complex temporal dynamics of the original data. Through the internal framework based on time-frequency consistency constraints, the periodic and trend features hidden in the original data are further captured, and the performance of the generator and discriminator is balanced to a certain extent. The quantitative and qualitative experimental results show that TFGAN outperforms TimeGAN and some existing advanced methods on various real benchmark datasets. In terms of quantitative indicators (prediction scores and discrimination scores) on the benchmark dataset, the performance of TFGAN is improved by $16.68 \%$ compared with the original model, and the performance on the Stock dataset is improved by $\mathbf{2 6. 4 3 \%}$.