时间序列生成模型的评估方法主要分为定性和定量两部分。定性的方法可 以评价生成数据多样性，但主观性较强；定量的方法中，一些方法可以评价 时间序列数据真实性但不能同时评价多样性，另一些同时评价多样性和真实 性的指标不能针对时间序列的特点进行评估。因此，本文提出了一种新的 TS-FID（Time Series Frechet Inception Distance）分数的评价指标，它结合了 时间序列表示学习模型 TS2vec 和图像生成数据评估方法 FID 分数，既可以针 对时间序列特有的时间依赖性进行评估，又兼具 FID 分数评估数据多样性和 真实性的能力。这是一个定量的，全面的评估指标，对时间序列生成数据的 评估具有普适性。在实验部分，本文使用 TS-FID 指标对 TimeGAN 生成模型 进行评估，与 PCA、t-SNE 可视化和 Wasserstein 距离等评估方法做对比，证 实了 TS-FID 分数是一个可信的指标。

The evaluation methods of time series generation models are mainly divided into two parts: qualitative and quantitative. Qualitative methods can evaluate generated data diversity but are more subjective; among quantitative methods, some methods can evaluate time series data veracity but not diversity at the same time, and other metrics that evaluate both diversity and veracity cannot be evaluated for the characteristics of time series. Therefore, this paper proposes a new evaluation metric, TS-FID score, which combines the time series representation learning model TS2vec and the image generation data evaluation method FID score, to evaluate both the time-dependence specific to time series and the ability of FID score to evaluate data diversity and truthfulness. This is a quantitative, comprehensive assessment metric that is generalisable to the evaluation of time series generated data. In the experimental part of the paper, the TS-FID metric is used to evaluate TimeGAN-generated models against evaluation methods such as PCA, t-SNE visualisation and Wasserstein distance, confirming that the TS-FID score is a credible metric.