首页|期刊导航|电力信息与通信技术|基于MAD-TimeGAN数据增强的小样本时间序列预测方法

基于MAD-TimeGAN数据增强的小样本时间序列预测方法OA

Small Sample Time Series Prediction Method Based on MAD-TimeGAN Data Enhancement

中文摘要英文摘要

针对小样本时间序列预测存在的数据不足问题,现有时间序列生成对抗网络(time-series generative adversarial networks,TimeGAN)模型在训练过程中,生成器一旦学会生成某些能够有效欺骗判别器的样本,其往往倾向于反复生成这些样本,而不再探索数据分布的其他区域,导致生成样本多样性不足,进而引发模式崩溃问题,为此,文章提出了一种多智能体时间序列生成对抗网络(multi-agent diverse time-series generative adversarial networks,MAD-TimeGAN).该网络使用多智能体生成对抗网络(multi-agent diverse generative adversarial networks,MAD-GAN)替换原始GAN网络,MAD-GAN包含多个生成器和一个判别器,多个生成器用于生成不同模式的数据,判别器用于判别数据的真伪并且判断生成的样本来自哪个生成器,通过引入该网络,不仅能够有效缓解TimeGAN存在的模式崩溃问题,还能增加生成样本的多样性,获得质量更高的生成样本.为验证所提方法的有效性,在多个实验数据集上进行数据增强实验并观察其生成数据用于预测时的预测效果.实验结果表明,与未进行改进的模型相比,所提数据增强方法能够更好地覆盖原始数据,提升预测精度.

In response to the insufficient data problem of small sample time series prediction,during the training process of the existing TimeGAN model,once the generator learns to generate some samples that can effectively deceive the discriminator,it tends to generate these samples repeatedly instead of exploring other areas of the data distribution,resulting in insufficient diversity of generated samples,which in turn causes pattern crash problems.To this end,this paper proposes a multi-agent time series generalized adversarial networks(MAD-TimeGAN).The network uses a multi-agent diverse generative adversarial network(MAD-GAN)to replace the original GAN network.MAD-GAN includes multiple generators and a discriminator.Multiple generators are used to generate data in different modes.The discriminator is used to determine the authenticity of the data and determine which generator the generated samples come from.By introducing this network,it can not only effectively alleviate the pattern crash problem of TimeGAN,but also increase the diversity of generated samples and obtain higher-quality generated samples.To verify the effectiveness of the proposed method,data enhancement experiments were performed on multiple experimental data sets and the prediction effect of the generated data was observed when the prediction was predicted.Experimental results show that compared with the unimproved model,the proposed data enhancement method can better cover the original data and improve prediction accuracy.

尼俊红;孙浩;荀超;陈延滔;石超君

华北电力大学电气与电子工程学院,河北省 保定市 071003华北电力大学电气与电子工程学院,河北省 保定市 071003国网福建省电力有限公司,福建省 福州市 350003国网福建省电力有限公司经济技术研究院,福建省 福州市 350003华北电力大学电气与电子工程学院,河北省 保定市 071003

信息技术与安全科学

多智能体时间序列预测小样本数据增强MAD-TimeGAN

multi-agenttime series predictionsmall sample sizedata enhancementMAD-TimeGAN

《电力信息与通信技术》 2026 (8)

27-35,9

国家自然科学基金项目(62206095)国网福建省电力有限公司经济技术研究院科技项目(SGFJJY00NYJS2400093).

10.16543/j.2095-641x.electric.power.ict.2026.08.03

评论