基于CDoubleGAN的电网时序暂态数据生成  

Transient instability time-series data generation in power grid based on CDoubleGAN

作　　者：张启飞[1] 陈润泽 张亶[1] 叶瑞涛 梁秀波[1] ZHANG Qi-fei;CHEN Run-ze;ZHANG Dan;YE Rui-tao;LIANG Xiu-bo(School of Software Technology,Zhejiang University,Ningbo 315103,China)

机构地区：[1]浙江大学软件学院,浙江宁波315103

出　　处：《计算机工程与设计》2025年第1期159-165,共7页Computer Engineering and Design

基　　金：国家电网有限公司总部科技基金项目(5100-202155426A-0-0-00)。

摘　　要：为解决电力系统人工智能应用中样本数量不足的问题,对时序数据生成方法进行研究,提出一种CDoubleGAN模型。结合编解码器和两对生成器-鉴别器,采用ARFNN替代RNN解决Lipschitz连续性问题,实现使用Wasserstein距离对目标函数的稳定优化。将数据类别标签融入模型中,生成特定类别的样本。在IEEE-39系统的实验结果表明,CDoubleGAN在类别生成上的准确度超过98%,与TimeGAN相比,生成的数据与原数据具有更高的相似度,更好保留了数据原始特性以应用于数据生产。To address the issue of insufficient sample sizes in artificial intelligence applications for power systems,time-series data generation methods were explored and the CDoubleGAN model was introduced.An encoder-decoder with two pairs of generators-discriminators was combined.ARFNN was employed instead of RNN to resolve the Lipschitz continuity problem,facilitating stable optimization of the objective function using the Wasserstein distance.Data category labels were integrated into the model to generate specific category samples.Experimental results on the IEEE-39 system indicate that CDoubleGAN achieves over 98%accuracy in category generation.Compared to the TimeGAN baseline,the data generated by CDoubleGAN more closely resemble the original data,better preserving the original characteristics of the data for use in data production.

关 键 词：人工智能 深度学习 电力系统 暂态稳定 数据生成 编解码器 生成对抗网络 时序数据 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]