电器与能效管理技术 ›› 2022, Vol. 0 ›› Issue (3): 23-29.doi: 10.16628/j.cnki.2095-8188.2022.03.004

• 研究与分析 • 上一篇    下一篇

基于改进生成对抗网络的非侵入式负荷预测样本不均衡的改善方法

裘星1, 尹仕红1, 张之涵1, 谢智伟1, 江敏丰1, 郭兴林1, 杨建明1, 郑建勇2   

  1. 1.深圳供电局有限公司, 广东 深圳 440304
    2.东南大学 电气工程学院, 江苏 南京 210096
  • 收稿日期:2021-12-21 出版日期:2022-03-30 发布日期:2022-04-28
  • 作者简介:裘星(1990—),男,工程师,主要从事电能计量技术检测研究工作。|尹仕红(1978—),女,高级工程师,主要从事电能计量技术检测研究工作。|张之涵(1991—),男,工程师,主要从事电能计量技术检测研究工作。
  • 基金资助:
    深圳供电局有限公司科技项目(090000KK52190185)

Improving Method of Non-Intrusive Load Forecasting Sample Imbalance Based on Improved Generative Confrontation Network

QIU Xin1, YIN Shihong1, ZHANG Zhihan1, XIE Zhiwei1, JIANG Minfeng1, GUO Xinglin1, YANG Jianming1, ZHENG Jianyong2   

  1. 1. State Grid Shenzhen Electric Power Company, Shenzhen 440304, China
    2. Southeast University School of Electrical Engineering, Nanjing 210096, China
  • Received:2021-12-21 Online:2022-03-30 Published:2022-04-28

摘要:

采用神经网络对负荷进行识别训练时,由于不常用负荷的原始样本数量较少而常用负荷样本数量较多,从而导致训练的样本分布不均衡。传统的生成对抗网络(GAN)和采用辅助分类器的生成对抗网络(ACGAN)在对样本进行扩充训练时,原始的模型会容易产生梯度消失等问题。在原始ACGAN的模型基础上,在其判别真假损失函数上添加梯度惩罚函数,使得判别器和生成器在相互优化过程中梯度下降更快,更容易达到纳什均衡,从而使原始样本的不均衡情况得到有效改善。通过公共数据集PLAID进行测试分析,并验证了所提方法的有效性。

关键词: 生成对抗网络, 辅助分类器, 负荷识别, 样本不均衡, 样本扩充

Abstract:

When the neural network is used to identify and train the load,because the number of original samples of uncommon load is small and the number of common load samples is large,the distribution of training samples is unbalanced.Using the traditional generative adversarial networks (GAN) and the auxiliary classifier GAN (ACGAN),the original model is prone to the problems such as gradient disappearance when the sample is expanded and trained.Based on the original ACGAN model,the gradient penalty function is added to the loss function for distinguishing true and false,so that the gradient of the discriminator and generator will fall faster during the mutual optimization process,and it is easier to achieve Nash equilibrium.The imbalance of the original sample has been effectively improved.Through the test and analysis of the public data set PLAID,the effectiveness of the proposed method is verified.

Key words: generative adversarial networks (GAN), auxiliary classifier, load identification, sample imbalance, sample expansion

中图分类号: