摘要
日益增长的用户侧用电数据为基于数据驱动的窃电检测方法奠定了基础,然而窃电检测数据固有的不平衡性质会影响该类方法的性能。针对窃电检测的数据不平衡问题,提出一种基于Wasserstein生成对抗网络(wasserstein generative adversarialnetwork,WGAN)的窃电样本过采样方法,通过WGAN生成器与判别器的对抗训练,神经网络能够学习到窃电负荷序列难以显式建模的时间相关性,生成与真实窃电样本具有相近分布的合成样本。采用WGAN训练期间生成的多组窃电样本对原始窃电样本进行增强,使用卷积神经网络在多组增强训练集上进行训练,选择在验证集上取得最高AUC值的增强训练集,最后在其上训练分类器实现窃电检测。所提方法在某电力公司提供的真实数据上进行了实验测试,结果表明相比于随机过采样、SMOTE和ADASYN等方法在检出率、误检率、F1测度以及AUC评价指标上有明显提升。
Increasing electricity data lay the foundation for electricity theft detection methods based on data-driving,but the inherent imbalance nature of data affects the performance of such methods.Aiming at the imbalance problem of electricity theft detection,this paper proposed an electricity theft sample oversampling method based on Wasserstein generative adversarial network(WGAN).Through adversarial training of WGAN generator and discriminator,neural network can learn the temporal correlation of electricity theft load sequence which is difficult to explicitly model and generate synthetic samples with distributions similar to real samples.The original samples are enhanced by multiple sets of samples generated during WGAN training.Convolutional neural network is used to train on multiple enhanced training sets,and the enhanced training set with the highest AUC value on verification sets is selected.Finally,detection is implemented on the selected set.The method is tested on the actual data provided by a power company.Result showed that compared with the methods of random oversampling,SMOTE and ADASYN,the detection rate,false detection rate,F1 measurement and AUC of the proposed method are significantly improved.
作者
王德文
杨凯华
WANG Dewen;YANG Kaihua(School of Control and Computer Engineering,North China Electric Power University,Baoding 071003,Hebei Province,China)
出处
《电网技术》
EI
CSCD
北大核心
2020年第2期775-782,共8页
Power System Technology
关键词
数据驱动
窃电检测
数据不平衡
生成对抗网络
过采样
data-driving
electricity theft detection
data imbalance
generative adversarial network
oversampling
