基于生成对抗网络的分布式光伏窃电数据增强方法

A data augmentation method for distributed photovoltaic electricity theft using generative adversarial network

由于分布式光伏窃电的稽查难度大,致使相关部门收集的窃电样本数量有限,无法满足基于数据驱动的窃电检测需求。通过数据增强的方式,提出一种基于Wasserstein生成对抗网络(WGAN)的分布式光伏窃电样本数据增强方法。首先,WGAN通过生成网络与判别网络的对抗训练,能够学习到光伏窃电数据序列难以显式建模的时间相关性,可以生成与真实窃电样本具有相近分布的新的窃电样本;然后,根据典型的光伏窃电模型,针对窃电样本的数据特征选用卷积神经网络(CNN)进行窃电检测;最后,通过算例分析,对比不同数据增强方法与分类器,表明WGAN生成的窃电样本能够符合真实样本的波动规律和历史数据的概率分布特征,进而有效改善分类器的检测性能。

Due to the difficulty of the inspection of distributed photovoltaic(PV)electricity theft,the number of electricity theft samples collected by relevant departments is limited,which cannot meet the needs of data-driven electricity theft detection.This paper proposes a data augmentation method for distributed PV electricity theft using Wasserstein generative adversarial network(WGAN).First,WGAN can explicitly learn the time correlation that is difficult to model in the PV electricity theft data sequence.Furthermore,it can generate new electricity theft samples with similar distributions to the real ones through the confrontation training of the generator and discriminator networks.Then,according to the typical PV electricity theft model and data characteristics,the convolutional neural network(CNN)is selected for electricity theft detection.Finally,through the case analysis,it is shown that the electricity theft samples generated by WGAN can conform to the fluctuation law of authentic samples and the probability distribution characteristics of historical data,thereby effectively improving the detection performance.