基于有监督样本生成和残差网络的窃电检测方法  

Electricity theft detection method based on supervised sample generation and residual network

作　　者：景彪 苏盛[2] JIN Biao;SU Sheng(Yuxi Power Supply Bureau of Yunnan Power Grid Corp.,Ltd.,Yuxi 653199,China;Changsha University of Science and Technology Hunan Province Key Laboratory of Smart Grids Operation and Control,Changsha 410014,China)

机构地区：[1]云南电网有限责任公司玉溪供电局,云南玉溪653199 [2]长沙理工大学智能电网运行与控制湖南省重点实验室,湖南长沙410014

出　　处：《供用电》2024年第11期91-98,104,共9页Distribution & Utilization

基　　金：国家自然科学基金项目(51777015)。

摘　　要：针对窃电检测中因窃电样本不足导致模型识别精度低的问题,提出了一种基于有监督样本生成和残差网络的窃电检测方法。首先,基于依据生成对抗网络的结构原理,设计并实现了辅助分类生成式对抗网络(auxiliary classifier generative adversarial network,ACGAN),旨在精确捕捉并模拟真实窃电样本的固有特征分布。在ACGAN的判别器模块中,通过引入窃电分类作为辅助任务嵌入,从而实现了有监督学习机制的有效引入,从而提高了生成窃电样本的质量。采用Hinge损失函数来优化网络训练过程中的损失,显著提升了生成窃电样本的逼真度与实用性。其次,利用ACGAN网络的生成器来扩充数据集,生成额外的窃电样本。最后,在ACGAN判别器网络基础上,通过添加残差结构来构建窃电检测模型,以增强模型的对大量窃电样本的学习能力。同时,为了减少窃电检测模型的训练时间,采用ACGAN判别器网络的权重参数对窃电检测模型进行初始化。实验结果表明,所提出的ACGAN网络能够有效地学习真实样本的概率分布。与现有方法相比,有监督样本生成方法有效提高了窃电模型的识别精度,在正确率、检出率、综合性能指标(F1分数)等性能指标上均表现更优。In response to the problem of low model recognition accuracy caused by insufficient stealing samples in electricity theft detection,this paper proposes a method for electricity theft detection with supervised sample generation and residual network.Firstly,based on the structural principle of generative adversarial networks,an Auxiliary Classifier Generative Adversarial Network(ACGAN)is constructed to learn the distribution of real theft samples.Add a theft classification network to the discriminator of ACGAN as an auxiliary training task to improve the quality of generating theft samples.Then,the generator of the ACGAN network is used to generate electricity theft samples to expand the dataset.Finally,a residual structure is added to the ACGAN discriminator network to construct a theft detection model,improving the model’s learning ability for a large number of theft samples.In addition,to reduce the training time of the theft detection model,a theft detection model with similar structure is constructed on the basis of the ACGAN discriminator network,and the weight parameters of the ACGAN discriminator network are used to initialize the model.The experiment shows that the proposed ACGAN network can effectively learn the probability distribution of real samples.The comparison with existing methods shows that the supervised sample generation method proposed in this paper effectively improves the recognition accuracy of the theft model,and performs better in indicators such as accuracy,detection rate,F1.

关 键 词：窃电检测 残差网络 辅助分类生成对抗网络 数据增强 深度学习 

分 类 号：TM743[电气工程—电力系统及自动化]