摘要
深度学习在窃电行为检测领域的研究中应用越来越多,但传统的基于神经网络的深度学习因需要大量的训练样本、调参过程复杂等原因应用十分受限。首次将深度森林分类算法引入窃电行为检测领域,利用其依赖训练样本量小、超参数少、计算效率高的优点,结合从电量、电压、电流、功率因数等数据提取的特征检测用户是否存在窃电嫌疑。通过某地区用电信息采集系统提供的负荷数据,验证了所提窃电行为检测模型的有效性。
Because of neural network based traditional deep learning methods requiring large numbers of training samples and complex parameter configuration,it is limited in electricity theft detection.This paper introduces deep forest classification algorithm into the electricity theft detection,and detects the electricity theft by extracting the features from the data about power consumption,voltage,current and power factor,based on the advantages of the method such as requiring less training samples,less hyper-parameters and higher computing efficiency.Finally,the effectiveness of detecting the electricity theft with the proposed method is proved by the load data obtained from power consumption information acquisition system in a region.
作者
杨学良
陶晓峰
熊霞
戚梦逸
孙萌
YANG Xueliang;TAO Xiaofeng;XIONG Xia;QI Mengyi;SUN Meng(NARI Technology Development(State Grid Electric Power Research Institue)Co.,Ltd.,Nanjing 211106,China)
出处
《智慧电力》
北大核心
2019年第10期85-92,共8页
Smart Power
基金
国家电网公司科技项目资助(521101180017)~~
关键词
窃电行为检测
深度森林
多粒度扫描
级联森林
特征增强
超参数调试
electricity theft detection
deep forest
multi-grained scanning
cascading forest
feature enhancing
hyper-parameter testing