摘要
针对现有窃电检测方法提取的用电特征有效性低,分类算法未注重样本类分布不平衡而导致窃电用户检出率不高的问题,提出一种堆叠稀疏自编码器与改进深度森林结合的窃电检测模型。堆叠稀疏自编码器用于从原始用电数据中提取高度抽象潜在的特征,深度森林算法对所得到的特征进行分类学习。引入Hellinger distance作为深度森林决策树的分裂指标以改进样本类别不平衡问题。实例分析表明,在DR值上所提模型与Deep Forest、RF和ANN相比,分别提高12.96%、13.68%和17.7%,有效提高了少数类窃电用户的检出率。
The effectiveness of electricity consumption features extracted by existing electricity theft detection methods is low. The classification algorithm does not pay attention to the imbalance of sample class distribution, resulting in low detection rate of electricity theft users. Aimed at this problem, a detection model based on hybrid stacked sparse autoencoder and improved deep forest is proposed. Stack sparse autoencoders were used to extract highly abstract potential features from the raw electricity consumption data, and deep forest algorithm was used to classify the obtained features. In order to solve the problem of unbalanced sample, the Hellinger distance was introduced as the splitting index of decision tree in deep forest. Case studies demonstrate that compared with the deep forest, RF and ANN classification models, the detection rate score of the proposed model is increased by 12.96%, 13.68% and 17.7%, which can improve detection rate for theft users.
作者
王耀聃
李红娇
詹清钦
Wang Yaodan;Li Hongjiao;Zhan Qingqin(College of Computer Science and Technology,Shanghai University of Electric Power,Shanghai 200090,China)
出处
《计算机应用与软件》
北大核心
2022年第12期64-72,158,共10页
Computer Applications and Software
基金
国家自然科学基金项目(61403247,61702321)
上海市信息安全综合管理技术研究重点实验室开放课题(AGK2015005)。
关键词
窃电检测
深度森林
稀疏自编码器
不平衡样本
深度学习
Electricity theft detection
Deep forest
Sparse autoencoder
Unbalanced sample
Deep learning
