文章摘要
燕续峰,翟少鹏,王治华,等.深度神经网络在非侵入式负荷分解中的应用[J].电力系统自动化,2019,43(1):126-132. DOI: 10.7500/AEPS20180629004.
YAN Xufeng,ZHAI Shaopeng,WANG Zhihua, et al.Application of Deep Neural Network in Non-intrusive Load Disaggregation[J].Automation of Electric Power Systems,2019,43(1):126-132. DOI: 10.7500/AEPS20180629004.
深度神经网络在非侵入式负荷分解中的应用
Application of Deep Neural Network in Non-intrusive Load Disaggregation
DOI:10.7500/AEPS20180629004
关键词: 非侵入式负荷监测  电器状态聚类  时间特性模型  深度神经网络
KeyWords: non-intrusive load monitoring  appliance state clustering  time characteristic model  deep neural network
上网日期:2018-11-20
基金项目:国家自然科学基金资助项目(51877134)
作者单位E-mail
燕续峰 上海交通大学电子信息与电气工程学院, 上海市 200240  
翟少鹏 上海交通大学电子信息与电气工程学院, 上海市 200240  
王治华 国网上海市电力公司电力调度控制中心, 上海市 200122  
王芬 上海交通大学电子信息与电气工程学院, 上海市 200240  
何光宇 上海交通大学电子信息与电气工程学院, 上海市 200240 hhhxxjj@163.com 
摘要:
      负荷监测是智能用电的一个重要环节,为了实现非侵入式负荷监测,提出了一种基于深度神经网络的非侵入式负荷分解方法。首先提出了改进的电器状态聚类算法,通过改进终止条件和增加消除冗余类判据使得聚类结果更符合电器实际运行情况。针对目前研究常用的隐马尔可夫模型的弱时间特性问题,提出了电器时间特性模型,综合考虑了电器运行特性和用户使用习惯,从时间角度对电器进行建模。构建了深度神经网络进行负荷分解,网络的输入综合考虑了电器状态及时间、功率信息,采用历史运行数据及时间特性模型生成数据训练网络参数。最后,在测试数据集上验证了方法的有效性和准确性。
Abstract:
      Load monitoring is an important part of intelligent electricity consumption. For the non-intrusive load monitoring, a deep neural network based non-intrusive load disaggregation method is proposed. Firstly, a modified iterative appliance state clustering algorithm is proposed. By modifying the stopping criteria and adding eliminating criteria of redundant clusters, the clustering results are more consistent with the actual appliance operation. An appliance time characteristic model is proposed considering weak time characteristics of hidden Markov models which are commonly used in the current study. The appliance characteristics and user habits are taken into consideration. The electrical appliances are modeled from the perspective of time. A deep neural network is constructed to perform load disaggregation. The input of the network includes appliance states, power and time information. The history data and the generated data based on models are used to train the network parameters. The effectiveness and accuracy of the method are verified on the data set.
查看全文(Free!)   查看附录   查看/发表评论  下载PDF阅读器