基于双分支特征集成LSTM的非侵入式负荷分解研究  

Research on Non-intrusive Load Disaggregation with Dual-branch Feature Integration LSTM

作　　者：韩小地 郭家兴 钱康 刘文瑞 HAN Xiaodi;GUO Jiaxing;QIAN Kang;LIU Wenrui(Guoneng Xinjiang Ganquanbao Integrated Energy Co.,Ltd.,Urumqi 830018,China;Jiangsu Keneng Electric Power Engineering Consulting Co.,Ltd.,Nanjing 210036,China;China Energy Engineering Group Jiangsu Electric Power Design Institute Co.,Ltd.,Nanjing 211102,China;School of and Electrical Power Engineering,Hohai University,Nanjing 210098,China)

机构地区：[1]国能新疆甘泉堡综合能源有限公司,新疆乌鲁木齐830018 [2]江苏科能电力工程咨询有限公司,江苏南京210036 [3]中国能源建设集团江苏省电力设计院有限公司,江苏南京211102 [4]河海大学电气与动力学院,江苏南京210098

出　　处：《山东电力技术》2024年第10期31-43,共13页Shandong Electric Power

基　　金：江苏省政府科技项目“可再生氢能制/储/管道掺混一体化场站成套设计与运行控制关键技术研发”(BE2022040)。

摘　　要：面向分钟级低频稳态负荷数据,提出了一种双分支特征集成长短记忆(longshort-term memory,LSTM)神经网络的非侵入式负荷分解(non-intrusive load disaggregation,NILD)方法。按照分钟级低频采样数据对NILD的影响,将待分解的用电设备分为2类:第一类是持续运行时间较长,分钟级低频采样数据能够有效表征状态变化的用电设备;第二类是持续运行时间短、采用分钟级低频采样数据存在特征淹没的用电设备。针对第一类设备,将事件检测与分钟级负荷稳态功率分别作为输入特征,构建特征集成LSTM模型进行负荷;针对第二类设备,计及低频采样数据的特征淹没,采用差分滤波对负荷功率数据进行前置处理,然后与事件检测结果一起作为特征集成LSTM模型的输入。采用两种数据集对所提方法进行性能评估,实例证明特征集成LSTM模型在利用低频稳态数据进行NILD时具有一定的优越性。Aiming at minute-level low-frequency steady-state load data,a dual-branch feature ensemble long short-term memory(LSTM)neural network method for non-intrusive load disaggregation(NILD)is proposed.According to the impact of minute-level low-frequency sampling data on NILD,the electrical devices to be disaggregated are divided into two categories.One is electrical devices that have a long continuous operation time and whose state changes can be effectively characterized by minute-level low-frequency sampling data.The second category is electrical devices that have a short continuous operation time and are characterized by feature submergence when using minute-level low-frequency sampling data.For the first category of devices,event detection and minute-level load steady-state power are used as input features to construct a feature ensemble LSTM model for load disaggregation.For the second category of devices,considering the feature submergence of low-frequency sampling data,differential filtering is used to preprocess the load power data,which is then used together with the results of event detection as the input of the feature ensemble LSTM model.Two datasets are used to evaluate the performance of the proposed method.The example proves that the feature ensemble LSTM model has certain advantages in using low-frequency steady-state data for NILD.

关 键 词：非侵入式负荷分解 低采样率 特征集成 事件检测 差分滤波 

分 类 号：TM714[电气工程—电力系统及自动化]