基于设备特征多层优选和CNN⁃NLSTM模型的非侵入式负荷分解  

Non‑intrusive load disaggregation based on multiple optimization of appliance features and CNN‑NLSTM model

作　　者：王家驹[1] 王竣平 白泰[1] 张然 丁熠辉 杨林[2] 张姝 WANG Jiaju;WANG Junping;BAI Tai;ZHANG Ran;DING Yihui;YANG Lin;ZHANG Shu(Metering Center,State Grid Sichuan Electric Power Company,Chengdu 610065,China;College of Electrical Engineering,Sichuan University,Chengdu 610065,China)

机构地区：[1]国网四川省电力公司计量中心,四川成都610045 [2]四川大学电气工程学院,四川成都610065

出　　处：《电力科学与技术学报》2023年第1期146-153,共8页Journal of Electric Power Science And Technology

基　　金：国家电网四川电力公司科技项目(52199720003P)。

摘　　要：非侵入式负荷分解技术可以有效挖掘用户侧设备信息,是电网开展用户负荷互动响应的基础。针对目前非侵入式负荷分解模型适应性较差及准确率较低等问题,提出一种基于设备特征多层优选的非侵入式负荷分解模型。首先,针对设备运行特性设计自适应滑动数据窗,进而获取到更加完整的设备功率片段,同时调整网络输入输出维度;其次,通过融合浅层卷积神经网络(CNN)与两层嵌套长短时记忆网络(NLSTM)提取并加深设备特征;然后,将其输入到改进的注意力机制中,通过调配特征权重,获得最优的设备特征序列;最后,在REDD数据集上进行实验分析,通过对设备特征多层选择、加深与复用在减小训练时间的同时,显著地提升负荷分解的准确率。Non‑intrusive load disaggregation technology can effectively mine the appliance information of customers,which is the basis to carry out interactive customer load response by the grid company.The conventional non‑intrusive load disaggregation technology has several drawbacks,such as limited scope of application and low accuracy.In this paper,a non‑intrusive load disaggregation model with multiple optimization selection of appliance characteristics is proposed.First,an adaptive sliding data window is designed for appliance operation characteristics to obtain a more complete power segment and to adjust the network input and output dimensions.Second,the appliance features can be extracted and deepened by fusing shallow convolutional neural networks(CNN)with two‑layer nested long and short‑term memory networks(NLSTM),which is further fed into an improved attention mechanism to obtain the optimum appliance feature sequence by adjusting the feature weights.Finally,experimental analysis on the REDD dataset shows that the multiple selection,deepening and reusing of appliance features can significantly improve the accuracy of load decomposition while reducing training time.

关 键 词：非侵入式负荷分解 自适应滑动窗 卷积神经网络 嵌套长短时记忆网络 改进注意力机制 

分 类 号：TM714[电气工程—电力系统及自动化]