摘要
建筑与工业、交通并列成为我国三大"耗能大户",公共楼宇能耗巨大但节能潜力不可估量,监测其负荷特征挖掘节能潜力应用前景广阔。非侵入式负荷监测与分解作为大数据环境下面向智能电网配电侧的一种高级应用,可通过电力端口信息挖掘用户用电行为,但传统算法主要针对家庭用户,且存在功率跟踪性差,训练时间长等问题。为此,文章面向楼宇用户提出一种基于深度学习和迁移学习的负荷分解模型。该模型将3种神经网络:多层感知器神经网络(multi-layer perceptron neural networks,MLP)、卷积神经网络(convolutional neural network,CNN)、长短时记忆网络(long short-term memory,LSTM)模块化并联连接,并融合网络特征,通过再学习模块重新学习融合特征与结果的映射关系;针对新楼宇数据量不足的问题,将特征学习网络模块特征冻结,运用迁移学习重新训练网络,在确保模型精度的同时降低深度学习所需数据量和训练时间。最后利用真实楼宇负荷数据划分出3种应用场景,利用所提模型开展分解应用并与3种传统深度学习算法分解结果作对比,结果表明:基于深度学习和迁移学习的楼宇负荷分解模型准确率高,泛化能力强,可快速有效地实现楼宇负荷分解。
Construction, industry and transportation are considered as the top three "energy consumers" in China.Public buildings consume huge amounts of energy but with immeasurable energy saving potential, so monitoring their load characteristics to explore their energy saving potential is a promising application. The non-intrusive load monitoring and decomposition, as an advanced application for the distribution side of smart grids in the big data environment, can tap into the users’ electricity consumption behaviors through the power port information. However, the traditional algorithms, mainly targeted at the home users, have the problems as poor power tracking, long training time, and the like. To this end, this paper proposes a load decomposition model based on the deep learning and transfer learning for the building users. The model connects the three neural networks(MLP, CNN and LSTM) in parallel in a modular way, fuses their network features, and re-learns the mapping relationship between the fused features and the results through a re-learning module. For the new buildings with insufficient data, by freezing the features of the feature learning network module, the transfer learning is applied to retrain the network, which reduces the amount of data required for the deep learning and decreases the training time. This ensures the accuracy of the model while reducing the amount of data and training time required for deep learning.Finally, the proposed model is used to decompose the three application scenarios classified with the actual building load data, and the results are compared with those of the three traditional deep learning algorithms. The comparison results show that the building load decomposition model based on the deep learning and transfer learning has a higher accuracy rate and stronger generalization ability, which can quickly and effectively achieve the building load decomposition.
作者
杨秀
吴吉海
孙改平
田英杰
王皓靖
李安
YANG Xiu;WU Jihai;SUN Gaiping;TIAN Yingjie;WANG Haojing;LI An(School of Electric Power Engineering,Shanghai University of Electric Power,Yangpu District,Shanghai 200090,China;State Grid Shanghai Electric Power Research Institute,Hongkou District,Shanghai 200080,China)
出处
《电网技术》
EI
CSCD
北大核心
2022年第3期1160-1168,共9页
Power System Technology
关键词
非侵入式负荷分解
深度学习网络
迁移学习
公共楼宇
non-intrusive load monitoring and decomposition
deep learning network
transfer learning
public buildings
