Research On Deep Learning Based On Non-intrusive Load Monitoring

Non-intrusive load monitoring is a kind of technology that predicts the signal or state of each electrical appliance in the house according to the aggregate signal of the residential power entrance.The former is called load disaggregation,and the latter is called load identification.Due to the limitations of technical development,many literatures also continue the tradition of the early stage and call the non-intrusive load disaggregation or non-intrusive load identification as non-intrusive load monitoring.In this paper,with the help of deep learning technology,the problem is modeled as regression task and classification task from two different perspectives,and the two different modeling perspectives both achieve non-intrusive load monitoring in a real sense,that is,both non-intrusive load identification and non-intrusive load disaggregation are realized.The main work of this paper is as follows:1.NILM is modeled as a multi-label classification task,and the non-intrusive load monitoring methods based on U-NET and SegNet models were proposed,respectively.NILM can be considered a multi-label classification task since multiple appliances can simultaneously operate,where each appliance can be treated as an individual label.The multi-label classifier can directly predict(identify)the state of each electrical appliance,and then the device power consumption was obtained indirectly by multiplying the state with corresponding average power consumption.Obviously,it uses the average power consumption corresponding to the state to replace the real power consumption.The decomposition performance depends not only on the performance of the network model,but also on the detail level of the division of the state of electric appliance.Therefore,using NILM as a classification task pays more attention to obtaining accurate load identification performance,and can better provide service guarantee for smart home,residential security,and other fields.U-Net and SegNet are introduced from image segmentation field.Based on the characteristic that each appliance can run at the same time in non-intrusive load monitoring,the Binary Cross Entropy is used as the loss function.By comparing with existing methods,the experiment proves that the multi-label classification based on U-Net or SegNet can also achieve good performance in non-intrusive load monitoring.2.NILM is modeled as a regression task,and a non-intrusive load monitoring method based on bidirectional temporal convolution is proposed.Treating NILM as a regression task,it can directly predict(decompose)the load signal of the target electrical appliance,and then compare with the target threshold value of the electrical appliance,so as to indirectly obtain the state of the electrical appliance.Therefore,this method focuses more on obtaining accurate load decomposition performance and providing more accurate load decomposition information for smart grid and power users.Firstly,the network structures in deep learning,such as standard causal convolution,causal dilated convolution,bidirectional dilated convolution,residual connection and TCN residual block,are briefly introduced.Then,the Bi-TCN residual block based on TCN is derived.It explains why Bi-TCN has larger receptive field and better performance than TCN.The Bi-TCN residual blocks were stacked,and then the different load characteristics of different layers of Bi-TCN residual blocks were combined by using the residues connection.Compared with existing methods,experimental results show that the proposed method achieves better performance.