Research On Non-intrusive Load Disaggregation Method With The On-off Recognition

With the continuous improvement of smart grid requirements for fine load-side management,residential user portrait,user-side energy conservation,public safety and other fields have also attracted attention.Load disaggregation is to obtain the information of power consumption and service time of each electrical equipment in the user’s house through sensors reading or inference algorithms,so as to provide data support for the establishment of demand response.The non-intrusive load disaggregation method can obtain the energy consumption data of the house appliances by analyzing the total electricity volume of the smart meter.Because it does not need to install any sensor in the house,it has the advantages of low cost investment,fast expansion and deployment,and so on,it has been paid more and more attention by the industry and academia.At present,scholars have realized non-intrusive load disaggregation through a variety of different algorithms,but the proposed algorithms have the problem of low accuracy of identification of the ON-OFF state of appliances and low accuracy of load disaggregation due to the defects of the algorithm itself.For this reason,an improved non-intrusive load disaggregation method based on the ON-OFF state identification of appliance is proposed to improve the ON-OFF state identification accuracy and improve the algorithm.The specific work is as follows:First of all,in order to solve the problem of low accuracy of load disaggregation caused by high error rate of unbalanced the ON-OFF state identification,this paper proposes a model to identify the ON-OFF state of appliances.The model is composed of the ON-OFF state recognition module and the ON-OFF state combination module.By using the synthetic minority oversampling technique(SMOTE)to resampling the training data,deep neural network(DNN)model to extract the ON-OFF state features,bidirectional long short term memory-conditional random field(Bi LSTM-CRF)model improved the ability to recognize unbalanced the ON-OFF state.Secondly,to solve the problem that the non-intrusive load disaggregation model based on HMM has too strong assumption conditions(i.e.non-aftereffect property of Markov chain and parameter invariance assumption),which leads to low accuracy of load disaggregation,the paper proposes a non-intrusive load disaggregation method based on the ON-OFF state recognition with improved HMM.The improved method is firstly divided into several sub-datasets of the ON-OFF state according to the results of the ON-OFF state recognition,and then load disaggregation is carried out for the devices in the ON state,which effectively improves the defects of ignoring the temporal characteristics and single calculation model parameters caused by the hypothesis of HMM.At last the experimental results show that the method can improve the accuracy of load disaggregation.Finally,in view of the deficiency of deep learning method insufficient feature extraction of electricity data,this paper proposes to use the self-attention mechanism to select information with high relevance to the current moment from the preprocessing features,highlight the weight of highly relevant data to realize the focus on the feature sequence,and then improve the accuracy of load disaggregation.Meanwhile,in order to further verify the generalization ability of the ON-OFF state recognition model,different partitioning methods of the ON-OFF state sub-datasets are proposed,and the validity of the proposed methods is verified by experiments in the load disaggregation model of one-dimensional convolutional neural network,and the models suitable for different the ON-OFF state recognition methods are obtained.