Research On Non-intrusive Load Monitoring Of Household Appliances

The non-intrusive load monitoring technology can decompose and obtain feeder load data from home bus data,and provide data support for advanced services such as user demand response and user behavior analysis by using fine-grained electrical appliance level load information,so as to promote sustainable development of smart grid.On the one hand,the current non-intrusive load monitoring method fails to fully integrate the task of appliance load decomposition with the task of state identification,and does not conduct targeted research on the unbalanced distribution of the original data state categories,resulting in large errors in load decomposition.On the other hand,as the types of appliances in practical application scenarios become richer and richer,more demanding requirements are put forward for the generalization of the model.Therefore,this paper investigates the load decomposition algorithm in non-intrusive load monitoring technology,and the specific work includes the following three points:(1)Aiming at the problem of outliers in power decomposition caused by state identification errors of target appliances in the current single-task load decomposition algorithm,a non-invasive load decomposition network integrating state recognition is constructed,and the load decomposition errors were corrected by the state recognition results.At the same time,the feature extraction module is optimized,and the hybrid attention mechanism is introduced to distribute the weight of intermediate features from different dimensions,which improves the expression ability of the network.The state recognition sub-network is further improved by using AM-Softmax to improve the appliance on and off critical state recognition accuracy.The experimental results show that the network has good performance on load decomposition and state recognition tasks.(2)For the problem of unbalanced distribution of electrical appliance on and off state category samples in the original data,a data enhancement method based on electrical appliance associated operation probability and electrical appliance similarity was proposed.The method uses G-SMOTE to expand the number of samples in which each type of appliance is in the on state,eliminating the influence of the imbalance of appliance switching state data categories on the training effect of the model.At the same time,based on the associated operation probability of all electrical appliances and the similarity of the same type of electrical appliances in different households,the relevant appliance operation window data are randomly intercepted and added to the bus load to increase the diversity of sample feature information.The experimental results show that the introduction of data enhancement can effectively improve the state recognition accuracy of the appliances.(3)The load decomposition algorithm based on data sharing is proposed to address the demand for generalizability of non-intrusive load decomposition models in practical application scenarios.The algorithm introduces a federal learning framework to realize data sharing among users in each region and joint training of models,which increases the number of training samples and improves the generalizability of models.At the same time,the aggregation strategy of model parameters in the federal learning framework is optimized,and weights are assigned based on the balance of local data distribution and data size to reduce the impact of non-independent and homogeneous distribution of user load data in each region on the model performance.The experimental results show that the algorithm can be better adapted to practical application scenarios and has good generalization.This method effectively uses the state identification results to correct the load decomposition errors,and introduces a data enhancement method to reduce the impact of the severe imbalance of the electrical switch state data samples on the model performance;in addition,the model training based on the federal learning framework enhances the generalization of the model,which is beneficial to the model deployment in the smart grid.