Cross-scene Real-time Non-intrusive Load Monitoring

Non-intrusive load monitoring uses a single sensor to effectively and accurately monitor the energy consumption of an individual appliance,which is of vital importance to the energy conservation and emission reduction issues of global concern.Besides,non-intrusive load monitoring can also help optimize energy management in a smart grid,improve household appliance design,and increase operational efficiency.Traditional load monitoring methods have the disadvantages of low resolution and poor real-time performance,especially in the smart home scene.The existing algorithm model can not meet the requirements of high accuracy and strong generalization ability.Deep learning performs well in related problems,especially for big data,which can effectively perform feature extraction to avoid complex and redundant data modeling.Based on efficient deep neural networks,this thesis proposes real-time non-intrusive cross-scene load monitoring methods.The method focuses on how to improve the generalization ability and robustness of the model in different power consumption scenarios.Since the appliance to be recognized is easily submerged in the background load under different scenarios,this thesis proposes a concatenate neural network that can resist the interference of background load noise.This method divides each sample into two parts: the background load part and the mixed load part.Then the features are extracted by a deep convolutional neural network.It can be found that the background load signal in the circuit model is almost stationary in a short time,thus the mixed load feature is converted into the target load feature through this result.Finally,the target load feature is recognized by a classifier.Experimental results show that this method can not only improve the effect of load type recognition,but also reduce the error of load energy disaggregation,and it performs well and robustly on household appliance datasets.At the same time,it is too difficult for traditional algorithm models to generalize in multi-mode appliances and cross-user scenarios.To solve the weakness,the thesis proposes a metric learning network for domain generalization.With the help of the idea of metric learning and domain generalization,an auxiliary similarity module is designed to measure the distance of data in different domains,so that different samples with the same label get a unified generalized expression in the feature space.The model can be readily applied to new scenes.Experimental results show that this method not only performs well under non-intrusive load monitoring tasks,but also rivals the stateof-the-art in the field of computer vision and performs well on multiple datasets.