Research On Non-intrusive Residential Load Disaggregation Method Based On Deep Learning

Smart grid is an important strategic direction of the second smart grid.In order to meet the technical requirements of smart grid for advanced measurement system,efficient and economic load monitoring has become an urgent technical problem.Non-intrusive load disaggregation can monitor the working state of power load within the whole user only by sampling and analyzing the total load data at the power entrance of the user,which avoids the sensor cost and user privacy problems in intrusive monitoring.Therefore,the field of load monitoring is mainly studied in this aspect.The existing non-intrusive load disaggregation methods mainly establish load disaggregation model by manually extracting load characteristics,which have many problems such as strong feature dependence and large disaggregation error,and can not meet the demand of real-time and efficient load monitoring.Deep learning is the most effective technology in various fields of intelligent technology.It has strong feature learning ability and has been formally applied in picture editing,pronunciation discrimination and other aspects,but it is not widely used in the field of load disaggregation.Therefore,based on deep learning method,this dissertation studies load disaggregation from two objectives of state disaggregation and power disaggregation.First of all,this dissertation selects the classic deep learning model of load disaggregation-long and short-term memory network.In order to better learn the internal characteristics of the total load sequence and judge the current load state,this dissertation introduces the Attention mechanism and bidirectional network to establish three modified models.REDD data set,which is commonly used in the field of load disaggregation,is selected for training and testing,and six evaluation indexes,such as the recall and precision,are used to comprehensively assess the disaggregation results of the modified model.The conclusion of this experiment indicates that the use of Attention mechanism and bidirectional network can optimize the disaggregation results of the original LSTM network.On the whole,the modified model can effectively disaggregate the load state,while the disaggregation results of the load power is relatively general.In view of its short training time and high application efficiency,it can be mainly applied to the scene that only needs to monitor the load state.Secondly,in view of the above problems,this dissertation introduces a new probability generation model in the field of image processing for the first time,which is called generation confrontation network,and proposes a load disaggregation model based on self attention generation confrontation network.Through the confrontation training between the generator and the discriminator,the generator generates the load sequence that the discriminator can’t distinguish the authenticity of it,so as to realize load disaggregation.The consequence is that the accuracy and accuracy of the model disaggregation are improved by more than 10%,and the absolute error of draw is reduced to about 50%.Based on the effective disaggregation of load state,the proposed method focuses on the disaggregation of load power.Finally,based on the research of load disaggregation,this dissertation deeply excavates the information of electrical appliance level power consumption obtained,and expounds five application scenarios from the perspectives of residents and power grid,including user abnormal behavior monitoring electrical appliance operation status monitoring and demand response potential analysis,which provides new ideas for user energy consumption management and demand side management of power grid.This dissertation has 39 pictures,19 tables and 83 references.