Non-intrusive Load Disaggregration Based On Deep Learning

Non-intrusive load disaggregration is to measure the total power data of household meters to estimate the power consumption of individual appliances in the household.The disaggregration results are helpful for real-time understanding of the current situation of power consumption,which may help users reduce energy consumption,and help grid companies manage grid distribution,identify faulty electrical appliances,and can investigate the use of electrical appliances,which is conducive to more scientific planning of electricity consumption,improve electricity efficiency and achieve the purpose of saving electricity,saving energy and reducing consumption.The existing non-intrusive load disaggregration is mainly based on the traditional load characteristics to establish the power load disaggregration model,but the traditional feature extraction method can not achieve high-precision real-time disaggregration,can not meet the needs of the current market grid smart development.At present,there are many non-intrusive load disaggregration methods based on deep learning,but the current deep learning has low accuracy in non-intrusive load disaggregration applications,easy to appear gradient disappearance,and large disaggregration error for low frequency electrical appliances.In view of the above problems,this thesis proposes two improved deep learning algorithms for non-intrusive load disaggregration.The main work of this thesis is as follows:1.Design a deep gate recurrent residual networks by combining a deep gru module and a resnet module.The deep gru module can better capture the dependence of time step distance in the time series when the network becomes deeper,so as to better identify the feature data.By integrating the network into the residual network,we can use the deep network to extract the deep load characteristics,and effectively reduce the calculation parameters to speed up the network training.The experimental results have been further improved.2.Aiming at the problems of low accuracy,easy to appear gradient disappearing and large disaggregration error of low frequency electrical appliances in the current application of deep learning in non-intrusive load disaggregration,a non-intrusive load disaggregration model based on group dilated residual network is proposed.In this model,we use dilated convolution to increase the receptive field,capture more data,and solve the problem that long time series data is difficult to learn.At the same time,we use sequence to sequence disaggregration method to ensure the efficiency of the algorithm and reduce the disaggregration error.