Deep Learning Based Non-invasive Load Disaggregation Technology

Under the pressing of ”carbon neutrality”,energy conservation and efficiency enhancement have become an urgent issue for the power system.Non-intrusive load disaggregation monitors the appliances inside the household non-intrusively by collecting and processing the load data sampled by the smart meter installed at the house incoming line.On the one hand,load disaggregation enables users to understand the household energy consumption,helps users reduce electricity cost? on the other hand,it collects and analyzes demand-side load data for the power grid,which is conducive to the demand-side response implementation.However,the existing non-intrusive load disaggregation method is not accurate enough.This thesis deploys deep learning technology to improve the effect of non-intrusive load disaggregation,the details as follows:(1)The problem of non-intrusive load disaggregation is studied and the state characteristics of the load are analyzed.The performance evaluation index of non-intrusive load disaggregation is deduced and summarized.The deep learning principles and models are studied,the activation,loss,and optimization functions of the neural network are deduced.Several deep learning models which suit the non-intrusive load disaggregation are studied for the following research.(2)A time-power hybrid load activation extraction algorithm is proposed.The existing load activation extraction algorithm has low extraction accuracy,incomplete and impure extraction of load activation problems,a time-power hybrid load activation extraction algorithm is designed.The algorithm improves the performance of the load activation extraction algorithm by establishing the correlation between time and load power.Simulations verify the effectiveness of the load activation extraction algorithm.(3)The convolutional long short-term memory load disaggregation model and the convolutional gated recurrent load disaggregation model are proposed.The time series convolutional network is applied to the non-intrusive load disaggregation problem and constructed the time series convolutional network load disaggregation model.To further improve the load characteristic extraction and representation capabilities of the model,based on the time series convolutional network model,the time series convolutional encoder load disaggregation model is innovatively designed.(4)The proposed load disaggregation models are tested on the REDD public data set.Two load disaggregation scenarios are designed,which are load state disaggregation and load power disaggregation.The results show that for the load state disaggregation scenario,the improved convolutional gated recurrent load disaggregation model proposed in this thesis has both accuracy and generalization performance,and can achieve the goal of load state disaggregation.Compared with the existing models,all indicators are better.Promote.Aiming at the task of load power disaggregation,the temporal convolutional encoder load disaggregation model proposed in this paper has extremely strong load information extraction and representation capabilities,which is greatly improved compared to other models.