Research On Non-intrusive Load Monitoring System Based On Deep Learning

With the rapid development of the smart grid,by using the non-intrusive load monitoring system（NILM）to collect various electricity consumption information of users,to further optimize and improve the smart power supply control system of the grid,so as to achieve smarter,more efficient and higher quality Power grid.Currently in NILM,the system function is mainly divided into two parts,namely the identification of electric equipment and short-term load forecasting.A deep learning algorithm is introduced in the power equipment identification part,which can deeply dig the relationship between the power consumption data of various equipment.When the amount of training data is sufficient,a very good recognition accuracy can be obtained.Short-term load forecasting is related to many factors such as time,climate and environment.The introduction of machine learning forecasting methods can well change the defects of traditional algorithms with large forecast errors and make load forecasting more accurate and practical.This article mainly carried out two parts of the research on NILM.In the load classification and identification part,firstly,the load characteristics are compared and selected.After demonstration,the active power is selected as the load identification characteristics,and the REDD data set is selected as the experimental verification data.In the recognition algorithm,the Deep Belief Network（DBN）algorithm in deep learning is introduced,and on this basis,a DBN network algorithm optimized based on the Sparrow search algorithm（SSA）is proposed to carry out the load.Classification recognition.The SSA algorithm is used to complete the assignment of initial weights and optimize the parameters of each layer of RBM,so that the DBN network can obtain a better training direction and further improve the accuracy.Then use KNN,SVM,BP neural network and DBN and SSA-DBN to conduct comparative experiments.The experimental results verify the accuracy and practicability of the DBN algorithm in power load identification,and further reflect the SSA optimized DBN algorithm identification The effect is improved significantly,up to about 5%.In the part of short-term load forecasting,in order to solve the problems of traditional algorithms in load forecasting with many limitations and low accuracy,a load forecasting method based on Long Short-Term Memory（LSTM）is proposed,and SVM and BP neural are used.Two kinds of network algorithms and their accuracy comparison experiments.Finally,the average absolute value error（MAE）and root mean square error（RMSE）are used to evaluate the results of each algorithm,and the advantages of the LSTM algorithm over other algorithms in short-term load forecasting are summarized.