Research On NILM Based On Multilevel Evidence Combination Of Base Classifiers Using Low-sampling Rate Data

Electricity has been widely used since the Second Industrial Revolution.After entering the 21^st century,with the development of Internet Technology,Smart Grid has emerged as the times require.As one of the important components of the Smart Grid,the Advance Metering Infrastructure?AMI?is the basis for energy management and improving energy efficiency.Refining the load monitoring to the electrical level is one of the important contents of AMI.For residential users,Non-Intrusive Load Monitoring?NILM?can be used.NILM can get the power consumption of main individual appliances by analyzing the voltage,current,power and other aggregated electricity information.Applying NILM can help to save costs and reduce interference to users.In this paper,we present an NILM method for residential environment.Under the condition that we only have low-sampling rate?1/60 Hz?data,the working state of the individual appliances is identified by mining the artificial features and hidden features of the raw data,and these identification results are combined by Dempster-Shafer Theory?DST?.Finally,we use Integrated Neural Network?INN?to estimate power of main individual appliances.To identify the working state of individual appliances,traditional methods usually use high-sampling rate voltage,current,harmonic wave and other aggregated electricity information,and are based on detected artificial event features or raw data.In this paper,we only have low-sampling rate data.To make the most of these valid information,we propose Fuzzy Decision Tree and Long Short-Term Memory?LSTM?creatively from the perspective of the artificial features and raw data respectively.Because the two classifiers have their own advantages in dealing with different appliances.In order to get a better load identification performance,we innovatively apply DST to solve NILM problem.The classifying results of each classifier are combined from individual appliances,classifying methods,and time series three levels,and the combined results achieve better performance than single classifier for most individual appliances.To estimate the power after getting the working state of the individual appliances,in this paper,we propose and use Integrated Neural Network?INN?to learn the working mode of individual appliance.The individual appliance working state and aggregated power at each moment are input to the INN,and we generate more accurate,detailed power estimation results.In this paper,we mainly use AMPds public dataset.The experimental results show that the proposed method can achieve good performance on load identification,higher power estimation accuracy,and the power details of some individual appliances are restored successfully.