| In the analysis of power load categories and consumption data,power companies can predict users’ power usage,carry out intelligent distribution,set dynamic electricity price to affect users’ power consumption behavior,and ultimately improve the stability of power grid.The traditional way to get the type and data of power load is to install data sensors on each load,which will result in a large number of sensors and then lead to high cost,heavy workload,difficult maintenance and other shortcomings.In order to overcome the above shortages,some scholars proposed a non-invasive method to detect and identify the type of domestic power load and its power consumption data via analyzing the electrical data on the home power supply main line.In this dissertation,a new event point detection algorithm is proposed,and a new multi classification machine learning algorithm is applied to non-invasive load detection and recognition.First of all,a laboratory data sampled platform is established to obtain the electrical equipment data from the home fire wire sensor and non-invasive power load identification simulation is achieved.Through the analysis,it is found that the online data REDD is consistent with the actual measured data,and its data volume is large and with credibility.This dissertation finally chooses the online data set REDD.Through the comparative analysis of 10 kinds of load being on and off to get electrical data in the data set,referring to other methods proposed by scholars in this field,this dissertation proposes a simplified BIC algorithm based on sliding window to identify the opening and closing points of load,and the accuracy of the algorithm can achieve over 97%.Secondly,through further analysis of REDD data,27 different steady states of the metioned10 kinds loads are identified,and 282 groups of steady state process data are extracted as samples.Through the observation and comparison of the steady-state data,combined with the relevant theories in the field of signal processing,this dissertation applies wavelet packet decomposition to the data feature extraction of load identification.Combined with the concept of information entropy,the time-domain,frequency-domain and the combination of time-domain and frequency-domain features of the load steady-state data are extracted from the wavelet packet decomposition coefficient.In addition,some basic time-domain features of the original signal are added.Finally,35 features for the load steady-state process identification are determined.Then combined with correlation coefficient analysis and principal component analysis,35 kinds of features are reduced to 13 kinds,and the characteristic database of load is established.Finally,according to the number of classification objects and features,an enhanced random forest classification method is selected.In order to achieve the comparison with the traditional random forest,282 sample data sets are used in the algorithm training and testing respectively according to the 7:3 ratio of training set and test set.At the same time,through the prediction results of the test set,two super parameters of the algorithm are optimized experimentally,and the highest accuracy of the algorithm on the test set is 96%.Finally,all algorithms are integrated,and the identification effect of the algorithm is verified by experiments on the unused data of REDD and the case of single load and multiple loads mixing.The accuracy of load identification is about 89%.One of the innovations of this paper is that through the simplification and combination of sliding window and algorithm,a simplified algorithm based on sliding window is proposed,which can gather the advantages of two mentioned algorithms.The second one is to use wavelet packet decomposition to mine more detailed and deep features of load steady-state data in time domain,frequency domain and time-frequency domain.The third one is to apply the enhanced random forest algorithm to load identification,and get better recognition effect and faster recognition speed. |
