Study On The Identification And Disaggregation Of Residential Load

First of all,this paper measured the electrical data of different combined loads,and analyzed the performance of current,power and harmonic feature in different loads one by one.Then,data preprocessing and feature selection are carried out for the collected data set.Through Pearson correlation coefficient,random forest importance evaluation,principal component analysis and wavelet de-noising,invalid sample feature variables are filtered out and more suitable feature variables are screened out.Secondly,in view of the problems existing in current residential load identification,this paper proposes a non-invasive load identification method based on multi-label random forest.Firstly,multi-label technology is used to adapt the traditional random forest,and the multi-label cart decision tree is constructed to form the multi-label random forest,then the multi-label random forest is optimized and adjusted to make the model reach a better state.The algorithm supports multi-label data set,which can deal with the load identification problem in complex situation,and further expand the scope of application of traditional tree algorithm.Simulation test results show that the method is fast and efficient,with high recognition accuracy,strong generalization ability,superior anti noise performance and low training cost.Compared with single label random forest and other algorithms,it has great advantages,which is helpful for further research,application and promotion of non-invasive residential load identification.Thirdly,this paper also applies deep forest which is a new algorithm to the load identification for the first time,and obtains a good recognition effect.The performance of deep forest is comparable to that of deep neural network,but the training cost is lower.Moreover,compared with the depth neural network,it needs less super parameters.Even under the default parameters,it can also show better results.Finally,this paper also studies the load disaggregation method.Hidden Markov model is a classical model of load disaggregation,which has many extensions.This paper presents an implementation of the condition factor hidden semi Markov model in load disaggregation.This paper used three kinds of intelligent optimization algorithms: genetic algorithm,discrete binary particle swarm optimization and simulated annealing algorithm.In the aspect of deep neural network,this paper alsooptimizes the structure of LSTM network by adding convolution layer in two bidirectional LSTM layers,and gives the realization of GRU which is another RNN network variant.Comparing the two networks,it is found that the disaggregation performance of LSTM is better than that of GRU,but the training cost of GRU is much lower than that of LSTM.At last,we compare the effect of de-noising auto encoder and two kinds of recurrent neural network,and find that the de-noising auto encoder is better than LSTM and GRU network in both disaggregation performance and training cost.The paper includes 51 figures,10 tables and 118 references.