Research On Non-intrusive Power Load Decomposition Based On Machine Learning

In recent years,the proportion of electric energy in energy consumption has increased constantly.Therefore,reasonable planning in electric energy use to avoid unnecessary energy waste has become an important issue for sustainable development of modern society.It has thus been considered to monitor household appliance loads and analyze the information of electricity use by home appliances in different times of day,thereby formulating electricity usage plans to achieve the goal of saving electric energy.In non-intrusive load decomposition,sensors are only installed at the user's electricity access point.Afterwards,the collected overall electricity use information can be analyzed to perform load identification and decomposition of electricity consumption behaviors,thereby obtaining electricity usage conditions for various loads.Compared to traditional intrusive load monitoring,this non-intrusive method boasts low cost and easy popularization,which is more suitable for residential load monitoring and of certain application value and significance.In this thesis,in order to further enhance the accuracy of non-intrusive load decomposition,studies were conducted from the perspectives of load status identification and power decomposition.On the basis of status code establishment,a non-intrusive load status identification algorithm based on stacking model fusion and a non-intrusive load decomposition algorithm integrating Attention and Conv Bi LSTM(IACBL)were proposed,respectively.The specific work is as follows:Firstly,relevant concepts on non-intrusive load decomposition were described with analysis of load characteristics,from which it was known that the power feature was simple to extract and had a wide range of applications.Therefore,power was selected as the load characteristic in this thesis.Aiming at the difficulty in labelling the operation status of electric equipment with a large amount of data during training,a medium filter was utilized to remove noise in the dataset.Also,a k-means algorithm introducing the silhouette coefficient and the sum of squared error as evaluation indicators was designed to determine the status number of loads and construct the status code.Secondly,to further improve the decomposition accuracy,the non-intrusive load status recognition algorithm based on stacking model fusion was proposed in this thesis from the perspective of load status identification.The stacking algorithm was consisted of two frames,the primary learner and the secondary learner.The first layer of primary learner was trained at first to obtain the prediction results,which were then used as the input in the second layer ofsecondary learner for training.Ultimately,the status of each load was recognized.The genetic algorithm was used to optimize the parameters of the included machine learning algorithms.Besides,two sets of experiments were designed that used only active power or both active and reactive power as load characteristics,and 6 loads in the public AMPds2 dataset were selected as the research object.The experimental results showed that both the load status identification accuracy and the F1 value of stacking model fusion were superior to those of the algorithm with a single model.Finally,a non-intrusive load decomposition model integrating Attention and Conv Bi LSTM(IACBL)was constructed from the perspective of power decomposition.First,the convolutional layer and the Bi LSTM algorithm were adopted for feature extraction.Afterwards,the weight of the extracted feature was optimized via the attention mechanism layer.Ultimately,the fully connected layer was used for classification,obtaining the status code and thus realizing power decomposition.In this chapter,by taking the active power as the load feature,6 loads from the public AMPds2 dataset were selected as the research object.In order to study the generalization performance of the model in different scenarios,another load was added to the previous experiment.The experimental results showed that the algorithm proposed in this chapter was better than other mainstream algorithms in terms of power decomposition rate.In addition,by adding the reactive power to the case with active power,the accuracy of power decomposition was improved to some extent.