Non-intrusive Load Monitoring Algorithm Based On Deep Learning

In the context of energy shortage and environmental pollution,in order to achieve energy conservation and emission reduction,carbon neutrality and carbon peak are the current research hotspots.Through the establishment of two-way interaction between power grid and households,load monitoring enables the power grid to perceive the demand of all kinds of power loads,improve the power supply efficiency of power grid and social economic benefits,and make an important contribution to the realization of carbon neutrality and carbon peak.Non-intrusive Load Monitoring(NILM)estimates the state or disaggregate power of individual appliances from the aggregate power reading on a home meter.With the application research of deep learning,the power disaggregate error has been reduced,but the parameter amount and computational complexity of the model have also increased significantly,increasing the cost and power consumption of the device in the application.Deep learning models rely on a large amount of data for training.In fact,it is easier to obtain unlabeled data that only contains the aggregate power.However,existing research has not yet fully utilized unlabeled data.To solve the above problems,this thesis proposes a non-intrusive load monitoring model based on deep learning to achieve the balance between disaggregate power error and computational complexity.This thesis takes family load as the decomposition object,and the main work is described as follows:(1)To obtain more accurate disaggregate power,a non-intrusive load monitoring model based on attention mechanism and residual network is proposed.The model is based on the Seq2 seq framework and consists of an embedding layer,encoder and decoder.Among them,the embedding layer maps high-dimensional sparse one-hot vectors to low-dimensional dense vectors.The encoder includes Bi GRU,attention mechanism and residual network.Among them,Bi GRU can extract power sequence features from the front and rear directions,the attention mechanism can focus on important moments,and the residual network can reduce the difficulty of model fitting.The decoder includes Bi GRU,dense layer with softmax function,outputs the probability distribution of the disaggregate power.The performance of the model is tested on the open source dataset,and the average absolute error of the model is low.(2)To reduce the computational complexity of the proposed model and make full use of unlabeled data,a training framework based on knowledge distillation and semi-supervised learning is proposed.The complex neural network with low disaggregate power error is called a teacher network,and a neural network with a simple structure is called a student network.The unlabeled data is processed by the pre-trained teacher network to obtain the time series probability distribution,which is used to guide the training of the student network.On the same data set,compared with the teacher network,the power decomposition effect of the student network is similar,and the number of parameters is reduced by 93.3%,which significantly reduces the computational complexity.(3)To further improve the disaggregate power accuracy of student networks,this thesis introduces ensemble learning,and a training framework based on knowledge distillation,semisupervised learning and ensemble learning is proposed.Multiple student networks are trained under the guidance of teacher networks,and then the mean of power distribution is combined to form a strong model with lower decomposition power error than that of a single student network.The strong model is taken as a teacher network,and so on,until the integrated strong model cannot obtain a lower disaggregate power error than the student network.The testing dataset and training dataset of all experiments in this thesis are from different households,so the model can be directly applied to new households.