Research On Non-intrusive Load Decomposition Method Of Household Appliances Based On Deep Learning

The construction of the ubiquitous power Internet of Things with a high degree of integration of power flow,information flow and business flow can effectively improve energy utilization efficiency and reduce energy waste,which is of great significance to promoting the energy supply revolution.The non-intrusive load decomposition technology decomposes the electricity consumption data collected by the user’s total electricity meter to obtain the user’s internal load information,helps the user to formulate a reasonable energy consumption plan,improves the flexibility of power grid scheduling,and provide data support for the establishment of demand response.The traditional non-intrusive load decomposition task mainly relies on artificial extraction of load features to construct a load decomposition model,which has problems such as difficulty in feature selection,complex model construction and strong feature dependence.Therefore,deep learning is introduced into the load decomposition task,and through its superior feature learning ability,it can reduce the difficulty of artificially designing models and improve the accuracy of load decomposition.The main work of this paper is as follows:Firstly,the basic principles and knowledge base of non-intrusive load decomposition are summarized,and a load decomposition model is constructed by using traditional methods,and many disadvantages of traditional methods are summarized through analysis.Therefore,deep learning methods are introduced in the non-intrusive load decomposition task.Then,aiming at the problem of insufficient feature extraction capability of current deep learning models,a non-intrusive load decomposition method based on Deep Residual Shrinkage Network and Attention Mechanism is proposed.With the help of the soft thresholding of the Deep Residual Shrinkage Network,the redundant information in the feature data is shrinked,and the feature information that is conducive to the decomposition task is retained,which can not only improve the feature extraction ability of the model,but also effectively reduce the problem of model degradation caused by the increase of the number of network layers in the deep neural network.In order to extract the interactive coupling features of the load,the "memory" function of the Gated Recurrent Unit is used to mine the interdependence between different time steps in the time series,and the introduction of the Attention Mechanism can make the model better learn the features of the time series information.In order to verify the feasibility of the proposed method,two power consumption scenarios are set according to the characteristics of different electrical appliances.The results show that the proposed method has stronger feature extraction ability than other load decomposition models.Finally,for the problem of poor model generalization ability caused by insufficient data samples,a non-intrusive load decomposition method based on Data Augmentation is proposed.Based on the Generative Adversarial Network,combining supervised learning and unsupervised learning to build a Time-Series Generative Adversarial Network,by mining the "time dynamics" relationship in the real data to generate time series data that conforms to the real data distribution,and the generated time series data is evaluated visually by t-distributed Stochastic Neighbor,the results show that the data generated by the model has a similar distribution to the real data in the sample space.In order to verify the improvement of the generalization ability of the model by the proposed method,the decomposition model constructed in Chapter 3 is used for verification.The results show that when the number of samples is constant,the proposed method can improve the model generalization ability to a certain extent.Before data augmentation,the average decomposition accuracy of the model after data augmentation is improved by 1.18%.