Research On Method Of Household Electrical Load Decomposition Based On Deep Learning

With the goal of "carbon neutrality" and "carbon peaking",the efficient use of energy is becoming more and more important.Load decomposition technology is used to collect power data from the household’s total electricity meter,decompose the power information of individual devices,and provide detailed information on household electricity consumption to customers,helping them plan and regulate the use of electricity and energy in the household,achieving efficient use and conservation of electricity and energy,while reducing carbon emissions.Load decomposition of households can help power companies develop more granular household electricity consumption information,and provide more detailed data support for analysis and detection of abnormal electricity consumption information of customers.Load decomposition technology has received more attention from researchers in the field of electric energy scheduling optimization and electric energy management,providing a strong support for achieving sustainable development of energy.Load decomposition is to extract the load characteristics of power-using equipment by establishing a load decomposition model to decompose the specific power change information.With the increase of power-using devices and the development of science and technology,the traditional load feature extraction methods cannot meet the large-scale data and high-precision load decomposition.Deep learning has excellent feature extraction and regression prediction capability in serial data processing,and therefore is widely used in load decomposition.However,there are still some problems in the current load decomposition methods based on deep learning,such as the low accuracy of load decomposition for a single household and the poor generalization ability of load decomposition models for multiple households.The following studies were conducted for the application of load decomposition.First,the load decomposition dataset,load device types are summarized and analyzed.Comprehensive evaluation indexes are given for the decomposition performance and load recognition accuracy of the load decomposition model.The basis of data and performance evaluation is laid for the subsequent training of the load decomposition model.Secondly,to address the problems of insufficient accuracy of model decomposition and low training efficiency,a load decomposition method based on temporal convolutional attention network is proposed by analyzing out the operational characteristics of electricity-using devices in the dataset and using the improved temporal convolutional network as a feature extraction model.In order to increase the ability of the model to process serial data information,the attention mechanism module is introduced.Through experiments,it is shown that the proposed load decomposition method effectively improves the decomposition efficiency and decomposition performance of the model in several evaluation indexes.Finally,a domain-based adaptive load decomposition method is proposed to address the problem of low load accuracy when a single model is used for load decomposition across households.Firstly,the data from different households are divided into source and target domains,and data pre-processing is performed for the divided data.A domain adaptive load decomposition model with feature extractor,domain discriminator and predictor is constructed.Since the target domain data is unlabeled,an unsupervised approach is used to train the model and extract to domain-invariant load features among different family data.Experimental results demonstrate that the proposed method improves the load decomposition accuracy of the load decomposition model on the cross-family problem.