Non-intrusive Load Monitoring Based On Variable Input LSTM And Transfer Learning

Non-intrusive power load decomposition aims to decompose characteristic data such as the power of household-to-household total ammeter into the power consumption of each household appliance.With the popularization of smart meters and the development of smart grid technology,non-intrusive power load decomposition,as a part of innovative energy management scheme,will play a very important role in both supply-side energy management and demand-side energy management.As one of the most important machine learning technologies,deep learning has also achieved positive results in the application of non-intrusive home power load decomposition.However,when deep learning algorithms such as cyclic neural network and noise reduction automatic coding machine are applied to power load decomposition,there are long network training time and complex state electrical appliances.Load decomposition accuracy is not high.In addition,using a deep learning network model for training the sampled data of users in one area,when applied to the decomposition of power load of users in another area,there is a problem that the accuracy is reduced and a large number of new label data need to be retrained.Firstly,the input form of the cyclic neural network is improved.The structure of LSTM(Long Short-Term Memory)is used to decompose the power load.The training and testing are carried out on the UK-DALE data set.The experimental results show that the training time of the network is not only greatly reduced.In order to solve the problem that a deep learning network model trained by user data in one area needs a lot of new label data to be retrained when applied to users in another area,this paper adopts migration learning to retrain the L which has been trained on UK-DALE data set in the LSTM-RNN network migrates to the REDD data set of the United States,so that only a small amount of REDD data can be used to improve the accuracy of the network's power prediction of American household appliances,and ensure that the existing models can also reliably predict households in new areas.In this paper,the experimental results are compared with several existing in-depth learning methods.The comparison indicators include F1 score,accuracy,accuracy,recall rate,total power relative error,average absolute error,which proves that the improvement proposed in this paper has obvious effect.