Non-intrusive Load Recognition Based On Scene Adaptation

Electricity is an important part of energy consumption.With the increasing shortage of global energy,it is very important to save electricity.Non-intrusive load identification technology can monitor the situation of residential electricity in real time.It can not only put forward rationalization proposals for users’ electricity saving,but also provide strategic support for intelligent operation of power grid.It has important research significance for the development of "source network load".At present,most of the non-intrusive load identification technologies are focused on improving the load feature modeling and recognition algorithms,but the difference of load in different application scenarios leads to the difficulty of maintaining the recognition effect of the model and algorithm.For this reason,a scene adaptive non-intrusive load identification algorithm is proposed.The frequency of data plays an important role in the recognition of load.Low frequency data are used in this paper.In order to ensure high real-time,the transient characteristics of short time and large amount of information are adopted and the extraction method is optimized.Data modeling is an important step of load identification.Based on the parallel data modeling method of active and reactive time series,a cascaded network identification load of CNN+LSTM is designed according to the structure characteristics of sample time and space and the basic idea of deep learning.The network parameters are optimized,and the results are combined with the results of load feature analysis and network training to adapt to the changes of the scene.The method of sample migration and model migration is proposed.In order to verify the feasibility of the proposed method,the public data set UKDALE is used to select 7 types of load,and two users are tested by self-test and mutual test,and 6 evaluation indexes are used to evaluate the results of the test.The experimental results include two parts: the result of load identification and the result of load decomposition.The results of load identification experiment show that the performance of CNN+LSTM network is better than that of SVM and CNN network.After applying the migration strategy,each index of the treadmill is better than the original model.The result of the load decomposition results shows that the indicators are excellent after the application of the migration strategy.