Research On Home Non-Invasive Load Monitoring Algorithm Based On Deep Learning

Non-intrusive load monitoring is a technology to obtain the power consumption of each load in the home by analyzing the power consumption information at the home user bus.In this paper,non-invasive load monitoring is divided into two directions according to its purpose: non-invasive load identification and non-invasive load decomposition,and the two directions are studied based on deep learning methods.From the point of view of non-invasive load identification,the event detection and load identification are studied.In terms of event detection,in order to solve the problems of small current electrical false detection and inaccurate steady-state moment detection in sliding window two-side CUSUM event detection algorithm,improvements and optimization were proposed based on the original CUSUM algorithm.The transient detection window was extended to be a compound computing window,the calculation method of event accumulative sum was improved,and the steady-state moment criterion was introduced.Through the combined switching experiments of three typical electrical appliances,the results show that the algorithm proposed in this paper improves the accuracy of identifying the low current load and the accuracy of determining the moment when the load enters the steady state.The data acquisition equipment is used to collect the electricity consumption data of a household in a city and some household appliances,so as to ensure the practical basis of the subsequent experimental research.The feature database was established based on plaid public data set and self-collected data to provide data support for load identification.In terms of load identification,based on the denoising autoencoder model,the reconstruction error is modified to cross entropy,and then the full convolution layer is replaced by the convolutional layer to obtain the full convolution classification autoencoder model.In order to cope with the similar current sampling waveforms in practice,which will affect the identification effect,a variety of feature fusion methods are introduced to carry out load identification,and an improved FCN-CAE load identification model is obtained.Finally,plaid data set was used to verify the effectiveness of the proposed load identification model,and then self-collected data were used for experiments.The results showed that the improved FCN-CAE load identification model could well solve the identification errors caused by waveform similarity in practice,and the identification accuracy could reach 98.7%,indicating that the model has good adaptability.Superior identification performance,high identification accuracy.From the perspective of non-invasive load decomposition,first of all,in order to solve the problem that errors occur in the process of decomposition of traditional seq2 seq load decomposition frame,which leads to the sum value of each load’s power consumption after decomposition being greater than the total load value,a sequentially to multi-sequence reconstruction model is introduced and improved on the basis of seq2 seq load decomposition frame.A load decomposition framework based on seq2 multiseq is proposed.Then,the traditional variational autoencoder is optimized to obtain the VAE load decomposition model which is improved.Finally,the VAE load decomposition model is obtained based on seq2 multiseq framework.Finally,Ampds data set and UK-DALE data set are adopted to verify the performance of the proposed load decomposition model.Then,through self-collected data,experiments are carried out.The results show that the improved load decomposition model has better performance than the typical load decomposition model,and the accuracy of load decomposition is significantly improved.The seq2 multiseq framework further improves the accuracy and stability of load decomposition.Aiming at the problem that the power decomposition curve obtained from load decomposition has a large error with the actual power curve when it enters the steady state,an application scheme combining the two technical routes of load identification and load decomposition is proposed.Firstly,a fuzzy identification algorithm based on deep learning is proposed to ensure that the identification results contain correct categories.Then,the results of load decomposition are corrected through the results of fuzzy identification,and the corrected results are output when the constraint conditions are met.Finally,the N value of the fuzzy identification result is obtained through data training,and then the fusion algorithm is tested by self-collected data.The results show that the fusion algorithm improves the accuracy of the load decomposition result when the appliance enters the steady state,makes the power decomposition curve obtained by decomposition more close to the reality,and reduces the power prediction error.Figure [75] Table [10] Reference [93]...