Non-intrusive Load Monitoring And Decomposition Technology Based On Deep Learning And Application Of Typical Scenarios

With the continuous development and progress of society,the demand of residents,industrial and commercial users for electricity consumption and electricity consumption experience is constantly increasing.In addition,under the background of the current policy of carbon peaking and carbon neutralization,higher requirements are put forward for the renovation and construction of our smart power grid.Non-intrusive load monitoring(NILM)technology,as a key technology in the advanced measurement system of power grid terminals,achieves the acquisition of fine grained power information at load terminals through software algorithms instead of traditional hardware laying.Compared to traditional mathematical methods,deep learning replaces the process of model construction with data-driven learning,which has a stronger feature extraction ability.Therefore,this paper takes deep learning neural network as the basic architecture and open source measured data set as the experimental sample,integrates the improved temporal convolutional network(TCN)and the long short-term memory(LSTM)neural network to study the non-intrusive load condition monitoring and decomposition power,and research is conducted on the internal household itemized power load information and application scenarios reflected by the output results.(1)Based on the similarity between load condition monitoring task and computer vision recognition problem,the pretreatment load power data is converted into visual image by gramian angular field(GAF)method,and dual dimension power information of time and power is endowed in two-dimensional image to ensure the integrity and relevance of time power sequence.Then,a GAF-TCN hierarchical neural network structure is proposed by expanding the perceptive field of the convolution kernel with the advantage of expansive causal convolution.By making full use of the powerful feature extraction ability of TCN for image information,the implicit relationship between total power data and sub-load state is captured,and the load state prediction of multiple electrical devices is completed.And the predicted results are used as data support for the next layer of load decomposition task.(2)Based on the pre-identification state results,the strongly correlated power sequence(SCPS)of each electrical appliance is formed by coupling with the total meter power data.The SCPS replaces the total meter data as the training set input,effectively improving the signal to noise ratio of the target signal in the mixed signal.Since the physical essence of load decomposition is the process of decoupling the total power data,which is consistent with the sequence conversion principle,a SCPS-Seq2Seq-TPA-LSTM neural network model integrating temporal pattern attention(TPA)mechanism is proposed on the basis of the sequence to sequence(Seq2Seq)structure.The model uses the parallel operation advantage of LSTM to grasp the dependence of the front and back time points in the time series,and constructs the encoding and decoding structure to complete the conversion process from the strongly correlated power sequence to the decomposed power sequence.(3)On the basis of obtaining the fine-grained power information of the household internal load,the household internal energy consumption and its electricity consumption behavior are deeply explored,and the electricity consumption behavior after the user participates in the demand response(DR)is optimized.First,the dynamic time warping(DTW)function is introduced to modify the K-means++algorithm to solve the problem of partial scaling and drift of partial power load curve caused by electrical characteristics.According to the clustering results and many load characteristics,the internal load structure of household users is divided and the electricity consumption behavior patterns are deeply explored.The model is established using the highest electricity consumption comfort and the lowest electricity cost as the dual objective optimization functions,and the income and impact of the household’s participation in the demand response application electricity behavior optimization are analyzed.