Research On Non-intrusive Load Monitoring Method For Household Users

Smart grid technology plays an important role in promoting energy structure optimization and energy efficiency improvement in the power grid.Non-intrusive load monitoring(NILM)is a sensing and measurement link in the smart grid architecture,and is also one of the key technologies in electricity management.NILM obtains equipment level electricity consumption information by analyzing total energy consumption data.It can help users understand the operation of loads,obtain detailed load energy consumption data,and formulate reasonable energy-saving plans.It can also help power companies understand the load composition,further carry out refined distribution work,and enhance the interactive response ability between the power grid and users.The existing NILM methods are mainly divided into event detection methods and nonevent detection methods.The former identifies loads based on load event detection,known as load identification;The latter achieves equipment level energy consumption monitoring through total energy consumption prediction technology,known as load decomposition.This article conducts research from three aspects: event detection,load identification,and load decomposition.The specific work is as follows:(1)A sliding window based bilateral cumulative sum(CUSUM)event detection method was designed.By analyzing the four stages of sliding window event occurrence,important parameters in the algorithm were designed to achieve load event change point detection.Solved the problem of missed and repeated detection of load events due to the wide range of load power values,different duration of switching transient processes,and large fluctuation levels during partial load operation.(2)A load identification model based on the Light Gradient Boosting Machine(Light GBM)was designed.In this model,the histogram method can utilize the advantages of different feature parameters to determine the order of internal nodes in the decision tree based on the information gain rate of the feature dictionary,thereby optimizing the optimal splitting point;The mutually exclusive feature binding method and leaf growth strategy can improve the training speed of the model and reduce computational costs.The identification performance of the model is verified by several groups of measured data validation.The experimental results show that the model has high multi class identification accuracy and fast training speed.(3)A load decomposition model was designed that combines a Dilated-Deep Residual Shrinkage Network(D-DRSN)with an improved sequence to point structure.First,after the conventional convolution layer,the cavity residual shrinkage module is introduced.The cavity convolution captures the feature dependency by expanding the receptive field,which solves the problem of learning long time series data.The residual shrinkage network strengthens and extracts deep load features,which solves the problem of gradient disappearance caused by the deepening of network layers;Secondly,the model moves the prediction point from the midpoint of the output window to the endpoint,enabling the model to learn more feature information and improving the accuracy of load decomposition.Using multiple datasets to validate the decomposition performance of the model,experimental results show that the improved model improves the accuracy of load decomposition and achieves better performance on multiple performance indicators.