Research And Application Of Event-based Non-invasive Load Monitoring Technology

With the rapid development of society,the global energy demand continues to increase,which poses a huge challenge to the efficiency of traditional energy supply systems.Non-intrusive load monitoring(NILM)is a technology that can monitor power consumption without installing too many sensors,which provides opportunities for low-cost energy monitoring,and event detection and load identification are important research points in NILM.However,the current event detection technology is too single for the frequency scale of the selected signal,which makes the detection process unable to balance accuracy and rapidity.In addition,the load recognition technology lacks typicality in constructing features,which limits the recognition performance of the recognition model.Meanwhile,the unsuitable model makes the feature information cannot be fully explored,which leads to bad recognition performance.This paper conducts innovative research on the above technical points and further expands the practical application of the proposed technology.The content includes the following four aspects:1)Aiming at the problem that the traditional event detection algorithm is too single for the frequency scale of the selected signal,this paper proposes an event detection algorithm based on different frequency scales.From the perspective of lowfrequency data with low granularity,a high-sensitivity Voting Improved Isolated Forest algorithm(VIIF)is designed to rapidly pre-detect suspected events.,and from the perspective of high-frequency data with high granularity,a Time Shift Downsampling Matching algorithm(TSDM)with high-accuracy is designed to verify suspected events.This method compares with the four methods on public and private datasets,and achieves better detection performance.2)Aiming at the problem that the lack of typicality of the electrical features will limit the recognition performance of the recognition model,this paper explores three novel types of graph features from different angles of electrical signals.This paper builds the weighted voltage–current(WVI)trajectory images and Markov Transition Field(MTF)images according to the static and dynamic characteristics of the V-I trajectory,respectively.And this paper uses the current spectral sequence to construct current spectral sequence-based GAF(I-GAF)images by a novel series coding method named Gramian Angular Field(GAF).Experiments verify the validity of the three graph features.3)To fully mine the information of the constructed graph features,this paper makes use of the efficient graph feature mining capability of convolutional deep learning models to design a residual convolutional network with multi-blocks to perform load classification tasks.Finally,a variety of advanced algorithms are compared on public and private datasets,and the results show that the recognition performance of the proposed recognition model is superior.4)In order to apply the theoretical technology in practice,this paper realizes data acquisition and storage functions through wireless communication between the data acquisition device and the host computer.Then,the algorithm models are deployed on the host computer to construct a complete non-intrusive load monitoring system.Finally,this system realizes real-time load status recognition tests.