Research On Denoising,Compression And Classification Of Compound Power Quality Disturbance Signals

As the total load of China’s power grid increases,the substation capacity of the power grid becomes larger and larger,and the total length of the line becomes longer and longer,and various frequent power quality disturbance events increase.Some sensitive users have higher and higher requirements for power quality.For example,some chip manufacturers need to be equipped with large-capacity UPS and dynamic voltage restorers in order to prevent the large-scale scrapping of chips caused by voltage sags and cause heavy economic losses.After the voltage sag occurs,the amplitude of the voltage can be maintained unchanged.Therefore,the study of power quality events is of great significance.Supported by the National Natural Science Foundation of China Project “Research on Composite Power Quality Disturbance Recognition and Fault Diagnosis Technology Based on Deep Learning”(Project No.52077089),this dissertation carefully studied power quality disturbance signal denoising,power quality data compression,and the multiclassification of power quality disturbance events.The main contributions of this dissertation are as follows.In terms of signal denoising,this paper proposes a joint denoising algorithm for power quality disturbance signals based on strong tracking extended Kalman filter,joint dictionary sparse decomposition,and variational mode decomposition.The algorithm first uses the number of times that the fading factor of the strong tracking filter is greater than 1 to preliminarily determine the category of the disturbance signal,and then adopts different denoising strategies for different disturbance signals: sparse decomposition and FFT joint denoising are adopted for voltage disturbance signals and harmonics containing Gaussian white noise;The method of segmenting,denoising,and finally concatenating the temporary rise,temporary drop,and interrupt signals using fading factors;For signals containing transient disturbances,the transient disturbance part is denoised using variational mode decomposition,while the peak signal is denoised by retaining the values of the peak points.The simulation results show that the algorithm significantly improves the signal-to-noise ratio after denoising.In the aspect of signal compression,a compression algorithm based on strong tracking high-order unscented Kalman filter and entropy of mixing coding is proposed.The highorder Kalman filter is used to separate the transient part and the steady part,and the amplitude and phase reserved in the steady state are compressed.The entropy of mixing coding is used to compress the transient: the detail coefficient is suppressed first,and then the long run coding is used to compress;For approximate coefficients,a segmented approximation method is adopted to improve redundancy,and Huffman encoding is used for compression.The simulation and experimental results demonstrate that the compression algorithm proposed in this paper improves the compression ratio and has a very low reconstruction error.The classification problem of composite power quality disturbances can be regarded as multi-label learning.In terms of multi-label classification,this article proposes an ensemble learning method that utilizes three different deep learning architectures,dual channel CNN-GRU,Res Net GRU,and Inception GRU,to train on the training set.Then,voting is used to determine the label of the disturbance signal.Simulation experiments have shown that the proposed algorithm outperforms traditional classifier algorithms and a single deep-learning algorithm in all five indicators.The classification experiment was conducted on the measured signals collected from the signal power source,and the results showed that the algorithm is superior to a single CNN network.In order to make full use of a large number of unlabeled power quality disturbance signals,an orthogonalization random mapping manifold regularization semi-supervised extreme learning machine classification algorithm was proposed.The algorithm uses an orthogonalization Halton sequence instead of the traditional pseudo-random number matrix as the initialization weight of online sequence extreme learning machine;Meanwhile,unlabeled data are fully utilized for data mining to further improve classification accuracy;The orthogonalization semi-supervised online sequence extreme learning machine,which is suitable for online learning,is also derived in this paper.The classification results of real fault recording signals demonstrate the effectiveness of the algorithm.