Research On Series Arc Faults Detection For Complex Branch Based On VMD Feature Enhancement And Deep Learning

Among the causes of fire,electrical fire continues to occupy a large proportion.Among them,the series fault arc has become one of the main causes of electrical fire,when its center temperature is up to 2000 K can quickly ignite nearby combustible.In the low-voltage AC power supply and distribution system,it is more difficult to detect the various types of load at the end of the power grid and the complex changes in connection mode.Moreover,the occurrence of such faults is random and hidden,which threatens the safe and stable operation of the power system.In this paper,series fault arcs occurring in complex branches with multiple loads in parallel under low-voltage AC system are taken as the research object,and its occurrence mechanism and detection technology are studied as follows:(1)According to UL1699 standard,a series fault arc occurrence experiment platform with variable branch number is built.By selecting different experimental loads and reasonably matching,parallel arc experiments with single,double,three,four,five and six loads were carried out to collect the dry current of each branch when the arc fault occurred and normal operation.The characteristics of arc are analyzed in time and frequency domain.In the time domain analysis,the waveform method is used to analyze the waveform of the fault arc and normal operation data.In the frequency domain analysis,the 1k Hz frequency spectrum of normal and fault data is selected for analysis and comparison.(2)The study of the theory of the neural network based on cycle,built for complex load series arc fault detection LSTM network model,the current data from the experiment under normal and fault 1:1,training set and testing set 5:1 ratio of input to the LSTM network,and optimize the LSTM network parameters in order to improve the detection precision.The results show that the average detection accuracy of the proposed method reaches 92.28% for each branch and 95.18% for the total data set,realizing the arc detection of series faults under complex loads.(3)Combined with variational modal decomposition(VMD),the data are decomposed into a group of intrinsic modal components with different central frequencies.By comparing the spectra of normal and fault data,it is found that9 k Hz～25k Hz contains abundant characteristic information.The IMF4～IMF6(9.3khz～25k Hz)components decomposed by VMD are amplified and added to other components for reconstruction to form the total data set I.In order to effectively distinguish the branch where the fault arc is located,the faults of the branch of the same load type are classified as one,and all the resulting data are classified as one.A total of4 categories are re-marked to form multi-classification dataset II.(4)In order to further improve the detection accuracy,VMD feature enhancement and CNN-LSTM model are proposed and improved accordingly.The enhanced VMD data are input into the network model for training and testing.The detection results show that the combination of VMD feature enhancement and CNN-LSTM model can effectively improve the accuracy of arc detection for series faults,and the accuracy of arc detection is more than 98% on six branch data sets.The detection accuracy is99.29% on total data set I and 98.17% on multi-classification data set II.This paper has47 figures,12 tables and references.