CN110376497B - Low-voltage distribution system series fault electric arc identification method based on full-phase deep learning - Google Patents

Low-voltage distribution system series fault electric arc identification method based on full-phase deep learning Download PDF

Info

Publication number
CN110376497B
CN110376497B CN201910739837.0A CN201910739837A CN110376497B CN 110376497 B CN110376497 B CN 110376497B CN 201910739837 A CN201910739837 A CN 201910739837A CN 110376497 B CN110376497 B CN 110376497B
Authority
CN
China
Prior art keywords
full
phase
deep learning
low
frequency spectrum
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201910739837.0A
Other languages
Chinese (zh)
Other versions
CN110376497A (en
Inventor
冷继伟
陈烜
段卫平
杜刚
肖屏
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
State Grid Sichuan Electric Power Service Co ltd
Original Assignee
State Grid Sichuan Electric Power Service Co ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by State Grid Sichuan Electric Power Service Co ltd filed Critical State Grid Sichuan Electric Power Service Co ltd
Priority to CN201910739837.0A priority Critical patent/CN110376497B/en
Publication of CN110376497A publication Critical patent/CN110376497A/en
Application granted granted Critical
Publication of CN110376497B publication Critical patent/CN110376497B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
    • G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
    • G01R31/12—Testing dielectric strength or breakdown voltage ; Testing or monitoring effectiveness or level of insulation, e.g. of a cable or of an apparatus, for example using partial discharge measurements; Electrostatic testing

Landscapes

  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Testing Of Short-Circuits, Discontinuities, Leakage, Or Incorrect Line Connections (AREA)

Abstract

The invention discloses a low-voltage distribution system series fault arc identification method based on full-phase deep learning, and solves the problems that in the existing low-voltage distribution network faults, the identification method aiming at series fault arcs is easily interfered by noise and frequency spectrum leakage, the identification effect is influenced, the identification efficiency is not high, and the stability is not high. The method comprises the following steps: under a low-voltage alternating current system, current signals are acquired for different loads in a low-voltage loop; carrying out full-phase discrete Fourier transform on the acquired current signal, carrying out full-phase frequency spectrum characteristic quantity extraction on the load, and constructing a full-phase frequency spectrum characteristic vector; constructing a deep learning neural network model based on Logistic regression, and performing deep learning training on full-phase frequency spectrum characteristic quantities under different loads and different running states until the model is converged; and (4) utilizing the trained model to complete the discrimination of different load types and the identification of whether the series fault arc occurs.

Description

Low-voltage distribution system series fault arc identification method based on full-phase deep learning
Technical Field
The invention relates to the technical field of low-voltage distribution system series fault arc identification, in particular to a low-voltage distribution system series fault arc identification method based on full-phase deep learning.
Background
In a low-voltage distribution system, when insulation of a line or equipment is aged or broken or electrical contact is poor, a fault arc is often accompanied, and further fire hazard is caused. The low-voltage fault arc is divided into a parallel fault arc and a series fault arc, the parallel fault arc is caused by a short-circuit fault between distribution network lines, the short-circuit current is large, and the protection device can correctly identify and act. The series fault arc is caused by poor contact or broken line of a line, the fault current value is 5-30A, and the sensitivity requirement of the protection device cannot be met. In an actual low-voltage distribution network fault, the most difficult to identify is also the series fault arc.
The detection methods for series fault arcs can be currently classified into three categories: firstly, a simulation circuit based on a Cassie arc model; judging by using physical characteristics such as light, heat, sound and the like generated when the electric arc occurs; analyzing the current and voltage signals when the series arc occurs, and identifying the fault arc by using a mathematical method. The third method mainly focuses on time-frequency characteristic analysis of current and voltage signals, mathematical analysis of characteristic vectors and the like due to wide application range and high reliability. The time domain characteristic analysis is to identify the series fault arc by using the characteristics of current, voltage waveform change rate, abrupt waveform slope value, zero rest time and the like, and has the defect that the time domain index generates larger error and interferes the judgment result due to the influence of environmental noise or the pollution of sampling data. The frequency domain characteristic analysis is to judge whether the series fault arc occurs by using the frequency spectrum characteristic, the harmonic and the inter-harmonic content of the signal on the frequency domain, and has the defects of being easily influenced by frequency spectrum leakage and interfering with a judgment result. The mathematical analysis method of the feature vector comprises an autoregressive parameter model, a fractal dimension and support vector machine, wavelet transformation, singular value decomposition and the like, the principle and the detection flow are complex, the stability in actual detection is not high, and the method cannot be well applied to engineering practice.
Disclosure of Invention
The technical problem to be solved by the invention is as follows: in the existing low-voltage distribution network fault, aiming at the problems that the identification method of the series fault electric arc is easily interfered by noise and frequency spectrum leakage, the identification effect is influenced, the identification efficiency is not high, and the stability is not high, the invention provides the low-voltage distribution system series fault electric arc identification method based on full-phase deep learning, which solves the problems.
The invention is realized by the following technical scheme:
the method for identifying the series fault arc of the low-voltage distribution system based on the full-phase deep learning comprises the following steps:
under a low-voltage alternating current system, current signals are acquired for different loads in a low-voltage loop;
carrying out full-phase discrete Fourier transform on the acquired current signal, carrying out full-phase frequency spectrum characteristic quantity extraction on the load, and constructing a full-phase frequency spectrum characteristic vector;
constructing a deep learning neural network model based on Logistic regression by using the established full-phase frequency spectrum characteristic vector, and performing deep learning training on the full-phase frequency spectrum characteristic vector under different loads and different running states until the model is converged;
performing full-phase spectrum analysis measurement on the segmented load current waveform to be identified by using a Logistic regression-based deep learning neural network model completed by learning training, completing the discrimination of different load types and the identification of whether series fault arcs occur, and outputting according to the model
Figure BDA0002163564350000021
It numbers each load according to the type, normal operation or fault arc state,
Figure BDA0002163564350000022
and is
Figure BDA0002163564350000023
Wherein u ≠ v, u, v ≠ 0, 1, …, 47, which indicates that the type number of the identified load is round (u +1)/2, the operation state is normal arc, and u is even number; or the running state is fault arc, and u is odd number; where round () denotes rounding.
The working principle is as follows: in the existing low-voltage distribution network faults, aiming at the problems that the identification method of the series fault arc is easily interfered by noise and frequency spectrum leakage, the identification effect is influenced, the identification efficiency is not high, and the stability is not high, the invention adopts the scheme to take the full-phase spectrum characteristic forming mechanism of the load current signal as an entry point, and extracts the characteristic quantities of the distorted signal and the normal signal on the full-phase spectrum by utilizing the frequency spectrum leakage and the frequency spectrum interference suppression performance of the full-phase spectrum; constructing a deep learning neural network model based on Logistic regression, and distinguishing input full-phase frequency spectrum characteristics by using a deep learning classification identification method based on Logistic regression to realize discrimination of load types and whether fault arcs occur; the nonlinear, adaptive and fault-tolerant capabilities of the neural network are utilized to intelligently and effectively identify the series fault arc. The method for identifying the series fault arc is not easily interfered by noise and frequency spectrum leakage, and has high identification efficiency and strong stability.
Furthermore, the low-voltage alternating current system comprises a power supply unit, a fault generation unit and a data sampling and analyzing unit, wherein the power supply unit adopts a low-voltage alternating current test power supply AC, the low-voltage alternating current test power supply AC is connected with one end of a primary coil of an isolation transformer through a switch K1, the other end of the primary coil of the isolation transformer is grounded, one end of a secondary coil of the isolation transformer is connected with an arc generator, the arc generator is connected with a resistor R1, and a resistor R1 is connected with a sliding rheostat RpVaristor RpThe other end of the secondary side coil of the isolation transformer is connected through a switch K2; the arc generator is used for generating stable series fault arc, and a switch K3 is connected in parallel with the arc generator; the resistor R1 is connected with a data acquisition unit DA, and the data acquisition unit DA is connected with a computer PC to realize analysis; wherein, the slide rheostat RpResistance R1Representing the load and the sampling resistance, respectively.
Further, the acquired current signal is subjected to full-phase discrete fourier transform, wherein a frequency offset formula of the final full-phase discrete fourier transform is as follows:
Figure BDA0002163564350000024
in the formula: g (k)i)、Y(ki) Respectively the serial number k of the frequency peak spectral lineiCorresponding discrete fourier transform and full-phase discrete fourier transform spectral vector values,
Figure BDA0002163564350000025
for respective corresponding phase spectrum values, betai-kiIs the peak line kiN is an arbitrary natural number.
Specifically, the derivation process of the frequency offset formula of the final full-phase discrete fourier transform is as follows:
for a low-voltage alternating-current system, current signals are acquired for different loads in a low-voltage loop, an actual current signal of the low-voltage alternating-current system is represented as linear superposition of a plurality of signal components, and a load current sampling value y (n delta t) is expressed by the following formula:
Figure BDA0002163564350000031
(1) in the formula: n is [ -N +1, N-1 [ ]]2N-1 is the analysis data length; Δ t is the sampling interval, ωi、Ai、piThe angular frequency, amplitude and initial phase of the i signal components, respectively, and M is the total number of signal components.
Respectively performing discrete Fourier transform and full-phase discrete Fourier transform (ignoring the influence of negative frequency) on the current signal in the formula (1), wherein the respective frequency spectrum expressions are as follows:
Figure BDA0002163564350000032
Figure BDA0002163564350000033
(2) (3) in the formula: g (k) is a discrete Fourier transform spectrum, Y (k) is a full-phase discrete Fourier transform spectrum, k is a line number, betai＝ωiΔ ω,. DELTA.ω is the angular frequency resolution,. omegaiIs the angular frequency value of the i frequency component, W (beta)i-k) is a frequency offset βi-a corresponding rectangular window spectral function at k, AiIs the amplitude of the i signal component, piIs the initial phase of the signal component of i, M is the total number of the signal components; wherein
Figure BDA0002163564350000034
Let frequency component omegaiThe serial number of the spectral line of the frequency spectrum amplitude is ki(4) in the formula: beta is ai-kiIs the peak line kiThe amount of frequency offset of (a) is,n is an arbitrary natural number, betai＝ωiΔ ω,. DELTA.ω is the angular frequency resolution,. omegaiIs the angular frequency value of the i frequency component. From this, the frequency offset at which the final full-phase discrete fourier transform is derived is:
Figure BDA0002163564350000035
(5) in the formula: g (k)i)、Y(ki) Respectively the serial number k of the frequency peak spectral lineiCorresponding discrete fourier transform and full-phase discrete fourier transform spectral vector values,
Figure BDA0002163564350000036
for respective corresponding phase spectrum values, betai-kiIs the peak line kiN is an arbitrary natural number.
The angular frequency value of the i signal frequency component is:
ωi＝βiΔω (6)
the amplitude corresponding to the i signal frequency component is:
Figure BDA0002163564350000041
(7) in the formula: g (k) is a discrete Fourier transform spectrum, Y (k) is a full-phase discrete Fourier transform spectrum, and k is a line number.
Equations (6) and (7) are the calculation equations of the angular frequency value and the amplitude value corresponding to the i signal frequency component based on the all-phase discrete fourier transform. As can be seen from equations (2) and (3), when spectrum leakage occurs, the amplitude of the all-phase discrete fourier transform is attenuated by multiples, which is also the root cause of the good spectrum leakage and spectrum interference suppression performance of the spectrum of the all-phase discrete fourier transform, so that the measured frequency components have high accuracy.
Further, the extraction of the full-phase frequency spectrum characteristic quantity of the load specifically comprises the extraction of current frequency spectrum characteristics when the load normally operates and current frequency spectrum characteristics under series fault arcs, and the full-phase spectrum measurement is carried out on current waveforms when the load normally operates and the series fault arcs occur, so that the frequency spectrum characteristics of different loads under different states are extracted; the discrimination of whether fault arcs occur to different loads is realized by the characteristic identification of the amplitude distribution of each subharmonic component and the frequency and amplitude distribution condition of the interharmonic component playing a leading role; the method comprises the steps of identifying the amplitude distribution of each subharmonic component and the frequency and amplitude distribution condition of the interharmonic component playing a leading role, wherein the extracted specific characteristics specifically comprise the leading interharmonic frequency, the leading interharmonic content and the harmonic content.
Further, constructing a full-phase spectral feature vector with dominant inter-harmonic frequencies fIh-mDominant inter-harmonic content AIh-mAnd harmonic content AH-mForm 150-dimensional feature vector X ═ fIh-m，AIh-m，AH-m]Where m is 1, 2, …, 50, Ih-m represents the mth inter-harmonic, and H-m represents the mth harmonic.
Further, a deep learning neural network model based on Logistic regression is constructed by using the established full-phase spectrum feature vector, which specifically comprises the following steps:
the deep learning method based on Logistic regression is characterized in that when a traditional neural network is adopted to perform discriminant analysis on fault arcs, complex samples cannot be accurately classified, meanwhile, the convergence of results is greatly influenced by initial values, and the results are prone to fall into the conditions of local optimization, overfitting or even non-convergence. The Logistic regression is not influenced by the selection of the initial value, and the result is very stable and has stronger robustness when the initial value does not meet the assumption, so the method adopts a deep learning method based on the Logistic regression to train the sample.
In order to discriminate whether fault arcs are generated by different loads, 50-th harmonic content A specified by IEC61000-4-7 standard is selectedH-mAs characteristic parameters (m is 1, 2, …, 50), a dominant inter-harmonic frequency f between the two harmonics is selected at the same timeIh-mContent A ofIh-mAs characteristic parameters, a common composition 150-dimensional feature vector X ═ fIh-m,AIh-m,AH-m]For the input parameters of deep learning of the neural network, the deep learning neural network comprises 5 hidden layers, and the output response variable of the deep learning neural network
Figure BDA0002163564350000042
Figure BDA0002163564350000051
It numbers each sample load according to the type, normal operation or fault arc state,
Figure BDA0002163564350000052
and is
Figure BDA0002163564350000056
Figure BDA0002163564350000057
The identification of the type number of the load is round (u +1)/2, the operation state is normal (u is even number) or fault arc (u is odd number), wherein round () represents rounding.
The Logistic regression model in the deep learning neural network model based on the Logistic regression is a nonlinear regression model and has the following distribution form:
Figure BDA0002163564350000054
(8) in the formula: e (X) represents the probability distribution, p is the predicted probability value for the corresponding input variable X, ε is the prediction error, β0And beta1Is the nonlinear regression coefficient of the model;
the Logistic regression is to calculate a predicted probability value p of Y ═ 1 for different levels of the input variable X, where the probability value takes 0.5 as a partition point, and the formula is:
Figure BDA0002163564350000055
(9) in the formula: u is 0, 1, …, 47; y isu1 indicates a normal state (u is even) or a fault arc state (u is odd).
The invention has the following advantages and beneficial effects:
1. the invention takes the full-phase spectrum characteristic forming mechanism of the load current signal as an entry point, extracts the characteristic quantity of the distorted signal and the normal signal on the full-phase spectrum, and is not easy to be interfered by noise, spectrum leakage and spectrum; the measuring method based on full-phase spectrum analysis can effectively extract current characteristic parameters of the load in various operating states, and provides a theoretical basis for identifying the fault arc;
2. the deep learning neural network model based on Logistic regression is constructed, the neural network has nonlinearity, adaptivity and fault-tolerant capability, and the model network can be stably and quickly converged; judging the input full-phase frequency spectrum characteristics by using a deep learning classification identification method based on Logistic regression, and realizing the discrimination of the load type and whether a fault arc occurs;
3. the method for identifying the series fault arc is reasonable in flow, convenient to operate and use, not prone to being interfered by noise and frequency spectrum leakage, high in identification efficiency and strong in stability.
Drawings
The accompanying drawings, which are included to provide a further understanding of the embodiments of the invention and are incorporated in and constitute a part of this application, illustrate embodiment(s) of the invention and together with the description serve to explain the principles of the invention. In the drawings:
fig. 1 is a flow chart of a series fault arc identification method for a low-voltage distribution system based on full-phase deep learning according to the invention.
FIG. 2 is a schematic diagram of a series fault arc test platform of the present invention.
Fig. 3(a) is a diagram of a normal current waveform of the linear load of the present invention.
Fig. 3(b) is a current spectrum diagram of the linear load according to the present invention.
Fig. 4(a) is a normal current waveform diagram of the nonlinear load according to the present invention.
Fig. 4(b) is a diagram of the normal current spectrum of the nonlinear load according to the present invention.
Fig. 5(a) is a current waveform diagram at the time of a linear load fault arc of the present invention.
Fig. 5(b) is a current spectrum diagram at the time of a linear load fault arc of the present invention.
Fig. 6(a) is a current waveform diagram at the time of a nonlinear load fault arc according to the present invention.
Fig. 6(b) is a current spectrum diagram of the nonlinear load fault arc of the present invention.
FIG. 7 is a deep learning neural network model based on Logistic regression according to the present invention.
FIG. 8 is a deep learning training error map of the neural network of the present invention.
FIG. 9 is a graph of the network training output error of the present invention and other methods.
FIG. 10(a) is a content trend chart of each harmonic in the present invention.
FIG. 10(b) is a graph showing the content trend of the harmonics of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail below with reference to examples and accompanying drawings, and the exemplary embodiments and descriptions thereof are only used for explaining the present invention and are not meant to limit the present invention.
Examples
As shown in fig. 1 to 10(b), the method for identifying the series fault arc of the low-voltage distribution system based on the full-phase deep learning comprises the following steps:
under a low-voltage alternating current system, current signals are acquired for different loads in a low-voltage loop;
carrying out full-phase discrete Fourier transform on the acquired current signal, carrying out full-phase frequency spectrum characteristic quantity extraction on the load, and constructing a full-phase frequency spectrum characteristic vector;
constructing a deep learning neural network model based on Logistic regression by using the established full-phase frequency spectrum characteristic vector, and performing deep learning training on the full-phase frequency spectrum characteristic vector under different loads and different running states until the model is converged;
and carrying out full-phase spectrum analysis measurement on the segmented load current waveform to be identified by using a Logistic regression-based deep learning neural network model which is completed by learning training, and completing the discrimination of different load types and the identification of whether series fault arcs occur.
Specifically, under a low-voltage alternating-current system, current signals are acquired for different loads in a low-voltage loop, the current signals are expressed as linear superposition of a plurality of signal components, and a load current sampling formula is as follows:
Figure BDA0002163564350000071
in the formula: n is [ -N +1, N-1 [ ]]2N-1 is the analysis data length; Δ t is the sampling interval, ωi、Ai、piThe angular frequency, amplitude and initial phase of the i signal components, respectively, and M is the total number of signal components.
Specifically, the acquired current signal is subjected to full-phase discrete fourier transform, wherein a frequency offset formula of the final full-phase discrete fourier transform is as follows:
Figure BDA0002163564350000072
in the formula: g (k)i)、Y(ki) Respectively the serial number k of the frequency peak spectral lineiCorresponding discrete fourier transform and full-phase discrete fourier transform spectral vector values,
Figure BDA0002163564350000073
for respective corresponding phase spectrum values, betai-kiIs the peak line kiN is an arbitrary natural number.
Specifically, the extraction of the full-phase frequency spectrum characteristic quantity of the load specifically comprises the extraction of current frequency spectrum characteristics when the load normally operates and current frequency spectrum characteristics under series fault arcs, and the extraction of the frequency spectrum characteristics of different loads under different states is realized by performing full-phase spectrum measurement on current waveforms when the load normally operates and the series fault arcs occur; and the discrimination of whether fault arcs occur to different loads is realized by the characteristic identification of the amplitude distribution of each subharmonic component and the frequency and amplitude distribution condition of the interharmonic component playing a leading role. The method comprises the steps of identifying the amplitude distribution of each subharmonic component and the frequency and amplitude distribution condition of the interharmonic component playing a leading role, wherein the extracted specific characteristics specifically comprise the leading interharmonic frequency, the leading interharmonic content and the harmonic content.
Extracting the full-phase frequency spectrum characteristic quantity of the load, specifically comprising extracting the current frequency spectrum characteristic when the load normally operates and the current frequency spectrum characteristic under the series fault arc, wherein the analysis process of the full-phase frequency spectrum characteristic quantity is as follows:
(1) current spectrum characteristic of load in normal operation
According to the above-mentioned full-phase spectrum forming mechanism, full-phase discrete fourier transform is performed on the linear load current signal and the non-linear load current signal corresponding to equation (1), and the frequency and the amplitude of the frequency component are measured, so as to obtain the time domain waveform and the frequency spectrum thereof, which are respectively shown in fig. 3(a), fig. 3(b), fig. 4(a), and fig. 4 (b).
As can be seen from fig. 3(a), 3(b), 4(a) and 4(b), when the linear load is in normal operation, the waveform is close to an ideal sine wave, the spectrogram only contains a fundamental wave component, and the full-phase spectrum measurement result only contains a fundamental wave amplitude; the waveform generates regular periodic distortion when the nonlinear load normally runs, each frequency spectrum line in a frequency spectrum graph forms a frequency spectrum band, and a full-phase spectrum measurement result comprises each subharmonic component and a small part of interharmonic components, wherein odd subharmonic components are more prominent. When the load normally operates, the spectrum characteristics of the linear load and the nonlinear load can be extracted by the full-phase spectrum measuring method, and different types of loads have different spectrum characteristics, so that different types of linear loads and nonlinear loads in a normal operation state can be distinguished.
(2) Current spectral signature under series fault arc
By building a series fault arc simulation test platform, time domain and frequency domain analysis is carried out on the conventional linear load and nonlinear load current signal waveforms under the series fault arc, and the obtained waveform diagram and frequency spectrum diagram are respectively shown in fig. 5(a), 5(b), 6(a) and 6 (b).
As can be seen from fig. 5(a), 5(b), 6(a) and 6(b), waveforms of both the linear load and the nonlinear load are irregularly distorted during the series fault arc, and the waveforms are accompanied by burrs, and full-phase spectrum measurement is performed on the waveforms, so that inter-harmonic components including sub-harmonic components and different frequencies are obtained. Comparing fig. 3(a), 3(b) and 4(a), 4(b) at the same time, the current waveform when the series fault arc occurs can be regarded as a multi-linear superposition of current distortion waveform in normal operation, which aggravates the distortion degree of the signal, forms a frequency spectrum band on the frequency spectrum diagram, and has inter-harmonic components and prominent harmonic components (such as 3 rd order and 5 th order harmonics) with obviously different frequencies and amplitudes.
By performing full-phase spectrum measurement on the current waveform in normal operation and when series fault arcs occur, spectrum characteristics of different loads in different states can be extracted. By means of the characteristic identification of the amplitude distribution of each subharmonic component and the frequency and amplitude distribution condition of the interharmonic component playing a leading role, the discrimination of whether fault arcs occur to different loads can be achieved.
The construction of the full-phase frequency spectrum feature vector is dominated by the inter-harmonic frequency fIh-mDominant inter-harmonic content AIh-mAnd harmonic content AH-mForm 150-dimensional feature vector X ═ fIh-m，AIh-m，AH-m]Where m is 1, 2, …, 50, Ih-m represents the mth inter-harmonic, and H-m represents the mth harmonic.
The method comprises the following steps of constructing a deep learning neural network model based on Logistic regression by using the built full-phase frequency spectrum feature vector, and specifically comprising the following steps:
150 dimensional eigenvector X ═ fIh-m，AIh-m，AH-m]As an input parameter for deep learning of the neural network, the deep learning neural network comprises 5 hidden layers, and the output of the deep learning neural networkResponse variable
Figure BDA0002163564350000081
It numbers each sample load according to the type, normal operation or fault arc state,
Figure BDA0002163564350000082
and is
Figure BDA0002163564350000083
Figure BDA0002163564350000084
The identification of the type number of the load is round (u +1)/2, the operation state is normal (u is even number) or fault arc (u is odd number), wherein round () represents rounding.
The Logistic regression model in the deep learning neural network model based on the Logistic regression is a nonlinear regression model and has the following distribution form:
Figure BDA0002163564350000091
in the formula: e (X) represents the probability distribution, p is the predicted probability value for the corresponding input variable X, ε is the prediction error, β0And beta1Is the nonlinear regression coefficient of the model;
the Logistic regression is to calculate a predicted probability value p of Y ═ 1 for different levels of the input variable X, where the probability value takes 0.5 as a partition point, and the formula is:
Figure BDA0002163564350000092
wherein u is 0, 1, …, 47; y isu1 indicates a normal state (u is even) or a fault arc state (u is odd).
As shown in fig. 7, for the neural network training process in the Logistic regression-based deep learning network model, a total of 24 loads such as a resistor box, an electric kettle, an incandescent lamp, a pistol drill, a microwave oven, a vacuum cleaner, an LED lamp, a fluorescent lamp, a dimming lamp and the like meeting the UL1699 standard under different powers are adopted as test samples, each load is numbered (0-47) according to the type, normal operation or fault arc state, 100 normal operation waveform samples and 100 fault arc waveform samples are taken for each load, and a total of 4800 samples are subjected to the neural network deep learning shown in fig. 7. The samples are converged finally, the condition that the traditional neural network is not converged does not occur, and the network learning accuracy reaches 100%. The mean square value of the training error of the neural network in deep learning based on Logistic regression is shown in fig. 8, and with the increase of the number of times of deep learning training, the gradient decline index of the network error is more ideal, and the network can be stably and rapidly converged.
The output errors using the conventional BP neural network (method 1), the genetic algorithm neural network (method 2) and the logistic regression deep learning neural network (method 3) are shown in fig. 9.
As can be seen from fig. 9, the Logistic regression-based deep learning neural network (method 3) has better convergence, and rapidity and stability in the convergence process are better than those of the other two methods. Meanwhile, the defects that the traditional BP neural network is slow in convergence and the neural network based on the genetic algorithm is unstable in convergence are avoided.
Therefore, a fault arc identification flow chart based on full phase spectrum and deep learning is shown in fig. 1.
Aiming at the low-voltage distribution system series fault arc identification method based on full-phase deep learning, in order to simulate the real electric quantity change condition when the series fault arc occurs, a series fault arc simulation test platform is set up according to UL1699-2008 AFCI standard as shown in figure 2, the low-voltage alternating current system comprises a power supply unit, a fault generation unit and a data sampling analysis unit, the power supply unit adopts a low-voltage alternating current test power supply AC, the low-voltage alternating current test power supply AC is connected with one end of a primary side coil of an isolation transformer through a switch K1, the other end of the primary side coil of the isolation transformer is grounded, one end of a secondary side coil of the isolation transformer is connected with an arc generator, the arc generator is connected with a resistor R1, and the resistor R1 is connected with a slide rheostat RpVaristor RpThrough switchK2 is connected with the other end of the secondary coil of the isolation transformer; rp、R10.1 Ω denotes a test load and a sampling resistance, respectively; the arc generator is used for generating stable series fault arc, and a switch K3 is connected in parallel with the arc generator; and the resistor R1 is connected with a data acquisition unit DA, and the data acquisition unit DA is connected with a computer PC to realize analysis.
The full-phase spectrum analysis and measurement are performed on 4-segment load current waveforms collected by the production operation center of the power company in the province of the national network of Sichuan province in which the constructed test platform is located, and the obtained variation trends of the content of each (the highest 10) harmonic and the dominant inter-harmonic are shown in fig. 10(a) and 10 (b).
Fig. 10(a) shows that the characteristic quantity based on the full-phase spectrum analysis is obvious, the harmonic content of the signal 1 and the harmonic content of the signal 3 are intensively distributed at odd harmonics, the harmonic content of the signal 1 is integrally lowest and is about 2.5% at the highest, and the harmonic content of the signal 3 is integrally higher and is about 26.6% at the highest; the signals 2 and 4 are distributed at each harmonic, and the overall level of the signal 4 is higher than that of the signal 2 in terms of harmonic content. As can be seen from fig. 10(b), the inter-harmonic contents of signals 2 and 4 are higher overall, and the inter-harmonic contents of signals 1 and 3 are lower overall. The test signal was further quantified and the results of the partial measurements based on the eigenvalues of the full phase spectrum are shown in table 1.
TABLE 1 partial measurement results based on full phase spectral analysis
Figure BDA0002163564350000101
The test results show that the measurement method based on the full-phase spectrum analysis can effectively extract the current characteristic parameters of the load in various running states and provide a theoretical basis for identifying the fault arc.
Dominant inter-harmonic frequencies f to be based on full-phase spectral measurementsIh-mContent A ofIh-mAnd harmonic content AH-mForming 150-dimensional feature vectors (m is 1, 2, …, 50) as input parameters of the neural network in fig. 7, which has completed Logistic regression deep learning, to obtainThe output of the 4-stage test signal and the identification result of the fault arc are shown in table 2. The test conclusion is consistent with the test preset conditions, the correctness and the effectiveness of the low-voltage distribution system series fault arc identification method based on the full-phase deep learning are verified, the identification of the load type and whether fault arc occurs is realized, the identification efficiency is high, and the method is high in stability.
TABLE 2 test results of signals of each segment
Figure BDA0002163564350000102
The above-mentioned embodiments are intended to illustrate the objects, technical solutions and advantages of the present invention in further detail, and it should be understood that the above-mentioned embodiments are merely exemplary embodiments of the present invention, and are not intended to limit the scope of the present invention, and any modifications, equivalent substitutions, improvements and the like made within the spirit and principle of the present invention should be included in the scope of the present invention.

Claims (4)

1. The method for identifying the series fault arc of the low-voltage distribution system based on the full-phase deep learning is characterized by comprising the following steps of:
under a low-voltage alternating current system, current signals are acquired for different loads in a low-voltage loop;
carrying out full-phase discrete Fourier transform on the acquired current signal, carrying out full-phase frequency spectrum characteristic quantity extraction on the load, and constructing a full-phase frequency spectrum characteristic vector;
constructing a deep learning neural network model based on Logistic regression by using the established full-phase frequency spectrum characteristic vector, and performing deep learning training on the full-phase frequency spectrum characteristic vector under different loads and different running states until the model is converged;
carrying out full-phase spectrum analysis measurement on the segmented load current waveform to be identified by using a Logistic regression-based deep learning neural network model completed by learning training, and discriminating different load types and judging whether to useIdentification of occurrence of series fault arc, output according to the model
Figure FDA0003548228300000011
It numbers each load according to the type, normal operation or fault arc state,
Figure FDA0003548228300000012
and is
Figure FDA0003548228300000013
Wherein u ≠ v, u, v ≠ 0, 1, …, 47, which indicates that the type number of the identified load is round (u +1)/2, the operation state is normal arc, and u is even number; or the running state is fault arc, and u is odd number; where round () denotes rounding;
and performing full-phase discrete Fourier transform on the acquired current signal, wherein the final frequency offset formula of the full-phase discrete Fourier transform is as follows:
Figure FDA0003548228300000014
in the formula: g (k)i)、Y(ki) Respectively the serial number k of the frequency peak spectral lineiCorresponding discrete fourier transform and full-phase discrete fourier transform spectral vector values,
Figure DEST_PATH_FDA0002163564340000016
for respective corresponding phase spectrum values, betai-kiIs the peak line kiThe frequency offset, N, is any natural number; Δ ω is angular frequency resolution;
the low-voltage alternating current system comprises a power supply unit, a fault generation unit and a data sampling and analyzing unit, wherein the power supply unit adopts a low-voltage alternating current test power supply AC, the low-voltage alternating current test power supply AC is connected with one end of a primary coil of an isolation transformer through a switch K1, the other end of the primary coil of the isolation transformer is grounded, and a secondary coil of the isolation transformerOne end of the arc generator is connected with the arc generator, the arc generator is connected with the resistor R1, the resistor R1 is connected with the slide rheostat RpVaristor RpThe other end of the secondary side coil of the isolation transformer is connected through a switch K2; the arc generator is used for generating stable series fault arc, and a switch K3 is connected in parallel with the arc generator; the resistor R1 is connected with a data acquisition unit DA, and the data acquisition unit DA is connected with a computer PC to realize analysis; wherein, the slide rheostat RpResistance R1Respectively representing a load and a sampling resistance;
the extraction of the full-phase frequency spectrum characteristic quantity of the load specifically comprises the extraction of current frequency spectrum characteristics when the load normally operates and current frequency spectrum characteristics under series fault arcs, and the extraction of the frequency spectrum characteristics of different loads under different states is realized by performing full-phase spectrum measurement on current waveforms when the load normally operates and the series fault arcs occur; the discrimination of whether fault arcs occur to different loads is realized by the characteristic identification of the amplitude distribution of each subharmonic component and the frequency and amplitude distribution condition of the interharmonic component playing a leading role;
the method comprises the steps of identifying the amplitude distribution of each subharmonic component, the frequency of the interharmonic component playing a leading role and the characteristics of the amplitude distribution condition, wherein the extracted specific characteristics specifically comprise the leading interharmonic frequency, the leading interharmonic content and the harmonic content;
the construction of the full-phase frequency spectrum feature vector is dominated by the inter-harmonic frequency fIh-mDominant inter-harmonic content AIh-mAnd harmonic content AH-mForm 150-dimensional feature vector X ═ fIh-m，AIh-m，AH-m]Where m is 1, 2, …, 50, Ih-m represents the mth inter-harmonic, and H-m represents the mth harmonic.
2. The method for identifying the series fault arc of the low-voltage distribution system based on the full-phase deep learning as claimed in claim 1, wherein the current signals are collected from different loads in the low-voltage loop under the low-voltage alternating current system, and the current signals are represented by linear superposition of a plurality of signal components, wherein the load current sampling formula is as follows:
Figure FDA0003548228300000021
in the formula: n is [ -N +1, N-1 [ ]]2N-1 is the analysis data length; Δ t is the sampling interval, ωi、Ai、piThe angular frequency, amplitude and initial phase of the i signal components, respectively, and M is the total number of signal components.
3. The method for identifying the series fault arc of the low-voltage distribution system based on the full-phase deep learning as claimed in claim 1, wherein the built full-phase frequency spectrum feature vector is used for building a deep learning neural network model based on Logistic regression, and the method comprises the following specific steps:
150 dimensional eigenvector X ═ fIh-m，AIh-m，AH-m]As an input parameter of the deep learning of the neural network, the deep learning neural network comprises 5 hidden layers, and an output response variable of the deep learning neural network is Y.
4. The method for identifying the series fault arc of the low-voltage distribution system based on the full-phase deep learning as claimed in claim 3, wherein the Logistic regression model in the Logistic regression based deep learning neural network model is a nonlinear regression model and has the following distribution form:
Figure FDA0003548228300000022
in the formula: e (X) represents the probability distribution, p is the predicted probability value for the corresponding input variable X, ε is the prediction error, β0And beta1Is the nonlinear regression coefficient of the model;
the Logistic regression is to calculate a predicted probability value p of Y ═ 1 for different levels of the input variable X, where the probability value takes 0.5 as a partition point, and the formula is:
Figure FDA0003548228300000023
wherein u is 0, 1, …, 47; y isu1 denotes normal state, u is an even number; or yuWith 1 indicating a fault arc condition and u being an odd number.
CN201910739837.0A 2019-08-12 2019-08-12 Low-voltage distribution system series fault electric arc identification method based on full-phase deep learning Active CN110376497B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910739837.0A CN110376497B (en) 2019-08-12 2019-08-12 Low-voltage distribution system series fault electric arc identification method based on full-phase deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910739837.0A CN110376497B (en) 2019-08-12 2019-08-12 Low-voltage distribution system series fault electric arc identification method based on full-phase deep learning

Publications (2)

Publication Number Publication Date
CN110376497A CN110376497A (en) 2019-10-25
CN110376497B true CN110376497B (en) 2022-05-13

Family

ID=68258858

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910739837.0A Active CN110376497B (en) 2019-08-12 2019-08-12 Low-voltage distribution system series fault electric arc identification method based on full-phase deep learning

Country Status (1)

Country Link
CN (1) CN110376497B (en)

Families Citing this family (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111008673A (en) * 2019-12-24 2020-04-14 国网天津市电力公司 Method for collecting and extracting malignant data chain in power distribution network information physical system
CN111222255A (en) * 2020-02-25 2020-06-02 湖南世优电力科技股份有限公司 Big data simulation system and method for electrical fault identification
CN111398750B (en) * 2020-03-30 2021-09-28 深圳供电局有限公司 Arc identification method and system for arc identification
CN113049922B (en) 2020-04-22 2022-11-15 青岛鼎信通讯股份有限公司 Fault arc signal detection method adopting convolutional neural network
CN111610416B (en) * 2020-05-25 2021-12-07 南京航空航天大学 Series arc fault intelligent circuit breaker
CN112147462A (en) * 2020-09-16 2020-12-29 国网江西省电力有限公司电力科学研究院 Power transmission line fault identification method based on deep learning
CN112904156A (en) * 2021-01-19 2021-06-04 重庆邮电大学 Fault arc detection method based on frequency domain classification
CN112924751A (en) * 2021-01-27 2021-06-08 西安航天动力试验技术研究所 Method for detecting reliability of measuring line based on spectrum analysis
CN114034960A (en) * 2021-11-25 2022-02-11 国网重庆市电力公司电力科学研究院 Fault arc detection method based on weak arc characteristics

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103543375A (en) * 2013-08-26 2014-01-29 上海交通大学 Method for detecting alternating-current fault arcs on basis of wavelet transformation and time-domain hybrid features
CN106093569A (en) * 2016-05-30 2016-11-09 中国民用航空总局第二研究所 Navigation signal measuring method, system and navigator inline diagnosis method, system
CN106483374A (en) * 2016-09-27 2017-03-08 福州大学 A kind of harmonic wave harmonic detection method based on Nuttall double window whole phase FFT
CN106597022A (en) * 2016-12-23 2017-04-26 北京化工大学 Accelerometer dynamic model parameter identification method based on all-phase spectrum estimation
CN108646149A (en) * 2018-04-28 2018-10-12 国网江苏省电力有限公司苏州供电分公司 Fault electric arc recognition methods based on current characteristic extraction
CN109738771A (en) * 2019-03-12 2019-05-10 长春工程学院 Tandem type fault electric arc test platform and its test method based on deep learning

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103543375A (en) * 2013-08-26 2014-01-29 上海交通大学 Method for detecting alternating-current fault arcs on basis of wavelet transformation and time-domain hybrid features
CN106093569A (en) * 2016-05-30 2016-11-09 中国民用航空总局第二研究所 Navigation signal measuring method, system and navigator inline diagnosis method, system
CN106483374A (en) * 2016-09-27 2017-03-08 福州大学 A kind of harmonic wave harmonic detection method based on Nuttall double window whole phase FFT
CN106597022A (en) * 2016-12-23 2017-04-26 北京化工大学 Accelerometer dynamic model parameter identification method based on all-phase spectrum estimation
CN108646149A (en) * 2018-04-28 2018-10-12 国网江苏省电力有限公司苏州供电分公司 Fault electric arc recognition methods based on current characteristic extraction
CN109738771A (en) * 2019-03-12 2019-05-10 长春工程学院 Tandem type fault electric arc test platform and its test method based on deep learning

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
"利用Logistic二次回归法提高判别分析效率";张庆丰;《中国优秀硕士学位论文全文数据库 基础科学辑》;20070815;正文第3-12、27-29页 *
"消除外部环境因素干扰的MOA 在线监测参数修正方法";李海峰 等;《高电压技术》;20180831;第44卷(第8期);第2580-2586页 *

Also Published As

Publication number Publication date
CN110376497A (en) 2019-10-25

Similar Documents

Publication Publication Date Title
CN110376497B (en) Low-voltage distribution system series fault electric arc identification method based on full-phase deep learning
Wang et al. ArcNet: Series AC arc fault detection based on raw current and convolutional neural network
CN110488161A (en) A kind of detection of multi-load series arc faults and localization method
Chang et al. Load identification in neural networks for a non-intrusive monitoring of industrial electrical loads
CN110648088B (en) Electric energy quality disturbance source judgment method based on bird swarm algorithm and SVM
CN111308272B (en) Positioning method for low-current ground fault section
Li et al. Condition assessment of power transformers using a synthetic analysis method based on association rule and variable weight coefficients
CN105548739B (en) A kind of arrester operating state signal processing method
Eghtedarpour et al. Intelligent identification of flicker source in distribution systems
Huda et al. Power quality signals detection using S-transform
CN107543962A (en) The computational methods of leading m-Acetyl chlorophosphonazo spectrum distribution
CN110647924A (en) GIS equipment state evaluation method based on support vector description and K-nearest neighbor algorithm
Alshahrani et al. Evaluation and classification of power quality disturbances based on discrete wavelet transform and artificial neural networks
Gu et al. Identification method for low-voltage arc fault based on the loose combination of wavelet transformation and neural network
Gong et al. Series arc fault identification method based on wavelet transform and feature values decomposition fusion DNN
CN113094983B (en) Online simulation method for multi-dimensional time-varying characteristics of direct-current fault electric arc of photovoltaic system
Vahidi et al. An approach to detection of high impedance fault using discrete wavelet transform and artificial neural networks
Abidullah et al. Real-time power quality disturbances detection and classification system
CN110174602B (en) Nonlinear load series arc fault determination method and application
Rong et al. Detection of arc grounding fault based on the features of fault voltage
CN115144703B (en) High-resistance grounding fault identification method based on zero-sequence differential current and energy moment indexes
CN112684311B (en) Characteristic quantity extraction method for identifying oil paper insulation partial discharge type of transformer
Artale et al. DC series arc faults in PV systems. Detection methods and experimental characterization
Reddy et al. On-line monitoring of winding parameters for single-phase transformers
Jieqiu et al. Arc fault identification method based on fractal theory and SVM

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant