CN112485622A - Partial discharge pattern recognition method based on GA-BPNN - Google Patents

Partial discharge pattern recognition method based on GA-BPNN Download PDF

Info

Publication number
CN112485622A
CN112485622A CN202011435765.XA CN202011435765A CN112485622A CN 112485622 A CN112485622 A CN 112485622A CN 202011435765 A CN202011435765 A CN 202011435765A CN 112485622 A CN112485622 A CN 112485622A
Authority
CN
China
Prior art keywords
neural network
partial discharge
formula
bpnn
threshold
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.)
Pending
Application number
CN202011435765.XA
Other languages
Chinese (zh)
Inventor
白晓斌
米建伟
梁军科
郭峰
李林
刘子瑞
薛军
郑建康
景晓东
常江
王亮
杨军
李飞
蒋文龙
王新辉
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.)
Baoji Power Supply Co Of State Grid Shaanxi Electric Power Co
Xi'an Huapu Electric Instruments Manufacturing Co ltd
State Grid Corp of China SGCC
State Grid Shaanxi Electric Power Co Ltd
Original Assignee
Baoji Power Supply Co Of State Grid Shaanxi Electric Power Co
Xi'an Huapu Electric Instruments Manufacturing Co ltd
State Grid Corp of China SGCC
State Grid Shaanxi Electric Power 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 Baoji Power Supply Co Of State Grid Shaanxi Electric Power Co, Xi'an Huapu Electric Instruments Manufacturing Co ltd, State Grid Corp of China SGCC, State Grid Shaanxi Electric Power Co Ltd filed Critical Baoji Power Supply Co Of State Grid Shaanxi Electric Power Co
Priority to CN202011435765.XA priority Critical patent/CN112485622A/en
Publication of CN112485622A publication Critical patent/CN112485622A/en
Pending legal-status Critical Current

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
    • G01R31/1227—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 of components, parts or materials

Abstract

The invention discloses a partial discharge mode recognition method based on GA-BPNN, which comprises the steps of firstly, utilizing a statistical characteristic parameter method to extract characteristic values of a partial discharge phase distribution map, constructing a BP neural network model according to characteristic values of a plurality of groups of partial discharge signals for training, finally utilizing a genetic algorithm to obtain the optimal initial weight and threshold of the BP neural network, and performing training again and performing partial discharge mode recognition; the optimal weight and the threshold of the BP neural network are obtained by utilizing the characteristics that the genetic algorithm has strong global search capability and is not easy to fall into the local optimal solution and searching the optimal solution through simulating the natural evolution process, and the advantages of high convergence speed and high identification accuracy are achieved by utilizing the good coupling of GA and BPNN compared with the BPNN method, so that the problems that the BP neural network is easily influenced by the initial weight and the threshold and falls into the local optimal solution are solved, and the learning efficiency and the pattern identification accuracy are effectively improved.

Description

Partial discharge pattern recognition method based on GA-BPNN
Technical Field
The invention relates to an electrical equipment detection technology, in particular to a partial discharge pattern recognition method based on GA (genetic algorithm) -BPNN (back propagation neural network).
Background
With the development of modern industry and the improvement of urbanization level, in order to meet the increasing power demand, the scale of the power system is continuously increased, and the operation safety of the power system is more and more emphasized. High-voltage electrical equipment such as a switch cabinet, a transformer and the like is used as important equipment in a power grid system, and stable operation of the high-voltage electrical equipment is reliable guarantee for normal operation of the power system. The most important fault type of high-voltage electrical equipment is insulation fault, and the main reason for the generation of the fault is partial discharge phenomenon of an equipment insulation structure. The high-voltage electrical equipment is subjected to partial discharge online detection, so that the insulation state of the equipment can be judged in time, the insulation defect of the equipment can be found in advance, and the equipment is maintained in time to avoid sudden failure and unnecessary loss. The main reason for the insulation failure of high-voltage equipment is the partial discharge phenomenon, and the insulation medium can generate different partial discharge types according to the damaged position and type. The partial discharge mode identification is used as a key technology of partial discharge online detection, can effectively judge the insulation defect type of the high-voltage equipment, and provides reference for comprehensively judging the running state and the insulation performance of the high-voltage electrical equipment.
The traditional partial discharge mode identification is mainly carried out on-site judgment through expert experience, and is unsatisfactory in the aspects of efficiency and accuracy. With the wide spread and development of machine learning, an Artificial Neural Network (ANN) is widely applied to various industries, and in the field of high-voltage electricity, a BP neural network becomes a main partial discharge pattern recognition method, so that the efficiency and the recognition accuracy are improved compared with the traditional expert experience.
However, the method for identifying the partial discharge mode by using the BP neural network still has the following problems: firstly, the method is easily influenced by an initial weight and a threshold; secondly, local optimal solution is easy to fall into, so that learning fails; convergence is slow, and learning efficiency is low; resulting in a low recognition accuracy.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a partial discharge pattern recognition method based on GA-BPNN, which has the advantages of high convergence rate and high recognition accuracy and can accurately recognize the partial discharge type of the high-voltage electrical equipment.
In order to achieve the purpose, the invention adopts the following technical scheme:
the partial discharge pattern recognition method based on the GA-BPNN comprises the following steps:
(1) determining a BP neural network structure according to the characteristic parameters and the partial discharge type;
(2) setting BP neural network parameters;
(3) designing genetic algorithm parameters;
(4) according to the principle that the proportion of a small data sample training set in neural network training is larger than 2/3, 450 groups of 500 groups of partial discharge signal data samples are used as training data to train the neural network, 50 groups of partial discharge signal data samples are used as test data to test the pattern recognition capability of the network, and the data are normalized;
(5) training a BP neural network, and calculating the fitness value F of each individual in the genetic algorithm population by using a global error function Ep;
(6) screening individuals with the best fitness value through genetic operation, selection, crossing and mutation, and calculating the fitness value of a new individual;
(7) judging whether the end condition is met, if so, decoding to obtain the optimal BP neural network input layer and hidden layer interlayer connection weight omegaijThe hidden layer and the output layer are connected with a weight omegajkHidden layer node threshold ApAnd output layer node threshold BpOtherwise, returning to the genetic algorithm to continuously search the best fitness value individual;
(8) and retraining the BP neural network by using the obtained optimal connection weight and the threshold value, and carrying out partial discharge pattern recognition on the partial discharge test data.
Further, in the step (1), 16 statistical characteristic parameters are extracted from the PRPD map by using a statistical characteristic parameter method to serve as an input vector X of the BP neural networkpThree common discharge types of internal discharge, creeping discharge and corona discharge are adopted as the output vector Y of the neural networkpTherefore, the corresponding number of nodes N of the input layer is 16, and the number of nodes M of the output layer is M＝3；
Determining the value range of the number Q of the hidden layer nodes by referring to empirical formulas (7), (8) and (9);
Figure BDA0002827329300000031
Q<N-1 (8)
Q＝log2N (9)
in the formula (7), a is a constant between 0 and 10, the optimal number Q of hidden layer nodes is determined by firstly referring to a formula to determine the value range and the initial number of hidden layer nodes, and is determined by a trial and error method and a training result when a neural network model is trained, and finally the number of hidden layer nodes is determined to be 10, namely the BP neural network structure is 16-10-3.
Further, the characteristic parameter extraction step in the step (1) is as follows:
(1.1) statistical parameter transformation of PRPD map
Period of power frequency
Figure BDA0002827329300000032
Is divided into a number N of phase intervals,
Figure BDA0002827329300000033
for each phase interval, the width is
Figure BDA0002827329300000034
The probability of discharging quantity or discharging times in the phase interval i is PiMean and standard deviation are μ and σ, respectively;
(1.2) extracting Skewness Skewness, Kurtosis, local peak number Pn, cross-correlation coefficient C and phase asymmetry
Figure BDA0002827329300000035
The discharge quantity factor Q has 6 statistical characteristics, and 16 statistical characteristic parameters are determined in positive and negative semi-Cycle (+), Cycle (-) and Cycle (-) respectively;
Figure BDA0002827329300000036
Figure BDA0002827329300000041
wherein the Skewness Skewness reflects the phase distribution of partial discharge
Figure BDA0002827329300000042
And
Figure BDA00028273293000000413
the left and right deflection degree of the two-dimensional spectrogram in the positive and negative semi-circle distribution shape relative to the normal distribution is obtained by the formula (1):
Figure BDA0002827329300000043
kurtosis reflects the phase distribution of partial discharge
Figure BDA00028273293000000415
And
Figure BDA00028273293000000416
the convex degree of the two-dimensional spectrogram in the positive and negative semi-circle distribution shape relative to the normal distribution is obtained by a formula (2):
Figure BDA0002827329300000045
the number of local peaks Pn reflects the phase distribution of the partial discharge
Figure BDA0002827329300000046
And
Figure BDA00028273293000000417
the number of local peaks on the two-dimensional spectrogram profile is at any peak point
Figure BDA0002827329300000047
Whether the peak point is a local peak point can be judged by the formula (3);
Figure BDA0002827329300000048
the cross-correlation coefficient C is a coefficient reflecting the phase distribution of partial discharge
Figure BDA0002827329300000049
And
Figure BDA00028273293000000418
the similarity of the distribution shapes of the two-dimensional spectrogram in the positive and negative half cycles is obtained by the formula (4):
Figure BDA00028273293000000410
in the formula (4), the reaction mixture is,
Figure BDA00028273293000000411
and
Figure BDA00028273293000000419
as a spectrogram in a phase interval
Figure BDA00028273293000000412
Average partial discharge amount;
degree of phase asymmetry
Figure BDA0002827329300000051
Reflecting the phase distribution of partial discharge
Figure BDA0002827329300000052
And
Figure BDA0002827329300000053
the difference of the two-dimensional spectrogram in the initial phases of the positive and negative half cycles is obtained by the formula (5):
Figure BDA0002827329300000054
in the formula (5), the reaction mixture is,
Figure BDA0002827329300000055
and
Figure BDA0002827329300000056
respectively the initial phases of the spectrogram in positive and negative half cycles;
discharge magnitude factor Q reflects partial discharge phase distribution
Figure BDA0002827329300000057
The difference of the discharge capacity of the three-dimensional spectrogram in positive and negative half cycles is obtained through a formula (6):
Figure BDA0002827329300000058
in the formula (6), the reaction mixture is,
Figure BDA0002827329300000059
and
Figure BDA00028273293000000510
respectively spectrogram in phase interval
Figure BDA00028273293000000511
The number of discharges of (c).
Further, determining an initial hidden layer threshold A in step (2)pAnd output layer threshold BpInitializing connection weights ω between input layer, hidden layer, and output layer neuronsij、ωjkGiving a learning rate eta and an error precision epsilon, and adopting a Sigmoid function to select an excitation function for the neural network, such as a formula (10);
Figure BDA00028273293000000512
in the formula (10), f (x) is a Sigmoid function, x is an argument, and e is a natural constant.
Further, the step (3) comprises:
coding string length: according to the target to be optimized, the individuals of the genetic algorithm are all the weights and thresholds of the whole BP neural network, and the connection weight omega from the input layer node to the hidden layer nodeijConnection weight omega from hidden layer node to output layer nodejkAnd hidden layer node threshold ApAnd output layer node threshold BpComposition is carried out;
the coding method comprises the following steps: a real number coding method is adopted, each real number takes an integer and five decimal places to ensure that the precision is all accurate to 0.00001, the value range of the weight is (-5,5), the value range of the threshold value is (-3,3), the weight and the threshold value of the neural network are connected in series according to a certain sequence, and each position on a coding string corresponds to the corresponding weight and the corresponding threshold value;
the operation parameters are as follows: initializing running parameters of genetic algorithm, including population size M, genetic algebra G and cross probability PcAnd the mutation probability Pm；
Fitness function: calculating an individual fitness value F of the genetic algorithm by using a global error function Ep of the BP neural network;
Figure BDA0002827329300000061
in formula (11), OkFor the calculation output of the k-th output node of the neural network, ypkA is a coefficient for the expected output of the corresponding node of the neural network; the fitness value of the genetic algorithm is in inverse proportional relation with the global error of the BP neural network, namely the smaller the global error is, the better the individual fitness is, and screening is carried out through the fitness value.
Further, according to the three-layer BP neural network structure 16-10-3 identified by the partial discharge pattern, 190 weight values and 13 threshold values are totally obtained, and the individual coding length of one genetic algorithm is 203.
Compared with the prior art, the invention has the following beneficial technical effects:
the method comprises the steps of firstly, extracting characteristic values of a Partial Discharge Phase distribution (PRPD) map by using a statistical characteristic parameter method, constructing a BP neural network model according to a plurality of groups of Partial Discharge signal characteristic values for training, and finally, obtaining the optimal initial weight and threshold of the BP neural network by using a genetic algorithm for training again and identifying a Partial Discharge mode; the optimal weight and the threshold of the BP neural network are obtained by utilizing the characteristics that the genetic algorithm has strong global search capability and is not easy to fall into the local optimal solution and searching the optimal solution through simulating the natural evolution process, and the advantages of high convergence speed and high identification accuracy are achieved by utilizing the good coupling of GA and BPNN compared with the BPNN method, so that the problems that the BP neural network is easily influenced by the initial weight and the threshold and falls into the local optimal solution are solved, and the learning efficiency and the pattern identification accuracy are effectively improved.
Specifically, the method comprises the following steps: (1) the BP neural network is used for carrying out pattern recognition on the partial discharge signals with the characteristic parameters and the discharge types of the partial discharge signals having the mapping relation, and compared with a traditional expert experience pattern recognition method, the method has higher recognition rate and efficiency. (2) A genetic algorithm is utilized to optimize the BP neural network, the optimal initial weight and threshold of the neural network are screened and determined, the dependency of the BP neural network identification performance on the initial value is overcome, the convergence speed is high, and the data processing capacity is high. (3) The algorithm can be applied to partial discharge online detection, and the insulation state of the high-voltage electrical equipment can be effectively judged by combining partial discharge online detection methods such as ultrasonic detection and earth electric wave detection.
Drawings
FIG. 1 is a GA-BPNN flow chart;
FIG. 2 is a graph of the number of iterations of the BPNN training;
FIG. 3 is a graph of the number of iterations of GA-BPNN training;
FIG. 4 is a BPNN identification error graph;
FIG. 5 is a graph of GA-BPNN recognition error.
Detailed Description
The present invention will be described in further detail with reference to the following examples, which are not intended to limit the invention thereto.
The invention provides a BP Neural Network (BPNN) partial discharge signal pattern recognition method based on Genetic Algorithm (GA) optimization, which is used for recognizing the type of partial discharge so as to judge the type of an insulation defect.
Because the time domain waveform and the time-frequency map data volume of the partial discharge signal are huge, it is very difficult to directly perform the pattern recognition on the partial discharge signal. Therefore, in order to effectively perform pattern recognition, the nature and characteristics of the partial discharge signal are characterized by using a small number of parameters, and it is necessary to extract characteristic parameters of the partial discharge signal. The common partial discharge characteristic parameter extraction method comprises the following steps: statistical characteristic parameters, waveform characteristic parameters, Weibull parameters, typing characteristic parameters and distance characteristic parameters.
The invention adopts a statistical characteristic parameter method to extract characteristic parameters of a local discharge signal, and the basic steps of characteristic value extraction are as follows:
(1) and carrying out statistical parameter transformation on the PRPD map. Period of power frequency
Figure BDA0002827329300000081
Is divided into a number N of phase intervals,
Figure BDA0002827329300000082
for each phase interval, the width is
Figure BDA0002827329300000083
The probability of discharging quantity or discharging times in the phase interval i is PiMean and standard deviation are μ and σ, respectively;
(2) extracting Skewness Skewness, Kurtosis, local peak number Pn, cross-correlation coefficient C and phase asymmetry
Figure BDA0002827329300000084
And the discharge quantity factor Q has 6 statistical characteristics, and 16 statistical characteristic parameters are determined in positive and negative half cycles (+), and cycles (-).
TABLE 1 statistical characteristic parameters
Figure BDA0002827329300000085
Table 1 is a table of statistical characteristic parameters, in which Skewness Skewness reflects the phase distribution of partial discharge
Figure BDA0002827329300000086
And
Figure BDA0002827329300000087
the left and right deviation degree of the distribution shape of the two-dimensional spectrogram in positive and negative half cycles relative to the normal distribution.
Figure BDA0002827329300000088
Kurtosis reflects the phase distribution of partial discharge
Figure BDA0002827329300000089
And
Figure BDA00028273293000000810
the convex degree of the two-dimensional spectrogram in the positive and negative semi-circle distribution shape relative to the normal distribution is realized.
Figure BDA0002827329300000091
The number of local peaks Pn reflects the phase distribution of the partial discharge
Figure BDA0002827329300000092
And
Figure BDA00028273293000000917
the number of local peaks on the two-dimensional spectrogram profile. At any peak point
Figure BDA0002827329300000093
Whether the peak point is a local peak point can be determined by equation (3).
Figure BDA00028273293000000919
And is
Figure BDA0002827329300000094
The cross-correlation coefficient C is a coefficient reflecting the phase distribution of partial discharge
Figure BDA0002827329300000095
And
Figure BDA00028273293000000918
the similarity of the distribution shapes of the two-dimensional spectrogram in the positive and negative half circles is determined.
Figure BDA0002827329300000096
In the formula (4), the reaction mixture is,
Figure BDA0002827329300000097
and
Figure BDA00028273293000000920
as a spectrogram in a phase interval
Figure BDA0002827329300000098
Average partial discharge amount.
Degree of phase asymmetry
Figure BDA0002827329300000099
Reflecting the phase distribution of partial discharge
Figure BDA00028273293000000910
And
Figure BDA00028273293000000921
the difference of the two-dimensional spectrogram in the initial phase of positive and negative half cycles.
Figure BDA00028273293000000911
In the formula (5), the reaction mixture is,
Figure BDA00028273293000000912
and
Figure BDA00028273293000000922
the initial phases of the spectra in the positive and negative half cycles, respectively.
Discharge magnitude factor Q reflects partial discharge phase distribution
Figure BDA00028273293000000913
Difference of discharge amount of the three-dimensional spectrogram in positive and negative half cycles.
Figure BDA00028273293000000914
In the formula (6), the reaction mixture is,
Figure BDA00028273293000000915
and
Figure BDA00028273293000000923
respectively spectrogram in phase interval
Figure BDA00028273293000000916
The number of discharges of (c).
According to FIG. 1, the basic steps of GA-BPNN are as follows:
(1) and determining the BP neural network structure according to the characteristic parameters and the partial discharge type. 16 statistical characteristic parameters are extracted from the PRPD map by using a statistical characteristic parameter method and are used as input vectors X of the BP neural networkpThree common discharge types: internal discharge, creeping discharge, corona discharge as output vector Y of neural networkpTherefore, the corresponding number of input layer nodes N is 16, and the number of output layer nodes M is 3.
For the selection of the number Q of hidden layer nodes, there is no unified determination method theoretically at present, and the value range of the number of hidden layer nodes is generally determined by referring to empirical formulas (7), (8) and (9).
Figure BDA0002827329300000101
Q<N-1 (8)
Q＝log2N (9)
In the formula (7), a is a constant between 0 and 10. The optimal number Q of hidden layer nodes is firstly determined by a reference formula to obtain a value range and an initial number of hidden layer nodes, is determined by a trial and error method and a training result when a neural network model is trained, and is finally determined to be 10, namely the BP neural network structure is 16-10-3.
(2) Setting BP neural network parameters and determining an initial hidden layer threshold ApAnd output layer threshold BpInitializing connection weights ω between input layer, hidden layer, and output layer neuronsij、ωjkAnd determining a neuron excitation function by giving a learning rate eta and an error precision epsilon, wherein the excitation function selected by the neural network is a Sigmoid function.
Figure BDA0002827329300000102
(3) Designing genetic algorithm parameters:
coding string length: according to the target to be optimized, the individuals of the genetic algorithm are all the weights and thresholds of the whole BP neural network, and the connection weight omega from the input layer node to the hidden layer nodeijConnection weight omega from hidden layer node to output layer nodejkAnd hidden layer node threshold ApAnd output layer node threshold BpAnd (4) forming. According to the three-layer BP neural network structure 16-10-3 identified by the partial discharge mode, 190 weight values and 13 threshold values are totally arranged, and the individual coding length of a genetic algorithm is 203.
The coding method comprises the following steps: common individual coding methods are binary coding, gray code coding, real number coding, and the like. The method for selecting real number coding in the GA-BP neural network has the following advantages: the method has high precision, is suitable for the problem of continuous variables, and avoids the problem of Hamming cliff; the method is suitable for representing numerical values with a large range and is suitable for parallel search in a large space; and thirdly, the mixed use of the genetic algorithm and the classical optimization method is facilitated. Each real number takes an integer and five decimal places to ensure that the precision is all accurate to 0.00001. The value range of the weight is (-5,5), the value range of the threshold is (-3,3), the weight and the threshold of the neural network are connected in series according to a certain sequence, and each position on the coding string corresponds to the corresponding weight and the corresponding threshold.
The operation parameters are as follows: initializing running parameters of genetic algorithm, including population size M, genetic algebra G and cross probability PcAnd the mutation probability Pm。
Fitness function: in order to obtain the optimal weight and threshold, an individual fitness value F of the genetic algorithm is calculated by using a global error function Ep of the BP neural network.
Figure BDA0002827329300000111
In formula (11), OkFor the calculation output of the k-th output node of the neural network, ypkOutputs are expected for respective nodes of the neural network. a is a coefficient. The fitness value of the genetic algorithm is in inverse proportional relation with the global error of the BP neural network, namely the smaller the global error is, the better the individual fitness is, and screening is carried out through the fitness value.
(4) According to the principle that the proportion of a small data sample training set in the neural network training is larger than 2/3, 450 groups of 500 groups of partial discharge signal data samples are used as training data to train the neural network, 50 groups of partial discharge signal data samples are used as test data to test the pattern recognition capability of the network, and the data are normalized.
(5) And training the BP neural network, and calculating the fitness value F of each individual in the genetic algorithm population by using the global error function Ep.
(6) Through genetic operation, selection, crossover and mutation, individuals with the best fitness value are screened out, and a new individual fitness value is calculated.
(7) Judging whether the end condition is satisfied, if it is fullIf the requirement is satisfied, decoding is carried out to obtain the optimal BP neural network input layer and hidden layer interlayer connection weight omegaijThe hidden layer and the output layer are connected with a weight omegajkHidden layer node threshold ApAnd output layer node threshold Bp. Otherwise, returning to the genetic algorithm to continuously search for the best fitness value individual.
(8) And training a BP neural network again by using the obtained optimal connection weight and the threshold value, and carrying out partial discharge pattern recognition on the partial discharge test data.
The effectiveness of the method of the invention is verified experimentally as follows:
and respectively carrying out partial discharge pattern recognition on 50 groups of test data by using the trained BPNN and GA-BPNN. The BPNN and GA-BPNN training iterations are shown in FIG. 2 and FIG. 3, and have faster convergence speed compared with the BPNN, GA-BPNN. The BPNN and GA-BPNN partial discharge identification error pairs are shown in FIGS. 4 and 5, and the identification capability pairs are shown in Table 2.
TABLE 2 BPNN vs. GA-BPNN network identification capability
Figure BDA0002827329300000121
The comparison of the two neural network identification capacities can be obtained, the BPNN has an unobvious mapping relation, high requirement on error precision and more iteration times of complex partial discharge signals, and the BP neural network pattern identification capacity and generalization capacity trained by randomly generating a threshold and a weight are weak and have low accuracy. The GA-BPNN determines the optimal weight and threshold through a genetic algorithm, has stronger pattern recognition capability on the local discharge type, has better pattern recognition and generalization capability, and better solves the problem that the BP neural network is easily influenced by the initial connection weight and threshold and is easily trapped into local optimization.
The present invention is described in detail with reference to the above embodiments, and those skilled in the art will understand that: modifications and equivalents may be made to the embodiments of the invention without departing from the spirit and scope of the invention, which is to be covered by the claims.

Claims (6)

1. The partial discharge pattern recognition method based on the GA-BPNN is characterized by comprising the following steps of:
(1) determining a BP neural network structure according to the characteristic parameters and the partial discharge type;
(2) setting BP neural network parameters;
(3) designing genetic algorithm parameters;
(4) according to the principle that the proportion of a small data sample training set in neural network training is larger than 2/3, 450 groups of 500 groups of partial discharge signal data samples are used as training data to train the neural network, 50 groups of partial discharge signal data samples are used as test data to test the pattern recognition capability of the network, and the data are normalized;
(5) training a BP neural network, and calculating the fitness value F of each individual in the genetic algorithm population by using a global error function Ep;
(6) screening individuals with the best fitness value through genetic operation, selection, crossing and mutation, and calculating the fitness value of a new individual;
(7) judging whether the end condition is met, if so, decoding to obtain the optimal BP neural network input layer and hidden layer interlayer connection weight omegaijThe hidden layer and the output layer are connected with a weight omegajkHidden layer node threshold ApAnd output layer node threshold BpOtherwise, returning to the genetic algorithm to continuously search the best fitness value individual;
(8) and retraining the BP neural network by using the obtained optimal connection weight and the threshold value, and carrying out partial discharge pattern recognition on the partial discharge test data.
2. The GA-BPNN-based partial discharge pattern recognition method of claim 1, wherein in step (1), 16 statistical feature parameters are extracted from PRPD map by statistical feature parameter method and used as input vector X of BP neural networkpThree common discharge types of internal discharge, creeping discharge and corona discharge are adopted as the output vector Y of the neural networkpThus correspond toThe number of nodes N of the input layer is 16, and the number of nodes M of the output layer is 3;
determining the value range of the number Q of the hidden layer nodes by referring to empirical formulas (7), (8) and (9);
Figure FDA0002827329290000021
Q<N-1 (8)
Q＝log2N (9)
in the formula (7), a is a constant between 0 and 10, the optimal number Q of hidden layer nodes is determined by firstly referring to a formula to determine the value range and the initial number of hidden layer nodes, and is determined by a trial and error method and a training result when a neural network model is trained, and finally the number of hidden layer nodes is determined to be 10, namely the BP neural network structure is 16-10-3.
3. A GA-BPNN-based partial discharge pattern recognition method according to claim 2, wherein the feature parameter extraction step in step (1) is as follows:
(1.1) statistical parameter transformation of PRPD map
Period of power frequency
Figure FDA0002827329290000023
Is divided into a number N of phase intervals,
Figure FDA0002827329290000024
for each phase interval, the width is
Figure FDA0002827329290000025
The probability of discharging quantity or discharging times in the phase interval i is PiMean and standard deviation are μ and σ, respectively;
(1.2) extracting Skewness Skewness, Kurtosis, local peak number Pn, cross-correlation coefficient C and phase asymmetry
Figure FDA0002827329290000026
The discharge quantity factor Q has 6 statistical characteristics, and 16 statistical characteristic parameters are determined in positive and negative semi-cycles Cycle (+), Cycle (-) and Cycle (-) respectively;
Figure FDA0002827329290000022
Figure FDA0002827329290000031
wherein the Skewness Skewness reflects the phase distribution of partial discharge
Figure FDA0002827329290000037
And
Figure FDA0002827329290000038
the left and right deflection degree of the two-dimensional spectrogram in the positive and negative semi-circle distribution shape relative to the normal distribution is obtained by the formula (1):
Figure FDA0002827329290000032
kurtosis reflects the phase distribution of partial discharge
Figure FDA0002827329290000039
And
Figure FDA00028273292900000310
the convex degree of the two-dimensional spectrogram in the positive and negative semi-circle distribution shape relative to the normal distribution is obtained by a formula (2):
Figure FDA0002827329290000033
the number of local peaks Pn reflects the phase distribution of the partial discharge
Figure FDA00028273292900000311
And
Figure FDA00028273292900000312
the number of local peaks on the two-dimensional spectrogram profile is at any peak point
Figure FDA00028273292900000313
Whether the peak point is a local peak point can be judged by the formula (3);
Figure FDA0002827329290000034
the cross-correlation coefficient C is a coefficient reflecting the phase distribution of partial discharge
Figure FDA00028273292900000314
And
Figure FDA00028273292900000315
the similarity of the distribution shapes of the two-dimensional spectrogram in the positive and negative half cycles is obtained by the formula (4):
Figure FDA0002827329290000035
in the formula (4), the reaction mixture is,
Figure FDA00028273292900000320
and
Figure FDA00028273292900000321
as a spectrogram in a phase interval
Figure FDA00028273292900000316
Average partial discharge amount;
degree of phase asymmetry
Figure FDA00028273292900000317
Reflecting the phase distribution of partial discharge
Figure FDA00028273292900000318
And
Figure FDA00028273292900000319
the difference of the two-dimensional spectrogram in the initial phases of the positive and negative half cycles is obtained by the formula (5):
Figure FDA0002827329290000036
in the formula (5), the reaction mixture is,
Figure FDA0002827329290000043
and
Figure FDA0002827329290000044
respectively the initial phases of the spectrogram in positive and negative half cycles;
discharge magnitude factor Q reflects partial discharge phase distribution
Figure FDA0002827329290000045
The difference of the discharge capacity of the three-dimensional spectrogram in positive and negative half cycles is obtained through a formula (6):
Figure FDA0002827329290000041
in the formula (6), the reaction mixture is,
Figure FDA0002827329290000047
and
Figure FDA0002827329290000048
respectively spectrogram in phase interval
Figure FDA0002827329290000046
The number of discharges of (c).
4. The GA-BPNN-based partial discharge pattern recognition method of claim 2, wherein the initial hidden layer threshold A is determined in step (2)pAnd output layer threshold BpInitializing connection weights ω between input layer, hidden layer, and output layer neuronsij、ωjkGiving a learning rate eta and an error precision epsilon, and adopting a Sigmoid function to select an excitation function for the neural network, such as a formula (10);
Figure FDA0002827329290000042
in the formula (10), f (x) is a Sigmoid function, x is an argument, and e is a natural constant.
5. A GA-BPNN-based partial discharge pattern recognition method according to claim 4, wherein step (3) comprises:
coding string length: according to the target to be optimized, the individuals of the genetic algorithm are all the weights and thresholds of the whole BP neural network, and the connection weight omega from the input layer node to the hidden layer nodeijConnection weight omega from hidden layer node to output layer nodejkAnd hidden layer node threshold ApAnd output layer node threshold BpComposition is carried out;
the coding method comprises the following steps: a real number coding method is adopted, each real number takes an integer and five decimal places to ensure that the precision is all accurate to 0.00001, the value range of the weight is (-5,5), the value range of the threshold value is (-3,3), the weight and the threshold value of the neural network are connected in series according to a certain sequence, and each position on a coding string corresponds to the corresponding weight and the corresponding threshold value;
the operation parameters are as follows: initializing running parameters of genetic algorithm, including population size M, genetic algebra G and cross probability PcAnd the mutation probability Pm；
Fitness function: calculating an individual fitness value F of the genetic algorithm by using a global error function Ep of the BP neural network;
Figure FDA0002827329290000051
in formula (11), OkFor the calculation output of the k-th output node of the neural network, ypkA is a coefficient for the expected output of the corresponding node of the neural network; the fitness value of the genetic algorithm is in inverse proportional relation with the global error of the BP neural network, namely the smaller the global error is, the better the individual fitness is, and screening is carried out through the fitness value.
6. The GA-BPNN-based partial discharge pattern recognition method of claim 5, wherein according to the three-layer BP neural network structure 16-10-3 of partial discharge pattern recognition, there are 190 weights and 13 thresholds, and then the individual code length of one genetic algorithm is 203.
CN202011435765.XA 2020-12-09 2020-12-09 Partial discharge pattern recognition method based on GA-BPNN Pending CN112485622A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011435765.XA CN112485622A (en) 2020-12-09 2020-12-09 Partial discharge pattern recognition method based on GA-BPNN

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011435765.XA CN112485622A (en) 2020-12-09 2020-12-09 Partial discharge pattern recognition method based on GA-BPNN

Publications (1)

Publication Number Publication Date
CN112485622A true CN112485622A (en) 2021-03-12

Family

ID=74941031

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011435765.XA Pending CN112485622A (en) 2020-12-09 2020-12-09 Partial discharge pattern recognition method based on GA-BPNN

Country Status (1)

Country Link
CN (1) CN112485622A (en)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113064032A (en) * 2021-03-26 2021-07-02 云南电网有限责任公司电力科学研究院 Partial discharge mode identification method based on map features and information fusion
CN113283579A (en) * 2021-04-27 2021-08-20 国网山东省电力公司烟台供电公司 Method for identifying partial discharge fault type through convolutional neural network
CN114047413A (en) * 2021-11-11 2022-02-15 国网上海市电力公司 GIS partial discharge identification method and system based on MIV and MEA-LVQ neural network
CN114113942A (en) * 2021-11-25 2022-03-01 国网四川省电力公司成都供电公司 Oil paper insulation partial discharge identification and characteristic correction method and experimental device
CN114217179A (en) * 2021-11-15 2022-03-22 国网辽宁省电力有限公司锦州供电公司 High-voltage switch cabinet partial discharge monitoring method

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102809718A (en) * 2012-07-25 2012-12-05 华南理工大学 Ultra-high-frequency partial discharge signal identification method for gas insulated switchgear (GIS)
CN103177288A (en) * 2013-03-05 2013-06-26 辽宁省电力有限公司鞍山供电公司 Transformer fault diagnosis method based on genetic algorithm optimization neural network
CN106503802A (en) * 2016-10-20 2017-03-15 上海电机学院 A kind of method of utilization genetic algorithm optimization BP neural network system
CN106503689A (en) * 2016-11-14 2017-03-15 哈尔滨理工大学 Neutral net local discharge signal mode identification method based on particle cluster algorithm
CN109034388A (en) * 2018-07-27 2018-12-18 湖北中烟工业有限责任公司 A kind of prediction model of cigarette material and mainstream smoke constituents based on Genetic Algorithm Optimized Neural Network
CN109459671A (en) * 2018-09-27 2019-03-12 国网浙江省电力有限公司湖州供电公司 A kind of switch cabinet state monitoring method of the improvement neural network based on genetic algorithm
CN109596955A (en) * 2018-12-30 2019-04-09 国网北京市电力公司 Shelf depreciation state determines method and device
CN109683066A (en) * 2018-11-08 2019-04-26 山东康威通信技术股份有限公司 Power cable typical defect local discharge signal recognition methods

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102809718A (en) * 2012-07-25 2012-12-05 华南理工大学 Ultra-high-frequency partial discharge signal identification method for gas insulated switchgear (GIS)
CN103177288A (en) * 2013-03-05 2013-06-26 辽宁省电力有限公司鞍山供电公司 Transformer fault diagnosis method based on genetic algorithm optimization neural network
CN106503802A (en) * 2016-10-20 2017-03-15 上海电机学院 A kind of method of utilization genetic algorithm optimization BP neural network system
CN106503689A (en) * 2016-11-14 2017-03-15 哈尔滨理工大学 Neutral net local discharge signal mode identification method based on particle cluster algorithm
CN109034388A (en) * 2018-07-27 2018-12-18 湖北中烟工业有限责任公司 A kind of prediction model of cigarette material and mainstream smoke constituents based on Genetic Algorithm Optimized Neural Network
CN109459671A (en) * 2018-09-27 2019-03-12 国网浙江省电力有限公司湖州供电公司 A kind of switch cabinet state monitoring method of the improvement neural network based on genetic algorithm
CN109683066A (en) * 2018-11-08 2019-04-26 山东康威通信技术股份有限公司 Power cable typical defect local discharge signal recognition methods
CN109596955A (en) * 2018-12-30 2019-04-09 国网北京市电力公司 Shelf depreciation state determines method and device

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
吕启深 等: "GA-BP神经网络在GIS典型缺陷局部放电识别中的应用", 电瓷避雷器, no. 04, pages 34 - 38 *
王国利 等: "AGA-BP神经网络用于变压器超高频局部放电模式识别", 电工电能新技术, vol. 22, no. 02, pages 6 - 9 *
田质广 等: "基于遗传算法的神经网络在发电机定子超高频局部放电模式识别中的应用", 大电机技术, no. 01, pages 36 - 40 *
苑津莎 等: "基于相关系数矩阵和概率神经网络的局部放电模式识别", 电力系统保护与控制, vol. 41, no. 13, pages 110 - 115 *
蒋雄伟 等: "基于遗传算法的神经网络在局部放电模式识别中的应用", 西安交通大学学报, vol. 33, no. 12, pages 1 - 4 *

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113064032A (en) * 2021-03-26 2021-07-02 云南电网有限责任公司电力科学研究院 Partial discharge mode identification method based on map features and information fusion
CN113064032B (en) * 2021-03-26 2022-08-02 云南电网有限责任公司电力科学研究院 Partial discharge mode identification method based on map features and information fusion
CN113283579A (en) * 2021-04-27 2021-08-20 国网山东省电力公司烟台供电公司 Method for identifying partial discharge fault type through convolutional neural network
CN114047413A (en) * 2021-11-11 2022-02-15 国网上海市电力公司 GIS partial discharge identification method and system based on MIV and MEA-LVQ neural network
CN114217179A (en) * 2021-11-15 2022-03-22 国网辽宁省电力有限公司锦州供电公司 High-voltage switch cabinet partial discharge monitoring method
CN114113942A (en) * 2021-11-25 2022-03-01 国网四川省电力公司成都供电公司 Oil paper insulation partial discharge identification and characteristic correction method and experimental device
CN114113942B (en) * 2021-11-25 2023-11-10 国网四川省电力公司成都供电公司 Recognition and characteristic correction method and experimental device for oil paper insulation partial discharge

Similar Documents

Publication Publication Date Title
CN112485622A (en) Partial discharge pattern recognition method based on GA-BPNN
CN109271975B (en) Power quality disturbance identification method based on big data multi-feature extraction collaborative classification
CN109978079A (en) A kind of data cleaning method of improved storehouse noise reduction self-encoding encoder
CN105388402A (en) Cross-linked cable partial discharge mode recognition method based on parameter optimization SVM (Support Vector Machine) algorithm
Zhang et al. Semisupervised momentum prototype network for gearbox fault diagnosis under limited labeled samples
CN110111606A (en) A kind of vessel traffic flow prediction technique based on EEMD-IAGA-BP neural network
CN110263834A (en) A kind of detection method of new energy power quality exceptional value
Zhao et al. A hybrid learning method for constructing compact rule-based fuzzy models
CN111126560A (en) Method for optimizing BP neural network based on cloud genetic algorithm
Liang et al. Dissolved gas analysis of transformer oil based on Deep Belief Networks
CN110348489B (en) Transformer partial discharge mode identification method based on self-coding network
CN108760813B (en) Gas turbine blade health monitoring system and method based on temperature signals
CN116561569A (en) Industrial power load identification method based on EO feature selection and AdaBoost algorithm
CN115563500A (en) Power distribution equipment partial discharge mode identification method, device and system based on data enhancement technology
CN115409317A (en) Transformer area line loss detection method and device based on feature selection and machine learning
CN113762591A (en) Short-term electric quantity prediction method and system based on GRU and multi-core SVM counterstudy
CN111461214A (en) Automatic fault diagnosis method for insulated pipe bus based on random forest algorithm
CN116087692B (en) Distribution network tree line discharge fault identification method, system, terminal and medium
CN114943430B (en) Smart grid key node identification method based on deep reinforcement learning
CN110852597B (en) Electricity consumption peak period resident load ratio calculation method based on generation of countermeasure network
Tian et al. Linear Load Arc Fault Identification Based on Neural Network
CN110032981B (en) Rotary machine fault identification method based on improved support vector machine
jun Wang et al. Application of Adaptive Genetic BP Model in Classification of Speech Characteristic Signal
Milani et al. A coevolutionary optimization approach with deep sparse autoencoder for the extraction of equipment degradation indicators
Jami’in et al. Wind Speed Time Series Modeling under Least Square Error and Genetic Algorithm

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