GIS equipment PD type determination method using improved marine predator algorithm to optimize SVM

Abstract

The application relates to the technical field of GIS equipment PD type determination, and particularly discloses a GIS equipment PD type determination method using an improved marine predator algorithm to optimize SVM. On one hand, the initial population of the original marine predator algorithm is randomly generated, which may cause uneven distribution of the initial population individual positions, reduce population diversity and affect the optimization speed; therefore, a chaos mapping strategy and a reverse learning mechanism are introduced to improve the initial population quality. On the other hand, the original marine predator algorithm has defects such as being prone to falling into local optimization and slow convergence speed; therefore, a multi-subpopulation disturbance idea is proposed. The prey population of the marine predator disturbed by the FADs effect is divided into two subpopulations according to the fitness function value; the subpopulation with a higher fitness function value is disturbed according to the self-adaptive Cauchy mutation, and the other subpopulation is subjected to differential evolution, so as to generate a multi-subpopulation disturbance solution which is brought into the next iteration.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of partial discharge (PD) type identification of gas insulated switchgear (GIS), and particularly relates to a GIS PD type identification method using an improved marine predator algorithm to optimize SVM. BACKGROUND

[0002] Gas Insulated Switchgear (GIS) refers to a metal-enclosed device that uses sulfur hexafluoride (SF6) gas as an insulating medium, either entirely or partially. These devices include disconnectors, circuit breakers, grounding switches, busbars, transformers, connecting pieces, arresters, and outgoing terminal ends, which are all encapsulated in a metal cavity, and the cavity is filled with SF6 gas with excellent insulating and arc extinguishing performance as an insulating and arc extinguishing medium. Therefore, GIS is also known as SF6 fully enclosed switchgear. Since the industrial application of GIS in the mid-20th century, it has been widely used in power transmission and transformation systems, and is widely used in high-voltage and ultra-high-voltage power grids, and its application in the field of extra-high voltage is also increasing. Various insulation defects inevitably exist inside GIS, which distort the electric field around them. In some places, the electric field is greatly enhanced, exceeding the breakdown field strength of the insulating medium, resulting in PD (partial discharge). Long-term development of PD may lead to serious insulation faults. On the one hand, PD is a sign of insulation deterioration, and by detecting PD, the insulation condition of GIS can be known, thereby providing a reference for the maintenance of GIS and preventing and stopping serious accidents. At the same time, PD is an important reason for insulation deterioration and must be taken seriously. Detecting PD in GIS has important practical significance for ensuring the safe and reliable operation of GIS.

[0003] Commonly used PD type identification methods for gas insulated switchgear include decision trees, neural networks, and support vector machines. The internal structure of the identifier constructed based on the principle of decision trees is clear, and through decision tree classification, the entire pattern recognition process can be clearly seen, making it easier to understand the relationship between input and output. However, in the identification of complex objects, the recognition ability of decision trees is not as good as that of neural networks and support vector machines. Support vector machines, as an intelligent recognition algorithm, can establish a highly nonlinear classification plane, greatly improving the recognition accuracy compared to traditional recognition algorithms. This recognition has a large amount of calculation and is suitable for computer-based automated recognition. The initial population of the original marine predator algorithm is randomly generated, which may lead to uneven distribution of individual positions in the initial population, reducing population diversity and affecting optimization speed. Moreover, the original marine predator algorithm has defects such as being prone to local optimization and slow convergence speed. SUMMARY

[0004] The application provides a GIS equipment PD type determination method using an improved marine predator algorithm to optimize an SVM, and solves the technical problem that the initial population of the original marine predator algorithm is randomly generated, which may cause uneven distribution of the initial population individual positions, reduce population diversity and affect the optimization speed; and the original marine predator algorithm has defects such as being prone to falling into local optimization and slow convergence speed.

[0005] To solve the above technical problems, the application provides a GIS equipment PD type determination method using an improved marine predator algorithm to optimize an SVM, and the method comprises the following steps:

[0006] S1, sample data is acquired, and the sample data is preprocessed to obtain a sample set;

[0007] S2, an original ecological marine predator algorithm is improved to obtain an improved marine predator algorithm; the improved marine predator algorithm comprises the following steps:

[0008] S21, an initial population is generated based on a chaotic mapping strategy and a reverse learning mechanism;

[0009] S22, an initial prey matrix and an elite predator matrix are constructed according to the initial population;

[0010] S23, the initial prey matrix and the elite predator matrix are iterated and optimized according to the iteration mode of the individual iteration determined according to the current iteration number;

[0011] S24, the prey population of the marine predator is disturbed by applying the FADs effect;

[0012] S25, the prey population of the marine predator disturbed by the FADs effect is divided into two subpopulations according to the average value of the fitness function, the subpopulation with a higher fitness function value is disturbed by adaptive Cauchy mutation, and the other subpopulation is subjected to differential evolution, to generate a multi-subpopulation disturbance solution, which is brought into the next iteration;

[0013] S26, the iteration of the high-speed ratio, the constant-speed ratio and the low-speed ratio stages is completed in sequence according to the algorithm time sequence architecture;

[0014] S27, when the iteration number reaches the maximum iteration number, the optimal parameter combination of the SVM training model is output, otherwise the iteration is continuously performed;

[0015] S3, the parameters in the SVM model are combined and optimized by the improved marine predator algorithm based on the training set in the sample set, to obtain a training model;

[0016] S4, the test set in the sample set is subjected to PD fault category identification based on the training model.

[0017] Further, the i-th individual of the initial population is composed of:

[0018] X(i) = [X i，1 , X i，2 ,..., X i，j ,..., X i，d ],

[0019] wherein i = 1, 2,..., n, n represents the population size, j = 1, 2,..., d, d refers to the dimension of a solution vector, i.e. the number of solving variables, X i，j represents the solving variables, i.e. the c, g parameters of the support vector machine, c represents the penalty factor, and g represents the kernel function parameter;

[0020] The step S21 specifically comprises the steps of:

[0021] S211, forming a pre-initial population by traversing the initial space based on a tent mapping, and the i-th individual thereof is represented as:

[0022] X 1 (i) = [X 1 i，1 , X 1 i，2 ,..., X 1 i，j ,..., X 1 i，d ], i = 1, 2,..., n;

[0023] S212, introducing a reverse learning mechanism to generate a reverse population corresponding to the pre-initial population, and the i-th individual thereof is represented as:

[0024]

[0025] S213, selecting n individuals with higher fitness function values from the pre-initial population and the reverse population corresponding thereto to form the initial population.

[0026] Further, the step S211 specifically comprises the steps of:

[0027] S2111, generating a chaotic particle sequence based on the tent mapping:

[0028]

[0029] wherein i = 1, 2,..., n, n represents the population size, j = 1, 2,..., d, d represents the dimension of the space, i.e. the number of control variables, x i，j represents the generated chaotic particle sequence;

[0030] S2112, mapping the chaotic particle sequence to the search space to obtain the jth dimension value of the ith individual in the pre-initial population as:

[0031] X 1 i，j = lb j +(ub j -lb j )x i，j ,

[0032] wherein ub j and lb j are the upper limit and lower limit of X i，j search, i.e. the upper and lower limits of the optimization parameters c and g;

[0033] S2113, obtaining the pre-initial population formed based on the tent mapping traversal initial space, and the ith individual is:

[0034] X 1 (i) = [X 1 i，1 , X 1 i，2 ,..., X 1 i，j ,..., X 1 i，d ], i = 1, 2,..., n.

[0035] Further, in the reverse learning of the reverse learning mechanism introduced in the step S212, the reverse solution of the ith individual X 1 (i) = [X 1 i，1 , X 1 i，2 ,..., X 1 i，j ,..., X 1 i，d ] of the pre-initial population is given by:

[0036]

[0037] wherein lb and ub are row vectors composed of lb j and ub j , lb is a two-dimensional row vector corresponding to the lower limit of the values of c and g, and ub is also a two-dimensional row vector corresponding to the upper limit of the values of c and g.

[0038] Further, in the step S22, the initial prey matrix (Prey) and the elite predator matrix (Elite) are respectively represented as:

[0039]

[0040]

[0041] Among them, X I Let P be the d-dimensional vector of the top predator, which is copied n times to construct the elite predator matrix. i E represents the i-th prey individual in the current iteration. i This represents the i-th elite predator individual in the current iteration.

[0042] Furthermore, step S23 specifically includes the following steps:

[0043] S231, in In the initial iteration phase, i.e., the high-speed ratio phase, the iteration is performed according to the following formula:

[0044]

[0045] Where R is a vector of random numbers uniformly distributed in the range [0, 1], P = 0.5, s i It is the movement step size, R B It is a normally distributed Brownian motion random vector. P represents the process of element-wise multiplication. inew P represents i The new value obtained after iteration is iter, which is the current iteration number, and Max_iter is the maximum iteration number;

[0046] S232, in During the mid-cycle phase, i.e., the constant-speed ratio phase, the iteration is performed according to the following formula:

[0047]

[0048]

[0049]

[0050] Among them, R L It is a random vector that follows a Lévy distribution, C F These are parameters that marine predators adapt to;

[0051] S233, in In the later stages of iteration, i.e., the low-speed ratio stage, the iteration is performed according to the following formula:

[0052]

[0053] Further, in the step S25, the fitness function value ranks n / 2 to the maximum rank are divided into a subgroup, recorded as prey.q, and the subgroup prey.q is disturbed by adaptive Cauchy mutation; the fitness function value ranks minimum to n / 2 are divided into a subgroup, recorded as prey.h, and the subgroup prey.h is subjected to differential evolution.

[0054] The subgroup prey.q is disturbed by adaptive Cauchy mutation, specifically including the steps of:

[0055] A1, the average of each column in the subgroup prey.q is calculated, that is, the average values avgp(1, j) of c and g of the n / 2 individuals in the subgroup prey.q;

[0056] A2, each element prey.q(i, j) in the subgroup prey.q is subtracted from the column average avgp(1, j) corresponding to the column to obtain r(f, j), and the maximum value of each column in r(i, j) is calculated and saved in Rmax(i, j);

[0057] A3, a new variable Xm(i, j) related to the relative iteration number and the relative distance is used to replace the independent variable in the standard Cauchy distribution function, so as to change the standard Cauchy distribution into adaptive Cauchy distribution Xm(i, j) is calculated by the following formula:

[0058]

[0059] Wherein, λ is a constant, Iter is the current iteration number, Max_iter is the maximum iteration number, The relative iteration number is called, and the relative distance is called r(i, j) / Rmax(i, j);

[0060] A4, the Xm(i, j) is brought into the standard Cauchy distribution function The adaptive Cauchy mutation formula is obtained:

[0061]

[0062] According to the adaptive Cauchy mutation formula, in the early stage of iteration, the performance of the individual is poor, and the large mutation value can cause sufficient disturbance and expand the solution space; with the increase of the iteration number, the mutation value gradually decreases, so as to ensure the smooth convergence of the problem to the optimal value. At the same time, adaptive mutation will obtain a larger mutation value when the extreme value of the group tends to be consistent, so as to enhance the search ability of the algorithm; when the search space of the group is large enough, the mutation value is reduced to avoid the fluctuation of the optimal value and speed up the convergence speed of the algorithm.

[0063] A5, the prey matrix is updated by the following formula:

[0064] Pi (i,j) = prey.q(i,j) + (-1 + 2*R)*avgp(1,j)*F(Xm(i,j)),

[0065] i = 1, 2...n / 2, j = 1, 2...d,

[0066] wherein, P i (i,j) represents the individual obtained after perturbing prey.q(i,j), R ∈ [0, 1] is a random number uniformly distributed.

[0067] Further, the subpopulation prey.h is subjected to differential evolution, specifically comprising the steps of:

[0068] B1, the average of each column in the subpopulation prey.h is calculated, that is, the average value avghp(1,j) of c and g of the n / 2 individuals in the subpopulation prey.h:

[0069] B2, each element prey.h(i,j) in the subpopulation prey.h is subjected to mutation operation according to the following formula, to obtain the mutation subpopulation U(i,j):

[0070] U(i,j) = prey.h(i,j)*R(-0.5,0.5) + avghp(1,j),

[0071]

[0072] wherein, R(-0.5,0.5) represents a random number uniformly distributed in (-0.5,0.5);

[0073] B3, the prey matrix is updated according to the following formula:

[0074]

[0075]

[0076] wherein, f1(i) and f2(i) respectively correspond to the fitness function values before and after mutation of the subpopulation prey.h.

[0077] Further, in the step S24, the mathematical model of FADs effect is as follows:

[0078]

[0079] wherein, r is a random number uniformly distributed in [0, 1]; P i represents the i-th prey individual at the current iteration, P inew represents the perturbation of P iThe new value obtained after disturbance. Subscripts r1 and r2 represent random indexes of different prey matrices; FADs represent the probability of FADs affecting the optimization process, taking 0.2; U is a binary vector containing 0 and 1.

[0080] Further, in the step S1, the sample data is preprocessed, specifically:

[0081] According to the idea of the three-ratio method in transformer oil chromatographic analysis, three concentration ratios c1, c2 and c3 are selected as characteristic quantities to form a sample set after normalization of the characteristic quantities, wherein c1 represents the concentration ratio of SOF2 to SO2F2, c2 represents the concentration ratio of CF4 to CO2, and c3 represents the ratio between the concentration sum of SOF2 and SO2F2 and the concentration sum of CF4 and CO2.

[0082] The application provides a GIS device PD type determination method using an improved marine predator algorithm to optimize SVM. On the one hand, the initial population of the original marine predator algorithm is randomly generated, which may cause uneven distribution of the initial population individual position, reduce population diversity and affect the optimization speed. The chaos mapping strategy and the reverse learning mechanism are introduced to improve the quality of the initial population. On the other hand, the original marine predator algorithm has defects such as being easy to fall into local optimum and slow convergence speed. A multi-subgroup disturbance idea is proposed. The prey population of the marine predator disturbed by the FADs effect is divided into two subgroups according to the fitness function value. The subgroup with high fitness function value is disturbed by adaptive Cauchy mutation, and the other subgroup is subjected to differential evolution, to generate a multi-subgroup disturbance solution, which is brought into the next iteration.

[0083] Overall, the GIS device PD type determination method using the improved marine predator algorithm to optimize SVM can improve the recognition efficiency and classification accuracy of the gas insulated combined electrical equipment PD fault type, and guarantee the safety and reliability of the modern power system. BRIEF DESCRIPTION OF DRAWINGS

[0084] Figure 1 is a flowchart of the GIS device PD type determination method using the improved marine predator algorithm to optimize SVM provided by the embodiment of the application;

[0085] Figure 2 is a flowchart of the improved marine predator algorithm optimizing support vector machine provided by the embodiment of the application;

[0086] Figure 3 is a particle swarm optimization algorithm optimizing support vector machine prediction result graph provided by the embodiment of the application, wherein (a) is a comparison graph of the test set true class and the predicted class, and (b) is a fitness curve graph;

[0087] Figure 4 is the prediction result figure of the original marine predator algorithm optimized support vector machine provided by the embodiment of the present application, wherein (a) is a comparison figure of the real class and the predicted class of the test set, and (b) is a fitness curve figure;

[0088] Figure 5 is the prediction result figure of the improved marine predator algorithm optimized support vector machine provided by the embodiment of the present application, wherein (a) is a comparison figure of the real class and the predicted class of the test set, and (b) is a fitness curve figure. DETAILED DESCRIPTION

[0089] The embodiments of the present application will be described in detail below with reference to the accompanying drawings. The embodiments are given only for the purpose of illustration and should not be understood as limiting the present application. The accompanying drawings are used for reference and illustration only and do not constitute a limitation on the scope of patent protection of the present application, because many changes can be made to the present application without departing from the spirit and scope thereof.

[0090] The embodiment of the present application provides a GIS device PD type determination method using an improved marine predator algorithm optimized SVM, as shown in the figure. Figure 1 As shown in the figure, the embodiment includes steps S1-S4.

[0091] S1, sample data is obtained, and the sample data is preprocessed to obtain a sample set;

[0092] S2, an improved original ecological marine predator algorithm is obtained;

[0093] S3, based on the training set in the sample set, the parameters in the SVM model are combined and optimized by the improved marine predator algorithm to obtain a training model;

[0094] S4, the test set in the sample set is identified based on the training model.

[0095] In step S1, the decomposition characteristic component content of SF6 under four different insulation defect caused PDs is selected as sample data in the embodiment, and the decomposition characteristic components include SO2F2, SOF2, CF4 and CO2. 32 groups of data are collected under each defect, and 128 groups of decomposition characteristic component content data under 4 kinds of PDs are obtained as sample data. 64 groups of data are used for classifier training, and another 64 groups of data are used for testing the performance of the classifier. The composition of the data set is shown in the following table 1:

[0096] Table 1. Training and test samples

[0097]

[0098] The sample data is preprocessed, specifically:

[0099] In the light of the three-ratio method in the chromatographic analysis of transformer oil, three concentration ratios c1, c2 and c3 are selected as characteristic quantities to form a sample set after normalization of the characteristic quantities, wherein c1 represents the concentration ratio of SOF2 to SO2F2, c2 represents the concentration ratio of CF4 to CO2, and c3 represents the ratio between the concentration sum of SOF2 and SO2F2 and the concentration sum of CF4 and CO2. For step S2, the improved marine predator algorithm steps include steps S21 to S27.

[0100] S21, generating an initial population based on a chaotic mapping strategy and a reverse learning mechanism.

[0101] For the initial population, the i-th individual is composed of X(i)=[X i，1 , X i，2 ,..., X i，j ,..., X i，d ], i=1, 2,..., n, n represents the population size, j=1, 2,..., d, d refers to the dimension of a solution vector, i.e., the number of variables to be solved, X i，j represents the variables to be solved, i.e., the c and g parameters of the support vector machine, c represents the penalty factor, and g represents the kernel function parameter.

[0102] Step S21 specifically includes steps S211 to S213:

[0103] S211, traversing the initial space based on the tent mapping to form a pre-initial population, and the i-th individual is represented as:

[0104] X 1 (i)=[X 1 i，1 , X 1 i，2 ,..., X 1 i，j ,..., X 1 i，d ], i=1, 2,..., n.

[0105] Step S211 specifically includes steps S2111 to S2113:

[0106] S2111, generating a chaotic particle sequence based on the tent mapping:

[0107]

[0108] wherein: i=1, 2,..., n, n represents the population size; j=1, 2,..., d, d represents the spatial dimension, i.e., the number of control variables; x i，j represents the generated chaotic particle sequence;

[0109] S2112, mapping the chaotic particle sequence to the search space to obtain the jth dimension value of the ith individual in the pre-initial population as:

[0110] X 1 i，j = lb j +(ub j -lb j )x i，j ,

[0111] wherein ub j and lb j are the upper and lower bounds of X i，j search, i.e. the upper and lower limits of the optimization parameter c or g;

[0112] S2113, obtaining the pre-initial population formed by traversing the initial space based on the tent mapping, and the ith individual of the pre-initial population is:

[0113] X 1 (i) = [X 1 i，1 , X 1 i，2 ,..., X 1 i，j ,..., x 1 i，d ], i = 1, 2,..., n.

[0114] S212, introducing a reverse learning mechanism to generate a reverse total population corresponding to the pre-initial total population, and the ith individual of the reverse total population is expressed as:

[0115]

[0116] S213, selecting n individuals with higher fitness function values from the pre-initial total population and the corresponding reverse total population to form an initial population. The fitness function takes the average recognition rate returned by the 3-fold cross-validation of the model on the training data.

[0117] S22, constructing an initial prey matrix and an elite predator matrix according to the initial population.

[0118] In step S22, the initial prey matrix (Prey) and the elite predator matrix (Elite) are respectively expressed as:

[0119]

[0120]

[0121] wherein X I is a d-dimensional top predator vector, which is copied n times to construct the elite predator matrix, and Pi represents the i prey individual at the current iteration number, E i represents the i elite predator individual at the current iteration number.

[0122] S23, according to the current iteration number, the individual iteration mode is determined, and the prey matrix and the elite predator matrix are iterated and optimized.

[0123] Step S23 specifically includes steps S231-S233:

[0124] S231, in The initial iteration of the iteration, that is, the high-speed ratio stage, is iterated according to the following formula:

[0125]

[0126] Wherein, R is a random number vector uniformly distributed in [0, 1], constant P = 0.5, s i is a moving step, R B is a Brown motion random vector with normal distribution, represents the process of multiplication by elements, P inew represents P i The new value obtained after iteration, iter is the current iteration number, and Max_iter is the maximum iteration number;

[0127] S232, in The middle iteration of the iteration, that is, the constant speed ratio stage, is iterated according to the following formula:

[0128]

[0129]

[0130]

[0131] Wherein, R L is a random vector following Lévy distribution, C F is the adaptive parameter of the marine predator;

[0132] S233, in The late iteration of the iteration, that is, the low-speed ratio stage, is iterated according to the following formula:

[0133]

[0134] S24, applying FADs effect to disturb the prey population of marine predators.

[0135] In the step S24, the mathematical model of FADs effect is as follows:

[0136]

[0137] where r is a random number uniformly distributed in [0, 1]; P inew denotes the new value obtained after perturbing P i . Subscripts r1 and r2 represent random indexes different from each other for the prey matrix; FADs represents the probability of FADs affecting the optimization process, taken as 0.2; U is a binary vector containing 0 and 1.

[0138] S25, the prey population of the marine predator disturbed by the effect of FADs is divided into two subpopulations according to the average value of the fitness function, and the subpopulation with a higher fitness function value is disturbed by adaptive Cauchy mutation, and the other subpopulation is subjected to differential evolution, to generate a multi-subpopulation disturbed solution, which is brought into the next iteration.

[0139] In step S25, the subpopulation with a fitness function value ranking from n / 2 to the largest is denoted as prey.q, and the subpopulation prey.q is disturbed by adaptive Cauchy mutation; the subpopulation with a fitness function value ranking from the smallest to n / 2 is denoted as prey.h, and the subpopulation prey.h is subjected to differential evolution;

[0140] The subpopulation prey.q is disturbed by adaptive Cauchy mutation, which includes the following steps:

[0141] A1, the average of each column in the subpopulation prey.q is calculated, i.e. the average value avgp(1, j) of c and g of the n / 2 individuals in the subpopulation prey.q:

[0142]

[0143] A2, the difference between each element prey.q(i, j) in the subpopulation prey.q and the column average avgp(1, j) corresponding to the column is taken and the absolute value (abs function) is taken to obtain r(i, j), and the maximum value (max function) of each column in r(i, j) is calculated and saved in Rmax(i, j):

[0144]

[0145] A3, a new variable Xm(i, j) related to the relative iteration number and the relative distance is used to replace the independent variable in the standard Cauchy distribution function, so as to change the standard Cauchy distribution into adaptive Cauchy distribution, Xm(i, j) is calculated as follows:

[0146]

[0147] where λ is a constant, taken as 9, Iter is the current iteration number, Max_iter is the maximum iteration number, The relative iteration number is called r(i, j) / Rmax(i, j), and the relative distance is called r(i, j) / Rmax(i, j);

[0148] A4, bring Xm(i, j) into the standard Cauchy distribution function The adaptive Cauchy mutation formula is obtained:

[0149]

[0150] According to the adaptive Cauchy mutation formula, in the early stage of iteration, the performance of the individual is poor, and a larger mutation value can cause sufficient disturbance and expand the solution space; with the increase of the number of iterations, the mutation value gradually decreases, thereby ensuring the smooth convergence of the problem to the optimal value. At the same time, the adaptive mutation will obtain a larger mutation value when the extreme values of the population tend to be consistent, thereby enhancing the search ability of the algorithm; and when the search space of the population is large enough, the mutation value is reduced to avoid the disturbance of the optimal value and speed up the convergence speed of the algorithm.

[0151] A5, update the prey matrix according to the following formula:

[0152] P i (i, j) = prey.q(i, j) + (-1 + 2*R)*avghp(1, j)*F(Xm(i, j)),

[0153] i = 1, 2...n / 2, j = 1, 2...d,

[0154] wherein, P i (i, j) represents an individual obtained by disturbing prey.q(i, j), and R ∈ [0, 1] is a random number uniformly distributed.

[0155] The subpopulation prey.h is subjected to differential evolution, which specifically includes the following steps:

[0156] B1, calculate the average of each column in the subpopulation prey.h, that is, the average value avghp(1, j) of c and g of the n / 2 individuals in the subpopulation prey.h:

[0157] B2, perform mutation operation on each element prey.h(i, j) in the subpopulation prey.h according to the following formula, to obtain the mutation subpopulation U(i, j):

[0158] U(i, j) = prey.h(i, j)*R(-0.5, 0.5) + avghp(1, j),

[0159]

[0160] wherein, R(-0.5, 0.5) represents a random number uniformly distributed in (-0.5, 0.5);

[0161] B3, prey matrix is updated as follows:

[0162]

[0163]

[0164] Wherein, f1(i) and f2(i) are respectively corresponding to the prey.h before and after mutation of the fitness function value.

[0165] S26, according to the algorithm timing architecture to complete the high-speed ratio, the speed ratio and the low-speed ratio stage iteration;

[0166] S27, when the number of iterations reaches the maximum number of iterations, output SVM training model optimal parameter combination, otherwise continue to execute iteration.

[0167] In step S3, the model parameters are set as follows:

[0168] When training the constructed model using the training set, k-fold cross validation is used to measure the performance of the classifier. According to the size of the sample, this paper takes k=3, that is, the fitness function takes the average recognition rate returned by 3-fold cross validation. When training the classifier, the optimization algorithm is used to optimize the penalty factor c and the kernel function parameter g of the support vector machine. The search range of c is taken as (0.1, 100), and the search range of g is taken as (0.01, 100). In the particle swarm optimization algorithm, the inertia factor w=1, the acceleration factors c1 and c2 are 1.6 and 1.5 respectively, the population size (population number) n=30, and the iteration number t=90; In the marine predator algorithm and the improved marine predator algorithm, FADs=0.2, P=0.5, the population size (population number) n=30, and the iteration number t=90.

[0169] In step S4, the experimental results are shown in Figs. 1-3 Figure 3 , Figs. 2-3 Figure 4 , Figs. 3-3 Figure 5 : wherein Figs. 1-3 Figure 3 (a), Figs. 2-3 Figure 4 (a), Figs. 3-3 Figure 5 (a) are respectively the identification result figures of the model for the test samples when using the particle swarm optimization algorithm, the marine predator algorithm and the improved marine predator algorithm to optimize the support vector machine parameters; Figs. 1-3 Figure 3 (b), Figs. 2-3 Figure 4 (b), Figs. 3-3 Figure 5 (b) are respectively the fitness curve figures of the model when using the particle swarm optimization algorithm, the marine predator algorithm and the improved marine predator algorithm to optimize the support vector machine parameters (the solid line represents the best fitness curve, and the dashed line represents the average fitness curve of all particles at the current iteration number).

[0170] Experimental result analysis:

[0171] From the attached Figure 3 (a), the attached Figure 4 (a), the attached Figure 5 (a) can be known, when using particle swarm optimization algorithm, marine predator algorithm and improved marine predator algorithm to optimize the parameters c and g of support vector machine respectively, the corresponding model prediction accuracy is 90.625%, 92.1875% and 93.75% respectively, the results show that the improved marine predator algorithm has the most significant optimization effect on the parameters of support vector machine; from the attached Figure 3 (b), the attached Figure 4 (b), the attached Figure 5 (b) can be known, when using improved marine predator algorithm to optimize the parameters of support vector machine, the fitness curve is best, the optimization algorithm convergence is improved while improving the optimization precision. It is shown that the method of using the improved marine predator algorithm to optimize the parameters of support vector machine to construct the classifier is feasible.

[0172] In summary, the GIS device PD type determination method for optimizing SVM by using the improved marine predator algorithm provided by the embodiment of the application. On the one hand, the initial population of the original marine predator algorithm is randomly generated, which may lead to uneven distribution of the initial population individual position, reduce the population diversity and affect the optimization speed. The chaos mapping strategy and the reverse learning mechanism are introduced to improve the quality of the initial population. On the other hand, the original marine predator algorithm has the defects of being easy to fall into local optimum and slow convergence speed. A multi-subpopulation disturbance idea is proposed. The prey population of the marine predator disturbed by the FADs effect is divided into two subpopulations according to the fitness function value. The subpopulation with higher fitness function value is disturbed by adaptive Cauchy mutation, and the other subpopulation is subjected to differential evolution, to generate multi-subpopulation disturbance solution for next iteration.

[0173] Overall, the GIS device PD type determination method for optimizing SVM by using the improved marine predator algorithm proposed by the application can improve the recognition efficiency and classification accuracy of the PD fault type of gas insulated combined electrical equipment, and ensure the safety and reliability of modern power systems.

[0174] The above embodiments are the preferred embodiments of the application, but the embodiments of the application are not limited by the above embodiments, and any changes, modifications, substitutions, combinations and simplifications made without departing from the spirit and principles of the application shall be equivalent replacement methods and shall be included in the protection scope of the application.

Claims

1. A method for determining the PD type of GIS equipment using an improved marine predator algorithm-optimized SVM, characterized in that, Including the following steps: S1. Obtain sample data and preprocess the sample data to obtain a sample set; S2. An improved marine predator algorithm is obtained by improving the original marine predator algorithm. The steps of the improved marine predator algorithm include: S21. Generate the initial population based on the chaotic mapping strategy and the reverse learning mechanism; S22. Construct an initial prey matrix and an elite predator matrix based on the initial population; S23. Determine the individual iteration method based on the current iteration number, iterate the initial prey matrix and elite predator matrix, and perform iterative optimization; S24. Apply the FADs effect to disturb the prey populations of marine predators; S25. Divide the prey population of marine predators after FADs perturbation into two subgroups based on fitness function values. Perturb the subgroup with higher fitness function values ​​using adaptive Cauchy mutation, and perform differential evolution on the other subgroup to generate a multi-subgroup perturbation solution, which is then used in the next iteration. S26. According to the algorithm timing architecture, the iterations of the high speed ratio, constant speed ratio and low speed ratio stages are completed in sequence; S27. When the number of iterations reaches the maximum number of iterations, output the optimal parameter combination of the SVM training model; otherwise, continue to execute the iteration. S3. Based on the training set in the sample set, the parameters in the SVM model are combined and optimized using the improved marine predator algorithm to obtain the training model. S4. Based on the training model, identify PD fault categories on the test set in the sample set.

2. The method for determining the PD type of GIS equipment using an improved marine predator algorithm to optimize SVM according to claim 1, characterized in that: For the initial population, the i-th individual is composed of X(i) = [X... i，1 X i，2 , ..., X i，j , ..., X i，d ], i = 1, 2...n, where n represents the population size, j = 1, 2...d, where d refers to the dimension of a solution vector, i.e., the number of variables to be solved, X i，j The parameters c and g represent the variables we are looking for, i.e., the parameters of the support vector machine. c represents the penalty factor, and g represents the kernel function parameters. Step S21 specifically includes the following steps: S211. Based on the tent mapping, the initial space is traversed to form a pre-initial population, whose i-th individual is represented as: X 1 (i)=[X 1 i，1 ,X 1 i，2 ,...,X 1 i，j ,...,X 1 i，d ],i=1,2,...,n; S212. Introduce a reverse learning mechanism to generate a reverse population corresponding to the pre-initial population, where the i-th individual is represented as: S213. Select n individuals with higher fitness function values ​​from the pre-initial population and its corresponding reverse population to form the initial population.

3. The method for determining the PD type of GIS equipment using an improved marine predator algorithm to optimize SVM according to claim 2, characterized in that, Step S211 specifically includes the following steps: S2111. Generating chaotic particle sequences based on tent mapping: Where: i = 1, 2, ..., n, n represents the population size; j = 1, 2, ..., d, d represents the spatial dimension, i.e., the number of control variables; x i，j This represents the resulting sequence of chaotic particles; S2112. Mapping the chaotic particle sequence to the search space, we obtain the j-th dimension value of the i-th individual in the pre-initialized population: x 1 i，j =lb j +(ub j -lb j )x i，j , Among them, ub j and lb j For X i，j The upper and lower bounds of the search, i.e. the upper and lower limits of the optimization parameters c or g; S2113. Obtain the pre-initial population formed by traversing the initial space based on the tent mapping, and its i-th individual is: X 1 (i)=[X 1 i，1 ,X 1 i，2 ,...,X 1 i，j ,...,X 1 i，d ],i=1,2,...,n。 4. The method for determining the PD type of GIS equipment using an improved marine predator algorithm to optimize SVM according to claim 3, characterized in that: In the reverse learning mechanism introduced in step S212, the i-th individual X of the pre-initialized population... 1 (i)=[X 1 i，1 X 1 i，2 , ..., X 1 i，j , ..., X 1 i，d The inverse solution of ] It is given by the following formula: Among them, lb and ub are composed of lb j and ub j The resulting row vectors are lb, a two-dimensional row vector corresponding to the lower bound of c and g, and ub, also a two-dimensional row vector corresponding to the upper bound of c and g.

5. The method for determining the PD type of GIS equipment using an improved marine predator algorithm to optimize SVM according to claim 4, characterized in that, In step S22, the initial prey matrix Prey and the elite predator matrix Elite are respectively represented as: Among them, X I Let P be the d-dimensional vector of the top predator, which is copied n times to construct the elite predator matrix. i E represents the i-th prey individual in the current iteration. i This represents the i-th elite predator individual in the current iteration.

6. The method for determining the PD type of GIS equipment using an improved marine predator algorithm to optimize SVM according to claim 5, characterized in that, Step S23 specifically includes the following steps: S231, in In the initial iteration phase, i.e., the high-speed ratio phase, the iteration is performed according to the following formula: Where R is a vector of random numbers uniformly distributed in the range [0, 1], P = 0.5, s i It is the movement step size, R B It is a normally distributed Brownian motion random vector. P represents the process of element-wise multiplication. inew This represents the new value of Pi obtained after iteration, where iter is the current iteration number and Max_iter is the maximum iteration number; S232, in During the mid-cycle phase, i.e., the constant-speed ratio phase, the iteration is performed according to the following formula: Among them, R L It is a random vector that follows a Lévy distribution, C F These are parameters that marine predators adapt to; S233, in In the later stages of iteration, i.e., the low-speed ratio stage, the iteration is performed according to the following formula:

7. The method for determining the PD type of GIS equipment using an improved marine predator algorithm to optimize SVM according to claim 6, characterized in that, In step S25, fitness function values ​​ranked from n / 2 to the highest are grouped into a subgroup, denoted as prey.q, and the subgroup prey.q is perturbed by adaptive Cauchy mutation; fitness function values ​​ranked from the lowest to n / 2 are grouped into a subgroup, denoted as prey.h, and the subgroup prey.h is subjected to class difference evolution. The subgroup prey.q is perturbed by adaptive Cauchy mutation, specifically including the following steps: A1. Calculate the mean of each column in the subgroup prey.q, that is, the average value of c and g of n / 2 individuals in the prey.q subgroup avgp(1,j); A2. Take the difference between each element prey.q(i,j) in the subgroup prey.q and the column mean avgp(1,j) corresponding to that column, and take the absolute value to obtain r(i,j). Calculate the maximum value of each column in r(i,j) and store it in Rmax(i,j). A3. Replace the independent variable in the standard Cauchy distribution function with a new variable Xm(i,j) that is related to the relative iteration number and relative distance, so as to transform the standard Cauchy distribution into an adaptive Cauchy distribution. Xm(i,j) is calculated by the following formula: Where λ is a constant, Iter is the current iteration number, and Max_iter is the maximum iteration number. This is called the relative iteration number, and r(i,j) / Rmax(i,j) is called the relative distance. A4. Substitute Xm(i,j) into the standard Cauchy distribution function. The adaptive Cauchy variation formula is obtained as follows: A5. Update the prey matrix using the following formula: Pi(i,j)=prey.q(i,j)+(-1+2*R)*avgp(1,j)*F(Xm(i,j)), i=1, 2...n / 2, j=1, 2...d, Where Pi(i,j) represents the individual obtained after perturbing prey.q(i,j), and R∈[0,1] is a uniformly distributed random number.

8. The method for determining the PD type of GIS equipment using an improved marine predator algorithm to optimize SVM according to claim 7, characterized in that, The subgroup prey.h undergoes a difference-like evolution, specifically including the following steps: B1. Calculate the mean of each column in the subgroup prey.h, that is, the average of c and g for n / 2 individuals in the subgroup prey.h: avghp(1,j) B2. Perform mutation operation on each element prey.h(i,j) in the subgroup prey.h according to the following formula to obtain the mutated subgroup U(i,j): U(i,j)=prey.h(i,j)*R(-0.5,0.5)+avghp(1,j), Where R(-0.5, 0.5) represents a uniformly distributed random number in (-0.5, 0.5); B3. Update the prey matrix using the following formula: Here, f1(i) and f2(i) correspond to the fitness function values ​​of the subgroup prey.h before and after mutation, respectively.

9. The method for determining the PD type of GIS equipment using an improved marine predator algorithm to optimize SVM according to claim 8, characterized in that, In step S24, the mathematical model of the FADs effect is as follows: Among them, P inew This represents the new value obtained after perturbing Pi, where r is a random number uniformly distributed in [0, 1]; the subscripts r1 and r2 represent random indices where the prey matrix is ​​different; FADs represent the probability that FADs affect the optimization process, which is 0.2; U is a binary vector containing 0 and 1.

10. The method for determining the PD type of GIS equipment using an improved marine predator algorithm to optimize SVM according to claim 1, characterized in that, In step S1, the sample data is preprocessed, specifically as follows: Following the approach of the three-ratio method in transformer oil chromatographic analysis, three concentration ratios c1, c2, and c3 were selected as characteristic quantities. After normalization of the characteristic quantities, a sample set was formed. Here, c1 represents the concentration ratio of SOF2 to SO2F2, c2 represents the concentration ratio of CF4 to CO2, and c3 represents the ratio between the sum of the concentrations of SOF2 and SO2F2 and the sum of the concentrations of CF4 and CO2.

Images