Transformer winding deformation classification method for optimizing support vector machine based on Lissajous ellipse graph and grey wolf algorithm

Abstract

The invention provides a transformer winding deformation classification method for optimizing a support vector machine based on a Lissajous ellipse graph and a grey wolf algorithm, and the method comprises the steps: S1, respectively collecting the voltage and current of a high-voltage side of a detected phase of a transformer, and the voltage and current of a low-voltage side of the detected phase of the transformer; s2, constructing a Lissajous ellipse graph by using the voltage amplitude difference of the high-voltage side and the low-voltage side of the measured phase of the transformer and the current amplitude of the high-voltage side; s3, establishing a characteristic parameter sample library representing different states of the transformer winding; s4, establishing a support vector machine based on the characteristic parameter sample library, and optimizing an intelligent transformer winding deformation classification model of the support vector machine based on a Lissajous ellipse graph and a grey wolf algorithm; and S5, acquiring the voltage and current of the high-voltage side and the low-voltage side of the detected phase of the field transformer, and quickly classifying the winding state of the field transformer by using the classification model. The grey wolf algorithm adopted to optimize the support vector machine has the advantages of high speed, high accuracy and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of transformer winding deformation classification, and in particular to a transformer winding deformation classification method using a support vector machine. Background Art

[0002] The power transformer is one of the key devices in the power system, and its health status is directly related to the working condition of the entire power system. When a power transformer fails, it will not only affect the stability of the regional power system, but also may cause large-scale power outages and even the risk of power supply chain unbundling.

[0003] Winding deformation defect is one of the main causes of power transformer failure. At present, the power industry mainly uses offline methods such as short-circuit impedance method, low-voltage pulse method, and frequency response method to detect this defect. The short-circuit impedance method means that when the load impedance is zero, the equivalent impedance inside the transformer will change with the deformation of the winding, so that the transformer winding state can be determined by comparing the transformer short-circuit impedance value before and after the fault. The frequency response method refers to loading an excitation signal at the head end of the transformer winding and testing the response signal at the end at the same time, so as to obtain a frequency response characteristic curve that characterizes the mechanical state of the transformer winding, and then determine the transformer winding state. The basic principle of the low-voltage pulse method is similar to the frequency response method. Both use the amplitude-frequency characteristic curve to determine the winding state, but its excitation signal is different from the frequency response method. It is mainly a steep pulse containing a large number of high-frequency components. Although the above-mentioned offline detection method has clear criteria, it requires the transformer to be shut down when it is applied on site, which will not only affect the normal operation of the power grid, but also cause great economic losses. In order to solve the above problems, scholars have proposed a method for online detection of winding state using transformer voltage and current graphs. This method collects the voltage and current data of the primary and secondary sides of the running transformer, establishes the voltage-current Lissajous ellipse figure that represents the mechanical state of the transformer winding, and realizes the online detection of the transformer winding state by comparing the transformer Lissajous ellipse figure and its characteristic parameters before and after the short-circuit impact. The field application results show that this method is highly practical, has a high safety factor, and does not require additional test equipment. However, the criterion for winding deformation proposed based on this method is still in the category of threshold method, so its applicability is poor and it does not have the ability to intelligently classify winding deformation. Summary of the invention

[0004] In view of the technical problems that the existing technology has poor applicability and lacks the ability to intelligently classify winding deformation, the present invention proposes a transformer winding deformation classification method based on Lissajous ellipse graphics and Grey Wolf algorithm optimized support vector machine. The voltage and current graphic analysis method adopted does not require additional equipment such as external sensors, and only uses the existing voltage transformer and current transformer of the power grid to realize online detection of the transformer winding status, thereby effectively reducing the maintenance cost of the transformer.

[0005] In order to achieve the above object, the technical solution of the present invention is implemented as follows: a transformer winding deformation classification method based on Lissajous ellipse graphics and gray wolf algorithm optimized support vector machine, the steps comprising:

[0006] S1: collect the voltage and current on the high voltage side of the transformer phase under test, as well as the voltage and current on the low voltage side;

[0007] S2: construct a Lissajous ellipse figure using the voltage amplitude difference between the high and low voltage sides of the measured phase of the transformer and the current amplitude on the high voltage side;

[0008] S3: obtaining characteristic parameters of the Lissajous ellipse figure of the transformer under different winding states, and establishing a sample library of characteristic parameters representing different states of the transformer winding;

[0009] S4: Establish a support vector machine based on a feature parameter sample library, and optimize the intelligent transformer winding deformation classification model of the support vector machine based on Lissajous ellipse graphics and the Grey Wolf algorithm;

[0010] S5: Collect the voltage and current on the high-voltage side of the measured phase of the on-site transformer, as well as the voltage and current on the low-voltage side, and use the intelligent transformer winding deformation classification model to quickly classify the on-site transformer winding status.

[0011] The method for constructing the Lissajous ellipse figure in step S2 is:

[0012] Using the voltages on the high and low voltage sides of the measured phase, the voltage amplitude difference ΔU=U1-U2 is calculated, and then used as the horizontal coordinate, and the high voltage side current amplitude I1 of the measured phase is used as the vertical coordinate to construct a Lissajous ellipse figure representing the mechanical state of the transformer winding;

[0013] Among them, the major axis a, minor axis b and inclination angle θ of the Lissajous ellipse obtained by the ellipse formula group are characteristic parameters that characterize the mechanical state of the winding; the characteristic parameter expression is:

[0014]

[0015] Among them, X c and Y cIt is the geometric center of the Lissajous ellipse. A, B, C and D are parameters. The parameters A, B, C and D are obtained according to the voltage u1 on the high voltage side, the voltage u2 on the low voltage side and the current i1 on the high voltage side.

[0016] The method for obtaining parameters A, B, C and D according to the voltage u1 on the high voltage side, the voltage u2 on the low voltage side and the current i1 on the high voltage side is:

[0017] Parameters A, B, C and D are related to the voltage u1 on the high voltage side, the voltage u2 on the low voltage side and the current i1 on the high voltage side:

[0018] Δu(t)=u1(t)-u2(t)=ΔUcos(ωt);

[0019]

[0020] Among them, ΔU=U1-U2, U1 is the amplitude of the high-voltage side voltage u1(t), U2 is the amplitude of the low-voltage side voltage u2, I1 is the amplitude of the high-voltage side input current i1(t), is the initial phase, t is the time, ω is the angular frequency;

[0021] After trigonometric transformation, we get:

[0022]

[0023] cos 2 (ωt)+sin 2 (ωt) = 1;

[0024] By eliminating the time parameter, the relationship between the voltage amplitude difference and the input current is obtained:

[0025]

[0026] The parameters A, B, C and D are:

[0027]

[0028] The method for establishing a characteristic parameter sample library characterizing different states of transformer windings in step S3 is:

[0029] The characteristic parameters of the Lissajous ellipse of the transformer under different winding states are obtained by using the ellipse formula group. The Lissajous characteristic parameter data are labeled with labels 1, 2, ..., m according to the corresponding winding deformation types, so as to establish a characteristic parameter sample library L = {L1, L2, ..., L n}, where the i-th sample is L i ={m,a i ,b i ,θ i}, n is the total sample size.

[0030] The characteristic parameters of the Lissajous ellipse are the major axis a, the minor axis b and the inclination angle θ;

[0031] Different winding states include normal, short circuit, axial displacement and broken wire.

[0032] The method for establishing a support vector machine based on a feature parameter sample library in step S4 is:

[0033] Map an m-dimensional input vector To a high-dimensional feature space, and construct an optimal hyperplane in the high-dimensional space. Different mappings construct different support vector machines. In the high-dimensional feature space, the method of dividing the hyperplane is:

[0034] f(x) = ω·x + b = 0;

[0035] Where ω is the normal vector, b is the intercept constant, and x is the feature parameter sample library L = {L1, L2, ..., L n} in the feature parameters;

[0036] The optimal formula for the support vector machine to distinguish categories is:

[0037]

[0038] In the formula, ξ is the slack variable, C is the penalty factor, and the constraints are:

[0039] y i [(ω·x i )+b]≥1-ξ i ,i=1,2…n;

[0040] If α i * is the optimal solution, b * is the classification threshold, and the constraint is Then the optimal classification surface is obtained:

[0041]

[0042] In the formula, x i is the i-th sample to be tested in the sample library L i ={m,a i ,b i ,θ i} characteristic parameter data {a i ,b i ,θ i},y i is the i-th sample to be tested L i ={m,a i ,b i,θ i} label i, K(x i x) is the radial kernel function, which is calculated as follows:

[0043]

[0044] Among them, g 2 is the optimal hyperparameter, sgn() is the sign function, and exp() is the exponential function.

[0045] The method for establishing the intelligent transformer winding deformation classification model based on Lissajous ellipse graphics and Grey Wolf algorithm optimized support vector machine in step S4 is:

[0046] S41: Establish a gray wolf algorithm model and set the classification error rate of the support vector machine as the fitness function of the gray wolf optimization algorithm;

[0047] S42: Initialize the wolf pack, set the wolf pack size to N, and initialize the wolf pack position in the feasible domain [lb, ub], and randomly generate a two-dimensional vector in the feasible domain. The two-dimensional vector is composed of the hyperparameter g of the support vector machine and the penalty factor C;

[0048] S43: Randomly generate wolf pack positions, and substitute the generated wolf pack positions into the support vector machine based on the feature parameter sample library to obtain the fitness value of the wolf pack. The top three wolves with the best fitness are set as α, β and γ in descending order of fitness, and their positions are X α , X β and X γ , the position of the remaining wolves is X;

[0049] S44: Determine position X α Whether the fitness value is less than the preset threshold λ, the position of the wolf pack is updated to form an optimized support vector machine classification model.

[0050] The formula for setting the classification error rate of the support vector machine as the fitness function of the gray wolf optimization algorithm in step S41 is:

[0051]

[0052] The fitness value of each wolf is represented by fitness, n is the number of samples to be tested, T i is the i-th sample to be tested L i The actual label, P i is the i-th sample L i The predicted label of .

[0053] The method for updating the position of the wolf pack in step S44 is:

[0054] Use the update formula to update the position:

[0055]

[0056]

[0057]

[0058]

[0059] Among them, E1, E2, and E3 are random numbers between 0 and 1, and L is the position update coefficient at time t; D α , D β , D γ are the direction vectors between α wolf, β wolf, γ wolf and the current wolf respectively;

[0060] After updating the position, the wolf pack is adjusted out of bounds and the wolf pack fitness value is recalculated. The new wolf pack fitness value is compared with the position X obtained last time. α , X β and X γ Compare the fitness values ​​of α , X β and X γ location;

[0061] Determine the updated position X α Is the fitness value less than λ? If not, set t = t + 1 and repeat the wolf position update until the updated position X α The fitness value is less than λ, and the position X obtained in the last iteration is α As the optimal position, it is converted into two parameters, C and g, and substituted into the support vector machine to form an optimized support vector machine classification model.

[0062] The method of using the classification model to quickly classify the on-site transformer winding state in step S5 is:

[0063] The voltage and current on the high-voltage side of the measured phase of the on-site transformer, as well as the voltage and current on the low-voltage side, are collected. The voltage amplitude difference between the two and the current amplitude on the high-voltage side are calculated to construct a Lissajous ellipse. The major axis a, minor axis b and tilt angle θ of the ellipse are input into the classification model, and the classification model is used to quickly classify the winding status of the on-site transformer.

[0064] The beneficial effects of the present invention are: 1) Compared with the currently commonly used short-circuit impedance method and frequency response analysis method, the voltage and current graphic analysis method adopted by the present invention does not require additional equipment such as external sensors, and only uses the existing voltage transformer and current transformer of the power grid to realize online detection of the transformer winding status, thereby effectively reducing the maintenance cost of the transformer.

[0065] 2) When there is a fault in the transformer winding, the inclination angle, major axis, and minor axis of the transformer Lissajous ellipse constructed based on the voltage-current graphical analysis method will change accordingly. Therefore, the mechanical state of the transformer winding can be more intuitively reflected through the Lissajous ellipse.

[0066] 3) This method uses the winding deformation classification model established by support vector machine, which has good generalization ability, avoids the problem of poor applicability of the threshold method, and solves engineering problems such as small samples, high dimensions, and nonlinearity in the research on transformer winding deformation classification.

[0067] 4) Compared with the winding deformation classification method using the commonly used grid search to optimize the support vector machine, the gray wolf algorithm used in this invention to optimize the support vector machine has the advantages of fast speed and high accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0069] Figure 1 The present invention is a flow chart of a transformer winding deformation classification method based on Lissajous ellipse graphics and Grey Wolf algorithm to optimize support vector machine.

[0070] Figure 2 This is the wiring diagram for phase A test using the voltage and current graphical analysis method.

[0071] Figure 3 This is the Lissajous ellipse figure when the winding is normal.

[0072] Figure 4 This is a comparison diagram of the Lissajous ellipse when the winding is normal and when there is a short circuit fault.

[0073] Figure 5 This is a comparison diagram of the Lissajous ellipse when the winding is normal and when there is an axial displacement fault.

[0074] Figure 6 This is a comparison diagram of the Lissajous ellipse when the winding is normal and when there is a wire break fault.

[0075] Figure 7 This is the recognition result of the winding deformation classification model based on Lissajous ellipse graphics and Grey Wolf algorithm to optimize the support vector machine.

[0076] Figure 8 is the recognition accuracy and recognition time of the test set.

[0077] Fig. 9 Recognition results of the winding deformation classification model based on grid search optimized support vector machine.

[0078] Fig.10 It is a flow chart of a specific embodiment. DETAILED DESCRIPTION

[0079] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0080] Example 1

[0081] like Figure 1 As shown, a transformer winding deformation classification method based on Lissajous ellipse graphics and Grey Wolf algorithm optimized support vector machine, the steps include:

[0082] S1: Collect the voltage and current on the high-voltage side of the measured phase of the transformer, as well as the voltage and current on the low-voltage side.

[0083] The voltage u1 and current i1 of the high-voltage side of the transformer and the voltage u2 of the low-voltage side are collected through the existing voltage transformer (PT) and current transformer (CT) of the power grid. Figure 2 shown.

[0084] S2: Select the voltage amplitude difference between the high and low voltage sides of the measured phase of the transformer and the current amplitude on the high voltage side to construct a Lissajous ellipse, and use its major axis, minor axis and tilt angle as characteristic parameters. This construction method can mainly reflect the change of the transformer body impedance and is more sensitive to the mechanical state of the winding. When the transformer winding is slightly deformed, fault diagnosis can also be achieved through the changes in the three characteristic parameters.

[0085] By using the voltages on the high and low voltage sides of the measured phase, the voltage amplitude difference ΔU=U1-U2 is calculated, and then it is used as the abscissa, and the high voltage side current amplitude I1 of the measured phase is used as the ordinate, and the Lissajous ellipse representing the mechanical state of the transformer winding can be constructed, as shown in Figure 3 As shown. Among them, the ellipse formula group is used to obtain Figure 3 The major axis a, minor axis b and inclination angle θ of the mesoellipse can be used as characteristic parameters to characterize the mechanical state of the winding. The characteristic parameter expressions are as follows:

[0086]

[0087]

[0088]

[0089] In formulas (1)-(3), X c and Y c is the geometric center of the Lissajous ellipse. The other parameters A, B, C and D are related to the voltages u1 and u2 on the high and low voltage sides, and the current i1 on the high voltage side, as follows:

[0090] Δu(t)=u1(t)-u2(t)=ΔUcos(ωt) (4)

[0091]

[0092] Among them, ΔU=U1-U2, U1 is the amplitude of the high-voltage side voltage u1, U2 is the amplitude of the low-voltage side voltage u2, I1 is the amplitude of the high-voltage side input current i1, is the initial phase, t is the time, and ω is the angular frequency. After trigonometric transformation, equation (5) can be transformed into equation (6):

[0093]

[0094] cos 2 (ωt)+sin 2 (ωt)=1 (8)

[0095] Based on equations (4), (7) and (8), by eliminating the time parameter, the relationship between the voltage amplitude difference and the input current is obtained:

[0096]

[0097] Therefore, from formula (9), the parameters A, B, C and D in formulas (1)-(3) are:

[0098]

[0099] S3: Obtain the characteristic parameters of the transformer Lissajous ellipse under different winding states, establish a sample library of characteristic parameters that characterize different states of the transformer winding, obtain the characteristic parameters under normal winding, short circuit fault, axial displacement and broken wire fault, and establish a sample library to provide training sets and test sets for the subsequent SVM model to classify winding faults.

[0100] The characteristic parameters of the transformer Lissajous ellipse under different winding states are obtained by using the above formulas (1)-(3), and then the above Lissajous characteristic parameter data are labeled with labels 1, 2, ..., m according to the corresponding winding deformation types, so as to establish a characteristic parameter sample library L = {L1, L2, ..., L n}, where the i-th sample is L i ={m,a i ,b i ,θ i}, n is the total sample size.

[0101] Among them, different winding states include normal, short circuit, axial displacement, and broken wire, etc. The characteristic parameters of the Lissajous ellipse are the major axis a, the minor axis b and the tilt angle θ.

[0102] S4: Establish a support vector machine based on the feature parameter sample library, and optimize the intelligent transformer winding deformation classification model of the support vector machine based on the Lissajous ellipse figure and the gray wolf algorithm. The model training and classification based on the sample library constructed by the feature parameters obtained by Lissajous are better than the traditional SVM (support vector machine) model classification effect, and due to the superiority of the Lissajous feature parameters, the model can also achieve good recognition of minor short-circuit faults.

[0103] The method of establishing a support vector machine is to map an m-dimensional input vector To the high-dimensional feature space, and construct an optimal hyperplane in the high-dimensional space. Different mappings construct different support vector machines. In the sample space, the method of dividing the hyperplane is:

[0104] f(x)=ω·x+b=0 (11)

[0105] Where ω is the normal vector, which determines the direction of the hyperplane; b is the intercept constant, which determines the distance between the hyperplane and the origin; x is the feature parameter sample library L = {L1, L2, ..., L n}. The optimal formula for the support vector machine to distinguish categories is:

[0106]

[0107] In the formula, ξ is the slack variable, C is the penalty factor, and the constraints are:

[0108] y i [(ω·x i )+b]≥1-ξ i ,i=1,2…n (13)

[0109] If α i* is the optimal solution, b * is the classification threshold, and the constraint is Then the optimal classification surface is obtained:

[0110]

[0111] In the formula, x i is the i-th sample to be tested in the sample library L i ={m,a i ,b i ,θ i} characteristic parameter data {a i ,b i ,θ i},y i is the i-th sample to be tested L i ={m,a i ,b i ,θ i} label m, K(x i x) is the radial kernel function, which is calculated as follows:

[0112]

[0113] Wherein, g is the optimal hyperparameter. The support vector machine algorithm currently mainly relies on grid search for the determination of the hyperparameter g and the penalty factor C, which requires exhaustive traversal in all hyperparameter combinations and selecting the optimal set of hyperparameters from all combinations, which is difficult to ensure the efficiency of recognition. Therefore, the present invention adopts the gray wolf algorithm to perform parameter optimization on the support vector machine (SVM).

[0114] The specific method of the Gray Wolf Algorithm for optimizing parameters of the support vector machine is:

[0115] The wolf pack is divided into four levels: α, β, γ, and ω. The three wolves with the best fitness are defined as α, β, and γ, and ω is the rest of the wolf pack. Among them, α is the dominant wolf and has a decisive role; β assists α in making decisions and obeys α; γ dominates the remaining ω wolf pack and obeys α and β; ω is the lowest level, obeys α, β, and γ, and plays a role in balancing internal relationships.

[0116] S41: Based on the characteristic parameter sample library L={L1,L2,…,L n The classification error rate of the support vector machine of} is set to the fitness function of the gray wolf optimization algorithm (see formula (16)), and λ is set to the classification error rate at which the gray wolf optimization algorithm terminates.

[0117]

[0118] In formula (16), the fitness value of each wolf is represented by fitness, n is the number of samples to be tested, Ti is the i-th sample L i The actual label, P i is the i-th sample L i The predicted label of .

[0119] S42: Initialize the wolf pack, set the wolf pack size to N, and initialize the wolf pack position in the feasible domain [lb, ub], that is, randomly generate a two-dimensional vector in the feasible domain, which is composed of the hyperparameter g of the support vector machine and the penalty factor C.

[0120] S43: Substitute the randomly generated wolf pack positions into the feature parameter sample library L = {L1, L2, ..., L n}, and thus obtain the fitness value of the wolf pack, and set the top three wolves with the best fitness to α, β and γ in descending order of fitness, and their positions are X α , X β and X γ , the position of the remaining wolves is X.

[0121] S44: Judgement X α Is the fitness value less than λ? If not, define the number of iterations t = 1, and update the position of the wolf pack. The update formula is shown in equations (17)-(20).

[0122]

[0123]

[0124]

[0125]

[0126] Among them, E1, E2, and E3 are random numbers between 0 and 1, L is the position update coefficient at time t, and D α , D β , D γ They are the direction vectors between α wolf, β wolf, γ wolf and the current wolf respectively.

[0127] After updating the position of the wolf pack, make an out-of-bounds adjustment and recalculate the wolf pack fitness value. α , X β and X γ Compare the fitness values ​​of α , X β and X γ location.

[0128] Determine the updated X αIs the fitness value of less than λ? If not, set t = t + 1 and repeat the wolf position update until the updated X α The fitness value is less than λ, and the X obtained in the last iteration is α As the optimal position, it is converted into two parameters, C and g, and substituted into the support vector machine to form an optimized support vector machine classification model.

[0129] S5: Collect the voltage and current on the high-voltage side of the measured phase of the on-site transformer, as well as the voltage on the low-voltage side, and construct a Lissajous ellipse with the calculated voltage amplitude difference and the current amplitude on the high-voltage side. Input the data of the major axis a, minor axis b and tilt angle θ of the ellipse into the classification model, and use the classification model to quickly classify the winding status of the on-site transformer.

[0130] Example 2

[0131] This patent uses a transformer physical model that can simulate various winding deformation faults to establish a transformer Lissajous ellipse figure test platform, and further explains the transformer winding deformation classification method based on Lissajous ellipse figures and gray wolf algorithm optimized support vector machine through the above test platform. The details are as follows:

[0132] First, a potential transformer (PT) and a current transformer (CT) are installed on both the high and low voltage sides of the measured phase of the transformer physical model, and then a 100V AC sinusoidal signal is loaded on the high voltage side of the physical model to simulate the operation of the transformer. Finally, the voltage u1 and current i1 on the high voltage side of the measured phase, as well as the voltage u2 and current i2 on the low voltage side are collected by an oscilloscope to establish a transformer Lissajous ellipse figure test platform, the wiring of which is as follows: Figure 2 shown.

[0133] By using the voltages on the high and low voltage sides of the measured phase, the voltage amplitude difference ΔU=U1-U2 is calculated, and then it is used as the abscissa, and the high voltage side current amplitude I1 of the measured phase is used as the ordinate, and the Lissajous ellipse representing the mechanical state of the transformer winding can be constructed, as shown in Figure 3 As shown. Among them, the formula (21)-(23) is used to obtain Figure 3 The major axis a, minor axis b and inclination angle θ of the mesoellipse can be used as characteristic parameters to characterize the mechanical state of the winding.

[0134]

[0135]

[0136]

[0137] In formulas (21)-(23), X c and Y c is the geometric center of the Lissajous ellipse. The other parameters A, B, C and D are related to the voltages u1 and u2 on the high and low voltage sides, and the current i1 on the high voltage side, as follows:

[0138]

[0139] Among them, ΔU = U1-U2 (where U1 is the amplitude of the high-voltage side voltage u1, and U2 is the amplitude of the low-voltage side voltage u2), I1 is the amplitude of the high-voltage side input current i1, θ is the initial phase, t is the time, and ω is the angular frequency. After trigonometric transformation and Euler formula calculation, formula (24) can be converted into:

[0140]

[0141] Therefore, from equation (25), the parameters A, B, C and D in equations (21)-(23) are:

[0142]

[0143] Since the transformer entity model is in a normal state, the major axis a, the minor axis b and the inclination angle θ of the obtained ellipse are characteristic parameters when the transformer winding is normal.

[0144] Any two high-voltage windings of the measured phase of the transformer physical model are short-circuited to simulate the winding short-circuit fault, and the Lissajous ellipse figure and its characteristic parameters when the winding short-circuit fault occurs are obtained through the above test platform and analytical formula. Then, by adjusting the number of short-circuited windings of the transformer model high-voltage winding, the Lissajous ellipse figure and its characteristic parameters under different winding short-circuit degrees are obtained.

[0145] A capacitor is connected in series between any two high-voltage windings of the measured phase of the transformer physical model to simulate the axial displacement fault of the winding. The Lissajous ellipse and its characteristic parameters when the winding has an axial displacement fault are obtained using the test platform and analytical formula. Then, by adjusting the series capacitance value between the two windings, the Lissajous ellipse and its characteristic parameters under different axial displacement degrees of the winding are obtained.

[0146] Disconnect the connection between any two adjacent discs of the high-voltage winding of the measured phase of the transformer physical model to simulate the winding disconnection fault. Use the test platform and analytical formula to obtain the Lissajous ellipse figure and its characteristic parameters when the winding has a disconnection fault. If the disconnection position of the high-voltage winding is adjusted, the Lissajous ellipse figure and its characteristic parameters when the winding is disconnected at different positions can be obtained.

[0147] Through the above tests, a total of 40 sets of transformer winding Lissajous ellipse figures and their characteristic parameter data were obtained, of which 10 sets were normal winding, short circuit, axial displacement, and broken wire faults. By comparing the Lissajous ellipse figures and their characteristic parameters of some faulty windings with those of normal windings, it can be seen that the differences between the Lissajous ellipse figures and their characteristic parameters of different winding faults are quite obvious, such as Figure 4-Figure 6 As shown in Table 1. Therefore, the characteristic parameters of Lissajous ellipse figures can be applied to the intelligent classification of transformer winding deformation.

[0148] Table 1 Changes in characteristic parameters of Lissajous ellipse figures under different winding faults

[0149]

[0150] Labels 1, 2, 3, and 4 are used to represent the four types of winding states: normal, short circuit, axial displacement, and broken wire. Then, the 40 sets of Lissajous characteristic parameter data are set with corresponding labels according to the corresponding winding types, so as to establish a characteristic parameter sample library L = {L1, L2, ..., L n Some characteristic parameter samples are shown in Table 2.

[0151] Table 2 Some characteristic parameter samples of winding under normal and fault conditions

[0152]

[0153] From the above feature parameter sample library L = {L1, L2, ..., L n}, 10 groups of normal winding data, 4 groups of winding short-circuit data, 5 groups of winding axial displacement data, and 6 groups of winding broken wire data were selected as training sets, and the remaining 6 groups of winding short-circuit data, 5 groups of winding axial displacement data, and 4 groups of winding broken wire data were used as test sets.

[0154] Using the above characteristic parameter sample library L = {L1, L2, ..., L n} training set and the Grey Wolf algorithm to optimize the support vector machine to establish an intelligent transformer winding deformation classification model, as follows.

[0155] The method for establishing the support vector machine is as follows: mapping an n-dimensional input vector To the high-dimensional feature space, and construct an optimal hyperplane in the high-dimensional space. Different mappings construct different support vector machines. In the sample space, the method of dividing the hyperplane is:

[0156] f(x)=ω·x+b=0 (27)

[0157] Where ω is the normal vector, which determines the direction of the hyperplane; b is the intercept constant, which determines the distance between the hyperplane and the origin; x is the feature parameter sample library L = {L1, L2, ..., L n}. The optimal formula for the support vector machine to distinguish categories is:

[0158]

[0159] In the formula, ξ is the slack variable, C is the penalty factor, and the constraints are:

[0160] y i [(ω·x i )+b]≥1-ξ i ,i=1,2…n (29)

[0161] If α i * is the optimal solution, b * is the classification threshold, and the constraint is Then the optimal classification surface is obtained:

[0162]

[0163] In the formula, x i is the i-th sample to be tested in the sample library L i ={m,a i ,b i ,θ i} characteristic parameter data {a i ,b i ,θ i},y i is the i-th sample to be tested L i ={m,a i ,b i ,θ i} label m, K(x i x) is the radial kernel function, which is calculated as follows:

[0164]

[0165] Wherein, sgn() is a sign function, exp() is an exponential function, and g is an optimal hyperparameter. The above-mentioned support vector machine algorithm currently mainly relies on grid search for the determination of hyperparameter g and penalty factor C, and it is necessary to exhaustively traverse all hyperparameter combinations and select the optimal set of hyperparameters from all combinations, which is difficult to ensure the efficiency of recognition. Therefore, the present invention adopts the gray wolf algorithm to perform parameter optimization on the support vector machine (SVM).

[0166] The specific method of the Gray Wolf Algorithm for optimizing parameters of the support vector machine is:

[0167] The wolf pack is divided into four levels: α, β, γ, and ω. The three wolves with the best fitness are defined as α, β, and γ, and ω is the rest of the wolf pack. Among them, α is the dominant wolf and has a decisive role; β assists α in making decisions and obeys α; γ dominates the remaining ω wolf pack and obeys α and β; ω is the lowest level, obeys α, β, and γ, and plays a role in balancing internal relationships.

[0168] Based on the feature parameter sample library L={L1,L2,…,L n The classification error rate of the support vector machine of the training set is set as the fitness function of the gray wolf optimization algorithm (see formula (32)), and λ is the classification error rate at which the gray wolf optimization algorithm terminates, which is set to 0.02.

[0169]

[0170] The fitness value of each wolf is represented by fitness, n is the characteristic parameter sample library L = {L1, L2, ..., L n}, T i is the i-th sample L i The actual label, P i is the i-th sample L i The predicted label of .

[0171] Initialize the wolf pack, set the wolf pack size to 50 and the dimension to 2, and initialize the wolf pack position in the feasible domain [0.01,100]. That is, randomly generate a two-dimensional vector in the feasible domain. The two-dimensional vector is the hyperparameter g and penalty factor C of the support vector machine.

[0172] Substitute the randomly generated wolf pack positions into the sample library based on the characteristic parameters L = {L1, L2, ..., L n}, and thus obtain the fitness value of the wolf pack, and set the top three wolves with the best fitness to α, β and γ in descending order of fitness, and their positions are X α , X β and X γ , the position of the remaining wolves is X.

[0173] Judge X α Is the fitness value less than λ? If not, define the number of iterations t = 1, and update the position of the wolf pack. The update formula is shown in equations (33)-(36).

[0174]

[0175]

[0176]

[0177]

[0178] Among them, E1, E2, and E3 are random numbers between 0 and 1, and L is the position update coefficient at time t.

[0179] After updating the position of the wolf pack, make an out-of-bounds adjustment and recalculate the wolf pack fitness value. α , X β and X γ Compare the fitness values ​​of α , X β and X γ location.

[0180] Judging from the above steps, we get X α Is the fitness value of less than λ? If not, set t = t + 1 and repeat the above steps to update the wolf position until the updated X α The fitness value is less than λ, and the X obtained in the last iteration is α As the optimal position, it is converted into two parameters, C and g, and substituted into the support vector machine to form an optimized support vector machine classification model.

[0181] Using the characteristic parameter sample library L = {L1, L2, ..., L n The test set of} verifies the correctness of the above transformer winding deformation classification model, such as Figure 7 and Figure 8 As shown. Figure 7 and Figure 8 It can be seen that the accuracy of the classification algorithm is good, reaching 100%, and the classification time is short. It only takes 0.6s to complete the model training and the classification of 15 groups of test data. Under the same conditions, the accuracy and running time of the winding deformation classification using grid search optimization support vector machine are 80% and 80% respectively (see Fig. 9 ) and 18s, and its performance is far inferior to the former. Therefore, the transformer winding deformation classification method based on Lissajous ellipse figures and support vector machine optimized by grey wolf algorithm has good field application value.

[0182] The process of the above embodiment is as follows Fig.10 shown.

[0183] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A transformer winding deformation classification method based on Lissajous ellipse graphics and gray wolf algorithm optimized support vector machine, characterized in that the steps include: S1: collect the voltage and current on the high voltage side of the transformer phase under test, as well as the voltage and current on the low voltage side; S2: construct a Lissajous ellipse figure using the voltage amplitude difference between the high and low voltage sides of the measured phase of the transformer and the current amplitude on the high voltage side; S3: obtaining characteristic parameters of the Lissajous ellipse figure of the transformer under different winding states, and establishing a sample library of characteristic parameters representing different states of the transformer winding; S4: Establish a support vector machine based on a feature parameter sample library, and optimize the intelligent transformer winding deformation classification model of the support vector machine based on Lissajous ellipse graphics and the Grey Wolf algorithm; S5: Collect the voltage and current on the high-voltage side of the measured phase of the on-site transformer, as well as the voltage and current on the low-voltage side, and use the intelligent transformer winding deformation classification model to quickly classify the on-site transformer winding status.

2. The transformer winding deformation classification method based on Lissajous ellipse figure and Grey Wolf algorithm optimized support vector machine according to claim 1 is characterized in that: The method for constructing the Lissajous ellipse figure in step S2 is: Using the voltages on the high and low voltage sides of the measured phase, the voltage amplitude difference ΔU=U1-U2 is calculated, and then used as the horizontal coordinate, and the high voltage side current amplitude I1 of the measured phase is used as the vertical coordinate to construct a Lissajous ellipse figure representing the mechanical state of the transformer winding; Among them, the major axis a, minor axis b and inclination angle θ of the Lissajous ellipse obtained by the ellipse formula group are characteristic parameters that characterize the mechanical state of the winding; the characteristic parameter expression is: Among them, X c and Y c It is the geometric center of the Lissajous ellipse. A, B, C and D are parameters. The parameters A, B, C and D are obtained according to the voltage u1 on the high voltage side, the voltage u2 on the low voltage side and the current i1 on the high voltage side.

3. The transformer winding deformation classification method based on Lissajous ellipse figure and Grey Wolf algorithm optimized support vector machine according to claim 2 is characterized in that: The method for obtaining parameters A, B, C and D according to the voltage u1 on the high voltage side, the voltage u2 on the low voltage side and the current i1 on the high voltage side is: Parameters A, B, C and D are related to the voltage u1 on the high voltage side, the voltage u2 on the low voltage side and the current i1 on the high voltage side: Δu(t)=u1(t)-u2(t)=ΔUcos(ωt); Among them, ΔU=U1-U2, U1 is the amplitude of the high-voltage side voltage u1(t), U2 is the amplitude of the low-voltage side voltage u2, I1 is the amplitude of the high-voltage side input current i1(t), is the initial phase, t is the time, ω is the angular frequency; After trigonometric transformation, we get: cos 2 (ωt)+sin 2 (ωt)=1; By eliminating the time parameter, the relationship between the voltage amplitude difference and the input current is obtained: The parameters A, B, C and D are:

4. The transformer winding deformation classification method based on Lissajous ellipse figure and Grey Wolf algorithm optimized support vector machine according to claim 2 or 3, characterized in that: The method for establishing a characteristic parameter sample library characterizing different states of transformer windings in step S3 is: The characteristic parameters of the transformer Lissajous ellipse under different winding states are obtained by using the ellipse formula group. The Lissajous characteristic parameter data are labeled with labels 1, 2, ..., m according to the corresponding winding deformation types, so as to establish a characteristic parameter sample library L = {L1, L2, ..., L n }, where the i-th sample is L i ={m,a i ,b i ,θ i }, n is the total sample size.

5. The transformer winding deformation classification method based on Lissajous ellipse figure and Grey Wolf algorithm optimized support vector machine according to claim 4 is characterized in that: The characteristic parameters of the Lissajous ellipse are the major axis a, the minor axis b and the inclination angle θ; Different winding states include normal, short circuit, axial displacement and broken wire.

6. The transformer winding deformation classification method based on Lissajous ellipse figures and Grey Wolf algorithm optimized support vector machine according to claim 5 is characterized in that: The method for establishing a support vector machine based on a feature parameter sample library in step S4 is: mapping an m-dimensional input vector To a high-dimensional feature space, and construct an optimal hyperplane in the high-dimensional space. Different mappings construct different support vector machines. In the high-dimensional feature space, the method of dividing the hyperplane is: f(x) = ω·x + b = 0; Where ω is the normal vector, b is the intercept constant, and x is the feature parameter sample library L = {L1, L2, ..., L n } in the feature parameters; The optimal formula for the support vector machine to distinguish categories is: In the formula, ξ is the slack variable, C is the penalty factor, and the constraints are: y i [(ω·x i )+b]≥1-ξ i ,i=1,2…n; If α i * is the optimal solution, b * is the classification threshold, and the constraint is Then the optimal classification surface is obtained: In the formula, x i is the i-th sample to be tested in the sample library L i ={m,a i ,b i ,θ i } characteristic parameter data {a i ,b i ,θ i },y i is the i-th sample to be tested L i ={m,a i ,b i ,θ i } label m, K(x i x) is the radial kernel function, which is calculated as follows: Among them, g 2 is the optimal hyperparameter, Sgn() is the sign function, and exp() is the exponential function.

7. The transformer winding deformation classification method based on Lissajous ellipse figures and Grey Wolf algorithm optimized support vector machine according to claim 6 is characterized in that: The method for establishing the intelligent transformer winding deformation classification model based on Lissajous ellipse graphics and Grey Wolf algorithm optimized support vector machine in step S4 is: S41: Establish a gray wolf algorithm model and set the classification error rate of the support vector machine as the fitness function of the gray wolf optimization algorithm; S42: Initialize the wolf pack, set the wolf pack size to N, and initialize the wolf pack position in the feasible domain [lb, ub], and randomly generate a two-dimensional vector in the feasible domain. The two-dimensional vector is composed of the hyperparameter g of the support vector machine and the penalty factor C; S43: Randomly generate wolf pack positions, and substitute the generated wolf pack positions into the support vector machine based on the feature parameter sample library to obtain the fitness value of the wolf pack. The top three wolves with the best fitness are set as α, β and γ in descending order of fitness, and their positions are X α , X β and X γ , the position of the remaining wolves is X; S44: Determine position X α Whether the fitness value of is less than the preset threshold λ, the position of the wolf pack is updated to form an optimized support vector machine classification model.

8. The transformer winding deformation classification method based on Lissajous ellipse figures and Grey Wolf algorithm optimized support vector machine according to claim 7 is characterized in that: The formula for setting the classification error rate of the support vector machine as the fitness function of the gray wolf optimization algorithm in step S41 is: The fitness value of each wolf is represented by fitness, n is the number of samples to be tested, T i is the i-th sample to be tested L i The actual label, P i is the i-th sample L i The predicted label of .

9. The transformer winding deformation classification method based on Lissajous ellipse figures and Grey Wolf algorithm optimized support vector machine according to claim 8, characterized in that: The method for updating the position of the wolf pack in step S44 is: Use the update formula to update the position: Among them, E1, E2, and E3 are random numbers between 0 and 1, and L is the position update coefficient at time t; D α , D β , D γ are the direction vectors between α wolf, β wolf, γ wolf and the current wolf respectively; After updating the position, the wolf pack is adjusted out of bounds and the wolf pack fitness value is recalculated. The new wolf pack fitness value is compared with the position X obtained last time. α , X β and X γ Compare the fitness values ​​of α , X β and X γ location; Determine the updated position X α Is the fitness value less than λ? If not, set t = t + 1 and repeat the wolf position update until the updated position X α The fitness value is less than λ, and the position X obtained in the last iteration is α As the optimal position, it is converted into two parameters, C and g, and substituted into the support vector machine to form an optimized support vector machine classification model.

10. The transformer winding deformation classification method based on Lissajous ellipse figures and Grey Wolf algorithm optimized support vector machine according to any one of claims 7 to 9, characterized in that: The method of using the classification model to quickly classify the on-site transformer winding state in step S5 is: The voltage and current on the high-voltage side of the measured phase of the on-site transformer, as well as the voltage and current on the low-voltage side, are collected. The voltage amplitude difference between the two and the current amplitude on the high-voltage side are calculated to construct a Lissajous ellipse. The major axis a, minor axis b and tilt angle θ of the ellipse are input into the classification model, and the classification model is used to quickly classify the winding status of the on-site transformer.

Images