A method for locating partial overheating fault prone areas of complicated working condition converter transformer

Abstract

The application provides a method for positioning partial overheating fault prone areas of complicated working condition converter transformers, and belongs to the technical field of converter transformers, and comprises the following steps: obtaining an optimal scaling factor group by using a genetic algorithm; scaling the electromagnetic field equation of the converter transformer to be measured by using the obtained optimal scaling factor group to obtain scaling factors of inductance and resistance; establishing a three-dimensional electromagnetic field model based on the optimal scaling factor group to determine the equivalent permeability and conductivity of the core; constructing three working condition models of no fault, multi-point grounding fault and inter-limb short circuit fault, and performing finite element simulation; using a convolutional neural network to extract features and classify the simulation results, and quickly identifying the core loss distribution under different fault working conditions; taking the core loss distribution as a thermal load to establish a temperature field finite element model, performing thermal simulation and obtaining a temperature distribution; positioning the partial overheating fault area; and determining the partial overheating prone area in the converter transformer to be measured according to the optimal scaling factor group.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of converter transformers, and particularly relates to a method for positioning a partial overheating fault prone area of a converter transformer under complex working conditions. BACKGROUND

[0002] As a key equipment in a power system, the working stability of a converter transformer is directly related to the safe operation of the power grid. However, with the increasing complexity of the power system and the growing load, the converter transformer often encounters various fault complex working conditions, such as multi-point grounding fault, inter-phase short circuit fault, etc. The faults under these complex working conditions can cause severe overheating in the internal local area of the converter transformer, greatly shorten the service life of the equipment, and even cause serious accidents. Therefore, accurately identifying and positioning these local overheating fault areas is crucial for the reliable operation and fault diagnosis of the converter transformer.

[0003] The existing converter transformer fault diagnosis technology mainly focuses on the monitoring and analysis of parameters such as vibration, acoustic emission, oil chromatography, etc. of the whole machine. These methods can roughly judge whether the equipment has failed, but it is difficult to accurately locate the specific area of the fault. Another method based on finite element simulation can analyze the electromagnetic characteristics and temperature distribution inside the converter transformer under different fault working conditions by establishing a detailed electromagnetic field and temperature field model. Since finite element analysis needs to minimize the grid size to obtain higher precision simulation results, the hardware configuration requirements for the computer are very high, and a large amount of computing time and storage space are required, which is inefficient. Moreover, the finite element simulation results lack automatic fault identification and classification functions. Usually, manual analysis and interpretation of the electromagnetic field and temperature field distribution obtained by finite element simulation are required, which cannot automatically and quickly identify the characteristic patterns under different fault working conditions, and the labor intensity is high. SUMMARY

[0004] Therefore, the application provides a method for positioning a partial overheating fault prone area of a converter transformer under complex working conditions, which can solve the technical problems that the existing technology requires high hardware configuration of the computer for fine finite element model calculation when positioning the partial overheating fault prone area of the converter transformer under complex working conditions by using finite element, and cannot automatically identify and classify faults.

[0005] The application is implemented as follows:

[0006] The first aspect of the application provides a method for positioning a partial overheating fault prone area of a converter transformer under complex working conditions, which comprises the following steps:

[0007] S10, a plurality of scaled-down models of the converter transformer to be tested are established, and a group of candidate scaling factors corresponding to the plurality of scaled-down models is obtained according to the similarity theory to determine the scaling rules; and an optimal scaling factor group is obtained through optimization.

[0008] S20, an electromagnetic field equation of the to-be-tested converter transformer is established, the electromagnetic field equation is scaled by using the optimal scaling factor group, an inductance scaling factor and a resistance scaling factor of the scaled model are obtained, and the inductance scaling factor and the resistance scaling factor are added to the optimal scaling factor group;

[0009] S30, according to the optimal scaling factor group, a three-dimensional electromagnetic field model of a laminated core of the converter transformer is established by using finite element analysis software, a scaled laminated core three-dimensional electromagnetic field model is obtained, and equivalent permeability and equivalent conductivity of the core are determined;

[0010] S40, a first model simulated under a fault-free condition, a second model simulated under a multi-point grounding fault condition and a third model simulated under an inter-lamination short-circuit fault condition are constructed, and the first, second and third models are subjected to finite element simulation;

[0011] S50, a convolutional neural network is used to extract features and classify the finite element simulation results of the first, second and third models, and to quickly identify core loss distribution under different fault conditions;

[0012] S60, the obtained core loss distribution is taken as a thermal load, temperature field finite element models corresponding to the first, second and third models are established, and a preset convective heat exchange boundary condition is applied to the temperature field models;

[0013] S70, the temperature field finite element models corresponding to the first, second and third models are subjected to thermal simulation calculation, and core temperature distributions under the three conditions are obtained;

[0014] S80, an unsupervised learning clustering algorithm is used to identify temperature abnormal areas under different fault conditions according to the obtained core temperature distributions under the three conditions, and to locate a local overheating fault area;

[0015] S90, according to the optimal scaling factor group, a local overheating prone area in the to-be-tested converter transformer is obtained based on the located local overheating fault area.

[0016] The complex conditions include a fault-free condition, a multi-point grounding fault condition and an inter-lamination short-circuit fault condition.

[0017] The first model simulated under the no-fault working condition, and the second model simulated under the multi-point grounding fault and the third model simulated under the inter-sheet short-circuit fault condition are specifically: for the no-fault working condition, the equivalent parameters obtained in the step S20 are used to perform non-fault electromagnetic field modeling on the scaled model, which is recorded as the first model; for the multi-point grounding fault condition, copper sheets are inserted in the non-fault modeling and short-circuit processing is performed thereon, thereby constructing an electromagnetic field model of the multi-point grounding fault, which is recorded as the second model; for the inter-sheet short-circuit fault condition, short-circuit regions and short-circuit fault points of the iron core are set in the non-fault modeling, thereby constructing an electromagnetic field model of the inter-sheet short-circuit fault, which is recorded as the third model.

[0018] The scaled factor group includes a size scaled factor, a current density scaled factor, an electric displacement vector scaled factor, an electric field intensity scaled factor and a magnetic field intensity scaled factor.

[0019] The method for optimizing the optimal scaled factor group is a genetic algorithm.

[0020] The step S10 specifically includes:

[0021] In step 101, a plurality of groups of geometric scaling mathematical models are established as scaled models according to size and structure parameters of the to-be-tested converter transformer, so as to reflect electromagnetic field distribution characteristics of the converter transformer under different scales.

[0022] In step 102, size scaled factors, current density scaled factors, electric displacement vector scaled factors, electric field intensity scaled factors and magnetic field intensity scaled factors are determined as a group of to-be-selected scaled factor combinations through similarity theory analysis.

[0023] Optionally, in step 103, a genetic algorithm is used for iterative optimization to find an optimal scaled factor combination that makes the scaled model most consistent with the actual converter transformer, so as to obtain an optimal scaled model for accurately describing electromagnetic field distribution of the to-be-tested converter transformer. Optionally, the method for optimizing the optimal scaled factor group can also be selected according to expert opinions.

[0024] The plurality of groups of geometric scaling mathematical models can reflect electromagnetic field distribution characteristics of the converter transformer under different scales.

[0025] The specific steps of the step S20 include:

[0026] In step 201, an equivalent electromagnetic field equation of the converter transformer is established based on geometric structure and material parameters of the converter transformer.

[0027] In step 202, the optimal scaled factor group obtained in the step S10 is used to scale the electromagnetic field equation, so as to obtain inductance scaling factors and resistance scaling factors of the scaled model.

[0028] Step 203, add the inductance scaling factor and the resistance scaling factor to the optimal scaling factor group to obtain a more complete scaling parameter set.

[0029] Wherein, the established equivalent electromagnetic field equation of the converter transformer reflects the geometric structure and material parameters of the converter transformer; the obtained inductance scaling factor and resistance scaling factor reflect the corresponding relationship between the scaled model and the actual converter transformer in electromagnetic characteristics; the established three-dimensional electromagnetic field model of the laminated core of the converter transformer can accurately describe the electromagnetic field distribution of the core under the scaling condition.

[0030] The step S30 specifically comprises:

[0031] Step 301, using the complete scaling parameter set obtained in step S20, a three-dimensional electromagnetic field model of the laminated core of the converter transformer is established in the finite element analysis software;

[0032] Step 302, through finite element analysis, the equivalent permeability and equivalent conductivity of the core under the scaling condition are obtained.

[0033] The step S40 specifically comprises:

[0034] Step 401, using the equivalent permeability and equivalent conductivity obtained in step S30, a first electromagnetic field model under a fault-free condition is established;

[0035] Step 402, on the basis of the first model, a short-circuit copper sheet is inserted to construct a second electromagnetic field model of a multi-point grounding fault;

[0036] Step 403, in the first model, a short-circuit region of the core and a short-circuit fault point are set to construct a third electromagnetic field model of an inter-lamination short-circuit fault. The first model under the fault-free condition established in step 401 reflects the electromagnetic field characteristics of the converter transformer under normal working condition; the electromagnetic field models of the multi-point grounding fault and the inter-lamination short-circuit fault constructed in steps 402 and 403 respectively simulate the electromagnetic field behavior of the converter transformer under different fault conditions.

[0037] Wherein, the clustering algorithm using unsupervised learning is k-means clustering.

[0038] Wherein, the steps of using a convolutional neural network to extract features and classify the finite element simulation results of the first, second and third models are as follows: a CNN model containing multiple convolutional layers, pooling layers and fully connected layers is constructed, the input is the electromagnetic field distribution image under the three working conditions of fault-free condition, multi-point grounding fault condition and inter-lamination short-circuit fault condition, after convolution and pooling operation, multi-scale features are extracted, then the fully connected layer is used to classify these features, and the core loss distribution under different fault conditions is output.

[0039] The step S50 specifically comprises:

[0040] Step 501, using a convolutional neural network to extract features from the finite element simulation results under the three working conditions obtained in step S40;

[0041] Step 502, using a convolutional neural network to classify the extracted features to quickly identify the core loss distribution characteristics under different fault conditions.

[0042] The step S60 specifically comprises:

[0043] Step 601, taking the core loss distribution obtained in step S50 as the thermal load, and constructing the temperature field finite element model corresponding to the first, second and third models respectively;

[0044] Step 602, in the temperature field model, a predetermined convective heat transfer boundary condition is applied to simulate the heat exchange process between the converter transformer and the surrounding environment.

[0045] The step S70 specifically comprises:

[0046] Step 701, performing thermal simulation calculation on the temperature field finite element models under the three working conditions established in step S60;

[0047] Step 702, obtaining the core temperature distribution under the three working conditions.

[0048] The step S80 specifically comprises:

[0049] Step 801, using a k-means clustering algorithm to identify the temperature anomaly area under different fault conditions according to the core temperature distribution under the three working conditions obtained in step S70;

[0050] Step 802, automatically finding the abnormal point group in the temperature distribution through the clustering algorithm to locate the local overheating fault area.

[0051] The step S90 specifically comprises:

[0052] Step 901, according to the optimal scaling factor group obtained in step S10, the local overheating fault area located in step S80 is mapped to the specific position of the actual converter transformer;

[0053] Step 902, determining the local overheating prone area in the converter transformer to be tested, providing an important reference for the design and fault diagnosis of the converter transformer.

[0054] Further, the initial population of the genetic algorithm is a plurality of scaling factor groups; the fitness function is used to calculate the deviation between the scaling models corresponding to the scaling factor groups, including magnetic field intensity deviation and electric field intensity deviation.

[0055] Wherein, the formula of the fitness function is: f(X) = ‖B sim (X) - B real ‖ 2 + λ‖E sim (X) - E real ‖ 2 ; wherein, B sim (X) and E sim (X) respectively represent the magnetic field strength and electric field strength obtained by finite element simulation under the scale factor combination X; B real and E real respectively represent the measured magnetic field strength and electric field strength of the actual converter transformer; λ is an adjustment parameter, used to balance the weight between the magnetic field and the electric field.

[0056] Further, the crossover operation in the genetic algorithm is realized by using an arithmetic crossover method, specifically, two parent individuals are randomly selected in the population to realize crossover by using a crossover coefficient to generate two new offspring individuals.

[0057] Further, the termination condition in the genetic algorithm is to terminate when the maximum number of iterations is reached, wherein the maximum number of iterations = 10 × number of scale factor groups × number of elements in each scale factor group.

[0058] Compared with the prior art, the beneficial effects of the complex working condition converter transformer local overheating fault prone area positioning method provided by the application are:

[0059] 1. The local overheating area of the converter transformer under different fault conditions can be accurately positioned. By establishing multiple groups of geometric scale models, using a genetic algorithm to find the optimal scale parameters, and combining finite element analysis, convolutional neural network feature extraction, and unsupervised clustering techniques, the temperature abnormal area under different fault conditions can be quickly identified, providing a reliable basis for fault diagnosis of the converter transformer.

[0060] 2. The modeling and simulation process is more efficient. This method fully utilizes the similarity theory, describes the electromagnetic field distribution of the converter transformer under different scales by establishing scale models, avoids the cumbersome full-size modeling process, and reduces the computer hardware configuration requirements. At the same time, the convolutional neural network is used to extract and classify the features of the finite element simulation results, greatly improving the speed of fault area identification and providing convenience for practical application.

[0061] 3. The method combines numerical simulation and machine learning techniques to enhance the accuracy of diagnosis. The method combines finite element analysis to model the electromagnetic field and temperature field of the converter transformer under different fault conditions, and then uses convolutional neural networks and clustering algorithms to automatically identify and classify the simulation results, fully utilizing the advantages of both techniques to improve the accuracy of local overheating fault area positioning.

[0062] 4. The method provides important support for the design optimization and fault diagnosis of converter transformers. By identifying the local overheating prone area, the structure optimization design of the converter transformer can be guided to reduce the risk of local overheating. At the same time, the method can also be applied to the fault diagnosis of in-service equipment, providing important reference for maintenance management.

[0063] In summary, the method for positioning the local overheating prone area of the converter transformer under complex working conditions proposed in the present application fully utilizes numerical simulation and machine learning techniques to achieve rapid and accurate positioning of the local overheating fault area of the converter transformer, solving the technical problems of existing technology in positioning the local overheating prone area of the converter transformer under complex working conditions using finite elements, which requires high hardware configuration of the computer for fine finite element modeling, and cannot automatically identify and classify faults. BRIEF DESCRIPTION OF DRAWINGS

[0064] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the description of the embodiments of the present application. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0065] Figure 1 The flowchart of the method provided by the present application is shown in Figure 1.

[0066] Figure 2 The magnetic flux simulation diagram under the multi-point grounding fault condition is shown in Figure 2.

[0067] Figure 3 The eddy current simulation diagram under the multi-point grounding fault condition is shown in Figure 3.

[0068] Figure 4 The magnetic flux simulation diagram under the inter-plate short circuit fault condition is shown in Figure 4.

[0069] Figure 5 The eddy current simulation diagram under the inter-plate short circuit fault condition is shown in Figure 5.

[0070] Figure 6 The eddy current loss simulation diagram under the inter-plate short circuit fault condition is shown in Figure 6.

[0071] Figure 7 The temperature distribution comparison diagram of the reduced scale model formed by the unoptimized and optimized reduced scale factor groups is shown in Figure 7.

[0072] Figure 8 is a temperature distribution diagram in a multi-point grounding fault state;

[0073] Figure 9 is a temperature distribution diagram in a interlamination short circuit fault state. DETAILED DESCRIPTION

[0074] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application.

[0075] As shown in Figure 1 is a flow chart of a method for positioning a partial overheating fault prone area of a complicated working condition converter transformer provided by the present application, and the method comprises the following steps:

[0076] S10, a plurality of scaled-down models of the converter transformer to be measured are established, a plurality of groups of candidate scaling factors corresponding to the scaled-down models are obtained according to a scaling rule determined according to a similarity theory, the groups of candidate scaling factors comprise a size scaling factor, a current density scaling factor, a potential displacement vector scaling factor, an electric field intensity scaling factor and a magnetic field intensity scaling factor; and an optimal group of scaling factors is obtained by using a genetic algorithm.

[0077] S20, an electromagnetic field equation of the converter transformer to be measured is established, the electromagnetic field equation is scaled by using the optimal group of scaling factors, an inductance scaling factor and a resistance scaling factor of the scaled-down model are obtained, and the inductance scaling factor and the resistance scaling factor are added to the optimal group of scaling factors.

[0078] S30, according to the optimal group of scaling factors, a three-dimensional electromagnetic field model of the laminated core of the converter transformer is established by using a finite element analysis software, a scaled laminated core three-dimensional electromagnetic field model is obtained, and the equivalent permeability and the equivalent conductivity of the core are determined.

[0079] S40, a first model under a fault-free working condition, a second model under a multi-point grounding fault and a third model under an interlamination short circuit fault working condition are constructed, and the first, second and third models are subjected to finite element simulation.

[0080] S50, the convolutional neural network is used to extract features and classify the finite element simulation results of the first, second and third models, and the core loss distribution under different fault working conditions is quickly identified.

[0081] S60, the obtained core loss distribution is taken as a thermal load, temperature field finite element models corresponding to the first, second and third models are established, and a preset convective heat exchange boundary condition is applied to the temperature field model.

[0082] S70, thermal simulation calculation is performed on the temperature field finite element models corresponding to the first, second and third models respectively to obtain the core temperature distribution under three working conditions;

[0083] S80, a clustering algorithm of unsupervised learning is used to identify the temperature abnormal area under different fault working conditions according to the obtained core temperature distribution under three working conditions, and the local overheating fault area is located;

[0084] S90, according to the optimal scaling factor group and based on the located local overheating fault area, the local overheating prone area in the to-be-tested converter transformer is obtained.

[0085] The specific implementation of the above steps is described in detail as follows:

[0086] The specific implementation of step S10 is as follows: first, a plurality of groups of geometric scaling mathematical models are established according to the size and structure parameters of the converter transformer. These scaling models can reflect the electromagnetic field distribution characteristics of the converter transformer under different scales. Through similarity theory analysis, various scaling factors are determined, including:

[0087] 1) size scaling factor: used for scaling the geometric size, reflecting the spatial scaling of electromagnetic field distribution.

[0088] 2) current density scaling factor: used for scaling the current density, reflecting the current scaling of electromagnetic field distribution.

[0089] 3) electric displacement vector scaling factor: used for scaling the electric displacement vector, reflecting the scaling of electric field distribution.

[0090] 4) electric field intensity scaling factor: used for scaling the electric field intensity, reflecting the scaling of electric field distribution.

[0091] 5) magnetic field intensity scaling factor: used for scaling the magnetic field intensity, reflecting the scaling of magnetic field distribution.

[0092] These scaling factors constitute a group of to-be-selected scaling factor combinations. Then, a genetic algorithm is used to find the optimal scaling factor combination that best matches the model with the actual converter transformer through iterative optimization. The purpose of this step is to find the optimal scaling model that can accurately describe the electromagnetic field distribution of the to-be-tested converter transformer.

[0093] The specific implementation of step S20 is as follows: first, based on the geometric structure and material parameters of the converter transformer, an equivalent electromagnetic field equation is established. Then, the optimal scaling factor group obtained in step S10 is used to scale the electromagnetic field equation, thereby obtaining the inductance scaling factor and the resistance scaling factor of the scaled model. These factors reflect the trend of the electromagnetic parameters when the model is scaled. The inductance scaling factor and the resistance scaling factor are added to the optimal scaling factor group to obtain a more complete scaling parameter set. The purpose of this step is to determine the correspondence between the scaled model and the actual converter transformer in terms of electromagnetic characteristics.

[0094] The specific implementation of step S30 is as follows: using the complete scaling parameter set obtained in step S20, a three-dimensional electromagnetic field model of the laminated core of the converter transformer is established in the finite element analysis software. Through finite element analysis, the equivalent permeability and equivalent conductivity of the core under scaling conditions can be obtained. The purpose of this step is to obtain accurate data of the core material parameters in the scaled model, laying a foundation for subsequent thermal field analysis.

[0095] The specific implementation of step S40 is as follows: first, a first model under a fault-free condition is constructed. By applying the equivalent parameters obtained in step S20 to the scaled model, an electromagnetic field model under fault-free conditions can be established. Second, for the multi-point grounding fault condition, a short-circuit copper sheet is inserted into the core based on the fault-free model to simulate the electromagnetic field behavior of the multi-point grounding fault, obtaining a second model. Third, for the inter-lamination short-circuit fault condition, a short-circuit region and a short-circuit fault point are set in the fault-free model to construct a third model. These three models correspond to typical conditions such as fault-free, multi-point grounding fault, and inter-lamination short-circuit fault, providing a basis for subsequent thermal field analysis and fault location.

[0096] The specific implementation of step S50 is as follows: a convolutional neural network is used to extract features and classify the finite element simulation results under the three conditions obtained in step S40. Convolutional neural networks are a deep learning algorithm that is good at extracting image features and classification. By learning and classifying the electromagnetic field distribution characteristics under the three conditions, the core loss distribution characteristics under different fault conditions can be quickly identified. The purpose of this step is to obtain the core loss distribution under different fault conditions, providing thermal load conditions for subsequent temperature field analysis.

[0097] The specific implementation of step S60 is as follows: the core loss distribution obtained in step S50 is used as the thermal load to construct the temperature field finite element model corresponding to the first, second, and third models. In the temperature field model, a pre-set convective heat exchange boundary condition is applied to simulate the heat exchange process between the converter transformer and the surrounding environment. The purpose of this step is to establish a temperature field distribution model of the converter transformer under different fault conditions, providing a basis for subsequent fault area identification.

[0098] The specific implementation of step S70 is: performing thermal simulation calculation on the temperature field finite element models in the three working conditions established in step S60 to obtain the core temperature distribution in the three working conditions. Through thermal simulation, the actual temperature field distribution of the converter transformer under different fault conditions can be simulated, thereby providing a basis for subsequent temperature abnormal area identification.

[0099] The specific implementation of step S80 is: using an unsupervised learning clustering algorithm, such as k-means clustering, to identify the temperature abnormal area under different fault conditions according to the core temperature distribution in the three working conditions obtained in step S70. The clustering algorithm can automatically find the abnormal point group in the temperature distribution, thereby locating the local overheating fault area. The purpose of this step is to quickly and accurately find the temperature abnormal area in the converter transformer under different fault conditions, thereby providing a basis for subsequent fault area positioning.

[0100] The specific implementation of step S90 is: determining the local overheating prone area in the converter transformer to be measured according to the optimal scaling factor group obtained in step S10 and the local overheating fault area located in step S80. By applying the optimal scaling factor group to the geometric model of the actual converter transformer, the fault area located in step S80 can be mapped to the specific position of the actual converter transformer, thereby obtaining the distribution of the local overheating prone area. The purpose of this step is to provide an important reference for the design and fault diagnosis of the converter transformer, and to point out the prone fault area that needs to be focused on.

[0101] In order to better understand the present application, the specific embodiments of the present application will be described in more detail below in conjunction with specific formulas:

[0102] The specific implementation of step S10 is as follows:

[0103] First, a plurality of geometric scaling mathematical models are established according to the size and structural parameters of the converter transformer to be measured. These scaling models can reflect the electromagnetic field distribution characteristics of the converter transformer under different scales. Through similarity theory analysis, various scaling factors can be determined, including:

[0104] 1) Size scaling factor k s : used for scaling the geometric size, reflecting the spatial scaling of the electromagnetic field distribution.

[0105] 2) Current density scaling factor k j : used for scaling the current density, reflecting the current scaling of the electromagnetic field distribution.

[0106] 3) Electric displacement vector scaling factor k d : used for scaling the electric displacement vector, reflecting the scaling of the electric field distribution.

[0107] 4) Electric field intensity scaling factor k e : Used to scale electric field intensity, reflecting the scaling of electric field distribution.

[0108] 5) Magnetic field intensity scaling factor k h : Used to scale magnetic field intensity, reflecting the scaling of magnetic field distribution.

[0109] These scaling factors form a set of candidate scaling factor combinations {k s , k j , k d , k e , k h}.

[0110] Then, a genetic algorithm is used to iteratively optimize this set of candidate scaling factors to find the optimal scaling factor combination that best fits the model to the actual converter transformer The genetic algorithm is an optimization algorithm that simulates the natural evolution process, through operations such as selection, crossover and mutation, iteratively updating the scaling factor combination until the optimal solution is reached. The purpose of this step is to find the optimal scaling model that can accurately describe the electromagnetic field distribution of the converter transformer to be tested. The specific implementation of using a genetic algorithm to optimize the scaling factor group will be described in detail below.

[0111] 1. Population initialization

[0112] First, a population containing multiple individuals needs to be initialized. Each individual corresponds to a set of scaling factor combinations, represented as X = {k s , k j , k d , k e , k h}. The initial population can be generated randomly or constructed based on prior knowledge. The population size is usually set to a large value to ensure the diversity of the population.

[0113] 2. Individual fitness evaluation

[0114] For each individual, the ability to adapt to the current problem needs to be evaluated, i.e. the fitness function value. The design of the fitness function is critical, as it reflects the degree of individual solving the target problem.

[0115] The fitness function is specifically:

[0116] f(X) = ‖B sim (X) - B real ‖ 2 + λ ‖E sim (X) - E real ‖ 2

[0117] Where:

[0118] X represents a scaling factor combination, including a size scaling factor k s , a current density scaling factor k j , a potential displacement vector scaling factor k d , an electric field strength scaling factor k e , and a magnetic field strength scaling factor k h .

[0119] B sim (X) represents the magnetic field strength distribution obtained by finite element simulation under the scaling factor combination X.

[0120] B real represents the magnetic field strength distribution measured by the actual converter transformer.

[0121] E sim (X) represents the electric field strength distribution obtained by finite element simulation under the scaling factor combination X.

[0122] E real represents the electric field strength distribution measured by the actual converter transformer.

[0123] λ is a tuning parameter used to balance the weights between the magnetic field and the electric field.

[0124] The essence of the fitness function is to calculate the deviation between the scaled model and the actual situation, including the magnetic field strength deviation and the electric field strength deviation. Specifically:

[0125] ‖B sim (X)-B real ‖ 2 The Euclidean distance square between the scaled model magnetic field strength and the actual magnetic field strength is calculated, reflecting the degree of deviation of the magnetic field distribution.

[0126] ‖E sim (X)-E real ‖ 2 The Euclidean distance square between the scaled model electric field strength and the actual electric field strength is calculated, reflecting the degree of deviation of the electric field distribution.

[0127] The sum of the two parts is the total deviation, which is the value of the fitness function.

[0128] The smaller the fitness function value, the smaller the deviation between the scaled model and the actual situation, and the higher the fitness of the individual. Conversely, the larger the fitness function value, the larger the deviation, and the lower the fitness.

[0129] In the evolution process of the genetic algorithm, individuals with smaller fitness function values (i.e., scaling factor combinations that can better describe the actual electromagnetic field distribution) will be preferentially selected, thereby gradually optimizing the best scaling factor combination.

[0130] By adjusting the value of λ, the weight of the two parts of magnetic field and electric field in the fitness function can be changed. When λ is large, the contribution of electric field distribution is large; when λ is small, the contribution of magnetic field distribution is large. Therefore, the value of λ can be set reasonably according to the needs of specific problems, and the purpose of balancing the two is achieved.

[0131] In summary, the design of the above fitness function can effectively evaluate the difference between the scale model and the actual situation, and quantify this difference as a numerical value, providing a target basis for the optimization process of the genetic algorithm. The reasonable design of the fitness function is crucial for the improvement of the performance of the algorithm.

[0132] The smaller the fitness function value, the smaller the deviation between the scale model and the actual situation, and the higher the fitness of the individual. For each individual in the population, the fitness function value needs to be calculated.

[0133] 3. Selection operation

[0134] According to the fitness function value of the individual, a certain selection strategy is used to select a batch of individuals from the current population as parent individuals for generating the next generation population. Common selection strategies include roulette selection, tournament selection, etc. Generally speaking, the probability of being selected by individuals with higher fitness function values is greater.

[0135] 4. Crossover operation

[0136] Two individuals are randomly selected from the parent individuals, and two new offspring individuals are generated by a certain crossover method. The purpose of crossover is to combine the excellent genes of the parent individuals together to produce better offspring individuals. Common crossover methods include single-point crossover, multi-point crossover, uniform crossover, etc.

[0137] For the scale factor combination X = {k s ,k j ,k d ,k e ,k h} in this example, the arithmetic crossover method can be used. Let the two parent individuals be X1 and X2, and the crossover coefficient be α ∈ [0, 1], then two offspring individuals can be generated:

[0138] X′1 = αX1 + (1-α)X2

[0139] X′2 = (1-α)X1 + αX2

[0140] Crossover operation can generate new scale factor combinations and increase the diversity of the population.

[0141] Specifically, for the scale factor combination X = {k s ,k j ,kd ,k e ,k h}, the crossover operation is performed in the arithmetic crossover manner.

[0142] First, two parent individuals are randomly selected from the current population, and assume that they are X1 and X2, respectively. and

[0143] Then, a crossover coefficient a is introduced, which is a real number between 0 and 1. The value of the crossover coefficient a determines how many genes the newly generated offspring individuals will inherit from the parent individuals. Generally, the crossover coefficient a can be set to a fixed value, such as 0.5, indicating that the new offspring individuals will each inherit half of the genes from the two parent individuals. Alternatively, the value of a can be randomly generated according to certain rules or probability distributions to increase the diversity of the population.

[0144] According to the crossover coefficient a, two new offspring individuals X'1 and X'2 can be generated, which are composed of the linear combination of the parent individuals X1 and X2, respectively. The specific calculation method is as follows:

[0145] X'1 = aX1 + (1-a)X2

[0146] X'2 = (1-a)X1 + aX2

[0147] Taking a = 0.5 as an example, assume that the parent individuals are:

[0148] X1 = {2.0, 1.5, 0.8, 3.2, 1.1}

[0149] X2 = {1.7, 1.2, 0.9, 2.8, 1.4}

[0150] Then, the two offspring individuals generated by the crossover operation are:

[0151] X'1 = 0.5 × {2.0, 1.5, 0.8, 3.2, 1.1} + 0.5 × {1.7, 1.2, 0.9, 2.8, 1.4} = {1.85, 1.35, 0.85, 3.0, 1.25}

[0152] X'2 = 0.5 × {1.7, 1.2, 0.9, 2.8, 1.4} + 0.5 × {2.0, 1.5, 0.8, 3.2, 1.1} = {1.85, 1.35, 0.85, 3.0, 1.25}

[0153] As can be seen, the newly generated offspring individuals X'1 and X'2 respectively inherit part of the genes from the parent individuals X1 and X'2, thereby realizing the recombination of genes.

[0154] It should be noted that in practice, a crossover probability p is usually set for the crossover operation, which controls the frequency of the crossover operation. For each pair of selected parent individuals, a crossover operation will occur with probability p, generating new offspring individuals; with probability 1-p, the original parent individuals are retained without crossover. The crossover probability p is usually set between 0.6 and 0.9, and needs to be adjusted and optimized according to the specific problem. c c c c

[0155] Through the above crossover operation, the algorithm can generate new scaling factor combinations, increase the diversity of the population, and thus help to find better solutions. At the same time, since the new offspring individuals inherit some of the excellent genes of the parent individuals, they are expected to have higher fitness, pushing the algorithm to evolve in a better direction.

[0156] 5. Mutation operation

[0157] In order to further increase the diversity of the population and prevent premature convergence to a local optimal solution, mutation operation is needed on the offspring individuals. Mutation is achieved by changing one or more positions of the individual's genes. For the scaling factor combination in this example, the following mutation method can be used:

[0158] For offspring individual X' = {k s ,k j ,k d ,k e ,k h}, with mutation probability p m , one or more scaling factors are randomly selected, such as k j , and a value is randomly taken within a mutation range around it, such as:

[0159]

[0160] where N represents a Gaussian distribution with mean 0 and standard deviation σ j . Through mutation operation, offspring individuals can be randomly changed within a certain range, increasing the diversity of the population.

[0161] 6. Population update

[0162] After selection, crossover and mutation operations, a batch of new offspring individuals will be obtained. According to certain replacement strategies, such as elite preservation strategy, full reset strategy, etc., individuals with higher fitness are selected from the parent population and offspring population to form the new generation of population.

[0163] 7. Termination condition

[0164] ​​​​Repeat steps 3-6 until the termination condition is met. The termination condition can be set to reach the maximum number of iterations, the fitness function value is less than a certain threshold, or the fitness function value of consecutive generations has no significant improvement, etc. When the termination condition is met, take the individual with the minimum fitness function value in the current population, i.e. the corresponding scaling factor combination as the optimal solution. In this scheme, the default maximum number of iterations is set to 10 times the number of scaling factor combinations times the number of elements in each scaling factor combination, or the maximum number of iterations can also be set to 300 by default.

[0165] Through the above genetic algorithm process, the initial population can be continuously evolved and gradually approach the optimal scaling factor combination. Genetic algorithm has global optimization ability and can effectively avoid falling into local minimum value. At the same time, through crossover and mutation operations, the diversity of the population can be maintained and the robustness of the algorithm can be enhanced.

[0166] It is worth noting that the performance of genetic algorithm is closely related to parameter settings, including population size, crossover probability, mutation probability, selection strategy, etc., which need to be debugged and optimized according to specific problems. In addition, the design of the fitness function is also crucial, which directly determines the convergence performance of the algorithm. In this example, the fitness function considers both magnetic field and electric field, and the weight of the two can be balanced by adjusting λ.

[0167] The specific implementation of step S20 is as follows:

[0168] First, based on the geometric structure and material parameters of the converter transformer, its equivalent electromagnetic field equation is established. The electromagnetic field equation can be expressed as:

[0169]

[0170]

[0171]

[0172]

[0173] where, is the magnetic field intensity, is the current density, is the electric displacement vector, is the magnetic induction intensity, is the electric field intensity, ρ is the charge density, and t represents time.

[0174] Then, the optimal scaling factor combination obtained in step S10 is used to scale the electromagnetic field equation, thereby obtaining the inductance scaling factor k L and the resistance scaling factor k R of the scaled model. These factors reflect the trend of electromagnetic parameter changes when the model is scaled down.

[0175]

[0176]

[0177] where L m and L p are the inductances of the scaled model and the actual converter transformer, respectively, R m and R p are the resistances of the scaled model and the actual converter transformer, respectively.

[0178] The inductance scaling factor k L and the resistance scaling factor k R are added to the optimal scaling factor group to obtain a more complete scaling parameter set The purpose of this step is to determine the correspondence between the scaled model and the actual converter transformer in electromagnetic characteristics.

[0179] The specific implementation of step S30 is as follows:

[0180] The complete scaling parameter set obtained in step S20 is used to build a three-dimensional electromagnetic field model of the laminated core of the converter transformer in the finite element analysis software. Specifically, the following three-dimensional finite element equation can be constructed:

[0181]

[0182] where A is the magnetic vector potential, μ r is the relative permeability, σ is the electrical conductivity, and J s is the source current density.

[0183] Through finite element analysis, the equivalent permeability μ r,eq and the equivalent conductivity σ eq of the core under scaling conditions can be obtained. These parameters reflect the changes in electromagnetic characteristics of the core material under reduced scale, laying the foundation for subsequent thermal field analysis.

[0184] The specific implementation of step S40 is as follows:

[0185] 1) For the fault-free condition, the equivalent parameters k L and k R obtained in step S20 are used to model the non-fault electromagnetic field of the scaled model. The electromagnetic field equation under the fault-free condition can be expressed as:

[0186]

[0187] 2) For the multi-point grounding fault condition, a short-circuit copper sheet is inserted in the no-fault modeling and short-circuit processing is performed to construct an electromagnetic field model of the multi-point grounding fault. Under the multi-point grounding fault, the electromagnetic field equation can be expressed as:

[0188]

[0189] where J fault is the fault current density.

[0190] 3) For the inter-limb short-circuit fault condition, a short-circuit region and a short-circuit fault point of the core are set in the no-fault modeling to construct an electromagnetic field model of the inter-limb short-circuit fault. Under the inter-limb short-circuit fault, the electromagnetic field equation can be expressed as:

[0191]

[0192] where J′ fault is the inter-limb short-circuit fault current density.

[0193] The three models correspond to typical conditions such as no fault, multi-point grounding fault and inter-limb short-circuit fault, respectively, providing a basis for subsequent thermal field analysis and fault location.

[0194] The specific implementation of step S50 is as follows:

[0195] A convolutional neural network (CNN) is used to extract features and classify the finite element simulation results under the three conditions obtained in step S40. The convolutional neural network is a deep learning algorithm that is good at extracting image features and classification.

[0196] Specifically, a CNN model containing multiple convolutional layers, pooling layers and fully connected layers can be constructed. The input is the electromagnetic field distribution image under the three conditions, which is subjected to a series of convolution and pooling operations to extract multi-scale features. Then, the fully connected layer is used to classify these features, and the core loss distribution under different fault conditions is output.

[0197] Let the parameters of the CNN model be θ, the input electromagnetic field distribution be X, and the true core loss distribution label be Y. The training objective of the CNN can be expressed as:

[0198]

[0199] where L is the loss function, such as cross-entropy loss, and N is the number of training samples. The parameter θ is optimized by the backpropagation algorithm so that the CNN model can accurately identify the core loss distribution features under different fault conditions.

[0200] The purpose of this step is to obtain the core loss distribution under different fault conditions to provide thermal load conditions for subsequent temperature field analysis.

[0201] The specific implementation of step S60 is as follows:

[0202] 1) The core loss distribution P obtained in step S50 is used as the heat load of the converter transformer. loss As the heat load, the finite element models of the temperature fields corresponding to the first, second and third models are respectively constructed. The temperature field equation can be expressed as:

[0203]

[0204] Wherein, p is the density, c p is the specific heat capacity, k is the thermal conductivity, and T is the temperature.

[0205] 2) In the temperature field model, a preset convective heat transfer boundary condition is applied to simulate the heat exchange process between the converter transformer and the surrounding environment. The convective heat transfer boundary condition can be expressed as:

[0206]

[0207] Wherein, h is the convective heat transfer coefficient, and T ∞ is the ambient temperature.

[0208] The purpose of this step is to establish the temperature field distribution model of the converter transformer under different fault conditions, and to provide a basis for subsequent fault area identification.

[0209] The specific implementation of step S70 is as follows:

[0210] 1) The temperature field finite element models under the three working conditions established in step S60 are respectively subjected to thermal simulation calculation, and the core temperature distributions T1(x,y,z), T2(x,y,z), T3(x,y,z) under the three working conditions are obtained.

[0211] The solving process of thermal simulation can be expressed as:

[0212] Under the initial condition T(x,y,z,0)=T0 and the boundary condition , the temperature field equation is solved to obtain the temperature distribution under the three working conditions.

[0213] 2) Through thermal simulation, the actual temperature field distribution of the converter transformer under different fault conditions can be simulated, which provides a basis for subsequent temperature abnormal area identification.

[0214] Specifically, the principle of the present application is:

[0215] 1. Application of similarity theory in scaled model establishment. As a large power equipment, the structure and size of converter transformer are usually complex. Full-scale modeling and simulation not only have large amount of calculation, but also it is difficult to obtain all the specific parameter information. The invention uses similarity theory to reflect the electromagnetic field distribution characteristics of converter transformer at different scales through the establishment of geometric scaled model, which greatly reduces the complexity of modeling and simulation.

[0216] 2. Application of genetic algorithm in solving optimal scaling parameters. The geometric structure and material parameters of converter transformer will affect its electromagnetic field distribution, so it is necessary to determine a set of optimal scaling parameters to make the scaled model accurately describe the actual electromagnetic field characteristics. The invention uses genetic algorithm for iterative optimization to obtain the optimal scaling parameter combination that best matches the actual converter transformer.

[0217] 3. Application of convolutional neural network in fault feature extraction. Under different fault conditions, the electromagnetic field and temperature distribution inside the converter transformer will change significantly, and these changes contain important information for fault diagnosis. The invention uses convolutional neural network to automatically extract these features from finite element simulation results, greatly improving the speed and accuracy of fault area identification.

[0218] 4. Application of unsupervised clustering in abnormal area identification. Under different fault conditions, the temperature distribution of local overheating areas in converter transformer is significantly different. The invention uses unsupervised learning clustering algorithm to automatically find abnormal point groups in these temperature distributions, accurately locating the fault area and providing basis for subsequent maintenance.

[0219] Among them, similarity theory is a theory and analysis method for studying the similarity of physical systems under different scales or conditions. It is based on the similarity of physical laws and dimensional analysis. Similarity theory provides an effective analysis tool for studying physical phenomena and has wide application in engineering technology. The core idea of similarity theory is that if two physical systems meet certain similarity criteria under certain conditions, there is a certain correspondence between their corresponding physical quantities, i.e., scaling relationship. This scaling relationship can be characterized by a set of dimensionless similarity numbers. By studying the values of these dimensionless similarity numbers, we can determine whether two physical systems are similar, and accordingly establish a scaled model to transfer the research object to more convenient conditions for research.

[0220] The basic concepts and principles of similarity theory mainly include the following aspects:

[0221] 1. Similarity principle

[0222] The principle of similarity is the foundation of similarity theory. It states that if two physical systems satisfy certain similarity criteria, their physical processes and phenomena are similar. The similarity criteria require that the relevant dimensionless similarity numbers of the two systems be equal.

[0223] 2. Dimensional analysis

[0224] Dimensional analysis is an important tool of similarity theory. By analyzing the dimensions of various physical quantities that affect a physical process, we can find their relationships and derive the corresponding dimensionless similarity numbers.

[0225] 3. Similarity criteria and dimensionless similarity numbers

[0226] Similarity criteria are the basic guidelines for establishing similar models. They require that the corresponding dimensionless similarity numbers of two similar systems be equal. Dimensionless similarity numbers are obtained through dimensional analysis and reflect the relationships between various physical quantities in the system.

[0227] 4. Scale model

[0228] A scale model is a similar model established by scaling down or up the actual physical system according to the similarity criteria and dimensionless similarity numbers. By studying the scale model, we can obtain the corresponding results of the actual system.

[0229] 5. Selection of similarity criteria

[0230] Selecting appropriate similarity criteria is the key to the application of similarity theory. Different physical processes require different similarity criteria, which requires a deep understanding and analysis of the physical process.

[0231] Similarity theory has a wide range of applications in engineering and technology, such as:

[0232] 1. In the field of fluid mechanics, such as studying fluid flow, heat and mass transfer, similarity theory is often used to establish scale models for experimental research.

[0233] 2. In the field of structural mechanics, such as in wind tunnel tests, similarity theory is used to establish scale models to study the response of actual structures under the action of air dynamics.

[0234] 3. In the field of electromagnetic fields, such as in the study of electromagnetic field distribution of large-scale power equipment, similarity theory can be used to establish scale models for analysis and optimization design.

[0235] 4. In the field of thermal engineering, such as in the study of the performance of heat exchangers, engines and other thermal devices, similarity theory is often used to establish scale models for experimental research and numerical simulation.

[0236] In the field of electromagnetic field, the similarity theory is mainly based on the dimensionless form of electromagnetic field equation, and the corresponding similarity criteria are derived by comparing the dimensionless coefficients, mainly including:

[0237] Current density similarity criterion: requires the current density scaling factor of two similar electromagnetic systems to be equal.

[0238] Electric field intensity similarity criterion: requires the electric field intensity scaling factor of two similar electromagnetic systems to be equal.

[0239] Magnetic field intensity similarity criterion: requires the magnetic field intensity scaling factor of two similar electromagnetic systems to be equal.

[0240] Electric displacement vector similarity criterion: requires the electric displacement vector scaling factor of two similar electromagnetic systems to be equal.

[0241] These dimensionless similarity numbers reflect the scaling relationship of current, electric field, magnetic field and other electromagnetic quantities at different scales.

[0242] The size scaling factor, current density scaling factor, electric displacement vector scaling factor, electric field intensity scaling factor and magnetic field intensity scaling factor mentioned in the scheme are derived from the related criteria in electromagnetic field similarity theory. By determining these scaling factors, a scaled model of the converter transformer can be established to study the electromagnetic field and thermal field distribution characteristics.

[0243] In summary, the core technical principles of the present application include: 1) using similarity theory to establish a scaled model to reduce the complexity of modeling and simulation; 2) using genetic algorithm to solve the optimal scaling parameters to accurately reflect the actual electromagnetic field characteristics; 3) using convolutional neural network to automatically extract fault features to improve the speed and accuracy of diagnosis; 4) applying unsupervised clustering to identify temperature abnormal areas and accurately locate local overheating fault points.

[0244] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

[0245] Two specific embodiments of the present application are provided below.

[0246] Embodiment 1:

[0247] A power company is responsible for managing a large converter substation, which uses multiple high-voltage and high-power converter transformers as main equipment. In order to ensure the stable operation of the converter transformer, the power company decides to use the complex working condition converter local overheating fault prone area positioning method proposed by the present application to carry out preventive diagnosis on the key equipment.

[0248] Firstly, the power company engineers established the following several groups of geometric scale models according to the specific parameters of the converter transformer:

[0249] -1:1 scale model, the size is completely consistent with the actual converter transformer;

[0250] -1:2 scale model, each geometric size is reduced to 1 / 2 of the original;

[0251] -1:5 scale model, each geometric size is reduced to 1 / 5 of the original.

[0252] Through similarity theory analysis, the following scale factors are determined:

[0253] Size scale factor k s = 0.5, 0.2

[0254] Current density scale factor k j = 2, 5

[0255] Potential displacement vector scale factor k d = 0.5, 0.2

[0256] Electric field strength scale factor k e = 2, 5

[0257] Magnetic field strength scale factor k h = 2, 5

[0258] Then, the genetic algorithm is used to optimize the solution of the selected scale factors, and the optimal scale factor combination is obtained to make the 1:2 and 1:5 scale models closest to the actual converter transformer:

[0259] The optimal scale factor combination of the 1:2 scale model is

[0260] The optimal scale factor combination of the 1:5 scale model is

[0261] Based on the above optimal scale factor combination, the engineers established the three-dimensional electromagnetic field model of the 1:2 and 1:5 scale models in the finite element analysis software. Through simulation calculation, the equivalent magnetic permeability and equivalent electrical conductivity data of the core are obtained, which lays a foundation for subsequent temperature field analysis:

[0262] The equivalent magnetic permeability μ of the 1:2 scale model r,eq = 3000, the equivalent electrical conductivity σ eq = 2 × 10 6 S / m

[0263] The equivalent magnetic permeability μ of the 1:5 scale model r,eq= 2800, equivalent conductivity σ eq = 1.8 x 10 6 S / m

[0264] Next, the engineers constructed corresponding electromagnetic field finite element models for the no-fault condition, multi-point grounding fault condition, and inter-sheet short circuit fault condition:

[0265] 1) No-fault condition model: Directly use the equivalent parameters of the scaled-down model for electromagnetic field modeling, without adding any faults.

[0266] 2) Multi-point grounding fault condition model: Insert several short-circuit copper sheets in the core based on the no-fault model to simulate the electromagnetic field behavior of the multi-point grounding fault. The size and position of the short-circuit copper sheets are as follows:

[0267] 1:2 scaled-down model: The short-circuit copper sheet size is 50 mm long, 30 mm wide, and 5 mm thick, with 3 short-circuit points.

[0268] 1:5 scaled-down model: The short-circuit copper sheet size is 20 mm long, 12 mm wide, and 2 mm thick, with 5 short-circuit points.

[0269] 3) Inter-sheet short circuit fault condition model: Set the short-circuit region and short-circuit fault point of the core in the no-fault model to simulate the electromagnetic field distribution of the inter-sheet short circuit fault:

[0270] 1:2 scaled-down model: The short-circuit region is 100 mm long and 80 mm wide, and the short-circuit fault point is located in the middle of the core.

[0271] 1:5 scaled-down model: The short-circuit region is 40 mm long and 32 mm wide, and the short-circuit fault point is located in the upper part of the core.

[0272] Based on the above three fault condition models, the engineers used convolutional neural networks to extract features and classify the finite element simulation results. Specifically, a CNN model containing 3 convolutional layers, 2 pooling layers, and 2 fully connected layers was used. The input is the electromagnetic field distribution image under the three conditions, and after training, the CNN model can accurately identify the core loss distribution characteristics under different fault conditions:

[0273] Core loss distribution P under no-fault condition loss,1 (x, y, z)

[0274] Core loss distribution P under multi-point grounding fault condition loss,2 (x, y, z)

[0275] Core loss distribution P under inter-sheet short circuit fault condition loss,3 (x, y, z)

[0276] The core loss distribution described above is taken as the thermal load, and the engineers build a finite element model of the temperature field under three working conditions. In the temperature field model, the convective heat transfer boundary condition is used to simulate the heat exchange between the converter transformer and the surrounding environment:

[0277] The convective heat transfer coefficient h = 20 W / m 2 ·K

[0278] The ambient temperature T ∞ = 40℃

[0279] Through thermal simulation calculation, the engineers obtain the core temperature distribution under three fault conditions:

[0280] 1) Temperature distribution T1(x,y,z) under no-fault condition

[0281] The highest temperature T 1,max = 70℃

[0282] The temperature distribution is shown in Table 1

[0283] 2) Temperature distribution T2(x,y,z) under multi-point grounding fault condition

[0284] The highest temperature T 2,max = 95℃

[0285] The temperature distribution is shown in Table 2

[0286] 3) Temperature distribution T3(x,y,z) under inter-sheet short circuit fault condition

[0287] The highest temperature T 3,max = 120℃

[0288] The temperature distribution is shown in Table 3

[0289] Table 1 Temperature distribution under no-fault condition

[0290]

[0291]

[0292] Table 2 Temperature distribution under multi-point grounding fault condition

[0293] Coordinates (x, y, z) T2 0，0，0 90 0，0，50 92 0，0，100 95 0，50，0 88 0，50，50 90 0，50，100 92 50，0，0 86 50，0，50 88 50，0，100 90 50，50，0 84 50，50，50 86 50，50，100 88

[0294] Table 3 Temperature distribution under inter-sheet short circuit fault condition

[0295] Coordinates (x, y, z) T3 0，0，0 115 0，0，50 118 0，0，100 120 0，50，0 112 0，50，50 115 0，50，100 117 50，0，0 110 50，0，50 112 50，0，100 115 50，50，0 107 50，50，50 110 50，50，100 112

[0296] Next, the engineers use the k-means clustering algorithm to analyze the temperature distribution under the above three working conditions, and quickly identify the local temperature abnormal area:

[0297] 1) Temperature distribution T1(x, y, z) under normal working condition is analyzed by clustering analysis, and no obvious temperature abnormal area is found.

[0298] 2) Temperature distribution T2(x, y, z) under multi-point grounding fault working condition is analyzed by clustering analysis, and three temperature abnormal areas in the middle of the core are found, corresponding to the positions of multiple short-circuit copper sheets.

[0299] 3) Temperature distribution T3(x, y, z) under inter-sheet short-circuit fault working condition is analyzed by clustering analysis, and one obvious temperature abnormal area is found in the upper part of the core, which is consistent with the position of the set short-circuit fault point.

[0300] Based on the above analysis results, the engineers locate the local overheating fault area combined with the optimal scaling factor group obtained in step S10, and correspond to the specific position of the actual converter transformer to obtain the following conclusions:

[0301] 1) In the 1:2 scaled model, the three temperature abnormal areas under the multi-point grounding fault working condition correspond to the three short-circuit fault areas in the middle of the core of the actual converter transformer.

[0302] 2) In the 1:5 scaled model, the one temperature abnormal area under the inter-sheet short-circuit fault working condition corresponds to the one short-circuit fault area in the upper part of the core of the actual converter transformer.

[0303] Through the application of the method of the present application, the power company engineers not only quickly and accurately locate the local overheating fault area of the converter transformer, but also provide an important reference for the optimal design and fault diagnosis of the equipment. For example, for the middle part of the core where the multi-point grounding fault is prone to occur, a more optimal insulation structure design can be considered; for the upper part of the core where the inter-sheet short-circuit fault is prone to occur, the corresponding part can be mechanically reinforced. At the same time, the diagnosis result also provides a reliable basis for subsequent maintenance and management, which is helpful for the development of preventive maintenance measures.

[0304] Example 2:

[0305] This embodiment is directed to a positioning analysis of the local overheating fault prone area of a 550kV three-phase three-winding converter transformer. The converter transformer adopts a laminated core structure, mainly composed of main columns, side columns and upper and lower yokes.

[0306] 1. Establishing a scaled model and determining an optimal scaling factor group (steps S10, S20)

[0307] According to the geometric size and structural parameters of the converter transformer, a plurality of groups of size scaling mathematical models are established. The scaling factors such as the size scaling factor ks, the current density scaling factor kj, the electric displacement vector scaling factor kd, the electric field intensity scaling factor ke, and the magnetic field intensity scaling factor kh are optimized by using a genetic algorithm to obtain an optimal scaling factor combination {ks*, kj*, kd*, ke*, kh*} which can best describe the electromagnetic field distribution of the actual converter transformer.

[0308] According to the optimal scaling factor combination, the electromagnetic field equation of the converter transformer is established, and scaling is performed to obtain the inductance scaling factor kL and the resistance scaling factor kR of the scaled model. The inductance scaling factor kL and the resistance scaling factor kR are added to the optimal scaling factor group to obtain {ks*, kj*, kd*, ke*, kh*, kL, kR}.

[0309] 2. Modeling and finite element simulation (steps S30, S40)

[0310] By using the above scaling parameters, a three-dimensional electromagnetic field model of the laminated core of the converter transformer is constructed in a finite element analysis software to determine the equivalent permeability and conductivity.

[0311] For three typical working conditions of no fault, multi-point grounding fault, and inter-lamination short circuit fault, corresponding electromagnetic field finite element models are established, and simulation is performed to obtain the magnetic flux density, current density, and eddy current loss distribution under each working condition, as shown in FIG. 2. Figures 2-6 In addition, FIG. 3 shows the temperature distribution comparison chart of the scaled model formed by the non-optimized and optimized scaling factor groups. Figure 7 The left side of FIG. 3 is the non-optimized scaling factor corresponding to the scaled model, and the generated temperature distribution chart is a non-fine temperature distribution. Figure 7 The right side of FIG. 3 is the scaled model of the optimized scaling factor, and the generated temperature distribution chart is a fine temperature distribution. Figure 7

[0312] Under the no fault condition, the maximum magnetic flux density of the core reaches 2.00T, which is located at the corner of the second and third layers of the core. Under the multi-point grounding fault condition, the maximum fault current density is 1.21×105A / m2 between the first and second levels of the core upper and lower yokes, and the corresponding eddy current loss density is also the maximum, which is 1.29×105W. Under the inter-lamination short circuit fault condition, the maximum eddy current density at the fault point is 2.76×105A / m2, which is much higher than that under the non-fault condition.

[0313] 3. Feature extraction and loss identification (step S50)

[0314] The convolutional neural network is used to extract features and classify the finite element simulation results such as the magnetic flux density, current density, and eddy current loss under the above three working conditions, and the core loss distribution characteristics under different fault working conditions are quickly identified. ​

[0315] 4. Temperature field simulation (steps S60, S70)

[0316] The core loss distribution obtained from step 3 is taken as the thermal load, and a temperature field finite element model is established for three working conditions, respectively, and a heat simulation calculation is performed by applying the convective heat transfer boundary condition.

[0317] Figures 8-9 The temperature distribution of the core under different working conditions is given. When there is no fault, the maximum temperature is 55.10°C; under the multi-point grounding fault, the overall temperature rises, and the maximum temperature of the main column, side column and upper yoke part is 57.85°C on average; the inter-limb short circuit fault only has the highest temperature at the local fault point, reaching 53.40°C, and the overall maximum temperature is basically not affected.

[0318] 5. Temperature anomaly area identification (step S80)

[0319] Using an unsupervised learning clustering algorithm, based on the core temperature distribution under the above three working conditions, the temperature anomaly area under the multi-point grounding fault and the inter-limb short circuit fault, i.e. the local overheating fault area, is identified.

[0320] As shown in Table 10, compared with the no-fault working condition, the maximum temperature of the core under the multi-point grounding fault is 58.20°C, with a temperature rise of 5.63%; and the maximum temperature under the inter-limb short circuit fault is 55.80°C, with a temperature rise of 1.27%. It can be judged that the multi-point grounding fault is more likely to cause overall overheating, and the inter-limb short circuit fault is more likely to cause local overheating.

[0321] 6. Local overheating prone area positioning (step S90)

[0322] According to the optimal scaling factor group, the temperature anomaly area identified above is mapped to the actual size of the converter transformer, and the local overheating prone area is located.

[0323] The analysis shows that for the 550kV converter transformer, the local overheating prone area of the multi-point grounding fault is located near the main column, side column and upper yoke part; and the local overheating prone area of the inter-limb short circuit fault is located inside the core, especially at the fault point.

[0324] As can be seen from the above examples, the local overheating fault prone area positioning method can effectively combine scaling modeling, electromagnetic field simulation, machine learning feature recognition and temperature field analysis techniques, realize the rapid and accurate positioning of the local overheating area of the converter transformer under different fault working conditions, and provide strong support for the fault diagnosis and prevention of the equipment.

Claims

1. A method for locating partial overheating fault prone areas of a complex operating condition converter transformer, characterized in that, The method comprises the following steps: S10, a plurality of groups of scaled-down models of the to-be-tested converter transformer are established, scaling rules are determined according to the similarity theory, and corresponding groups of to-be-selected scaling factors of the plurality of groups of scaled-down models are obtained; An optimal group of scaling factors is obtained through optimization; S20, an electromagnetic field equation of the to-be-tested converter transformer is established, the electromagnetic field equation is scaled by using the optimal group of scaling factors, inductance scaling factors and resistance scaling factors of the scaled-down models are obtained, and the inductance scaling factors and the resistance scaling factors are added to the optimal group of scaling factors; S30, according to the optimal group of scaling factors, a three-dimensional electromagnetic field model of the scaled laminated core is obtained by using a finite element analysis software to model the laminated core of the converter transformer, and the equivalent permeability and the equivalent conductivity of the core are determined; S40, a first model simulated under a fault-free condition, a second model simulated under a multi-point grounding fault condition and a third model simulated under an inter-lamination short-circuit fault condition are constructed, and finite element simulations are performed on the first, second and third models; S50, a convolutional neural network is used to extract features and classify the finite element simulation results of the first, second and third models, and core loss distributions under different fault conditions are identified; S60, the obtained core loss distributions are taken as thermal loads, temperature field finite element models corresponding to the first, second and third models are established, and preset convective heat exchange boundary conditions are applied to the temperature field models; S70, thermal simulation calculations are performed on the temperature field finite element models corresponding to the first, second and third models, and core temperature distributions under the three conditions are obtained; S80, an unsupervised learning clustering algorithm is used to identify temperature abnormal areas under different fault conditions according to the obtained core temperature distributions under the three conditions, and a local overheating fault area is located; S90, according to the optimal group of scaling factors, a local overheating prone area in the to-be-tested converter transformer is obtained based on the located local overheating fault area.

2. The method for locating the partial overheating fault prone area of the complex operating mode converter transformer according to claim 1, characterized in that, The first model simulated under the fault-free condition, the second model simulated under the multi-point grounding fault condition and the third model simulated under the inter-lamination short-circuit fault condition are constructed as follows: for the fault-free condition, the equivalent parameters obtained in the step S20 are used to perform non-fault electromagnetic field modeling on the scaled-down model, which is recorded as the first model; for the multi-point grounding fault condition, copper sheets are inserted in the non-fault modeling and are subjected to short-circuit processing, and an electromagnetic field model of the multi-point grounding fault is constructed, which is recorded as the second model; for the inter-lamination short-circuit fault condition, short-circuit areas and short-circuit fault points of the core are set in the non-fault modeling, and an electromagnetic field model of the inter-lamination short-circuit fault is constructed, which is recorded as the third model.

3. The method of claim 1, wherein the method is characterized by, The group of scaling factors comprises a size scaling factor, a current density scaling factor, an electric displacement vector scaling factor, an electric field intensity scaling factor and a magnetic field intensity scaling factor.

4. The method of claim 1, wherein the method is characterized by, The method for obtaining the optimal group of scaling factors through optimization is a genetic algorithm.

5. The method of claim 1, wherein the method is characterized by, The unsupervised learning clustering algorithm used is a k-means clustering.

6. The method of claim 1, wherein, The steps of adopting the convolutional neural network to extract features and classify the finite element simulation results of the first, second and third models are as follows: a CNN model comprising multiple convolutional layers, pooling layers and fully connected layers is constructed, input is the electromagnetic field distribution images under the three working conditions of the no-fault working condition, the multi-point grounding fault working condition and the inter-sheet short-circuit fault working condition, multi-scale features are extracted through convolution and pooling operations, then the fully connected layers are used to classify the features, and the core loss distribution under different fault working conditions is output.

7. The method of claim 4, wherein the method is characterized by, The initialization population of the genetic algorithm is a plurality of scale factor groups; the fitness function is used to calculate the deviation between the scale models corresponding to the scale factor groups, including magnetic field intensity deviation and electric field intensity deviation.

8. The method according to claim 7, wherein, The formula of the fitness function is specifically: f(X) =‖B sim (X)‖ -‖B sim (X)‖ real ‖ 2 +λ‖E sim (X)‖ -‖E real (X)‖ 2 ; wherein, B sim (X) and E sim (X) respectively represent the magnetic field intensity and the electric field intensity obtained by finite element simulation under the scale factor combination X; B real and E real respectively represent the magnetic field intensity and the electric field intensity measured by the actual converter transformer; and λ is an adjustment parameter, used to balance the weight between the magnetic field and the electric field.

9. The method of claim 7, wherein the method is characterized by, The crossover operation in the genetic algorithm is realized in an arithmetic crossover manner, specifically, two parent individuals in the population are randomly selected to realize crossover by using a crossover coefficient to generate two new offspring individuals.

10. The method of claim 7, wherein the method is characterized by, The termination condition in the genetic algorithm is to terminate when the maximum number of iterations is reached, wherein the maximum number of iterations = 10 x the number of scale factor groups x the number of elements in each scale factor group.

Images