Fault-excited ferromagnetic resonance overvoltage characteristic determination method

By combining theoretical models and machine learning technology, a joint model that can describe the ferromagnetic resonant overvoltage characteristics under different fault conditions was established, which solved the problem of fusion of theoretical models and experimental data in the prior art, and improved the prediction of complex fault conditions Capability and model applicability.

CN119064844AActive Publication Date: 2024-12-03ELECTRIC POWER RESEARCH INSTITUTE OF STATE GRID NINGXIA ELECTRIC POWER COMPANY +8

Patent Information

Application Number
CN202411150511.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-21
Publication Date
2024-12-03
Estimated Expiration
2044-08-21

AI Technical Summary

Technical Problem

The existing methods are difficult to achieve the fusion of theoretical models and experimental data, resulting in insufficient ability to describe the ferromagnetic resonant overvoltage characteristics in complex fault conditions.

Method used

Using a method combining theoretical model with machine learning model, the electrical data is stimulated by collecting multiple faults of electromagnetic voltage transformers, ferromagnetic resonance equations and correction models are established, and the parameter correction is used to form a system of can describe different things. Joint model of ferromagnetic resonant overvoltage characteristics in fault conditions.

Benefits of technology

This method can more accurately predict the overvoltage characteristics under fault excitation, improve the applicability and robustness of the model, effectively describe the ferromagnetic resonant overvoltage characteristics under the superposition of single faults and multiple faults, and enhance the power system Safety and reliability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119064844A_ABST
    Figure CN119064844A_ABST
Patent Text Reader

Abstract

The invention provides a fault-excited ferromagnetic resonance overvoltage characteristic determination method, which belongs to the technical field of voltage transformers and comprises the following steps: acquiring electrical data of a voltage transformer with a single fault, and establishing a ferromagnetic resonance equation set and a correction model. An initial model is trained based on the data set. Secondly, the method collects electrical data of the voltage transformer with various faults, establishes a second training data set, and uses the data set to carry out fine tuning on the initial model to obtain a Lora model capable of adapting to various faults. Finally, the initial model and the Lora model serve as a combined model, single or multiple fault vectors are input, and an equation set describing ferromagnetic resonance overvoltage characteristics excited by faults is output; according to the method, the machine learning technology is fused, the overvoltage characteristics under different fault conditions can be described more accurately and comprehensively, and the problem that the description capability of the overvoltage characteristics under the complex fault condition is insufficient in an existing method is solved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of voltage transformers, and in particular relates to a method for determining ferromagnetic resonance overvoltage characteristics caused by a fault. Background Art

[0002] As a common voltage detection device in power systems, electromagnetic voltage transformers play an important role in power grid protection, control, and measurement. However, in actual operation, electromagnetic voltage transformers are susceptible to various faults, such as winding short circuit, core fault, secondary side grounding, etc. These faults can cause ferromagnetic resonance and cause serious overvoltage faults. Overvoltage faults can not only damage the voltage transformer itself, but also endanger the safe and reliable operation of the entire power system. Therefore, accurately analyzing and predicting the ferromagnetic resonance overvoltage characteristics under fault excitation is of great significance for taking timely protection measures and reducing the harm of faults.

[0003] At present, the research on ferromagnetic resonance overvoltage characteristics mainly focuses on two aspects: one is the analysis method based on theoretical models, and the other is the experience summary based on experimental measurements.

[0004] Theoretical analysis methods usually establish mathematical models to describe the ferromagnetic resonance process, such as voltage equation, current equation, flux equation and impedance equation, and try to use these equations to characterize the overvoltage characteristics caused by ferromagnetic resonance. This method has good physical significance, but due to the nonlinearity and complexity of the ferromagnetic resonance process, it is difficult to fully reflect the actual situation by relying solely on theoretical models, and there is a certain prediction error.

[0005] The test measurement method is to artificially create faults, conduct test measurements on actual electromagnetic voltage transformers, collect data such as voltage, current, and frequency under fault excitation, and summarize empirical rules. This method can obtain fault information more accurately, but it cannot be extended to various fault conditions, and the measurement data lacks consistency.

[0006] In other words, it is difficult for existing methods to achieve the integration of theoretical models and experimental data, and there is a technical problem of insufficient ability to describe overvoltage characteristics under complex fault conditions. Summary of the invention

[0007] In view of this, the present invention provides a method for determining the ferromagnetic resonance overvoltage characteristics induced by a fault, which can solve the technical problems that the existing methods are difficult to achieve the integration of theoretical models and experimental data and have insufficient ability to describe the overvoltage characteristics under complex fault conditions.

[0008] The present invention is achieved in that:

[0009] The present invention provides a method for determining a ferromagnetic resonance overvoltage characteristic caused by a fault, comprising the following steps:

[0010] S10, collecting electrical data of ferromagnetic resonance excited by various faults of an electromagnetic voltage transformer with a known single fault, including voltage, current, frequency, power factor, temperature, overvoltage, harmonic content, and core saturation, and recording the data as a first data set;

[0011] S20, establishing a ferromagnetic resonance equation group taking into account the electrical data, including a voltage equation, a current equation, a magnetic flux equation, and an impedance equation;

[0012] S30, establishing a combined model of ferromagnetic resonance overvoltage characteristics, including the ferromagnetic resonance equation group and a correction model, wherein the correction model is used to correct the parameters of the ferromagnetic resonance equation group, and the correction model is a neural network with a multi-input and multi-output structure, including a voltage equation parameter correction subnetwork, a current equation parameter correction subnetwork, a flux equation parameter correction subnetwork, an impedance equation parameter correction subnetwork, and a summary correction subnetwork;

[0013] S40, based on the first data set, establishing a first training data set, wherein the training input is voltage, current, frequency, power factor, temperature, harmonic content, and core saturation, and the training output is overvoltage amplitude, duration, frequency characteristics, and correction values ​​of parameters of each equation; using the first training data set to perform fitting training on the ferromagnetic resonance overvoltage characteristic combination model, and obtain the ferromagnetic resonance overvoltage characteristic combination model after fitting training, which is recorded as the first model;

[0014] S50, collecting electrical data of ferromagnetic resonance excited by multiple different faults of an electromagnetic voltage transformer known to have at least one fault, and recording the data as a second data set;

[0015] S60, using the second data set to establish a second training data set, wherein the training input is voltage, current, frequency, power factor, temperature, harmonic content, core saturation and fault type, and the training output is overvoltage amplitude, duration, frequency characteristics and correction values ​​of each equation parameter; using the second training data set to fine-tune the first model to obtain a Lora model;

[0016] S70, using the combined model and the Lora model as a joint model, inputting a fault vector of a single fault or a plurality of faults superimposed, and obtaining a group of ferromagnetic resonance equations for describing ferromagnetic resonance overvoltage characteristics excited by the fault.

[0017] The fault vector is specifically represented as a vector including multiple binary bits, each bit representing a possible fault type, and is used to describe a single or multiple superimposed fault conditions.

[0018] Among them, the voltage equation parameter correction subnetwork has a training input of voltage data and basic circuit parameters, and a training output of voltage equation parameter correction values, which is used to correct the voltage equation parameters according to actual voltage data, and the structure is a three-layer feedforward neural network.

[0019] Furthermore, the current equation parameter correction subnetwork has training inputs of current data and basic circuit parameters, and training outputs of current equation parameter correction values, which are used to correct current equation parameters according to actual current data, and the structure is a three-layer feedforward neural network.

[0020] Furthermore, the flux equation parameter correction subnetwork has a training input of flux data and core saturation, and a training output of flux equation parameter correction values, which is used to correct flux equation parameters according to actual flux data, and has a structure of a three-layer feedforward neural network.

[0021] Furthermore, the impedance equation parameter correction subnetwork has a training input of impedance data and temperature and a training output of impedance equation parameter correction values, which is used to correct the impedance equation parameters according to actual impedance data, and has a structure of a three-layer feedforward neural network.

[0022] Furthermore, the training input of the summary correction subnetwork is the correction value output by each subnetwork of the correction model excluding the summary correction subnetwork, and the training output is a comprehensive correction parameter for integrating the correction values ​​provided by each subnetwork. The structure is a two-layer feedforward neural network.

[0023] Furthermore, the Lora model includes a fault basis identification subnetwork and a correction model of a ferromagnetic resonance equation group of superimposed fault-induced ferromagnetic resonance.

[0024] Furthermore, the fine-tuning refers to adjusting the parameters of the hidden layer and output layer of each sub-network of the correction model in the first model, so as to adjust the weights and biases of these layers through the gradient descent method to adapt to various types of fault superposition fault situations.

[0025] Furthermore, the fault basis identification subnetwork has electrical data and fault type as training input, and fault characteristics as training output, which is used to identify and classify different types of faults, and its structure is a four-layer feedforward neural network; the correction model of the ferromagnetic resonance equation group of superimposed fault-excited ferromagnetic resonance has characteristics and fault characteristics of each subnetwork output as training input, and correction parameters as training output, which is used to correct the ferromagnetic resonance equation group according to the fault situation, and its structure is a five-layer feedforward neural network.

[0026] Furthermore, the step of obtaining the correction value of the equation parameter specifically includes:

[0027] S41, parameter input: select multiple groups of different initial parameter values ​​as input, these parameters include but are not limited to basic circuit parameters such as voltage, current, frequency, power factor, and environmental factors such as temperature and core saturation;

[0028] S42, solving the equation group: Substituting the parameter values ​​selected in S41 into the ferromagnetic resonance equation group, solving the equation group to obtain the output results, including the overvoltage amplitude, duration, frequency characteristics, etc.;

[0029] S43, actual data collection: according to the parameter conditions input in S1, the corresponding fault conditions are simulated on the actual electromagnetic voltage transformer, and multiple sets of electrical data are collected, including the actual measured overvoltage amplitude, duration, frequency characteristics, etc.;

[0030] S44, error calculation: compare the theoretical output results of the equation group in S42 with the actual data collected in S43, and calculate the errors of various indicators;

[0031] S45. Determine parameter correction values: Based on the error calculated in S44, use an optimization algorithm to iteratively adjust the parameters of the equation group until the error is reduced to an acceptable range, and use the difference between the final determined parameters and the initial parameters as the equation parameter correction value.

[0032] The optimization algorithm here uses the gradient descent method. This process compares the theoretical calculation results with the actual measurement data, continuously adjusts the equation parameters, and finally obtains the correction parameters that can accurately describe the ferromagnetic resonance overvoltage characteristics of fault excitation. This method takes into account both the theoretical model and the actual data, which can effectively improve the accuracy and applicability of the model.

[0033] Compared with the prior art, the method for determining the overvoltage characteristics of ferromagnetic resonance under fault excitation provided by the present invention has the following beneficial effects: 1. Combining the theoretical model with the machine learning model makes up for the defect that a single model cannot fully describe the complex ferromagnetic resonance process. The theoretical model provides the necessary physical basis for machine learning, while the machine learning model corrects and improves the theoretical model with measured data, so that the overall model can more accurately predict the overvoltage characteristics under fault excitation.

[0034] 2. For single fault and multiple fault superposition, the initial model and LORA fine-tuning model are established respectively, which can fully and accurately describe the ferromagnetic resonance overvoltage characteristics under different fault conditions. Moreover, the LORA model will not directly change the weight parameters of the initial model, so that the single fault information judgment ability of the initial model can be better retained, and the judgment ability of a single fault will not be weakened due to the influence of multiple fault superposition. Compared with the existing single fault analysis method, the method of the present invention has stronger adaptability and robustness.

[0035] 3. In the parameter correction stage, a systematic parameter selection strategy was adopted, including determining the parameter range, Latin hypercube sampling, considering extreme cases and typical working conditions, etc., which made full use of the measured data and improved the accuracy and applicability of the model parameters.

[0036] 4. The entire method is process-oriented and highly automated, and can be easily applied to operation monitoring and fault diagnosis of electromagnetic voltage transformers, providing strong support for improving the safety and reliability of power systems.

[0037] In general, the method for determining the fault-induced ferromagnetic resonance overvoltage characteristics of the present invention fully utilizes the advantages of theoretical analysis and experimental measurement, integrates machine learning technology, and can more accurately and comprehensively describe the overvoltage characteristics under different fault conditions. It solves the technical problem that the existing methods are difficult to achieve the integration of theoretical models and experimental data, and have insufficient ability to describe the overvoltage characteristics under complex fault conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A flow chart of the method provided by the present invention;

[0039] Figure 2 A flowchart of the steps for obtaining corrected values ​​of equation parameters;

[0040] Figure 3 It is the time domain waveform of overvoltage;

[0041] Figure 4 It is the time domain waveform diagram of the current;

[0042] Figure 5 It is the time domain waveform diagram of magnetic flux;

[0043] Figure 6 This is the time domain waveform of impedance. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0045] like Figure 1 FIG. 1 is a flow chart of a method for determining a ferromagnetic resonance overvoltage characteristic caused by a fault provided by the present invention. The method comprises the following steps:

[0046] S10, collecting electrical data of ferromagnetic resonance excited by various faults of an electromagnetic voltage transformer with a known single fault, including voltage, current, frequency, power factor, temperature, overvoltage, harmonic content, and core saturation, and recording the data as a first data set;

[0047] S20, establishing a ferromagnetic resonance equation group considering electrical data, including a voltage equation, a current equation, a flux equation, and an impedance equation;

[0048] S30, establishing a combined model of ferromagnetic resonance overvoltage characteristics, including a ferromagnetic resonance equation group and a correction model, wherein the correction model is used to correct the parameters of the ferromagnetic resonance equation group, and the correction model is a neural network with a multi-input and multi-output structure, including a voltage equation parameter correction subnetwork, a current equation parameter correction subnetwork, a flux equation parameter correction subnetwork, an impedance equation parameter correction subnetwork, and a summary correction subnetwork;

[0049] S40, based on the first data set, establishing a first training data set, wherein the training input is voltage, current, frequency, power factor, temperature, harmonic content, and core saturation, and the training output is overvoltage amplitude, duration, frequency characteristics, and correction values ​​of parameters of each equation; using the first training data set to perform fitting training on the ferromagnetic resonance overvoltage characteristic combination model, and obtain the ferromagnetic resonance overvoltage characteristic combination model after fitting training, which is recorded as the first model;

[0050] S50, collecting electrical data of ferromagnetic resonance excited by multiple different faults of an electromagnetic voltage transformer known to have at least one fault, and recording the data as a second data set;

[0051] S60, using the second data set to establish a second training data set, wherein the training input is voltage, current, frequency, power factor, temperature, harmonic content, core saturation and fault type, and the training output is overvoltage amplitude, duration, frequency characteristics and correction values ​​of each equation parameter; using the second training data set to fine-tune the first model to obtain a Lora model;

[0052] S70, using the combined model and the Lora model as a joint model, inputting a fault vector of a single fault or a plurality of faults superimposed, and obtaining a ferromagnetic resonance equation group for describing ferromagnetic resonance overvoltage characteristics excited by the fault.

[0053] like Figure 2 As shown, the step of obtaining the equation parameter correction value specifically includes:

[0054] S41, parameter input: select multiple groups of different initial parameter values ​​as input, these parameters include but are not limited to basic circuit parameters such as voltage, current, frequency, power factor, and environmental factors such as temperature and core saturation;

[0055] S42, solving the equation group: Substituting the parameter values ​​selected in S41 into the ferromagnetic resonance equation group, solving the equation group to obtain the output results, including the overvoltage amplitude, duration, frequency characteristics, etc.;

[0056] S43, actual data collection: according to the parameter conditions input in S1, the corresponding fault conditions are simulated on the actual electromagnetic voltage transformer, and multiple sets of electrical data are collected, including the actual measured overvoltage amplitude, duration, frequency characteristics, etc.;

[0057] S44, error calculation: compare the theoretical output results of the equation group in S42 with the actual data collected in S43, and calculate the errors of various indicators;

[0058] S45, determining parameter correction values: Based on the error calculated in S44, use an optimization algorithm (such as gradient descent method) to iteratively adjust the parameters of the equation group until the error is reduced to an acceptable range. The difference between the final determined parameters and the initial parameters is used as the equation parameter correction value.

[0059] The specific implementation methods of the above steps are described in detail below:

[0060] Step S10: Collecting electrical data of ferromagnetic resonance excited by multiple different faults of an electromagnetic voltage transformer with a known single fault. The specific implementation of this step includes the following sub-steps:

[0061] Step 10.1 Determine the rated parameters of the electromagnetic voltage transformer to be tested. Including the rated voltage U n , rated frequency f n 、Rated phase angle φ n wait.

[0062] Step 10.2 Design a device to simulate faults. For example, you can use a trigger circuit, a short-circuit switch, etc. to simulate various fault conditions on the primary or secondary side of the voltage transformer, such as short-circuit faults, ground faults, and open circuit faults.

[0063] Step 10.3 collects various electrical quantity data of the electromagnetic voltage transformer during the simulated fault, including voltage v, current i, frequency f, power factor cosφ, temperature T, overvoltage peak value U p , harmonic content THD, core saturation S, etc. In order to obtain more comprehensive data, a high-speed data acquisition instrument can be used with a sampling frequency of no less than 1kHz.

[0064] Step 10.4 organizes the collected electrical data into a first data set D1 for use in subsequent steps. The data set should contain detailed electrical characteristic data under various single fault conditions.

[0065] Through this step, the detailed electrical parameter data of the electromagnetic voltage transformer under a known single fault condition was obtained, laying the foundation for the subsequent establishment of the ferromagnetic resonance equation group and training of the neural network model.

[0066] Step S20: Establish a ferromagnetic resonance equation group taking into account the electrical data. The specific implementation of this step includes the following sub-steps:

[0067] Step 20.1 According to the working principle of electromagnetic voltage transformer, establish a set of equations describing the relationship between voltage v, current i, magnetic flux ψ and impedance Z. These equations should include voltage equation, current equation, magnetic flux equation and impedance equation, covering the key physical processes of ferromagnetic resonance phenomenon.

[0068] Step 20.2 introduces the influence of environmental factors such as temperature T, frequency f, and core saturation S into these basic equations to more accurately describe the ferromagnetic resonance characteristics under actual working conditions.

[0069] Step 20.3 For the parameters in the equation, such as inductance L, capacitance C, resistance R, etc., use measurement or fitting methods to obtain their values. For some parameters that are difficult to measure directly, such as the nonlinear coefficient k v , k i , k h etc. can be determined by fitting experimental data.

[0070] Step 20.4 organizes the established voltage equation, current equation, flux equation and impedance equation to form a complete set of equations describing the ferromagnetic resonance process. This set of equations will serve as the theoretical basis for the subsequent ferromagnetic resonance overvoltage characteristic prediction model.

[0071] The following is a detailed description of each equation:

[0072] 1. Voltage equation:

[0073]

[0074] Where, v is the secondary voltage of the voltage transformer (unit V); t is the time (unit s); C is the equivalent capacitance (unit F), obtained by measurement; i is the current (unit A); R s is the equivalent series resistance (unit: Ω), obtained by measurement; L s is the equivalent series inductance (unit: H), obtained by measurement; k v is the voltage nonlinear coefficient, obtained by fitting the experimental data; V s is the fundamental voltage amplitude (unit V), obtained by Fourier analysis; ω is the angular frequency (unit rad / s), ω=2πf, f is the system frequency; φ is the fundamental phase angle (unit rad), obtained by Fourier analysis; V n is the nth harmonic voltage amplitude (unit V), obtained by Fourier analysis; φ n is the nth harmonic phase angle (in rad), obtained by Fourier analysis.

[0075] 2. Current equation:

[0076] i=i m +i e +i h ;

[0077]

[0078] In the formula, i m is the magnetizing current (unit A); i e is the eddy current loss current (unit A); i h is the hysteresis loss current (unit A); L m is the main magnetizing inductance (unit: H), obtained by measurement; k i is the current nonlinear coefficient, obtained by fitting the experimental data; ψ is the magnetic flux (unit: Wb); C e is the equivalent eddy current capacitance (unit F), obtained by measurement; G e is the equivalent eddy current conductance (unit: S), obtained by measurement; I n is the nth harmonic current amplitude (unit A), obtained by Fourier analysis; θ n is the nth harmonic current phase angle (unit: rad), obtained by Fourier analysis; k h is the hysteresis loss coefficient, which is obtained by fitting the experimental data.

[0079] 3. Magnetic flux equation:

[0080] ψ=∫vdt;

[0081]

[0082] Where, B is the magnetic induction intensity (unit T); H is the magnetic field intensity (unit A / m); N is the number of winding turns; A e is the effective cross-sectional area of ​​the core (unit: m 2 ), obtained by measurement; l e is the effective magnetic path length of the core (unit: m), obtained by measurement; μ0 is the vacuum permeability, 4π×10 -7 H / m; μ r is the relative magnetic permeability, dimensionless; B r is the remanence (unit T), obtained through experimental measurement; H c is the coercive force (unit: A / m), obtained through experimental measurement; μ i is the initial magnetic permeability, dimensionless, obtained through experimental measurement; k T is the temperature coefficient, obtained by fitting experimental data; T is the current temperature (unit K); T0 is the reference temperature (unit K), usually room temperature 293.15K; k fis the frequency coefficient, which is obtained by fitting the experimental data; f is the current frequency (unit Hz); f0 is the reference frequency (unit Hz), usually the rated frequency 50Hz or 60Hz.

[0083] 4. Impedance equation:

[0084] Z = R + jX;

[0085]

[0086] Where Z is the total impedance (unit Ω); R is the equivalent resistance (unit Ω); X is the equivalent reactance (unit Ω); R0 is the DC resistance (unit Ω), obtained by measurement; α is the temperature coefficient (unit 1 / K), obtained by fitting experimental data; k R is the influence coefficient of frequency on resistance, obtained by fitting experimental data; X0 is the reference reactance (unit Ω), obtained by measurement; L0 is the reference inductance (unit H), obtained by measurement; k X is the influence coefficient of magnetic saturation on reactance, obtained by fitting experimental data; B s is the saturation magnetic induction intensity (unit T), obtained by experimental measurement; S is the core saturation, dimensionless; Z eq is the equivalent impedance (unit: Ω); Z p is the parallel impedance (unit Ω); R p is the parallel resistance (unit Ω), obtained by measuring

[0087] C p is the parallel capacitance (unit: F), obtained by measurement.

[0088] These equations form a complex set of ferromagnetic resonance equations, covering the relationship between voltage, current, flux and impedance. The equations contain multiple parameters, taking into account the influence of factors such as temperature, frequency, magnetic saturation, etc., and introduce nonlinear terms and high-order derivative terms to more accurately describe the ferromagnetic resonance phenomenon. The solution of these equations requires the combination of numerical methods, such as the Runge-Kutta method or the finite element method, and iterative calculations may be required to deal with nonlinear terms.

[0089] Through this step, a set of ferromagnetic resonance equations covering the relationship between voltage, current, magnetic flux and impedance was established, providing a theoretical basis for subsequent model construction.

[0090] Step S30: Establishing a ferromagnetic resonance overvoltage characteristic combined model. The specific implementation of this step includes the following sub-steps:

[0091] Step 30.1 constructs a combined model of ferromagnetic resonance overvoltage characteristics. The model includes two parts: ferromagnetic resonance equations and a correction model.

[0092] Step 30.2: The ferromagnetic resonance equation group is the complete equation group established in step S20, which is used to describe the relationship between various physical quantities in the ferromagnetic resonance process.

[0093] Step 30.3 The correction model is a multi-input and multi-output neural network structure, which includes 5 sub-networks: voltage equation parameter correction sub-network, current equation parameter correction sub-network, flux equation parameter correction sub-network, impedance equation parameter correction sub-network and summary correction sub-network.

[0094] Step 30.4 The input of the voltage equation parameter correction subnetwork is voltage data and basic circuit parameters, and the output is the correction value of the voltage equation parameter, which is used to correct the voltage equation. This subnetwork adopts a 3-layer feedforward neural network structure.

[0095] Step 30.5 The input of the current equation parameter correction subnetwork is the current data and basic circuit parameters, and the output is the correction value of the current equation parameter, which is used to correct the current equation. This subnetwork also adopts a 3-layer feedforward neural network structure.

[0096] Step 30.6 The input of the flux equation parameter correction subnetwork is the flux data and core saturation, and the output is the correction value of the flux equation parameter, which is used to correct the flux equation. This subnetwork also adopts a 3-layer feedforward neural network structure.

[0097] Step 30.7 The input of the impedance equation parameter correction subnetwork is impedance data and temperature, and the output is the correction value of the impedance equation parameter, which is used to correct the impedance equation. This subnetwork also adopts a 3-layer feedforward neural network structure.

[0098] Step 30.8 summarizes the input of the correction sub-network as the correction values ​​output by the other four sub-networks, and outputs a comprehensive correction parameter, which is used to integrate the correction information provided by each sub-network. This sub-network adopts a 2-layer feedforward neural network structure.

[0099] Step 30.9: Through the combined model, the theoretical equation group established in step S20 can be integrated with the actual measurement data to improve the model's ability to describe the ferromagnetic resonance overvoltage characteristics caused by faults.

[0100] In general, this step establishes a joint model consisting of a group of ferromagnetic resonance equations and a parameter-corrected neural network, which utilizes the advantages of theoretical models and measured data to improve the accuracy of describing the fault-induced ferromagnetic resonance overvoltage characteristics.

[0101] Step S40: Establish a first training data set based on the first data set, and perform fitting training on the ferromagnetic resonance overvoltage characteristic combination model. The specific implementation of this step includes the following sub-steps:

[0102] Step 40.1 selects training input features from the first data set D1. The training inputs include voltage v, current i, frequency f, power factor cosφ, temperature T, harmonic content THD, core saturation S, etc.

[0103] Step 40.2 Training output includes overvoltage peak value U p , overvoltage duration t p , overvoltage frequency characteristics f p And the corrected values ​​of the parameters of each equation.

[0104] Step 40.3 uses the above training input and output samples to perform fitting training on the ferromagnetic resonance overvoltage characteristic combination model established in step S30. The training algorithm can use a gradient descent method or other optimization algorithms, and the goal is to minimize the error between the training samples and the model output.

[0105] After the training in step 40.4 is completed, a fitted and trained ferromagnetic resonance overvoltage characteristic combination model is obtained, which is recorded as M1. This model can accurately describe the ferromagnetic resonance overvoltage characteristics under a known single fault condition.

[0106] Through this step, an initial ferromagnetic resonance overvoltage characteristic prediction model M1 was obtained based on the training of the first data set D1, laying the foundation for subsequent processing of more complex fault conditions.

[0107] Step S50: Collect electrical data of ferromagnetic resonance excited by multiple different faults of an electromagnetic voltage transformer known to have at least one fault, and record it as a second data set. This step is similar to step S10 and will not be described in detail.

[0108] Step S60: Using the second data set, establish a second training data set, and fine-tune the ferromagnetic resonance overvoltage characteristic combination model established in step S30. The specific implementation of this step includes the following sub-steps:

[0109] Step 60.1 selects training input features from the second data set D2. The training inputs include voltage v, current i, frequency f, power factor cosφ, temperature T, harmonic content THD, core saturation S, and fault type F.

[0110] Step 60.2 The training output still includes the overvoltage peak value U p , overvoltage duration t p , overvoltage frequency characteristics f p And the corrected values ​​of the parameters of each equation.

[0111] Step 60.3 uses the LORA (Low-Rank Adaptation) fine-tuning algorithm to adjust the parameters of the initial model M1 trained in step S40. Specifically, the LORA fine-tuning optimizes the parameters of the hidden layer and output layer of the modified model in M1 by the gradient descent method to adapt to the situation where multiple types of faults are superimposed.

[0112] After LORA fine-tuning in step 60.4, a refined model M2 for the composite fault situation is obtained, which is called the LORA model. This model can better describe the impact of multiple fault superposition on the ferromagnetic resonance overvoltage characteristics.

[0113] Through this step, the initial model M1 was LORA-fine-tuned using the second data set D2, and a more robust ferromagnetic resonance overvoltage characteristic prediction model M2 was obtained, which can cope with complex fault conditions.

[0114] Step S70: Taking the combined model M1 and the LORA model M2 as a joint model, inputting a fault vector of a single fault or a plurality of faults superimposed, and obtaining a ferromagnetic resonance equation group for describing the ferromagnetic resonance overvoltage characteristics excited by the fault. The specific implementation of this step includes the following sub-steps:

[0115] Step 70.1 Fault Vector It is a vector containing multiple binary bits, each bit represents a possible fault type and is used to describe single or multiple superimposed fault conditions.

[0116] Step 70.2 sets the fault vector As input, the LORA model M2 obtained by training in step S60 can obtain the corrected ferromagnetic resonance equation group parameters.

[0117] Step 70.3 combines the ferromagnetic resonance equations established in step S30 with the output of the LORA model M2 to obtain a complete ferromagnetic resonance equation group that can accurately describe the fault-induced ferromagnetic resonance overvoltage characteristics.

[0118] Step 70.4 LORA model M2 includes two sub-networks: a fault basis identification sub-network and a ferromagnetic resonance equation group correction model of superimposed fault-induced ferromagnetic resonance.

[0119] Step 70.5 The input of the fault basis identification sub-network is electrical data and fault type, and the output is fault characteristics, which are used to identify and classify different types of faults. This sub-network adopts a 4-layer feedforward neural network structure.

[0120] Step 70.6 is a correction model for the ferromagnetic resonance equations of superimposed fault-induced ferromagnetic resonance, the input of which is the characteristics of each sub-network output and the fault characteristics, and the output is a correction parameter, which is used to correct the ferromagnetic resonance equations according to the fault situation. The sub-network adopts a 5-layer feedforward neural network structure.

[0121] Through this step, the ferromagnetic resonance equation group established in step S30 is integrated with the LORA model M2 trained in step S60 to form a joint model that can describe the ferromagnetic resonance overvoltage characteristics under single fault or multiple fault superposition conditions.

[0122] The following is a specific implementation of steps S41-S45:

[0123] Step S41: parameter input. Multiple groups of different initial parameter values ​​are selected as input, and these parameters include but are not limited to basic circuit parameters such as voltage v, current i, frequency f, power factor cosφ, and environmental factors such as temperature T and core saturation S.

[0124] In order to achieve the optimal parameter correction effect, the specific steps for selecting initial parameters are as follows:

[0125] 1. Determine the parameter range:

[0126] According to the working principle and actual operation of electromagnetic voltage transformer, determine the reasonable range of each parameter, such as voltage 0.8U n 1.2U n , frequency 45Hz~65Hz, power factor 0.8 lagging 1.0~0.8 leading, temperature -40℃~80℃, core saturation 0.6~1.2.

[0127] 2. Using Latin hypercube sampling method:

[0128] The Latin hypercube sampling method is used to generate uniformly distributed sample points within a certain parameter range, ensuring that the samples in the multidimensional parameter space have good representativeness and uniformity.

[0129] 3. Consider extreme cases:

[0130] On the basis of sampling, some extreme parameter combinations are added, such as the maximum voltage and the maximum temperature, the minimum frequency and the maximum core saturation, etc., to test the performance of the model under boundary conditions.

[0131] 4. Includes typical working conditions:

[0132] Add some known typical operating condition parameter combinations, which often appear in actual operation or have special significance.

[0133] 5. Consider parameter correlation:

[0134] There may be correlations between some parameters, such as an increase in temperature may cause a change in core saturation.

[0135] 6. Dynamic Adjustment:

[0136] After initial training, analyze the performance of the model in different parameter intervals. For intervals with poor performance, increase the number of sample points in that interval and perform focused sampling.

[0137] 7. Cross Validation:

[0138] The selected parameter group is divided into training set and validation set, and the cross-validation method is used to evaluate the performance of the model under different parameter combinations to ensure that the selected parameter group has good generalization ability.

[0139] 8. Iterative optimization:

[0140] Based on the initial training results, optimization algorithms (such as Bayesian optimization) are used to guide subsequent parameter selection, and gradually find the parameter combination that best reflects the system characteristics.

[0141] Through the above steps, multiple sets of different initial parameter values ​​can be systematically selected as input to cover various possible working conditions and fault conditions, which helps to train a model with strong robustness and good generalization ability. This method can improve the accuracy of parameter correction and ensure that the model can maintain good performance in the face of various actual situations.

[0142] Step S42: Solving the equations. Substitute the parameter values ​​selected in step S41 into the ferromagnetic resonance equations established in step S20, and obtain the output results of the equations by numerical solution methods (such as the Runge-Kutta method), including the overvoltage peak value U p , overvoltage duration t p , overvoltage frequency characteristics f p wait.

[0143] Step S43: actual data collection. According to the parameter conditions input in step S41, the corresponding fault conditions are simulated on the actual electromagnetic voltage transformer, and multiple sets of electrical data are collected, including the actual measured overvoltage peak value U p , overvoltage duration t p , overvoltage frequency characteristics f p wait.

[0144] Step S44: Error calculation. Compare the theoretical output results of the equation group in step S42 with the actual data collected in step S43, and calculate various indicators (overvoltage peak value U p , overvoltage duration t p , overvoltage frequency characteristics f p ) error.

[0145] Step S45: Determine parameter correction value.

[0146] 1. Based on the error calculated in step S44, use an optimization algorithm (such as gradient descent method) to iteratively adjust the parameters of the equation group until the error is reduced to an acceptable range.

[0147] 2. The difference between the final determined parameters and the initial parameters is used as the equation parameter correction value.

[0148] 3. This process continuously adjusts the equation parameters by comparing the theoretical calculation results with the actual measurement data, and finally obtains the corrected parameters that can accurately describe the ferromagnetic resonance overvoltage characteristics of fault excitation.

[0149] 4. This method not only takes into account the theoretical model, but also combines it with actual data, which can effectively improve the accuracy and applicability of the model.

[0150] In general, steps S41-S45 achieve the goal of comparing theoretical calculation results with measured data, using optimization algorithms to continuously adjust the parameters of the equation group, and finally obtaining a set of correction parameters that can accurately describe the fault-induced ferromagnetic resonance overvoltage characteristics. This method fully utilizes the advantages of theoretical models and experimental data, and improves the accuracy and robustness of the model.

[0151] Specifically, the principle of the present invention is to establish a joint prediction framework combining theoretical models and machine learning models. The specific principles are as follows:

[0152] Firstly, according to the working principle of electromagnetic voltage transformer, a complete set of equations describing the relationship between voltage, current, magnetic flux and impedance is established, including voltage equation, current equation, magnetic flux equation and impedance equation. These equations take into account the influence of environmental factors such as temperature, frequency and core saturation, and can more accurately describe the physical laws in the ferromagnetic resonance process.

[0153] However, it is difficult to fully reflect the overvoltage characteristics under actual fault conditions by relying solely on these theoretical equations, and there is a certain prediction error. Therefore, the present invention introduces a correction model based on machine learning, which consists of five sub-networks: a voltage equation parameter correction sub-network, a current equation parameter correction sub-network, a flux equation parameter correction sub-network, an impedance equation parameter correction sub-network, and a summary correction sub-network.

[0154] These subnetworks can dynamically correct the parameters of the theoretical equations according to the measured voltage, current, magnetic flux, and impedance data, so that the prediction results of the overall model can better match the actual observed values. Specifically, each subnetwork adopts a feedforward neural network structure, with the input being the measured data of the corresponding physical quantity and the output being the parameter correction value of the physical quantity equation. The summary correction subnetwork integrates the correction results of each subnetwork and gives a comprehensive parameter correction.

[0155] Through the integration of the theoretical model and the machine learning model, the method of the present invention can give full play to the advantages of both. The theoretical model provides the physical basis and overall framework, while the machine learning model uses the measured data to continuously correct and optimize the model parameters, so that the overall model can better describe the complex fault-induced ferromagnetic resonance process.

[0156] In order to further improve the adaptability of the method, the present invention also uses LORA fine-tuning technology. On the basis of the initial model training, the second data set containing multiple fault types is used to fine-tune the parameters of the correction sub-network of the model. This can not only adapt to single fault situations, but also cope with complex situations where multiple faults are superimposed, thereby greatly improving the robustness of the model.

[0157] In addition, when selecting the initial parameter values, the present invention adopts a series of strategies, such as determining the parameter range, Latin hypercube sampling, considering extreme cases and typical working conditions, etc., making full use of the measured data and improving the accuracy of parameter correction. Through iterative optimization, a set of correction parameters that can accurately describe the fault-induced ferroresonance overvoltage characteristics is finally obtained.

[0158] In general, the method of the present invention fully integrates theoretical analysis and machine learning technology, uses measured data to continuously correct and improve the theoretical model, and finally constructs a joint model that can comprehensively and accurately predict the fault-induced ferromagnetic resonance overvoltage characteristics.

[0159] In order to better understand and implement the present invention, a specific embodiment 1 of the present invention is provided below. The steps of this embodiment 1 are specifically described as follows:

[0160] Step S10: Collecting electrical data of ferromagnetic resonance excited by various faults of an electromagnetic voltage transformer with a known single fault.

[0161] In this step, it is necessary to collect detailed electrical data of the voltage transformer under a known single fault condition to provide a basis for the subsequent theoretical model establishment and neural network training. Specifically, it includes the following aspects:

[0162] 1. Determine the rated parameters of the voltage transformer to be tested, including the rated voltage U n , rated frequency f n and rated phase angle φn wait.

[0163] 2. Design devices to simulate faults, such as trigger circuits, short-circuit switches, etc., to simulate various fault conditions on the primary or secondary side of the voltage transformer, such as short-circuit faults, ground faults, open circuit faults, etc.

[0164] 3. Collect various electrical quantity data of the voltage transformer during simulated fault, including voltage v, current i, frequency f, power factor cosφ, temperature T, overvoltage peak value U p , harmonic content THD and core saturation S, etc. In order to obtain more comprehensive data, a high-speed data acquisition instrument can be used with a sampling frequency of no less than 1kHz.

[0165] 4. Organize the collected electrical data into the first data set D1 for use in subsequent steps. This data set should contain detailed electrical characteristic data under various single fault conditions.

[0166] Step S20: Establishing a set of ferromagnetic resonance equations taking into account the electrical data.

[0167] In this step, it is necessary to establish a set of equations describing the relationship between voltage v, current i, magnetic flux ψ and impedance Z to characterize the key physical process of ferromagnetic resonance phenomenon. These equations include:

[0168] Voltage equation:

[0169]

[0170] In the formula, C is the equivalent capacitance, R s is the equivalent series resistance, L s is the equivalent series inductance, k v is the voltage nonlinear coefficient. V s is the fundamental voltage amplitude, ω is the angular frequency, and φ is the fundamental phase angle. n and φ n are the nth harmonic voltage amplitude and phase angle respectively.

[0171] Current equation:

[0172] i=i m +i e +i h

[0173]

[0174] In the formula, i m is the magnetizing current, i e is the eddy current loss current, i h is the hysteresis loss current. m Main magnetizing inductance, k iis the current nonlinear coefficient. ψ is the magnetic flux. C e is the equivalent eddy current capacitance, G e is the equivalent eddy current conductance. n and θ n are the amplitude and phase angle of the nth harmonic current, k h is the hysteresis loss coefficient.

[0175] Flux equation:

[0176] ψ=∫vdt

[0177]

[0178]

[0179] Where N is the number of winding turns, A is e is the effective cross-sectional area of ​​the core, l e is the effective magnetic path length of the core. μ0 is the vacuum magnetic permeability, μ r is the relative magnetic permeability, B r is the residual magnetism, H c is the coercive force, μ i is the initial magnetic permeability. k T is the temperature coefficient, k f is the frequency coefficient.

[0180] Impedance equation:

[0181] Z=R+jX

[0182]

[0183] In the formula, R is the equivalent resistance, X is the equivalent reactance. R0 is the DC resistance, α is the temperature coefficient, k R is the influence coefficient of frequency on resistance. X0 is the reference reactance, L0 is the reference inductance, k X is the influence coefficient of magnetic saturation on reactance, B s is the saturation magnetic induction intensity. S is the core saturation. Z eq is the equivalent impedance, Z p is the parallel impedance, R p is the parallel resistance, C p The parallel capacitor.

[0184] These equations cover the relationship between voltage, current, magnetic flux and impedance, and take into account the influence of factors such as temperature, frequency, and magnetic saturation. At the same time, nonlinear terms and high-order derivative terms are introduced to more accurately describe the ferromagnetic resonance phenomenon.

[0185] When establishing these equations, it is necessary to use measurement or fitting methods to obtain the values ​​of various parameters, such as capacitance C, resistance R s 、Inductance Ls , main magnetizing inductance L m , initial magnetic permeability μ i , Temperature coefficient k T , frequency coefficient k f , the influence coefficient of frequency on resistance k R , the influence coefficient k of magnetic saturation on reactance X For some parameters that are difficult to measure directly, such as the nonlinear coefficient k v , k i , k h , can be determined by fitting experimental data.

[0186] By establishing this complete set of ferromagnetic resonance equations, a theoretical foundation has been laid for the subsequent ferromagnetic resonance overvoltage characteristic prediction model. However, relying solely on these theoretical equations, it is difficult to fully reflect the overvoltage characteristics under actual fault conditions, and there are certain prediction errors. Therefore, it is necessary to combine these theoretical models with machine learning technology to improve the accuracy and robustness of the prediction.

[0187] Step S30: Establishing a ferromagnetic resonance overvoltage characteristic combination model.

[0188] In this step, a combined model of ferromagnetic resonance overvoltage characteristics consisting of a theoretical equation group and a correction model is constructed. The theoretical equation group is the complete equation group established in step S20, which is used to describe the relationship between various physical quantities in the ferromagnetic resonance process; and the correction model is a multi-input and multi-output structure based on a neural network, which is used to dynamically adjust the parameters of the theoretical equation group, so that the overall model can more accurately predict the overvoltage characteristics under fault excitation.

[0189] The revised model includes the following five sub-networks:

[0190] 1. Voltage equation parameter correction subnetwork:

[0191] The input is voltage v data and basic circuit parameters (C, R s , L s , k v wait)

[0192] The output is the correction value of the voltage equation parameter, which is used to adjust the voltage equation

[0193] This sub-network adopts a 3-layer feedforward neural network structure.

[0194] 2. Current equation parameter correction subnetwork:

[0195] The input is the current i data and basic circuit parameters (L m , k i , C e , G e , kh wait)

[0196] The output is the correction value of the current equation parameter, which is used to adjust the current equation

[0197] The same 3-layer feedforward neural network structure is used

[0198] 3. Flux equation parameter correction subnetwork:

[0199] The input is the flux ψ data and the core saturation S

[0200] The output is the correction value of the flux equation parameter, which is used to adjust the flux equation

[0201] This sub-network adopts a 3-layer feedforward neural network structure, and the activation function is Sigmoid

[0202] 4. Impedance equation parameter correction subnetwork:

[0203] The input is impedance Z data and temperature T

[0204] The output is the correction value of the impedance equation parameter, which is used to adjust the impedance equation

[0205] The sub-network also uses a 3-layer feedforward neural network, and the activation function is Tanh

[0206] 5. Summarize and correct the sub-network:

[0207] The input is the correction value of the output of the other 4 sub-networks

[0208] The output is a comprehensive correction parameter, which is used to integrate the correction information provided by each sub-network

[0209] This subnetwork contains 2 hidden layers, and the activation function is ReLU

[0210] Through the integration of this theoretical model and machine learning model, we hope to bring out the advantages of both: the theoretical model provides the physical basis and overall framework, while the neural network model uses measured data to continuously correct and optimize the model parameters, so that the overall model can better describe the complex fault-induced ferromagnetic resonance process.

[0211] Step S40: establishing a first training data set based on the first data set, and performing fitting training on the ferromagnetic resonance overvoltage characteristic combination model.

[0212] In this step, training input features are selected from the first data set D1, including voltage v, current i, frequency f, power factor cosφ, temperature T, harmonic content THD and core saturation S, etc.

[0213] The training output includes the following aspects:

[0214] 1. Overvoltage peak value U p

[0215] 2. Overvoltage duration t p

[0216] 3. Overvoltage frequency characteristic f p

[0217] 4. Correction values ​​of various equation parameters, such as voltage equation parameters (C, R s ,L s ,k v ) correction value, current equation parameter (L m ,k i ,C e ,G e ,k h ) correction value, flux equation parameter (B r ,H c ,μ i ,k T ,k f ) correction value, impedance equation parameters (R0, α, k R ,X0,L0,k X ,B s )’s correction value.

[0218] The ferromagnetic resonance overvoltage characteristic combination model established in step S30 is fitted and trained using a gradient descent method or other optimization algorithms, with the goal of minimizing the error between the training sample and the model output. After the training is completed, a well-fitted and trained ferromagnetic resonance overvoltage characteristic combination model M1 is obtained.

[0219] The model M1 can accurately describe the ferroresonance overvoltage characteristics under a known single fault condition. However, in order to further improve the adaptability of the model, it is necessary to use a second data set containing multiple fault types for fine-tuning.

[0220] Step S60: Using the second data set, establish a second training data set, and fine-tune the ferromagnetic resonance overvoltage characteristic combination model established in step S30.

[0221] In this step, training input features are selected from the second data set D2, including voltage v, current i, frequency f, power factor cosφ, temperature T, harmonic content THD, core saturation S, and fault type F. The training output still includes the overvoltage peak value U p , overvoltage duration t p , overvoltage frequency characteristics f p And the corrected values ​​of the parameters of each equation.

[0222] In order to adapt to the situation of multiple fault superposition, the LORA (Low-Rank Adaptation) fine-tuning algorithm is used to adjust the parameters of the initial model M1 trained in step S40. Specifically, the LORA fine-tuning will optimize the parameters of the hidden layer and output layer of the modified model in M1 by gradient descent method to adapt to the composite fault situation.

[0223] After LORA fine-tuning, a refined model M2 for composite fault conditions is obtained, called the LORA model. This model M2 can better describe the impact of single faults and multiple fault superposition on the ferroresonance overvoltage characteristics.

[0224] LoRa model M2 includes two key sub-networks:

[0225] 1. Fault basis identification sub-network:

[0226] Inputs are electrical data v, i, f, cosφ, T, THD, S and fault type F

[0227] The output is the fault signature, which is used to identify and classify different types of faults.

[0228] This subnetwork adopts a 4-layer feedforward neural network structure, with 32, 16, and 8 hidden layer nodes, and the activation function is ReLU.

[0229] 2. Corrected model of ferromagnetic resonance equations for superimposed fault-induced ferromagnetic resonance:

[0230] The input is the fault characteristics output by the fault basis identification subnetwork and the parameter correction characteristics output by each subnetwork.

[0231] The output is the correction parameters of the ferromagnetic resonance equations

[0232] This subnetwork adopts a 5-layer feedforward neural network structure, with 64, 32, 16, 8, and 4 hidden layer nodes, and the activation function is Sigmoid.

[0233] Through the collaboration of these two sub-networks, LoRa model M2 can achieve the following functions:

[0234] 1. First, use the fault basis identification subnetwork to identify and classify different types of faults to obtain the fault feature vector.

[0235] 2. Then, these fault characteristics and the outputs of each parameter correction sub-network obtained by training in step S40 are input into the correction model of the ferromagnetic resonance equation group of superimposed fault-induced ferromagnetic resonance.

[0236] 3. The correction model will output the correction values ​​of each parameter of the ferromagnetic resonance equation group established in step S20 according to the fault situation.

[0237] 4. Finally, the theoretical equations are integrated with the output of the LORA model to obtain a complete model that can accurately describe the fault-induced ferromagnetic resonance overvoltage characteristics.

[0238] In this way, the advantages of theoretical models and machine learning models are fully utilized, which not only retains the theoretical basis with clear physical meaning, but also can continuously optimize and improve the model parameters in combination with measured data, greatly improving the prediction accuracy of overvoltage characteristics under complex fault conditions.

[0239] Step S41: Parameter input

[0240] In this step, multiple sets of different initial parameter values ​​need to be selected as input, including but not limited to basic circuit parameters such as voltage v, current i, frequency f, power factor cosφ, and environmental factors such as temperature T and core saturation S.

[0241] In order to achieve the optimal parameter correction effect, the following strategies are adopted:

[0242] 1. Determine the parameter range:

[0243] According to the working principle and actual operation of the electromagnetic voltage transformer, determine the reasonable range of variation of each parameter, for example:

[0244] Voltage v: 0.8U n ~1.2U n

[0245] Frequency f: 45Hz~65Hz

[0246] Power factor cosφ: 0.8 lagging ~ 1.0 ~ 0.8 leading

[0247] Temperature T: -40℃~80℃

[0248] Core saturation S: 0.6~1.2

[0249] 2. Using Latin hypercube sampling method:

[0250] Latin hypercube sampling is used to generate uniformly distributed sample points within a certain parameter range. This method ensures that the samples in the multidimensional parameter space are well representative and uniform.

[0251] 3. Consider extreme cases:

[0252] On the basis of sampling, some extreme parameter combinations are added, such as the maximum voltage and the maximum temperature, the minimum frequency and the maximum core saturation, etc., to test the performance of the model under boundary conditions.

[0253] 4. Includes typical working conditions:

[0254] Add some known typical operating condition parameter combinations, which often appear in actual operation or have special significance.

[0255] 5. Consider parameter correlation:

[0256] There may be correlations between some parameters, such as temperature increase may cause core saturation changes. When selecting parameter combinations, these correlations should be considered to generate parameter combinations that meet actual conditions.

[0257] 6. Dynamic Adjustment:

[0258] After initial training, analyze the performance of the model in different parameter intervals. For intervals with poor performance, increase the number of sample points in that interval and perform focused sampling.

[0259] 7. Cross Validation:

[0260] The selected parameter group is divided into training set and validation set, and the cross-validation method is used to evaluate the performance of the model under different parameter combinations to ensure that the selected parameter group has good generalization ability.

[0261] 8. Iterative optimization:

[0262] Based on the initial training results, optimization algorithms (such as Bayesian optimization) are used to guide subsequent parameter selection, and gradually find the parameter combination that best reflects the system characteristics.

[0263] Through the above steps, multiple sets of different initial parameter values ​​can be systematically selected as input to cover various possible working conditions and fault conditions, which helps to train a model with strong robustness and good generalization ability. This method can improve the accuracy of parameter correction and ensure that the model can maintain good performance in the face of various actual situations.

[0264] Step S42: Solving the equations

[0265] In this step, the parameter values ​​selected in step S41 are substituted into the ferromagnetic resonance equations established in step S20, and the output results of the equations are obtained by a numerical solution method (such as the Runge-Kutta method), including the overvoltage peak value U p , overvoltage duration t p and overvoltage frequency characteristic f p wait.

[0266] Step S43: Actual data collection

[0267] According to the parameter conditions input in step S41, the corresponding fault conditions are simulated on the actual electromagnetic voltage transformer, and multiple sets of electrical data are collected, including the actual measured overvoltage peak value U p, overvoltage duration t p and overvoltage frequency characteristic f p wait.

[0268] Step S44: Error calculation

[0269] In this step, the theoretical output results of the equation group in step S42 are compared with the actual data collected in step S43 to calculate various indicators (overvoltage peak value U p , overvoltage duration t p , overvoltage frequency characteristics f p ) error.

[0270] Step S45: Determine parameter correction value

[0271] 1. Based on the error calculated in step S44, use an optimization algorithm (such as gradient descent method) to iteratively adjust the parameters of the equation group until the error is reduced to an acceptable range.

[0272] 2. The difference between the final determined parameters and the initial parameters is used as the equation parameter correction value.

[0273] 3. This process continuously adjusts the equation parameters by comparing the theoretical calculation results with the actual measurement data, and finally obtains the corrected parameters that can accurately describe the ferromagnetic resonance overvoltage characteristics of fault excitation.

[0274] In general, in this embodiment 1, steps S41-S45 achieve the goal of comparing theoretical calculation results with measured data, using an optimization algorithm to continuously adjust the parameters of the equation group, and finally obtaining a set of correction parameters that can accurately describe the fault-induced ferroresonance overvoltage characteristics. This method fully utilizes the advantages of theoretical models and experimental data, and improves the accuracy and robustness of the model.

[0275] In order to further better understand and implement the present invention, the following provides Example 2 of a specific application scenario of the present invention: A certain power department recently discovered an abnormality in the electromagnetic voltage transformer of a 220kV substation. The secondary side voltage of the voltage transformer often has serious overvoltage faults, which poses a serious threat to the safe and stable operation of the entire power grid. The technicians decided to use the method for determining the ferromagnetic resonance overvoltage characteristics of fault excitation proposed by the present invention to conduct in-depth analysis and diagnosis of the voltage transformer.

[0276] First, the technicians collected the electrical data of the voltage transformer under a known single fault condition according to step S10. Specifically, it includes:

[0277] Voltage transformer rated voltage U n =220kV, rated frequency f n =50Hz.

[0278] Under three single fault conditions, namely short circuit fault, ground fault and open circuit fault, the voltage v, current i, frequency f, power factor cosφ, temperature T, overvoltage peak U p , harmonic content THD and core saturation S, etc., together form the first data set D1. Among them, the value range of each physical quantity is as follows:

[0279] Voltage v: 198kV~242kV

[0280] Current i:80A~120A

[0281] Frequency f: 48Hz~52Hz

[0282] Power factor cosφ: 0.85 lagging ~ 0.95 leading

[0283] Temperature T: 20℃~40℃

[0284] Overvoltage peak value U p :1.2U n ~1.8U n

[0285] Harmonic content THD: 2%~8%

[0286] Core saturation S: 0.7~1.1

[0287] By analyzing the first data set D1, the technicians established a theoretical equation group describing the ferromagnetic resonance process of the voltage transformer, including a voltage equation, a current equation, a flux equation, and an impedance equation, as described in step S20. Taking the voltage equation as an example, its expression is:

[0288]

[0289] Among them, C = 0.01μF is the equivalent capacitance, R s =0.2Ω is the equivalent series resistance, L s =0.5mH is the equivalent series inductance, k v =0.02 is the voltage nonlinear coefficient. Fundamental voltage amplitude V s =195kV, fundamental angular frequency ω = 2π × 50rad / s, fundamental phase angle φ = 0.1rad. The nth harmonic voltage amplitude V n and phase angle φ n It is obtained through Fourier analysis.

[0290] Similarly, the parameters of other equations are also determined and calibrated based on the first data set D1.

[0291] However, it is difficult to fully describe the overvoltage characteristics under actual fault conditions with only these theoretical equations. Therefore, the technicians establish a combined model of ferromagnetic resonance overvoltage characteristics consisting of a theoretical equation group and a correction model according to step S30.

[0292] The revised model includes five sub-networks, all of which adopt a feedforward neural network structure:

[0293] Voltage equation parameter correction subnetwork:

[0294] The input is voltage v data and basic circuit parameters, C, R s , L s , k v wait.

[0295] The output is the correction value of the voltage equation parameter, which is used to adjust the voltage equation.

[0296] This subnetwork contains 3 hidden layers, with the number of hidden layer nodes being 16, 8, and 4 respectively. The activation function is ReLU, and the number of nodes in the final output layer is 4.

[0297] Current equation parameter correction subnetwork:

[0298] The input is current i data and basic circuit parameters, L m , k i , C e , G e , k h wait.

[0299] The output is the correction value of the current equation parameter, which is used to adjust the current equation.

[0300] This sub-network also contains 3 hidden layers with 16, 8, and 4 nodes respectively, the activation function is ReLU, and the number of nodes in the output layer is 5.

[0301] Flux equation parameter correction subnetwork:

[0302] The input is the magnetic flux ψ data and the core saturation S.

[0303] The output is the correction value of the flux equation parameters, which is used to adjust the flux equation.

[0304] The number of hidden layer nodes of this sub-network is 12, 6, and 3, the activation function is Sigmoid, and the number of output layer nodes is 3.

[0305] Impedance equation parameter correction subnetwork:

[0306] The input is impedance Z data and temperature T.

[0307] The output is the correction value of the impedance equation parameters, which is used to adjust the impedance equation.

[0308] The number of hidden layer nodes of this sub-network is 10, 5, and 2, the activation function is Tanh, and the number of output layer nodes is 4.

[0309] Summarize the modified subnetwork:

[0310] The input is the corrected value of the output of the other 4 sub-networks.

[0311] The output is a comprehensive correction parameter, which is used to integrate the correction information provided by each sub-network.

[0312] This subnetwork contains 2 hidden layers with 8 and 4 nodes, the activation function is ReLU, and the number of nodes in the output layer is 1.

[0313] By training the first data set D1, the initial ferromagnetic resonance overvoltage characteristic combination model M1 was obtained.

[0314] In order to further improve the adaptability of the model, the technicians collected a second data set D2 containing multiple fault types according to steps S50 and S60. The data set includes electrical data under the conditions of short circuit fault, ground fault, open circuit fault, and superimposed fault of 1 short circuit fault + 1 ground fault. The value range of each physical quantity is as follows:

[0315] Voltage v: 190kV~250kV

[0316] Current i:70A~130A

[0317] Frequency f: 47Hz~53Hz

[0318] Power factor cosφ: 0.8 lagging ~ 0.98 leading

[0319] Temperature T: 15℃~45℃

[0320] Overvoltage peak value U p :1.3U n ~2.0U n

[0321] Harmonic content THD: 3%~10%

[0322] Core saturation S: 0.6~1.2

[0323] Based on the second data set D2, the technicians fine-tuned the initial model M1 using LORA and obtained a refined model M2 for complex fault conditions. The LORA model includes two key sub-networks:

[0324] Fault basis identification subnetwork:

[0325] The inputs are electrical data v, i, f, cosφ, T, THD, S and fault type F.

[0326] The output is fault signature, which is used to identify and classify different types of faults.

[0327] This sub-network adopts a 4-layer feedforward neural network structure, with the number of hidden layer nodes being 32, 16, and 8 respectively, and the activation function being ReLU.

[0328] Corrected model of ferromagnetic resonance equations for superimposed fault-induced ferromagnetic resonance:

[0329] The input is the features and fault features output by each sub-network.

[0330] The output is the correction parameters to the ferromagnetic resonance equations.

[0331] This sub-network adopts a 5-layer feedforward neural network structure, with the number of hidden layer nodes being 64, 32, 16, 8, and 4, and the activation function being Sigmoid.

[0332] Through LORA fine-tuning, model M2 can better adapt to the complex situations of single fault and multiple fault superposition.

[0333] In order to test the performance of the LORA model, technicians designed the following scenario:

[0334] In a 220kV substation, a short circuit fault occurs on the primary side of the voltage transformer, and a ground fault also occurs on the secondary side. The first and second digits represent short circuit fault and ground fault respectively, and the third and fourth digits do not consider open circuit fault and other faults for the time being.

[0335] By inputting this fault vector into the joint model of step S70, the corrected parameters of the ferromagnetic resonance equation group can be obtained. Taking the voltage equation as an example, the corrected parameters are as follows:

[0336] C=0.008μF,R s =0.22Ω,L s =0.55mH,k v =0.025

[0337] Fundamental voltage amplitude V s =180kV, fundamental phase angle φ = 0.15rad. The 2nd to 5th harmonic voltage amplitudes and phase angles also change accordingly.

[0338] Substituting the corrected parameters into the theoretical equations and solving them numerically, we can obtain the time domain waveform of the overvoltage as follows: Figure 3As shown in the figure, it can be seen that in the combined case of short circuit fault + ground fault, the voltage transformer has a serious overvoltage fault, with a peak voltage as high as 1.8 times the rated voltage and a duration of about 0.1 second. This overvoltage may not only cause damage to the voltage transformer itself, but also endanger the safe and stable operation of the entire power grid.

[0339] In order to more comprehensively analyze the ferromagnetic resonance characteristics of this complex fault situation, the technicians further checked the changes in other physical quantities:

[0340] Current waveform Figure 4 As shown, the peak current reaches 2 times the rated current.

[0341] The magnetic flux waveform is as follows Figure 5 As shown, the peak value of magnetic flux reaches 1.8Wb, which is close to the saturation of the core.

[0342] Impedance changes as Figure 6 As shown in the figure, the impedance amplitude decreases rapidly at the beginning of the fault and then recovers gradually.

[0343] Comprehensive analysis of the above results shows that in the combined case of short circuit fault + ground fault, the voltage transformer has a serious ferromagnetic resonance overvoltage fault. The overvoltage peak reaches 1.8U n , lasting about 0.1 seconds, and the current, magnetic flux and impedance also changed dramatically. This overvoltage fault seriously threatens the safety of the voltage transformer and the entire power grid, and timely protection measures must be taken. This discovery reflects the effect of the method of the present invention.

[0344] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for determining the ferromagnetic resonance overvoltage characteristic of a fault, characterized in that: The following steps are involved: S10, collecting electrical data of ferromagnetic resonance excited by various faults of an electromagnetic voltage transformer with a known single fault, including voltage, current, frequency, power factor, temperature, overvoltage, harmonic content, and core saturation, and recording the data as a first data set; S20, establishing a ferromagnetic resonance equation group including the electrical data, including a voltage equation, a current equation, a magnetic flux equation, and an impedance equation; S30, establishing a combined model of ferromagnetic resonance overvoltage characteristics including the ferromagnetic resonance equation group and a correction model, wherein the correction model is used to correct the parameters of the ferromagnetic resonance equation group, and the correction model is a neural network with a multi-input and multi-output structure, including a voltage equation parameter correction subnetwork, a current equation parameter correction subnetwork, a flux equation parameter correction subnetwork, an impedance equation parameter correction subnetwork, and a summary correction subnetwork; S40, based on the first data set, establishing a first training data set, wherein the training input is voltage, current, frequency, power factor, temperature, harmonic content, and core saturation, and the training output is overvoltage amplitude, duration, frequency characteristics, and correction values ​​of parameters of each equation; using the first training data set to perform fitting training on the ferromagnetic resonance overvoltage characteristic combination model, and obtain the ferromagnetic resonance overvoltage characteristic combination model after fitting training, which is recorded as the first model; S50, collecting electrical data of ferromagnetic resonance excited by multiple different faults of an electromagnetic voltage transformer known to have at least one fault, and recording the data as a second data set; S60, using the second data set to establish a second training data set, wherein the training input is voltage, current, frequency, power factor, temperature, harmonic content, core saturation and fault type, and the training output is overvoltage amplitude, duration, frequency characteristics and correction values ​​of each equation parameter; using the second training data set to fine-tune the first model to obtain a Lora model; S70, using the combined model and the Lora model as a joint model, inputting a fault vector of a single fault or a plurality of faults superimposed, and obtaining a group of ferromagnetic resonance equations for describing ferromagnetic resonance overvoltage characteristics excited by the fault.

2. The method for determining the ferromagnetic resonance overvoltage characteristic caused by a fault according to claim 1, characterized in that: The voltage equation parameter correction subnetwork has a training input of voltage data and basic circuit parameters, and a training output of voltage equation parameter correction values, which is used to correct the voltage equation parameters according to actual voltage data. The structure is a three-layer feedforward neural network.

3. The method for determining the ferromagnetic resonance overvoltage characteristic caused by a fault according to claim 2, characterized in that: The current equation parameter correction subnetwork has a training input of current data and basic circuit parameters, and a training output of current equation parameter correction values, which is used to correct the current equation parameters according to actual current data. The structure is a three-layer feedforward neural network.

4. The method for determining the ferromagnetic resonance overvoltage characteristic caused by a fault according to claim 3, characterized in that: The flux equation parameter correction subnetwork has a training input of flux data and core saturation, and a training output of flux equation parameter correction values, which is used to correct flux equation parameters according to actual flux data. The structure is a three-layer feedforward neural network.

5. The method for determining the ferromagnetic resonance overvoltage characteristic caused by a fault according to claim 4, characterized in that: The impedance equation parameter correction subnetwork has a training input of impedance data and temperature and a training output of impedance equation parameter correction values, which is used to correct impedance equation parameters according to actual impedance data. The structure is a three-layer feedforward neural network.

6. The method for determining the ferromagnetic resonance overvoltage characteristic caused by a fault according to claim 5, characterized in that: The training input of the summary correction subnetwork is the correction value output by each subnetwork of the correction model excluding the summary correction subnetwork, and the training output is a comprehensive correction parameter used to integrate the correction values ​​provided by each subnetwork. The structure is a two-layer feedforward neural network.

7. The method for determining the ferromagnetic resonance overvoltage characteristic caused by a fault according to claim 6, characterized in that: The Lora model includes a fault basis identification subnetwork and a correction model of a ferromagnetic resonance equation group of superimposed fault-induced ferromagnetic resonance.

8. The method for determining the ferromagnetic resonance overvoltage characteristic caused by a fault according to claim 7, characterized in that: The said fine-tuning refers to adjusting the parameters of the hidden layer and the output layer of each sub-network of the correction model in the first model, so as to adjust the weights and biases of these layers by the gradient descent method to adapt to the fault conditions of multiple types of fault superposition.

9. The method for determining the ferromagnetic resonance overvoltage characteristic caused by a fault according to claim 7, characterized in that: The fault basis identification subnetwork has electrical data and fault type as training inputs, and fault characteristics as training outputs, which are used to identify and classify different types of faults. Its structure is a four-layer feedforward neural network. The correction model of the ferromagnetic resonance equation group for superimposed fault-induced ferromagnetic resonance has characteristics and fault characteristics of each subnetwork output as training inputs, and correction parameters as training outputs, which are used to correct the ferromagnetic resonance equation group according to the fault situation. Its structure is a five-layer feedforward neural network.

10. A method for determining a fault-induced ferromagnetic resonance overvoltage characteristic according to any one of claims 1 to 9, characterized in that: The step of obtaining the equation parameter correction value specifically includes: S41, parameter input: select multiple groups of different initial parameter values ​​as input, these parameters include at least one of the following: voltage, current, frequency, power factor, temperature, core saturation; S42, solving the equation group: Substituting the parameter values ​​selected in S41 into the ferromagnetic resonance equation group, solving the equation group to obtain the output results, including the overvoltage amplitude, duration, and frequency characteristics; S43, actual data collection: according to the parameter conditions input in S1, the corresponding fault conditions are simulated on the actual electromagnetic voltage transformer, and multiple sets of electrical data are collected, including the actual measured overvoltage amplitude, duration, and frequency characteristics; S44, error calculation: compare the theoretical output results of the equation group in S42 with the actual data collected in S43, and calculate the errors of various indicators; S45. Determine parameter correction values: Based on the error calculated in S44, use an optimization algorithm to iteratively adjust the parameters of the equation group until the error is reduced to a predetermined range, and use the difference between the final determined parameters and the initial parameters as the equation parameter correction value.

Citation Information

Patent Citations

  • Ferromagnetic resonance overvoltage type identification method and system

    CN105404734A

  • Power distribution network voltage transformer ferromagnetic resonance judgment method and system

    CN108020805A

  • Method for detecting and diagnosing ferromagnetic resonance of isolated neutral system in real time

    CN109459657A

  • Method and device for suppressing oscillation of capacitor voltage transformer

    CN111786388A

  • Ferromagnetic resonance fault pre-judgment method for mutual inductor in power supply system

    CN117826059A

Cited By

  • Effectiveness verification method for current-limiting and resonance-eliminating device of electromagnetic voltage transformer

    CN119064845A

  • A method for verifying the effectiveness of a current limiting and detuning device for an electromagnetic voltage transformer

    CN119064845B