Transformer core and clamp grounding dual current measurement method

Abstract

The application provides a transformer core and clamp grounding double-current measurement method, and belongs to the technical field of transformer cores and clamps, and comprises the following steps: adopting a first current transformer and a second current transformer to monitor the grounding currents of a transformer core and a clamp respectively; recording the preprocessed first voltage sequence and second voltage sequence respectively; obtaining the first net voltage sequence and the second net voltage sequence after noise reduction; performing difference calculation after alignment to obtain a numerical sequence representing the actual grounding currents of the transformer core and the clamp, and recording the numerical sequence as a grounding current sequence; adopting a pre-trained grounding current calibration model to perform noise reduction and calibration on the grounding current sequence to obtain a standard grounding current sequence; optimizing the model parameters according to the grounding current sequence to make the grounding current calibration model more suitable for noise reduction and calibration on the grounding current sequence; obtaining the standard grounding current sequence obtained through continuous collection and calculation, inputting the standard grounding current sequence into a pre-trained core and clamp grounding judgment model to obtain a judgment result of whether the transformer core and the clamp are grounded.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of transformer core and clamp, and particularly relates to a transformer core and clamp grounding double-current measurement method. BACKGROUND

[0002] The power system is an important infrastructure for modern social production and life, and its safe and reliable operation is crucial for promoting economic development and maintaining social stability. As a key device in the power system, the operation status of the transformer directly affects the operation quality and power supply reliability of the entire power grid. Therefore, real-time monitoring of the transformer is necessary, especially the monitoring of the grounding state of the transformer core and clamp. Transformer core and clamp grounding refers to the case where the transformer core is accidentally connected to the grounding device. When this happens, the voltage of the transformer core and clamp will decrease, causing the core to saturate when working, resulting in safety hazards such as overheating, vibration, and noise. At the same time, it will also cause zero-sequence current to circulate on the grounding line, causing a series of hazards to power equipment and systems, such as equipment damage, reclosing failure, and power metering error. Therefore, timely detection and positioning of transformer core and clamp grounding faults are of great significance to ensure the safe and economic operation of the power system.

[0003] Currently, the detection of the grounding state of the transformer core and clamp mainly uses the electrical method, but due to the existence of complex space electromagnetic fields around the transformer oil tank wall, the leakage magnetic field of the transformer itself and the electromagnetic field of the surrounding equipment will cause the current measurement results of the core and clamp grounding line to have large fluctuations. The randomness of the space electromagnetic field greatly interferes with the measurement of the core and clamp grounding current. It is difficult to shield the space electromagnetic field during field measurement, so the current measurement lacks accuracy. SUMMARY

[0004] Therefore, the present application provides a transformer core and clamp grounding double-current measurement method, which can solve the technical problem of lack of accuracy in measuring the grounding current due to the randomness of the space electromagnetic field interfering with the measurement of the core and clamp grounding current when detecting the grounding state of the transformer core and clamp using the electrical method.

[0005] The present application is implemented as follows:

[0006] The first aspect of the present application provides a transformer core and clamp grounding double-current measurement method, which includes the following steps:

[0007] S10, the first current transformer and the second current transformer are used to monitor the grounding current of the transformer core and the clamp respectively, wherein the first current transformer monitors the current on the grounding line of the core and the clamp, and the current is referred to as the first current; the first current is actually the sum of the grounding current of the core and the clamp and the current generated by the magnetic flux leakage, Ir+Ig; the second current transformer only monitors the interference current Ig generated by the magnetic flux leakage;

[0008] S20, the current signals detected by the first current transformer and the second current transformer are converted into corresponding voltage signals;

[0009] S30, the converted voltage signals are preprocessed, including filtering, amplification, and analog-to-digital conversion, to obtain voltage digital signals, which are referred to as the first voltage sequence and the second voltage sequence respectively;

[0010] S40, the pre-trained voltage denoising model is used to denoise the first voltage sequence and the second voltage sequence respectively to obtain the first net voltage sequence and the second net voltage sequence after denoising;

[0011] S50, after aligning the first net voltage sequence and the second net voltage sequence, difference calculation is performed to obtain a numerical sequence representing the actual grounding current of the core and the clamp, referred to as the grounding current sequence;

[0012] S60, the pre-trained grounding current calibration model is used to denoise and calibrate the grounding current sequence to obtain a standard grounding current sequence; wherein the grounding current calibration model further includes optimizing the model parameters according to the grounding current sequence during the denoising and calibration process, so that the grounding current calibration model is more suitable for denoising and calibrating the grounding current sequence;

[0013] S70, the standard grounding current sequence obtained by continuous acquisition and calculation is input into the pre-trained core and clamp grounding judgment model to obtain the judgment result of whether the transformer core and the clamp are grounded.

[0014] On the basis of the above technical solutions, the transformer core and clamp grounding double current measurement method of the application can be further improved as follows:

[0015] The specific steps of S10 include: arranging a first current transformer on the grounding line of the core and the clamp outside the transformer, connecting the primary winding of the first current transformer to the grounding line of the core and the clamp, and obtaining the current signal of the total current Ir+Ig flowing through the grounding line of the core and the clamp by electromagnetic induction principle; the primary winding of the second current transformer is arranged opposite to the primary winding of the first current transformer, and is used to obtain the current signal of the interference current Ig generated by the magnetic flux leakage.

[0016] The specific steps of S30 include: filtering, amplifying and analog-to-digital converting the voltage signal corresponding to the first current transformer to obtain a first voltage digital sequence; filtering, amplifying and analog-to-digital converting the voltage signal corresponding to the first current transformer to obtain a second voltage digital sequence.

[0017] The establishment and training steps of the voltage denoising model include:

[0018] Step 1, collect a large amount of standard voltage data and corresponding noisy voltage data as first training data;

[0019] Step 2, normalizing the first training data for pretreatment;

[0020] Step 3, selecting a BP neural network as a model structure;

[0021] Step 4, training the BP neural network with the pretreated first training data, and iteratively updating the model parameters with gradient descent as an optimization algorithm to obtain the voltage denoising model.

[0022] The specific steps of S50 include: aligning the first net voltage sequence and the second net voltage sequence obtained in step S40 on the time axis; and performing point-by-point subtraction operation on the first net voltage sequence and the second net voltage sequence to obtain a ground current sequence representing the actual core and clamp ground current.

[0023] The establishment and training steps of the ground current calibration model include:

[0024] Step 1, collect standard ground current data and original ground current data as second training data;

[0025] Step 2, normalizing and detrending the second training data for pretreatment;

[0026] Step 3, selecting a long short-term memory neural network as a model structure;

[0027] Step 4, training the long short-term memory neural network with the pretreated second training data, and iteratively updating the model parameters with gradient descent as an optimization algorithm to obtain the ground current calibration model.

[0028] The establishment and training steps of the core and clamp ground judgment model include:

[0029] Step 1, collect standard ground current data under normal grounding and non-grounding conditions as third training data, and label the grounding state as a training set label;

[0030] Step 2, normalizing and denoising preprocessing of the third training data;

[0031] Step 3, selecting a support vector machine model as a model structure;

[0032] Step 4, training the support vector machine model using the preprocessed third training data, using gradient descent as an optimization algorithm to iteratively update the model parameters, to obtain the core and clamp grounding judgment model.

[0033] The method used to optimize the model parameters is: grey wolf hunting algorithm + genetic algorithm.

[0034] Further, when training the model, regularization techniques and learning rate scheduling strategies are introduced to alleviate overfitting problems and speed up convergence.

[0035] In step S70, when the grounding state of the standard current sequence is judged, a model integration framework is constructed, that is: artificial neural network, support vector machine and random forest three models are used to classify the sequence at the same time; the output results of the three models are weighted and fused to obtain the final grounding judgment result.

[0036] Compared with the prior art, the transformer core and clamp grounding double current measurement method provided by the application has the beneficial effects that: the project adopts a double current transformer monitoring method, that is, one current transformer monitors the sum of the grounding current and the leakage magnetic current (Ir+Ig) of the core and clamp, and the other current transformer monitors the interference current Ig generated by the leakage magnetic current, so the actual measurement value of the transformer core and clamp grounding current is the difference between the currents detected by the two current transformers. By effectively excluding the interference of the space electromagnetic field of the transformer, the monitoring accuracy of the core and clamp grounding current in the strong electromagnetic field is improved, the problem of large on-site current test data is solved, and the effectiveness of the transformer multi-point grounding fault detection is improved.

[0037] Further, the method combines advanced machine learning algorithms and signal processing techniques, can intelligently extract feature patterns from raw measurement data, and accurately judge the grounding state of the transformer core and clamp. Compared with artificial experience analysis, this method reduces the influence of subjective judgment, improves the objectivity and accuracy of detection. At the same time, through modeling training, the method can adapt to data under various complex working conditions, and has strong robustness.

[0038] Therefore, the technical scheme of the application solves the technical problem that when the electrical method is used to detect the grounding of the transformer core and clamp, the randomness of the space electromagnetic field causes great interference to the measurement of the core and clamp grounding current, resulting in lack of accuracy in measuring the grounding current. BRIEF DESCRIPTION OF DRAWINGS

[0039] 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 also be obtained by those skilled in the art without any creative effort on the basis of these drawings.

[0040] Figure 1 It is a schematic diagram of double-current detection of the present application;

[0041] Figure 2 It is a flow chart of the method provided by the present application;

[0042] Figure 3 It is a schematic diagram of single-coil measurement wiring and current vector diagram;

[0043] Figure 4 It is a schematic diagram of double-coil measurement wiring and current vector diagram. DETAILED DESCRIPTION

[0044] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the following will describe the technical solutions in the embodiments of the present application clearly and completely with reference to the drawings in the embodiments of the present application.

[0045] As Figure 1 shown, it is a deployment schematic diagram of a transformer core and clamp grounding double-current measurement method provided by the present application, Figure 2 It is a flow chart of the method, and the method comprises the following steps:

[0046] S10, a first current transformer and a second current transformer are used to monitor the grounding currents of the transformer core and the clamp, wherein the first current transformer monitors the sum of the core and clamp grounding current Ir and the interference current Ig generated by the leakage magnetic field, i.e. Ir+Ig, and the second current transformer only monitors the interference current Ig generated by the leakage magnetic field;

[0047] S20, the current signals detected by the first current transformer and the second current transformer are converted into corresponding voltage signals;

[0048] S30, the converted voltage signals are preprocessed, including filtering, amplification, analog-to-digital conversion, to obtain voltage digital signals, which are respectively recorded as a first voltage sequence and a second voltage sequence;

[0049] S40, a pre-trained voltage denoising model is used to perform denoising processing on the first voltage sequence and the second voltage sequence respectively, to obtain a first denoised voltage sequence and a second denoised voltage sequence;

[0050] S50, difference calculation is performed after the first net voltage sequence and the second net voltage sequence are aligned, to obtain a numerical sequence representing actual core and clamp grounding current, recorded as a grounding current sequence;

[0051] S60, a pre-trained grounding current calibration model is used to denoise and calibrate the grounding current sequence to obtain a standard grounding current sequence; wherein the grounding current calibration model further includes optimizing model parameters according to the grounding current sequence in the process of denoising and calibrating the grounding current sequence, so that the grounding current calibration model is more suitable for denoising and calibrating the grounding current sequence;

[0052] S70, the standard grounding current sequence obtained by continuous acquisition and calculation is obtained, and the pre-trained core and clamp grounding judgment model is input into the standard grounding current sequence to obtain the judgment result of whether the transformer core and clamp are grounded.

[0053] Among them, the first current transformer and the second current transformer both use a Hall effect current sensor, and the acquisition and calculation unit uses a single-chip microcomputer.

[0054] Preferably, in order to avoid excessive current, a current limiting impedance unit is arranged between the transformer grounding line and the ground, which can automatically / manual adjust the current limiting impedance according to the grounding current size to temporarily limit the current within the regulations when the core and clamp grounding current exceeds the standard, which provides protection for the safe operation of the transformer. The current limiting circuit adopts a parallel connection form, wherein Z1, Z2, Z3, Z4, Z5. The single-chip microcomputer realizes the switching of each branch impedance through the switch. Among them, the Z5 branch does not have a switch connected in series. This design can avoid the open circuit of the core grounding lead. Normally, K1 is closed, that is, the switch K1 shorts all the current limiting resistors, and the current limiting resistance is 0Ω at this time. When the system detects that the current exceeds the standard (≥0.1A), each resistor is sequentially connected to the circuit: the system first opens K1, closes K2, K3, K4, K5, and connects the resistance values of Z1, Z2, Z3, Z4, Z5 in parallel to the grounding line; if the system detects that the current is still excessive, then open K2, at this time it is equivalent to connecting the resistance values of Z2, Z3, Z4, Z5 in parallel to the grounding line; and the process is repeated until the current is limited to below 0.1A.

[0055] The method of inputting the shunt winding proposed above is equivalent to connecting a resistance and an inductance in series in the loop, and in particular, the inductance of the winding greatly limits the high-frequency current, further reduces the effective value of the fault current, and has low thermal capacity requirements, solving the problem of high thermal capacity requirements relying only on series resistance. The on-off of the current limiting impedance unit adopts switch control of the on-off of the branch, and the impedance is switched by multiple levels to quickly limit the grounding current within the standard (less than 0.1 A). The traditional current limiting device is only a series resistance, and the series impedance of the present application greatly limits the high-frequency current due to the presence of inductance, and the current limiting effect is obvious due to the traditional current limiting resistance.

[0056] The specific implementation of the method steps provided by the present application is described in detail below:

[0057] The specific implementation of step S10 is as follows: first, a first current transformer is arranged on the outer core and clamp ground wire of the transformer, and the primary winding of the first current transformer is connected to the ground wire of the core and clamp. The current flowing through the core and clamp ground wire includes the sum of the current of the transformer core and clamp ground and the interference current generated by the transformer leakage magnetic field, i.e. Ir+Ig. The first current transformer can obtain the first current signal of the total current through electromagnetic induction principle. Then a second current transformer is arranged near the grounding point of the core and clamp inside the transformer, and the secondary winding of the second current transformer is not closed, but is only used to detect the magnetic field strength at that position. The magnetic field strength detected by the second current transformer is mainly caused by the interference current Ig generated by the leakage magnetic field near the core and clamp. The second current transformer can convert the signal of the interference current Ig into a voltage signal output through electromagnetic induction principle. The purpose of this step is to detect the signals of the total current Ir+Ig and the interference current Ig through the two current transformers respectively, to provide raw data for the subsequent steps.

[0058] The specific implementation of step S20 is as follows: the current signals detected by the first current transformer and the second current transformer in step S10 are respectively converted into corresponding voltage signals through a circuit. Specifically, for the current signal Ir+Ig detected by the first current transformer, it can be converted into a corresponding voltage signal through a precision resistor; for the current signal Ig detected by the second current transformer, the voltage signal induced by its secondary winding can be directly used as the output. The purpose of this step is to convert the current signal into a voltage signal which is more convenient for subsequent processing.

[0059] The specific implementation of step S30 is to preprocess the first voltage signal and the second voltage signal converted in step S20, including filtering, amplification, analog-to-digital conversion and the like. The purpose of filtering is to remove high-frequency noise and harmonic components in the voltage signal, and a classic filter circuit such as a notch filter or a Butterworth filter can be used. The purpose of amplification is to amplify the voltage signal to a suitable range to facilitate subsequent acquisition and processing, and an operational amplifier circuit can be used. The purpose of analog-to-digital conversion is to convert the analog voltage signal into a digital signal to facilitate subsequent processing using digital signal processing techniques, and an analog-to-digital conversion circuit such as a SAR analog-to-digital converter can be used. After preprocessing, a digital sequence corresponding to the first voltage signal is obtained, denoted as the first voltage sequence; and a digital sequence corresponding to the second voltage signal is obtained, denoted as the second voltage sequence. The purpose of this step is to preprocess the original voltage signal to obtain a more accurate digital voltage sequence for subsequent processing.

[0060] The specific implementation of step S40 is to use a pre-trained voltage denoising model to denoise the first voltage sequence and the second voltage sequence obtained in step S30. Denoising refers to using machine learning algorithms to filter noise and repair distortion from the original signal to obtain a more accurate clean signal. Common voltage denoising models include wavelet transform denoising models, autoregressive filter models, adaptive linear neural network models, etc. These models all need to be trained with a large amount of standard data to optimize parameters and obtain better denoising performance. For example, for the wavelet transform denoising model, the training data can be first decomposed by wavelet transform to obtain wavelet coefficients of different scales; then the wavelet coefficients are thresholded according to a set threshold function to remove high-frequency noise components; and finally the thresholded wavelet coefficients are reconstructed to obtain the denoised clean signal. This step removes noise components from the original first voltage sequence and second voltage sequence through the voltage denoising model to obtain the first clean voltage sequence and the second clean voltage sequence, providing a more accurate data basis for subsequent calculations.

[0061] The specific implementation of step S50 is: first, align the first net voltage sequence and the second net voltage sequence obtained in step S40 on the time axis, that is, ensure that each time point of the two sequences is corresponding. Then, perform point-by-point subtraction operation on the two sequences, that is, subtract the second net voltage sequence from the first net voltage sequence, to obtain a numerical sequence representing the actual core and clamp ground current, denoted as a ground current sequence. Since the first net voltage sequence represents the signal of the total current Ir+Ig, and the second net voltage sequence represents only the signal of the interference current Ig, the difference between the two can remove the influence of the interference current Ig, thereby obtaining the digital signal of the actual core and clamp ground current Ir. The purpose of this step is to separate the actual core and clamp ground current signal from the signals of the total current and the interference current through subtraction operation.

[0062] The specific implementation of step S60 is: using a pre-trained ground current calibration model to further denoise and calibrate the ground current sequence obtained in step S50 to obtain a standard ground current sequence. The so-called ground current calibration model is a model based on a machine learning algorithm, which can perform denoising and amplitude calibration on the ground current sequence to make it more accurate and reliable. Common ground current calibration models include autoregressive models, long short-term memory neural network models, support vector machine regression models, etc. These models also need to be trained with a large amount of standard data to optimize parameters, thereby obtaining good denoising and calibration performance. For example, for a long short-term memory neural network model, the ground current sequence can be used as the input of the model, and through the training of the network, an ideal standard ground current sequence, i.e., the model output, can be fitted. While denoising and calibrating the ground current sequence, the model also has the function of adaptively adjusting parameters according to the input sequence, so that the model is more consistent with the actual ground current characteristics, further improving the accuracy of denoising and calibration. The purpose of this step is to remove noise and distortion components from the original ground current sequence and calibrate the amplitude through the processing of the ground current calibration model, and finally obtain an accurate standard ground current sequence, providing a reliable data basis for subsequent core and clamp grounding judgment.

[0063] The specific implementation of step S70 is: first, obtain a standard ground current sequence in a longer time period by continuously performing steps S10 to S60. Then, use a pre-trained core and clamp grounding judgment model to input the standard ground current sequence into the model, and the judgment result of whether the transformer core and clamp are grounded can be obtained.

[0064] The so-called core and clamp grounding judgment model is a model based on machine learning algorithm, which learns the characteristic patterns of grounding current sequences under normal grounding and ungrounding conditions through training on a large amount of historical data. Common core and clamp grounding judgment models include artificial neural network models, support vector machine models, random forest models, etc. In the training stage, these models will take a large number of standard grounding current sequences under normal grounding and ungrounding conditions as input, and the corresponding grounding state as output label. Through feature extraction and model learning, the internal laws of grounding current sequences under different grounding states are captured. For example, for an artificial neural network model, the grounding current sequence can be input to the input layer of the network, and through feature extraction and nonlinear mapping of the hidden layer, the probability value of the grounding state is obtained in the output layer. If the probability value is higher than the set threshold, it is judged as the grounding state, otherwise it is the ungrounded state. For the support vector machine model, an optimal classification hyperplane is constructed to effectively separate the grounding current sequences under different grounding states, thereby realizing the grounding state judgment.

[0065] It should be noted that in order to improve the judgment accuracy of the model, the training data needs to be carefully labeled, and appropriate feature engineering methods are used to extract effective features, and the hyperparameters of the model are optimized combined with artificial experience. In actual application, appropriate model structure and training strategy need to be selected according to specific circumstances to ensure the generalization ability of the model. In addition, the judgment result of the model also needs to be interactively verified and continuously improved with the actual situation on site.

[0066] The specific establishment and training steps of the voltage noise reduction model, the grounding current calibration model, and the core and clamp grounding judgment model can be summarized as follows:

[0067] 1. Establishment and training of voltage noise reduction model

[0068] (1) Collect a large number of standard voltage data and corresponding noisy voltage data as training set;

[0069] (2) Preprocess the training data, such as normalization, missing value completion, etc.;

[0070] (3) Select appropriate noise reduction model structure, such as wavelet transform, autoregressive filter or neural network, etc.;

[0071] (4) Define the loss function of the model, usually using mean square error, etc.;

[0072] (5) Use optimization algorithms (such as gradient descent) to learn and iteratively optimize model parameters;

[0073] (6) Evaluate the model performance on the validation set and perform hyperparameter tuning;

[0074] (7) Evaluate the final model on the test set to obtain the generalization performance.

[0075] 2. Establishment and training of grounding current calibration model

[0076] (1) Collect a large amount of standard grounding current data and corresponding raw grounding current data as training set, wherein the raw grounding current data is the current data collected directly through current transformer or other current detection device, and the standard grounding current data is the standard grounding current obtained after noise reduction and calibration processing of the raw grounding current data;

[0077] (2) Preprocess the training data, such as normalization, detrending, etc.;

[0078] (3) Select appropriate calibration model structure, such as autoregressive model, LSTM neural network or SVR, etc.;

[0079] (4) Define the loss function of the model, such as mean square error, smooth L1 loss, etc.;

[0080] (5) Use optimization algorithm to learn and iteratively optimize the model parameters;

[0081] (6) Evaluate the model performance on the validation set and perform hyperparameter tuning;

[0082] (7) Evaluate the final model on the test set to obtain the generalization performance.

[0083] 3. Establishment and training of core and clamp grounding judgment model

[0084] (1) Collect a large amount of standard grounding current data under normal grounding and non-grounding conditions, and label the corresponding state as training set;

[0085] (2) Preprocess the training data, such as normalization, noise reduction, etc.;

[0086] (3) Select appropriate classification model structure, such as artificial neural network, support vector machine or random forest, etc.;

[0087] (4) Define the loss function of the model, such as cross-entropy loss, etc.;

[0088] (5) Perform feature engineering to extract effective features from the grounding current sequence;

[0089] (6) Use optimization algorithm to learn and iteratively optimize the model parameters;

[0090] (7) Evaluate the model performance on the validation set and perform hyperparameter tuning;

[0091] (8) Evaluate the final model on the test set to obtain the generalization performance.

[0092] During the training process of the above model, some regularization techniques (such as L1 / L2 regularization, dropout, etc.) are usually needed to prevent overfitting. At the same time, some ensemble learning methods (such as Boosting, Bagging, etc.) can also be tried to further improve the model performance. It is worth noting that due to the quality and quantity of the actual data, sometimes semi-supervised learning, transfer learning, etc. strategies need to be used to make full use of existing data resources.

[0093] Regarding the grey wolf hunting + genetic algorithm used to optimize the model parameters, the specific implementation steps are as follows:

[0094] 1. Grey wolf hunting algorithm (GWO) part:

[0095] (1) Initialize a wolf population containing candidate solutions (i.e. model parameter values);

[0096] (2) Calculate the fitness value (such as mean square error) of each wolf (candidate solution);

[0097] (3) According to the fitness value, divide the individuals in the wolf population into four categories: α, β, δ and ω, representing the leaders, leader candidates, predators and cubs in the wolf population;

[0098] (4) Move the ω wolves (cubs) closer to the positions of α, β and δ, i.e. update the parameter values of ω, so that they gradually approach the optimal solution;

[0099] (5) Repeat (2) ~ (4) until the termination condition is met (such as reaching the maximum number of iterations or convergence condition).

[0100] 2. Genetic algorithm (GA) part:

[0101] (1) Take the optimal solution obtained by GWO as the initial population;

[0102] (2) Encode the individuals in the population (binary or real number encoding);

[0103] (3) Calculate the fitness value of each individual;

[0104] (4) According to the fitness value, select parent individuals with a certain probability;

[0105] (5) Perform crossover and mutation operations on the selected parent individuals to generate new offspring individuals;

[0106] (6) Replace some old individuals with offspring individuals to form a new generation population;

[0107] (7) Repeat (3) ~ (6) until the termination condition is met.

[0108] Since GWO has strong global search ability and GA is good at local search, combining the two can achieve better optimization results. Specifically, first use GWO to find a promising area in the solution space, then use GA to further search in detail in this area to find better solutions.

[0109] In practical applications, the parameters of GWO and GA need to be optimized to obtain better performance. For example, for GWO, the size of the wolf pack, step factor, etc. need to be set; for GA, the population size, crossover probability, mutation probability, etc. need to be set. At the same time, the appropriate switching time of the combination of the two also needs to be determined, for example, the maximum number of iterations of GWO can be set as the switching condition.

[0110] In addition, in order to improve computational efficiency, parallel computing and other techniques can be used to distribute model training and parameter optimization tasks to multiple computing nodes for simultaneous execution. During parameter optimization, the parameters of the algorithm can also be dynamically adjusted according to the real-time loss function value to speed up convergence.

[0111] Regarding the setting of thresholds, the following needs to be considered:

[0112] 1. Threshold in voltage noise reduction model, such as threshold function in wavelet transform noise reduction. Generally speaking, the threshold should not be too large or too small, otherwise it will cause signal distortion or noise residue. Common threshold functions include hard threshold, soft threshold, Gaussian threshold, etc. The value range of the threshold can refer to the empirical formula, such as λ = σ√(2lnN) (N is the signal length, σ is the noise standard deviation).

[0113] 2. Amplitude calibration threshold in grounding current calibration model. If the amplitude of the model output grounding current deviates greatly from the actual value, amplitude calibration is needed. A maximum allowable deviation threshold can be set, such as ±10%, when the deviation between the predicted value and the true value exceeds the threshold, the predicted value is linearly scaled and calibrated.

[0114] 3. Classification threshold in core and clamp grounding judgment model. Generally, 0.5 is used as the default threshold, that is, when the grounding probability output by the model is greater than 0.5, it is determined as a grounding state. But also according to the actual needs, the threshold can be adjusted appropriately to obtain higher accuracy or recall rate.

[0115] 4. In the data preprocessing stage, it may be necessary to set an abnormal value threshold, and data exceeding the threshold will be considered as an outlier and will be removed or corrected to avoid interference with the model. The value of the abnormal threshold can be based on the distribution characteristics of the data, and the mean ± 3 times the standard deviation is usually used as an empirical threshold.

[0116] 5. During model training, an early stopping threshold can also be set. Training stops when the loss function value on the validation set does not decrease or decreases below the threshold for several consecutive epochs, thus avoiding overfitting. This threshold is usually set to a small positive value, such as 0.0001.

[0117] It should be noted that most of the threshold values ​​mentioned above are empirical values. In practical applications, they need to be optimized based on specific data and scenarios to achieve the best results. Setting appropriate thresholds is crucial for improving model performance and generalization ability.

[0118] In summary, this invention proposes a novel method for detecting grounding current in transformer cores and clamps using dual current transformers, combined with machine learning models for noise reduction, calibration, and grounding status determination. Compared to traditional methods, this method offers higher anti-interference capabilities and detection accuracy, effectively addressing issues such as leakage flux interference and signal distortion, thus providing strong technical support for online monitoring and fault diagnosis of transformers.

[0119] Since the method of the present invention involves some calculation steps, in order to better implement the solution of the present invention, the specific implementation of the above method is described in more detail below using formulas or variables:

[0120] Step S10: Obtain the grounding current and leakage flux interference current of the core and clamps.

[0121] In this step, two current transformers are needed to obtain the total current on the core and the grounding wire of the clamp, and the interference current generated by leakage flux, respectively. Specifically:

[0122] A first current transformer is installed on the grounding wire of the transformer core and clamps, with its primary winding connected in series with the grounding wire. The current flowing through the grounding wire is the same as the grounding current Icore. r and leakage magnetic interference current I g The sum, i.e., I total =I r +I g Therefore, the first current transformer can obtain I total The current signal.

[0123] The primary winding of the second current transformer is positioned opposite to the primary winding of the first current transformer, with the secondary winding not closed, forming a loop antenna. The magnetic field strength at this location is primarily due to the interference current I generated by leakage flux. g This is caused by electromagnetic induction, therefore the second current transformer can detect I. g The current signal.

[0124] Using two current transformers, I can be obtained separately. total and Ig Two current signals, providing the basis data for the subsequent steps.

[0125] Step S20: Current signal converted into voltage signal

[0126] In this step, the current signal obtained in step S10 needs to be converted into the corresponding voltage signal for subsequent digital processing. Specifically:

[0127] The first current transformer detects I total The current signal is converted into the corresponding first voltage signal V1 through the precision resistor R1, and the conversion formula is:

[0128] V1 = I total · R1 = (I r + I g ) · R1

[0129] The second current transformer detects I g The voltage signal induced by the secondary winding of the current signal is output as the second voltage signal V2. According to the law of electromagnetic induction, the relationship between this voltage signal V2 and I g is:

[0130]

[0131] Where N is the number of turns of the secondary winding, Φ is the magnetic flux generated by I g , and L is the equivalent inductance of the secondary winding.

[0132] Through the above conversion, the voltage signals V1 and V2 corresponding to I total and I g are obtained, which are ready for subsequent digital processing.

[0133] Step S30: Digitalization of voltage signal

[0134] In this step, the analog voltage signals V1 and V2 converted in step S20 need to be digitized in order to be analyzed and processed using digital signal processing techniques in the subsequent steps. Specifically:

[0135] V1 is filtered, amplified, and analog-to-digital converted to obtain the first voltage digital sequence {V1[n]}.

[0136] The filter uses a notch filter or a Butterworth filter , etc., to remove high-frequency noise.

[0137] Amplification uses an operational amplifier circuit with a gain of G v , i.e., V′1 = G v V1.

[0138] Analog-to-digital conversion employs a SAR or sigma-delta (ΣΔ) analog-to-digital converter to sample and quantize the analog signal into a digital code V1[n].

[0139] V2 is subjected to similar filtering, amplification and analog-to-digital conversion processing as described above to obtain a second voltage digital sequence {V2[n]}.

[0140] After the above digital processing, the digital sequences {V1[n]} and {V2[n]} corresponding to V1 and V2 are obtained, laying a foundation for subsequent digital signal processing.

[0141] Step S40: Voltage signal denoising

[0142] In this step, the voltage digital sequences {V1[n]} and {V2[n]} obtained in step S30 need to be denoised to remove noise interference and improve signal quality. The following denoising model can be used:

[0143] Wavelet transform denoising model:

[0144] Wavelet transform is performed on {V1[n]} to obtain wavelet coefficients W1.

[0145] Threshold denoising is performed on the wavelet coefficients W1, and the threshold function is where N is the signal length, and σ is the noise standard deviation.

[0146] Wavelet inverse transform is performed on the denoised wavelet coefficients to obtain the first clean voltage sequence after denoising

[0147] {V2[n]} is subjected to similar wavelet transform denoising processing to obtain the second clean voltage sequence

[0148] Autoregressive filter model:

[0149] {V1[n]} is modeled as an autoregressive process where p is the model order, and w[n] is white noise.

[0150] Yule-Walker or Burg algorithm is used to estimate the autoregressive parameters {φ k}.

[0151] A linear prediction filter is constructed {V1[n]} is filtered to obtain the clean voltage sequence

[0152] {V2[n]} is subjected to similar autoregressive filtering processing to obtain

[0153] Adaptive linear neural network model:

[0154] Construct a neural network with adaptive linear filter structure.

[0155] Input {V1[n]} and {V2[n]} into the neural network for training, and the network weight is adaptively adjusted to minimize the residual sum of squares.

[0156] Use the trained neural network to perform predictive filtering on {V1[n]} and {V2[n]} to output the net voltage sequence and

[0157] Through the above noise reduction model, the noise components in {V1[n]} and {V2[n]} can be effectively removed, and a clearer net voltage sequence and Lay the foundation for subsequent current separation and calibration.

[0158] Step S50: Current separation

[0159] In this step, the net voltage sequence and obtained in step S40 is used to separate the actual core and clamp ground current sequence {I r [n]} is separated out. The specific method is as follows:

[0160] Align and on the time axis to ensure that their timestamps correspond one-to-one.

[0161] Subtract and point by point:

[0162]

[0163] According to the voltage conversion formula in step S20, we can get:

[0164]

[0165] Therefore, {V r [n]} actually represents a voltage sequence proportional to the actual core and clamp ground current {I r [n]}. By further scaling, we can get the standard ground current sequence {I r [n]}.

[0166] Step S60: Ground current calibration

[0167] In this step, the ground current sequence {V r [n]} obtained in step S50 needs to be further denoised and amplitude calibrated to obtain the standard ground current sequence The following model can be used:

[0168] Autoregressive (AR) model:

[0169] Model {V r [n]} as an autoregressive process Estimate the autoregressive parameters {φ k} using Yule-Walker or Burg algorithm.

[0170] Construct a linear prediction filter Filter {V r [n]} to obtain the net current sequence

[0171] According to the amplitude range of the actual ground current, scale to obtain the standard sequence

[0172] Long Short-Term Memory Neural Network (LSTM) model:

[0173] Build a recurrent neural network with LSTM units, input {V r [n]} and output predicted standard current

[0174] Train the LSTM network on the training set so that it can complete both denoising and amplitude calibration tasks.

[0175] Evaluate the performance of the LSTM model on the test set to obtain the final standard ground current sequence

[0176] Support Vector Machine (SVM) regression model:

[0177] Map {V r [n]} to a high-dimensional feature space and build an SVM regression model to fit the nonlinear mapping relationship of the standard ground current.

[0178] Train the SVM regression model on the training set and use the kernel trick to simplify the calculation.

[0179] Evaluate the performance of the model on the test set and output the standard ground current sequence

[0180] The above model can denoise and calibrate the grounding current sequence while adaptively adjusting the model parameters according to the real-time characteristics of the input sequence, so that the model can better fit the characteristics of the actual grounding current and improve the accuracy of denoising and calibration.

[0181] Step S70: Core and clamp grounding state judgment

[0182] In this step, the standard grounding current sequence obtained continuously needs to be used to Judge whether the transformer core and clamp are currently in the grounding state. The following model can be used:

[0183] Artificial neural network (ANN) model:

[0184] A multi-layer feedforward neural network is constructed, and the input is the grounding current sequence in a period of time For example, the last 1000 time steps.

[0185] Feature extraction is performed on the input sequence to generate a feature vector x, including statistical features such as mean, variance, and peak factor.

[0186] The feature vector x is input into the neural network, and the output is a binary classification result y ∈ {0, 1}, where 0 represents non-grounding and 1 represents grounding.

[0187] The neural network is trained on the labeled training set, and the cross-entropy loss function Where p i is the grounding probability of the network output.

[0188] The model is evaluated on the test set, and the final grounding judgment result is output according to the set classification threshold (such as 0.5).

[0189] Support vector machine (SVM) model:

[0190] Map the input Sequence to a high-dimensional feature space to construct an SVM classifier.

[0191] The SVM model is trained on the training set to automatically learn the optimal classification hyperplane to separate the non-grounding and grounding samples.

[0192] Introduce kernel trick to simplify high-dimensional mapping calculation, use Gaussian kernel or polynomial kernel, etc.

[0193] Evaluate the SVM model on the test set, and output the current grounding state according to the distance or probability to the classification hyperplane.

[0194] Random forest (RF) model:

[0195] A random forest classifier is constructed by building a decision tree ensemble, each decision tree is trained on a different subset of features

[0196] Statistical features, wavelet features, etc. are extracted from the input sequence to form a feature vector x.

[0197] The feature vector x is input into the random forest model, and each decision tree independently generates a classification result, and then a majority vote is performed to obtain the final classification

[0198] The random forest model is trained on the labeled training set, and the optimal decision tree parameters are automatically determined by the Bayesian risk minimization criterion.

[0199] The discriminant performance of the model is evaluated on the test set, and the final grounding judgment result is output according to the set confidence threshold.

[0200] In the implementation of the present application, each step can be flexibly combined to fully utilize existing machine learning algorithms and signal processing techniques. Some optional implementation modes are given below:

[0201] For the noise reduction model of step S40, wavelet transform and autoregressive filtering are combined. First, wavelet transform is used to remove high-frequency noise in {V1[n]} and {V2[n]}, and then autoregressive filtering is used to further remove residual low-frequency noise.

[0202] For the current calibration model of step S60, an autoregressive model can be used to preliminarily reduce noise in {V r [n]}, and then the noise-reduced sequence is input into the LSTM network for fine calibration.

[0203] For the grounding state judgment model of step S70, a model integration framework can be constructed to combine the outputs of ANN, SVM and RF models, and the final judgment result is obtained through weight weighting.

[0204] For actual working conditions with continuously changing noise levels, an adaptive filter is introduced to enable the parameters of the noise reduction model to adaptively adjust with changes in noise statistical characteristics.

[0205] A transfer learning strategy is adopted to pre-train the model using historical data of other transformers, and then fine-tune it with a small amount of data of the current transformer to alleviate the problem of lack of training data.

[0206] During model training, regularization techniques such as L1 / L2 regularization, dropout, etc. are introduced to prevent overfitting, and appropriate learning rate scheduling strategies are adopted to speed up convergence.

[0207] ​Through the above extended embodiments, the applicability and detection accuracy of the application in actual engineering can be further improved. The application utilizes advanced machine learning and signal processing technology to realize intelligent monitoring of the grounding state of the transformer core and the clamp.

[0208] It should be noted that the variables, subscripts and constants involved in the above description are explained as follows:

[0209]

[0210]

[0211] Specifically, the principle of the application is that due to the uncertainty of the leakage magnetic field around the running transformer and the dispersion of the test sample, it is difficult to correct the grounding current data, and the conventional electromagnetic shielding method cannot meet the test requirements of the grounding current. Therefore, by changing the structure of the test current transformer, a same coil is connected after the original current transformer coil (a method of balancing the interference current by a double-coil current transformer), the induced currents in the two coils are superimposed in opposite directions, the interference caused by the leakage magnetic field is balanced by the opposite currents, and the real situation of the core and clamp grounding current can be objectively reflected.

[0212] The core and clamp grounding current is generally composed of the actual core and clamp downlead current Ir and the interference current Ig caused by the leakage magnetic field in the measurement. According to the different structures and operating conditions, the leakage magnetic field around the transformer has great randomness, so that the interference current Ig has randomness. At the same time, due to the anisotropy of the leakage magnetic field, the direction of the interference current Ig in the measurement process is also random, and even the same current transformer cannot eliminate the influence of the leakage magnetic field in the value. Among them, the actual measurement value I = Ir + Ig, in the formula, Ir is the actual core and clamp downlead current, and Ig is the interference current caused by the leakage magnetic field. The single-coil measurement wiring diagram and current vector diagram are shown in Figure 3 .

[0213] Figure 4 The double-coil measurement wiring diagram and current vector diagram are shown. After adopting the double-coil current transformer balancing interference current measurement method, the actual measurement value I3 = I1 + I2 + I4, I1 is the real core and clamp downlead current Ir, I2 is the interference current Ig caused by the leakage magnetic field, and I4 is the balancing current -Ig generated by the second completely same coil of the current transformer. Since the positions and directions of the current transformers for measuring currents I2 and I4 are almost the same, and the polarities are opposite, I2 ≈ -I4, the actual measurement value I3 ≈ I1, and the influence of the leakage magnetic field is eliminated on the circuit.

[0214] In addition, the application also relates to the principles of the following specific technologies:

[0215] 1. Adaptive current separation principle:

[0216] Due to the interference of leakage magnetic current in the transformer, it is necessary to decompose the measured current signal into two parts of actual core and clamp grounding current and leakage magnetic current. By reasonably designing the circuit structure and operation model, two voltage signals proportional to the two components can be obtained. By using adaptive signal processing technology (such as adaptive filtering, adaptive wavelet transform, etc.), the core and clamp grounding current component can be automatically separated.

[0217] 2. Intelligent current calibration principle:

[0218] Due to the existence of circuit parameter drift, quantization noise and other factors, the separated core and clamp grounding current may have deviation in amplitude from the actual value. A fine current calibration model can be established by machine learning technology, such as long short-term memory neural network (LSTM), support vector machine regression (SVR), etc. The separated current is input into the model, and the standardized current sequence after noise reduction and amplitude calibration is output.

[0219] 3. Pattern recognition principle:

[0220] After obtaining the standardized grounding current sequence, intelligent pattern recognition algorithm can be applied to judge the grounding state. According to the statistical characteristics and time-frequency characteristics of the current sequence, artificial neural network (ANN), support vector machine (SVM) or random forest (RF) classifier is constructed to distinguish the grounding / ungrounding two patterns. Through model automatic learning of pattern characteristics in training data, the current grounding state of the transformer can be accurately identified.

[0221] In summary, the core principle of the method is to use the double current method to filter out the spatial leakage magnetic interference, realize the accurate monitoring of grounding current in strong electromagnetic interference environment, and combine machine learning with advanced signal processing technology to adaptively analyze and process the original electrical measurement signal, so as to intelligently judge whether the transformer core and clamp have grounded, and accurately evaluate the grounding condition. Compared with the traditional manual detection and measurement method, the method has the advantages of automation, intelligence and real-time, and can efficiently realize the online monitoring of transformer core and clamp grounding fault.

[0222] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any skilled person 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.

Claims

1. A method of measuring transformer core and clamp ground dual current, characterized by, The method comprises the following steps: S10, using a first current transformer and a second current transformer to monitor the grounding current of the transformer core and the clamp, wherein the first current transformer monitors the current on the core and clamp grounding line of the transformer core and clamp, and the second current transformer only monitors the interference current generated by the leakage magnetic field of the transformer core and clamp; S20, converting the current signals of the first current and the interference current detected by the first current transformer and the second current transformer into corresponding voltage signals respectively; S30, preprocessing each voltage signal obtained by conversion, including filtering, amplifying and analog-to-digital conversion, to obtain corresponding voltage digital signals, which are respectively recorded as first voltage sequence and second voltage sequence; S40, using a pre-trained voltage denoising model to denoise the first voltage sequence and the second voltage sequence respectively to obtain first and second denoised voltage sequences; S50, aligning the first and second denoised voltage sequences and calculating the difference to obtain a numerical sequence representing the actual core and clamp grounding current, recorded as grounding current sequence; S60, using a pre-trained grounding current calibration model to denoise and calibrate the grounding current sequence to obtain a standard grounding current sequence; wherein the grounding current calibration model uses optimized model parameters of the grounding current calibration model according to the grounding current sequence in the process of denoising and calibrating the grounding current sequence; the method used for optimizing model parameters is: grey wolf hunting algorithm + genetic algorithm; find a promising area in the solution space by using grey wolf hunting algorithm GWO, and then use genetic algorithm GA to further search in the area to find a better solution; S70, repeatedly executing step S60 to obtain multiple standard grounding current sequences, and using a pre-trained core and clamp grounding judgment model to receive the input standard grounding current sequence to obtain a judgment result of whether the transformer core and clamp are grounded.

2. The method of claim 1, wherein, The specific steps of S10 include: arranging a first current transformer on the core and clamp grounding line of the transformer, connecting the primary winding of the first current transformer to the core and clamp grounding line, and obtaining the current signal of the first current flowing through the core and clamp grounding line by electromagnetic induction principle; The primary winding of the second current transformer is arranged opposite to the primary winding of the first current transformer, and is used to obtain the current signal of the interference current generated by the leakage magnetic field.

3. The method of claim 1, wherein the transformer core and clamp ground dual current measurement method is characterized by, The specific steps of S30 include: filtering, amplifying and analog-to-digital converting the voltage signals corresponding to the first current transformer and the second current transformer respectively to obtain first voltage sequence and second voltage sequence.

4. The method of claim 1, wherein, The steps of establishing and training the voltage denoising model include: Step 1, collect a large amount of standard voltage data and corresponding noisy voltage data as first training data; Step 2, pre-process the first training data by normalization; Step 3, select BP neural network as the model structure; Step 4, the BP neural network is trained by using the pretreated first training data, gradient descent is used as an optimization algorithm to iteratively update the model parameters, and the voltage denoising model is obtained.

5. The method of claim 1, wherein the transformer core and clamp ground dual current measurement method is characterized by, The specific steps of the S50 include: aligning the first net voltage sequence and the second net voltage sequence obtained in the step S40 in the time axis; performing point-by-point subtraction operation on the first net voltage sequence and the second net voltage sequence to obtain the ground current sequence representing the actual core and clamp ground current.

6. The method of claim 1, wherein, The establishment and training steps of the ground current calibration model include: Step 1, collect standard ground current data and original ground current data as second training data; Step 2, pre-process the second training data by normalization and detrending; Step 3, select a long short-term memory neural network as the model structure; Step 4, train the long short-term memory neural network by using the pretreated second training data, and iteratively update the model parameters by using gradient descent as an optimization algorithm to obtain the ground current calibration model.

7. The method of claim 1, wherein the transformer core and clamp ground dual current measurement method is characterized by, The establishment and training steps of the core and clamp grounding judgment model include: Step 1, collect standard ground current data under normal grounding and non-grounding conditions as third training data, and label the grounding state as training set label; Step 2, pre-process the third training data by normalization and denoising; Step 3, select a support vector machine model as the model structure; Step 4, train the support vector machine model by using the pretreated third training data, and iteratively update the model parameters by using gradient descent as an optimization algorithm to obtain the core and clamp grounding judgment model.

8. The method of claim 4 or 6 or 7, wherein the transformer core and clamp grounding dual current measurement method is characterized by, When training the model, regularization techniques and learning rate scheduling strategies are introduced to alleviate the overfitting problem and speed up the convergence speed.

9. The method of claim 1, wherein, When judging the grounding state of the standard current sequence in the step S70, a model integration framework is constructed, specifically: three models of artificial neural network, support vector machine and random forest are used to classify the standard ground current sequence at the same time; the output results of the three models are weighted and fused to obtain the final grounding judgment result.

Images