Automatic intelligent compensation method and system for current transformer

Abstract

The present invention relates to the technical field of current transformers, and in particular to an automatic intelligent compensation method and system for current transformers. The method comprises inputting the operating parameters of the current transformer into a first machine learning model, outputting saturation, obtaining specific features, inputting temperature data, saturation, and specific features into a second machine learning model, outputting a current compensation amount, and performing current compensation on a compensation winding within the current transformer body. By real-time monitoring and predicting changes in the saturation of the current transformer, the compensation current can be adjusted more accurately, thereby avoiding measurement errors caused by changes in saturation and improving the accuracy and stability of the current transformer in different working environments. By performing correlation analysis on feature data and removing redundant features, the data input into the machine learning model is ensured to be more accurate and efficient, which not only improves the prediction accuracy but also reduces the computational complexity and improves the system operation efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of current transformers, and in particular to an automatic intelligent compensation method and system for current transformers. Background Art

[0002] In power systems, current transformers (CTs) are crucial measurement and protection devices, and their accuracy directly impacts the safe and stable operation of the system. Traditional CT compensation methods rely primarily on manual calibration and hardware adjustments, such as adjusting the transformer's internal core structure or using additional compensation coils to reduce errors. In recent years, with the rapid development of electronic technology and intelligent algorithms, microprocessor-controlled compensation systems have emerged that monitor the CT output in real time and apply preset compensation algorithms to make adjustments.

[0003] Chinese patent application publication number CN112816754A discloses a current transformer current compensation method and device. The method detects the primary current amplitude of the current transformer and then uses a preset current compensation model to process the primary current amplitude and transformation ratio to obtain the true theoretical secondary current amplitude of the current transformer. The secondary current amplitude is an empirical value obtained by learning from historical data. Since the influence of the primary current amplitude is taken into account, it is closer to the actual secondary current amplitude. Therefore, selecting the range of the detection device according to this value is not likely to damage the detection device.

[0004] As mentioned in the above application, the existing intelligent compensation system has improved the accuracy of the current transformer to a certain extent, but it still has the following shortcomings: most systems rely on pre-set compensation parameters and are difficult to adapt to changes in transformer characteristics caused by factors such as ambient temperature and load changes during actual operation. Secondly, the compensation algorithm used is relatively simple and fails to fully utilize the advantages of modern signal processing technology and machine learning algorithms, which limits the further improvement of compensation accuracy. Summary of the Invention

[0005] In order to solve the above problems, the present invention provides an automatic intelligent compensation method and system for a current transformer.

[0006] The present invention adopts the following technical solution, an automatic intelligent compensation method for a current transformer, comprising the following steps:

[0007] Real-time collection of operating data of the current transformer, wherein the operating data of the current transformer includes temperature data and operating parameters;

[0008] Inputting the working parameters into a pre-built first machine learning model for predicting saturation, and outputting the saturation;

[0009] Processing the working parameters, extracting feature data, establishing a time series feature data set and a numerical feature data set, performing correlation analysis on features in the time series feature data set and the numerical feature data set, removing redundant features, and marking the remaining features as specific features;

[0010] The temperature data, saturation and specific features are input into a pre-built second machine learning model for predicting current compensation, and the current compensation amount is output. Current compensation is performed on the compensation winding in the current transformer body based on the output current compensation amount.

[0011] As a further description of the above technical solution: the operating parameters include input current, output current, output voltage, input current frequency and load characteristics, the load characteristics include continuous load, impact load or instantaneous load, and the load characteristics are obtained through a load sensor or detecting the current waveform.

[0012] As a further description of the above technical solution: the method of processing the working parameters, extracting feature data, and establishing a time series feature data set and a numerical feature data set includes:

[0013] Each time the working parameters are collected, a timestamp is recorded for each sampling point, and the collected input current, output current, output voltage, and input current frequency are associated with their timestamps and stored in a time series format to generate time series data;

[0014] Calculate the average value, standard deviation, and root mean square value of the input current, output current, and output voltage to obtain time domain characteristics;

[0015] The input current, output current and output voltage are converted from the time domain to the frequency domain through fast Fourier transform, the frequency components in the signal are extracted, the power distribution of the signal at different frequencies is calculated to obtain the frequency components and power spectrum density, and the frequency domain characteristics are obtained;

[0016] Calculate the ratio of the current transformer input current to the rated current, as well as the ratio of the output current to the input current, to obtain instantaneous characteristics;

[0017] Establishing a time series feature data set based on the acquired input current, output current, output voltage, and input current frequency, wherein the time series feature data set stores time series data;

[0018] A numerical feature data set is established using the acquired time domain features, frequency domain features, and instantaneous features. The numerical feature data set stores numerical data.

[0019] As a further description of the above technical solution: a method for performing correlation analysis on features in a time series feature data set and removing redundant features includes:

[0020] Calculate the mutual information value between any two features in the time series feature data set;

[0021] The calculation formula of the mutual information value is:

[0022]

[0023] Where I(X;Y) is the mutual information between feature X and feature Y, p(x,y) is the joint probability of feature X and feature Y taking the values ​​x and y, and p(x) and p(y) are the marginal probabilities of feature X and feature Y.

[0024] By calculating the mutual information value between all feature pairs, a mutual information matrix is ​​constructed. Each element I(X;Y) in the matrix represents the mutual information value between feature X and feature Y.

[0025] Example: For a dataset M with n features, the mutual information matrix is ​​an n×n matrix, where each element I(n1;n2) represents the mutual information value between features n1 and n2;

[0026] The mutual information matrix is ​​used to evaluate the correlation between each feature and other features, and redundant features are removed based on the correlation between two features.

[0027] As a further description of the above technical solution: the method for removing redundant features based on the correlation between two features includes:

[0028] A correlation threshold is preset. When the mutual information value between two features is greater than or equal to the preset correlation threshold, there is a high degree of redundancy between the two features, and one of them is randomly removed. Conversely, when the mutual information value between the two features is less than the preset correlation threshold, both features are retained.

[0029] As a further description of the above technical solution: a method for performing correlation analysis on features in a numerical feature data set, removing redundant features, and marking the remaining features as specific features includes: removing redundant features by obtaining correlation coefficients for the features in the numerical feature data set, and applying Lasso regression to perform feature importance evaluation on the remaining feature data, retaining features with evaluation values ​​not zero, and marking the retained features as specific features.

[0030] As a further description of the above technical solution: a method for eliminating redundant features by obtaining correlation coefficients for features in the numerical feature data set, and applying Lasso regression to the remaining feature data to perform feature importance assessment includes:

[0031] Normalize the features in the numerical feature data set and scale the values ​​to the range of [0,1];

[0032] Calculate the correlation coefficient between every two features in the numerical feature data set and generate a correlation coefficient matrix;

[0033] The calculation formula of the correlation coefficient is:

[0034]

[0035] In the formula, r(X i ,X j ) is feature X i and feature X j Correlation coefficient of and is the mean of the two features, n is the number of features in the numerical feature data set, X ik and X jk For the kth sample in feature X i and X j The value of ;

[0036] A correlation coefficient threshold is preset. If the absolute value of the correlation coefficient between two features is greater than the preset correlation coefficient threshold, one feature is retained and the other redundant feature is deleted. This process is repeated until the correlation coefficients between the remaining features are all lower than the correlation coefficient threshold.

[0037] Obtaining a numerical feature data set after eliminating redundant features and a target variable, wherein the target variable is a current compensation amount;

[0038] Select the Lasso regression model, the objective function of the Lasso is:

[0039]

[0040] In the formula, α is the regularization parameter, which controls the degree of feature sparsity, and β j is the regression coefficient of the jth feature, y i is the true value of the i-th sample, n is the number of samples, p is the number of features, is the predicted value of the i-th sample.

[0041] Use cross-validation to select the best α value;

[0042] Get the regression coefficient β of the Lasso model j :

[0043] If β j =0, indicating feature X j Considered unimportant, feature X is deleted j ;

[0044] If β j ≠0, the larger its absolute value is, the more characteristic X jThe more significant the impact on the target variable, the corresponding features are retained and marked as specific features.

[0045] As a further description of the above technical solution: the method for constructing the second machine learning model includes:

[0046] Initialize the second machine learning model structure. The second machine learning model structure adopts an MLP type multi-layer forward network structure with three input layers, two hidden layers, and one output layer. The input layer includes the first input layer, the second input layer, and the third input layer. The number of nodes in the first input layer is 1, corresponding to temperature data. The number of nodes in the second input layer is 1, corresponding to saturation. The nodes in the third input layer are [1, c], which are specific features, and c is the number of specific features. The hidden layer includes the first hidden layer and the second hidden layer. The number of nodes in the first hidden layer is 128, and the Relu function is used as the activation function. The number of nodes in the second hidden layer is 64, and the Relu function is used as the activation function. The output layer is the first output layer. The number of nodes in the first output layer is 1, corresponding to the predicted current compensation amount.

[0047] After initializing the second machine learning model structure, the temperature data, saturation and specific features, as well as the current compensation corresponding to the temperature data, saturation and specific features, are used to train the second machine learning model. The temperature data, saturation and specific features, as well as the current compensation corresponding to the temperature data, saturation and specific features are obtained from a database. The database records the current compensation amount of each current transformer under normal working conditions and is used to optimize the model. The optimizer is Adam, the loss function is the MSE loss function, the batch size is set to 32, the number of iterations is 200 rounds, and the training ends when the loss function converges.

[0048] As a further description of the above technical solution: the input current is the actual current flowing through the current transformer, which is directly measured by the current sensor or the primary side of the current transformer; the output current is the secondary current output by the current transformer, which is measured by the secondary side of the current transformer; the output voltage is the secondary terminal voltage of the current transformer, which is collected and obtained by the voltage sensor; the input current frequency is obtained by measuring the frequency of the input current with a frequency meter or a digital oscilloscope.

[0049] An automatic intelligent compensation system for a current transformer, used to implement the automatic intelligent compensation method for a current transformer, comprises:

[0050] A data acquisition module collects the working data of the current transformer in real time, wherein the working data of the current transformer includes temperature data and working parameters;

[0051] a data analysis module, inputting the working parameters into a pre-built first machine learning model for predicting saturation, and outputting the saturation;

[0052] a feature extraction module, processing the working parameters, extracting feature data, establishing a time series feature data set and a numerical feature data set, performing correlation analysis on features within the time series feature data set and the numerical feature data set, removing redundant features, and marking the remaining features as specific features;

[0053] The data processing module inputs the temperature data, saturation and specific features into a pre-built second machine learning model for predicting the current compensation amount, outputs the current compensation amount, and performs current compensation on the compensation winding in the current transformer body based on the output current compensation amount.

[0054] Beneficial effects:

[0055] The automatic intelligent compensation method and system for a current transformer provided by the present invention collect the working parameters of the current transformer in real time, predict the saturation based on the working parameters of the current transformer, then extract features of the working parameters of the current transformer, establish a feature set, perform correlation analysis on the features in the feature set, remove redundant features, and then input the temperature data, saturation and feature data into a pre-built second machine learning model for predicting the current compensation amount, output the current compensation amount, and perform current compensation on the compensation winding in the current transformer body based on the output current compensation amount, so that the automatic intelligent compensation method of the current transformer can more accurately adjust the compensation current by real-time monitoring and predicting the change of the saturation of the current transformer, thereby avoiding measurement errors caused by saturation changes and improving the accuracy of the current transformer in the absence of Accuracy and stability under the same working environment. Secondly, by performing correlation analysis on the feature data and removing redundant features, the data input into the machine learning model is ensured to be more accurate and efficient, which not only improves the prediction accuracy, but also reduces the computational complexity and improves the system operation efficiency. Finally, by collecting the working parameters of the current transformer in real time, it can quickly respond to environmental changes and accurately adjust the compensation amount. Under the influence of temperature changes or other external factors, it ensures that the current transformer can always maintain the best working state, avoids performance degradation caused by unstable working parameters, and overcomes the problem that existing technologies are difficult to adapt to the changes in transformer characteristics caused by factors such as ambient temperature and load changes in actual operation, and the compensation algorithm used is relatively simple, and fails to fully utilize the advantages of modern signal processing technology and machine learning algorithms, which limits the compensation accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The present invention will be further explained below in conjunction with the accompanying drawings and examples:

[0057] Figure 1 A schematic flow chart of an automatic intelligent compensation method for a current transformer provided by an embodiment of the present invention;

[0058] Figure 2A flow chart of a method for processing the working parameters, extracting feature data, and establishing a time series feature data set and a numerical feature data set provided by an embodiment of the present invention;

[0059] Figure 3 This is a module connection diagram of the automatic intelligent compensation system for current transformers provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0060] In order to make the technical means, creative features, objectives and effects of the present invention easier to understand, the present invention is further described below with reference to specific diagrams. It should be noted that the embodiments and features in the embodiments of this application can be combined with each other unless they conflict.

[0061] Example 1

[0062] See also Figure 1 、 Figure 2 The embodiment of the present invention provides a technical solution: an automatic intelligent compensation method for a current transformer, applied to a current transformer body, comprising the following steps:

[0063] Real-time collection of operating data of the current transformer, wherein the operating data of the current transformer includes temperature data and operating parameters;

[0064] It should be noted that the ambient temperature affects the magnetic properties, saturation characteristics and output signal of the current transformer. When the current transformer works at different ambient temperatures, its current compensation amount is different.

[0065] The operating parameters include input current, output current, output voltage, input current frequency and load characteristics;

[0066] The load characteristics include continuous load, impact load or instantaneous load, and the load characteristics are obtained through a load sensor or detecting a current waveform.

[0067] It should be noted that the input current is the actual current flowing through the current transformer. The current transformer generates an output signal based on this current, which is directly measured through the primary side of the current sensor or current transformer.

[0068] The output current is the secondary current output by the current transformer, and the output current is measured through the secondary side of the current transformer;

[0069] The primary side is the high current side of the current transformer, which is directly connected in series with the circuit to be measured. The secondary side is the low current side of the current transformer, which is connected to the measuring instrument or protection device.

[0070] The output voltage is the voltage at the secondary end of the current transformer. The output current generates a voltage through the resistive load, which is collected by the voltage sensor.

[0071] Input current frequency: Measure the frequency of the input current using a frequency meter or digital oscilloscope;

[0072] Inputting the working parameters into a pre-built first machine learning model for predicting saturation, and outputting the saturation;

[0073] It should be noted that saturation refers to the degree to which the core of a current transformer reaches magnetic saturation under high current. At this point, the core's magnetic flux density no longer increases with increasing input current, causing the proportional relationship between the current transformer's output signal and the actual current to be distorted.

[0074] The construction of the first machine learning model is based on the LSTM network structure. The LSTM network structure includes an input layer, two LSTM layers; an output layer, the input of the input layer is a working parameter, a set of input working parameters and the corresponding saturation is a sample, the sample batch of the input layer is 128, and the dimension is 4; the two LSTM layers include the first LSTM layer and the second LSTM layer, and the two LSTM layers each include 4 components, including an input gate, a forget gate, an output gate and a memory unit; the activation function of the two LSTM layers is tanh; the input dimension of the first LSTM layer is 4, and the output dimension is 128, the input dimension of the second LSTM layer is 128, and the output dimension is 64; the input dimension of the output layer is 64, and the output dimension is 3, the activation function is softmax, and the loss function is cross entropy.

[0075] The training method of the first machine learning model includes:

[0076] A set of working parameters and the corresponding saturation are combined into a set of historical training sets. G sets of historical training sets are extracted from the database as the training set to train the first machine learning model. The Adam optimizer is used, the batch size is set to 128, the initial learning rate is set to 0.001, and the sliding average method is used to update the parameter momentum and the input and output parameter matrix w. The update equation is:

[0077] vf=momentum×(vf-1)+(1-momentum)×Δwf;

[0078] wf+1=wf-η×vf;

[0079] Where f represents the sequence number of the current iteration step under the sliding average, vf represents the estimated gradient of the current iteration step f under the sliding average, (vf-1) represents the estimated gradient of the previous step, Δwf represents the actual gradient descent amount of this step, wf represents the parameter value of the current iteration step f, η is the learning rate that controls the step size, and wf+1 represents the updated parameter value, that is, the learning rate η is used to update wf in the opposite direction of vf;

[0080] It is worth noting that the parameter momentum determines the weights of (vf-1) and Δwf. That is, the higher the momentum, the more vf depends on (vf-1) in the previous step, and the stronger the smoothing effect. The lower the momentum, the more vf depends on the current gradient Δwf, and the stronger the ability to track the actual gradient changes.

[0081] The calculation formula of the loss function loos is:

[0082]

[0083] Where m is the serial number of a single historical training set, h m ∈[0, 1], h m is the probability distribution of the true label of the mth group of historical training sets, p m is the probability distribution of the mth group of historical training sets after prediction by the first machine learning model; m≤G; G is the number of historical training sets;

[0084] Special note:

[0085] h m＝1 When the term is maximum, p is required m Maximum, when the prediction is correct, loos is minimum;

[0086] h m＝0 The time term is the largest, requiring 1-p m Maximum, that is, when the prediction error is large, loos is the largest;

[0087] When loos converges, the training is stopped and the training is completed.

[0088] Processing the working parameters, extracting feature data, establishing a time series feature data set and a numerical feature data set, performing correlation analysis on features in the time series feature data set and the numerical feature data set, removing redundant features, and marking the remaining features as specific features;

[0089] The method of processing the working parameters, extracting feature data, and establishing a time series feature data set and a numerical feature data set includes:

[0090] When collecting working parameters, an accurate timestamp is recorded for each sampling point, and the collected input current, output current, output voltage, and input current frequency are associated with their timestamps and stored in a time series format to generate time series data;

[0091] Calculate the average value, standard deviation, and root mean square value of the input current, output current, and output voltage to obtain time domain characteristics;

[0092] It should be noted that the calculation formulas for the average value, standard deviation and root mean square value are all existing technologies and will not be repeated here. The standard deviation is used to measure the degree of signal fluctuation, and the root mean square value is used to represent the energy of the current waveform.

[0093] The input current, output current and output voltage are converted from the time domain to the frequency domain through fast Fourier transform, the frequency components in the signal are extracted, the power distribution of the signal at different frequencies is calculated to obtain the frequency components and power spectrum density, and the frequency domain characteristics are obtained;

[0094] Calculate the ratio of the current transformer input current to the rated current, as well as the ratio of the output current to the input current, to obtain instantaneous characteristics;

[0095] Establishing a time series feature data set based on the acquired input current, output current, output voltage, and input current frequency, wherein the time series feature data set stores time series data;

[0096] Establishing a numerical feature data set based on the acquired time domain features, frequency domain features and instantaneous features, wherein the numerical feature data set stores numerical data;

[0097] Methods for performing correlation analysis on features within a time series feature data set and removing redundant features include:

[0098] Calculate the mutual information value between any two features in the time series feature data set;

[0099] The calculation formula of the mutual information value is:

[0100]

[0101] Where I(X;Y) is the mutual information between feature X and feature Y, p(x,y) is the joint probability of feature X and feature Y taking the values ​​x and y, and p(x) and p(y) are the marginal probabilities of feature X and feature Y.

[0102] By calculating the mutual information value between all feature pairs, a mutual information matrix is ​​constructed. Each element I(X;Y) in the matrix represents the mutual information value between feature X and feature Y.

[0103] Example: For a dataset M with n features, the mutual information matrix is ​​an n×n matrix, where each element I(n1;n2) represents the mutual information value between features n1 and n2.

[0104] The mutual information matrix is ​​used to evaluate the correlation between each feature and other features, and redundant features are removed based on the correlation between two features.

[0105] Based on the correlation between two features, methods for removing redundant features include:

[0106] A correlation threshold is preset. When the mutual information value between two features is greater than or equal to the preset correlation threshold, there is a high degree of redundancy between the two features, and one of them is randomly removed. Conversely, when the mutual information value between the two features is less than the preset correlation threshold, both features are retained.

[0107] The method of performing correlation analysis on features in a numerical feature data set, removing redundant features, and marking the remaining features as specific features includes: removing redundant features by obtaining correlation coefficients for the features in the numerical feature data set, applying Lasso regression to perform feature importance evaluation on the remaining feature data, retaining features with evaluation values ​​not zero, and marking the retained features as specific features.

[0108] The method of eliminating redundant features by obtaining correlation coefficients for features in the numerical feature data set and applying Lasso regression to the remaining feature data to perform feature importance assessment includes:

[0109] Normalize the features in the numerical feature data set and scale the values ​​to the range of [0,1];

[0110] Calculate the correlation coefficient between every two features in the numerical feature data set and generate a correlation coefficient matrix;

[0111] The calculation formula of the correlation coefficient is:

[0112]

[0113] In the formula, r(X i ,X j ) is the feature X i and feature X j Correlation coefficient of and is the mean of the two features, n is the number of features in the numerical feature data set, X ik and X jk For the kth sample in feature X i and X j The value on .

[0114] A correlation coefficient threshold is preset. If the absolute value of the correlation coefficient between two features is greater than the preset correlation coefficient threshold, one feature is retained and the other redundant feature is deleted. This process is repeated until the correlation coefficients between the remaining features are all lower than the correlation coefficient threshold.

[0115] Obtaining a numerical feature data set after eliminating redundant features and a target variable, wherein the target variable is a current compensation amount;

[0116] Select the Lasso regression model, the objective function of the Lasso is:

[0117]

[0118] In the formula, α is the regularization parameter, which controls the degree of feature sparsity, and β j is the regression coefficient of the jth feature, y i is the true value of the i-th sample, n is the number of samples, p is the number of features, is the predicted value of the i-th sample.

[0119] Use cross-validation to select the best α value;

[0120] Get the regression coefficient β of the Lasso model j :

[0121] If β j =0, indicating feature X j Considered unimportant, feature X is deleted j ;

[0122] If β j ≠0, the larger its absolute value is, the more characteristic X j The more significant the impact on the target variable, the corresponding features are retained and marked as specific features.

[0123] Methods for selecting the best value of α using cross-validation include:

[0124] Define a range of candidate values ​​for α;

[0125] Perform cross validation for each α value and calculate the mean square error of the validation set;

[0126] Choose the value of α that minimizes the mean squared error.

[0127] The temperature data, saturation and specific features are input into a pre-built second machine learning model for predicting current compensation, and the current compensation amount is output. Current compensation is performed on the compensation winding in the current transformer body based on the output current compensation amount.

[0128] The method for constructing the second machine learning model includes:

[0129] Initialize the second machine learning model structure. The second machine learning model structure adopts an MLP type multi-layer forward network structure with three input layers, two hidden layers, and one output layer. The input layer includes the first input layer, the second input layer, and the third input layer. The number of nodes in the first input layer is 1, corresponding to temperature data. The number of nodes in the second input layer is 1, corresponding to saturation. The nodes in the third input layer are [1, c], which are specific features, and c is the number of specific features. The hidden layer includes the first hidden layer and the second hidden layer. The number of nodes in the first hidden layer is 128, and the Relu function is used as the activation function. The number of nodes in the second hidden layer is 64, and the Relu function is used as the activation function. The output layer is the first output layer. The number of nodes in the first output layer is 1, corresponding to the predicted current compensation amount.

[0130] After initializing the second machine learning model structure, the temperature data, saturation and specific features, as well as the current compensation corresponding to the temperature data, saturation and specific features, are used to train the second machine learning model. The temperature data, saturation and specific features, as well as the current compensation corresponding to the temperature data, saturation and specific features are obtained from the database. The database records the current compensation amount of each current transformer under normal working conditions and is used to optimize the model. The optimizer is Adam, the loss function is the MSE loss function, the batch size is set to 32, the number of iterations is 200 rounds, and the training ends when the loss function converges, indicating that the training is completed.

[0131] It should be noted that the Adam optimizer is one of the five major optimizers commonly used in machine learning, and its full name is Adaptive Moment Estimation.

[0132] The second machine learning model training method includes:

[0133] Convert temperature data, saturation and specific features into a corresponding set of feature vectors;

[0134] Temperature data, saturation and specific features are used as inputs of a machine learning model, and the machine learning model uses the current compensation amount corresponding to each set of temperature data, saturation and specific features as output, and the current compensation amount actually corresponding to each set of temperature data, saturation and specific features as a prediction target, and minimizing the loss function value of the machine learning model as a training target; training is stopped when the loss function value of the second machine learning model is less than or equal to a preset target loss value, and the second machine learning model is a deep feedforward neural network.

[0135] The loss function value of the second machine learning model is the mean square error;

[0136] Mean square error is one of the commonly used loss functions. By transforming the loss function formula Minimization is used to train the model so that the machine learning model can better fit the data, thereby improving the performance and accuracy of the model.

[0137] In the loss function, MSE is the loss function value of the second machine learning model, x is the feature vector group number; e is the number of feature vector groups; y x is the current compensation amount output by the second machine learning model corresponding to the x-th group of feature vectors, is the current compensation amount actually corresponding to the xth group of eigenvectors.

[0138] In this embodiment, the working parameters of the current transformer are collected in real time, the saturation is predicted based on the working parameters of the current transformer, and then the working parameters of the current transformer are extracted, and a feature set is established. The features in the feature set are subjected to correlation analysis, and redundant features are removed. Then, the temperature data, saturation and feature data are input into a pre-built second machine learning model for predicting the current compensation amount, and the current compensation amount is output. Based on the output current compensation amount, current compensation is performed on the compensation winding in the current transformer body, so that the automatic intelligent compensation method of the current transformer can more accurately adjust the compensation current by real-time monitoring and predicting the change of the saturation of the current transformer, thereby avoiding measurement errors caused by saturation changes and improving the accuracy of the current transformer in different working environments. Secondly, by performing correlation analysis on the feature data and removing redundant features, the data input into the machine learning model is ensured to be more accurate and efficient, which not only improves the prediction accuracy, but also reduces the computational complexity and improves the system operation efficiency. Finally, by collecting the working parameters of the current transformer in real time, it can quickly respond to environmental changes and accurately adjust the compensation amount. Under the influence of temperature changes or other external factors, it ensures that the current transformer can always maintain the best working state, avoids the performance degradation caused by unstable working parameters, and overcomes the problem that the existing technology is difficult to adapt to the changes in transformer characteristics caused by factors such as ambient temperature and load changes in actual operation, and the compensation algorithm used is relatively simple, which fails to fully utilize the advantages of modern signal processing technology and machine learning algorithms, limiting the compensation accuracy.

[0139] Example 2

[0140] See also Figure 3 , an automatic intelligent compensation system for a current transformer, used to implement the automatic intelligent compensation method for a current transformer, the system comprising:

[0141] A data acquisition module collects the working data of the current transformer in real time, wherein the working data of the current transformer includes temperature data and working parameters;

[0142] a data analysis module, inputting the working parameters into a pre-built first machine learning model for predicting saturation, and outputting the saturation;

[0143] a feature extraction module, processing the working parameters, extracting feature data, establishing a time series feature data set and a numerical feature data set, performing correlation analysis on features within the time series feature data set and the numerical feature data set, removing redundant features, and marking the remaining features as specific features;

[0144] The data processing module inputs the temperature data, saturation and specific features into a pre-built second machine learning model for predicting the current compensation amount, outputs the current compensation amount, and performs current compensation on the compensation winding in the current transformer body based on the output current compensation amount.

[0145] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above-described embodiments. The above-described embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. An automatic intelligent compensation method for a current transformer, applied to a current transformer body, is characterized in that: The following steps are involved: Real-time collection of operating data of the current transformer, wherein the operating data of the current transformer includes temperature data and operating parameters; Inputting the working parameters into a pre-built first machine learning model for predicting saturation, and outputting the saturation; Processing the working parameters, extracting feature data, establishing a time series feature data set and a numerical feature data set, performing correlation analysis on features in the time series feature data set and the numerical feature data set, removing redundant features, and marking the remaining features as specific features; The temperature data, saturation and specific features are input into a pre-built second machine learning model for predicting current compensation, and the current compensation amount is output. Current compensation is performed on the compensation winding in the current transformer body based on the output current compensation amount.

2. The automatic intelligent compensation method for current transformer according to claim 1, characterized in that: The operating parameters include input current, output current, output voltage, input current frequency and load characteristics. The load characteristics include continuous load, impact load or instantaneous load. The load characteristics are obtained by a load sensor or detecting a current waveform.

3. The automatic intelligent compensation method for current transformer according to claim 2, characterized in that: The method of processing the working parameters, extracting feature data, and establishing a time series feature data set and a numerical feature data set includes: Each time the working parameters are collected, a timestamp is recorded for each sampling point, and the collected input current, output current, output voltage, and input current frequency are associated with their timestamps and stored in a time series format to generate time series data; Calculate the average value, standard deviation, and root mean square value of the input current, output current, and output voltage to obtain time domain characteristics; The input current, output current and output voltage are converted from the time domain to the frequency domain through fast Fourier transform, the frequency components in the signal are extracted, the power distribution of the signal at different frequencies is calculated to obtain the frequency components and power spectrum density, and the frequency domain characteristics are obtained; Calculate the ratio of the current transformer input current to the rated current, as well as the ratio of the output current to the input current, to obtain instantaneous characteristics; Establishing a time series feature data set based on the acquired input current, output current, output voltage, and input current frequency, wherein the time series feature data set stores time series data; A numerical feature data set is established using the acquired time domain features, frequency domain features, and instantaneous features. The numerical feature data set stores numerical data.

4. The automatic intelligent compensation method for current transformer according to claim 3, characterized in that: Methods for performing correlation analysis on features within a time series feature data set and removing redundant features include: Calculate the mutual information value between any two features in the time series feature data set; The mutual information matrix is ​​used to evaluate the correlation between each feature and other features, and redundant features are removed based on the correlation between two features.

5. The automatic intelligent compensation method for current transformer according to claim 4, characterized in that: The method for removing redundant features based on the correlation between two features includes: A correlation threshold is preset. When the mutual information value between two features is greater than or equal to the preset correlation threshold, there is a high degree of redundancy between the two features, and one of them is randomly removed. Conversely, when the mutual information value between the two features is less than the preset correlation threshold, both features are retained.

6. The automatic intelligent compensation method for current transformer according to claim 1, characterized in that: The method of performing correlation analysis on features in a numerical feature data set, removing redundant features, and marking the remaining features as specific features includes: removing redundant features by obtaining correlation coefficients for the features in the numerical feature data set, applying Lasso regression to perform feature importance evaluation on the remaining feature data, retaining features with evaluation values ​​not zero, and marking the retained features as specific features.

7. The automatic intelligent compensation method for current transformer according to claim 6, characterized in that: The method of eliminating redundant features by obtaining correlation coefficients for features in the numerical feature data set and applying Lasso regression to the remaining feature data for feature importance evaluation includes: Normalize the features in the numerical feature data set and scale the values ​​to the range of [0,1]; Calculate the correlation coefficient between every two features in the numerical feature data set and generate a correlation coefficient matrix; A correlation coefficient threshold is preset. If the absolute value of the correlation coefficient between two features is greater than the preset correlation coefficient threshold, one feature is retained and the other redundant feature is deleted. This process is repeated until the correlation coefficients between the remaining features are all lower than the correlation coefficient threshold. Obtaining a numerical feature data set and a target variable after removing redundant features, wherein the target variable is a current compensation amount; Select the Lasso regression model; Get the regression coefficient β of the Lasso model j : If β j = 0, delete feature X j ; If β j ≠0, retain the corresponding features and mark them as specific features.

8. The automatic intelligent compensation method for current transformer according to claim 7, characterized in that: The method for constructing the second machine learning model includes: Initialize the second machine learning model structure. The second machine learning model structure adopts an MLP type multi-layer forward network structure with three input layers, two hidden layers, and one output layer. The input layer includes the first input layer, the second input layer, and the third input layer. The first input layer corresponds to temperature data, the second input layer corresponds to saturation, and the third input layer node is [1, c], which is a specific feature, and c is the number of specific features. The hidden layer includes the first hidden layer and the second hidden layer, both of which use the ReLU function as the activation function. The output layer is the first output layer, and the first output layer corresponds to the predicted current compensation amount. After initializing the second machine learning model structure, the second machine learning model is trained using temperature data, saturation, specific features, and current compensation corresponding to the temperature data, saturation, and specific features. The optimizer is Adam, and the loss function is the MSE loss function. The training ends when the loss function converges.

9. The automatic intelligent compensation method for current transformer according to claim 2, characterized in that: The input current is the actual current flowing through the current transformer, which is directly measured by a current sensor or the primary side of the current transformer. The output current is the secondary current output by the current transformer, which is measured by the secondary side of the current transformer. The output voltage is the secondary terminal voltage of the current transformer, which is collected and obtained by a voltage sensor. The input current frequency is obtained by measuring the frequency of the input current with a frequency meter or a digital oscilloscope.

10. Automatic intelligent compensation system for current transformer, characterized in that: A method for implementing the automatic intelligent compensation of a current transformer according to any one of claims 1 to 9, the system comprising: A data acquisition module collects the working data of the current transformer in real time, wherein the working data of the current transformer includes temperature data and working parameters; a data analysis module, inputting the working parameters into a pre-built first machine learning model for predicting saturation, and outputting the saturation; a feature extraction module, processing the working parameters, extracting feature data, establishing a time series feature data set and a numerical feature data set, performing correlation analysis on features within the time series feature data set and the numerical feature data set, removing redundant features, and marking the remaining features as specific features; The data processing module inputs the temperature data, saturation and specific features into a pre-built second machine learning model for predicting the current compensation amount, outputs the current compensation amount, and performs current compensation on the compensation winding in the current transformer body based on the output current compensation amount.

Images