A method and system for detecting saturation of a current transformer based on machine learning

Abstract

The application discloses a kind of based on machine learning's current transformer saturation detection method and system.Method includes: based on power system electromagnetic transient simulation program generates CT saturation waveform under different conditions;Classified, obtain the CT saturation waveform sample set after labeling;Each sample waveform in CT saturation waveform sample set is handled, respectively extracts time domain feature, frequency domain feature and transient feature, constructs multi-domain feature set;Improved gradient boosting tree model is used to supervise learning training to multi-domain feature set, and constructs CT saturation detection classifier;Obtain the current waveform data to be detected, execute feature extraction operation, obtain the multi-domain feature vector to be detected;The multi-domain feature vector to be detected is input into the CT saturation detection classifier that has been trained, and the classification result of CT saturation degree and prediction confidence are output.The application realizes the automation grading of CT saturation state, effectively solves the problem of low saturation efficiency and poor accuracy of traditional identification method.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for power system protection, specifically to a method and system for detecting saturation of current transformers based on machine learning. Background Technology

[0002] Current transformers (CTs) are critical equipment in power systems, and their measurement accuracy and transmission characteristics directly affect the safety and reliability of system operation. When a large-current short-circuit fault occurs in the system, the CT core may enter a saturated state due to a sharp increase in magnetic flux density, resulting in severe distortion of the secondary current waveform, which cannot accurately reflect the primary current. This distortion can seriously affect the correct operation of relay protection devices that rely on secondary current (such as differential protection and distance protection), potentially leading to protection failure or maloperation, thereby expanding the fault range and endangering the stability of the power grid.

[0003] Currently, the detection and identification of CT saturation on-site mainly relies on professional technicians manually analyzing fault waveforms. Technicians need to rely on experience to observe waveform clipping, distortion, and other characteristics to determine whether saturation has occurred and its severity. This method is highly dependent on personal experience, the analysis results are subjective, lack unified and objective standards, and is difficult to adapt to the protection and maintenance requirements of smart grids.

[0004] In recent years, although some studies have attempted to use single waveform characteristics (such as harmonic content) or simple threshold judgment methods for saturation detection, these methods have poor adaptability, are difficult to accurately quantify the degree of saturation, and cannot meet the high requirements of modern intelligent operation and maintenance platforms for detection accuracy and robustness. Therefore, there is an urgent need for an intelligent technical solution that can automatically, accurately, and reliably detect and classify the degree of CT saturation. Summary of the Invention

[0005] The purpose of this invention is to address the aforementioned problems in the existing technology by providing a machine learning-based method and system for detecting current transformer saturation. By fusing multi-domain features of the current waveform and utilizing an optimized machine learning model, the invention achieves automated classification of CT saturation status, effectively solving the problems of low efficiency and poor accuracy in traditional protection and maintenance operations that rely on manual experience to identify CT saturation.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0007] Firstly, a machine learning-based method for detecting saturation in current transformers includes the following steps:

[0008] Based on the electromagnetic transient simulation program of the power system, the fault current magnitude, primary system time constant, CT secondary load, and fault initiation angle parameters are adjusted to simulate and generate CT saturation waveforms under different operating conditions. The CT saturation waveforms are then categorized according to the severity of saturation to obtain a categorized CT saturation waveform sample set.

[0009] Each sample waveform in the CT saturated waveform sample set is processed to extract time-domain features, frequency-domain features, and transient features to construct a multi-domain feature set.

[0010] An improved gradient boosting tree model is used to conduct supervised learning training on the multi-domain feature set to construct a CT saturation detection classifier;

[0011] Acquire the waveform data of the current to be detected, perform feature extraction operation to obtain the multi-domain feature vector to be detected; input the multi-domain feature vector to be detected into the trained CT saturation detection classifier, and output the classification result and prediction confidence of the CT saturation degree.

[0012] Furthermore, the severity of saturation is reflected by waveform distortion rate and / or saturation onset time.

[0013] Furthermore, the time-domain features include peak factor, kurtosis, skewness, and rate of change of zero-crossing slope; the frequency-domain features include the fundamental-to-harmonic amplitude ratio and total harmonic distortion rate; and the transient features include wavelet transform modulus maxima, absolute value integral of current derivative, number of waveform abrupt change points, and amplitude.

[0014] Furthermore, the time-domain feature calculation method is as follows:

[0015] Peak factor: ,in The peak value of the waveform. The effective value of the waveform. This represents the instantaneous amplitude at the i-th sampling point in the signal sequence. Indicates the number of sampling points involved in the calculation;

[0016] kurtosis: ,in The mean of the waveform. Standard deviation;

[0017] Skewness: ;

[0018] Rate of change of slope at zero point: ,in This represents the instantaneous value difference between the (n+1)th sampling point near its zero-crossing point and the adjacent nth sampling point. This represents the instantaneous value difference between the nth sampling point near its zero-crossing point and the adjacent (n-1)th sampling point. This indicates the time interval between sampling points.

[0019] Furthermore, the frequency domain features are calculated as follows:

[0020] Fundamental to Harmonic Amplitude Ratio: Perform a Fast Fourier Transform on the preprocessed waveform to calculate the ratio of each harmonic amplitude to the fundamental amplitude. , , ..., );

[0021] Total harmonic distortion: ,in The fundamental amplitude, The amplitude of the second harmonic. The value is the amplitude of the 3rd harmonic, and n is the highest harmonic number.

[0022] Furthermore, the transient characteristics are calculated as follows:

[0023] Wavelet transform modulus maxima: The waveform is decomposed into 5 levels using the db4 wavelet basis function, and the modulus maxima of the detail coefficients in the 3rd to 5th levels are taken;

[0024] Integral of the absolute value of the current derivative: , This refers to the transient period of the fault.

[0025] Number and amplitude of waveform abrupt changes: Perform first-order difference on the waveform. ,when When the threshold is exceeded, it is identified as a mutation point. The number of mutation points is counted and the maximum mutation amplitude is calculated.

[0026] Furthermore, the improved gradient boosting tree model is improved through hyperparameter optimization, adaptive feature weights, and an early stopping mechanism.

[0027] Furthermore, the specific improvements to the gradient boosting tree model include: using a five-fold cross-validation method to optimize key hyperparameters such as learning rate, maximum tree depth, and number of weak learners; assigning initial weights greater than the baseline weights to transient features based on domain knowledge, and assigning baseline weights to time-domain and frequency-domain features; and introducing an early stopping strategy to monitor the accuracy of the validation set during training, and automatically terminating training if no improvement is achieved after a specified number of iterations.

[0028] The construction of the CT saturation detection classifier includes: iteratively constructing a gradient boosting tree using a forward step-by-step addition method. In each iteration, the negative gradient between the current model prediction result and the true label is calculated first, and then a new decision tree is trained to fit the gradient. The weights are determined through linear search, and the model is gradually optimized. The final classifier is formed by integrating multiple decision trees, which can effectively capture the complex nonlinear relationship between features in multiple domains and achieve high-precision recognition of CT saturation state.

[0029] Secondly, a machine learning-based current transformer saturation detection system includes:

[0030] The sample generation and labeling module is used to simulate and generate CT saturation waveforms under different operating conditions by adjusting the fault current magnitude, primary system time constant, CT secondary load, and fault initiation angle parameters based on the power system electromagnetic transient simulation program; and to label the CT saturation waveforms according to the severity of saturation to obtain a labeled CT saturation waveform sample set.

[0031] The multi-domain feature extraction module is used to process each sample waveform in the CT saturated waveform sample set, extracting time-domain features, frequency-domain features, and transient features respectively, and constructing a multi-domain feature set; the time-domain features include peak factor, kurtosis, skewness, and rate of change of slope at the zero crossing point;

[0032] The model training module is used to perform supervised learning training on the multi-domain feature set using an improved gradient boosting tree model to construct a CT saturation detection classifier.

[0033] The online detection module is used to acquire the waveform data of the current to be detected, perform feature extraction operation to obtain the multi-domain feature vector to be detected, input the multi-domain feature vector to be detected into the pre-trained CT saturation detection classifier, and output the classification result and prediction confidence of the CT saturation degree.

[0034] Thirdly, the present invention also provides an electronic device comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the steps of the machine learning-based current transformer saturation detection method as described in the first aspect.

[0035] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the machine learning-based current transformer saturation detection method as described in the first aspect.

[0036] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the machine learning-based current transformer saturation detection method as described in the first aspect.

[0037] Beneficial effects: This invention realizes the transformation of CT saturation detection from manual experience to automatic recognition, improving detection efficiency and objectivity; through multi-domain feature fusion and an improved Gradient Boosting Decision Tree (GBDT) model, it improves detection accuracy and adaptability to operating conditions; and it can be integrated into existing intelligent operation and maintenance platforms, and directly applied to scenarios such as protection action analysis and equipment status assessment, which helps to improve the level of intelligence in power system protection operation and maintenance. Attached Figure Description

[0038] Figure 1 This is a flowchart of a machine learning-based current transformer saturation detection method. Detailed Implementation

[0039] The technical solutions in the embodiments of the present invention will now be clearly and completely described in conjunction with the accompanying drawings.

[0040] Figure 1 This invention provides a flowchart of a machine learning-based current transformer saturation detection method. The method mainly includes: generating current transformer (CT) saturation waveforms under different operating conditions based on a power system simulation program, and labeling them according to their saturation degree to obtain a labeled CT saturation waveform sample set; extracting time-domain, frequency-domain, and transient features from the sample waveforms to construct a multi-domain feature set; training the multi-domain feature set using an improved GBDT model to construct a CT saturation detection classifier; acquiring the current waveform to be detected, performing feature extraction, inputting it into the classifier, and outputting the CT saturation degree detection result. The specific steps are as follows:

[0041] S1, Sample Simulation and Labeling: Using power system simulation software, simulate the CT saturation waveform under different operating conditions.

[0042] Based on a power system electromagnetic transient simulation program, the fault current magnitude, primary system time constant, CT secondary load, and fault initiation angle parameters are adjusted to simulate and generate CT saturation waveforms under different operating conditions. In this embodiment of the invention, the simulation parameters are set as follows: fault current magnitude: 5kA~30kA; primary system time constant: 0.05s~0.2s; CT secondary load: 0.5Ω~2.0Ω; fault initiation angle: 0°~90°.

[0043] The CT saturated waveforms are categorized according to their severity of saturation, resulting in a categorized CT saturated waveform sample set. The severity of saturation can be reflected by indicators such as waveform distortion rate and saturation onset time. Preferably, the categorization includes four types: "normal," "mildly saturated," "moderately saturated," and "severely saturated," with at least 1000 samples in each category to ensure sample diversity and representativeness. In this embodiment, the sample categories are divided according to waveform distortion rate: Normal: distortion rate <2%; Mildly saturated: distortion rate 2%~10%; Moderately saturated: distortion rate 10%~20%; Severely saturated: distortion rate >20%.

[0044] S2, Feature Extraction and Construction: Each waveform sample in the CT saturation waveform sample set is processed to extract time-domain, frequency-domain, and transient features, constructing a multi-domain feature set. This multi-domain feature set is derived from in-depth analysis of the physical mechanisms of CT saturation. When CT saturates, the core flux saturation leads to distortion of the secondary current waveform, increased harmonics, and transient abrupt changes. These phenomena correspond to time-domain, frequency-domain, and transient features, respectively. Time-domain features (such as peak factor and kurtosis) directly reflect waveform amplitude changes and distribution patterns; waveform clipping during saturation increases the peak factor and causes kurtosis to deviate from the normal value. Frequency-domain features (such as harmonic ratio and THD) capture harmonic distortion caused by core nonlinearity; harmonic content increases significantly during saturation. Transient features (such as wavelet mode maxima and abrupt change points) target rapid changes in the initial stage of saturation, features often ignored in traditional methods. Compared to existing technologies that rely solely on single-domain features (such as harmonic content), this invention, through multi-domain fusion, can more comprehensively and robustly quantify the degree of saturation, solving the problem of poor adaptability in traditional methods and improving detection accuracy.

[0045] Time-domain characteristics: peak factor, kurtosis, skewness, and rate of change of zero-crossing slope, reflecting the shape and variation characteristics of the current waveform in the time dimension;

[0046] Frequency domain characteristics: fundamental frequency to harmonic amplitude ratio, total harmonic distortion (THD), capturing the frequency distribution and harmonic interference of the current signal;

[0047] Transient characteristics: wavelet transform modulus maxima, absolute value integral of current derivative, number and amplitude of waveform abrupt change points, extract key change information in the transient process of current signal.

[0048] Specifically, after waveform normalization, the relevant waveform characteristics can be calculated using the following method:

[0049] Temporal characteristics:

[0050] Peak factor: ,in The peak value of the waveform. The effective value of the waveform. This represents the instantaneous amplitude at the i-th sampling point in the signal sequence. Indicates the number of sampling points involved in the calculation;

[0051] kurtosis: ,in The mean of the waveform. The standard deviation is (the kurtosis of a normal distribution is 3, and the kurtosis deviates significantly when saturated).

[0052] Skewness: This reflects the symmetry of the waveform; the asymmetry of the waveform increases during saturation.

[0053] Rate of change of slope at zero crossings: Calculate the slope of the waveform at two adjacent zero crossings. Take the absolute value of the slope difference, where This represents the instantaneous value difference between the (n+1)th sampling point near its zero-crossing point and the adjacent nth sampling point. This represents the instantaneous value difference between the nth sampling point near its zero-crossing point and the adjacent (n-1)th sampling point. Indicates the time interval between sampling points;

[0054] Frequency domain characteristics:

[0055] Fundamental to Harmonic Amplitude Ratio: Perform a Fast Fourier Transform (FFT) on the preprocessed waveform to calculate the ratio of each harmonic amplitude to the fundamental amplitude. , , ..., );in , , ...these represent the amplitudes of the fundamental wave, the second harmonic, the third harmonic, etc.

[0056] Total harmonic distortion: (n is the highest harmonic order, taken as 13).

[0057] Transient characteristics:

[0058] Wavelet transform modulus maxima: The waveform is decomposed into 5 levels using the db4 wavelet basis function, and the modulus maxima of the detail coefficients in the 3rd to 5th levels are taken;

[0059] Integral of the absolute value of the current derivative: , ( The fault transient period is taken as 0.01s to 0.05s.

[0060] Number and amplitude of waveform abrupt changes: Perform first-order difference on the waveform. ,when When the value exceeds the threshold (0.1 times the peak value), it is determined as a mutation point. The number of mutation points is counted and the maximum mutation amplitude is calculated.

[0061] S3, Model Training and Construction: The improved Gradient Boosting Tree (GBDT) model is used to conduct supervised learning training on the multi-domain feature set constructed in step S2 to construct a CT saturation detection classifier.

[0062] The improvements to the gradient boosting tree model in this invention include:

[0063] Hyperparameter optimization: The five-fold cross-validation method is used to optimize key hyperparameters such as learning rate, maximum tree depth, and number of weak learners.

[0064] Adaptive feature weights: Based on domain knowledge, transient features are assigned initial weights greater than the baseline weights, while time-domain and frequency-domain features are assigned baseline weights; for example, the baseline weight is 1.0, and the initial weights of transient features are set to 1.2-1.5 to guide the model to focus on features with stronger discriminative power.

[0065] Early stopping mechanism: An early stopping strategy is introduced to monitor the accuracy of the validation set during training. If the accuracy does not improve after a specified number of iterations, training is automatically terminated to prevent overfitting and ensure the model's generalization ability.

[0066] In this embodiment of the invention, GradientBoostingClassifier from the Scikit-learn library is used as the base model. The improvements are as follows: 5-fold cross-validation and grid search are used to optimize the hyperparameters, ultimately determining a learning rate of 0.05, a maximum tree depth of 4, and a number of weak learners of 200; the initial weights for transient features are set to 1.3, and the weights for time-domain and frequency-domain features are set to 1.0; an early stopping mechanism is enabled, terminating training after 20 consecutive rounds without improvement.

[0067] Before model training, the multi-domain feature set is standardized to convert each feature into a distribution with a mean of 0 and a standard deviation of 1. This eliminates the problem of inconsistent numerical scales caused by differences in the units of measurement of different features, and avoids certain features from having a dominant influence on model training due to their large values. This ensures that all features have an equal contribution during the training process.

[0068] An improved gradient boosting tree model training process was adopted, introducing three optimization measures on the basis of traditional GBDT. First, grid search combined with five-fold cross-validation was used to optimize key hyperparameters such as learning rate, maximum tree depth, and number of weak learners. Through multiple training and validation cycles, the optimal parameter combination was selected, and the learning rate was finally determined to be 0.05, the maximum tree depth to be 4, and the number of weak learners to be 200. Second, based on knowledge of the power system domain, a relatively high initial weight of 1.3 was set for transient features, and the time domain and frequency domain features were set to a baseline weight of 1.0, to guide the model to focus more on features sensitive to saturation transient responses. In addition, an early stopping mechanism was introduced to monitor the accuracy of the validation set during training. If no improvement was seen for several consecutive rounds (e.g., 20 rounds), training was terminated early to prevent overfitting and ensure that the model has good generalization ability.

[0069] The gradient boosting tree is constructed iteratively using a forward step-by-step addition method. In each iteration, the negative gradient between the current model prediction and the true label is calculated first. Then, a new decision tree is trained to fit this gradient, and its weights are determined through linear search to gradually optimize the model. The final classifier is an ensemble of multiple decision trees, which can effectively capture the complex nonlinear relationships between multi-domain features and achieve high-precision identification of CT saturation state. S4, Online Detection and Application: In practical application scenarios, the waveform data of the current to be detected is acquired, and the feature extraction operation in step S2 is repeated to obtain the multi-domain feature vector to be detected. This vector is input into the pre-trained CT saturation detection classifier, and the model outputs the classification result and prediction confidence of the CT saturation degree. The detection results can be used for the evaluation of the action behavior of power system relay protection or as a basis for the investigation of potential hazards in the operation status of the CT itself.

[0070] As an example, the trained CT saturation detection classification model is deployed on the intelligent operation and maintenance server of the substation. It automatically reads the fault recording waveform, calculates multi-dimensional features based on the feature extraction method, and inputs them into the loaded classifier. The model outputs the results (e.g., "moderate saturation, confidence level 96%)" for protection action behavior evaluation and to guide equipment inspection.

[0071] This invention realizes the transformation of CT saturation detection from manual experience to automatic recognition, improving detection efficiency and objectivity; through multi-domain feature fusion and an improved GBDT model, it enhances detection accuracy and adaptability to operating conditions; and it can be integrated into existing intelligent operation and maintenance platforms, and directly applied to scenarios such as protection action analysis and equipment status assessment, which helps to improve the level of intelligence in power system protection operation and maintenance.

[0072] Based on the same technical concept as the above-described method embodiments, another embodiment of the present invention provides a current transformer saturation detection system based on machine learning, comprising:

[0073] The sample generation and labeling module is used to simulate and generate CT saturation waveforms under different operating conditions by adjusting the fault current magnitude, primary system time constant, CT secondary load, and fault initiation angle parameters based on the power system electromagnetic transient simulation program; and to label the CT saturation waveforms according to the severity of saturation to obtain a labeled CT saturation waveform sample set.

[0074] The multi-domain feature extraction module is used to process each sample waveform in the CT saturated waveform sample set, extract time-domain features, frequency-domain features and transient features respectively, and construct a multi-domain feature set;

[0075] The model training module is used to perform supervised learning training on the multi-domain feature set using an improved gradient boosting tree model to construct a CT saturation detection classifier.

[0076] The online detection module is used to acquire the waveform data of the current to be detected, perform feature extraction operation to obtain the multi-domain feature vector to be detected, input the multi-domain feature vector to be detected into the pre-trained CT saturation detection classifier, and output the classification result and prediction confidence of the CT saturation degree.

[0077] The severity of saturation is reflected by waveform distortion rate and / or saturation onset time.

[0078] The time-domain features include peak factor, kurtosis, skewness, and rate of change of zero-crossing slope; the frequency-domain features include the fundamental-to-harmonic amplitude ratio and total harmonic distortion rate; the transient features include wavelet transform modulus maxima, absolute value integral of current derivative, number and amplitude of waveform abrupt change points.

[0079] The time-domain feature calculation method is as follows:

[0080] Peak factor: ,in The peak value of the waveform. The effective value of the waveform. This represents the instantaneous amplitude at the i-th sampling point in the signal sequence. Indicates the number of sampling points involved in the calculation;

[0081] kurtosis: ,in The mean of the waveform. Standard deviation;

[0082] Skewness: ;

[0083] Rate of change of slope at zero point: ,in This represents the instantaneous value difference between the (n+1)th sampling point near its zero-crossing point and the adjacent nth sampling point. This represents the instantaneous value difference between the nth sampling point near its zero-crossing point and the adjacent (n-1)th sampling point. This indicates the time interval between sampling points.

[0084] The frequency domain features are calculated as follows:

[0085] Fundamental to Harmonic Amplitude Ratio: Perform a Fast Fourier Transform on the preprocessed waveform to calculate the ratio of each harmonic amplitude to the fundamental amplitude. , , ..., );

[0086] Total harmonic distortion: ,in The fundamental amplitude, The amplitude of the second harmonic. The value is the amplitude of the 3rd harmonic, and n is the highest harmonic number.

[0087] The transient characteristics are calculated as follows:

[0088] Wavelet transform modulus maxima: The waveform is decomposed into 5 levels using the db4 wavelet basis function, and the modulus maxima of the detail coefficients in the 3rd to 5th levels are taken;

[0089] Integral of the absolute value of the current derivative: , This refers to the transient period of the fault.

[0090] Number and amplitude of waveform abrupt changes: Perform first-order difference on the waveform. ,when When the threshold is exceeded, it is identified as a mutation point. The number of mutation points is counted and the maximum mutation amplitude is calculated.

[0091] The improved gradient boosting tree model is improved through hyperparameter optimization, adaptive feature weights, and an early stopping mechanism.

[0092] The specific improvements to the gradient boosting tree model include: using a five-fold cross-validation method to optimize key hyperparameters such as learning rate, maximum tree depth, and number of weak learners; assigning initial weights greater than the baseline weights to transient features based on domain knowledge, and assigning baseline weights to time-domain and frequency-domain features; and introducing an early stopping strategy to monitor the accuracy of the validation set during training, and automatically terminating training if no improvement is achieved after a specified number of iterations.

[0093] The construction of the CT saturation detection classifier includes: iteratively constructing a gradient boosting tree using a forward step-by-step addition method. In each iteration, the negative gradient between the current model prediction result and the true label is calculated first, and then a new decision tree is trained to fit the gradient. The weights are determined through linear search, and the model is gradually optimized. The final classifier is formed by integrating multiple decision trees, which can effectively capture the complex nonlinear relationship between features in multiple domains and achieve high-precision recognition of CT saturation state.

[0094] It should be understood that the machine learning-based current transformer saturation detection system in this embodiment can implement all the technical solutions in the above method embodiments. The functions of each functional module can be specifically implemented according to the methods in the above method embodiments. The specific implementation process can be referred to the relevant descriptions in the above method embodiments, which will not be repeated here.

[0095] Another embodiment of the present invention provides an electronic device, including: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the programs, when executed by the processors, implement the steps of the machine learning-based current transformer saturation detection method as described above.

[0096] Another embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the machine learning-based current transformer saturation detection method described above.

[0097] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus (systems), electronic devices, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0098] This invention is described with reference to a flowchart of a method according to embodiments of the invention. It should be understood that each step in the flowchart and combinations thereof can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the process. Figure 1 A device for a function specified in one or more processes.

[0099] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 The function specified in one or more processes.

[0100] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 Steps of a specified function in one or more processes.

Claims

1. A machine learning-based method for detecting saturation in current transformers, characterized in that, Includes the following steps: Based on the electromagnetic transient simulation program of the power system, the fault current magnitude, primary system time constant, CT secondary load, and fault initiation angle parameters are adjusted to simulate and generate CT saturation waveforms under different operating conditions. The CT saturation waveforms are then categorized according to the severity of saturation to obtain a categorized CT saturation waveform sample set. Each waveform in the CT saturated waveform sample set is processed to extract time-domain features, frequency-domain features, and transient features to construct a multi-domain feature set. The time-domain features include peak factor, kurtosis, skewness, and rate of change of zero-crossing slope. The frequency-domain features include the ratio of fundamental to harmonic amplitude and total harmonic distortion. The transient features include wavelet transform modulus maxima, integral of absolute value of current derivative, number of waveform abrupt change points, and amplitude. An improved gradient boosting tree model is used to conduct supervised learning training on the multi-domain feature set to construct a CT saturation detection classifier. The improvements to the gradient boosting tree model include: using a five-fold cross-validation method to optimize key hyperparameters such as learning rate, maximum tree depth, and number of weak learners; assigning initial weights greater than the baseline weights to transient features based on domain knowledge, and assigning baseline weights to time-domain and frequency-domain features; and introducing an early stopping strategy to monitor the accuracy of the validation set during training, and automatically terminating training if the accuracy does not improve after a specified number of iterations. The construction of the CT saturation detection classifier includes: iteratively constructing a gradient boosting tree using a forward step-by-step addition method. In each iteration, the negative gradient between the current model prediction result and the true label is calculated first, and then a new decision tree is trained to fit the negative gradient. The weights are determined through linear search, and the model is gradually optimized. The final classifier is formed by integrating multiple decision trees, which can effectively capture the complex nonlinear relationship between features in multiple domains and achieve high-precision recognition of CT saturation state. Acquire the waveform data of the current to be detected, perform feature extraction operation to obtain the multi-domain feature vector to be detected; input the multi-domain feature vector to be detected into the trained CT saturation detection classifier, and output the classification result and prediction confidence of the CT saturation degree.

2. The method according to claim 1, characterized in that, The severity of saturation is reflected by waveform distortion rate and / or saturation onset time.

3. The method according to claim 1, characterized in that, The time-domain feature calculation method is as follows: Peak factor: ,in The peak value of the waveform. The effective value of the waveform. This represents the instantaneous amplitude at the i-th sampling point in the signal sequence. Indicates the number of sampling points involved in the calculation; kurtosis: ,in The mean of the waveform. Standard deviation; Skewness: ; Rate of change of slope at zero point: ,in This represents the instantaneous value difference between the (n+1)th sampling point near its zero-crossing point and the adjacent nth sampling point. This represents the instantaneous value difference between the nth sampling point near its zero-crossing point and the adjacent (n-1)th sampling point. This indicates the time interval between sampling points.

4. The method according to claim 1, characterized in that, The frequency domain features are calculated as follows: Fundamental to Harmonic Amplitude Ratio: Perform a Fast Fourier Transform on the preprocessed waveform to calculate the ratio of each harmonic amplitude to the fundamental amplitude. , , ..., ); Total harmonic distortion: ,in The fundamental amplitude, The amplitude of the second harmonic. The value is the amplitude of the 3rd harmonic, and n is the highest harmonic number.

5. The method according to claim 1, characterized in that, The transient characteristics are calculated as follows: Wavelet transform modulus maxima: The waveform is decomposed into 5 levels using the db4 wavelet basis function, and the modulus maxima of the detail coefficients in the 3rd to 5th levels are taken; Integral of the absolute value of the current derivative: , This refers to the transient period of the fault. Number and amplitude of waveform abrupt changes: Perform first-order difference on the waveform. , This represents the instantaneous amplitude at the i-th sampling point in the signal sequence. Represents the instantaneous amplitude at the (i+1)th sampling point in the signal sequence, when When the threshold is exceeded, it is identified as a mutation point. The number of mutation points is counted and the maximum mutation amplitude is calculated.

6. A current transformer saturation detection system based on machine learning, characterized in that, include: The sample generation and labeling module is used to simulate and generate CT saturation waveforms under different operating conditions by adjusting the fault current magnitude, primary system time constant, CT secondary load, and fault initiation angle parameters based on the power system electromagnetic transient simulation program; and to label the CT saturation waveforms according to the severity of saturation to obtain a labeled CT saturation waveform sample set. The multi-domain feature extraction module is used to process each sample waveform in the CT saturated waveform sample set, extracting time-domain features, frequency-domain features, and transient features respectively, and constructing a multi-domain feature set. The time-domain features include peak factor, kurtosis, skewness, and rate of change of zero-crossing slope. The frequency-domain features include the fundamental-to-harmonic amplitude ratio and the total harmonic distortion rate. The transient features include the wavelet transform modulus maxima, the absolute value integral of the current derivative, the number of waveform abrupt change points, and their amplitude. The model training module is used to perform supervised learning training on the multi-domain feature set using an improved gradient boosting tree model to construct a CT saturation detection classifier. The specific improvements to the gradient boosting tree model include: using a five-fold cross-validation method to optimize key hyperparameters such as learning rate, maximum tree depth, and number of weak learners; assigning initial weights greater than the baseline weights to transient features based on domain knowledge, and assigning baseline weights to time-domain and frequency-domain features; and introducing an early stopping strategy to monitor the accuracy of the validation set during training, and automatically terminating training if no improvement is achieved after a specified number of iterations. The construction of the CT saturation detection classifier includes: iteratively constructing a gradient boosting tree using a forward step-by-step addition method. In each iteration, the negative gradient between the current model prediction result and the true label is calculated first, and then a new decision tree is trained to fit the negative gradient. The weights are determined through linear search, and the model is gradually optimized. The final classifier is formed by integrating multiple decision trees, which can effectively capture the complex nonlinear relationship between features in multiple domains and achieve high-precision recognition of CT saturation state. The online detection module is used to acquire the waveform data of the current to be detected, perform feature extraction operation to obtain the multi-domain feature vector to be detected, input the multi-domain feature vector to be detected into the pre-trained CT saturation detection classifier, and output the classification result and prediction confidence of the CT saturation degree.

7. An electronic device, comprising: One or more processors; Memory; And one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the steps of the machine learning-based current transformer saturation detection method as described in any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the steps of the machine learning-based current transformer saturation detection method as described in any one of claims 1-5.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the machine learning-based current transformer saturation detection method as described in any one of claims 1-5.

Images