Distribution transformer excitation surge current identification method and system based on Sym5 wavelet transform and XGBoost model, and storage medium

Abstract

The invention relates to the technical field of distribution transformers, in particular to a distribution transformer excitation surge current identification method and system based on Sym5 wavelet transform and an XGBoost model and a storage medium, and the method comprises the steps: carrying out the three-layer multi-resolution decomposition of an original current signal I (t) through employing a Sym5 wavelet basis, obtaining a D3-layer detail component, and extracting the time-frequency features of the D3-layer detail component; extracting a time domain feature and a frequency domain feature of the original current signal I (t); constructing a multi-dimensional feature vector set based on the time-frequency features, the time-domain features and the frequency-domain features; and normalizing the multi-dimensional feature vector set, and inputting the normalized multi-dimensional feature vector set into an XGBoost model for result prediction. According to the method, through collaborative optimization of Sym5 wavelets and XGBoost, feature positioning deviation caused by asymmetry of wavelet bases in the prior art is overcome, and meanwhile, the adaptability to complex harmonic scenes is enhanced.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of distribution transformers, and particularly relates to a distribution transformer excitation inrush current identification method and system based on a Sym5 wavelet transform and an XGBoost model and a storage medium. BACKGROUND

[0002] During the closing process of a distribution transformer, the steady-state magnetic flux generated by the residual magnetism of the core and the external grid voltage is superimposed to form a transient magnetic bias to achieve smooth transition of the magnetic flux. When the closing angle is 0° or 180°, the saturation of the core magnetic flux will trigger an excitation inrush current of 6-8 times the rated current. Although the inrush current is short in duration, its high amplitude is easy to be misjudged as an internal fault current by the main protection, resulting in protection misoperation, regional power outage and decrease in action accuracy. In the prior art, a Chinese patent application with the publication number CN107765065A proposes an identification method based on a fundamental wave attenuation factor, but the algorithm is sensitive to noise and the increase of the closing angle will significantly reduce the sensitivity of the criterion; a Chinese patent application with the publication number CN110568248A further combines a differential current fundamental wave phase and a second harmonic component for comprehensive identification, but there is still a problem of insufficient adaptability in a complex harmonic scenario.

[0003] To sum up, there is an urgent need for a distribution transformer excitation inrush current identification method and system based on a Sym5 wavelet transform and an XGBoost model and a storage medium to solve the problems in the prior art. SUMMARY

[0004] The application aims to provide a distribution transformer excitation inrush current identification method based on a Sym5 wavelet transform and an XGBoost model, and aims to solve the problems of the excitation inrush current identification method in the prior art. The specific technical solutions are as follows:

[0005] A distribution transformer excitation inrush current identification method based on a Sym5 wavelet transform and an XGBoost model, comprising:

[0006] The original current signal I(t) is decomposed into D3 layer detail components by using a Sym5 wavelet base for three-layer multi-resolution decomposition, and the time-frequency features of the D3 layer detail components are extracted;

[0007] The time-domain features and the frequency-domain features of the original current signal I(t) are extracted;

[0008] A multi-dimensional feature vector set is constructed based on the time-frequency features, the time-domain features and the frequency-domain features;

[0009] After the multi-dimensional feature vector set is normalized, the normalized multi-dimensional feature vector set is input into the XGBoost model to predict the result.

[0010] Preferably, the time-frequency features of the D3 layer detail components are specifically as follows:

[0011] The extreme point sequence of the D3 layer detail component is obtained, the amplitude change gradient between adjacent extreme points is calculated, and the absolute value mean μ of the amplitude change gradient is counted G , variance , and positive and negative gradient proportion R ± ;

[0012] The extreme point time interval sequence in the extreme point sequence is extracted, the probability density function of the extreme point time interval is fitted, and the distribution kurtosis K and skewness S Δt are extracted Δt ;

[0013] The D3 layer detail component is segmented by time window, and the energy entropy E entropy of each window is calculated entropy , the energy entropy E E of all windows is counted, and the mean μ and standard deviation σ E are obtained;

[0014] The time domain features and frequency domain features of the original current signal I(t) are extracted, specifically:

[0015] The mean μ, standard deviation σ, and skewness S of the current sampling value in the original current signal I(t) are obtained I ; I I ;

[0016] The original current signal I(t) is subjected to fast Fourier transform, the amplitudes of the fundamental wave, 2nd harmonic, 3rd harmonic, 4th harmonic, and 5th harmonic are extracted, the energy proportion of each harmonic is calculated, and the maximum harmonic energy proportion and total harmonic distortion rate THD are obtained ;

[0017] The constructed multi-dimensional feature vector set is

[0018] Preferably, the extreme point sequence {(t i ,A i )} is obtained by detecting the extreme points of the D3 layer detail component D3(t) according to formula (2), wherein i=1, 2,..., N, t i represents the time position of the i-th extreme point, A i represents the absolute value of the amplitude of the i-th extreme point, and N represents the number of extreme points ;

[0019] |D3(t i )|>max(|D3(t i-10 )|,...,|D3(t i-1 )|,|D3(t i+1 )|,...,|D3(t i+10 )|) and |D3(t i| > 0.15 I base (2),

[0020] wherein, I base is the current base value; t i-10 is the time window; |D3(t i+10 )| is the absolute value of the amplitude of the D3 layer detail component at time t i ; |D3(t i )| is the absolute value of the amplitude of the D3 layer detail component at time t i-5 . i-5

[0021] Preferably, the amplitude variation gradient between adjacent extreme points in the extreme point sequence is calculated according to formula (3):

[0022]

[0023] wherein, G i is the amplitude variation gradient between the i-th extreme point and the i+1-th extreme point.

[0024] Preferably, the energy entropy E entropy of the window is calculated according to formula (4) and formula (5):

[0025]

[0026] wherein, M is the number of sampling points in the window, P a is the normalized energy proportion of the a-th sampling point, |D3(a)| is the absolute value of the amplitude of the a-th sampling point in the D3 layer detail component, and |D3(m)| is the absolute value of the amplitude of the m-th sampling point in the D3 layer detail component.

[0027] Preferably, the mean value μ I , the standard deviation σ I and the skewness S I of the current sampling values in the original current signal I(t) are obtained according to formula (6), formula (7) and formula (8):

[0028]

[0029] wherein, T is the total number of sampling points of the signal, and i(t) is the current sampling value at time t.

[0030] Preferably, the energy proportion of each harmonic is calculated according to formula (9):

[0031]

[0032] The total harmonic distortion rate THD is calculated according to formula (10):

[0033] ​

[0034] Wherein: 1 represents the fundamental wave, 2 represents the 2nd harmonic, 3 represents the 3rd harmonic, 4 represents the 4th harmonic, 5 represents the 5th harmonic, |H n |H is the absolute value of the amplitude of the harmonic represented by n. q |H is the absolute value of the amplitude of the fundamental wave or harmonic represented by q.

[0035] Preferably, the three-layer multi-resolution decomposition structure includes a D1 detail component layer, a D2 detail component layer, a D3 detail component layer and an A3 approximation component layer, and the decomposition scale parameter a=2 j Separate signal frequency bands layer by layer through down-sampling; wherein the D1 detail component layer focuses on 0.5-1kHz high-frequency noise and transient peak, the D2 detail component layer extracts 0.25-0.5kHz frequency band to capture the inrush current intermittent angle feature, the D3 detail component layer is aimed at the 62.5-125Hz target frequency band, the A3 approximation component layer retains the 0-62.5Hz low-frequency fundamental wave component, and j=1, 2, 3.

[0036] The application further provides a power distribution transformer inrush current identification system based on Sym5 wavelet transform and an XGBoost model, including a memory and a processor, the memory stores a computer program, and the computer program is executed to perform the method.

[0037] The application further provides a storage medium, which stores a computer program, and the computer program is executed to perform the method.

[0038] The technical scheme of the application has the following beneficial effects:

[0039] The identification method of the application utilizes the approximate symmetry and 5th-order vanishing moment characteristics of the Sym5 wavelet, optimizes the multi-resolution decomposition precision of the current signal, and suppresses the time-domain positioning deviation in the high-frequency transient feature extraction process from the root; in combination with the D3 layer detail component of the Sym5 wavelet decomposition, the amplitude gradient, time distribution density and energy entropy value of the extreme value sequence are extracted, and the time-domain statistical features and frequency-domain harmonic content of the original current signal are fused to construct a multi-dimensional feature vector set; finally, through the gain splitting mechanism and feature importance weighting strategy of the XGBoost model, high-discrimination features are adaptively selected, a nonlinear classification boundary of the inrush current and the fault current is established, and the problem of insufficient adaptability of traditional shallow models to complex harmonic scenes is solved. Through the collaborative optimization of the Sym5 wavelet and the XGBoost, the application solves the feature positioning deviation problem caused by the phase distortion of the wavelet base in the traditional technology, and the defect of insufficient adaptability of a single criterion to complex harmonic scenes, and provides a technical scheme with time-frequency analysis precision and machine learning generalization ability for inrush current identification.

[0040] The recognition method of the application solves the problem that the traditional method usually relies on single or few features (such as the second harmonic ratio, the waveform skewness), and is susceptible to noise and harmonic interference. The multi-dimensional feature vector set F of the application has the following characteristics:

[0041] 1) The extreme value amplitude gradient (μ G , R + ) reflects the high-frequency mutation characteristics of the inrush waveform, the extreme value time interval distribution (K Δt , S Δt ) captures the periodic discontinuous mode of the inrush, the energy entropy mean (μ E ) and the energy entropy standard deviation σ E quantify the energy distribution disorder of the transient component, and the inrush has a higher entropy value due to non-stationarity. The D3 layer wavelet detail component feature can solve the problem of insufficient extraction of high-frequency transient characteristics in the traditional method.

[0042] 2) The mean (μ I ), standard deviation (σ I ) and skewness (S I ) of the current sampling value depict the overall symmetry and amplitude distribution of the waveform. The inrush shows significant skewness due to the magnetic saturation, and the time-domain statistics of the original signal enhance the sensitivity to the overall distortion of the waveform.

[0043] 3) The maximum harmonic energy ratio The total harmonic distortion rate THD can comprehensively evaluate the harmonic content and stability. The THD of the inrush is usually > 20%, and the frequency domain harmonic feature can overcome the misjudgment risk caused by abnormal harmonic content (such as fault current also containing high-order harmonics) due to new energy grid connection.

[0044] The recognition method of the application introduces a multi-dimensional feature vector set F to achieve the following technical effects:

[0045] Precise separation of high-frequency transient characteristics: Sym5 wavelet D3 layer decomposition + extreme value gradient analysis, improving the time-domain positioning accuracy (±0.2ms);

[0046] Multi-dimensional cross verification: time domain (waveform distortion), frequency domain (harmonic), and energy (entropy value) triple criteria complement each other, reducing the misjudgment rate;

[0047] Adaptive machine learning: XGBoost automatically optimizes feature weights, adapts to different closing angles, residual magnetism, harmonic interference and other complex conditions, significantly improving the classification accuracy of the inrush current and the fault current.

[0048] In addition to the purposes, features and advantages described above, the application has other purposes, features and advantages. The application will be further described below with reference to the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0049] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are intended to explain the present invention and do not constitute an undue limitation of the present invention. In the accompanying drawings:

[0050] Figure 1 It is a schematic diagram of Sym5 wavelet decomposition principle and feature extraction;

[0051] Figure 2 This is a flow chart of gain optimization classification model training and real-time discrimination based on XGBoost;

[0052] Figure 3 It is a simulation system model of the identification method of the present invention. DETAILED DESCRIPTION

[0053] To facilitate understanding of the present invention, the present invention will be described more fully below, along with preferred embodiments thereof. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and comprehensive understanding of the disclosure of the present invention.

[0054] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0055] Example 1:

[0056] like Figure 1 and Figure 2 As shown, this embodiment provides a method for identifying magnetizing inrush current of a distribution transformer based on Sym5 wavelet transform and XGBoost model, comprising the following steps:

[0057] The original current signal I(t) is decomposed into three layers with multi-resolution by using the Sym5 wavelet basis to obtain the D3 layer detail component and extract the time-frequency characteristics of the D3 layer detail component.

[0058] Extract the time domain characteristics and frequency domain characteristics of the original current signal I(t);

[0059] Construct a multidimensional feature vector set based on time-frequency features, time domain features and frequency domain features;

[0060] The multidimensional feature vector set is normalized and then input into the XGBoost model for result prediction.

[0061] Furthermore, the identification method of this embodiment will be described in detail below:

[0062] In this embodiment, a high-precision current transformer with a sampling rate of 10 kHz is preferably used to collect current signals in real time under the conditions of transformer no-load closing and short-circuit fault, and the original signals are normalized (amplitude is scaled to the interval [-1, 1]) to eliminate the influence of dimensional differences.

[0063] In this embodiment, the Sym5 wavelet basis function is selected, and its mathematical expression is:

[0064]

[0065] where h K is the 10th order filter coefficient of the Sym5 wavelet (derived from the Daubechies wavelet family), which suppresses high-frequency noise interference through its 5th order vanishing moment characteristic, and reduces phase shift using approximate symmetry. t represents the time independent variable of the wavelet function, and Sym (2t-k) represents the basis function of the mother wavelet at scale 2 and translation k, and Sym5 (t) represents a 5th order symmetric wavelet generated by linear combination of 10 filter coefficients.

[0066] As shown in Figure 1 , the three-layer multi-resolution decomposition structure in this embodiment includes a D1 detail component layer, a D2 detail component layer, a D3 detail component layer, and an A3 approximation component layer, which separates the signal frequency band by layer down-sampling through the decomposition scale parameter a=2 j (where j=1, 2, 3); wherein the D1 detail component layer focuses on 0.5-1 kHz high-frequency noise and transient peak, and is used to filter electromagnetic interference; the D2 detail component layer extracts 0.25-0.5 kHz frequency band to capture the intermittent angle characteristics of the excitation inrush current; the D3 detail component layer is aimed at the 62.5-125 Hz target frequency band, and accurately separates the high-frequency transient component containing the waveform distortion core feature; and the A3 approximation component layer retains the 0-62.5 Hz low-frequency fundamental component, which is used to exclude direct current bias interference in subsequent harmonic feature extraction.

[0067] Through the symmetric structure optimization of the Sym5 wavelet basis, the time offset of the extreme points of the detail component is strictly controlled within ±0.2 ms, and the time domain feature positioning accuracy is improved by 4 times, providing a high-fidelity time-frequency feature basis for subsequent analysis.

[0068] Preferably, the time-frequency features of the D3 layer detail component are specifically:

[0069] 1) Obtain the extreme point sequence of the D3 layer detail component, calculate the amplitude change gradient between adjacent extreme points, and calculate the absolute value mean G , variance and positive and negative gradient proportion R ± of the amplitude change gradient, which are specifically:

[0070] Perform extreme point detection on the detail component D3(t) of the D3 layer to obtain the extreme point sequence {(t i ,A i )}, where i = 1, 2, ..., N, t i Indicates the time position of the i-th extreme point, A i represents the absolute value of the amplitude of the ith extreme point, N represents the number of extreme points; the extreme point sequence {(t i ,A i )} characterizes the transient distortion characteristics related to the magnetizing inrush current in the detail component of the D3 layer. In this embodiment, the conditions for defining the extreme point are:

[0071] |D3(t i )|>max(|D3(t i-10 )|,...,|D3(t i-1 )|,|D3(t i+1 )|,...,|D3(t i+10 )|)和|D3(t i )|>0.15I base (2),

[0072] Among them, I base is the current base value (take the rated current effective value); t i-10 to t i+10 represents the time width of the time window (preferably 1 ms in this embodiment), which is used to eliminate random noise interference; |D3(t i )| indicates t i The absolute value of the amplitude of the detail component of the D3 layer at the moment |D3(t i-10 )| indicates t i-10 The absolute value of the amplitude of the detail component of the D3 layer at time instant.

[0073] Furthermore, we can obtain the extreme point sequence {(t i ,A i )} and then calculate the amplitude change gradient between adjacent extreme points:

[0074]

[0075] Among them, G i is the amplitude change gradient between the i-th extreme point and the i+1-th extreme point.

[0076] Get the extreme point sequence {(t i ,A i After calculating the amplitude change gradient between adjacent extreme points in the equation, the absolute value mean μ of the amplitude change gradient can be obtained statistically. G ,variance and the positive and negative gradient ratio R ± , as the amplitude dynamic change feature.

[0077] 2) Extract the extreme point time interval sequence from the extreme point sequence, fit the probability density function of the extreme point time interval, and extract the distribution kurtosis K Δt and skewness S Δt , specifically:

[0078] Calculate the time interval Δt between adjacent extreme points i =t i+1 -t i , fit the probability density function p(Δt) of the extreme point time interval through kernel density estimation, and extract the distribution kurtosis K Δt and skewness S Δt , characterizing the periodic characteristics of the discontinuity angle.

[0079] 3) Divide the D3 layer detail components into time windows and calculate the energy entropy E of each window entropy , count the energy entropy E of all windows entropy Get the mean μ E and standard deviation σ E , specifically:

[0080] Calculate the energy entropy E of the window according to formula (4) and formula (5) entropy :

[0081]

[0082] Among them, M is the number of sampling points in the window, P a represents the normalized energy proportion of the a-th sampling point, |D3(a)| represents the absolute value of the amplitude of the a-th sampling point in the detail component of the D3 layer, and |D3(m)| represents the absolute value of the amplitude of the m-th sampling point in the detail component of the D3 layer.

[0083] Count the energy entropy E of all windows entropy The mean μ can be obtained E and standard deviation σ E .

[0084] Preferably, the time domain features and frequency domain features of the original current signal I(t) are extracted, specifically:

[0085] 1) Obtain the mean μ of the current sampling value in the original current signal I(t) I , standard deviation σ I and skewness S I , specifically:

[0086]

[0087] Wherein, T represents the total number of sampling points of the signal, and i(t) represents the current sampling value at time t.

[0088] 2) Perform fast Fourier transform on the original current signal I(t), extract the amplitude of the fundamental wave (50Hz), 2nd harmonic, 3rd harmonic, 4th harmonic and 5th harmonic, calculate the energy proportion of each harmonic, and obtain the maximum harmonic energy proportion and total harmonic distortion THD;

[0089] Specifically, the energy proportion of each harmonic is expressed as:

[0090]

[0091] In formula (9), 1 represents the fundamental wave, 2 represents the second harmonic, 3 represents the third harmonic, 4 represents the fourth harmonic, 5 represents the fifth harmonic, |H n | is the absolute value of the amplitude of the harmonic represented by n, |H q | is the absolute value of the amplitude of the fundamental wave or harmonic represented by q.

[0092] After calculating the energy proportion of each harmonic according to formula (9), the maximum energy proportion can be selected. As features in the subsequent multidimensional feature vector set.

[0093] Furthermore, the total harmonic distortion THD is expressed as:

[0094]

[0095] Furthermore, a multidimensional feature vector set is constructed based on time-frequency features, time domain features, and frequency domain features. Specifically:

[0096] In getting the characteristic μ G 、 R ± , K Δt 、S Δt 、μ E , σ E 、μ I , σ I 、S I 、 After averaging and THD, the above features can be integrated into a 12-dimensional feature vector set.

[0097] μ in the 12-dimensional feature vector set F G 、 R ± , K Δt 、S Δt 、μ E and σ E is the time-frequency feature of the detail component of the D3 layer, μ I , σ I and S Iare time-domain features of the original current signal I(t), and THD is a time-domain feature and a frequency-domain feature of the original current signal I(t). Each feature in the 12-dimensional feature vector set F is normalized and then input into the XGBoost model for result prediction. In this embodiment, Z-score standardization is used to normalize each feature in the 12-dimensional feature vector set F, which is represented as:

[0098]

[0099] wherein f u represents any one feature in the 12-dimensional feature vector set F, μ u represents the mean of the feature f u in the input sample, and σ u represents the standard deviation of the feature f u in the input sample.

[0100] As shown in FIG. 8, the training and real-time discrimination of the XGBoost model will be described as follows: Figure 2

[0101] 1) Model training and feature adaptive screening

[0102] (1) Structure and operation mechanism of the XGBoost model

[0103] In this embodiment, the XGBoost algorithm is used to construct a classification model, and the core thereof is realized through a gradient boosting framework and a regularization mechanism to accurately identify the field inrush current. The model takes a classification and regression tree (CART) as a base learner, constructs multiple decision trees through recursive feature splitting, and stores a weight value at each leaf node of the tree for predicting the sample category. The final output of the model is a linear weighted sum of the prediction results of all base learners, i.e.,

[0104]

[0105] wherein η is a learning rate for controlling the contribution weight of a single tree; f k is a prediction function of the kth tree; F i is a feature vector input by the sample i, K represents the total number of base learners (decision trees), and represents the final prediction output value of the sample i.

[0106] The training target of the model is to minimize the sum of the loss function and the regularization term. For a binary classification task, a log loss function (Log Loss) is selected to measure the classification performance by the deviation between the prediction probability and the true label:

[0107]

[0108] y​i It represents the true label of sample i, where the value is 1 for magnetizing inrush current (positive class) and 0 for fault current (negative class).

[0109] At the same time, the complexity of the model is constrained by the regularization term:

[0110]

[0111] Where T is the number of leaf nodes in the tree, w is the node weight, γ and λ are the split gain threshold and L2 regularization coefficient, respectively, used to suppress overfitting. Ω(f k ) represents the kth decision tree f k The complexity penalty term is used to suppress model overfitting.

[0112] In the process of tree structure generation, the second-order Taylor expansion is used to accelerate the optimization, and the first-order gradient g is used to i and the second-order gradient h i Approximate loss function. Specifically, the node splitting of each tree aims to maximize the gain, and the splitting gain threshold = 0.5: to suppress abnormal splitting caused by harmonic interference. The gain formula is defined as the difference between the gradient statistics before and after the split minus the regularization penalty:

[0113]

[0114] Among them, G L and G R is the sum of the first-order gradients of the left and right child nodes after splitting, H L and H R is the sum of the corresponding second-order gradients, represents the square of the first-order gradient of the parent node before splitting, H parent Represents the sum of the second-order gradients of the parent node before splitting.

[0115] The optimal weight of the leaf node is calculated by the gradient statistics:

[0116]

[0117] in, represents the optimal weight of the jth feature or node, G j represents the first-order gradient statistic of the jth feature or node, H j Represents the second-order gradient statistic of the j-th feature or node.

[0118] To balance the model's generalization capabilities and computational efficiency, the maximum tree depth was set to 6 (balancing feature interaction depth with real-time requirements), and a forward-stepping algorithm was used to generate decision trees one by one. In each iteration, a new tree was generated based on the prediction residuals of the current model, and a weight update strategy was used to gradually converge to the optimal solution. The D3 layer features derived from the Sym5 wavelet decomposition were combined with XGBoost's gain splitting mechanism to form a closed loop. Sym5's ±0.2ms time-domain localization accuracy provided XGBoost with high-fidelity time-frequency features.

[0119] (2) Model training and feature adaptive screening process

[0120] Use MATLAB / SIMULINK to build a simulation system model, such as Figure 3 As shown in the figure, the simulation system model includes a power supply module, a transformer module, a fault injection module, and a high-precision signal acquisition system, providing high-reliability input for subsequent feature extraction and classification model training. A large amount of sample data is obtained through the simulation system model. First, a 12-dimensional feature vector set F is obtained, and a training set containing excitation inrush current samples and short-circuit fault samples is constructed. The sample label is defined as a binary classification variable (0: fault current, 1: excitation inrush current). The hyperparameters of the XGBoost model are initialized, including the learning rate (0.1), the maximum tree depth (6), the L2 regularization coefficient (1.0), and the minimum gain threshold for leaf node splitting (0.5) to balance the model complexity and generalization ability. Secondly, the weighted information gain and coverage dual indicators are used to evaluate the feature importance: by calculating the cumulative gain value of the feature when the tree node is split and its sample coverage ratio, the top 8-dimensional features with the highest contribution to classification (such as the extreme gradient mean, the maximum harmonic energy ratio, etc.) are screened, and redundant features are eliminated to reduce the model dimension. Compared to traditional neural networks, XGBoost uses a gain-splitting algorithm to dynamically evaluate feature importance and automatically select the top eight features that contribute most to classification, whereas neural networks rely on manually designed feature selection layers. This mechanism enables the model to maintain a classification accuracy of over 85% in complex harmonic scenarios.

[0121] 2) Gradient boosting decision tree generation and nonlinear rule optimization

[0122] Based on the screened feature set, a group of gradient boosting decision trees is constructed through iteration. A binary classification logarithmic loss function is defined as the optimization objective of the model, and the second-order Taylor expansion is used to approximate the loss function to calculate the optimal solution of the leaf node weight. A greedy algorithm is used to select the feature and threshold with the maximum split gain. Specifically, in each iteration, a new decision tree is generated based on the prediction residual of the current model, and the tree structure is dynamically optimized through the split gain formula to maximize the classification information purity at each split. After 100 iterations, the prediction results of each base learner are weighted and summed to form a nonlinear classification rule, thereby establishing a dynamic decision boundary between the magnetizing inrush current and the fault current in the high-dimensional feature space.

[0123] 3) Real-time discrimination and classification logic deployment

[0124] The real-time collected current signal is sequentially subjected to Sym5 wavelet decomposition and feature extraction to generate a standardized 12-dimensional feature vector set F, which is input into the pre-trained XGBoost model to calculate the classification probability. The discrimination threshold is set to 0.85 (an increase of 35% in anti-malfunction ability compared to the standard 0.5 threshold), and if the output probability is higher than the threshold, it is determined as a magnetizing inrush current and the protection device is blocked; otherwise, the fault protection action is triggered. To realize reliable discrimination of continuous signals, a sliding window mechanism is used for segmented processing of the current data, with a window length of 100 ms and a step length of 20 ms, ensuring that feature calculation and classification decision are completed within the time limit of relay protection action. When the XGBoost output probability is > 0.9 for 3 consecutive windows, the sliding window step length is automatically adjusted to 10 ms to realize fast response of the fault current and zero malfunction rate of inrush current identification.

[0125] The identification method of the embodiment solves the problem that traditional methods usually rely on a single or small number of features (such as the second harmonic ratio and waveform skewness), which are easily affected by noise and harmonic interference. In the 12-dimensional feature vector set F of the embodiment:

[0126] 1) The extreme value amplitude gradient (μ G 、 R + ) reflects the high-frequency abrupt change characteristics of the inrush current waveform, the extreme value time interval distribution (K Δt 、S Δt ) captures the unique periodic discontinuous pattern of the inrush current, the energy entropy mean (μ E ) and the energy entropy standard deviation σ E quantify the energy distribution disorder degree of the transient component, and the D3 layer wavelet detail component feature can solve the problem of insufficient extraction of high-frequency transient characteristics in traditional methods;

[0127] 2) The mean (μ I ), standard deviation (σ I ), and skewness (S I ) of the current sample valueThe overall symmetry and amplitude distribution of the waveform are characterized. The inrush current shows significant skewness due to magnetic saturation. The time-domain statistics of the original signal enhance the sensitivity to the overall waveform distortion.

[0128] 3) Maximum harmonic energy ratio Total harmonic distortion (THD) can comprehensively assess harmonic content and stability. The THD of inrush current is typically >20%. Frequency domain harmonic characteristics can overcome the risk of misjudgment due to abnormal harmonic content caused by renewable energy grid connection (e.g., fault current also containing high-order harmonics).

[0129] The recognition method of this embodiment introduces a 12-dimensional feature vector set F to achieve the following technical effects:

[0130] Accurately separate high-frequency transient features: Sym5 wavelet D3 layer decomposition + extreme value gradient analysis to improve time domain positioning accuracy (±0.2ms);

[0131] Multi-dimensional cross-validation: The three criteria of time domain (waveform distortion), frequency domain (harmonics), and energy (entropy) complement each other to reduce the false positive rate;

[0132] Adaptive machine learning: XGBoost automatically optimizes feature weights to adapt to complex operating conditions such as different closing angles, residual magnetism, and harmonic interference, significantly improving the classification accuracy of excitation inrush current and fault current.

[0133] Example 2:

[0134] This embodiment provides a distribution transformer excitation inrush current identification system based on Sym5 wavelet transform and XGBoost model. The system includes a memory and a processor. The memory stores a computer program. When the computer program is run, it executes the identification method in Example 1.

[0135] Example 3:

[0136] This embodiment provides a storage medium, in which a computer program is stored. When the computer program is executed, the identification method in Embodiment 1 is executed.

[0137] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A method for identifying magnetizing inrush current of distribution transformers based on Sym5 wavelet transform and XGBoost model, characterized in that: include: The original current signal I(t) is decomposed into three layers with multi-resolution by using the Sym5 wavelet basis to obtain the D3 layer detail component and extract the time-frequency characteristics of the D3 layer detail component. Extract the time domain characteristics and frequency domain characteristics of the original current signal I(t); Construct a multidimensional feature vector set based on time-frequency features, time domain features and frequency domain features; The multidimensional feature vector set is normalized and then input into the XGBoost model for result prediction.

2. The method according to claim 1, characterized in that The time-frequency features of the D3 layer detail components are: Get the extreme point sequence of the detail component of the D3 layer, calculate the amplitude change gradient between adjacent extreme points, and calculate the absolute value mean μ of the amplitude change gradient G ,variance And the positive and negative gradient ratio R ± ; Extract the extreme point time interval sequence from the extreme point sequence, fit the probability density function of the extreme point time interval, and extract the distribution kurtosis K Δt and skewness S Δt ; The D3 layer detail components are segmented into time windows and the energy entropy E of each window is calculated. entropy , count the energy entropy E of all windows entropy Get the mean μ E and standard deviation σ E ; Extract the time domain and frequency domain features of the original current signal I(t), specifically: Get the mean μ of the current sampling value in the original current signal I(t) I , standard deviation σ I and skewness S I ; Perform fast Fourier transform on the original current signal I(t), extract the amplitude of the fundamental wave, 2nd harmonic, 3rd harmonic, 4th harmonic and 5th harmonic, calculate the energy proportion of each harmonic, and obtain the maximum harmonic energy proportion and total harmonic distortion THD; The constructed multidimensional feature vector set is 3. The method according to claim 2, characterized in that According to formula (2), the extreme point detection of the detail component D3(t) of the D3 layer is performed to obtain the extreme point sequence {(t i ,A i )}, where i = 1, 2, ..., N, t i Indicates the time position of the i-th extreme point, A i represents the absolute value of the amplitude of the i-th extreme point, and N represents the number of extreme points; |D3(t i )| > max(|D3(t i-10 )|,..., |D3(t i-1 )|, |D3(t i+1 )|,..., |D3(t i+10 )|) and |D3(t i )| > 0.15I base (2), Among them, I base is the current base value; t i-10 to t i+10 Indicates the width of the time window; |D3(t i )| indicates t i The absolute value of the amplitude of the detail component of the D3 layer at the moment |D3(t i-5 )| indicates t i-5 The absolute value of the amplitude of the detail component of the D3 layer at time instant.

4. The method according to claim 3, characterized in that According to formula (3), the amplitude change gradient between adjacent extreme points in the extreme point sequence is calculated: Among them, G i is the amplitude change gradient between the i-th extreme point and the i+1-th extreme point.

5. The method according to claim 2, characterized in that Calculate the energy entropy E of the window according to formula (4) and formula (5) entropy : Among them, M is the number of sampling points in the window, P a represents the normalized energy proportion of the a-th sampling point, |D3(a)| represents the absolute value of the amplitude of the a-th sampling point in the detail component of the D3 layer, and |D3(m)| represents the absolute value of the amplitude of the m-th sampling point in the detail component of the D3 layer.

6. The method according to claim 2, characterized in that According to formula (6), formula (7) and formula (8), the mean value μ of the current sampling value in the original current signal I(t) is obtained: I , standard deviation σ I and skewness S I : Wherein, T represents the total number of sampling points of the signal, and i(t) represents the current sampling value at time t.

7. The method according to claim 2, characterized in that The energy proportion of each harmonic is calculated according to formula (9): Calculate the total harmonic distortion THD according to formula (10): Among them: 1 represents the fundamental wave, 2 represents the second harmonic, 3 represents the third harmonic, 4 represents the fourth harmonic, 5 represents the fifth harmonic, |H n | is the absolute value of the amplitude of the harmonic represented by n, |H q | is the absolute value of the amplitude of the fundamental wave or harmonic represented by q.

8. The method according to any one of claims 1 to 7, characterized in that The three-layer multi-resolution decomposition structure includes a D1 detail component layer, a D2 detail component layer, a D3 detail component layer and an A3 approximate component layer. j The signal frequency bands are separated by downsampling layer by layer; among them, the D1 detail component layer focuses on the 0.5-1kHz high-frequency noise and transient spikes, the D2 detail component layer extracts the 0.25-0.5kHz frequency band to capture the excitation inrush current discontinuity angle characteristics, the D3 detail component layer targets the 62.5-125Hz target frequency band, and the A3 approximate component layer retains the 0-62.5Hz low-frequency fundamental component, j = 1, 2, 3.

9. A distribution transformer excitation inrush current identification system based on Sym5 wavelet transform and XGBoost model, characterized in that: The method comprises a memory and a processor, wherein a computer program is stored in the memory, and when the computer program is executed, the method according to any one of claims 1 to 8 is executed.

10. A storage medium, characterized in that: The storage medium stores a computer program, which, when executed, executes the method according to any one of claims 1 to 8.

Images