Power distribution network grounding fault line selection method based on image level

Abstract

The invention relates to a power distribution network grounding fault line selection method based on an image level, and the method comprises the following steps: obtaining historical waveform data of each line through power grid data, building a power distribution network fault feature system, and carrying out the training to obtain a grounding fault line selection model; acquiring real-time waveform data, uploading the real-time waveform data to the grounding fault line selection model, predicting a line state, and stopping power supply and positioning a line fault position when a line fault occurs; fault waveform data during line fault classification are analyzed by using a deep learning method, and the fault type of the line is obtained; and the line fault position and the fault type are sent to a power grid system, and the power grid system makes a fault decision for the line. The fault state of each line in the power distribution network system can be monitored in real time through the grounding fault line selection model, then the fault type of the line is obtained by using the deep learning method and the random forest classifier, and finally the purpose of making a fault decision on the line rapidly, fully automatically and accurately is achieved.

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and in particular to an image-based ground fault line selection method for a distribution network. Background Art

[0002] With the development of national new energy technology, the construction of distribution network system is also gradually progressing. At the same time, the complexity of the distribution network system, the increase in lines, and the diversification of operation modes have led to frequent distribution network failures. Among them, single-phase grounding faults are the most frequent faults, accounting for about 80% of all fault types. When a single-phase grounding fault occurs, it can run with the fault for 1 to 2 hours, but if it is in a fault state for a long time, it may cause line insulation damage, which will not only affect the economic benefits of the power company, but also endanger the safety of residents.

[0003] The fifth harmonic method is used to obtain the amplitude and phase of the fifth harmonic of the zero-sequence current of each line in the neutral point ungrounded system and the neutral point grounded system through the arc suppression coil to obtain the fault line selection criterion. The fifth harmonic fault signal is small and the accuracy requirement is high, so this method is only suitable as an auxiliary means for comprehensive line selection.

[0004] The prior art mentions utilizing transient information during arc extinguishing to establish a lumped parameter equivalent model of asymmetric line parameters. In practical applications, fault tracking, fault line selection and fault arc extinguishing are achieved by injecting and removing full compensation current. However, the prior art has limited ability to accurately identify transition resistance and track the degree of fault in real time during arc extinguishing, and this needs to be improved.

[0005] In order to solve the problem of limited ability to track the degree of fault in real time, this paper proposes a distribution network ground fault line selection method based on image level. This method uses the waveform data collected at the line end to compare the morphological differences of zero-sequence voltage and zero-sequence current under different lines according to the differences in waveform image morphology under different lines. The method uses an adaptive algorithm to analyze the specific fault type and the location of the fault, thereby completing the fault line selection.

[0006] In the detection of single-phase fault lines, the fifth harmonic component method in the steady-state method is the mainstream. Its principle is that when a single-phase grounding fault occurs, the distribution law of its fifth harmonic capacitor current is basically consistent with the fundamental capacitor current law of the neutral point ungrounded system. The fifth harmonic zero-sequence current on the fault line is basically equal to the sum of the fifth harmonic currents of the non-fault line in amplitude, and its phase lags behind the fifth harmonic zero-sequence voltage by 90°, which is completely opposite to the phase of the non-fault line. The fifth harmonic zero-sequence current in the non-fault line leads the fifth harmonic zero-sequence voltage by 90°. Based on this feature, fault line selection can be achieved.

[0007] However, in actual power grids, the fifth harmonic line selection method is affected by the quality of the power grid. When polluted by harmonics, the sensitivity and accuracy of grounding line selection will be reduced, and the applicability of this method will also decrease.

[0008] The method of fault line selection analysis at the image level is the perceptual hash algorithm. This method converts the fault waveform image of each line into a grayscale image through wavelet transformation. Figure 2 The values ​​are converted into a binary matrix as the hash value of the image. The hash value can be used to calculate the Hamming distance between the two feeders, and the image difference of the faulty line can be calculated through the Hamming distance.

[0009] This method reduces the algorithm execution time and can quickly complete the fault line selection in the single-phase grounding fault of the distribution network. However, this method is highly sensitive to images, and changes in the image structure may cause changes in the hash value, so it still has defects. Summary of the invention

[0010] The purpose of the present invention is to provide a distribution network ground fault line selection method based on image level, which can automatically monitor the abnormal conditions of the lines in the distribution network system and provide solution decisions, save human resources and improve detection efficiency.

[0011] The present invention is implemented by the following technical solution: a distribution network ground fault line selection method based on an image level, which comprises the following steps:

[0012] Step 1, obtain the historical waveform data of each line through the power grid data, process the historical waveform data through the wavelet transform analysis method, establish the distribution network fault feature system as training data, and use the training data to train the convolutional neural network AlexNet to obtain the ground fault line selection model;

[0013] Step 2, real-time waveform data of each line is obtained in real time, and the real-time waveform data is processed by wavelet transform analysis method and then uploaded to the ground fault line selection model. The ground fault line selection model predicts the state of each line, stops the power supply of the line when a line fails, and locates the line fault position;

[0014] Step 3, using a deep learning method to analyze and classify the fault waveform data when the line fails, and obtain the type of fault of the line;

[0015] Step 4: Send the line fault location and fault type to the power grid system, which makes a fault decision for the line, promptly restores power supply to non-line fault locations, and dispatches maintenance personnel to maintain the line fault location.

[0016] Compared with the previous technology, the beneficial effects of the present invention are:

[0017] 1. The present invention can monitor the fault status of each line in the distribution network system in real time through the ground fault line selection model, and then use the deep autoencoder and random forest classifier to analyze and obtain the fault type of the line, and quickly stop the power supply to the faulty line, saving manpower and material resources, and improving the monitoring efficiency with very high accuracy.

[0018] 2. The present invention can cover the entire distribution network and collect historical data for model analysis and learning, so that faults in a wider range can be detected, and the sample size is large enough to greatly improve the comprehensiveness and accuracy of fault detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is the main flow chart of the present invention;

[0020] Figure 2 Flowchart for establishing the ground fault line selection model;

[0021] Figure 3 Flowchart for analyzing fault types for deep learning methods. DETAILED DESCRIPTION

[0022] The present invention is described in detail below in conjunction with the accompanying drawings, but the protection content of the present invention is not limited to the following:

[0023] like Figures 1 to 3 As shown in FIG. 1 , a distribution network ground fault line selection method based on an image level includes the following steps:

[0024] Step 1, obtain the historical waveform data of each line through the power grid data, process the historical waveform data through the wavelet transform analysis method, establish the distribution network fault feature system as training data, and use the training data to train the convolutional neural network AlexNet to obtain the ground fault line selection model;

[0025] Step 2, real-time waveform data of each line is obtained in real time, and the real-time waveform data is processed by wavelet transform analysis method and then uploaded to the ground fault line selection model. The ground fault line selection model predicts the state of each line, stops the power supply of the line when a line fails, and locates the line fault position;

[0026] Step 3, using a deep learning method to analyze and classify the fault waveform data when the line fails, and obtain the type of fault of the line;

[0027] Step 4: Send the line fault location and fault type to the power grid system, which makes a fault decision for the line, promptly restores power supply to non-line fault locations, and dispatches maintenance personnel to maintain the line fault location.

[0028] Furthermore, the power grid system can make a fault decision for the line according to the fault type of the line.

[0029] The present invention can monitor the fault status of each line in the distribution network system in real time through the ground fault line selection model, and then use the deep learning method to analyze the waveform data of the fault section to obtain the fault type of the line, and quickly respond to the faulty line by stopping power supply. The entire process is automatically monitored by the algorithm model, which saves manpower and has very high accuracy.

[0030] When a line fault occurs, the presence of the fault must first be detected. Parameters such as zero-sequence current and zero-sequence voltage are usually used to identify ground faults in the distribution network. It should be noted that the ground fault line selection model can analyze waveform data and calculate the amount of waveform mutation based on waveform changes (usually refers to sudden changes in voltage or current, such as drops, jumps or fluctuations, which may be indications of faults);

[0031] Mutation amount C t The calculation formula is: C t =max{0,C t-1 +(x t -μ)-k}; where x t is real-time waveform data, C t is the cumulative sum of the waveform data C, t is the time, μ is the expected value of the current or voltage, and k is the mutation amount C t control limits.

[0032] When C t When >0 and increasing, it indicates that mutation may occur and further analysis or measures are needed;

[0033] When C t =0, it means there is no significant change and the line is in normal state.

[0034] Furthermore, the historical waveform data includes waveform data when each line is faulty and waveform data when the line is not faulty;

[0035] The waveform data are arranged in chronological order. The waveform data are waveform data of voltage or current changing with time, including waveforms before and after line faults. Because the waveform data of the line will produce zero-sequence components when a line fault occurs, the ground fault line selection model receives and processes the waveform data of each line in real time, and compares the waveform data of each section during the fault period according to the characteristic that the characteristics of the fault line are opposite to those of the normal line, so as to find the fault line.

[0036] Furthermore, the wavelet transform analysis method in step 1 and step 2 includes the following steps:

[0037] Step S1, performing denoising processing on historical waveform data or real-time waveform data to reduce the influence of high-frequency noise;

[0038] Step S2, selecting Haar wavelet and determining wavelet decomposition level j according to the frequency characteristics of historical waveform data or real-time waveform data;

[0039] Step S3, performing wavelet decomposition on the historical waveform data or the real-time waveform data to decompose it into approximate parts A of different scales. j and details D j ;

[0040] Step S4, respectively calculate the approximate part A j and details D j The characteristic quantity E of the wavelet coefficients at each level j ; Feature quantity E j The calculation formula is: or

[0041] Among them, D j (k) is the kth wavelet coefficient of the jth level detail signal; A j (k) is the kth wavelet coefficient of the jth level approximation signal; N j is the number of wavelet coefficients in the j-th level detail signal or the j-th level approximation signal;

[0042] Step S5: Decompose the feature values ​​E obtained by wavelet decomposition at each level j Arrange in chronological order and combine into multiple waveform feature matrices based on time-frequency analysis. Different waveform feature matrices correspond to different frequency bandwidths, which fully characterize the time-frequency characteristics of the waveform.

[0043] Step S6, normalizing the values ​​in each waveform feature matrix so that all values ​​are scaled to a range of 0 to 255, and mapped to pixel values ​​of the image to generate multiple pixel matrix images.

[0044] Furthermore, within the range of 0 to 255, the larger the pixel value, the brighter it is in the pixel matrix diagram, and the smaller the pixel value, the darker it is in the pixel matrix diagram.

[0045] Furthermore, the denoising process is implemented by using a median filtering method, which is a prior art and will not be described in detail.

[0046] Furthermore, in step 1, establishing a distribution network fault characteristic system includes the following steps:

[0047] Step 101, establishing a multi-output classification label β according to the fault status of the historical waveform data;

[0048] Step 102 , a pixel matrix image obtained from historical waveform data when the line is faulty is marked as β=1, and a pixel matrix image obtained from historical waveform data when the line is not faulty is marked as β=0.

[0049] It should be noted that the purpose of the distribution network ground fault feature system is to analyze the feature set of various fault types in the historical waveform data (here the features are the waveform features of current and voltage). With the help of these features, the type, location and impact of the fault can be identified, so that appropriate measures can be taken to isolate and repair the fault. The convolutional neural network AlexNet can learn a large amount of historical waveform data, and then automatically extract features that can effectively characterize the fault. These features can include instantaneous changes in current and voltage, waveform features, and other statistics, and finally predict / determine whether the line has a fault.

[0050] like Figure 2 As shown, in step 1, the steps of using training data to train the convolutional neural network AlexNet to obtain a ground fault line selection model include:

[0051] Step 103, defining the architecture of the convolutional neural network AlexNet, including the number of neurons in the input layer, hidden layer, and output layer and their activation functions;

[0052] Step 104, randomly initialize the weight parameters in the model, then input the distribution network grounding fault feature system, perform forward propagation through the convolutional neural network AlexNet, and obtain the output;

[0053] Step 105, introduce the cross entropy function f loss Calculate the output of the convolutional neural network AlexNet With the actual label y i The difference between loss The calculation formula is Among them, y i is the correct probability of the i-th type of fault in the actual label; is the predicted probability of the i-th type of fault under the fault feature system;

[0054] Step 106, calculate the cross entropy function f through the back propagation algorithm loss The gradient of each parameter is calculated, and then the gradient descent method is used to update the weight parameters in the convolutional neural network AlexNet to minimize the cross entropy function f loss ;

[0055] Step 107, repeat the forward propagation and cross entropy function f loss , back propagation and weight parameter update process until the performance of the convolutional neural network AlexNet reaches the expected level, and then all parameters are saved to obtain the ground fault line selection model.

[0056] It should be noted that after the pixel matrix image of the distribution network grounding fault feature system is input into the convolutional neural network AlexNet, it is first convolved through the first convolutional layer. The convolutional layer contains multiple convolution kernels, each of which extracts different image features, such as edges and colors, and then the maximum pooling layer is used to reduce the size of the features and retain important information; after the convolutional layer is processed, the pixel matrix image is flattened and input into the fully connected layer; finally, the Softmax layer outputs the probability distribution based on the multi-output classification label β.

[0057] In step 105, by calculating the cross entropy f loss A scalar value can be obtained, which represents the predicted distribution of the fault feature system and the actual label y i The difference (degree of deviation) between the (ideal model) distributions. Cross entropy f loss The smaller it is, the more the fault feature system is consistent with the actual label y i The deviation is small, and the system performance is good; the cross entropy f loss The larger the value is, the greater the deviation between the fault feature system and the ideal model is, and the model may need to be adjusted or improved. That is, the cross entropy f is calculated. loss The advantage is that it helps optimize the model, and the cross entropy f loss The value of can also be used as an important indicator to evaluate the performance of the model. By comparing the cross entropy values ​​of different models, the ground fault line selection model with better performance can be selected.

[0058] Further, the deep learning method includes a deep autoencoder and a random forest classifier. The deep autoencoder is good at automatically extracting high-level features from the data without manually defining the features. This feature extraction is adaptive because the model can adjust and learn the feature representation that is most helpful for the classification task according to the different input data. Therefore, the deep autoencoder can automatically extract discriminative features.

[0059] Random forest is an ensemble learning method based on decision trees, which has strong robustness and generalization ability. It reduces the overfitting problem that a single model may cause by building multiple decision trees and combining their results. Random forest performs well in processing high-dimensional data and can handle complex features extracted by deep autoencoders. By adaptively combining the judgments of multiple trees, random forest can flexibly respond to different fault characteristics and further improve the intelligent adaptive ability of the model.

[0060] Therefore, combining the two allows the model to not only adapt to different fault data, but also provide stable and reliable classification results in various situations.

[0061] Step 3 includes the following steps:

[0062] Step 301, sending a plurality of pixel matrix images A obtained by processing historical waveform data by a wavelet transform analysis method to a deep autoencoder model;

[0063] Step 302, training a deep autoencoder model, which is composed of a multi-layer encoder and a decoder, wherein the encoder compresses the pixel matrix image A layer by layer and maps the high-dimensional data of the pixel matrix image A to a low-dimensional feature space;

[0064] Step 303, the decoder restores the low-dimensional data A mapped into the low-dimensional feature space to high-dimensional data of the pixel matrix image A, minimizing the reconstruction error so that the low-dimensional data A can effectively represent the features of the pixel matrix image;

[0065] Step 304, after the training is completed, save the encoder in the deep autoencoder model; and establish a multi-output classification label α for the low-dimensional data A.

[0066] Step 3 also includes the following steps:

[0067] Step 305, construct a random forest classifier, read the low-dimensional data A and the multi-output classification label α as classification features;

[0068] Step 306, multiple pixel matrix images B obtained by processing the real-time waveform data by the wavelet transform analysis method are sent to the encoder,

[0069] Step 307, the encoder compresses the pixel matrix image B layer by layer, maps the high-dimensional data B of the pixel matrix image B to a low-dimensional feature space, and then sends the low-dimensional data B mapped into the low-dimensional feature space to the random forest classifier;

[0070] Step 308 , the low-dimensional data B is classified by voting of multiple decision trees in the random forest to obtain the value of the multi-output classification label α and confirm the fault type of the line fault location.

[0071] Furthermore, the fault types include metallic grounding, low resistance grounding and high resistance grounding within 800Ω;

[0072] When the metallic property is grounded, the multi-output classification label of the low-dimensional feature α = 0;

[0073] When the small resistor is grounded, the multi-output classification label of the low-dimensional feature α = 1;

[0074] When a large resistor within 800Ω is grounded, the multi-output classification label of the low-dimensional feature α=2.

[0075] In the present invention, the final fault decision can also cooperate with the monitoring camera to realize the visual monitoring of the line fault area, so that the operation and maintenance personnel can observe whether the fault decision is made correctly.

[0076] Although the present invention is illustrated and described by specific embodiments and their alternatives, it should be understood that various changes and modifications within the spirit and scope of the present invention can be implemented. Therefore, it should be understood that the present invention is not limited in any sense except by the limitations of the attached claims and their equivalents.

Claims

1. A distribution network ground fault line selection method based on image level, characterized in that: It includes the following steps: Step 1, obtain the historical waveform data of each line through the power grid data, process the historical waveform data through the wavelet transform analysis method, establish the distribution network fault feature system as training data, and use the training data to train the convolutional neural network AlexNet to obtain the ground fault line selection model; Step 2, real-time waveform data of each line is obtained in real time, and the real-time waveform data is processed by wavelet transform analysis method and then uploaded to the ground fault line selection model. The ground fault line selection model predicts the state of each line, stops the power supply of the line when a line fails, and locates the line fault position; Step 3, using a deep learning method to analyze and classify the fault waveform data when the line fails, and obtain the type of fault of the line; Step 4: Send the line fault location and fault type to the power grid system, which makes a fault decision for the line, promptly restores power supply to non-line fault locations, and dispatches maintenance personnel to maintain the line fault location.

2. The method for selecting ground fault lines in distribution network based on image level according to claim 1, characterized in that: The historical waveform data includes waveform data when each line fails and waveform data when the line is not failed; And the waveform data are arranged in chronological order.

3. The method for selecting ground fault lines in distribution network based on image level according to claim 2, characterized in that: The wavelet transform analysis method in step 1 and step 2 comprises the following steps: Step S1, performing denoising processing on historical waveform data or real-time waveform data to reduce the influence of high-frequency noise; Step S2, selecting Haar wavelet and determining wavelet decomposition level j according to the frequency characteristics of historical waveform data or real-time waveform data; Step S3, performing wavelet decomposition on the historical waveform data or the real-time waveform data to decompose it into approximate parts A of different scales. j and details D j ; Step S4, respectively calculate the approximate part A j and details D j The characteristic quantity E of the wavelet coefficients at each level j ; Feature quantity E j The calculation formula is: or Among them, D j (k) is the kth wavelet coefficient of the jth level detail signal; A j (k) is the kth wavelet coefficient of the jth level approximation signal; N j is the number of wavelet coefficients in the j-th level detail signal or the j-th level approximation signal; Step S5: Decompose the feature values ​​E obtained by wavelet decomposition at each level j Arrange in chronological order and combine into multiple waveform feature matrices based on time-frequency analysis. Different waveform feature matrices correspond to different frequency bandwidths, which fully characterize the time-frequency characteristics of the waveform. Step S6, normalizing the values ​​in each waveform feature matrix so that all values ​​are scaled to a range of 0 to 255, and mapped to pixel values ​​of the image to generate multiple pixel matrix images.

4. The method for selecting ground fault lines in distribution network based on image level according to claim 3 is characterized in that: In step 1, establishing a distribution network fault characteristic system includes the following steps: Step 101, establishing a multi-output classification label β according to the fault status of the historical waveform data; Step 102 , a pixel matrix image obtained from historical waveform data when the line is faulty is marked as β=1, and a pixel matrix image obtained from historical waveform data when the line is not faulty is marked as β=0.

5. The method for selecting ground fault lines in distribution network based on image level according to claim 4, characterized in that: In step 1, the steps of using the training data to train the convolutional neural network AlexNet to obtain the ground fault line selection model include: Step 103, defining the architecture of the convolutional neural network AlexNet, including the number of neurons in the input layer, hidden layer, and output layer and their activation functions; Step 104, randomly initialize the weight parameters in the model, then input the distribution network grounding fault feature system, perform forward propagation through the convolutional neural network AlexNet, and obtain the output; Step 105, introduce the cross entropy function f loss Calculate the output y of the convolutional neural network AlexNet predicti With the actual label y i The difference between loss The calculation formula is Among them, y i is the correct probability of the i-th type of fault in the actual label; y predicti is the predicted probability of the i-th type of fault under the fault feature system; Step 106, calculate the cross entropy function f through the back propagation algorithm loss The gradient of each parameter is calculated, and then the gradient descent method is used to update the weight parameters in the convolutional neural network AlexNet to minimize the cross entropy function f loss ; Step 107, repeat the forward propagation and cross entropy function f loss , back propagation and weight parameter update process until the performance of the convolutional neural network AlexNet reaches the expected level, and then all parameters are saved to obtain the ground fault line selection model.

6. The method for selecting a ground fault line in a distribution network based on an image layer according to claim 3 is characterized in that ,Step 3 includes the following steps: Step 301, sending a plurality of pixel matrix images A obtained by processing historical waveform data by a wavelet transform analysis method to a deep autoencoder model; Step 302, training a deep autoencoder model, which is composed of a multi-layer encoder and a decoder, wherein the encoder compresses the pixel matrix image A layer by layer and maps the high-dimensional data of the pixel matrix image A to a low-dimensional feature space; Step 303, the decoder restores the low-dimensional data A mapped into the low-dimensional feature space to high-dimensional data of the pixel matrix image A, minimizing the reconstruction error so that the low-dimensional data A can effectively represent the features of the pixel matrix image; Step 304, after the training is completed, save the encoder in the deep autoencoder model; and establish a multi-output classification label α for the low-dimensional data A.

7. The method for selecting a ground fault line in a distribution network based on an image layer according to claim 6, characterized in that ,Step 3 also includes the following steps: Step 305, construct a random forest classifier, read the low-dimensional data A and the multi-output classification label α as classification features; Step 306, sending a plurality of pixel matrix images B obtained by processing the real-time waveform data by the wavelet transform analysis method to the encoder, Step 307, the encoder compresses the pixel matrix image B layer by layer, maps the high-dimensional data B of the pixel matrix image B to a low-dimensional feature space, and then sends the low-dimensional data B mapped into the low-dimensional feature space to the random forest classifier; Step 308 , the low-dimensional data B is classified by voting of multiple decision trees in the random forest to obtain the value of the multi-output classification label α and confirm the fault type of the line fault location.

8. The method for selecting ground fault lines in distribution network based on image level according to claim 7, characterized in that: The fault types include metallic grounding, low resistance grounding and high resistance grounding within 800Ω; When the metallic property is grounded, the multi-output classification label of the low-dimensional feature α = 0; When a small resistor is grounded, the multi-output classification label of the low-dimensional feature α = 1; When a large resistor within 800Ω is grounded, the multi-output classification label of the low-dimensional feature α=2.

Images