Power distribution network single-phase earth fault line selection method and system

Abstract

The invention provides a power distribution network single-phase earth fault line selection method and system, and particularly relates to the technical field of power system fault detection, and the method carries out the signal decomposition and noise screening of a power distribution network zero-sequence current signal, and obtains a denoised signal. And then converting the de-noised signal into a two-dimensional matrix image by using Grubrum angle difference field transformation. The image is input into a teacher network containing a wavelet convolution module to extract a fault feature vector. And knowledge distillation is carried out on the feature vectors through a student network containing a depth separable convolution head module, a fault line discrimination result is output, and fault line selection operation is carried out according to the result. According to the method, the technical problem of how to construct a single-phase earth fault line selection method considering high-precision identification and low calculation complexity in a power distribution network environment with high noise interference and weak characteristic difference is solved, the fault detection precision is remarkably improved, and the calculation complexity is reduced; and the performance bottleneck of a traditional method in a complex noise and high-resistance fault scene is effectively solved.

Description

Technical Field

[0001] The present invention relates to the technical field of power system fault detection, and in particular, to a method and system for selecting a single-phase grounding fault line in a distribution network. Background Art

[0002] With the continuous expansion of the scale of China's distribution network, the non-effectively grounded neutral system (small current grounding system) is widely used due to its high power supply reliability. However, the single-phase grounding fault accounts for more than 80%. If the fault line cannot be quickly and accurately disconnected, it is extremely easy to cause interphase short circuits, equipment burnout and other accidents. However, the distribution network lines are complexly distributed, the fault signals are weak and easily affected by noise. Especially in the scenarios of high-resistance grounding and resonant grounding, the fault characteristics are very similar to those of non-fault lines. The traditional line selection methods based on artificial feature extraction or single signal processing are difficult to effectively distinguish the fault line, resulting in low fault recognition accuracy and high misjudgment rate, which seriously restricts the safe operation level of the distribution network.

[0003] The current technology mainly improves the line selection accuracy by signal denoising (such as wavelet transform) combined with deep learning models (such as ResNet, YOLO): signal processing technology attempts to suppress noise interference, but traditional denoising methods are difficult to distinguish noise from high-frequency fault components; time-frequency analysis tools (such as S transform) can extract time-domain and frequency-domain features, but they are not sensitive enough to phase differences; although complex neural networks can capture deep features, they have a large number of parameters (such as Faster R-CNN reaches 40.9M) and high computational complexity (such as SSD reaches 90.6G), and cannot be adapted to low-computing-power terminal devices. And lightweight models (such as MobileNet) have insufficient accuracy due to weak feature extraction ability, forming a sharp contradiction between accuracy and efficiency.

[0004] In summary, how to construct a single-phase grounding fault line selection method that takes into account high-precision recognition and low computational complexity in a distribution network environment with high noise interference and weak feature differences is an urgent problem to be solved. Summary of the Invention

[0005] The main purpose of the present invention is to provide a method and system for selecting a single-phase grounding fault line in a distribution network, so as to at least solve the technical problem of how to construct a single-phase grounding fault line selection method that takes into account high-precision recognition and low computational complexity in a distribution network environment with high noise interference and weak feature differences. Thus, through the collaborative design of noise suppression, feature enhancement and lightweight models, the fault detection accuracy is significantly improved and the computational complexity is reduced, effectively solving the performance bottleneck of traditional methods in complex noise and high-resistance fault scenarios.

[0006] To achieve the above object, the present invention provides a method and system for selecting a single-phase grounding fault line in a distribution network.

[0007] In a first aspect, the present invention provides a method for selecting a single-phase grounding fault line in a distribution network, including:

[0008] Perform signal decomposition and noise screening on the zero-sequence current signal of the distribution network to obtain a denoised zero-sequence current signal;

[0009] Process the denoised zero-sequence current signal through Gram angular difference field transformation to generate a corresponding two-dimensional matrix image;

[0010] Input the two-dimensional matrix image into a teacher network including a wavelet convolution module for feature extraction to obtain a fault feature vector;

[0011] Perform knowledge distillation on the fault feature vector through a student network including a depthwise separable convolution head module, and output a fault line discrimination result;

[0012] Perform a fault line selection operation according to the discrimination result.

[0013] Specifically, the performing signal decomposition and noise screening on the zero-sequence current signal of the distribution network to obtain a denoised zero-sequence current signal includes:

[0014] Decompose the zero-sequence current signal by using an improved adaptive noise complete ensemble empirical mode decomposition algorithm to obtain a plurality of intrinsic mode function components;

[0015] Calculate the multi-scale frequency entropy value of each intrinsic mode function component, and use the median of the entropy values as a threshold to screen out the noise-dominated components;

[0016] Denoise the noise-dominated components through a recursive least squares algorithm, and reconstruct them with the non-noise-dominated components to obtain the denoised zero-sequence current signal.

[0017] Specifically, the processing the denoised zero-sequence current signal through Gram angular difference field transformation to generate a corresponding two-dimensional matrix image includes:

[0018] Normalize the denoised zero-sequence current signal to the interval [-1, 1] to obtain a normalized signal

[0019] Map the normalized signal to a polar coordinate system to generate an angular component Wherein, is the i-th element of the normalized signal;

[0020] Construct a Gram angular difference field matrix according to the angular component to generate the two-dimensional matrix image, and the Gram angular difference field matrix satisfies the relational expression:

[0021]

[0022] Specifically, inputting the two-dimensional matrix image into a teacher network including a wavelet convolution module for feature extraction to obtain a fault feature vector includes:

[0023] Decompose the two-dimensional matrix image using the Haar wavelet basis function in the wavelet convolution module to separate the low-frequency component and the high-frequency component;

[0024] Perform 3×3 depth convolution kernel processing on the low-frequency component and the high-frequency component respectively, and reconstruct them into a fused feature map through inverse wavelet transform;

[0025] Output the fused feature map as the fault feature vector.

[0026] Specifically, performing knowledge distillation on the fault feature vector through a student network including a depthwise separable convolution head module and outputting a fault line discrimination result includes:

[0027] Perform depthwise separable convolution operations on the fault feature vector to generate a localization branch feature and a classification branch feature;

[0028] Pass the localization branch feature through two depthwise separable convolutions and a bounding box regression layer in sequence to output a fault location prediction result;

[0029] Pass the classification branch feature through two depthwise separable convolutions and a fully connected layer in sequence to output a fault class probability distribution;

[0030] Generate the fault line discrimination result according to the fault location prediction result and the fault class probability distribution.

[0031] Specifically, calculating the multi-scale frequency entropy value of each intrinsic mode function component includes:

[0032] Perform coarse-graining processing on each intrinsic mode function component to generate multi-scale subsequences;

[0033] Perform Fourier transform on each multi-scale subsequence to calculate the probability distribution and Shannon entropy value of the frequency domain signal;

[0034] Take the average of the Shannon entropy values of all scales to obtain the multi-scale frequency entropy value.

[0035] In a second aspect, the present invention provides a single-phase grounding fault line selection system for a distribution network. The line selection system applies the line selection method described in the first aspect. The line selection system includes:

[0036] A signal processing module for performing signal decomposition and noise screening processing on the zero-sequence current signal of the distribution network to obtain a denoised zero-sequence current signal;

[0037] The Gram transform module, connected to the signal processing module, is configured to perform Gram angular difference field transform processing on the denoised zero-sequence current signal to generate a corresponding two-dimensional matrix image;

[0038] The teacher network module, connected to the Gram transform module, is configured to input the two-dimensional matrix image into a teacher network including a wavelet convolution module for feature extraction to obtain a fault feature vector;

[0039] The student network module, connected to the teacher network module, is configured to perform knowledge distillation on the fault feature vector through a student network including a depthwise separable convolution head module and output a fault line discrimination result;

[0040] The line selection execution module, connected to the student network module, is configured to perform a fault line selection operation according to the discrimination result.

[0041] Specifically, the signal processing module includes:

[0042] The decomposition subunit is configured to decompose the zero-sequence current signal by using an improved adaptive noise complete ensemble empirical mode decomposition algorithm to obtain a plurality of intrinsic mode function components;

[0043] The entropy calculation subunit, connected to the decomposition subunit, is configured to calculate the multi-scale frequency entropy value of each intrinsic mode function component and screen out the noise-dominated components with the median of the entropy values as the threshold;

[0044] The denoising subunit, connected to the entropy calculation subunit, is configured to denoise the noise-dominated components by using a recursive least squares algorithm;

[0045] The reconstruction subunit, connected to the denoising subunit, is configured to reconstruct the denoised noise-dominated components and non-noise-dominated components to obtain the denoised zero-sequence current signal.

[0046] Specifically, the Gram transform module includes:

[0047] The normalization subunit is configured to normalize the denoised zero-sequence current signal to the interval [-1, 1] to obtain a normalized signal

[0048] The angle component generation subunit, connected to the normalization subunit, is configured to map the normalized signal to the polar coordinate system to generate an angle component Wherein, is the i-th element of the normalized signal;

[0049] A matrix construction subunit, connected to the angle component generation subunit, for constructing a Gram angular difference field matrix according to the angle component to generate the two-dimensional matrix image, and the Gram angular difference field matrix satisfies the relational expression:

[0050]

[0051] Specifically, the teacher network module includes:

[0052] A wavelet decomposition subunit, which includes the wavelet convolution module, and is used to decompose the two-dimensional matrix image by using the Haar wavelet basis function in the wavelet convolution module to separate the low-frequency component and the high-frequency component;

[0053] A convolution processing subunit, connected to the wavelet decomposition subunit, for respectively performing 3×3 depth convolution kernel processing on the low-frequency component and the high-frequency component;

[0054] A reconstruction subunit, connected to the convolution processing subunit, for reconstructing the processed low-frequency component and high-frequency component into a fused feature map through inverse wavelet transform and outputting the fused feature map as the fault feature vector.

[0055] The single-phase grounding fault line selection method and system provided by this application aim to improve the accuracy and efficiency of fault line selection. This method is directed at the zero-sequence current signal of the distribution network, implements signal decomposition and noise screening to obtain the denoised zero-sequence current signal. Then, using the Gram angular difference field transformation technology, the denoised signal is converted into a two-dimensional matrix image for subsequent analysis. This image is sent to a teacher network containing a wavelet convolution module to extract the fault feature vector. Subsequently, through a student network containing a depthwise separable convolution head module, knowledge distillation is performed on the fault feature vector, and the discrimination result of the fault line is output. Finally, based on this discrimination result, precise fault line selection operations are performed, thereby quickly locating and handling single-phase grounding faults in the distribution network. This method solves the technical problem of how to construct a single-phase grounding fault line selection method that takes into account high-precision identification and low computational complexity in a distribution network environment with high noise interference and weak feature differences, significantly improves the fault detection accuracy and reduces the computational complexity, and effectively solves the performance bottleneck of traditional methods in complex noise and high-resistance fault scenarios. Description of the Drawings

[0056] The specification drawings forming a part of this application are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention. In the drawings:

[0057] Figure 1 Schematic flowchart of the single-phase grounding fault line selection method for the distribution network provided by this application;

[0058] Figure 2 Schematic diagram of the single-phase grounding fault line selection system for the distribution network provided by this application.

[0059] Through the above-mentioned drawings, specific embodiments of this application have been shown, and there will be more detailed descriptions hereinafter. These drawings and textual descriptions are not intended to limit the scope of the concept of this application in any way, but to illustrate the concept of this application to those skilled in the art by referring to specific embodiments. Detailed implementation manners

[0060] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in this application will be clearly and completely described below in conjunction with the drawings in this application. Obviously, the described embodiments are some, but not all, of the embodiments of this application. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of this application.

[0061] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments of the invention described herein can be implemented in an order different from those illustrated or described herein.

[0062] In this invention, words such as "exemplary" or "for example" are used to represent examples, illustrations, or explanations. Any embodiment or design solution described as "exemplary" or "for example" in this application should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Rather, the use of words such as "exemplary" or "for example" is intended to present relevant concepts in a specific manner.

[0063] The single-phase grounding fault line selection method and system provided by this application preprocess the zero-sequence current signal of the distribution network through signal decomposition and noise screening to obtain a pure signal. The denoised signal is transformed into a two-dimensional matrix image by using the Gram angular difference field transformation to visually present the signal characteristics. The teacher network containing a wavelet convolution module is used to perform deep feature extraction on the image to generate a fault feature vector. The student network containing a depthwise separable convolution head module is used to perform knowledge distillation on the feature vector to refine key information and output the discriminant result of the fault line. Finally, the fault line selection is performed according to the discriminant result to achieve fast and accurate fault location and processing.

[0064] The following uses specific embodiments to elaborate in detail on the technical solution of this application and how the technical solution of this application solves the above technical problems. The following several specific embodiments can be combined with each other, and the same or similar concepts or processes may not be elaborated in some embodiments. The embodiments of this application will be described below in conjunction with the accompanying drawings.

[0065] Figure 1 It is a schematic flow diagram of the single-phase grounding fault line selection method for the distribution network provided by this application, aiming to elaborate in detail on the single-phase grounding fault line selection method for the distribution network. As Figure 1 shown, the single-phase grounding fault line selection method for the distribution network provided in this embodiment includes:

[0066] S101: Perform signal decomposition and noise screening on the zero-sequence current signal of the distribution network to obtain the denoised zero-sequence current signal.

[0067] Specifically, performing signal decomposition and noise screening on the zero-sequence current signal of the distribution network to obtain the denoised zero-sequence current signal includes:

[0068] Decompose the zero-sequence current signal using an improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) algorithm to obtain multiple intrinsic mode function components;

[0069] Calculate the multi-scale frequency entropy value of each intrinsic mode function component. The specific calculation of the multi-scale frequency entropy value of each intrinsic mode function component includes: performing coarse-graining processing on each intrinsic mode function component to generate multi-scale subsequences; performing Fourier transform on each multi-scale subsequence to calculate the probability distribution and Shannon entropy value of the frequency-domain signal; taking the average of the Shannon entropy values of all scales to obtain the multi-scale frequency entropy value. Use the median of the entropy values as the threshold to screen out the noise-dominated components;

[0070] Denoise the noise-dominated components through a recursive least squares algorithm and reconstruct them with the non-noise-dominated components to obtain the denoised zero-sequence current signal.

[0071] The specific implementation includes the following steps:

[0072] Step 1: Decompose the zero-sequence current signal using an improved ICEEMDAN (i.e., complete ensemble empirical mode decomposition with adaptive noise algorithm).

[0073] (1) Input signal: Collect the zero-sequence current signal at the distribution network bus, with a sampling frequency of 10 kHz and a signal length of 1024 points, denoted as the discrete sequence x(n), where n = 1, 2,..., 1024.

[0074] (2) Improved ICEEMDAN decomposition process:

[0075] a. Initialization parameters:

[0076] Set the number of decomposition layers to 12 layers (K = 12), and the adaptive white noise amplitude coefficient β k = 0.2×σ(r k-1 ), where σ(r k-1 ) represents the standard deviation of the residual signal of the previous layer.

[0077] Add 100 times of Gaussian white noise to each group (M = 100).

[0078] b. Iterative decomposition:

[0079] At the k-th iteration (k = 1, 2,..., 12): Generate M groups of Gaussian white noise whose amplitude is controlled by β k .

[0080] Construct the noisy signal: where r k-1 (n) is the residual of the previous layer.

[0081] Perform empirical mode decomposition (EMD) on each noisy signal to extract the first-order mode component

[0082] Calculate the mode component of the k-th layer:

[0083] Update the residual: r k (n) = r k-1 (n) - d k (n).

[0084] c. Output results: 12 intrinsic mode function components {d1(n), d2(n),..., d 12 (n)} and the residual component r 12 (n).

[0085] d. Improvement point: The improved ICEEMDAN algorithm dynamically adjusts the noise injection intensity through the adaptive white noise amplitude coefficient β k = 0.2×σ(r k-1 ), compared with the traditional fixed β k = 0.2, reducing the over-decomposition problem of high-frequency components.

[0086] Step 2: Multi-scale frequency entropy calculation and noise screening

[0087] 2.1 Coarse-graining processing: Generate subsequences with scale factors s = 1, 2, 3, 4, 5 for each intrinsic mode function component d k (n):

[0088] The length of the subsequence is [1024 / s]. For example, when s = 2, a 512-point sequence is generated.

[0089] Each subsequence is calculated by moving average:

[0090] where j = 1, 2, …, [1024 / s].

[0091] 2.2 Calculation of frequency-domain Shannon entropy:

[0092] For each subsequence Perform a fast Fourier transform (FFT) to calculate its power spectrum P(f).

[0093] After normalizing the power spectrum, calculate the entropy value through the Shannon entropy formula Calculate the entropy value.

[0094] 2.3 Multi-scale frequency entropy value: For each d k (n), take the average entropy of 5 scales:

[0095]

[0096] 2.4 Screening of noise components:

[0097] Calculate the median MFE of all MFE(k) median .

[0098] If MFE(k) > 1.2 × MFE median , it is determined that d k (n) is the noise-dominated component.

[0099] Step 3: Recursive least squares (RLS) denoising and reconstruction

[0100] 3.1 RLS filter parameters:

[0101] The filter order is 4, the forgetting factor is 0.98, and the regularization parameter is 0.01.

[0102] 3.2 Denoising process:

[0103] For each noise component d k (n), update the filter weights point by point through the RLS algorithm to predict and suppress the noise.

[0104] 3.3 Signal reconstruction:

[0105] Combine the denoised noise components, non-noise components, and residual components to obtain the denoised zero-sequence current signal:

[0106]

[0107] where D signalDenote the set of non-noise dominant components; d k (n) represents the time-domain signal of the non-noise component; D noise Denote the set of noise dominant components; d k ′(n) represents the denoised noise component, obtained from the noise component processed by the recursive least squares (RLS) filter; r 12 (n) represents the residual component.

[0108] In this step, through dynamic noise suppression and multi-scale feature analysis, the recognition rate of weak fault signals is significantly improved, providing high-quality input for subsequent Gram angular difference field transformation and deep learning models.

[0109] S102: Process the denoised zero-sequence current signal through Gram angular difference field transformation to generate a corresponding two-dimensional matrix image.

[0110] Specifically, the process of processing the denoised zero-sequence current signal through Gram angular difference field transformation to generate a corresponding two-dimensional matrix image includes:

[0111] Normalize the denoised zero-sequence current signal to the interval [-1, 1] to obtain a normalized signal

[0112] Map the normalized signal to the polar coordinate system to generate an angular component Where, is the i-th element of the normalized signal;

[0113] Construct a Gram angular difference field matrix according to the angular component to generate the two-dimensional matrix image, and the Gram angular difference field matrix satisfies the relationship:

[0114]

[0115] When this step is specifically implemented, it specifically includes the following steps:

[0116] Step 1: Normalization of zero-sequence current signal

[0117] 1.1 Input signal: Receive the denoised zero-sequence current signal x denoised (n), with a signal length of 1024 points.

[0118] 2.1 Normalization processing:

[0119] Adopt the maximum-minimum normalization method to linearly map the signal to the interval [-1, 1]:

[0120]

[0121] Output the normalized signal sequence

[0122] Step 2: Generation of the Polar Coordinate Angle Component

[0123] 2.1 Angle Mapping:

[0124] For each normalized signal point calculate the polar angle:

[0125] where

[0126] generate the angle component sequence Θ = {θ i , θ2,..., θ 1024}.

[0127] Step 3: Construction of the Gram Angle Difference Field (GADF) Matrix

[0128] 3.1 Matrix Element Definition:

[0129] Construct a 1024×1024 Gram angle difference field matrix GADF, whose element G i,j is:

[0130] G i,j = sin(θ i - θ j ), i, j = 1, 2,..., 1024

[0131] Calculation process:

[0132] Row index i: corresponds to the angle θ i at time point t i .

[0133] Column index j: corresponds to the angle θ j at time point t j .

[0134] Matrix symmetry: Since sin(θ i - θ j ) = -sin(θ j - θ i ), the matrix is skew-symmetric.

[0135] Image generation:

[0136] Map the value of each element of the GADF matrix to a grayscale pixel value:

[0137]

[0138] Output a 1024×1024 pixel two-dimensional grayscale image as the input to the subsequent deep learning model.

[0139] In this step, through the Gram angular difference field transformation, the phase difference features of the time series signal are encoded into two-dimensional spatial textures, solving the problem that traditional time-frequency methods are insensitive to weak phase features and providing a structured input for the high-precision feature extraction of the subsequent teacher network.

[0140] S103: Input the two-dimensional matrix image into a teacher network containing a wavelet convolution module for feature extraction to obtain a fault feature vector.

[0141] Specifically, the step of inputting the two-dimensional matrix image into a teacher network containing a wavelet convolution module for feature extraction to obtain a fault feature vector includes:

[0142] Decompose the two-dimensional matrix image using the Haar wavelet basis function in the wavelet convolution module to separate the low-frequency component and the high-frequency component;

[0143] Perform 3×3 depth convolution kernel processing on the low-frequency component and the high-frequency component respectively, and reconstruct them into a fused feature map through inverse wavelet transform;

[0144] Output the fused feature map as the fault feature vector.

[0145] Specific implementation specifically includes the following steps:

[0146] Step 1: Decompose the image using the Haar wavelet basis function

[0147] 1.1 Input data: Receive the Gram angular difference field (GADF) two-dimensional matrix image generated in S102, with a size of 1024×1024 pixels and a gray value range of 0-255.

[0148] 1.2 Wavelet decomposition process:

[0149] Definition of the Haar wavelet basis function:

[0150] Low-pass filter coefficients:

[0151] High-pass filter coefficients:

[0152] Image decomposition:

[0153] Perform two-dimensional discrete wavelet transform (DWT) on the GADF image to separate it into four sub-bands: low-frequency component LL: size 512×512, retaining the macroscopic structure of the image.

[0154] Horizontal high-frequency component LH: size 512×512, capturing vertical edge features.

[0155] Vertical high-frequency component HL: size 512×512, capturing horizontal edge features.

[0156] Diagonal high-frequency component HH: Size 512×512, capturing diagonal texture features.

[0157] Step 2: Depth convolution processing and feature fusion

[0158] 2.1 Low-frequency component processing:

[0159] 2.1.1 Input the LL sub-band and process it using 3×3 depthwise separable convolution:

[0160] Depth convolution: 512 3×3 convolution kernels, stride 1, padding 1, calculated channel by channel.

[0161] Pointwise convolution: 1×1 convolution kernel expands the number of channels to 256, and the activation function is ReLU.

[0162] 2.1.2 Output the low-frequency feature map

[0163] 2.2 High-frequency component processing:

[0164] 2.2.1 Combine the LH, HL, and HH sub-bands into a multi-channel input (3 channels, 512×512) and use 3×3 depthwise separable convolution:

[0165] Depth convolution: 3 3×3 convolution kernels, stride 1, padding 1, processed channel by channel.

[0166] Pointwise convolution: 1×1 convolution kernel expands the number of channels to 128, and the activation function is ReLU.

[0167] 2.2.2 Output the high-frequency feature map

[0168] 2.3 Inverse wavelet transform reconstruction:

[0169] 2.3.1 Input F low and F hight into the inverse discrete wavelet transform (IDWT):

[0170] Low-frequency upsampling: Perform bilinear interpolation on F low to restore the size to 1024×1024.

[0171] High-frequency fusion: Split F hight into three sub-bands, LH, HL, and HH, and upsample them to 1024×1024 respectively.

[0172] Reconstruction fusion: Superimpose and fuse the upsampled low-frequency and the three high-frequency sub-bands according to the inverse Haar wavelet transform rule to obtain the fused feature map

[0173] 2.3.2 Output the fused feature map

[0174] Step 3: Generate the fault feature vector

[0175] 3.1 Global average pooling:

[0176] Perform global average pooling (GlobalAverage Pooling) on the fused feature map to reduce the 1024×1024 feature map of each channel to a 1×1 scalar, generating a 384-dimensional vector F pooled 。

[0177] 3.2 Compression by the fully connected layer:

[0178] Input the 384-dimensional vector F pooled to the fully connected layer, and compress it to 128 dimensions through the fully connected layer:

[0179] V fault =W fc ·F pooled +b fc

[0180] where the weight matrix bias

[0181] Output a 128-dimensional fault feature vector V fault 。

[0182] In this step, through the collaborative design of multi-resolution analysis and lightweight convolution, hierarchical extraction and efficient compression of fault features are realized, providing high-discriminative feature inputs for the subsequent knowledge distillation of the student network.

[0183] S104: Perform knowledge distillation on the fault feature vector through a student network including a depthwise separable convolution head module, and output a fault line discrimination result.

[0184] Specifically, performing knowledge distillation on the fault feature vector through a student network including a depthwise separable convolution head module and outputting a fault line discrimination result includes:

[0185] Performing depthwise separable convolution operations on the fault feature vector to generate a localization branch feature and a classification branch feature;

[0186] Processing the localization branch feature through two depthwise separable convolutions and a bounding box regression layer in sequence to output a fault location prediction result;

[0187] Processing the classification branch feature through two depthwise separable convolutions and a fully connected layer in sequence to output a fault category probability distribution;

[0188] Generate the discriminant result of the faulty line according to the predicted result of the fault location and the probability distribution of the fault category.

[0189] Among them, when training the student network, the Adam optimizer (learning rate 0.001, batch size 32) is used, and it is iterated 50 rounds on a data set containing 10,000 groups of fault samples.

[0190] Specific implementation includes the following steps:

[0191] Step 1: Construction of the depthwise separable convolution head module

[0192] 1.1 Input data: A 128-dimensional fault feature vector output by the teacher network

[0193] 1.2 Feature branch generation:

[0194] 1.2.1 Location branch:

[0195] Reshape V fault into an 8×8×2 feature map (spatial dimension 8×8, number of channels 2).

[0196] Perform depthwise separable convolution operations:

[0197] Depth convolution: 2 3×3 convolutional kernels, stride 1, padding 1, processed channel by channel, output size 8×8×2.

[0198] Pointwise convolution: A 1×1 convolutional kernel expands the number of channels to 16, and the activation function is ReLU, outputting the location branch features

[0199] 1.2.2 Classification branch:

[0200] Input V fault into the fully connected layer, expanding it to a 256-dimensional vector, and the activation function is ReLU.

[0201] Reshape it into a 16×16×1 feature map and perform depthwise separable convolution:

[0202] Depth convolution: 1 3×3 convolutional kernel, stride 1, padding 1, output size 16×16×1. Pointwise convolution: A 1×1 convolutional kernel expands to 32 channels, outputting the classification branch features

[0204] Step 2: Location branch processing and location prediction

[0205] 2.1 Convolution processing:

[0206] Perform depthwise separable convolution on F loc twice in sequence:

[0207] First convolution: 16 3×3 depthwise convolution kernels, stride 1, padding 1, output size 8×8×16.

[0208] Second convolution: 16 3×3 depthwise convolution kernels, stride 1, padding 1, output size 8×8×16.

[0209] The ReLU activation function is used for all.

[0210] 2.2 Bounding box regression:

[0211] Input the final feature map into the bounding box regression layer (4 1×1 convolution kernels), and output the fault location parameters:

[0212]

[0213] x certer , y certer : The center coordinates of the fault area (normalized to [0,1]).

[0214] w, h: The width and height of the fault area (normalized to [0,1]).

[0215] Step 3: Classification branch processing and probability prediction

[0216] 3.1 Convolution processing:

[0217] 3.1.1 For F cls Perform depthwise separable convolution twice in sequence:

[0218] First convolution: 32 3×3 depthwise convolution kernels, stride 1, padding 1, output size 16×16×32.

[0219] Second convolution: 32 3×3 depthwise convolution kernels, stride 1, padding 1, output size 16×16×32.

[0220] 3.1.2 The ReLU activation function is used for all.

[0221] 3.2 Classification by fully connected layer:

[0222] After global average pooling, input it into the fully connected layer, and output the probability distribution of fault categories:

[0223] P cls = Softmax(W fc ·F pooled + b fc ),

[0224] where C: The number of fault categories (e.g., single-phase grounding fault, two-phase short circuit, etc.).

[0225] Step 4: Generation of fault discrimination results

[0226] 4.1 Result Fusion:

[0227] Location Screening: Extract the top K candidate bounding boxes with the highest confidence from B pred (e.g., K = 5).

[0228] Classification Matching: Extract the class probabilities of the corresponding regions from P cls according to the positions of the candidate bounding boxes.

[0229] 4.2 Final Output:

[0230] The discrimination result of the faulty line is the class with the highest probability and its location:

[0231]

[0232] Among them, TopK(B pred ) represents screening out the top K positions with the highest confidence from all 8×8 = 64 candidate positions. represents the probability distribution of the faulty line classes output by the classification branch, where the probability of the k-th candidate bounding box corresponds to C faulty line classes. represents traversing all K candidate bounding boxes (k = 1, 2,..., K) and selecting the class with the largest probability value as the final discrimination result.

[0233] In this step, through the depthwise separable convolution head module and knowledge distillation technology, the model is lightweighted on the premise of ensuring accuracy, meeting the requirements of real-time performance and embedded deployment of the distribution network fault line selection system.

[0234] S105: Perform the fault line selection operation according to the discrimination result.

[0235] Among them, the specific implementation steps of S105 specifically include the following steps:

[0236] Step 1: Validation of the effectiveness of the discrimination result

[0237] 1.1 Input data: Receive the discrimination result of the faulty line output by S104, including the faulty line class (such as "single-phase grounding") and the location parameter B pred = [x certer , y certer , w, h].[[]]

[0238] 1.2 Result verification:

[0239] Logical verification: Check whether the faulty line location is within the distribution network topology range (e.g., x certer

[0240] ∈ [0] and y certer ∈ [0, 1]). If it exceeds the range, it is marked as an invalid result.

[0241] Confidence Threshold: It is determined to be valid only when the classification probability P cls ≥ 0.95, otherwise the re - inspection mechanism is triggered.

[0242] Step 2: Tripping Instruction Generation and Execution

[0243] 2.1 Tripping Instruction Generation:

[0244] Target Line Mapping: According to the fault location (x certer , y certer ), match the distribution network topology database to determine the fault line ID (such as line L - 23).

[0245] Instruction Encoding: Generate a GOOSE tripping message according to the IEC 61850 standard, including the following fields: {

[0246] "cmd": "trip",

[0247] "line_id": "L - 23",

[0248] "timestamp": "2024 - 03 - 20T14:23:05.123Z",

[0249] "priority": 1

[0250] }

[0251] 2.2 Instruction Issuance:

[0252] Send the tripping instruction to the intelligent circuit breaker corresponding to the target line (such as model ABBREF615) through fiber - optic Ethernet, and the communication protocol is MMS (Manufacturing Message Specification).

[0253] Circuit Breaker Response Time Requirement: The delay from receiving the instruction to performing the tripping action ≤ 100ms.

[0254] Step 3: Fault Log Recording and Topology Update

[0255] 3.1 Log Recording:

[0256] Record the fault event in the SCADA system, and the fields include: fault time, line ID, fault category, location coordinates, classification probability, processing status.

[0257] The storage format is CSV, and an example entry: 2024 - 03 - 20 14:23:05, L - 23, single - phase grounding, 0.35, 0.72, 0.96, tripped.

[0258] 3.2 Topology Update:

[0259] After disconnecting the faulty line, the Dijkstra algorithm is called to recalculate the power flow path of the distribution network and update the topological connection matrix \(T\in\{0,1\}\) N×N (where \(N\) is the number of nodes) to ensure the restoration of power supply in the non-faulty area.

[0260] Power supply restoration time requirement: From the completion of tripping to topological update \(\leq30\) seconds.

[0261] 3.3 Coordinate - Line ID mapping rule:

[0262] The mapping rule between the fault location coordinates \((x\) certer , y\) certer ) and the distribution network line ID is as follows:

[0263] Rasterize the distribution network topology map into a \(100\times100\) grid;

[0264] According to the coordinates \((x\) certer \times100, y\) certer \times100) to locate to the corresponding grid;

[0265] Query the preset grid - line relationship table (such as grid \((35,72)\) corresponding to line L - 23);

[0266] Grid - line relationship table

[0267] Grid coordinate range (x, y) Corresponding line ID (30-40,70-80) L-23 (50-60,20-30) L-17

[0268] This step realizes the rapid isolation of the faulty line and system restoration through standardized instruction generation, strict verification mechanism and automatic topology update, ensuring the power supply reliability and operation and maintenance efficiency of the distribution network.

[0269] This embodiment provides a method for selecting the faulty line in a distribution network single-phase grounding fault. This method performs signal decomposition and noise screening on the zero-sequence current signal of the distribution network to obtain a pure signal after denoising. Subsequently, the Gram angular difference field transformation technology is used to transform the denoised signal into a two-dimensional matrix image, which can intuitively reflect the signal characteristics. Then, the image is input into a teacher network containing a wavelet convolution module for deep feature extraction to generate a fault feature vector. Through a student network containing a depthwise separable convolution head module, knowledge distillation is performed on the fault feature vector to further refine the key information and output the discrimination result of the faulty line. Finally, based on the discrimination result, precise faulty line selection operations are performed to quickly locate and handle the fault, ensuring the safe and stable operation of the distribution network. This method combines signal processing technology, image conversion, deep learning and knowledge distillation, providing an efficient solution for selecting the faulty line in a distribution network single-phase grounding fault.

[0270] Figure 2 Schematic diagram of the distribution network single-phase grounding fault line selection system provided by this application, as Figure 2As shown in the figure, the single-phase grounding fault line selection system for a distribution network provided in this embodiment uses the line selection method described in Figure 1 the embodiment. The system includes:

[0271] A signal processing module, which is used to decompose and filter the noise of the zero-sequence current signal of the distribution network to obtain a denoised zero-sequence current signal;

[0272] A Gram transform module, connected to the signal processing module, which is used to perform Gram angular difference field transform on the denoised zero-sequence current signal to generate a corresponding two-dimensional matrix image;

[0273] A teacher network module, connected to the Gram transform module, which is used to input the two-dimensional matrix image into a teacher network including a wavelet convolution module for feature extraction to obtain a fault feature vector;

[0274] A student network module, connected to the teacher network module, which is used to perform knowledge distillation on the fault feature vector through a student network including a depthwise separable convolution head module and output a fault line discrimination result;

[0275] A line selection execution module, connected to the student network module, which is used to perform fault line selection operations according to the discrimination result.

[0276] Specifically, the signal processing module includes:

[0277] A decomposition subunit, which is used to decompose the zero-sequence current signal by using an improved adaptive noise complete ensemble empirical mode decomposition algorithm to obtain a plurality of intrinsic mode function components;

[0278] An entropy calculation subunit, connected to the decomposition subunit, which is used to calculate the multi-scale frequency entropy value of each intrinsic mode function component and screen the noise-dominated components with the median of the entropy values as the threshold;

[0279] A denoising subunit, connected to the entropy calculation subunit, which is used to denoise the noise-dominated components through a recursive least squares algorithm;

[0280] A reconstruction subunit, connected to the denoising subunit, which is used to reconstruct the denoised noise-dominated components and non-noise-dominated components to obtain the denoised zero-sequence current signal.

[0281] Specifically, the Gram transform module includes:

[0282] A normalization subunit, which is used to normalize the denoised zero-sequence current signal to the interval [-1, 1] to obtain a normalized signal

[0283] An angle component generation subunit, connected to the normalization subunit, for mapping the normalized signal to a polar coordinate system to generate an angle component Wherein, is the i-th element of the normalized signal;

[0284] A matrix construction subunit, connected to the angle component generation subunit, for constructing a Gram angular difference field matrix according to the angle component to generate the two-dimensional matrix image, and the Gram angular difference field matrix satisfies the relation:

[0285]

[0286] Specifically, the teacher network module includes:

[0287] A wavelet decomposition subunit, which includes the wavelet convolution module, and is used to decompose the two-dimensional matrix image by using the Haar wavelet basis function in the wavelet convolution module to separate the low-frequency component and the high-frequency component;

[0288] A convolution processing subunit, connected to the wavelet decomposition subunit, for performing 3×3 depth convolution kernel processing on the low-frequency component and the high-frequency component respectively;

[0289] A reconstruction subunit, connected to the convolution processing subunit, for reconstructing the processed low-frequency component and high-frequency component into a fused feature map through inverse wavelet transform and outputting the fused feature map as the fault feature vector.

[0290] When the system is specifically implemented, it includes:

[0291] 1. A signal processing module, for performing signal decomposition and noise screening processing on the zero-sequence current signal of the distribution network to obtain a denoised zero-sequence current signal. The signal processing module includes:

[0292] 1.1 A decomposition subunit, which decomposes the zero-sequence current signal by using an improved adaptive noise complete ensemble empirical mode decomposition algorithm to obtain a plurality of intrinsic mode function components;

[0293] 1.2 Entropy calculation sub - unit, electrically connected to the decomposition sub - unit, is used to calculate the multi - scale frequency entropy values of each intrinsic mode function component, specifically including: performing coarse - graining processing on each component to generate multi - scale subsequences, performing Fourier transform on the subsequences to calculate the frequency - domain probability distribution and Shannon entropy values, and taking the average of the entropy values at all scales; screening out the noise - dominant components with the median of the entropy values as the threshold;

[0294] 1.3 Denoising sub - unit, electrically connected to the entropy calculation sub - unit, performs point - by - point filtering denoising on the noise - dominant components through the recursive least - squares algorithm;

[0295] 1.4 Reconstruction sub - unit, electrically connected to the denoising sub - unit, superimposes and reconstructs the denoised noise - dominant components and non - noise - dominant components into the denoised zero - sequence current signal.

[0296] 2. Gram transform module, communicatively connected to the signal processing module, the Gram transform module is used to convert the denoised zero - sequence current signal into a two - dimensional matrix image, including:

[0297] 2.1 Normalization sub - unit, linearly normalizes the zero - sequence current signal to the interval [-1, 1] to generate a normalized signal

[0298] 2.2 Angle component generation sub - unit, electrically connected to the normalization sub - unit, converts the normalized signal into a polar - coordinate angle sequence Θ={θ , θ2,..., θ i , θ n} through mapping;

[0299] 2.3 Matrix construction sub - unit, electrically connected to the angle component generation sub - unit, calculates the Gram angle difference field matrix GADF according to the angle components, where the matrix element G i,j = sin(θ i - θ j ), and converts the matrix into a grayscale image of 1024×1024 pixels.

[0300] 3. Teacher network module, communicatively connected to the Gram transform module, includes:

[0301] 3.1 Wavelet decomposition sub - unit, the wavelet decomposition sub - unit includes the wavelet convolution module, the wavelet convolution module is built - in with Haar wavelet basis functions, and decomposes the input two - dimensional matrix image into low - frequency components (LL), horizontal high - frequency components (LH), vertical high - frequency components (HL), and diagonal high - frequency components (HH);

[0302] 3.2 Convolution processing subunit, electrically connected to the wavelet decomposition subunit, which extends the channels of the low-frequency components to 256 using a 3×3 depthwise separable convolution kernel, and extends the channels of the combined high-frequency components to 128 using a 3×3 depthwise separable convolution kernel;

[0303] 3.3 Reconstruction subunit, electrically connected to the convolution processing subunit, which reconstructs the processed low-frequency and high-frequency components into a fused feature map of 1024×1024×384 through inverse wavelet transform, and compresses it into a 128-dimensional fault feature vector through global average pooling and a fully connected layer.

[0304] 4. Student network module, communicatively connected to the teacher network module, including:

[0305] 4.1 Depthwise separable convolution head module, which branches the 128-dimensional fault feature vector into localization features and classification features;

[0306] 4.2 Localization branch processing unit, which extracts spatial features through two 3×3 depthwise separable convolutions and outputs the fault location coordinates through a bounding box regression layer;

[0307] 4.3 Classification branch processing unit, which extracts classification features through two 3×3 depthwise separable convolutions and outputs the fault category probability distribution through a fully connected layer;

[0308] 4.4 Result fusion unit, which generates the fault line discrimination result by combining the localization confidence and the classification probability.

[0309] 5. Line selection execution module, communicatively connected to the student network module, including:

[0310] 5.1 Instruction generation unit, which matches the distribution network topology according to the discrimination result and generates a tripping instruction that conforms to the IEC 61850 standard;

[0311] 5.2 Communication and distribution unit, which sends the instruction to the intelligent circuit breaker of the target line through fiber optic Ethernet;

[0312] 5.3 Log recording unit, which stores the fault event data and updates the distribution network topology connection matrix.

[0313] In this system, each module realizes data transmission through electrical connection or communication connection. Among them, the denoised zero-sequence current signal is transmitted between the signal processing module and the Gram transform module, the teacher network module outputs the fault feature vector to the student network module, and the line selection execution module triggers the circuit breaker action according to the discrimination result. The system realizes the efficient fusion of noise suppression and feature lightweight model through multi-module collaboration, ensuring the accuracy and real-time performance of fault line selection.

[0314] Other embodiments of the present application will be readily apparent to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. The present application is intended to cover any variations, uses, or adaptations of the present application, which follow the general principles of the present application and include well-known knowledge or conventional technical means in the technical field not disclosed in the present application. The specification and examples are only illustrative, and the true scope and spirit of the present application are pointed out by the claims.

[0315] It should be understood that the present application is not limited to the exact structures described above and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of the present application is only limited by the appended claims.

Claims

1. A single-phase grounding fault line selection method for a distribution network, characterized in that, Including: Performing signal decomposition and noise screening on the zero-sequence current signal of the distribution network to obtain a denoised zero-sequence current signal; Processing the denoised zero-sequence current signal through Gram angular difference field transformation to generate a corresponding two-dimensional matrix image; Inputting the two-dimensional matrix image into a teacher network including a wavelet convolution module for feature extraction to obtain a fault feature vector; Performing knowledge distillation on the fault feature vector through a student network including a depthwise separable convolution head module and outputting a fault line discrimination result; Performing a fault line selection operation according to the discrimination result.

2. The single-phase grounding fault line selection method for a distribution network according to claim 1, characterized in that, The performing signal decomposition and noise screening on the zero-sequence current signal of the distribution network to obtain a denoised zero-sequence current signal includes: Decomposing the zero-sequence current signal by using an improved adaptive noise complete ensemble empirical mode decomposition algorithm to obtain a plurality of intrinsic mode function components; Calculating the multi-scale frequency entropy value of each intrinsic mode function component, and screening out the noise-dominated components by taking the median of the entropy values as the threshold; Denosing the noise-dominated components by using a recursive least squares algorithm, and reconstructing with non-noise-dominated components to obtain the denoised zero-sequence current signal.

3. The single-phase grounding fault line selection method for a distribution network according to claim 1, characterized in that The processing the denoised zero-sequence current signal through Gram angular difference field transformation to generate a corresponding two-dimensional matrix image includes: Normalize the denoised zero-sequence current signal to the range of [-1, 1] to obtain a normalized signal Map the normalized signal to the polar coordinate system to generate an angular component where is the i-th element of the normalized signal; Constructing a Gram angular difference field matrix according to the angular components to generate the two-dimensional matrix image, and the Gram angular difference field matrix satisfies the relation:

4. The single-phase grounding fault line selection method for a distribution network according to claim 1, characterized in that, The inputting the two-dimensional matrix image into a teacher network including a wavelet convolution module for feature extraction to obtain a fault feature vector includes: Decomposing the two-dimensional matrix image by using the Haar wavelet basis function in the wavelet convolution module to separate the low-frequency components and high-frequency components; Performing 3×3 depth convolution kernel processing on the low-frequency components and high-frequency components respectively, and reconstructing into a fused feature map through inverse wavelet transform; Outputting the fused feature map as the fault feature vector.

5. The single-phase grounding fault line selection method for a distribution network according to claim 1, wherein, The performing knowledge distillation on the fault feature vector through a student network including a depthwise separable convolution head module and outputting a fault line discrimination result includes: Performing depthwise separable convolution operation on the fault feature vector to generate a localization branch feature and a classification branch feature; Processing the localization branch feature through two depthwise separable convolutions and a bounding box regression layer in sequence to output a fault location prediction result; Processing the classification branch feature through two depthwise separable convolutions and a fully connected layer in sequence to output a fault category probability distribution; Generating the fault line discrimination result according to the fault location prediction result and the fault category probability distribution.

6. The single-phase grounding fault line selection method for a distribution network according to claim 2, characterized in that, The calculating the multi-scale frequency entropy value of each intrinsic mode function component includes: Performing coarse-graining processing on each intrinsic mode function component to generate multi-scale subsequences; Performing Fourier transform on each multi-scale subsequence to calculate the probability distribution and Shannon entropy value of the frequency domain signal; Taking the average of the Shannon entropy values of all scales to obtain the multi-scale frequency entropy value.

7. A single-phase grounding fault line selection system for a distribution network, characterized in that, The line selection system applies the line selection method according to any one of claims 1-6, and the line selection system includes: A signal processing module, which is used to perform signal decomposition and noise screening on the zero-sequence current signal of the distribution network to obtain the denoised zero-sequence current signal; A Gram transform module, connected to the signal processing module, which is used to perform Gram angular difference field transform on the denoised zero-sequence current signal to generate a corresponding two-dimensional matrix image; A teacher network module, connected to the Gram transform module, which is used to input the two-dimensional matrix image into a teacher network including a wavelet convolution module for feature extraction to obtain a fault feature vector; A student network module, connected to the teacher network module, which is used to perform knowledge distillation on the fault feature vector through a student network including a depthwise separable convolution head module and output a fault line discrimination result; A line selection execution module, connected to the student network module, which is used to perform fault line selection operations according to the discrimination result.

8. The system according to claim 7, wherein The signal processing module includes: A decomposition subunit, which is used to decompose the zero-sequence current signal by using an improved adaptive noise complete ensemble empirical mode decomposition algorithm to obtain a plurality of intrinsic mode function components; An entropy calculation subunit, connected to the decomposition subunit, which is used to calculate the multi-scale frequency entropy value of each intrinsic mode function component and screen the noise-dominated components with the median of the entropy values as the threshold; A denoising subunit, connected to the entropy calculation subunit, which is used to denoise the noise-dominated components by using a recursive least squares algorithm; A reconstruction subunit, connected to the denoising subunit, which is used to reconstruct the denoised noise-dominated components and non-noise-dominated components to obtain the denoised zero-sequence current signal.

9. The system according to claim 7, wherein The Gram transform module includes: A normalization subunit, configured to normalize the denoised zero-sequence current signal to the interval [-1, 1] to obtain a normalized signal An angle component generation subunit, connected to the normalization subunit, for mapping the normalized signal to a polar coordinate system to generate an angle component wherein is the i-th element of the normalized signal; A matrix construction subunit, connected to the angle component generation subunit, which is used to construct a Gram angular difference field matrix according to the angle components to generate the two-dimensional matrix image, and the Gram angular difference field matrix satisfies the relation:

10. The system according to claim 7, wherein The teacher network module includes: A wavelet decomposition subunit, the wavelet decomposition subunit includes the wavelet convolution module, and the wavelet decomposition subunit is used to decompose the two-dimensional matrix image by using the Haar wavelet basis function in the wavelet convolution module to separate the low-frequency components and high-frequency components; A convolution processing subunit, connected to the wavelet decomposition subunit, which is used to perform 3×3 depth convolution kernel processing on the low-frequency components and high-frequency components respectively; A reconstruction subunit, connected to the convolution processing subunit, which is used to reconstruct the processed low-frequency components and high-frequency components into a fused feature map through inverse wavelet transform and output the fused feature map as the fault feature vector.

Images