Analysis and Fault Location Method of Distribution Network Grounding Line Selection Based on Signal Feature Identification

Abstract

The present invention discloses a method for analyzing the grounding line selection and fault location of a distribution network based on signal feature identification, which includes the following steps: extracting the fault signals of monitoring points; using the AGCN-TCN model to classify the fault signals; calculating the local outlier factor of the monitoring points, decomposing the fault signals of each monitoring point according to a sliding window, assigning the decomposed sliding window frequency domain as the frequency domain fault index, constructing a comprehensive fault location determination index based on the local outlier factor and the frequency domain fault index, and determining the location of the fault point through the comprehensive fault location determination index; the present invention innovatively constructs an AGCN-TCN model for a distributed distribution network, realizes the feature extraction and analysis of fault signals from the perspectives of space and time, and avoids decision-making mistakes caused by the failure of a single monitoring device; the present invention defines a comprehensive fault location determination index, discriminates the fault location from the dual perspectives of time-frequency domain, and avoids the inaccuracy of a single method.

Description

Technical Field

[0001] The present invention relates to the technical field of distributed distribution network grounding line selection analysis, and specifically to a grounding line selection analysis and fault location method for distribution networks based on signal feature identification. Background Art

[0002] With the access of a high proportion of new source-load equipment, the proportion of distributed distribution networks in traditional distribution networks is increasing. The influx of a large number of devices and the expansion of the scope bring serious randomness and volatility, and distribution network faults show a trend of diversification and expansion. The topology of the distribution network becomes more complex and dynamic. Traditional grounding line selection methods may not be able to effectively handle faults in this new distribution network environment, resulting in delays in fault location and isolation speed, and affecting the stability and power supply reliability of the system.

[0003] As a common type of fault in distribution networks, grounding faults pose a great threat to the safety of power equipment and personnel. Especially in distributed distribution networks, the detection and line selection problems of grounding faults are more complex because the fault locations may be relatively scattered, and the access of different distributed power sources makes the system operation mode constantly change. Therefore, studying the grounding line selection method in distributed distribution networks can not only improve the accuracy and response speed of fault diagnosis, but also ensure rapid and effective isolation after a fault occurs, avoid the spread of the fault, and thus improve the reliability and safety of the distribution network.

[0004] Single-phase grounding faults can account for up to 80% of the faults in the system. If the distribution network remains in an abnormal operating state with a single-phase grounding fault for a long time, the possibility of the fault developing into a more serious fault will also increase, and it is easy to trigger more serious faults. At present, the line selection methods for single-phase grounding faults in small current grounding systems can be divided into steady-state line selection schemes, transient line selection schemes, external injection methods, and comprehensive line selection schemes according to the line selection criteria. Among them, the line selection schemes based on steady-state fault characteristics mainly include active component method, zero-sequence admittance method, zero-sequence reactive component method, group amplitude comparison and phase comparison method, fifth harmonic method, polarity comparison method, etc.; the line selection schemes based on transient fault characteristics mainly include first half-wave method, energy comparison method, wavelet analysis method, etc. However, the above methods are not applicable to the increasingly extensive distributed distribution networks. Therefore, this patent proposes a comprehensive line selection method assisted by a deep neural network intelligent algorithm, which can extract the data feature information of distributed distribution networks and realize the formulation of accurate and reasonable grounding line selection strategies. Summary of the Invention

[0005] Aiming at the deficiencies of the prior art, the present invention provides a grounding line selection analysis and fault location method for distribution networks based on signal feature identification, aiming to solve the problems in the background art.

[0006] To achieve the above object, the present invention provides the following technical solutions: A method for analyzing and fault locating of grounding line selection in a distribution network based on signal feature identification, comprising the following steps:

[0007] Step S1: Extract the fault signals of multiple monitoring points around the target distribution network and perform preprocessing;

[0008] Step S2: Use the AGCN-TCN model to classify the preprocessed fault signals;

[0009] The AGCN-TCN model extracts spatial features using the improved graph neural convolutional network AGCN, extracts temporal features using the temporal convolutional network TCN, fuses the spatial features and temporal features through a fully connected layer to obtain fused features, inputs the fused features into the classification layer, and matches the category of the corresponding fault signal with the corresponding wiring method;

[0010] The improved graph neural convolutional network AGCN improves its own adjacency matrix, that is, introducing distance weights into the adjacency matrix;

[0011] Step S3: Use the LOF outlier monitoring algorithm to calculate the local outlier factor of the monitoring points, decompose the preprocessed fault signals of each monitoring point according to a sliding window, assign the decomposed sliding window frequency domain as the frequency domain fault index, construct a comprehensive fault location determination index based on the local outlier factor and the frequency domain fault index of the monitoring points, and judge the location of the fault point through the comprehensive fault location determination index.

[0012] Further, the improved adjacency matrix of the improved graph neural convolutional network AGCN is expressed as:

[0013] ;

[0014] In the formula, is the element constituting the adjacency matrix ; represents the monitoring point and the monitoring point the physical distance between; represents a hyperparameter used to control the attenuation rate.

[0015] Further, the comprehensive fault location determination index is expressed as:

[0016] ;

[0017] In the formula, represents the set threshold of the local outlier factor; represents the local outlier factor;

[0018] When the monitoring point it means the monitoring point is determined as the fault point when the monitoring point indicates that the monitoring point is not determined as the fault point.

[0019] Furthermore, the preprocessing in step S1 includes decomposing the fault signal into intrinsic mode components by variational mode decomposition; denoising the intrinsic mode components by an improved wavelet threshold method;

[0020] The improved wavelet threshold method is to improve the threshold function of the original wavelet threshold method;

[0021] The improved threshold function is expressed as:

[0022] ;

[0023] In the formula, and represent the upper threshold and the lower threshold respectively; represents the input data of the improved threshold function.

[0024] Furthermore, the specific process of calculating the local outlier factor of the monitoring point by using the LOF outlier monitoring algorithm is as follows:

[0025] Calculate the neighborhood distance between the monitoring point and the monitoring point . Let the monitoring point and the monitoring point both monitor fault signals. The neighborhood distance of the monitoring point is expressed as: where ; represents the th fault signal monitored in the monitoring point ; represents the th

[0026] fault signal monitored in the monitoring point ; Calculate the distance neighborhood of the monitoring point

[0027] ;

[0028] In the formula, represents the distance neighborhood of the monitoring point , that is, the set of all monitoring points that satisfy . Indicates the monitoring point The distance, that is, the number of nearest neighbor points considered when calculating the distance neighborhood; Indicates the set of monitoring points;

[0029] Calculate the monitoring point to the monitoring point The reachable distance, expressed as:

[0030] ;

[0031] ;

[0032] In the formula, Indicates the monitoring point to the monitoring point The reachable distance; Indicates the monitoring point to the monitoring point The 5th reachable distance; Indicates the monitoring point to the monitoring point The 5th reachable distance; Indicates the monitoring point and the monitoring point The neighboring distance between; Indicates the monitoring point to itself The distance to the nearest monitoring point within the neighboring distance;

[0033] The monitoring point Local reachability density Is expressed as:

[0034] ;

[0035] In the formula, Indicates the monitoring point The For example, the number of neighborhood points, satisfying ;

[0036] Calculate the local outlier factor of the monitoring point Expressed as:

[0037] ;

[0038] In the formula, Indicates the local outlier factor of the monitoring point ; Indicates the monitoring point Local reachability density.

[0039] Furthermore, the specific process of decomposing the preprocessed fault signals at each monitoring point by a sliding window and assigning the obtained sliding window frequency domain as the frequency domain fault index is as follows: Fourier decompose the preprocessed fault signals at each monitoring point by a sliding window, and assign the obtained sliding window frequency domain after Fourier decomposition as the frequency domain fault index. When the current sliding window frequency domain assignment is greater than the frequency domain assignment of the previous sliding window, the current sliding window is regarded as the fault start point. It is 1; when the current sliding window frequency domain assignment is less than the frequency domain assignment of the previous sliding window, the current sliding window is regarded as the fault end point. It is 0, where the sliding window is the fault point.

[0040] Furthermore, the specific process of denoising the intrinsic mode components by using the improved wavelet threshold method is as follows:

[0041] Perform wavelet decomposition on the noisy data signal, that is, the intrinsic mode component. Determine the wavelet decomposition scale according to the scale and sampling frequency of the noisy data signal and obtain the wavelet decomposition coefficients. Among them, the wavelet decomposition coefficients include the approximation coefficients and the detail coefficients. The approximation coefficients represent the low-frequency components of the noisy data signal, and the detail coefficients represent the high-frequency components of the noisy data signal.

[0042] Use the improved threshold function to perform threshold quantization processing on the obtained wavelet decomposition coefficients to obtain the processed wavelet coefficients; reconstruct the signal based on the wavelet coefficients after threshold processing to obtain the denoised signal, that is, the denoised intrinsic mode component.

[0043] Furthermore, use the median absolute deviation to calculate the noise standard deviation, and use the data of the last layer of the detail coefficients to estimate the noise standard deviation; the median absolute deviation and the noise standard deviation are expressed as:

[0044] ;

[0045] In the formula, represents the median of the data of the last layer of the detail coefficients; represents the noise standard deviation;

[0046] Calculate the lower threshold in the fixed form :

[0047] ;

[0048] In the formula, represents the length of the noisy data signal; represents the proportionality coefficient of the upper threshold and the lower threshold.

[0049] Furthermore, the specific process of using the AGCN-TCN model to identify and classify the preprocessed fault signals is as follows:

[0050] Step S2.1: Construct an improved adjacency matrix A considering the connection distance according to the connection situation and connection distance information of the monitoring points;

[0051] Step S2.2: Use the preprocessed fault signals to construct a data set; divide the data set into a training set and a test set according to a ratio;

[0052] Step S2.3: Input the training set and the improved adjacency matrix A into the AGCN-TCN model to train the model, perform iterative updates according to the set number of training epochs, determine the parameters to be trained in the model, and after training is completed, obtain the optimal AGCN-TCN model;

[0053] Step S2.4: Input the test set and the improved adjacency matrix A into the optimal AGCN-TCN model. The improved graph neural convolutional network AGCN extracts the spatial connection between the distribution network topology structure and the fault signals as spatial features; input the test set into the time-domain convolutional network TCN to extract the temporal features of the monitored fault signals; connect the spatial features and the temporal features through a fully connected layer to obtain fused features, and input the fused features into the classification layer to match the category of the corresponding fault signal with the corresponding wiring method.

[0054] Compared with the existing technologies, the present invention has the following beneficial effects: The present invention innovatively constructs an AGCN-TCN model for distributed distribution networks, realizes the feature extraction and analysis of fault signals from the perspectives of space and time, and avoids decision-making mistakes caused by the failure of a single monitoring device; the present invention defines a comprehensive index for fault location determination, discriminates the fault location from the dual perspectives of time-frequency domain, avoids the inaccuracy of a single method, and improves the accuracy and response speed of fault location; overall, the present invention realizes the intelligent line selection analysis of grounding faults, significantly improves the fault handling efficiency and reliability of the power system in complex fault scenarios, and has important practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flowchart of the method of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0056] As Figure 1 shown, the present invention provides a technical solution: a method for distribution network grounding line selection analysis and fault location based on signal feature identification, including the following steps:

[0057] Step S1: Extract the fault signals of multiple monitoring points around the target distribution network and perform preprocessing.

[0058] Step S1.1: Install monitoring devices (such as current transformers, zero-sequence current sensors, etc.) at monitoring points around the target distribution network to extract fault signals (characteristic signals such as zero-sequence current analysis, zero-sequence voltage analysis, current amplitude and phase).

[0059] Step S1.2: Use variational mode decomposition (VMD) to decompose the fault signal into intrinsic mode components.

[0060] Variational mode decomposition (VMD) is a signal decomposition method based on the variational principle. It realizes signal decomposition by minimizing the difference between the signal and the reconstruction function. Its core idea is to decompose the signal into a series of local mode functions, where each mode function corresponds to a specific frequency component. In addition, variational mode decomposition has self-adaptability. Specifically, it can adjust the decomposition process according to the data signal frequency to improve the decomposition accuracy. The variational mode decomposition solution process is mainly divided into two processes: variational model establishment and variational model solution.

[0061] Step S1.21: Variational model establishment: Assume the fault signal is composed of components (intrinsic mode components) with different center frequencies and limited bandwidths. Each component has the smallest estimated bandwidth sum, and the constraint condition is that the sum of all intrinsic mode components is equal to the fault signal .

[0062] First, define each intrinsic mode component as an amplitude-frequency modulation and frequency modulation signal:

[0063] ;

[0064] In the formula, and respectively represent the instantaneous amplitude and instantaneous phase of ; represents time.

[0065] After initializing , perform processing with the Hilbert transform to obtain the analytic signal :

[0066] ;

[0067] In the formula, represents the Dirac function; represents the complex exponential function; represents the estimated center frequency; represents the convolution operation; represents the imaginary unit.

[0068] The introduction of the analytical signal can retain only the positive frequency components without loss and increase the signal representation ability. Mix it with its estimated center frequency, and modulate the spectrum of to the corresponding baseband.

[0069] Construct the optimal constrained variational model, expressed as:

[0070] ;

[0071] In the formula, represents the center frequency corresponding to the th intrinsic mode component; represents the partial derivative operation on ; represents the square operation of the two-norm.

[0072] Step S1.22: Solve the variational model: Introduce the penalty term and the Lagrange multiplier , and transform the variational problem into an unconstrained variational problem. The optimal constrained variational model can be equivalent to:

[0073] ;

[0074] In the formula, represents the Lagrangian function; represents the inner product operation.

[0075] Use the alternating direction multiplier method to alternately update and optimize each intrinsic mode component as well as the corresponding center frequency and Lagrange multiplier to obtain the optimal parameter solution; through the Fourier transform, the specific solution formula is as follows:

[0076] ;

[0077] In the formula, represents the noise tolerance; represents the frequency-domain representation of the th intrinsic mode component after the th iteration; represents the spectrum of the fault signal; represents the frequency-domain representation of the parameter or weight related to the th intrinsic mode component; represents the estimated center frequency of the th intrinsic mode component after the th iteration; represents the estimated parameter or weight of the th intrinsic mode component after the th iteration; represents at the The sum of the frequency-domain representations of all the intrinsic mode components after the -th iteration; Denotes the differential.

[0078] Obtained by calculation optimal values of the intrinsic mode components, and finally intrinsic mode components of the fault signal are decomposed by variational mode decomposition (VMD). And their central frequencies can be expressed as:

[0079] ;

[0080] In the formula, denotes the residual, which is the part that cannot be classified as intrinsic mode components, usually low-frequency signals or noise.

[0081] The fault signal is decomposed into multiple intrinsic mode components, representing different frequency components in the data. Usually, the high-frequency intrinsic mode components contain most of the noise components. Therefore, it is necessary to perform wavelet threshold denoising on the high-frequency intrinsic mode components.

[0082] The essence of wavelet threshold denoising is to filter the signal. Based on the strong correlation of wavelets, the wavelet decomposition of the noisy signal obtains the decomposition coefficients of the original signal and the noise. Its core principle is to decompose the signal into coefficients of different scales through wavelet transform. Noise usually appears in the high-frequency components, while the main information of the signal is usually concentrated in the low-frequency components. Therefore, wavelet transform can separate the signal and the noise, enabling the noise part to be removed through the thresholding method.

[0083] In terms of the threshold function, traditional wavelet threshold denoising includes hard threshold, soft threshold, and compromise threshold, etc. The hard function has two discontinuities, which will cause local oscillations in the denoised signal; although the soft function is continuous and the denoised signal is smoother, there will always be a fixed error between the estimated wavelet coefficients and the wavelet coefficients after signal decomposition, resulting in an error between the reconstructed signal obtained by wavelet inverse transform and the true signal, causing distortion of the reconstructed signal.

[0084] Step S1.3: Denoise the intrinsic mode components using the improved wavelet threshold method, which improves the threshold function of the original wavelet threshold method, ensuring continuity while reducing the error between the true signal and the reconstructed signal.

[0085] Step S1.31: Perform wavelet decomposition on the noisy data signal (intrinsic mode component), select the Morlet wavelet basis function, determine the wavelet decomposition scale according to the scale and sampling frequency of the noisy data signal, and obtain the wavelet decomposition coefficients. Herein, the wavelet decomposition coefficients specifically refer to the coefficients obtained through the discrete wavelet transform (DWT) decomposition process, which are usually a multi-level and multi-band result, mainly including approximation coefficients and detail coefficients. The approximation coefficients represent the low-frequency components of the noisy data signal, and the detail coefficients represent the high-frequency components of the noisy data signal.

[0086] Step S1.32: Perform threshold quantization processing on the decomposed wavelet decomposition coefficients using the improved threshold function to obtain the processed wavelet coefficients; reconstruct the signal based on the wavelet coefficients after threshold processing to obtain the denoised signal, that is, the denoised intrinsic mode component.

[0087] Improved threshold function Is expressed as:

[0088] ;

[0089] In the formula, and respectively represent the upper threshold and the lower threshold; represents the input data of the improved threshold function.

[0090] The improved threshold function retains the advantages of the hard threshold function, avoids fixed errors, and at the same time takes into account the advantages of the soft threshold function in terms of smoothness and stability to a certain extent, and has good performance.

[0091] The dynamic determination of the upper threshold and the lower threshold is realized by using the wavelet decomposition coefficients. The specific method is as follows: First, estimate the noise standard deviation. Since the traditional mean-based standard deviation is more sensitive to outliers, when there are some large abnormal noise points in the data, the standard deviation will be affected by these extreme values, resulting in an overestimated value of the standard deviation. In addition, some noises do not fully conform to the Gaussian distribution, and the standard deviation may not effectively describe the distribution of the noise. Therefore, the median absolute deviation (MAD) is used instead of the traditional standard deviation to calculate the noise standard deviation, which can reduce the sensitivity to extreme noise values and speed up the calculation process.

[0092] Use the data of the last layer of the detail coefficients to estimate the noise standard deviation; the median absolute deviation and the noise standard deviation Are expressed as:

[0093] ;

[0094] In the formula, The median of the data of the last layer representing the detail coefficient; Represents the noise standard deviation; Is the proportionality constant between the median absolute deviation MAD and the noise standard deviation in the normally distributed data; Represents the input data In the th data point.

[0095] Calculate the lower threshold in the fixed form :

[0096] ;

[0097] In the formula, Represents the length of the input data; Represents the proportionality coefficient between the upper threshold and the lower threshold.

[0098] Step S2: Use the AGCN-TCN model to perform fault identification and classification on the preprocessed fault signal.

[0099] Among them, the deep learning intelligent algorithm of AGCN-TCN excavates the potential connections between multiple monitoring points by improving the graph neural convolutional network AGCN, avoiding decision-making mistakes caused by the failure of a single monitoring device; uses the time-domain convolutional network TCN to extract the time-domain features of the fault signal.

[0100] First, draw a simplified topological graph of the monitoring points through the adjacency situation between each monitoring point. Suppose there is a set of graph structure data containing monitoring points, and each monitoring point has a preprocessed fault signal with a dimension of , forming a -dimensional feature matrix . First, it is necessary to construct an adjacency matrix with a dimension of :

[0101] ;

[0102] In the formula, Is the element that constitutes the adjacency matrix .

[0103] Theoretically, the graph neural convolutional network GCN performs a Fourier transform on the graph structure data and projects it into the Fourier domain, and then introduces a spectral framework to perform convolution in the frequency domain. In order to process the graph structure data after the Fourier transform, first normalize the Laplacian matrix of the target graph :

[0104] ;

[0105] In the formula, is the identity matrix; introduce the adjacency matrix of the diagonal matrix is to realize the renormalization of the adjacency matrix. Subsequently, perform spectral decomposition on the Laplacian matrix: , where is the transpose of is the eigenvector matrix of is the diagonal matrix of eigenvalues.

[0106] Introduce the graph convolution operator to perform convolution operation in the spectral domain on the graph structure data after Fourier transform:

[0107] ;

[0108] In the formula, represents the convolution kernel.

[0109] Through the above operations, obtain the updated features. In order to reduce the computational complexity in the Laplacian matrix decomposition process, use Chebyshev polynomials to construct an approximate model of the graph neural convolutional network GCN :

[0110] ;

[0111] In the formula, ; represents the adjacency matrix after adding self-loops to each node; on this basis, the propagation process of the graph neural convolutional network GCN is:

[0112] ;

[0113] In the formula, represents the node feature matrix of the th layer; represents the graph convolution function of the th layer; represents the node feature matrix of the th layer; represents the weight matrix of the th layer; represents the activation function.

[0114] The adjacency matrix in the traditional graph neural convolutional network GCN It is fixed and can only represent the connection situation between monitoring points with 0 and 1, unable to express the distance or relationship strength between monitoring points. However, the distance between each monitoring point in the distribution network is also an important consideration factor of spatial characteristics. Therefore, the present invention introduces distance weights into the adjacency matrix of the traditional GCN. Specifically, a weighted adjacency matrix is constructed through an exponential decay function. Through the exponential decay function, the weights between nodes change continuously and are negatively correlated with the distance. As the distance increases, the influence (weight) gradually decreases, enabling a more refined modeling of the actual relationship between nodes. The improved adjacency matrix is expressed as:

[0115] ;

[0116] In the formula, represents the physical distance between monitoring point and monitoring point ; represents a hyperparameter used to control the decay rate. The improved adjacency matrix can finely express the relationship strength according to factors such as distance.

[0117] In order to extract the temporal features of the fault signal, the present invention adopts a temporal convolutional network TCN. The temporal convolutional network TCN consists of several one-dimensional convolutional basic layers and residual blocks, including causal dilated convolution and residual connection. It mainly combines the historical information and existing information in the time series through dilated causal convolution. The resulting model only depends on the historical state and does not consider future moments, and its input-output response relationship is a typical causal constraint.

[0118] The specific process of the AGCN-TCN model for fault identification and classification of the preprocessed fault signal is as follows:

[0119] Step S2.1: According to the connection situation and connection distance information of the monitoring points, construct an improved adjacency matrix A considering the connection distance.

[0120] Step S2.2: Use the preprocessed fault signal to construct a data set; divide the data set into a training set and a test set according to 8:2.

[0121] Step S2.3: Input the training set and the improved adjacency matrix A into the AGCN-TCN model to train the model, perform iterative updates according to the set number of training rounds, determine the parameters to be trained in the model. After training, obtain the optimal AGCN-TCN model.

[0122] Step S2.4: Input the test set and the improved adjacency matrix A into the optimal AGCN-TCN model. The improved graph neural convolutional network AGCN extracts the spatial connection between the distribution network topology structure and the monitored fault signal as spatial features ; Input the test set into the Time Domain Convolutional Network (TCN) to extract the temporal features of the monitored fault signals. ; Connect the spatial features and the temporal features through a fully connected layer to obtain the fused features, and input the fused features into the classification layer to match the category of the corresponding fault signal with the corresponding wiring method.

[0123] Step S3: Use the LOF (Local Outlier Factor) outlier monitoring algorithm to calculate the local outlier factor of the monitoring points. Decompose the preprocessed fault signals of each monitoring point according to a sliding window, and assign the sliding window frequency domain obtained after decomposition as the frequency domain fault index. Construct a comprehensive fault location determination index based on the local outlier factor of the monitoring points and the frequency domain fault index, and judge the location of the fault point through the comprehensive fault location determination index.

[0124] The LOF outlier monitoring algorithm uses the k-nearest neighbor distance, the distance neighborhood, reachability distance, and local reachability density and other concepts to represent the outlier factor. The specific steps are as follows:

[0125] Step S3.1: Calculate the k-nearest neighbor distance between the monitoring point and the monitoring point . Use the Euclidean distance as the calculation standard. Assume that both the monitoring point and the monitoring point monitor n fault signals. The Euclidean distance between the monitoring point and the monitoring point is: , represents the i-th fault signal monitored in the monitoring point ; represents the j-th

[0126] fault signal monitored in the monitoring point i-th k-distance and the k-distance neighborhood of the monitoring point

[0127] The k-distance neighborhood is a range centered on the monitoring point with a radius of the k-nearest neighbor distance , which is expressed as:

[0128] ;

[0129] In the formula, represents the monitoring point 's distance neighborhood, that is, all monitoring points that satisfy ; is the set; represents the th distance of the monitoring point, that is, the number of the nearest neighbor points considered when calculating the distance neighborhood; represents the set of monitoring points.

[0130] Step S3.3: Calculate the th reach-distance from the monitoring point to the monitoring point

[0131] ;

[0132] ;

[0133] In the formula, represents the th reach-distance from the monitoring point to the monitoring point when calculating the reachability of the monitoring point to the monitoring point represents the th reach-distance from the monitoring point to the monitoring point here specifically refers to the direct distance; represents the th reach-distance from the monitoring point to the monitoring point here specifically refers to the direct distance; represents the th neighboring distance from the monitoring point

[0134] The local reachability density of the monitoring point is expressed as:

[0135] ;

[0136]

[0136] In the formula, represents the number of th example neighboring points of the monitoring point, satisfying .

[0137] Step S3.4: Calculate the local outlier factor of the monitoring point , which is expressed as:

[0138] ;

[0139] In the formula, represents the local outlier factor of the monitoring point ; represents the local reachability density of the monitoring point ; when approaches 1, it indicates that the monitoring point is a normal monitoring point.

[0140] When is less than 1, it indicates that the monitoring points are relatively dense and are also normal monitoring points; when is much greater than 1, it means that the density around the monitoring point is small. The greater the local outlier factor, the higher the probability of abnormality of the monitoring point .

[0141] Step S3.5: Perform Fourier decomposition on the preprocessed fault signals of each monitoring point according to a sliding window, and assign the sliding window frequency domain obtained after Fourier decomposition as the frequency domain fault index . When the current sliding window frequency domain assignment is greater than the frequency domain assignment of the previous sliding window, it is regarded as the start point of the fault, is 1; when the current sliding window frequency domain assignment is less than the frequency domain assignment of the previous sliding window, it is regarded as the end point of the fault, is 0, where the sliding window is the fault point.

[0142] Step S3.6: Construct a comprehensive index for fault location determination based on the local outlier factor and frequency domain fault index of the monitoring point, and judge the location of the fault point through the comprehensive index for fault location determination.

[0143] The comprehensive index for fault location determination is expressed as:

[0144] ;

[0145] In the formula, represents the set threshold of the local outlier factor; when the monitoring point , it indicates that the monitoring point is determined to be a fault point, and when the monitoring point , it indicates that the monitoring point is not determined to be a fault point.

[0146] Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for grounding line selection analysis and fault location in a distribution network based on signal feature identification, characterized in that: The steps include: Step S1: extracting fault signals of multiple monitoring points around the target distribution network and preprocessing them; Step S2: Use the AGCN-TCN model to perform fault identification and classification on the preprocessed fault signal; The AGCN-TCN model uses the improved graph neural convolutional network AGCN to extract spatial features and the time domain convolutional network TCN to extract temporal features. The spatial features and temporal features are fused through the fully connected layer to obtain fused features. The fused features are input into the classification layer to match the category of the corresponding fault signal with the corresponding wiring method. The improved graph neural convolutional network AGCN improves its own adjacency matrix, that is, introduces distance weights into the adjacency matrix; Step S3: Calculate the local outlier factor of the monitoring point using the LOF outlier monitoring algorithm, decompose the preprocessed fault signal of each monitoring point according to the sliding window, assign the sliding window frequency domain value obtained after the decomposition as the frequency domain fault index, and construct a comprehensive index for fault location determination based on the local outlier factor of the monitoring point and the frequency domain fault index, and determine the location of the fault point through the comprehensive index for fault location determination; The comprehensive index for fault location determination is expressed as: ； In the formula, Indicates the threshold of the local outlier factor; represents the local outlier factor; represents the frequency domain fault indicator; When monitoring point When When the monitoring point is judged as a fault point , indicating the monitoring point Not identified as a fault point.

2. The method for grounding line selection analysis and fault location in a distribution network based on signal feature identification according to claim 1 is characterized in that: The improved adjacency matrix of the improved graph neural convolutional network AGCN is expressed as: ； In the formula, To construct the adjacency matrix Elements of Indicates monitoring point and monitoring points The physical distance between Represents a hyperparameter that controls the rate of decay.

3. The method for grounding line selection analysis and fault location in a distribution network based on signal feature identification according to claim 2 is characterized in that: The preprocessing in step S1 includes decomposing the fault signal into eigenmode components using variational mode decomposition; The improved wavelet threshold method is used to reduce the noise of the intrinsic mode component; The improved wavelet threshold method is to improve the threshold function of the original wavelet threshold method; Improved threshold function It is expressed as: ； In the formula, and denote the upper and lower thresholds respectively; Represents the input data to the improved threshold function.

4. The method for grounding line selection analysis and fault location in a distribution network based on signal feature identification according to claim 3 is characterized in that: The specific process of calculating the local outlier factor of the monitoring point using the LOF outlier monitoring algorithm is as follows: Calculate monitoring points and monitoring points Between Proximity distance , set up monitoring points and monitoring points All monitoring Fault signal, monitoring point and monitoring points of The proximity distance is expressed as: , Indicates monitoring point The monitoring A fault signal; Indicates monitoring point The monitoring A fault signal; Calculate monitoring points No. Distance and Distance neighborhood, expressed as: ； In the formula, Indicates monitoring point of Distance neighborhood, that is, satisfying All monitoring points A collection of; Indicates monitoring point No. Distance, that is, in calculating The number of nearest neighbors to consider when calculating the distance neighborhood; represents a collection of monitoring points; Calculate monitoring points To the monitoring point No. The reachable distance is expressed as: ； ； In the formula, Indicates monitoring point To the monitoring point No. Reachable distance; Indicates monitoring point To the monitoring point The 5th reachable distance; Indicates monitoring point To the monitoring point The 5th reachable distance; Indicates monitoring point and monitoring points Between proximity distance; Indicates monitoring point To oneself The distance to the nearest monitoring point within the neighborhood; Monitoring Points The local reachable density It is expressed as: ； In the formula, Indicates monitoring point No. For example, the number of neighborhood points satisfies ; Calculate monitoring points The local outlier factor of is expressed as: ； In the formula, Indicates monitoring point The local outlier factor of Indicates monitoring point The local reachable density of .

5. The method for grounding line selection analysis and fault location in distribution network based on signal feature identification according to claim 4 is characterized in that: The specific process of decomposing the pre-processed fault signal of each monitoring point according to the sliding window and assigning the sliding window frequency domain value obtained after the decomposition as the frequency domain fault index is as follows: the pre-processed fault signal of each monitoring point is decomposed by Fourier according to the sliding window and the sliding window frequency domain value obtained after the Fourier decomposition is assigned as the frequency domain fault index , when the frequency domain value of the current sliding window is greater than the frequency domain value of the previous sliding window, the current sliding window is regarded as the starting point of the fault. is 1; when the frequency domain value of the current sliding window is less than the frequency domain value of the previous sliding window, the current sliding window is regarded as the end point of the fault. is 0, where the sliding window is the fault point.

6. The method for grounding line selection analysis and fault location in a distribution network based on signal feature identification according to claim 5 is characterized in that: The specific process of using the improved wavelet threshold method to reduce the noise of the intrinsic mode component is as follows: Performing wavelet decomposition on the noisy data signal, i.e., the intrinsic mode component, determining the wavelet decomposition scale according to the scale and sampling frequency of the noisy data signal and obtaining the wavelet decomposition coefficient, wherein the wavelet decomposition coefficient includes an approximation coefficient and a detail coefficient, the approximation coefficient represents the low-frequency component of the noisy data signal, and the detail coefficient represents the high-frequency component of the noisy data signal; The wavelet decomposition coefficients obtained by decomposition are subjected to threshold quantization processing using an improved threshold function to obtain processed wavelet coefficients; the signal is reconstructed based on the wavelet coefficients after threshold processing to obtain a denoised signal, namely, a denoised intrinsic mode component.

7. The method for grounding line selection analysis and fault location in a distribution network based on signal feature identification according to claim 6 is characterized in that: The median absolute deviation is used to calculate the noise standard deviation, and the data of the last layer of detail coefficients is used to estimate the noise standard deviation; the median absolute deviation and noise standard deviation It is expressed as: ； In the formula, The median of the data of the last layer representing the detail coefficients; represents the noise standard deviation; It is the proportionality constant between the median absolute deviation MAD and the noise standard deviation in the normally distributed data; Represents input data The data points; Calculate the lower threshold in fixed form : ； In the formula, Indicates the length of the noisy data signal; Indicates the proportionality factor between the upper and lower thresholds.

8. The method for grounding line selection analysis and fault location in a distribution network based on signal feature identification according to claim 7 is characterized in that: The specific process of using the AGCN-TCN model to perform fault identification and classification on the preprocessed fault signal is as follows: Step S2.1: According to the connection status and connection distance information of the monitoring points, an improved adjacency matrix A considering the connection distance is constructed; Step S2.2: construct a data set using the preprocessed fault signal; divide the data set into a training set and a test set in proportion; Step S2.3: Input the training set and the improved adjacency matrix A into the AGCN-TCN model to train the model, perform iterative updates according to the set number of training rounds, determine the parameters to be trained in the model, and after the training is completed, obtain the optimal AGCN-TCN model; Step S2.4: Input the test set and the improved adjacency matrix A into the optimal AGCN-TCN model, and the improved graph neural convolutional network AGCN extracts the spatial connection between the distribution network topology and the fault signal as the spatial feature; The test set is input into the time domain convolutional network (TCN) to extract the time series features of the monitored fault signal; the spatial features and the time series features are connected through the fully connected layer to obtain the fused features, which are then input into the classification layer to match the category of the corresponding fault signal with the corresponding wiring method.

Images