A power distribution network single-phase ground fault line selection method, system, device and medium

Abstract

The present application relates to the field of distribution network fault line selection in power system relay protection technology, specifically a distribution network single-phase grounding fault line selection method, system, device and medium, including collecting zero sequence time sequence signals of each candidate line after grounding fault, implementing filtering to obtain transient signals after filtering; dividing the transient signals after filtering into transient shock analysis interval and afterglow analysis interval, determining transient main peak point and constructing rising window, extracting afterglow sequence; calculating transient shock sharpness, afterglow fluctuation dispersion and afterglow fractal complexity of each candidate line, constructing transient three-dimensional feature vector to form original feature matrix; performing normalization processing to obtain normalized feature matrix; improving isolated forest algorithm and inputting normalized feature matrix; calculating isolated abnormal score of each candidate line, identifying single-phase grounding fault line and healthy line according to isolated abnormal score and outputting determination result. The present application improves adaptability and automation degree of fault line selection process.

Description

Technical Field

[0001] This invention relates to the field of distribution network fault location in power system relay protection technology, and particularly to a method, system, device and medium for single-phase grounding fault location in distribution networks. Background Technology

[0002] When a single-phase ground fault occurs in a distribution network with an ungrounded neutral point, grounded via an arc suppression coil, or grounded via high resistance, the fault current is usually relatively weak, and the system can continue to operate for a certain period of time. If the faulty line cannot be identified in a timely and accurate manner, the fault area may expand, increasing the difficulty of operation, maintenance, and troubleshooting. Therefore, accurately identifying the line experiencing a single-phase ground fault is of great significance for ensuring the safe and stable operation of the distribution network.

[0003] Existing fault location technologies for single-phase grounding faults in distribution networks are ill-suited to the fault identification needs of distribution networks with ungrounded neutral points, grounded via arc suppression coils, or grounded via high resistance. Current mainstream fault location methods include steady-state amplitude and phase comparison methods, harmonic methods, single time-domain or frequency-domain feature discrimination methods, and intelligent fault location methods relying on manual thresholds, clustering, and nearest neighbor identification. All of these methods have significant technical shortcomings.

[0004] From the perspective of steady-state characteristic detection, existing technologies mostly rely on steady-state zero-sequence amplitude, phase, harmonics, and other electrical quantities to complete fault identification. This is highly susceptible to factors such as arc suppression coil compensation, high-resistance grounding, line parameter fluctuations, power frequency interference, harmonic noise, and load disturbances. Under conditions of weak fault signals, the difference in steady-state characteristics between faulty and healthy lines will significantly decrease, leading to a substantial drop in fault identification stability. From the perspective of transient characteristic utilization, the zero-sequence transient signal of a fault includes multiple complex processes such as the initial impulse and the afterglow evolution, containing multi-dimensional waveform information such as impulse polarity, afterglow amplitude fluctuations, and micro-oscillations. Existing discrimination methods based on single amplitude, single slope, or single statistical index cannot fully explore the differentiated characteristics of transient waveforms, making it difficult to accurately distinguish between faulty and healthy lines. Furthermore, different transient characteristics have inconsistent numerical ranges and dimensions. Directly participating in discrimination without equalization processing will cause features with excessively large values ​​to dominate the line selection results, further reducing the accuracy of the discrimination. In addition, line selection methods such as manual thresholding, manual weighting, clustering, and nearest neighbor identification have poor adaptability. Related parameters need to be repeatedly tuned based on the on-site line structure and operating conditions. Once the power grid structure and fault conditions change, the original parameters become inapplicable, resulting in insufficient generalization ability of the algorithms. Furthermore, the number of fault samples obtainable in actual field conditions is limited, failing to provide sufficient and complete fault and line sample labels for supervised classification algorithms. This further restricts the practical application effect of various line selection methods, easily leading to inaccurate fault identification and expanded fault range, significantly increasing the difficulty of power grid operation and maintenance and fault diagnosis. Summary of the Invention

[0005] To overcome the problems of existing single-phase grounding fault location methods in distribution networks being susceptible to arc suppression coil compensation, high-resistance grounding, and various disturbances, having insufficient ability to characterize single transient or steady-state features, lacking balanced feature dimensions, having poor generalization of manually set parameters, and having limited field fault samples, resulting in limited stability and accuracy in fault line identification, this invention provides a single-phase grounding fault location method for distribution networks, comprising the following steps:

[0006] After a ground fault occurs in the distribution network, the zero-sequence time sequence signal of each candidate line is sampled to form the original transient signal set of each candidate line;

[0007] The original transient signals of each candidate line are filtered to obtain the filtered transient signals;

[0008] The filtered transient signal is divided into intervals according to the sampling time sequence. The first half of the sampling interval with the earlier sampling time sequence is used as the transient impact analysis interval, and the second half of the sampling interval with the later sampling time sequence is used as the afterglow analysis interval. An upward window is constructed in the transient impact analysis interval, and the afterglow sequence is extracted from the afterglow analysis interval.

[0009] The transient impact sharpness is calculated based on the rising window, and the afterglow fluctuation dispersion and afterglow fractal complexity are calculated based on the afterglow sequence. The transient three-dimensional feature vectors of each candidate line are constructed by the transient impact sharpness, afterglow fluctuation dispersion and afterglow fractal complexity. The transient three-dimensional feature vectors of all candidate lines are combined to form the original feature matrix.

[0010] Min-Max normalization is performed on each feature dimension of the original feature matrix to obtain the normalized feature matrix;

[0011] An improved isolated forest algorithm is proposed, which includes randomly generating root nodes of each isolated tree, optimizing the segmentation dimension based on the variance of the node feature dimension, generating adaptive feature weights according to the proportion of feature dimension variance, and recursively segmenting and constructing isolated trees after selecting the actual segmentation threshold based on the criterion of maximizing the variance between weighted subsets. Multiple isolated trees form an improved isolated forest.

[0012] The normalized feature matrix is ​​input into the improved isolated forest algorithm to calculate the isolated anomaly score of each candidate line. Based on the isolated anomaly score, the line with a single-phase ground fault and the line with good condition are identified and the judgment result is output.

[0013] Preferably, each candidate line is an outgoing line from the same bus participating in the fault selection; when sampling the zero-sequence time sequence signal of each candidate line, a fault transient effective window is constructed with the fault triggering time as the starting point, which includes the initial impact process of the fault and the afterglow evolution process after the impact.

[0014] Preferably, filtering is implemented using a second-order custom zero-phase 50Hz notch filter. The filtering parameters of the second-order custom zero-phase 50Hz notch filter are determined by the filter pole radius, the power frequency to be suppressed, the quality factor, the sampling frequency, and the normalized angular frequency, and the expression is:

[0015] ;

[0016] ;

[0017] ;

[0018] ;

[0019] in, Let be the radius of the filter poles. The power frequency to be suppressed. The quality factor of the filter. The sampling frequency of the zero-sequence timing signal. For normalized angular frequency, , These are the numerator and denominator coefficient vectors of the filter, respectively.

[0020] Preferably, the interval division is specifically as follows: the first half of the effective sampling interval of the fault transient window is used as the transient impact analysis interval, and the second half of the sampling interval is used as the afterglow analysis interval. The sampling point with the largest absolute value of the signal amplitude in the transient impact analysis interval is the transient main peak point. The rising window is a continuous sampling interval obtained by backtracking from the transient main peak point and setting a sampling interval. The afterglow sequence is taken from the signals corresponding to all sampling points in the afterglow analysis interval.

[0021] Preferably, the transient impact sharpness is:

[0022] ;

[0023] in, For the first The sharpness of transient impacts on the candidate routes. For the first The amplitude of the filtered transient signal at the sampling position of the transient main peak point for each candidate line. For the first The maximum signed slope between adjacent sampling points within the rising window of the candidate lines, expressed as:

[0024] ;

[0025] in, The time interval between adjacent sampling points. , The first The candidate routes are numbered at the sampling points. , The filtered transient signal value at the location;

[0026] The afterglow fluctuation dispersion is:

[0027] ;

[0028] in, For the first Afterglow fluctuation dispersion of the candidate routes. For the first The candidate route afterglow sequence number Standard deviation of each subinterval For sub-interval numbering, , For the first The mean absolute amplitude of the transient signal after filtering of the candidate lines;

[0029] The fractal complexity of the afterglow is:

[0030] ;

[0031] in, For the first The fractal complexity of the afterglow of the candidate routes, In the first The original fractal dimension obtained by double log-linear fitting on the candidate lines.

[0032] Preferably, the improved isolated forest algorithm is as follows:

[0033] The root node sample set of each isolated tree is generated by random subsampling. The segmentation dimension is optimized based on the feature dimension variance of the node sample. The adaptive feature weight is obtained based on the feature dimension variance ratio. The candidate threshold with the largest variance among the weighted subsets is selected as the actual segmentation threshold. Based on the actual segmentation threshold, multiple isolated trees are recursively segmented to form an improved isolated forest.

[0034] In the process of building an isolated forest, if the maximum feature variance of a node is not lower than the preset variance trigger threshold, the feature dimension with the largest variance is selected as the preferred splitting dimension; when the variances of multiple feature dimensions are equal or the variance difference is less than the preset variance tolerance, a splitting dimension is randomly selected from the candidate feature dimensions; when the variances of all feature dimensions are lower than the preset variance trigger threshold and the effective feature dimension set is not empty, a supplementary splitting dimension is randomly selected; otherwise, the current node is set as a leaf node.

[0035] The adaptive feature weights are generated by the proportion of variance of each feature dimension of the samples within the node. On the preferred segmentation dimension, a set of candidate segmentation thresholds is constructed using the midpoint of adjacent and dissimilar feature values. For each candidate segmentation threshold, the node samples are temporarily divided into left and right candidate subsets and the variance between weighted subsets is calculated. The candidate threshold corresponding to the maximum value of the variance between weighted subsets is taken as the actual segmentation threshold. When multiple candidate thresholds are close, one is randomly selected as the actual segmentation threshold.

[0036] Preferably, the method further includes calculating the isolated path length and the average isolated path length. The isolated path length is the number of branches that the candidate route passes through from the root node of the isolated tree to the leaf node, and the average isolated path length is the statistical average of the isolated path lengths of the route in all isolated trees.

[0037] The formula for scoring isolated anomalies is:

[0038] ;

[0039] in, For the first The isolated anomaly score of the candidate routes, To compare the number of tree samples with a single isolated tree The relevant average isolation path length normalization correction factor is expressed as follows:

[0040] ;

[0041] in, It is Euler's constant;

[0042] The candidate line with the lowest isolated anomaly score is identified as a single-phase ground fault line, while the remaining lines are identified as healthy lines.

[0043] The present invention also provides a single-phase grounding fault location system for a distribution network, used to perform the above method, comprising:

[0044] Signal acquisition module: used to sample the zero-sequence time-sequence signals of each candidate line after a ground fault occurs in the distribution network, and form the original transient signal set of each candidate line;

[0045] Signal preprocessing module: used to filter the original transient signals of each candidate line to obtain the filtered transient signals;

[0046] Feature extraction module: This module divides the filtered transient signal into intervals according to the sampling time sequence. The first half of the sampling interval with the earlier sampling time sequence is used as the transient impact analysis interval, and the second half of the sampling interval with the later sampling time sequence is used as the afterglow analysis interval. A rising window is constructed in the transient impact analysis interval, and the afterglow sequence is extracted from the afterglow analysis interval. The transient impact sharpness is calculated based on the rising window, and the afterglow fluctuation dispersion and afterglow fractal complexity are calculated based on the afterglow sequence. A transient three-dimensional feature vector is constructed from the transient impact sharpness, afterglow fluctuation dispersion, and afterglow fractal complexity of each candidate line. The transient three-dimensional feature vectors of all candidate lines are combined to form the original feature matrix.

[0047] Feature normalization module: used to perform Min-Max normalization on each feature dimension of the original feature matrix to obtain a normalized feature matrix;

[0048] Fault identification module: Improved isolated forest algorithm, including randomly generating root nodes of each isolated tree, optimizing the segmentation dimension based on the variance of node feature dimension, generating adaptive feature weights according to the proportion of feature dimension variance, and recursively segmenting and constructing isolated trees after selecting the actual segmentation threshold based on the maximum variance between weighted subsets. Multiple isolated trees form an improved isolated forest. The normalized feature matrix is ​​input into the improved isolated forest algorithm to calculate the isolated anomaly score of each candidate line.

[0049] Results output module: Used to distinguish between single-phase ground fault lines and healthy lines based on isolated anomaly scores and output the judgment results.

[0050] The present invention also provides a single-phase grounding fault selection device for a power distribution network, comprising: a memory and a processor, wherein the memory is used to store a computer program, and the computer program is used to execute the above method when loaded by the processor.

[0051] The present invention also provides a readable storage medium storing a computer program adapted to execute the above-described method when loaded by a processor.

[0052] The beneficial effects of this invention are as follows:

[0053] This invention fully utilizes the inherent differences in transient waveforms of faults to improve the ability to distinguish faulty lines from healthy lines. Instead of relying on steady-state zero-sequence amplitude, phase, or harmonic characteristics that are easily affected by arc suppression coil compensation, high-resistance grounding, and changes in line parameters, this invention addresses the differences between faulty lines (stronger initial impact, steeper rising edge, smoother afterglow waveform) and healthy lines (continuous small oscillations, higher waveform dispersion, and greater microscopic complexity). It extracts transient impact sharpness, afterglow fluctuation dispersion, and afterglow fractal complexity. These three-dimensional features describe the line waveform state from three dimensions: transient impact, linear statistical dispersion, and nonlinear topology, reducing misjudgments caused by using only a single amplitude, phase, or statistical indicator. The transient impact sharpness and afterglow fluctuation dispersion are derived from the application of the solar flare "initial impact-afterglow evolution" theory to the analysis of zero-sequence transient waveforms in distribution networks, while the afterglow fractal complexity is constructed based on fractal detection theory. The physical meaning of the above characteristics corresponds to the transient waveform change law after a single-phase ground fault occurs in the distribution network, and can be used to characterize the difference between the faulty line and the healthy line, thus it is suitable for fault line selection in the distribution network.

[0054] This invention can suppress power frequency interference and improve the adaptability of processing transient sequences in the field: This invention uses a second-order custom zero-phase notch filter to perform 50Hz notch filtering on the zero-sequence transient signals of each candidate line. Compared with conventional high-order notch filters, this filtering method can adapt to the transient sequences acquired in the field, avoid length exceeding the limit when processing short sequences, and ensure that the filtered signal has no phase shift and no obvious waveform distortion, providing a comparable signal basis for subsequent calculations of transient impulse sharpness, afterglow fluctuation dispersion, and afterglow fractal complexity.

[0055] This invention reduces the impact of feature dimension differences and computational anomalies on fault line selection results: It combines three types of transient features with different physical meanings to form an original feature matrix, and then performs Min-Max normalization processing on each feature, mapping each feature dimension to the same numerical range, thus avoiding the dominance of a single feature with a large numerical magnitude in subsequent discrimination. For situations that may occur during the calculation of afterglow fractal complexity, such as insufficient sequence length, logarithmic operation anomalies, fractal dimensions exceeding the physical range, and non-numerical or infinite values, corresponding fault-tolerant rules are set to ensure that the feature values ​​input to the intelligent discrimination process meet the computational requirements, thereby improving the stability of the fault line selection process.

[0056] This invention reduces errors in invalid segmentation and isolation path calculation under small sample conditions, achieving intelligent fault line selection without relying on fault sample labels and manual classification thresholds. The invention employs a three-dimensional feature-driven unsupervised improved isolated forest AI algorithm. Based on the characteristics of faulty lines exhibiting sparse outlier distribution and healthy lines relatively dense clustering in the transient three-dimensional feature space, an isolated forest is constructed by randomly sampling line samples. At each node, the feature dimension with the largest sample variance is preferentially selected, and adaptive feature weights are determined based on the node variances of the three feature dimensions. The candidate threshold with the largest variance among weighted subsets is used as the actual segmentation threshold. When feature variances are low or evaluation results are tied, constrained random selection is used to maintain the structural differences between different isolated trees. The faulty line is determined based on the average isolated path length and isolated anomaly score of each line. This method does not require pre-accumulation of a large number of labeled fault samples, does not rely on manually setting classification thresholds, feature weights, or the number of clusters, and can automatically output the faulty line number and judgment result, improving the stability, intelligence, and automation of single-phase grounding fault selection in small sample distribution networks. Attached Figure Description

[0057] Figure 1 This is an overall flowchart of an embodiment of the present invention;

[0058] Figure 2 This is the original transient signal diagram of the first candidate line obtained from the simulation of a single-phase grounding fault in a 66kV distribution network according to an embodiment of the present invention.

[0059] Figure 3 This is the original transient signal diagram of the second candidate line obtained from the simulation of a single-phase grounding fault in a 66kV distribution network according to an embodiment of the present invention.

[0060] Figure 4 This is the original transient signal diagram of the third candidate line obtained from the simulation of a single-phase grounding fault in a 66kV distribution network according to an embodiment of the present invention.

[0061] Figure 5 This is the original transient signal diagram of the fourth candidate line obtained from the simulation of a single-phase grounding fault in a 66kV distribution network according to an embodiment of the present invention.

[0062] Figure 6 This is the original transient signal diagram of the 5th candidate line obtained from the simulation of a single-phase grounding fault in a 66kV distribution network according to an embodiment of the present invention.

[0063] Figure 7 This is the original transient signal diagram of the 6th candidate line obtained from the simulation of a single-phase grounding fault in a 66kV distribution network according to an embodiment of the present invention. Detailed Implementation

[0064] Example 1:

[0065] This invention provides a method for selecting single-phase grounding fault lines in a distribution network. This method integrates transient multidimensional features with an intelligent discrimination algorithm. It includes performing a second-order custom zero-phase 50Hz notch filter on the zero-sequence transient signals of each candidate line to extract transient impact sharpness, afterglow fluctuation dispersion, and afterglow fractal complexity. Min-Max normalization is then used to balance the transient three-dimensional features. Subsequently, an unsupervised improved isolated forest artificial intelligence algorithm driven by three-dimensional features is used to select segmentation features based on the variance of each feature dimension within the current node. The segmentation threshold with the largest weighted variance among candidate subsets is selected based on the adaptive weights of the transient three-dimensional features. Multiple isolated trees are constructed, and the average isolated path length and isolated anomaly score of each candidate line are calculated. The line with the smallest isolated anomaly score is identified as the single-phase grounding fault line, thereby improving the adaptability and automation of the fault selection process.

[0066] like Figure 1 As shown, the method for locating a single-phase ground fault in a distribution network specifically includes the following steps:

[0067] S1, Signal Acquisition:

[0068] After the relay protection device determines that a ground fault has occurred in the distribution network, it collects the zero-sequence time sequence signal of each candidate line. Starting from the fault triggering time, it extracts the effective window of the fault transient state from the zero-sequence time sequence signal, which includes the initial impact process of the fault and the afterglow evolution process after the impact, and obtains the original transient signal set of each candidate line.

[0069] The candidate lines are all outgoing lines connected to the same busbar and participating in the selection of the grounding fault line in this distribution network. The effective window for fault transients is calculated by continuously capturing samples along the positive time direction, starting from the sampling point corresponding to the fault triggering time determined by the relay protection device. The effective signal interval formed by each sampling point. Each candidate line uses the same sampling frequency, the same window length, and the corresponding effective signal interval to ensure the comparability of the subsequently extracted transient features.

[0070] Let the total number of candidate lines participating in fault location be . , No. The original transient signal of the candidate lines The expression is:

[0071] ;

[0072] in, Candidate route number, Number the sampling points. This represents the total number of sampling points, i.e., the total number of sampling points of the original transient signal within the effective signal interval.

[0073] S2, Signal Preprocessing:

[0074] To address the issue of trigger length exceeding limits that conventional high-order notch filters are prone to when processing transient sequences in the field, a second-order custom zero-phase 50Hz notch filter was used to filter the original transient signals of each candidate line, resulting in the filtered transient signals for each candidate line. The filtered transient signal has no phase shift, no obvious waveform distortion, and can adapt to transient sequences.

[0075] The filter parameters are determined according to the following expression:

[0076] ;

[0077] ;

[0078] ;

[0079] ;

[0080] in, Let be the radius of the filter poles. The power frequency to be suppressed. The quality factor of the filter. The sampling frequency of the zero-sequence timing signal. For normalized angular frequency, , These are the numerator and denominator coefficient vectors of the filter, respectively.

[0081] S3, Signal interval division:

[0082] Filtered transient signals of each candidate line The sampling interval is divided into two parts: the first half of the sampling interval (the first half of all sampling points in the effective window of the fault transient state) is used as the transient impact analysis interval to extract transient impact features, and the second half of the sampling interval (the second half of all sampling points in the effective window of the fault transient state) is used as the afterglow analysis interval to extract afterglow features. The signal sequence formed by all sampling points in the afterglow analysis interval is the afterglow sequence.

[0083] Within the transient impact analysis interval, the transient main peak point is determined, and an ascending window is constructed by tracing back the sampling interval from the main peak point. Specifically:

[0084] Calculate the absolute value of the amplitude at each sampling point, and determine the sampling point with the largest absolute value. This sampling point is then designated as the transient main peak point. The peak amplitude of the candidate routes is:

[0085] ;

[0086] in, For the first The peak amplitude of the candidate route, i.e., the first The candidate routes were sampled at the transient main peak point. The amplitude of the filtered transient signal at the corresponding location; For the first Sampling points (sampling locations) corresponding to the transient main peak points of the candidate lines;

[0087] The rising window refers to the continuous sampling interval located before the transient peak point, used to calculate the transient impact slope; assuming the rising window traces backward from the transient peak point... Then the sampling interval is... The starting sampling point (starting position) of the rising window for each candidate line is:

[0088] ;

[0089] Sampling points To the sampling point This forms an upward window.

[0090] The afterglow analysis interval is used to describe the amplitude fluctuations and micro-oscillation complexity of the waveform after the initial impact of the fault; let the starting sampling point of the afterglow analysis interval be:

[0091] ;

[0092] in, The integer symbol;

[0093] The total number of sampling points in the filtered transient signal sequence after second-order 50Hz notch filtering. If it remains unchanged, then the first The afterglow sequence of the candidate lines is as follows:

[0094] .

[0095] S4. Multidimensional feature extraction of signals:

[0096] The transient impact sharpness, afterglow fluctuation dispersion, and afterglow fractal complexity of each candidate line are calculated. Transient impact sharpness describes the steepness and polarity of the initial fault impact; afterglow fluctuation dispersion describes the amplitude fluctuation of the afterglow waveform after the impact; and afterglow fractal complexity describes the complexity of the afterglow waveform's wrinkles and fragmented oscillations. The transient impact sharpness and afterglow fluctuation dispersion are derived from the application of solar flare impact and afterglow evolution theory to transient analysis of distribution network faults, while the afterglow fractal complexity originates from fractal detection theory. These three features correspond to time-domain impact, linear statistical dispersion, and nonlinear topology, respectively, and their physical meanings are independent. This avoids misjudgments caused by relying solely on a single amplitude, phase, or statistical index. The three-dimensional features are used together to characterize the transient waveform differences between faulty and healthy lines.

[0097] Calculate the first The signed slope between adjacent sampling points within the rising window of the candidate lines :

[0098] ;

[0099] in, The time interval between adjacent sampling points, with the maximum signed slope represented as... , , The first The candidate routes are numbered at the sampling points. , The filtered transient signal value at the location;

[0100] Calculate the transient impact sharpness of each candidate line:

[0101] ;

[0102] in, For the first The transient impact sharpness on the candidate lines.

[0103] The first The afterglow sequence of the candidate routes is divided into multiple sub-intervals, denoted as . Calculate the standard deviation of the subintervals:

[0104] ;

[0105] in, For the first The candidate route afterglow sequence number Standard deviation of each subinterval As sub-interval numbering, this embodiment divides the afterglow sequence into 8 sub-intervals. , The sign function for calculating the standard deviation;

[0106] Calculate the first The mean absolute amplitude of the transient signal after filtering of the candidate lines :

[0107] ;

[0108] Calculate the first Afterglow fluctuation dispersion of candidate lines :

[0109] .

[0110] At multiple grid scales, the first Perform box coverage on the afterglow sequences of the candidate lines, and calculate the minimum number of boxes required to cover the waveform of the afterglow sequence. Establish the following double logarithmic relationship:

[0111] ;

[0112] in, For grid scale, In the first The original fractal dimension obtained by double log-linear fitting on the candidate lines. Let be the fitting constant. Using the _ , The absolute amplitude mean of the transient signal after filtering of the candidate lines Normalization yields the afterglow fractal complexity. :

[0113] .

[0114] The transient impact sharpness, afterglow fluctuation dispersion, and afterglow fractal complexity obtained for each candidate route are combined to construct a transient three-dimensional feature vector for each candidate route:

[0115] ;

[0116] in, For the first The transient three-dimensional feature vectors of the candidate lines.

[0117] Arrange the transient three-dimensional feature vectors of all candidate routes in order of route number to form the original feature matrix. :

[0118] .

[0119] S5, Signal Feature Equalization Optimization:

[0120] For the original feature matrix Min-Max normalization is performed on each feature dimension:

[0121] ;

[0122] in, The first characteristic matrix of the original feature matrix The candidate routes, the first The normalized feature values ​​corresponding to each feature dimension in this embodiment ; The first characteristic matrix of the original feature matrix The candidate routes, the first The original feature values ​​corresponding to each feature dimension; For all candidate lines in the original feature matrix, the first line is the first line. Minimum value in each feature dimension; For all candidate lines in the original feature matrix, the first line is the first line. The maximum value in each feature dimension;

[0123] By eliminating differences in the dimensions and numerical ranges of different features, a normalized feature matrix is ​​obtained. :

[0124] .

[0125] S6. Building an isolated forest:

[0126] Normalized feature matrix An improved unsupervised isolated forest algorithm driven by input 3D features is proposed, which constructs isolated forests by combining random subsampling, non-random feature selection, and adaptive weighted threshold selection.

[0127] When constructing each isolated tree, a sample of trees is randomly selected from all candidate routes; let the total number of isolated trees be . The number of samples for building a single isolated tree is Constructing the first When there is an isolated tree, from all Randomly selected without replacement from the candidate routes The sample set of each line is used to form the root node sample set of the isolated tree. By retaining random subsampling between trees, different isolated trees have different combinations of tree construction samples; when At that time, it is preferable to use This is to avoid all isolated trees receiving the exact same tree construction sample.

[0128] For any current node in the isolated tree, let the set of line sample numbers contained in that node be . The number of node samples is Calculate the mean and variance of the current node sample across the three normalized feature dimensions:

[0129] ;

[0130] ;

[0131] in, , The sample set of the current node Within, all samples in the first... The sample mean and variance across each feature dimension.

[0132] When the maximum variance among the three feature dimensions is not lower than the preset variance trigger threshold When selecting the feature dimension with the largest variance of the current node's samples, choose it as the preferred segmentation dimension. :

[0133] ;

[0134] When there are two or more feature dimensions with equal variances, or when the difference in variance between feature dimensions is less than the preset variance tolerance, a feature dimension is randomly selected from the corresponding candidate feature dimensions to maintain the diversity of the isolated tree structure.

[0135] When the variance in all three feature dimensions is lower than the preset variance trigger threshold At that time, construct the set of effective feature dimensions for the current node. :

[0136] ;

[0137] in, This is a preset minimum used to determine whether the feature dimensions still have valid value differences. If the set of valid feature dimensions... If the set of valid feature dimensions is not empty, a feature dimension is randomly selected as a supplementary segmentation dimension. If the set of valid feature dimensions is empty, it means that there are no valid differences in the three feature dimensions of each line sample within the current node that can be used for further segmentation. The current node is set as a leaf node and segmentation is no longer continued.

[0138] Calculate adaptive feature weights based on the variance proportions of each feature dimension of the node:

[0139] ;

[0140] in, The sample set of the current node All samples in the first Sample variance across each feature dimension For sample set The current node Adaptive weights corresponding to each feature dimension, satisfying the following:

[0141] ;

[0142] The adaptive feature weights are automatically determined by the actual distribution of the samples at the current node and are not part of the manually preset fault classification weights. When the sum of the variances of the three feature dimensions is not greater than a preset minimum, the current node is set as a leaf node.

[0143] The current node sample is divided in the preferred segmentation dimension. Different eigenvalues Arranged in ascending order, denoted as:

[0144] ;

[0145] Construct a candidate segmentation threshold set using the midpoint of two adjacent distinct feature values. :

[0146] ;

[0147] For any candidate segmentation threshold ∈ Based on the current node sample in the preferred segmentation dimension The value of will temporarily divide the current node's samples into a left candidate subset. and right candidate subset :

[0148] ;

[0149] ;

[0150] Let the left candidate subset be Right candidate subset The sample sizes are respectively , The two candidate subsets are in the first The mean values ​​for each feature dimension are respectively and Candidate segmentation threshold The corresponding weighted subset variance is:

[0151] ;

[0152] The candidate threshold that maximizes the variance among weighted subsets is selected as the actual segmentation threshold for the current node:

[0153] ;

[0154] When the variances among the weighted subsets corresponding to multiple candidate segmentation thresholds are equal, or the difference is less than the preset calculation tolerance, one of the corresponding candidate thresholds is randomly selected as the actual segmentation threshold.

[0155] The left and right candidate subsets are two unlabeled sample subsets temporarily formed by the candidate segmentation threshold, and do not represent the pre-labeled faulty and healthy sample categories. Therefore, the calculation of the weighted variance between subsets does not require faulty and healthy line labels.

[0156] Based on the actual segmentation threshold The current node is divided into a left child node and a right child node, and the feature optimization and threshold optimization processes described above are recursively performed on the two child nodes respectively. When a node contains only one line sample, or when there are no effective feature dimensions within the node that can form two non-empty child nodes, the node is set as a leaf node. After completing the recursive segmentation of all nodes, an isolated tree is formed.

[0157] Repeatedly execute the processes of random subsampling, node feature variance calculation, segmentation feature optimization, weighted segmentation threshold selection, and recursive segmentation to construct a system composed of... Three-dimensional features composed of isolated trees drive unsupervised improvement of isolated forests.

[0158] S7. Perform intelligent fault diagnosis:

[0159] The normalized three-dimensional feature vectors of each candidate line are input into each isolated tree of the constructed improved isolated forest. The isolation path length, average isolation path length, and isolation anomaly score of each candidate line are calculated. The candidate line with the smallest isolation anomaly score is identified as a single-phase ground fault line, and the remaining candidate lines are marked as healthy lines. This result is output as the fault determination result.

[0160] Define the isolation path length: For any candidate route, starting from the root node of the isolated tree, based on the relationship between the normalized feature value of the route in the current node's segmentation dimension and the actual segmentation threshold, proceed sequentially to the left or right child node until a leaf node is reached. The number of branches traversed by the route from the root node to the leaf node of the isolated tree is defined as the isolation path length of the route in that isolated tree.

[0161] Calculate the average isolation path length:

[0162] ;

[0163] in, For the first The average isolation path length of the candidate routes in the entire isolated forest. For the first The candidate routes are in the first The length of the isolation path in an isolated tree; in this embodiment, there are 300 isolated trees in an isolated forest, that is... The smaller the average isolation path length, the easier it is for the line to be separated from other lines in the transient three-dimensional feature space by fewer node segmentations, and the higher its degree of outlier; the larger the average isolation path length, the closer the transient three-dimensional feature distribution of the line is to the other lines, and the higher its probability of being a sound line.

[0164] Calculate the isolated anomaly score for each candidate route:

[0165] ;

[0166] in, For the first The isolated anomaly score of the candidate routes, To compare the number of tree samples with a single isolated tree The relevant average isolation path length normalization correction coefficient is used to correct the isolation path length and eliminate the influence of differences in the number of tree-built samples on the isolation path length. The expression is:

[0167] ;

[0168] in, It is Euler's constant, and ;

[0169] Compare the isolated anomaly scores of all candidate lines, and determine the candidate line with the lowest isolated anomaly score as the single-phase ground fault line:

[0170] ;

[0171] in, The line number is for a single-phase ground fault. argmin represents the total number of candidate lines participating in fault location, and argmin is the score for obtaining isolated anomalies. The operator for the line number corresponding to the minimum value;

[0172] After the fault line is determined, the line number, transient impact sharpness, afterglow fluctuation dispersion, afterglow fractal complexity, and fault determination result of each candidate line are output. Among them, the line with the smallest isolated anomaly score is marked as the fault line, and the remaining candidate lines are marked as healthy lines.

[0173] Example 2:

[0174] This embodiment selects a 66kV distribution network to conduct a simulation test on the single-phase grounding fault selection method of the distribution network described in Embodiment 1. This distribution network contains 6 lines.

[0175] Let the total number of candidate lines participating in fault location be . The total number of sampling points of the original transient signal within the effective signal interval The sampling frequency is 10kHz, the first The original transient signals of the candidate lines are:

[0176] .

[0177] The parameters for the second-order notch filter are set as follows: the power frequency to be suppressed. The sampling frequency of the zero-sequence timing signal is 50Hz. The quality factor of the notch filter is 10kHz. The value is 30, and the formula is:

[0178] ;

[0179] ;

[0180] ;

[0181] .

[0182] No. The filtered transient signal of the candidate lines is .

[0183] In this embodiment If it is 20, then the first The starting position of the rising window for the candidate lines is:

[0184] ;

[0185] The starting sampling point for the afterglow analysis interval is:

[0186] ;

[0187] No. The afterglow sequence of the candidate lines is as follows:

[0188] .

[0189] In this embodiment, the transient impact sharpness of the six candidate lines is calculated as follows:

[0190] .

[0191] The first The afterglow sequence of the candidate routes is divided into 8 sub-intervals. The standard deviation of each sub-interval is calculated.

[0192] .

[0193] In this embodiment, the calculated afterglow fluctuation dispersion of the six candidate lines is as follows:

[0194] .

[0195] In this embodiment, the calculated afterglow fractal complexity of the six candidate lines is as follows:

[0196] .

[0197] Arrange the transient three-dimensional feature vectors of all candidate lines in order of line number to form the original feature matrix:

[0198] .

[0199] The minimum values ​​of the three feature dimensions are calculated as follows:

[0200] ;

[0201] The maximum values ​​of the three feature dimensions are as follows:

[0202] ;

[0203] The normalized feature matrix is:

[0204] .

[0205] In this embodiment, the total number of candidate lines is M=6, the total number of isolated trees is T=300, and the number of tree samples for a single isolated tree is n=5. When constructing each isolated tree, 5 line samples are randomly selected without replacement from the 6 candidate lines to form the root node sample set of that isolated tree. A preset variance trigger threshold is used. The effective value of the feature is extremely small. We construct a 3D feature-driven unsupervised improved isolated forest consisting of 300 isolated trees.

[0206] Based on the average isolation path length of each candidate route, the normalization correction coefficient in this embodiment is... Number of tree samples in the calculation formula It is 5. Calculate the isolated anomaly score for each candidate line, and determine the line with the smallest isolated anomaly score as the single-phase ground fault line based on the isolated anomaly score.

[0207] In this embodiment, the fault location results are shown in the table below:

[0208] .

[0209] This embodiment also provides a single-phase grounding fault location system for a distribution network, used to perform the above method, including:

[0210] Signal acquisition module: used to acquire the zero-sequence time sequence signals of each candidate line after a ground fault occurs in the distribution network, and form the original transient signal set of each candidate line;

[0211] Signal preprocessing module: used to filter the original transient signals of each candidate line to obtain the filtered transient signals;

[0212] Feature extraction module: This module divides the filtered transient signal into intervals, namely transient impact analysis interval and afterglow analysis interval. In the transient impact analysis interval, it determines the transient main peak point and constructs a rising window. It extracts the afterglow sequence from the afterglow analysis interval. Based on the rising window, it calculates the transient impact sharpness. Based on the afterglow sequence, it calculates the afterglow fluctuation dispersion and afterglow fractal complexity. It constructs a transient three-dimensional feature vector from the transient impact sharpness, afterglow fluctuation dispersion, and afterglow fractal complexity of each candidate line. The transient three-dimensional feature vectors of all candidate lines are combined to form the original feature matrix.

[0213] Feature normalization module: used to perform Min-Max normalization on each feature dimension of the original feature matrix to obtain a normalized feature matrix;

[0214] Fault identification module: used to construct an improved isolated forest algorithm by forming multiple isolated trees. The normalized feature matrix is ​​input into the improved isolated forest algorithm to calculate the isolated anomaly score of each candidate line.

[0215] Results output module: Used to distinguish between single-phase ground fault lines and healthy lines based on isolated anomaly scores and output the judgment results.

[0216] This embodiment also provides a single-phase grounding fault location device for a distribution network, including: a memory and a processor. The memory is used to store a computer program, and the computer program is used to execute the above method when loaded by the processor.

[0217] This embodiment also provides a readable storage medium storing a computer program that is adapted to execute the above-described method when loaded by a processor.

[0218] This invention starts from the impact process of the zero-sequence transient waveform of a fault and the evolution of the afterglow after the impact, constructing multi-dimensional features that can characterize the transient impact intensity, the fluctuation of the afterglow amplitude, and the complexity of the afterglow micro-oscillations. The transient impact sharpness and afterglow fluctuation dispersion are derived from the analytical theory of solar flare impact and afterglow evolution, and applied to the analysis of zero-sequence transient waveforms in distribution networks; the afterglow fractal complexity is constructed using fractal detection theory. These features can describe the waveform differences between faulty and healthy lines from different dimensions, and are suitable for single-phase grounding fault selection in distribution networks. Based on this, the isolated forest unsupervised artificial intelligence anomaly detection algorithm is used to identify outliers in the transient feature space, reducing the reliance on fault sample labels, manual classification thresholds, and clustering. A simulation test was conducted using a 66kV distribution network construction example to verify the effectiveness of the invention.

[0219] This invention has been described through embodiments. Those skilled in the art will understand that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of this invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, this invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of this invention.

Claims

1. A method for selecting the fault location of a single-phase grounding fault in a distribution network, characterized in that, Includes the following steps: After a ground fault occurs in the distribution network, the zero-sequence time sequence signal of each candidate line is sampled to form the original transient signal set of each candidate line; The original transient signals of each candidate line are filtered to obtain the filtered transient signals; The filtered transient signal is divided into intervals according to the sampling time sequence. The first half of the sampling interval with the earlier sampling time sequence is used as the transient impact analysis interval, and the second half of the sampling interval with the later sampling time sequence is used as the afterglow analysis interval. An upward window is constructed in the transient impact analysis interval, and the afterglow sequence is extracted from the afterglow analysis interval. The transient impact sharpness is calculated based on the rising window, and the afterglow fluctuation dispersion and afterglow fractal complexity are calculated based on the afterglow sequence. The transient three-dimensional feature vectors of each candidate line are constructed by the transient impact sharpness, afterglow fluctuation dispersion and afterglow fractal complexity. The transient three-dimensional feature vectors of all candidate lines are combined to form the original feature matrix. Min-Max normalization is performed on each feature dimension of the original feature matrix to obtain the normalized feature matrix; An improved isolated forest algorithm is proposed, which includes randomly generating root nodes of each isolated tree, optimizing the segmentation dimension based on the variance of the node feature dimension, generating adaptive feature weights according to the proportion of feature dimension variance, and recursively segmenting and constructing isolated trees after selecting the actual segmentation threshold based on the criterion of maximizing the variance between weighted subsets. Multiple isolated trees form an improved isolated forest. The normalized feature matrix is ​​input into the improved isolated forest algorithm to calculate the isolated anomaly score of each candidate line. Based on the isolated anomaly score, the line with a single-phase ground fault is distinguished from the healthy line, and the judgment result is output.

2. The method for selecting the fault location of a single-phase grounding fault in a distribution network according to claim 1, characterized in that, Each candidate line is an outgoing line from the same bus participating in the fault selection process; when sampling the zero-sequence time sequence signal of each candidate line, a fault transient effective window is constructed with the fault triggering time as the starting point, which includes the initial impact process of the fault and the afterglow evolution process after the impact.

3. The method for selecting the fault location of a single-phase grounding fault in a distribution network according to claim 1, characterized in that, The filtering is implemented using a second-order custom zero-phase 50Hz notch filter. The filter parameters of the second-order custom zero-phase 50Hz notch filter are determined by the filter pole radius, the power frequency to be suppressed, the quality factor, the sampling frequency, and the normalized angular frequency. The expression is as follows: ； ； ； ； in, Let be the radius of the filter poles. The power frequency to be suppressed. This is the quality factor of the filter. The sampling frequency of the zero-sequence timing signal. For normalized angular frequency, , These are the numerator and denominator coefficient vectors of the filter, respectively.

4. The method for selecting the fault location of a single-phase grounding fault in a distribution network according to claim 2, characterized in that, The interval division is specifically defined as follows: the first half of the effective sampling window of the fault transient is used as the transient impact analysis interval, and the second half of the sampling interval is used as the afterglow analysis interval. The sampling point with the largest absolute value of signal amplitude in the transient impact analysis interval is the transient main peak point. The rising window is a continuous sampling interval obtained by backtracking from the transient main peak point and setting a sampling interval. The afterglow sequence is taken from the signals corresponding to all sampling points in the afterglow analysis interval.

5. The method for selecting the line of a single-phase grounding fault in a distribution network according to claim 4, characterized in that, The transient impact sharpness is: ； in, For the first The sharpness of transient impacts on the candidate routes. For the first The amplitude of the filtered transient signal at the sampling position of the transient main peak point of each candidate line. For the first The maximum signed slope between adjacent sampling points within the rising window of the candidate lines, wherein the expression for the signed slope is: ； in, The time interval between adjacent sampling points. , The first The candidate routes are numbered at the sampling points. , The filtered transient signal value at the location; The afterglow fluctuation dispersion is: ； in, For the first Afterglow fluctuation dispersion of the candidate routes For the first The candidate route afterglow sequence number The standard deviation of each subinterval For sub-interval numbering, , For the first The mean absolute amplitude of the transient signal after filtering of the candidate lines; The afterglow fractal complexity is: ； in, For the first The fractal complexity of the afterglow of the candidate routes, In the first The original fractal dimension obtained by double log-linear fitting on the candidate lines.

6. The method for selecting the fault location of a single-phase grounding fault in a distribution network according to claim 1, characterized in that, The improved isolated forest algorithm is specifically as follows: The root node sample set of each isolated tree is generated by random subsampling. The segmentation dimension is optimized based on the feature dimension variance of the node sample. The adaptive feature weight is obtained based on the feature dimension variance ratio. The candidate threshold with the largest variance among the weighted subsets is selected as the actual segmentation threshold. Based on the actual segmentation threshold, multiple isolated trees are recursively segmented to form an improved isolated forest. In the process of building an isolated forest, if the maximum feature variance of a node is not lower than the preset variance trigger threshold, the feature dimension with the largest variance is selected as the preferred splitting dimension; when the variances of multiple feature dimensions are equal or the variance difference is less than the preset variance tolerance, a splitting dimension is randomly selected from the candidate feature dimensions; when the variances of all feature dimensions are lower than the preset variance trigger threshold and the effective feature dimension set is not empty, a supplementary splitting dimension is randomly selected; otherwise, the current node is set as a leaf node. The adaptive feature weights are generated by the proportion of variance of each feature dimension of the samples within the node. On the preferred segmentation dimension, a set of candidate segmentation thresholds is constructed using the midpoints of adjacent and dissimilar feature values. For each candidate segmentation threshold, the node samples are temporarily divided into left and right candidate subsets and the variance between weighted subsets is calculated. The candidate threshold corresponding to the maximum value of the variance between weighted subsets is taken as the actual segmentation threshold. When multiple candidate thresholds are close, one is randomly selected as the actual segmentation threshold.

7. The method for selecting the fault location of a single-phase grounding fault in a distribution network according to claim 6, characterized in that, It also includes calculating the isolated path length and the average isolated path length, wherein the isolated path length is the number of branches that the candidate route passes through from the root node of the isolated tree to the leaf node, and the average isolated path length is the statistical average of the isolated path length of the route in all isolated trees; The formula for scoring isolated anomalies is: ； in, For the first The isolated anomaly score of the candidate routes, To the number of tree samples compared to a single isolated tree The relevant average isolation path length normalization correction factor is expressed as follows: ； in, Here is Euler's constant; The candidate line with the lowest isolated anomaly score is identified as a single-phase ground fault line, and the remaining lines are identified as healthy lines.

8. A single-phase ground fault location system for a distribution network, used to perform the method described in any one of claims 1-7, characterized in that, include: Signal acquisition module: used to sample the zero-sequence time sequence signal of each candidate line after a ground fault occurs in the distribution network, and form the original transient signal set of each candidate line; Signal preprocessing module: used to filter the original transient signals of each candidate line to obtain the filtered transient signals; Feature extraction module: This module divides the filtered transient signal into intervals according to the sampling time sequence. The first half of the sampling interval with the earlier sampling time sequence is used as the transient impact analysis interval, and the second half of the sampling interval with the later sampling time sequence is used as the afterglow analysis interval. A rising window is constructed in the transient impact analysis interval, and the afterglow sequence is extracted from the afterglow analysis interval. The transient impact sharpness is calculated based on the rising window, and the afterglow fluctuation dispersion and afterglow fractal complexity are calculated based on the afterglow sequence. A transient three-dimensional feature vector is constructed from the transient impact sharpness, afterglow fluctuation dispersion, and afterglow fractal complexity of each candidate line. The transient three-dimensional feature vectors of all candidate lines are combined to form the original feature matrix. Feature normalization module: used to perform Min-Max normalization on each feature dimension of the original feature matrix to obtain a normalized feature matrix; Fault identification module: Improved isolated forest algorithm, including randomly generating root nodes of each isolated tree, optimizing the segmentation dimension based on the variance of node feature dimension, generating adaptive feature weights according to the proportion of feature dimension variance, and recursively segmenting and constructing isolated trees after selecting the actual segmentation threshold based on the maximum variance between weighted subsets. Multiple isolated trees form an improved isolated forest. The normalized feature matrix is ​​input into the improved isolated forest algorithm to calculate the isolated anomaly score of each candidate line. Results output module: Used to distinguish between single-phase ground fault lines and healthy lines based on isolated anomaly scores and output the judgment results.

9. A single-phase grounding fault location device for a power distribution network, characterized in that, include: A memory and a processor, wherein the memory is used to store a computer program, and the computer program is used to execute the method of any one of claims 1-7 when loaded by the processor.

10. A readable storage medium, characterized in that, The storage medium stores a computer program that is adapted to execute the method described in any one of claims 1-7 when loaded by a processor.

Images