Power distribution network fault location method and location system

Abstract

The application discloses a power distribution network fault positioning method and positioning system, and relates to the technical field of power distribution network fault positioning. The application collects multiple groups of historical data of different faults of a power distribution network, obtains missing intervals of voltage and current sequences in each group of historical data, positions the missing intervals to the front, middle and rear of the entire sequence, combines the length of the missing intervals, the distance between the starting position of the missing interval and the nearest missing interval in front, the distance between the ending position of the missing interval and the nearest missing interval behind, obtains different interpolation methods of different missing intervals, marks abnormal values of the completed interpolation sequence, divides the sequence into Sfront and Srear with equal length, respectively uses ELM to repair the abnormal values, obtains standard data sets belonging to the fault, carries out clustering processing on the data in the standard data sets, eliminates irrelevant clusters, and positions to a specific position according to the similarity between the fault and each group of faults in the target cluster. The application effectively improves the fault positioning precision of the power distribution network.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network fault location, and in particular to a distribution network fault location method and location system. Background Art

[0002] During the operation of the distribution network, the occurrence of faults will affect the reliability and stability of the power supply. Accurate and rapid fault location is crucial to reducing power outage losses. Due to the complex operating environment, the voltage, current and other data collected by sensors often have missing values ​​and outliers. If handled improperly, it will lead to distortion of fault characteristics. The existing preprocessing methods have obvious defects: when processing missing values, traditional interpolation techniques adopt a unified strategy, such as linear interpolation and mean filling, and do not combine the missing position, interval length and the correlation characteristics of the previous and next data for interpolation, resulting in filling deviation. In outlier processing, traditional methods often directly eliminate abnormal data, such as the interquartile range method, but ignore that it may be a key representation of the fault, resulting in the loss of fault feature information and the inability to obtain the true fault pattern. In addition, the existing methods do not incorporate the data time series dependency. The processed data is prone to deviate from the actual trend, which further leads to fault feature vector calculation deviation, causing cluster centroid offset or cluster division errors, and ultimately causing fault misjudgment or decreased positioning accuracy. Summary of the Invention

[0003] In response to the deficiencies of the prior art, the present invention provides a distribution network fault location method and location system, which solves the problem of inaccurate fault location caused by improper processing of missing data values ​​and abnormal values.

[0004] To achieve the above objectives, the present invention is implemented through the following technical solutions: a distribution network fault location method, comprising:

[0005] S1. Obtain the missing intervals of the voltage and current sequences in each set of historical data and locate them at the front, middle, and back of the entire sequence. Combined with the missing interval length L, the distance Dfront from the starting point of the missing interval to the nearest missing interval before it, and the distance Dback from the ending point of the missing interval to the nearest missing interval after it, different interpolation methods are obtained for different missing intervals. Outliers are marked on the interpolated sequences and the sequences are divided into a front group Sfront and a back group Sback of equal length. The extreme learning machine (ELM) bidirectional prediction is used to repair the outliers.

[0006] S2. Construct discriminant features based on the processed voltage and current sequences and perform normalization processing to obtain a standard data set belonging to this type of fault. Cluster the data in the standard data set to find clusters containing light and heavy faults. Eliminate irrelevant clusters based on the minimum time interval and time similarity, calculate the centroid of the remaining clusters, calculate the distance di from the existing fault to the centroid of the remaining clusters of different faults to locate the target cluster, and then calculate the similarity between the fault and each group of faults in the target cluster to locate the specific location.

[0007] As a further solution of the present invention, S0 is also included before S1, and S0 is used to collect multiple groups of historical data of different faults in the distribution network. The specific historical data include voltage sequence and current sequence within the fault duration, as well as fault level, timestamp, and specific location.

[0008] As a further embodiment of the present invention, the method for locating the specific position of the deletion interval in the entire sequence is:

[0009] Calculate the total length M of the sequence and define the starting index start and end of the missing interval. If the missing interval contains only one value, then start is equal to end.

[0010] If end<=M*A, the missing interval is located at the front of the sequence;

[0011] If start>M*A and end<=M*(A+B), the missing interval is located in the middle of the sequence;

[0012] If start>M*(A+B), the missing interval is located at the end of the sequence, where A, B, C∈(0,1) and A+B+C=1.

[0013] As a further solution of the present invention, if the missing interval is located at the front of the sequence, the selected interpolation method specifically includes:

[0014] If L<=Lth and max{Dfront, Dback}<=Dth, select the nearest neighbor interpolation method;

[0015] If L<=Lth and max{Dfront, Dback}>Dth, select linear interpolation;

[0016] If L>Lth and max{Dbefore, Dafter}<=Dth, select the interpolation method used for the missing interval closest to the current missing interval, that is, the interpolation method used for the missing interval corresponding to min{Dbefore, Dafter};

[0017] If L>Lth and max{Dfront, Dback}>Dth, select the linear extrapolation method.

[0018] As a further solution of the present invention, if the missing interval is located in the middle of the sequence, the selected interpolation method specifically includes:

[0019] If L<=Lth and max{Dfront, Dback}<=Dth, select the nearest neighbor interpolation method;

[0020] If L<=Lth and max{Dbefore,Dafter}>Dth, choose the weighted moving average method;

[0021] If L>Lth and max{Dfront, Drear}<=Dth, the interpolation method used by the missing interval closest to the current missing interval, i.e.

[0022] If L>Lth and max{Dfront, Drear}>Dth, Lagrange interpolation or cubic spline interpolation is selected.

[0023] As a further scheme of the present application, if the missing interval is positioned at the rear part of the sequence, the selected interpolation method specifically comprises:

[0024] If L<=Lth and max{Dfront, Drear}<=Dth, nearest neighbor interpolation is adopted;

[0025] If L<=Lth and max{Dfront, Drear}>Dth, linear interpolation is adopted.

[0026] If L>Lth and max{Dfront, Drear}<=Dth, the interpolation method used by the missing interval closest to the current missing interval, i.e.

[0027] If L>Lth and max{Dfront, Drear}>Dth, exponential smoothing method can be adopted.

[0028] As a further scheme of the present application, the specific operation method for repairing abnormal values by using the extreme learning machine ELM bidirectional prediction is as follows:

[0029] The window size W is set, the local mean μ and the standard deviation σ are calculated, if a point xt satisfies |xt-μ|>k*σ, the point is marked as a potential abnormal point.

[0030] The sequence is equally divided into a front group Sfront and a rear group Srear, the interquartile range method is used to detect the abnormal points in Sfront and Srear respectively, the abnormal points in Sfront and Srear are removed, and the pure training sets Sfront_pure and Srear_pure are obtained.

[0031] The ELM model is trained by using Sfront_pure and Srear_pure respectively, and the trained ELM model is used to predict Srear and Sfront respectively, and the predicted value sequences Srear_prediction and Sfront_prediction are obtained.

[0032] For each point yt in the rear group, yt_pred is in Srear_prediction, |yt-yt_pred| is calculated, if the point exceeds the specified threshold and the point is marked as a potential abnormal point, yt_pred is used to replace yt, and similarly, the above operation is also performed on Sfront.

[0033] The repaired Sfront and Srear are combined into a complete sequence, and the normal points not marked are retained, wherein k is a proportion coefficient.

[0034] As a further solution of the present invention, t>Tmax and S' <S0和t<Tmin且S’<S0时的簇剔除,其中,t为轻、重故障之间最小的时间间隔,S’为时间相似度,Tmin、Tmax、S0分别为时间阈值下限、时间阈值上限、时间相似度阈值。

[0035] As a further solution of the present invention, the cluster corresponding to min(di) is found as the target cluster, and the fault corresponding to the target cluster is used as the specific fault type that occurred in the distribution network during that period. The cosine similarity Si between the fault feature vector and the feature vector of each historical fault in the target cluster is calculated, and the fault location corresponding to min(Si) is obtained as the real-time location of the current fault.

[0036] The distribution network fault location system includes: a data acquisition module, a missing value processing module, an abnormal value processing module, and a fault location module;

[0037] Data acquisition module, which collects multiple sets of historical data of different faults in the distribution network and transmits them to the missing value processing module;

[0038] The missing value processing module obtains the missing intervals of the voltage and current series in each set of historical data, locates them at the front, middle, and back of the entire series, and combines the length of the missing interval, the distance from the starting point of the missing interval to the nearest missing interval, and the distance from the end point of the missing interval to the nearest missing interval to obtain different interpolation methods for different missing intervals. The processed series are then transmitted to the outlier processing module.

[0039] The outlier processing module marks the outliers in the interpolated sequence and divides the sequence into a front group S before and a back group S after of equal length. It then uses the extreme learning machine (ELM) bidirectional prediction to repair the outliers and transmits the repaired sequence to the fault location module.

[0040] The fault location module constructs discriminant features based on the processed voltage and current sequences, obtains a standard data set belonging to this type of fault, clusters the data in the standard data set, finds clusters containing light and heavy faults, eliminates irrelevant clusters based on the minimum time interval and time similarity, calculates the centroid of the remaining clusters, calculates the distance from the existing fault to the centroid of the remaining clusters of different faults to locate the target cluster, and then calculates the similarity between the fault and each group of faults in the target cluster to locate the specific location of the fault.

[0041] The present invention provides a distribution network fault location method and location system, which have the following advantages compared with the prior art:

[0042] (1) The present application can more accurately fill in missing values, reduce data bias and improve data quality by selecting different interpolation methods for different missing value positions.

[0043] (2) When dealing with abnormal values, the present application uses limit learning machine to predict the sequence before and after, uses the predicted value to replace the abnormal value, avoids the problem caused by directly eliminating the abnormal value, better mines the hidden relationship between data, and improves the accuracy of fault judgment.

[0044] (3) The present application can more accurately locate the fault type and specific position of the power distribution network by screening and calculating the centroid of the clustering result and comparing the fault feature vectors, thereby improving the accuracy and efficiency of fault location. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 The steps of the present application are as follows:

[0046] Figure 2 The system principle block of the present application is as follows. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0048] As Figure 1 , the present application provides a power distribution network fault positioning method, comprising:

[0049] For each type of fault in the power distribution network, a plurality of sets of historical data are widely collected, which are marked in detail with fault levels, which can be divided into three levels of light, medium and heavy, and the specific position and timestamp of the fault occurrence are accurately recorded, and each set of data completely covers the voltage [v1, v2,..., vn] and current [i1, i2,..., in] after the fault occurs.

[0050] Taking a certain city power distribution network as an example, in a fault caused by a tree branch touching the line, the monitoring equipment records the voltage sequence [215, 210, null, 200, 195] after the fault occurs, the current sequence [80, 90, null, 110, 120], the timestamp is "2024-10-15 14:30:00", and the fault level is evaluated as light, and the fault position is located near the 3rd pole of a certain street, wherein null represents a missing value.

[0051] In actual monitoring, missing values ​​often appear in voltage and current series due to factors such as environmental interference and equipment failure. One existing approach is to directly interpolate these series using various interpolation methods. However, a series may sometimes contain more than one missing value, and the length of the missing values ​​can also affect interpolation accuracy. Another approach is to use existing machine learning models to learn from existing data for interpolation. However, this approach does not easily extract temporal features between data, resulting in incomplete features learned by the model and large deviations in subsequent interpolated predictions.

[0052] Therefore, the present invention divides the sequence into three parts: front, middle, and back. The specific division rules are as follows: first calculate the total length M of the sequence, then define the starting indexes start and end of the missing interval. If the missing interval contains only one value, then start is equal to end. The front part means that the missing interval is completely located in the front A region of the sequence, that is, end<=M*A, close to the beginning of the sequence; the middle part means that the missing interval is located in the middle B region of the sequence, that is, start>M*A and end<=M*(A+B), which is in the middle position away from the beginning and end of the sequence; the back part means that the missing interval is completely located in the back C region of the sequence, that is, start>M*(A+B), close to the end of the sequence; if the missing interval spans multiple regions, the missing interval is divided according to the actual situation, and the front, middle, and back parts are determined for different divided parts, where A, B, C∈(0,1) and A+B+C=1;

[0053] For example, the voltage sequence of a distribution network when a fault occurs is [215, 210, null, null, null, 233, 244, 255, 213]. The total length M of the sequence is 9. The starting index of the missing interval is start=3 and end=5, A=1 / 3, B=1 / 3, and C=1 / 3. The front part is end<=3, the middle part is start>3 and end<=6, and the back part is start>6. The missing interval is [3, 5]. Index 3 is in the front part, and indices 4 and 5 are in the middle part. Therefore, the missing interval can be divided and the missing values ​​in the front and middle parts can be processed using different methods to make the interpolation more accurate.

[0054] After determining the position of the missing interval in the entire sequence, calculate the missing interval length L, the front distance D from the starting point of the missing interval to the nearest missing interval, and the back distance D from the end point of the missing interval to the nearest missing interval. Combine these three parameters to determine the interpolation method applicable to different intervals.

[0055] The missing interval length L reflects the amount of data that needs to be interpolated. When L is small, it means that there is less missing data, and interpolation can rely more on a small amount of nearby known data. When L is large, it means that there is more missing data, and it is necessary to consider more macro characteristics such as the overall trend and periodicity of the data when interpolating.

[0056] Dpre reflects the richness of the front-side data that can be used as a reference during interpolation. The larger the Dpre, the more data available for reference, and the better the historical trend of the data can be grasped during interpolation. The smaller the Dpre, the limited data available, and interpolation may only be based on a few data adjacent to the missing interval.

[0057] D posterior reflects the amount of back-end data that can be used as a reference for interpolation. The larger D posterior is, the richer the back-end data is, which can provide more information about the direction of subsequent data for interpolation. The smaller D posterior is, the less back-end data is, and the degree of reliance on back-end data during interpolation will be reduced.

[0058] When the missing interval has an index at the front of the sequence, if L <= Lth and max{Dfront, Dback} <= Dth, the nearest neighbor interpolation method is used. The nearest neighbor interpolation method is a simple and direct method that selects the value of the known data point closest to the missing value to fill the missing value. If L <= Lth and max{Dfront, Dback} > Dth, it indicates that there is rich data information in the front or back of the missing interval, and the linear interpolation method can be used. For example, if the total length M of a voltage sequence is 10 and the missing interval is [2,3], that is, the second and third data are missing, then L is 2. If Lth is set to 3, D before is 1, D after is 2, and Dth is 3, then the missing interval is located at the front of the sequence. At the same time, there is insufficient data for reference on both the front and back sides of the missing interval, and the missing interval is short. In this case, the nearest neighbor interpolation method is used to fill the second missing value with the value of the first data and the third missing value with the value of the fourth data. This method is simple to calculate and can quickly and effectively fill missing values ​​in short missing scenarios to ensure data consistency.

[0059] If L>Lth and max{Dfront, Dback}<=Dth, it means that there are many values ​​to be filled in the missing interval, but the amount of data on the front and back sides of the missing interval is very small. Therefore, simply interpolating and inferring through the data on the front and back sides will lead to large errors. At this time, the interpolation method used for the missing interval closest to the current missing interval is selected, that is, the interpolation method used for the missing interval corresponding to min{Dfront, Dback}. If L>Lth and max{Dfront, Dback}>Dth, the linear extrapolation method is used. The linear extrapolation method predicts the missing values ​​based on the growth or decline trend of the data on the side corresponding to max{Dfront, Dback}. For example, a current sequence has a total length of M of 15, a missing interval of [3, 6], L of 4, which is greater than the Lth value of 3, and D of 6, which is greater than the Dth value of 3. If the trailing data shows a linear growth trend, and the seventh data is known to be 120 and the eighth data is 130, the missing value can be predicted through linear calculation. According to the linear growth law, the data growth rate is first calculated, and then the missing value is predicted based on this rate and the known data. This method is suitable for front-end missing situations where the missing interval is long but the leading or trailing data has a clear trend, avoiding the large errors caused by simple interpolation.

[0060] When the missing interval has an index in the sequence, if L<=Lth and max{Dfront,Dback}<=Dth, the nearest neighbor interpolation method is used. If L<=Lth and max{Dfront,Dback}>Dth, it indicates that there is rich data information on the front or back side of the missing interval. The weighted moving average method can be used. This method combines the trend and weight of the previous and next data to calculate the missing value. By assigning different weights to data at different time points, it can balance the overall trend and local fluctuations of the data. When the data fluctuates greatly but there is a certain trend, this method can better comprehensively consider various data characteristics and obtain a more reasonable missing value estimate.

[0061] If L>Lth and max{Dfront, Dback}<=Dth, select the interpolation method used for the missing interval closest to the current missing interval, that is, the interpolation method used for the missing interval corresponding to min{Dfront, Dback}. If L>Lth and max{Dfront, Dback}>Dth, Lagrange interpolation or cubic spline interpolation can be used. Lagrange interpolation calculates missing values ​​by constructing a polynomial by using known previous and next data points. Its advantage is that it can use multiple known data points to fit the data curve by constructing a polynomial, thereby more accurately calculating the missing values. Cubic spline interpolation can ensure the smoothness of the interpolation curve and is more in line with the actual changes in the data. When processing data with a certain degree of continuity and smoothness, cubic spline interpolation can make the interpolated curve smoother, avoid mutation points, and more accurately reflect the inherent laws of the data. For example, if the total length M of a sequence is 20, the missing interval is [8,9], L is 5, which is greater than the Lth value of 3, D is 5 before and 6 after, both greater than the Dth value of 3, then Lagrange interpolation is used to construct a polynomial using the known points before and after. This can fully consider the changing trend of the surrounding data and obtain a more accurate missing value.

[0062] When the missing interval has an index at the end of the sequence, if L<=Lth and max{Dfront,Dback}<=Dth, the nearest neighbor interpolation method is used; if L<=Lth and max{Dfront,Dback}>Dth, the linear interpolation method is used;

[0063] If L > Lth and max{Dfront, Dback} <= Dth, select the interpolation method for the missing interval closest to the current missing interval, that is, the interpolation method for the missing interval corresponding to min{Dfront, Dback}. If L > Lth and max{Dfront, Dback} > Dth, use exponential smoothing. This method predicts missing values ​​based on the historical trends and weights of the preceding or following data, better adapting to data fluctuations. It assigns greater weight to recent data and less weight to distant data, placing greater emphasis on recent data trends. For example, if the total length of a current series, M, is 18, the missing interval is [15, 18], L is 4 (greater than Lth, 3), and Dfront is 6 (greater than Dth, 3). In this case, exponential smoothing can fully utilize the historical trends of the preceding data and predict missing values ​​through weighted calculations, making the prediction more consistent with the actual data variation. This method is particularly suitable for scenarios with trendy data and long missing intervals in the following data.

[0064] After the voltage and current series are reasonably interpolated using the above method, these series need to be processed for outliers. Traditional outlier processing methods are mostly based on statistical rules, distance metrics, and density, and directly remove outliers that do not conform to the rules. However, in distribution network data, these outliers are likely caused by faults. If they are directly removed, subsequent fault diagnosis will be more difficult and the difficulty of mining hidden relationships between data will be increased. This paper adopts a new outlier processing method, using extreme learning machine (ELM) bidirectional prediction to repair outliers. The specific operation method is as follows:

[0065] Set the window size W, calculate the local mean μ and standard deviation σ, and if a point xt satisfies |xt-μ|>k*σ, mark the point as a potential outlier;

[0066] The sequences are divided into a front group Spre and a back group Spost of equal length. The interquartile range method is used to detect outliers in Spre and Spost respectively. The outliers in Spre and Spost are removed to obtain the pure training sets Spre_pure and Spost_pure.

[0067] Use Sbefore_pure and Sbefore_pure to train ELM models respectively, and use the trained ELM models to predict Sbefore and Sbefore respectively, and obtain the predicted value sequences Sbefore_prediction and Sbefore_prediction;

[0068] For each point yt∈S after the posterior group, yt_pred∈S after_prediction, calculate |yt-yt_pred|. If the point exceeds the specified threshold and is marked as a potential outlier, yt is replaced by yt_pred. Similarly, the above operation is performed on S before.

[0069] The repaired pre-S and post-S sequences are merged into a complete sequence, and the unmarked normal points are retained.

[0070] Based on the processed voltage and current sequences, discriminant features are constructed, including voltage mean, voltage standard deviation, current mean, current standard deviation, and duration. These discriminant features can comprehensively reflect the data characteristics at the time of the fault and provide a basis for fault diagnosis. Min-max normalization is performed on multiple data sets to convert data of different ranges and magnitudes to a unified scale, eliminating the influence of data dimension and obtaining a standard data set for this type of fault.

[0071] Each set of data in the standard data set has its corresponding timestamp and specific location. Clustering is performed on the data in the standard data set. For example, the K-Means clustering algorithm is used to obtain m clusters. The corresponding fault level in each cluster is recorded.

[0072] Identify clusters containing minor and major faults, sort the data in these clusters according to their corresponding timestamps, calculate the minimum time interval t between minor and major faults, compare t with the specified time threshold T, and calculate the temporal similarity S' of the two sets of data using dynamic time warping (DTW).

[0073] Let t>Tmax and S' <S0时的簇剔除,若轻故障数据与重故障数据之间的时间间隔t大于规定的时间阈值Tmax,这表明轻、重故障发生的时间间隔较长,它们之间的关联性较弱,而时间相似度S小于设定的相似度阈值S0,意味着这两组数据在时间分布上差异较大,在这种情况下,该簇中的轻、重故障数据可能来自不同的故障事件或故障发展阶段,它们被聚类到一起可能是由于聚类算法的误差或数据的异常波动导致的,此时需要剔除此簇,因为它不能准确反映出具有相似特征的故障数据集合,保留该簇可能会干扰后续对故障的分析和判断;

[0074] t <Tmin且S’<S0时的簇剔除,虽然轻、重故障数据之间的时间间隔t小于时间阈值Tmin,说明故障发生时间较为接近,但时间相似度S却小于设定阈值S0,说明它们在时间分布模式上差异明显,这可能意味着在短时间内发生的轻、重故障并非同一故障发展过程中的不同阶段,而是由不同原因导致的独立故障,这种情况下,该簇内的数据特征不一致,也不符合聚类的初衷,若保留该簇,可能会在后续分析中产生混淆,因此需要将其剔除,以保证聚类结果的准确性和可靠性;

[0075] For example, there are two clusters, one of which contains minor fault data and major fault data. The minor fault timestamp is 10:00 and the major fault timestamp is 11:30. The time threshold Tmax is set to 1 and Tmin is set to 0.5. The time interval t is calculated to be 1.5 hours. The time similarity S calculated by the DTW algorithm is 0.6. The similarity threshold S0 is set to 0.8. This cluster needs to be removed.

[0076] After eliminating all non-compliant clusters, the centroids of the remaining clusters are calculated. When locating a specific fault in the distribution network during a certain period of time, data within that period is collected in real time, and the characteristic vector within that period is calculated. The Euclidean distance di is then calculated between the characteristic vector and the centroid of each cluster in all faults. The Euclidean distance can measure the similarity between two vectors. The di are sorted to find the fault corresponding to the minimum di, i.e., min(di). This means the specific fault type that occurred in the distribution network during that period is obtained.

[0077] After finding the specific fault type, continue to calculate the similarity Si between the fault feature vector and the feature vector of each fault in the corresponding cluster. For example, using the cosine similarity algorithm, obtain min (Si). The fault location marked by fault i is used as the real-time location of the current fault, thereby achieving accurate positioning of the distribution network fault.

[0078] like Figure 2 ,A distribution network fault location system includes: a data acquisition module, a missing value processing module, an abnormal value processing module, and a fault location module;

[0079] Data acquisition module, which collects multiple sets of historical data of different faults in the distribution network and transmits them to the missing value processing module;

[0080] The missing value processing module obtains the missing intervals of the voltage and current series in each set of historical data, locates them at the front, middle, and back of the entire series, and combines the length of the missing interval, the distance from the starting point of the missing interval to the nearest missing interval, and the distance from the end point of the missing interval to the nearest missing interval to obtain different interpolation methods for different missing intervals. The processed series are then transmitted to the outlier processing module.

[0081] The outlier processing module marks the outliers in the interpolated sequence and divides the sequence into a front group S before and a back group S after of equal length. It then uses the extreme learning machine (ELM) bidirectional prediction to repair the outliers and transmits the repaired sequence to the fault location module.

[0082] The fault location module constructs discriminant features based on the processed voltage and current sequences, obtains a standard data set belonging to this type of fault, clusters the data in the standard data set, finds clusters containing light and heavy faults, eliminates irrelevant clusters based on the minimum time interval and time similarity, calculates the centroid of the remaining clusters, calculates the distance from the existing fault to the centroid of the remaining clusters of different faults to locate the target cluster, and then calculates the similarity between the fault and each group of faults in the target cluster to locate the specific location of the fault.

[0083] Some of the data in the above formulas are dimensionless and numerically calculated. Meanwhile, the contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0084] The above embodiments are only used to illustrate the technical method of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A distribution network fault location method, characterized in that: include: S1. Obtain the missing intervals of the voltage and current sequences in each set of historical data and locate them at the front, middle, and back of the entire sequence. Combined with the missing interval length L, the distance Dfront from the starting point of the missing interval to the nearest missing interval before it, and the distance Dback from the ending point of the missing interval to the nearest missing interval after it, different interpolation methods are obtained for different missing intervals. Outliers are marked on the interpolated sequences and the sequences are divided into a front group Sfront and a back group Sback of equal length. The extreme learning machine (ELM) bidirectional prediction is used to repair the outliers. S2. Construct discriminant features based on the processed voltage and current sequences and perform normalization processing to obtain standard data sets belonging to different types of faults. Cluster the data in the standard data sets to find clusters containing light and heavy faults. Eliminate irrelevant clusters based on the minimum time interval and time similarity, calculate the centroid of the remaining clusters, calculate the distance di from the existing fault to the centroid of the remaining clusters of different faults to locate the target cluster, and then calculate the similarity between the fault and each group of faults in the target cluster to locate the specific location.

2. The distribution network fault location method according to claim 1, characterized in that: Before S1, S0 is also included, and S0 is used to collect multiple sets of historical data of different faults in the distribution network. The specific historical data include voltage sequence and current sequence within the fault duration, as well as fault level, timestamp, and specific location.

3. The distribution network fault location method according to claim 1, characterized in that: The method for locating the specific position of the missing interval in the entire sequence is: Calculate the total length M of the sequence and define the starting index start and end of the missing interval. If the missing interval contains only one value, then start is equal to end. If end<=M*A, the missing interval is located at the front of the sequence; If start>M*A and end<=M*(A+B), the missing interval is located in the middle of the sequence; If start>M*(A+B), the missing interval is located at the end of the sequence, where A, B, C∈(0,1) and A+B+C=1.

4. The distribution network fault location method according to claim 3, characterized in that: If the missing interval is located at the front of the sequence, the interpolation method selected includes: If L<=Lth and max{Dfront, Dback}<=Dth, select the nearest neighbor interpolation method; If L<=Lth and max{Dfront, Dback}>Dth, select linear interpolation; If L>Lth and max{Dbefore, Dafter}<=Dth, select the interpolation method used for the missing interval closest to the current missing interval, that is, the interpolation method used for the missing interval corresponding to min{Dbefore, Dafter}; If L>Lth and max{Dfront, Dback}>Dth, select the linear extrapolation method; Where Dth is the distance threshold and Lth is the length threshold.

5. The distribution network fault location method according to claim 3, characterized in that: If the missing interval is located in the middle of the sequence, the interpolation method selected includes: If L<=Lth and max{Dfront, Dback}<=Dth, select the nearest neighbor interpolation method; If L<=Lth and max{Dbefore,Dafter}>Dth, choose the weighted moving average method; If L>Lth and max{Dbefore, Dafter}<=Dth, select the interpolation method used for the missing interval closest to the current missing interval, that is, the interpolation method used for the missing interval corresponding to min{Dbefore, Dafter}; If \(L > L_{th}\) and \(\max\{D_{front}, D_{back}\}>D_{th}\), select Lagrange interpolation or cubic spline interpolation.

6. The distribution network fault location method according to claim 3, characterized in that: When the missing interval is located at the rear of the sequence, the specific interpolation methods selected include: If \(L\leq L_{th}\) and \(\max\{D_{front}, D_{back}\}\leq D_{th}\), use the nearest neighbor interpolation method; If \(L\leq L_{th}\) and \(\max\{D_{front}, D_{back}\}>D_{th}\), use the linear interpolation method; If \(L > L_{th}\) and \(\max\{D_{front}, D_{back}\}\leq D_{th}\), select the interpolation method used for the missing interval with the closest distance to the current missing interval, that is, the interpolation method used for the missing interval corresponding to \(\min\{D_{front}, D_{back}\}\); If \(L > L_{th}\) and \(\max\{D_{front}, D_{back}\}>D_{th}\), the exponential smoothing method can be used.

7. The distribution network fault location method according to claim 1, characterized in that: The specific operation method for repairing outliers using the extreme learning machine ELM bidirectional prediction is as follows: Set the window size \(W\), calculate the local mean \(\mu\) and standard deviation \(\sigma\). If a point \(x_t\) satisfies \(|x_t - \mu|>k*\sigma\), then mark this point as a potential outlier; Divide the sequence into the front group \(S_{front}\) and the back group \(S_{back}\) with equal length. Use the interquartile range method to detect the outliers within the groups for \(S_{front}\) and \(S_{back}\) respectively, and remove the outliers in \(S_{front}\) and \(S_{back}\) to obtain the pure training sets \(S_{front\_pure}\) and \(S_{back\_pure}\); Train the ELM model with \(S_{front\_pure}\) and \(S_{back\_pure}\) respectively, and use the trained ELM models to predict \(S_{back}\) and \(S_{front}\) respectively to obtain the predicted value sequences \(S_{back\_prediction}\) and \(S_{front\_prediction}\); For each point \(y_t\in S_{back}\) in the back group and \(y_{t\_pred}\in S_{back\_prediction}\), calculate \(|y_t - y_{t\_pred}|\). If this point exceeds the specified threshold and this point is marked as a potential outlier, then replace \(y_t\) with \(y_{t\_pred}\). Similarly, perform the above operations on \(S_{front}\); Merge the repaired \(S_{front}\) and \(S_{back}\) into a complete sequence, and retain the unmarked normal points, where \(k\) is the proportionality coefficient.

8. The distribution network fault location method according to claim 1, characterized in that: Remove the clusters when \(t > T_{max}\) and \(S'<S_0\) and \(t < T_{min}\) and \(S'<S_0\) in each type of fault, where \(t\) is the minimum time interval between light and heavy faults, \(S'\) is the time similarity, and \(T_{min}\), \(T_{max}\), and \(S_0\) are the lower time threshold, upper time threshold, and time similarity threshold respectively.

9. The distribution network fault location method according to claim 1, characterized in that: After removing all the non - compliant clusters, calculate the centroid of the remaining clusters. When locating the specific fault of the distribution network at a certain time period, collect the data within this time period in real - time, obtain the feature vector within this time period, calculate the Euclidean distance \(d_i\) between it and the centroids of each cluster in all faults, select the fault corresponding to \(\min(d_i)\) as the specific fault that occurs in the distribution network during this time period, and calculate the cosine similarity \(S_i\) between the feature vector of this fault and the feature vectors of each historical fault segment in the target cluster, and obtain the fault position corresponding to \(\min(S_i)\) as the real - time position of the current fault.

10. A distribution network fault location system, configured to execute the distribution network fault location method according to any one of claims 1 to 9, characterized in that: Include: Data acquisition module, missing value processing module, outlier processing module, fault location module; The data acquisition module collects multiple groups of historical data on different faults of the distribution network and transmits them to the missing value processing module; The missing value processing module obtains the missing intervals of the voltage and current series in each set of historical data, locates them at the front, middle, and back of the entire series, and combines the length of the missing interval, the distance from the starting point of the missing interval to the nearest missing interval, and the distance from the end point of the missing interval to the nearest missing interval to obtain different interpolation methods for different missing intervals. The processed series are then transmitted to the outlier processing module. The outlier processing module marks the outliers in the interpolated sequence and divides the sequence into a front group S before and a back group S after of equal length. It then uses the extreme learning machine (ELM) bidirectional prediction to repair the outliers and transmits the repaired sequence to the fault location module. The fault location module constructs discriminant features based on the processed voltage and current sequences, obtains a standard data set belonging to this type of fault, clusters the data in the standard data set, finds clusters containing light and heavy faults, eliminates irrelevant clusters based on the minimum time interval and time similarity, calculates the centroid of the remaining clusters, calculates the distance from the existing fault to the centroid of the remaining clusters of different faults to locate the target cluster, and then calculates the similarity between the fault and each group of faults in the target cluster to locate the specific location of the fault.

Images