A data sample screening method and a screening device for judging power grid fault types

Abstract

The present invention discloses a data sample screening method and a screening device for power grid fault type judgment. The waveforms of actual faults are evaluated from aspects such as intuitive current and voltage values, the operation conditions of relay protection devices, and the fault judgment result types in neural network intelligent detection. The distance score D, the relay protection device feedback score E, the intelligent detection score F and their corresponding weights are calculated respectively, so as to calculate the confidence level of fault type judgment of fault recording signal data. The data samples are screened according to the confidence level of fault type judgment, making the data samples used for training the neural network more reliable. At the same time, it plays a role in expanding the waveform library, solving the technical problem in the prior art that there is a lack of effective technical means for screening the data samples for power grid fault type judgment, making it difficult to improve the richness and reliability of the data samples and affecting the accuracy of the neural network in judging the power grid fault type.

Description

Technical Field

[0001] The present invention relates to the technical field of sample screening, and in particular, to a method and a filter for screening data samples for judging power grid fault types. Background Art

[0002] In order to ensure the reliability and stability of the power system, predicting an upcoming power fault in advance and taking corresponding preventive measures can effectively prevent the occurrence of power accidents and reduce economic losses. With the development of artificial intelligence technology, the prediction of power grid fault types mainly uses power data samples to train a neural network model, and the trained neural network model is used to predict the power grid fault types. The accuracy of the neural network model for predicting power grid fault types is related to the richness and reliability of the sample in the data sample library. Therefore, the selection of data samples for judging power grid fault types is of great significance. In the prior art, the sample data for training the model is obtained by manually collecting the power waveform data generated when the power grid fails, and there is no reliability analysis and screening of the sample data, and there is easily a lack of fault sample data, which affects the accuracy of the neural network for judging power grid fault types. Summary of the Invention

[0003] The present invention provides a method and a filter for screening data samples for judging power grid fault types, which are used to solve the technical problem that the prior art lacks technical means for effectively screening data samples for judging power grid fault types, making it difficult to improve the richness and reliability of data samples and affecting the accuracy of the neural network for judging power grid fault types.

[0004] In view of this, the first aspect of the present invention provides a method for screening data samples for judging power grid fault types, including:

[0005] Collecting fault recording signal data, relay protection device action information, and the fault types marked by the corresponding neural network intelligent detection when the power grid fails;

[0006] Calculating the distance score D from the fault recording signal data to the fault types marked by the neural network intelligent detection according to the fault recording signal data;

[0007] Calculating the relay protection device feedback score E according to the relay protection device action information;

[0008] Obtaining the intelligent detection score F of the neural network intelligent detection of the fault type;

[0009] Calculating the weights of the distance score D, the relay protection device feedback score E, and the intelligent detection score F based on the analytic hierarchy process;

[0010] Calculate the confidence level of the fault type judgment for the fault recording signal data according to the distance score D, the feedback score E of the relay protection device, the intelligent detection score F, and the corresponding weights. The calculation formula is as follows:

[0011] σ = D·θ 1 + E·θ 2 + F·θ 3

[0012] where σ is the confidence level of the fault type judgment, and θ 1 is the weight of the distance score D, and θ 2 is the weight of the feedback score E of the relay protection device, and θ 3 is the weight of the intelligent detection score F;

[0013] If the confidence level of the fault type judgment σ is greater than the threshold σ 0 , then put the corresponding fault recording signal data into the data sample library for training the neural network model.

[0014] Optionally, calculate the feedback score E of the relay protection device according to the action information of the relay protection device, including:

[0015] Calculate the score E for evaluating the fault type based on the opening and closing conditions of the three-phase circuit breakers of each relay protection device sent back from the relay protection device end when the fault occurs according to the action information of the relay protection device 1 ;

[0016] Calculate the score E for evaluating the fault type based on the protection messages collected from the relay protection device monitoring device when the fault occurs according to the action information of the relay protection device 2 ;

[0017] Add the score E 1 and the score E 2 to obtain the feedback score E of the relay protection device.

[0018] Optionally, calculate the distance score D from the fault recording signal data to the fault type marked by the neural network intelligent detection according to the fault recording signal data, including:

[0019] Use the K-means clustering analysis method to cluster each feature quantity of the classified fault recording signal data to obtain the eigenvalue centroid of the waveform signal of each fault type;

[0020] Calculate the distance from the fault recording signal data when the power grid fails to the eigenvalue centroid of the fault type marked by the neural network intelligent detection to obtain the distance score D.

[0021] Optionally, the distance score D is the Euclidean distance score.

[0022] Optionally, calculating the weights of the distance score D, the relay protection device feedback score E, and the intelligent detection score F based on the analytic hierarchy process, including:

[0023] Define the target layer A, criterion layer C, and measure layer P of the analytic hierarchy process. The target layer A is whether the fault type detected by the neural network intelligence is credible. The criterion layer C is several factors in the actual power grid that affect the distance score D, the relay protection device feedback score E, and the intelligent detection score F. The measure layer P is the weights that the distance score D, the relay protection device feedback score E, and the intelligent detection score F should each account for in the confidence level calculation;

[0024] Evaluate the importance of several factors in the actual power grid that affect the distance score D, the relay protection device feedback score E, and the intelligent detection score F respectively, and establish a judgment matrix B;

[0025] Construct a judgment matrix C for the measure layer factors subordinate to several factors that affect the distance score D, the relay protection device feedback score E, and the intelligent detection score F l ;

[0026] Calculate the eigenvectors of the judgment matrix B and the judgment matrix C l ;

[0027] Calculate the weights of the distance score D, the relay protection device feedback score E, and the intelligent detection score F according to the eigenvectors of the judgment matrix B and the judgment matrix C l .

[0028] The second aspect of the present invention provides a data sample filter for power grid fault type judgment, including:

[0029] A data collection module for collecting fault recording signal data, relay protection device action information, and the corresponding fault types labeled by neural network intelligent detection when a power grid fails;

[0030] A first scoring module for calculating the distance score D from the fault recording signal data to the fault type labeled by neural network intelligent detection according to the fault recording signal data;

[0031] A second scoring module for calculating the relay protection device feedback score E according to the relay protection device action information;

[0032] A third scoring module for obtaining the intelligent detection score F of the neural network intelligent detection fault type;

[0033] A weight calculation module for calculating the weights of the distance score D, the relay protection device feedback score E, and the intelligent detection score F based on the analytic hierarchy process;

[0034] A confidence level calculation module, which is used to calculate the confidence level of the fault type judgment of the fault recording signal data according to the distance score D, the protection device feedback score E, the intelligent detection score F, and the corresponding weights. The calculation formula is:

[0035] σ = D·θ 1 + E·θ 2 + F·θ 3

[0036] where σ is the confidence level of the fault type judgment, and θ 1 is the weight of the distance score D, and θ 2 is the weight of the protection device feedback score E, and θ 3 is the weight of the intelligent detection score F;

[0037] A sample screening module, which is used to put the corresponding fault recording signal data into the data sample library for training the neural network model if the confidence level σ of the fault type judgment is greater than the threshold σ 0 .

[0038] Optionally, the second score module is specifically used for:

[0039] Calculating the score E for evaluating the fault type based on the opening and closing conditions of the three-phase circuit breakers of each protection device sent back from the protection device end when the fault occurs according to the protection device action information 1 ;

[0040] Calculating the score E for evaluating the fault type based on the protection messages collected from the protection device monitoring device when the fault occurs according to the protection device action information 2 ;

[0041] Adding the score E 1 and the score E 2 to obtain the protection device feedback score E.

[0042] Optionally, the first score module is specifically used for:

[0043] Using the K-means clustering analysis method to cluster the characteristic quantities of the classified fault recording signal data to obtain the eigenvalue centroids of the waveform signals of each fault type;

[0044] Calculating the distance from the fault recording signal data when the power grid fails to the eigenvalue centroid of the fault type marked by the neural network intelligent detection to obtain the distance score D.

[0045] Optionally, the distance score D is the Euclidean distance score.

[0046] Optionally, the weight calculation module is specifically used for:

[0047] Define the target layer A, criterion layer C, and measure layer P of the analytic hierarchy process. The target layer A is whether the fault type detected by the neural network is credible. The criterion layer C is several factors in the actual power grid that affect the distance score D, relay protection device feedback score E, and intelligent detection score F. The measure layer P is the weights that the distance score D, relay protection device feedback score E, and intelligent detection score F should each account for in the confidence level calculation.

[0048] Conduct importance assessments on several factors in the actual power grid that affect the distance score D, relay protection device feedback score E, and intelligent detection score F respectively, and establish a judgment matrix B.

[0049] Construct a judgment matrix C for the measure layer factors subordinate to several factors that affect the distance score D, relay protection device feedback score E, and intelligent detection score F l ;

[0050] Calculate the eigenvectors of the judgment matrix B and the judgment matrix C l ;

[0051] Calculate the weights of the distance score D, relay protection device feedback score E, and intelligent detection score F according to the eigenvectors of the judgment matrix B and the judgment matrix C l ;

[0052] As can be seen from the above technical solutions, the data sample screening method and filter provided by the present invention for power grid fault type judgment have the following advantages:

[0053] The data sample screening method and filter provided by the present invention for power grid fault type judgment evaluate the waveforms of actual faults from aspects such as intuitive current and voltage values, the operation of relay protection devices, and the fault judgment result types in neural network intelligent detection, calculate the distance score D, relay protection device feedback score E, and intelligent detection score F and their corresponding weights respectively, thereby calculating the confidence level of fault type judgment for fault recording signal data, screening data samples based on the confidence level of fault type judgment, making the data samples for training the neural network more reliable, connecting the relationship between the generation of internal data in the distribution network and the improvement of the accuracy of the neural network, and compared with the traditional fault removal and subsequent manual maintenance steps, it is equivalent to adding a filter that plays a feedback role, used to automatically batch mark the fault types and screen the recorded fault waveforms, and at the same time plays the role of expanding the waveform library, solving the technical problem that the prior art lacks effective technical means for screening data samples for power grid fault type judgment, making it difficult to improve the richness and reliability of data samples and affecting the accuracy of the neural network in judging power grid fault types. Description of the Drawings

[0054] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the accompanying drawings required for the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other related drawings can also be obtained based on these drawings.

[0055] Figure 1 It is a schematic flow chart of a data sample screening method for power grid fault type judgment provided by the present invention;

[0056] Figure 2 It is a logical framework diagram of a data sample screening method for power grid fault type judgment provided by the present invention;

[0057] Figure 3 It is a schematic structural diagram of a data sample screener for power grid fault type judgment provided by the present invention. Detailed implementation manners

[0058] In order to enable those skilled in the art to better understand the solution of the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0059] For ease of understanding, please refer to Figure 1 and Figure 2 , an embodiment of a data sample screening method for power grid fault type judgment provided by the present invention includes:

[0060] Step 101, collect the fault recording signal data, relay protection device action information, and the corresponding fault types labeled by neural network intelligent detection when the power grid fails.

[0061] It should be noted that when the power grid fails, collect the fault recording signal data (i.e., fault waveforms), the action conditions of the relay protection devices, and the fault types labeled by neural network intelligent detection. The fault recording signal data can be used as the fault waveform data sample to be analyzed after data processing. The neural network here is a pre-trained fault type judgment model. The fault recording signal data can be one or more waveform data of three-phase voltage, current, and zero-sequence voltage and current signals.

[0062] Step 102, calculate the distance score D from the fault recording signal data to the fault types labeled by neural network intelligent detection according to the fault recording signal data.

[0063] It should be noted that after obtaining the fault recording signal data, the corresponding fault type can be detected through a neural network. Therefore, the distance score D from the fault recording signal data to this fault type, that is, the intuitive numerical score, can be calculated by means of cluster analysis. Specifically, the K-means cluster analysis method is used to numerically cluster each characteristic quantity of one or more input waveform signals in the classified waveform sample library, and the centroid μ of each input waveform signal characteristic value of each fault type is obtained. k , calculate the Euclidean distance from the fault recording signal data to the centroid μ of the characteristic value k , that is, the distance score D.

[0064] When a certain input signal has m characteristic quantities, it is regarded as a clustering calculation in an m-dimensional space. For example, the effective value, peak value, mutation amount or the first derivative (mutation degree) of the waveform of the three-phase voltage, current, and zero-sequence voltage and current signals. Assume that there are two points of the waveform characteristic quantities of this type as (x 1 , y 1 , z 1 ) and (x 2 , y 2 , z 2 ). Then, the distance between the two points is:

[0065]

[0066] Set the objective function of the K-means cluster analysis as:

[0067]

[0068] Among them, r nk takes a value of 0 or 1. When the waveform sample vector n is classified into group k, r nk takes a value of 1, otherwise r nk takes a value of 0, x n is the coordinate vector of a certain waveform sample, μ k is the coordinate vector of the cluster center, that is, the centroid of the characteristic value, N is the number of all waveform samples used for clustering a certain type of characteristics, and K is the total number of classes existing in a certain cluster. The goal of the objective function J is to minimize the distance between the observed vector and the cluster center, that is, when the clustering is completed, the sum of the distances from each group of observed vectors to the cluster center is the smallest.

[0069] Use the iterative method to solve r nk and μ k to make the objective function J obtain the optimal solution, that is, obtain the minimum value. The centroid μ of the characteristic value of the optimal objective function can be obtained. k .

[0070] After a fault actually occurs, the recorded wave signal data collected from the power grid dispatching end is processed to obtain a fault waveform data sample, and the eigenvalue centroid μ of the fault waveform data sample corresponding to the fault type is calculated. k The Euclidean distance D to μ, that is, the distance score D:

[0071]

[0072] Among them, X i represents the coordinate of the i-th feature quantity of the actual input signal, and μ ik represents the coordinate of the centroid of the input signal of this fault type at the i-th feature quantity.

[0073] Step 103: Calculate the relay protection device feedback score E according to the relay protection device action information.

[0074] It should be noted that in the present invention, the relay protection device feedback score E is used to measure the possibility of occurrence of each fault type from the perspective of the relay protection device feedback. The relay protection device feedback score E can be obtained by a computer program reading the opening and closing conditions of the three-phase circuit breakers of each relay protection device sent back from the relay protection device end, and comparing the pre-conceived fault situation with the fault type result detected by the neural network intelligent detection to obtain the score E 1 , and the score E obtained by comparing the fault type determined from the protection message collected by the relay protection device monitoring device at the time of the fault with the fault type result detected by the neural network intelligent detection 2 is superimposed.

[0075] Specifically, considering that the relay protection device structure of the non-direct grounding system or small-resistance grounding system with a voltage level of 110 kV and below is simpler than that of the transmission line with a voltage level of 220 kV and above, there is mechanical linkage on the ABC three phases, and it is not easy to distinguish single-phase and multi-phase through the opening and closing of the switch. Therefore, when applied to the power grid at the level below 110 kV, the evaluation score E 1 is 0.

[0076] Since the selection phase instruction reaching two or more phases will trip all three phases, only five situations and their possible fault causes are considered: no-phase tripping, single-phase tripping (phase A, phase B, phase C), and three-phase tripping simultaneously. Since whether the reclosing of the switch is executed reflects the severity of the fault and has little significance for judging the fault type, it is not classified and discussed here.

[0077] The specific possible fault situations are allocated as follows, and they are numbered in sequence:

[0078] No-phase tripping: 01 - There is a fault on the line, but the fault is not caused by current change; 02 - There is no fault on the line, and the artificial intelligence gives a false alarm;

[0079] Single-phase tripping: There is a fault in the line, which may be 03 - single-phase grounding fault, 04 - phase-to-phase short circuit fault; 05 - there is no fault in the line, and the relay protection equipment malfunctions.

[0080] Three-phase tripping simultaneously: There is a fault in the line, which may be 06 - single-phase grounding fault, 07 - phase-to-phase short circuit fault, 08 - two-phase grounding fault, 09 - three-phase grounding fault, 10 - three-phase short circuit fault; 11 - there is no fault in the line, and the measuring device malfunctions or maloperates.

[0081] The following situations are all based on the premise that the artificial intelligence detection has results. First, detect the tripping situation of the relay protection equipment:

[0082] If there is no single-phase tripping, detect whether situations 01 and 02 occur.

[0083] If there is single-phase tripping, detect whether situations 03, 04, and 05 occur.

[0084] If there is three-phase tripping, detect whether situations 06, 07, 08, 09, 10, and 11 occur.

[0085] The situation number consistent with the artificial intelligence detection result indicates that through the opening situation of the circuit breaker of the relay protection device, it can be seen that it is the same as or there is the possibility of the artificial intelligence detection result, and the score E is counted. 1 It is:

[0086]

[0087] Otherwise (that is, it is different from the artificial intelligence detection result), then deduct the score with the same value as the score.

[0088] When calculating the score E using the logic states of the switches of each device 1 After that, further calculate the score E using the protection messages collected by the relay protection equipment monitoring device when a fault occurs. 2 .

[0089] Taking the distribution network as an example, the current mainstream relay protection methods are overcurrent three-stage protection and zero-sequence protection. When the equipment malfunctions or maloperates due to a fault, the action reason will be displayed on its control panel, such as delayed tripping or exceeding the set amplitude. If it is a zero-sequence protection warning, it means there is a grounding fault. If it is the same as or there is the possibility of the artificial intelligence detection result, record its score (E 2 ), otherwise deduct the score with the same value as the score. (The scoring rule is the same as E 1The same. The possible situations are single-phase grounding, two-phase grounding, and three-phase grounding. If the overcurrent stage-three protection gives a warning, it indicates a short-circuit fault. The scoring rule is the same as that for the zero-sequence protection warning. The possible situations are generally phase-to-phase short circuit, three-phase short circuit, two-phase grounding, and three-phase grounding. In a specific grid structure, depending on the magnitude relationship between the starting current and the fault current in the setting calculation, the fault situations that may occur can be appropriately adjusted.

[0090] The calculation formula for the feedback score E of the relay protection device is:

[0091] E = E 1 + E 2

[0092] Step 104: Obtain the intelligent detection score F of the neural network for intelligent detection of the fault type.

[0093] It should be noted that the neural network intelligent detection of the fault type, as a judgment of the assumed correct fault type, occupies a certain evaluation weight by itself. The direct score is a constant, that is, the score F is a preset constant term.

[0094] Step 105: Calculate the weights of the distance score D, the relay protection device feedback score E, and the intelligent detection score F based on the analytic hierarchy process.

[0095] It should be noted that after obtaining the three score situations of the intuitive value (i.e., the distance score D), the relay protection device feedback (i.e., the score E), and the neural network intelligent detection (i.e., the score F), according to the analytic hierarchy process (AHP), evaluate their influence weights in the entire fault judgment confidence level, which are θ 1 、θ 2 、θ 3 .

[0096] Specifically, in the present invention, the target layer A of the analytic hierarchy process is "whether the fault type detected by the neural network is credible", the criterion layer C is "several factors n in the actual power grid that affect the distance score D, the relay protection device feedback score E, and the intelligent detection score F", and the measure layer P is "the weights that the distance score D, the relay protection device feedback score E, and the intelligent detection score F should respectively occupy in the confidence level calculation". According to the actual situation of the power grid, the importance of the n criterion layer factors that affect the distance score D, the relay protection device feedback score E, and the intelligent detection score F is respectively evaluated, and their importance is respectively expressed as ω 1 ,ω 2 ,……,ω n . Then, using the pairwise comparison method, a judgment matrix B is established:

[0097]

[0098] For the three measure-level factors subordinate to the \(l\)-th influencing factor among these \(n\) criterion-level factors, the pairwise comparison method is used to construct the judgment matrix \(C\). l :

[0099]

[0100] Among them, \(\theta\) 1 is the weight of the distance score \(D\), and \(\theta\) 2 is the weight of the relay protection equipment feedback score \(E\), and \(\theta\) 3 is the weight of the intelligent detection score \(F\).

[0101] Then, the root method is used to approximately calculate the eigenvectors \(w\) and \(w\) l of the judgment matrices \(B\) and \(C\). Taking the calculation of the eigenvector \(w\) of the judgment matrix \(B\) (the corresponding weights of the \(C\) layer for the \(A\) layer purpose, also known as the relative weights of the \(C\) layer when \(A - C\) layer) as an example, the elements in the matrix \(B\) are named \(a\) l , \(i = 1, 2, \cdots, n\), \(j = 1, 2, \cdots, n\). ij Calculate the geometric mean of the elements in each row of the judgment matrix \(B\)

[0102] to obtain the vector \(u=(u,\cdots,u\) ) n . T .

[0103] Calculate to obtain the vector as the approximate value of the required eigenvector \(w\).

[0104] Use the principles of linear algebra to find the maximum eigenvalue of the matrix \(B\), and calculate Among them, \((B\omega)\) i is the value of the \(i\)-th element of \(B\omega\).

[0105] According to the judgment matrix consistency index, judge whether it meets the consistency. If it does, then otherwise recalculate.

[0106] When \(A - C\) layer, the calculated relative weights of the \(C\) layer are:

[0107] w=(w 1 ,\cdots,w n ) T

[0108] When \(C - P\) layer (the corresponding weights of the \(P\) layer for the \(C\) layer purpose, taking the \(C\) corresponding to the \(l\)-th criterion-level influencing factor as an example) l , the expressions for calculating the three relative weights of the \(P\) layer are:

[0109] w l =(w 1l ,w2l , w 3l ) T , l = 1, …, n

[0110] According to w and w l , calculate the relative weight expressions for all aspects of the A-P layer as follows:

[0111]

[0112] The relative weights of all aspects of the P layer calculated when obtaining the A-P layer are:

[0113] θ = (θ 1 , θ 2 , θ 3 ) T .

[0114] Step 106: Calculate the confidence level for fault type judgment of the fault recording signal data based on the distance score D, the feedback score E of the protection device, the intelligent detection score F, and the corresponding weights.

[0115] The calculation formula for the confidence level for fault type judgment of the fault recording signal data is:

[0116] σ = D·θ 1 + E·θ 2 + F·θ 3 = D·θ 1 +(E 1 + E 2 )·θ 2 + F·θ 3

[0117] where σ is the confidence level for fault type judgment, θ 1 is the weight of the distance score D, θ 2 is the weight of the feedback score E of the protection device, and θ 3 is the weight of the intelligent detection score F.

[0118] Step 107: If the confidence level for fault type judgment σ is greater than the threshold σ 0 , then put the corresponding fault recording signal data into the data sample library for training the neural network model.

[0119] It should be noted that if σ ≥ σ 0 (set threshold), then label the same fault type as the neural network detection for this data sample and output the supplementary training set for neural network training. If σ < σ 0 (set threshold), then temporarily store this sample in the storage area that needs to be manually checked, and then decide whether to use or discard it according to the fault type determined by manual inspection.

[0120] The data sample filter for power grid fault type judgment provided by the present invention evaluates the waveforms of actual faults from aspects such as intuitive current and voltage values, the operation of relay protection devices, and the types of fault judgment results in neural network intelligent detection, calculates the distance score D, the relay protection device feedback score E, the intelligent detection score F and their corresponding weights respectively, so as to calculate the confidence level of fault type judgment of fault recording signal data, and filters data samples according to the confidence level of fault type judgment, making the data samples for training the neural network more reliable, connecting the relationship between the generation of internal data in the distribution network and the improvement of the accuracy of the neural network, and compared with the traditional fault removal and subsequent manual maintenance steps, it is equivalent to adding a filter that plays a feedback role, which is used to automatically batch mark the fault types and screen and record the fault waveforms, and at the same time plays the role of expanding the waveform library, solving the technical problem that the prior art lacks effective technical means for screening data samples for power grid fault type judgment, making it difficult to improve the richness and reliability of data samples and affecting the accuracy of the neural network in judging power grid fault types.

[0121] For ease of understanding, please refer to Figure 3 , a data sample filter for power grid fault type judgment is provided in the present invention, including:

[0122] A data collection module for collecting fault recording signal data, relay protection device action information and the corresponding fault types marked by neural network intelligent detection when a fault occurs in the power grid;

[0123] A first scoring module for calculating the distance score D from the fault recording signal data to the fault type marked by neural network intelligent detection according to the fault recording signal data;

[0124] A second scoring module for calculating the relay protection device feedback score E according to the relay protection device action information;

[0125] A third scoring module for obtaining the intelligent detection score F of the neural network intelligent detection fault type;

[0126] A weight calculation module for calculating the weights of the distance score D, the relay protection device feedback score E and the intelligent detection score F based on the analytic hierarchy process;

[0127] A confidence level calculation module for calculating the confidence level of fault type judgment of the fault recording signal data according to the distance score D, the relay protection device feedback score E, the intelligent detection score F, and their corresponding weights. The calculation formula is:

[0128] σ = D·θ 1 +E·θ 2 +F·θ 3

[0129] Among them, σ is the confidence level for fault type judgment, and θ 1 is the weight of the distance score D, and θ 2 is the weight of the relay protection equipment feedback score E, and θ 3 is the weight of the intelligent detection score F;

[0130] The sample screening module is used to, if the confidence level σ for fault type judgment is greater than the threshold σ 0 , then put the corresponding fault recorder signal data into the data sample library for training the neural network model.

[0131] The second scoring module is specifically used for:

[0132] Calculating the score E for fault type evaluation based on the three-phase breaker opening conditions of each relay protection equipment sent back from the relay protection equipment end when the fault occurs according to the relay protection equipment action information 1 ;

[0133] Calculating the score E for fault type evaluation based on the protection messages collected from the relay protection equipment monitoring device when the fault occurs according to the relay protection equipment action information 2 ;

[0134] Adding the score E 1 and the score E 2 to obtain the relay protection equipment feedback score E.

[0135] The first scoring module is specifically used for:

[0136] Using the K-means clustering analysis method to cluster each feature quantity of the classified fault recorder signal data to obtain the eigenvalue centroids of the waveform signals of each fault type;

[0137] Calculating the distance from the fault recorder signal data when the power grid fails to the eigenvalue centroids of the fault types marked by the neural network intelligent detection to obtain the distance score D.

[0138] The distance score D is the Euclidean distance score.

[0139] The weight calculation module is specifically used for:

[0140] Defining the target layer A, criterion layer C, and measure layer P of the analytic hierarchy process. The target layer A is whether the fault types detected by the neural network intelligent detection are credible. The criterion layer C is several factors in the actual power grid that affect the distance score D, relay protection equipment feedback score E, and intelligent detection score F. The measure layer P is the weights that the distance score D, relay protection equipment feedback score E, and intelligent detection score F should each account for in the confidence level calculation;

[0141] Evaluating the importance of several factors in the actual power grid that affect the distance score D, relay protection equipment feedback score E, and intelligent detection score F respectively, and establishing the judgment matrix B;

[0142] Construct a judgment matrix C for the measure layer factors that are subordinate to several factors of the influence distance score D, the relay protection device feedback score E, and the intelligent detection score F. l ;

[0143] Calculate the eigenvectors of the judgment matrix B and the judgment matrix C. l ;

[0144] Calculate the weights of the distance score D, the relay protection device feedback score E, and the intelligent detection score F according to the eigenvectors of the judgment matrix B and the judgment matrix C. l The data sample filter for power grid fault type judgment provided by the present invention evaluates the waveforms of actual faults from aspects such as the intuitive current and voltage values, the operation conditions of relay protection devices, and the types of fault judgment results in neural network intelligent detection, calculates the distance score D, the relay protection device feedback score E, and the intelligent detection score F and their corresponding weights respectively, so as to calculate the confidence level of fault type judgment of the fault recording signal data, and filters the data samples according to the confidence level of fault type judgment, making the data samples used for training the neural network more reliable, connecting the relationship between the generation of internal data in the distribution network and the improvement of the accuracy of the neural network. And compared with the traditional fault removal and subsequent manual maintenance steps, it is equivalent to adding a filter that plays a feedback role, which is used to automatically batch mark the fault types and filter and record the fault waveforms, and at the same time plays the role of expanding the waveform library, solving the technical problem that the prior art lacks effective technical means for screening the data samples used for power grid fault type judgment, making it difficult to improve the richness and reliability of the data samples and affecting the accuracy of the neural network in judging the power grid fault type.

[0145] The data sample filter for power grid fault type judgment provided by the present invention has the same working principle as the data sample screening method for power grid fault type judgment in the embodiment of the data sample screening method for power grid fault type judgment described above, and will not be elaborated here.

[0146] As mentioned above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0147] ​

Claims

1. A method for screening data samples for power grid fault type judgment, characterized in that, it includes: Collect fault recording signal data, relay protection device action information, and corresponding fault types labeled by neural network intelligent detection when the power grid fails; Calculate the distance score D from the fault recording signal data to the fault type labeled by neural network intelligent detection according to the fault recording signal data; Calculate the relay protection device feedback score E according to the relay protection device action information; Obtain the intelligent detection score F of the neural network intelligent detection fault type; Calculate the weights of the distance score D, the relay protection device feedback score E, and the intelligent detection score F based on the analytic hierarchy process; Calculate the confidence level of the fault type judgment of the fault recording signal data according to the distance score D, the relay protection device feedback score E, the intelligent detection score F, and the corresponding weights. The calculation formula is: σ = D·θ 1 + E·θ 2 + F·θ 3 Among them, σ is the confidence level for fault type judgment, and θ 1 is the weight of the distance score D, and θ 2 is the weight of the relay protection device feedback score E, and θ 3 is the weight of the intelligent detection score F; If the confidence level σ of the fault type judgment is greater than the threshold value σ 0 , the corresponding fault recording signal data will be put into the data sample library for training the neural network model; Calculate the relay protection device feedback score E according to the relay protection device action information, including: Calculate the score E for fault type assessment based on the opening conditions of the three-phase circuit breakers of each relay protection device sent back from the relay protection device end when the fault occurs according to the action information of the relay protection device 1 ; Calculate the score E of the fault type assessment based on the protection messages collected from the relay protection device monitoring device when the fault occurs according to the action information of the relay protection device 2 ; Add the score E 1 and the score E 2 to obtain the feedback score E of the protection device.

2. The method for screening data samples for power grid fault type judgment according to claim 1, characterized in that, Calculating the distance score D from the fault recording signal data to the fault type labeled by neural network intelligent detection according to the fault recording signal data includes: Using the K-means clustering analysis method to cluster each feature quantity of the classified fault recording signal data to obtain the eigenvalue centroid of the waveform signal of each fault type; Calculate the distance from the fault recording signal data when the power grid fails to the eigenvalue centroid of the fault type labeled by neural network intelligent detection to obtain the distance score D.

3. The method for screening data samples for power grid fault type judgment according to claim 1 or 2, characterized in that, The distance score D is the Euclidean distance score.

4. The method for screening data samples for power grid fault type judgment according to claim 1, characterized in that, Calculating the weights of the distance score D, the relay protection device feedback score E, and the intelligent detection score F based on the analytic hierarchy process includes: Define the target layer A, criterion layer C, and measure layer P of the analytic hierarchy process. The target layer A is whether the fault type detected by the neural network is credible. The criterion layer C is several factors affecting the distance score D, the relay protection device feedback score E, and the intelligent detection score F in the actual power grid. The measure layer P is the weights that the distance score D, the relay protection device feedback score E, and the intelligent detection score F should occupy respectively in the confidence level calculation; Conduct importance evaluations on several factors affecting the distance score D, the relay protection device feedback score E, and the intelligent detection score F in the actual power grid respectively, and establish a judgment matrix B; Construct a judgment matrix C for the measure layer factors that are subordinate to several factors affecting the distance score D, the feedback score E of the relay protection device, and the intelligent detection score F l ; Calculate the eigenvectors of judgment matrix B and judgment matrix C l ; According to the eigenvectors of judgment matrix B and judgment matrix C l calculate the weights of the distance score D, the relay protection device feedback score E, and the intelligent detection score F.

5. A data sample screening device for power grid fault type judgment, characterized in that, it includes: A data collection module for collecting fault recording signal data, relay protection device action information, and corresponding fault types labeled by neural network intelligent detection when the power grid fails; A first scoring module for calculating the distance score D from the fault recording signal data to the fault type labeled by neural network intelligent detection according to the fault recording signal data; A second scoring module for calculating the relay protection device feedback score E according to the relay protection device action information; A third scoring module for obtaining the intelligent detection score F of the neural network intelligent detection fault type; A weight calculation module, which is used to calculate the weights of the distance score D, the relay protection device feedback score E, and the intelligent detection score F based on the analytic hierarchy process; A confidence level calculation module, which is used to calculate the confidence level of the fault type judgment of the fault recording signal data according to the distance score D, the relay protection device feedback score E, the intelligent detection score F, and the corresponding weights. The calculation formula is: σ = D·θ 1 + E·θ 2 + F·θ 3 Among them, σ is the confidence level for fault type judgment, θ 1 is the weight of the distance score D, θ 2 is the weight of the relay protection device feedback score E, θ 3 is the weight of the intelligent detection score F; A sample screening module, which is used to put the corresponding fault recording signal data into the data sample library for training the neural network model if the confidence level σ of the fault type judgment is greater than the threshold value σ 0 , The second scoring module is specifically used for: Calculate the score E for fault type evaluation based on the opening conditions of the three-phase circuit breakers of each relay protection device sent back from the relay protection device end when the fault occurs according to the action information of the relay protection device 1 ; Calculate the score E for fault type assessment based on the protection messages collected from the protection and control equipment monitoring device when the fault occurs according to the action information of the protection and control equipment 2 ; Add the score E 1 and the score E 2 to obtain the feedback score E of the protection device.

6. The data sample filter for power grid fault type judgment according to claim 5, characterized in that The first scoring module is specifically used for: Using the K-means clustering analysis method to cluster each feature quantity of the classified fault recording signal data to obtain the eigenvalue centroid of the waveform signal of each fault type; Calculating the distance from the fault recording signal data when the power grid fails to the eigenvalue centroid of the fault type marked by the neural network intelligent detection to obtain the distance score D.

7. The data sample filter for power grid fault type judgment according to claim 5 or 6, characterized in that The distance score D is the Euclidean distance score.

8. The data sample filter for power grid fault type judgment according to claim 1, characterized in that The weight calculation module is specifically used for: Defining the target layer A, the criterion layer C, and the measure layer P of the analytic hierarchy process. The target layer A is whether the fault type detected by the neural network intelligent detection is credible. The criterion layer C is several factors in the actual power grid that affect the distance score D, the relay protection device feedback score E, and the intelligent detection score F. The measure layer P is the weights that the distance score D, the relay protection device feedback score E, and the intelligent detection score F should account for in the confidence level calculation respectively; Conducting importance evaluations on several factors in the actual power grid that affect the distance score D, the relay protection device feedback score E, and the intelligent detection score F respectively, and establishing a judgment matrix B; Construct a judgment matrix C for the measure layer factors that are subordinate to several factors affecting the distance score D, the feedback score E of the relay protection device, and the intelligent detection score F l ; Calculate the eigenvectors of judgment matrix B and judgment matrix C l ; According to the eigenvectors of judgment matrix B and judgment matrix C l calculate the weights of the distance score D, the relay protection device feedback score E, and the intelligent detection score F.

Images