Distribution network latent fault type identification method and system

By obtaining ground fault data in the distribution network, calculating latent fault characteristics and building ELM and SVM models, combined with the D-S evidence fusion theory, the problem of latent fault type identification in the distribution network is solved, and detection accuracy and maintenance are improved Efficiency enables effective identification of latent faults.

CN119089256APending Publication Date: 2024-12-06ELECTRIC POWER RES INST OF GUANGXI POWER GRID CO LTD
View PDF 0 Cites 6 Cited by

Patent Information

Application Number
CN202410966574.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-18
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

The existing distribution network latent fault type identification methods are difficult to effectively identify latent faults in the distribution network, especially because the fault current is small, the fault traces are difficult to capture, and the fault mechanism is complex, making it difficult to distinguish different types of latent faults.

Method used

By obtaining the distribution network grounding fault data, using signal processing technology to calculate latent fault characteristics, using multiple variance analysis to verify the characteristics, constructing the ultimate learning machine (ELM) and support vector machine (SVM) models, combining the D-S evidence fusion theory to identify the results Fusion is performed to extract effective multi-domain features to identify latent fault types.

Benefits of technology

It improves the accuracy and maintenance efficiency of latent fault type detection, realizes the feature extraction and verification of different latent faults in different time, frequency and time-frequency domains, and improves the recognition effect.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119089256A_ABST
    Figure CN119089256A_ABST
Patent Text Reader

Abstract

The invention discloses a distribution network latent fault type identification method and system, and relates to the technical field of distribution network protection, and the method comprises the steps: obtaining grounding fault data in a distribution network, and calculating latent fault features; verifying the calculated fault features by using multivariate variance analysis, and extracting effective multi-domain features of the latent fault; based on the screened effective multi-domain features, constructing a latent fault type recognition model of an extreme learning machine ELM and a latent fault type classification model of a support vector machine SVM, and outputting recognition and classification results; and fusing identification and classification results to obtain a distribution network latent fault type identification result. According to the method, the latent fault type identification model based on the extreme learning machine and the support vector machine is established based on the extracted fault features, the identification results of the two models are fused, the comprehensive identification model is superior to a single model, the proposed comprehensive model can improve the latent fault type detection accuracy and maintenance efficiency, and the method is suitable for popularization and application. The method has a good identification effect.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

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

[0002] The distribution network is an infrastructure that provides consumers with reliable and high-quality electricity. The distribution network is large in size and has many points. Operational safety and equipment status are crucial. Failure of any component in the distribution system may cause power supply disruption or power outage. According to statistics, most faults in the distribution network occur on the distribution line, of which single-phase grounding faults account for about 70% of all faults in the distribution line.

[0003] In order to ensure power supply reliability, most distribution systems use a neutral point non-direct grounding system. A permanent single-phase grounding fault in this system will cause the non-fault phase voltage to rise to the original Times, long-term operation with faults can easily lead to more serious faults such as two-phase ground short circuit. These small current ground faults are often induced by insulation degradation. The insulation degradation of equipment is a gradual process. Before a permanent ground fault occurs, the distribution equipment with insulation degradation often shows a transient ground fault, which is called a latent fault. The latent faults of the distribution network are latent and periodic. The fault current is very small, and the fault traces are difficult to capture. The occurrence mechanism is complex. At present, it is difficult to effectively identify different types of latent faults. Summary of the invention

[0004] In view of the problems existing in the existing distribution network latent fault type identification method, the present invention is proposed. Therefore, the problem to be solved by the present invention is how to provide a distribution network latent fault type identification method and system.

[0005] In order to solve the above technical problems, the present invention provides a technical solution:

[0006] In a first aspect, the present invention provides a method for identifying latent fault types in a distribution network, which includes obtaining grounding fault data in the distribution network, and calculating latent fault characteristics using signal processing technology; verifying the calculated fault characteristics using multivariate variance analysis, and extracting effective multi-domain features of the latent faults; based on the screened effective multi-domain features, constructing a latent fault type recognition model of an extreme learning machine ELM and a latent fault type classification model of a support vector machine SVM, and outputting recognition and classification results; fusing the recognition and classification results to obtain a latent fault type identification result for the distribution network.

[0007] As a preferred solution of the method for identifying potential fault types of distribution networks described in the present invention, the fault characteristics include the ratio of maximum zero-sequence current to maximum current change, the degree of zero-sequence current distortion, the degree of randomness, the information dimension and the time-frequency characteristics based on wavelet transform.

[0008] As a preferred solution of the method for identifying the type of latent fault in the distribution network of the present invention, the calculation formula of the ratio of the maximum zero-sequence current to the maximum current change is:

[0009] ΔI 0max =max(I n+1 -I n )

[0010]

[0011] Among them, I 0ratio is the ratio of the maximum zero-sequence current to the maximum current change, I 0max is the maximum zero-sequence current, ΔI 0max is the maximum change in current during the initial transient phase, I n and I n+1 They represent the samples sampled at the nth and n+1th times after the fault occurs respectively; the calculation formula of the zero-sequence current distortion degree is:

[0012]

[0013] Wherein, THD is the waveform distortion of zero-sequence current, h represents the harmonic number, I(h) represents the root mean square value of the h-th harmonic current, M represents the highest harmonic number, and the logarithm of THD, i.e., lg(THD), is used to represent the degree of zero-sequence current distortion; the degree of randomness includes using the randomness of waveform distortion to describe the steady-state characteristics of faults caused by different types of latent faults, and is defined as the standard deviation of the current waveform distortion, and the calculation formula is:

[0014]

[0015] Where k represents the kth cycle of the zero-sequence current in the steady state, N represents the number of cycles in the steady state, μ represents the average value of the zero-sequence current distortion degree lg (THD) of N cycles of the zero-sequence current waveform, and THD k represents the total harmonic distortion of the steady-state zero-sequence current at the kth cycle; the information dimension includes constructing an m-dimensional vector according to the original one-dimensional time series data of the zero-sequence current through phase space reconstruction, for I(t)=[i t1 ,i t1 ,…,i tN ], when m = 2, the two-dimensional vector representation for the time delay τ is obtained as:

[0016]

[0017] Where I(t) represents the original one-dimensional time series data of zero-sequence current, i t1 -i tN Represents the zero-sequence current at different times, N represents the number of time series, N m Represents the number of m-dimensional vector time series; obtain the phase space trajectory of zero-sequence current, and calculate the information dimension using a two-dimensional graph through the phase space of zero-sequence current. The calculation formula is:

[0018]

[0019]

[0020] In the formula, D is the information dimension, r represents the square root of the total number of divided boxes, T r is information entropy, p i is the probability of the phase space trajectory in the ith frame.

[0021] As a preferred solution of the method for identifying the type of latent fault in the distribution network described in the present invention, the time-frequency characteristics based on wavelet transform include obtaining the zero-sequence current waveform of the system fault, determining the time of occurrence of the fault and the end time of the transient process of the fault; selecting the wavelet basis function and the number of decomposition layers based on the characteristics of the signal, performing wavelet transform on the zero-sequence current waveform within the cycle, and obtaining coefficients at different scales; calculating the energy of the wavelet transform signal of each layer and the zero-sequence current energy characteristics before and after the fault, further analyzing the time-frequency characteristics of the signal according to the calculated energy characteristics, integrating the extracted time-frequency characteristics into a feature vector, and calculating the fault characteristics extracted according to the wavelet transform as follows:

[0022]

[0023] Where f(t) represents the zero-sequence current waveform of the system fault; I m represents the zero-sequence current peak value of the fault; t0 represents the initial time of f(t), t total Indicates the end time of the selected waveform, t s Indicates the time when the fault occurs, t w Indicates the end time of the fault transient process, t mid represents the middle time of f(t), t k Indicates t mid and t total The selected time between m represents the peak time of f(t), t h Indicates t m After that, the first half peak value of the zero-sequence current, n represents the number of wavelet transform layers, E i Indicates that in [t s , tw ], E1 represents the energy of the i-th wavelet transform signal of the zero-sequence current in the period of [t s , t w ] is the energy of the first-layer wavelet transform signal of the zero-sequence current in the period of j Indicates that in [t0, t k ], e1 represents the energy of the j-th wavelet transform signal of the zero-sequence current in the period [t0, t k ] is the energy of the first-layer wavelet transform signal of the zero-sequence current in the period of ]; U0 represents the line voltage before the fault, M s represents the maximum slope of f(t), M1 represents the slope from t0 to t w The energy of the zero-sequence current during the fault, M total From t0 to t total The energy of the zero-sequence current during the fault, M mid From t0 to t mid The energy of the zero-sequence current during the fault, M k From t0 to t k The energy of the zero-sequence current of the fault during the process.

[0024] As a preferred solution of the method for identifying the potential fault type of the distribution network of the present invention, the construction of the potential fault type identification model of the extreme learning machine ELM includes the following steps: randomly selecting weights ω and bias B from the input of the hidden layer, and constructing a single hidden layer neural network with L hidden neurons and an activation function, the expression is:

[0025]

[0026] In the formula, ω k and β k Represents the weight vectors from the input layer to the hidden layer and from the hidden layer to the output layer respectively; b k represents the bias of the kth hidden neuron, g(<ω k ,x i >+b k ) represents relative to the input sample x i The output of the kth hidden neuron; calculate the hidden layer output matrix H, the calculation formula is:

[0027]

[0028]

[0029]

[0030] H·β=Y

[0031] In the formula, H represents the hidden layer output matrix of the single hidden layer feedforward neural network, and β represents the output weight matrix. The output weight β from the hidden layer to the output is calculated according to the minimum norm and least squares function, and the weight matrix is ​​expressed as:

[0032] β=H + ·Y

[0033] In the formula, H + It represents the Moore-Penrose generalized inverse of the matrix H, and the particle swarm optimization algorithm is used to optimize the initial weights and biases.

[0034] As a preferred solution of the method for identifying the type of latent fault in the distribution network of the present invention, the steps of constructing the latent fault type classification model of the support vector machine SVM are as follows: normalizing the input vector containing the latent fault characteristics; selecting the radial basis function as the kernel function, for the training set {(x1, y1),…, (x i ,y i ),…,(x N ,y N )},x i ∈R D ,y i ∈{-1,1}, mapping the input vector x i , through the kernel function From the original space R D Transfer to high-dimensional feature space, in high-dimensional space, use the principle of structural risk minimization to construct the optimal decision function, the expression is:

[0035] f(x)=sgn(ω·x+b)

[0036] In the formula, ω represents the weight vector and b is a constant. The core parameters c and g of the model are selected, where c represents the penalty factor, which indicates the importance of abnormal data in the class, and g is the parameter coefficient of the kernel function. The optimal decision function is transformed into an optimization problem, which is expressed as follows:

[0037]

[0038] Among them, c is the fault tolerance penalty coefficient, ξ i is the relaxation factor, and N represents the total number of samples. The optimal linear decision function is solved by the Lagrangian method and the Kuhn-Tucker theorem. The number of cross-validation times of the model is set to obtain the optimal core parameters, and the multi-classification model is trained using the obtained optimal core parameters, and the classification accuracy is tested.

[0039] As a preferred solution of the method for identifying the type of latent fault in the distribution network described in the present invention, the fusion of the identification and classification results includes the following steps: using the DS evidence fusion theory to fuse the two model identification results, and establish a set O of all results of the identification framework decision problem; selecting evidence and constructing a basic probability distribution, using the diagnosis results of ELM and SVM as two independent evidences; using the quality function to convert the model identification results into a BPA that satisfies the DS evidence theory, and constructing the evidence quality function of the latent fault type and uncertainty θ, the expression is:

[0040]

[0041] m i (θ)=1-α i

[0042] Among them, α i is the reliability coefficient of the i-th piece of evidence; m i (F j ) represents the proposition F j The BPA of the ith piece of evidence; m i (θ) represents the BPA of the ith piece of evidence with uncertainty; Y ij represents the diagnostic output probability of the i-th evidence of the j-th fault type, and N represents the number of training sample sets. Combining the recognition results of the extreme learning machine ELM and the support vector machine SVM, the BPA is calculated based on the Dempster combination rule, and the calculation formula is:

[0043]

[0044]

[0045] Among them, A, B and C are three subsets of the recognition frame O, and K is a normalized constant. The recognition diagnosis rules are constructed to perform latent fault decision diagnosis to determine the latent fault type. The rules are expressed as:

[0046] Rule 1:

[0047] Rule 2: m(F type )>m(θ)

[0048] Among them, F type represents the fault type finally determined by combining the identification results of ELM and SVM; F l Indicates the fault type when the BPA of the proposition evidence is maximum.

[0049] In a second aspect, the present invention provides a distribution network latent fault type identification system, which includes: an acquisition module, used to obtain grounding fault data in the distribution network, and calculate the latent fault characteristics using signal processing technology; a verification module, used to verify the calculated fault characteristics using multivariate variance analysis, and extract effective multi-domain features of the latent fault; a construction module, used to construct a latent fault type identification model of an extreme learning machine ELM and a latent fault type classification model of a support vector machine SVM based on the screened effective multi-domain features, and output the identification and classification results; an identification module, used to fuse the identification and classification results to obtain the distribution network latent fault type identification results.

[0050] In a third aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein: when the processor executes the computer program, the steps of a method for identifying latent fault types of a distribution network are implemented.

[0051] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, the steps of a method for identifying potential fault types of a distribution network are implemented.

[0052] The beneficial effect of the present invention is to extract various characteristics of different latent faults in different time, frequency and time-frequency domains through the acquired ground fault data, and verify the fault characteristics using multivariate variance analysis. Based on the extracted fault characteristics, a latent fault type recognition model based on extreme learning machine and support vector machine is established. The recognition results of the two models are fused using DS evidence fusion theory, and the comprehensive recognition model based on DS evidence theory is superior to each single model. The proposed comprehensive model based on DS evidence theory can improve the detection accuracy and maintenance efficiency of latent fault types, and has a good recognition effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0054] Figure 1 The figure is a flow chart of the method for identifying potential fault types in distribution networks.

[0055] Figure 2 It is a schematic diagram of zero-sequence current waveform of a system fault of the present invention;

[0056] Figure 3 It is a schematic diagram of the change of fitness relative to the number of iterations in the PSO of the present invention. DETAILED DESCRIPTION

[0057] In order to make the above-mentioned purposes, features and advantages of the present invention more understandable, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0058] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0059] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0060] Example 1

[0061] Reference Figure 1 , which is the first embodiment of the present invention, and provides a method for identifying a type of potential fault in a distribution network, comprising:

[0062] S1: Obtain ground fault data in the distribution network and calculate latent fault characteristics using signal processing technology;

[0063] Specifically, ground fault data is obtained, and various characteristics of different latent faults in different time, frequency and time-frequency domains are calculated;

[0064] The fault characteristics include the ratio of the maximum zero-sequence current to the maximum current change, the degree of zero-sequence current distortion, the degree of randomness, the information dimension, and the time-frequency characteristics based on wavelet transform.

[0065] The ratio of the maximum zero-sequence current of C1 to the maximum current change I 0ratio : Maximum zero sequence current I 0max It can reflect the overall change state of the fault relative to the fault medium. The maximum change in current during the initial transient stage ΔI 0max Represents the dynamic state of the fault, as the continuous samples (I n ,I n+1 ) is calculated. 0max With ΔI 0maxThe ratio of can reflect the fault characteristics and is not affected by the load level and network status.

[0066] ΔI 0max =max(I n+1 -I n )

[0067]

[0068] Among them, I n and I n+1 They represent the samples taken at the nth and n+1th times after the fault occurs.

[0069] C2 Zero-sequence current distortion degree: The difference in grounding medium is mapped to the nonlinear difference in transition impedance, which affects the harmonics of zero-sequence current. The distortion of the waveform compared with the sine wave is usually expressed by total harmonic distortion (THD):

[0070]

[0071] Wherein h represents the harmonic number, I(h) represents the root mean square value of the hth harmonic current. M represents the highest harmonic number, which is set to 10 in this embodiment. The logarithm of THD, i.e., lg(THD), is used to describe the distortion degree of zero-sequence current.

[0072] C3 Randomness: The randomness of waveform distortion is used to describe the steady-state characteristics of faults caused by different types of latent faults, and is defined as the standard deviation of the current waveform distortion, as shown in formula (4):

[0073]

[0074] Wherein, k represents the kth cycle of the zero-sequence current in the steady state, N represents the number of cycles in the steady state, which is set to 10 in this embodiment. μ represents the average value of the distortion lg (THD) of the zero-sequence current waveform in N cycles, where THD k It represents the total harmonic distortion of the steady-state zero-sequence current at the kth cycle, which is the same as the above THD I(h) Both represent total harmonic distortion, but they represent different current waveforms and are calculated in the same way.

[0075] C4 Information dimension: Through phase space reconstruction, an m-dimensional vector can be constructed based on the original one-dimensional time series data of zero-sequence current. t1 ,i t1 ,…,i tN ], when m = 2, for the time delay τ, the following two-dimensional vector can be obtained:

[0076]

[0077] I(t) represents the original one-dimensional time series data of zero-sequence current, i t1 -i tN Represents the zero-sequence current at different times, N represents the number of time series, N m Represents the number of m-dimensional vector time series.

[0078] Using these two vectors as the horizontal and vertical coordinates, the phase space trajectory of the zero-sequence current is obtained. Using the two-dimensional graph, the information dimension can be calculated as:

[0079]

[0080] Where r represents the square root of the total number of divided boxes. For example, if the division is 2×2, r=2, and if the division is 5×5, r=5. r is the information entropy, which is calculated as: where p i represents the probability of the phase space trajectory in the ith box.

[0081] C5–C11 time-frequency characteristics based on wavelet transform:

[0082] Wavelet transform has good adaptability to the time-frequency window and can reflect the characteristic information of the waveform at the initial stage of the fault. According to wavelet transform, seven features (C5–C11) are extracted as shown in formula (7).

[0083]

[0084] Where, f(t) represents the zero-sequence current waveform of the system fault; I m represents the zero-sequence current peak value of the fault; t0 represents the initial time of f(t), t total Indicates the end time of the selected waveform, t s Indicates the time when the fault occurs, t w Indicates the end time of the fault transient process, t mid represents the middle time of f(t), t k Indicates t mid and t total The selected time between is used to distinguish the time-frequency characteristics of zero-sequence current before and after the transient process, t m represents the peak time of f(t), t h Indicates t m After that, the first half peak value of the zero-sequence current is obtained, and n represents the number of wavelet transform layers. i Indicates that in [t s , t w ], E1 represents the energy of the i-th wavelet transform signal of the zero-sequence current in the period of [t s , t w] is the energy of the first-layer wavelet transform signal of the zero-sequence current in the period of j Indicates that in [t0, t k ], e1 represents the energy of the j-th wavelet transform signal of the zero-sequence current in the period [t0, t k ] is the energy of the first-layer wavelet transform signal of the zero-sequence current in the period of ]; U0 represents the line voltage before the fault, M s represents the maximum slope of f(t). M1 represents the slope from t0 to t w The energy of the zero-sequence current during the fault, M total From t0 to t total The energy of the zero-sequence current during the fault, M mid From t0 to t mid The energy of the zero-sequence current during the fault, M k From t0 to t k The energy of the fault zero-sequence current during the process. In this process, f(t) is a waveform of 1.1 cycles (the first cycle after the fault contains abundant transient components, so 0.1 cycle before the fault and 1 cycle after the fault are taken).

[0085] Features [C1, C2, C3, C4, C5, C6, C7, C8, C9, C10, C11] constitute the candidate waveform feature set.

[0086] S2: Use multivariate variance analysis to verify the calculated fault features and extract effective multi-domain features of latent faults;

[0087] Specifically, the calculated fault features are collected, and these raw data are preprocessed to divide the data into several groups according to different fault types or health states, each group representing a specific fault mode or normal operating state.

[0088] The null hypothesis is established, that is, it is assumed that there is no significant difference in the characteristic means among all groups; the alternative hypothesis is that the characteristic means of at least one group are significantly different from those of other groups.

[0089] Perform MANOVA analysis, using the multivariate analysis of variance method, to test whether the characteristic vectors under different fault types or states come from the same distribution, that is, to verify whether the null hypothesis is established. Calculate statistics to evaluate the significance of differences between groups.

[0090] If the result of MANOVA rejects the null hypothesis, it means that there are indeed significant differences in the characteristics of different fault types or states. At this time, it is necessary to further analyze which characteristics contribute most to the differences between groups.

[0091] By observing the main effects and interaction effects of fault characteristics, as well as statistical significance, we can identify those multi-domain features that can effectively distinguish different fault types. These features are the effective features of latent faults.

[0092] S3: Based on the screened effective multi-domain features, a latent fault type recognition model of the extreme learning machine ELM and a latent fault type classification model of the support vector machine SVM are constructed, and the recognition and classification results are output;

[0093] Specifically, establishing a latent fault type identification model based on extreme learning machine includes the following steps:

[0094] Training process: (1) Randomly select weights ω and bias B from the input of the hidden layer.

[0095] Let X = {x1, x2, …, x N x i ∈R D , i=1,2,…,N} represents a training set of N samples, D represents the dimension, R D represents D-dimensional input samples, Y = {y1, y2, ..., y N y i ∈R P} represents the vectorized label, P represents the number of categories, R P represents the output of the P-type output layer. A single hidden layer neural network with L hidden neurons and activation function g(x) can be expressed as:

[0096]

[0097] Among them, ω k and β k Represents the weight vector from the input layer to the hidden layer and from the hidden layer to the output layer respectively. k represents the bias of the kth hidden neuron, g(<ω k ,x i >+b k ) represents relative to the input sample x i The output of the kth hidden neuron.

[0098] (2) Calculate the hidden layer output matrix H.

[0099] Formula (8) can be expressed in a compact form H·β=Y as follows:

[0100]

[0101]

[0102] Where H represents the hidden layer output matrix of a single hidden layer feedforward neural network, and β represents the output weight matrix. After the user specifies the learning rate and momentum, the back propagation learning algorithm can be used to find the optimal weights and biases of a single hidden layer feedforward neural network.

[0103] (3) Calculate the output weight β from the hidden layer to the output.

[0104] Compared with the number of training samples, a single hidden layer feedforward neural network can achieve the approximation ability of a specified error level with much fewer hidden layer neurons. According to the minimum norm and least squares function, the weight matrix can be solved as:

[0105] β=H + ·Y

[0106] Among them, H + represents the Moore-Penrose generalized inverse of the matrix H. Compared with the back propagation learning algorithm, this processing can directly obtain the minimum norm of the weights with the best generalization performance without involving local minima. Considering the impact of the modified output weights on model identification, the particle swarm optimization algorithm is used to optimize its initial weights and biases.

[0107] Classification process:

[0108] (1) Use formula (9) to calculate the new input X U The hidden layer output matrix H U ;

[0109] (2) Get the class label Y of the new instance U =H U β.

[0110] The steps to establish a latent fault classification model based on support vector machine are:

[0111] (1) Normalizing the input vector containing eleven waveform features;

[0112] (2) Select the kernel function type;

[0113] For the training set {(x1,y1),…,(x i ,y i ),…,(x N ,y N )},x i ∈R D ,y i ∈{-1,1}, the basic idea of ​​support vector machine SVM is to map the input vector x i , through the kernel function From the original space R D Transfer to a high-dimensional feature space. In this high-dimensional space, the optimal decision function is constructed using the principle of structural risk minimization.

[0114] f(x)=sgn(ω·x+b)

[0115] Where ω represents the weight vector and b is a constant.

[0116] (3) Select the core parameters c and g of the model, where c represents the penalty factor, indicating the importance of abnormal data in the class, and g is the parameter coefficient of the kernel function;

[0117] Formula (12) can be transformed into an optimization problem:

[0118]

[0119] Where c is the fault tolerance penalty coefficient (c>0), ξ i is the relaxation factor (ξ i ≥0), N represents the total number of samples.

[0120] This type of optimization problem can be solved by Lagrangian method and Kuhn-Tucker theorem, and finally the optimal linear decision function is obtained. In this embodiment, a one-to-one solution method is used for multi-classification, while a traditional support vector machine algorithm is used for binary classification. In this embodiment, radial basis function is selected as the kernel function, and model parameters are set here, including c and kernel function parameter g.

[0121] (4) Setting the number of cross-validation times for the model is mainly to obtain the optimal core parameters c and g;

[0122] (5) Using the obtained optimal c and g, the multi-classification model is trained using the svmtrain function in Libsvm, and the classification accuracy is tested using the svmppredict function in Libsvm.

[0123] S4: Fusion the identification and classification results to obtain the distribution network potential fault type identification results.

[0124] Specifically, the Dempster-Shafer (DS) evidence fusion theory is used to fuse the recognition results of the two models. The steps of fusing the recognition results of the two models using the Dempster-Shafer (DS) evidence fusion theory are as follows:

[0125] (1) Establish a set O of all possible outcomes of the identification framework decision problem;

[0126] (2) Select evidence and construct a basic probability assignment (BPA);

[0127] The diagnostic results of ELM and SVM are used as two independent pieces of evidence. First, they are transformed into BPA that satisfies the DS evidence theory using the quality function. In the DS evidence theory, the reliability coefficient α is used to measure the reliability of the evidence, and its value is between 0 and 1.

[0128] The evidence quality function of latent fault type and uncertainty θ is:

[0129]

[0130] m i (θ)=1-α i

[0131] Among them, α i is the reliability coefficient of the ith piece of evidence; m i (F j ) represents the proposition F j The BPA of the ith piece of evidence; m i (θ) represents the BPA of the ith piece of evidence with uncertainty; Y ij represents the diagnostic output probability of the i-th evidence of the j-th fault type. The N in formula (14) has the same meaning as that in formula (8), and both represent the number of training sample sets.

[0132] Evidence fusion combines the recognition results of ELM and SVM and calculates BPA based on Dempster's combination rule. The formula is:

[0133]

[0134] Among them, A, B and C are three subsets of the recognition frame O, and K is a normalization constant.

[0135] Decision diagnosis, using the following rules to determine the fault type:

[0136] Rule 1:

[0137] Rule 2: m(F type )>m(θ)

[0138] Among them, F type represents the fault type finally determined by combining the identification results of ELM and SVM; F l It represents the fault type when the BPA of the proposition evidence is the largest; Rule 1 stipulates that the determined fault type has the highest BPA; Rule 2 stipulates that the BPA of the determined fault type needs to be greater than the uncertainty.

[0139] Furthermore, the present embodiment also provides a distribution network latent fault type identification system, including: an acquisition module, used to obtain grounding fault data in the distribution network, and calculate the latent fault characteristics using signal processing technology; a verification module, used to verify the calculated fault characteristics using multivariate variance analysis, and extract effective multi-domain features of the latent fault; a construction module, used to construct a latent fault type identification model of an extreme learning machine ELM and a latent fault type classification model of a support vector machine SVM based on the screened effective multi-domain features, and output the identification and classification results; an identification module, used to fuse the identification and classification results to obtain the distribution network latent fault type identification results.

[0140] This embodiment also provides a computer device, which is suitable for the method of identifying the type of potential faults in a distribution network, and includes: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement all or part of the steps of the method described in the embodiment of the present invention as proposed in the above embodiment.

[0141] This embodiment also provides a storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the method in any optional implementation of the above embodiment is executed. Wherein, the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, referred to as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, referred to as EEPROM), erasable programmable read-only memory (Erasable Programmable Read Only Memory, referred to as EPROM), programmable read-only memory (Programmable Red-Only Memory, referred to as PROM), read-only memory (Read-Only Memory, referred to as ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0142] The storage medium proposed in this embodiment and the data storage method proposed in the above embodiment belong to the same inventive concept. The technical details not fully described in this embodiment can be found in the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0143] Example 2

[0144] Reference Figure 2 and Figure 3 , which is the second embodiment of the present invention, and this embodiment provides a method for identifying potential fault types in a distribution network. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0145] Collect ground fault data from the fault indicator with a sampling rate of 4096Hz in a county distribution network. A total of 61 cases of lightning arrester breakdown, 61 cases of line discharge to crossbar, 61 cases of cable insulation damage, 55 cases of insulator breakdown (flashover), 59 cases of tree contact, and 145 cases of other ground faults were collected. Different latent fault types and corresponding symbols are shown in the table:

[0146] Table 1 Different latent fault types and corresponding symbols

[0147] symbol Latent fault type F1 Lightning arrester breakdown F2 Line to crossbar discharge F3 Cable insulation damage F4 Insulator breakdown F5 Tree Contacts F6 Other types

[0148] For the steady-state process of zero-sequence current after a single-phase grounding fault in a medium-voltage distribution network, the three-dimensional steady-state fault features in the zero-sequence current are extracted according to the differences in the steady-state waveforms of different latent faults for identification. Among them, the selected feature quantities are: C1-the ratio of the maximum zero-sequence current to the maximum current change, C2-the degree of zero-sequence current distortion, and C3-the degree of randomness.

[0149] Through phase space reconstruction, a 2D vector is constructed based on the original one-dimensional time series data of zero-sequence current to obtain the zero-sequence current phase space trajectory, from which the fault features are obtained for identification. The selected feature quantity is C4-information dimension.

[0150] For the transient process of zero-sequence current after a single-phase ground fault in a medium-voltage distribution network, a wavelet transform with good adaptability to the time-frequency window is used to reflect the characteristic information of the waveform in the initial stage of the fault. According to the wavelet transform, the seven features (C5–C11) shown are extracted. Features [C1, C2, C3, C4, C5, C6, C7, C8, C9, C10, C11] constitute the candidate waveform feature set.

[0151] Multivariate variance analysis was used to verify the fault characteristics and extract effective multi-domain features reflecting different latent faults. Using SPSS software, MANOVA was used to analyze the relationship between 11 waveform characteristics and grounding type latent fault types. The F statistics of different waveform characteristics adopted in the present invention are shown in the table:

[0152] Table 2 F statistics of different waveform characteristics

[0153] Waveform characteristics F-Data Sorting C3 289.64 1 C2 235.4 2 C1 120.25 3 C4 115.25 4 C10 108.65 5 C11 83.13 6 C5 79.84 7 C9 39.94 8 C6 18.96 9 C8 10.78 10 C7 10.61 11

[0154] The significance level α value results of the F-data of the 11 extracted waveform features all meet 0.01<α≤0.05, which is statistically significant. These 11 waveform features have a strong correlation with the latent fault type, so they can be used as feature quantities reflecting different latent fault types;

[0155] Define the ELM model parameters. Eleven waveform features are taken as input, so the number of neurons in the input layer is set to 11. Combined with the empirical formula, the number of hidden layer neurons is set to 13. The six latent faults shown in Table 1 are taken as output, and the number of neurons in the output layer is set to 6. The activation function of the hidden layer neurons is selected as the Sigmoid function, and the output vector is processed by the Softmax function, converting the value of each element in the output vector to between 0 and 1.

[0156] The training set and test set are divided, the feature vector data consisting of 11 waveform features are input, and the vector label of each latent fault is defined. The particle swarm optimization algorithm is used to optimize the initial weights and deviations to complete the training and optimization of the model.

[0157] During the construction of the model, the number of training samples for each latent fault was set to 40, and the remaining samples belonged to the test set. All 11 features were extracted as input. The particle swarm optimization algorithm (PSO) was used to optimize the initial weights and biases. The initial values ​​of c1 and c2 were set to 1.5, the size of the particle swarm was set to 30, and the fitness function was set to the sum of squared errors between the output of the ELM label and the actual fault label. The number of hidden layer neurons was set to 11, the activation function was set to the sig function, and the fitness function changed with the number of iterations, as shown in Figure 3 shown.

[0158] The test set data is used to evaluate the model performance. In various test cases, when the number of iterations reaches 500, the model is basically optimized. The recognition accuracy of each case is shown in the table:

[0159] Table 3 Recognition accuracy

[0160]

[0161] Normalize the input vector containing eleven waveform features. Select the radial basis function as the kernel function. Select the core parameters c and g of the model, where c represents the penalty factor, indicating the importance of abnormal data in the class, and g is the parameter coefficient of the kernel function.

[0162] The parameter c and kernel function parameter g in the model are optimized using the cross-validation method. The number of cross-validation of the model is set mainly to obtain the optimal core parameters c and g, which is set to 10 in the present invention. All 11 features are extracted as input. In order to facilitate the fusion of the results of different methods, the training set and test set of the SVM model are the same as those of the ELM model. Parameters c and g are limited to [-5,5], and the step size is 0.1. Using the triple cross-validation method, the best recognition result (92.8%) is obtained when c=0.5 and g=4.

[0163] Using the obtained optimal c and g, the multi-classification model is trained using the svmtrain function in Libsvm, and the classification accuracy is tested using the svmppredict function of Libsvm.

[0164] The recognition results of the two models are fused using the Dempster-Shafer (DS) evidence fusion theory.

[0165] The set O of all possible outcomes of the identification framework decision problem is established. The identification framework refers to different latent fault types, expressed as:

[0166] d={F1,F2,F3,F4,F5,F6}

[0167] Select evidence and construct basic probability assignment (BPA). Use the diagnostic results of ELM and SVM as two independent pieces of evidence. First, use the quality function to transform them into BPA that satisfies the DS evidence theory. In the DS evidence theory, the reliability of evidence is measured by the reliability coefficient α, which ranges from 0 to 1. The recognition model based on ELM and SVM shows that the recognition accuracy is higher than 80% in this study, so the reliability coefficient is set to 0.8.

[0168] Evidence fusion combines the recognition results of ELM and SVM and calculates BPA based on Dempster's combination rule.

[0169] For each set of input data, the fusion model will output a seven-dimensional vector, representing the six fault types and the degree of uncertainty in determining the fault type. The recognition accuracy of each fault type is shown in the table. It can be seen from the table that the overall recognition accuracy of the fusion model is 94.1%, which is higher than the ELM-based recognition model and the SVM-based recognition model. The data size and recognition accuracy corresponding to each latent fault type in the test set are shown in the table:

[0170] Table 4 Data size and recognition accuracy corresponding to each latent fault type

[0171] Latent fault type Test Data Recognition accuracy (%) F1 21 100 F2 21 100 F3 21 90.5 F4 15 93.3 F5 19 89.5 F6 105 93.3 Overall 202 94.1

[0172] In summary, the present invention extracts various characteristics of different latent faults in different time, frequency and time-frequency domains through the acquired ground fault data, and verifies the fault characteristics using multivariate variance analysis. Based on the extracted fault characteristics, a latent fault type recognition model based on extreme learning machine and support vector machine is established. The recognition results of the two models are fused using DS evidence fusion theory, and the comprehensive identification model based on DS evidence theory is superior to each single model. The proposed comprehensive model based on DS evidence theory can improve the detection accuracy and maintenance efficiency of latent fault types, and has a good recognition effect.

[0173] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. 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 solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for identifying potential fault types in a distribution network, characterized in that: include, Obtain ground fault data in the distribution network and calculate latent fault characteristics using signal processing technology; Multivariate variance analysis is used to verify the calculated fault features and extract effective multi-domain features of latent faults; Based on the screened effective multi-domain features, the latent fault type recognition model of extreme learning machine ELM and the latent fault type classification model of support vector machine SVM are constructed, and the recognition and classification results are output; The identification and classification results are integrated to obtain the identification results of the potential fault types of the distribution network.

2. The method for identifying the type of potential fault in a distribution network according to claim 1, characterized in that: The fault characteristics include the ratio of the maximum zero-sequence current to the maximum current change, the degree of zero-sequence current distortion, the degree of randomness, the information dimension, and the time-frequency characteristics based on wavelet transform.

3. The method for identifying the type of potential fault in a distribution network according to claim 2, characterized in that: The calculation formula of the ratio of the maximum zero-sequence current to the maximum current change is: ΔI 0max =max(I n+1 -I n ) Among them, I 0ratio is the ratio of the maximum zero-sequence current to the maximum current change, I 0max is the maximum zero-sequence current, ΔI 0max is the maximum change in current during the initial transient phase, I n and I n+1 They represent the samples taken at the nth and n+1th times after the fault occurs respectively; The calculation formula of the zero-sequence current distortion degree is: Wherein, THD is the waveform distortion of zero-sequence current, h represents the harmonic number, I(h) represents the root mean square value of the h-th harmonic current, M represents the highest harmonic number, and the logarithm of THD, i.e., lg(THD), is used to represent the degree of zero-sequence current distortion; The randomness degree includes using the randomness of waveform distortion to describe the steady-state characteristics of faults caused by different types of latent faults, and is defined as the standard deviation of the current waveform distortion, and the calculation formula is: Where k represents the kth cycle of the zero-sequence current in the steady state, N represents the number of cycles in the steady state, μ represents the average value of the zero-sequence current distortion degree lg (THD) of N cycles of the zero-sequence current waveform, and THD k Indicates the total harmonic distortion of the steady-state zero-sequence current at the kth cycle; The information dimension includes constructing an m-dimensional vector according to the original one-dimensional time series data of zero-sequence current through phase space reconstruction. t1 ,i t1 ,…,i tN ], when m = 2, the two-dimensional vector representation for the time delay τ is obtained as: Where I(t) represents the original one-dimensional time series data of zero-sequence current, i t1 -i tN Represents the zero-sequence current at different times, N represents the number of time series, N m Represents the number of m-dimensional vector time series; The phase space trajectory of the zero-sequence current is obtained, and the information dimension is calculated using a two-dimensional graph through the phase space of the zero-sequence current. The calculation formula is: In the formula, D is the information dimension, r represents the square root of the total number of divided boxes, T r is information entropy, p i is the probability of the phase space trajectory in the i-th frame.

4. The method for identifying the type of potential fault in a distribution network according to claim 3, characterized in that: The time-frequency characteristics based on wavelet transform include obtaining the zero-sequence current waveform of the system fault, determining the fault occurrence time and the end time of the fault transient process; The wavelet basis function and the number of decomposition layers are selected based on the characteristics of the signal, and the zero-sequence current waveform is subjected to wavelet transformation within a period to obtain coefficients at different scales. Calculate the energy of each layer of wavelet transform signal and the zero-sequence current energy characteristics before and after the fault. Further analyze the time-frequency characteristics of the signal based on the calculated energy characteristics, integrate the extracted time-frequency characteristics into a feature vector, and calculate the fault characteristics extracted by wavelet transform as follows: Where f(t) represents the zero-sequence current waveform of the system fault; I m represents the zero-sequence current peak value of the fault; t0 represents the initial time of f(t), t total Indicates the end time of the selected waveform, t s Indicates the time when the fault occurs, t w Indicates the end time of the fault transient process, t mid represents the middle time of f(t), t k Indicates t mid and t total The selected time between m represents the peak time of f(t), t h Indicates t m After that, the first half peak value of the zero-sequence current, n represents the number of wavelet transform layers, E i Indicates that in [t s , t w ], E1 represents the energy of the i-th wavelet transform signal of the zero-sequence current in the period of [t s , t w ] is the energy of the first-layer wavelet transform signal of the zero-sequence current in the period of j Indicates that in [t0, t k ], e1 represents the energy of the j-th wavelet transform signal of the zero-sequence current in the period [t0, t k ] is the energy of the first-layer wavelet transform signal of the zero-sequence current in the period of ]; U0 represents the line voltage before the fault, M s represents the maximum slope of f(t), M1 represents the slope from t0 to t w The energy of zero-sequence current during the fault, M total From t0 to t total The energy of zero-sequence current during the fault, M mid From t0 to t mid The energy of zero-sequence current during the fault, M k From t0 to t k The energy of the zero-sequence current of the fault during the process.

5. The method for identifying the type of potential fault in a distribution network according to claim 4, characterized in that: The construction of the latent fault type identification model of the extreme learning machine ELM comprises the following steps: The weights ω and bias B are randomly selected from the input of the hidden layer to construct a single hidden layer neural network with L hidden neurons and activation function, expressed as: In the formula, ω k and β k Represents the weight vectors from the input layer to the hidden layer and from the hidden layer to the output layer respectively; b k represents the bias of the kth hidden neuron, g(<ω k ,x i >+b k ) represents relative to the input sample x i The output of the kth hidden neuron; Calculate the hidden layer output matrix H, the calculation formula is: H·β=Y In the formula, H represents the hidden layer output matrix of the single hidden layer feedforward neural network, and β represents the output weight matrix; The output weight β from the hidden layer to the output is calculated according to the minimum norm and least squares function, and the weight matrix is ​​expressed as: β=H + ·Y In the formula, H + It represents the Moore-Penrose generalized inverse of the matrix H, and the particle swarm optimization algorithm is used to optimize the initial weights and biases.

6. The method for identifying the type of potential fault in a distribution network according to claim 5, characterized in that: The steps for constructing the latent fault type classification model of the support vector machine SVM are: Normalizing the input vector containing the latent fault features; Select radial basis function as kernel function, for the training set {(x1,y1),…,(x i ,y i ),…,(x N ,y N )},x i ∈R D ,y i ∈{-1,1}, mapping the input vector x i , through the kernel function From the original space R D Transfer to high-dimensional feature space, in high-dimensional space, use the principle of structural risk minimization to construct the optimal decision function, the expression is: f(x)=sgn(ω·x+b) In the formula, ω represents the weight vector, and b is a constant; Select the core parameters c and g of the model, where c represents the penalty factor, indicating the importance of abnormal data in the class, and g is the parameter coefficient of the kernel function. The optimal decision function is transformed into an optimization problem, expressed as: Among them, c is the fault tolerance penalty coefficient, ξ i is the relaxation factor, N represents the total number of samples; The optimal linear decision function is obtained by solving the problem through Lagrange method and Kuhn-Tucker theorem; The number of cross-validation times of the model is set to obtain the optimal core parameters, the obtained optimal core parameters are used to train the multi-classification model, and the classification accuracy is tested.

7. The method for identifying the type of potential fault in a distribution network according to claim 6, characterized in that: The fusing of the recognition and classification results comprises the following steps: The DS evidence fusion theory is used to fuse the recognition results of the two models and establish the set O of all the results of the recognition framework decision problem; Select evidence and construct basic probability assignments, using the diagnostic results of ELM and SVM as two independent pieces of evidence; The quality function is used to transform the model identification results into BPA that satisfies the DS evidence theory, and the evidence quality function of the latent fault type and uncertainty θ is constructed, which is expressed as: m i (θ)=1-α i Among them, α i is the reliability coefficient of the ith piece of evidence; m i (F j ) represents the proposition F j The BPA of the ith piece of evidence; m i (θ) represents the BPA of the ith piece of evidence with uncertainty; Y ij represents the diagnostic output probability of the i-th evidence of the j-th fault type, and N represents the number of training sample sets; Combining the recognition results of extreme learning machine ELM and support vector machine SVM, BPA is calculated based on Dempster combination rule. The calculation formula is: Among them, A, B and C are three subsets of the recognition frame O, and K is a normalization constant; Construct identification and diagnosis rules to conduct latent fault decision diagnosis and determine the type of latent fault. The rules are expressed as: Rule1: Rule2:m(F type )>m(θ) Among them, F type represents the fault type finally determined by combining the identification results of ELM and SVM; F l Indicates the fault type when the BPA of the proposition evidence is maximum.

8. A distribution network latent fault type identification system, based on the distribution network latent fault type identification method according to any one of claims 1 to 7, characterized in that: include, An acquisition module is used to acquire ground fault data in the distribution network and calculate latent fault characteristics using signal processing technology; A verification module is used to verify the calculated fault features using multivariate variance analysis and extract effective multi-domain features of latent faults; A construction module is used to construct a latent fault type recognition model of an extreme learning machine ELM and a latent fault type classification model of a support vector machine SVM based on the screened effective multi-domain features, and output recognition and classification results; The identification module is used to fuse the identification and classification results to obtain the identification results of the potential fault types of the distribution network.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for identifying the type of latent fault in a distribution network according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for identifying the type of latent fault in a distribution network described in any one of claims 1 to 7 are implemented.

Citation Information

Cited By

  • Medium-voltage power distribution single-phase earth fault fusion identification model

    CN119917981A

  • Transformer fault detection method and system

    CN120831534A

  • Post porcelain insulator internal defect detection method based on laser ultrasound

    CN121208136A

  • A laser-ultrasound-based method for detecting internal defects in post porcelain insulators

    CN121208136B

  • Cable insulation multi-parameter dynamic monitoring and fault pre-diagnosis system and implementation method

    CN121254012A