Substation short-circuit fault diagnosis method based on adaptive iterative feature parameter library

Through the adaptive iterative feature parameter library and BiLSTM-Attention network model, the problems of malfunction and calculation complexity in substation line short circuit fault diagnosis are solved, fast and accurate fault identification is achieved, and decomposition parameters are automatically selected, reducing the calculation burden.

CN118734144BActive Publication Date: 2024-11-19SKILL TRAINING CENT OF STATE GRID JIANGSU ELECTRIC POWER CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411231331.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-04
Publication Date
2024-11-19
Estimated Expiration
2044-09-04

AI Technical Summary

Technical Problem

The prior art has problems of malfunction or delayed action in the diagnosis of short-circuit faults in substations, especially in complex power grid structures, and traditional methods require manual selection of parameters, resulting in high computational complexity and low diagnostic efficiency.

Method used

Using an adaptive iterative feature parameter library method, the variational modal decomposition is optimized by improving the Northern Goshawk algorithm, and combined with the BiLSTM-Attention network model, the feature extraction and diagnosis of fault signals are achieved. This method includes fault signal preprocessing, variational modal decomposition, correlation coefficient denoising, adaptive feature library update and BiLSTM-Attention model training, automatically selecting the number of decomposition layers and punishment factors, and filtering out the features with the highest contribution rate for diagnosis.

Benefits of technology

It significantly reduces the computational complexity, improves diagnostic accuracy and efficiency, avoids the disadvantages of manually selecting parameters, and achieves fast and accurate identification of short-circuit faults in substation line.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118734144B_ABST
    Figure CN118734144B_ABST
Patent Text Reader

Abstract

The present invention discloses a substation short-circuit fault diagnosis method based on an adaptive iterative feature parameter library, using a fault recorder to collect fault signals of substation lines, and preprocessing the fault signals; using an improved northern goshawk algorithm to optimize the penalty factor α and the number of decomposition layers K in variational modal decomposition, and using the optimized variational modal decomposition to decompose the preprocessed fault signal into intrinsic modal components; using correlation coefficients to denoise and reconstruct the fault signal; using the features of the reconstructed fault signal to construct a feature parameter library; constructing a BiLSTM-Attention network model and training it using the fault signal features; inputting the features with the highest contribution rate into the trained BiLSTM-Attention network model to realize the diagnosis of substation line short-circuit faults. The present invention effectively realizes the diagnosis of line short-circuit fault states, and can be applied to substation line short-circuit fault diagnosis, fault inversion and other fields.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a line fault diagnosis technology for a transformer substation, and in particular to a transformer substation short circuit fault diagnosis method based on an adaptive iterative characteristic parameter library. Background Art

[0002] With the improvement of industrial automation, substation access equipment is becoming increasingly complex. After a substation fails, the dispatch center will receive a variety of fault alarms. Therefore, the fault type is determined through fault diagnosis methods and reported to the dispatcher. Then the dispatch center assigns operation and maintenance personnel to repair the fault based on the information and restore the normal operation of the system.

[0003] Common faults in substations include line faults, busbar faults, and transformer faults. Line faults include short circuit faults and open circuit faults, among which short circuit faults have the highest probability of occurrence. The equipment in the line connection station affects the stability of the entire station at all times, and in severe cases, it will cause the system to crash. Therefore, substation line short circuit fault diagnosis is of great significance to maintaining the safe operation of the entire station.

[0004] Traditional short-circuit fault diagnosis methods rely on relay protection devices and differential protection, usually based on preset current thresholds. When the detection current exceeds the threshold, the protection action will be triggered to isolate the faulty line. However, these methods may have problems with false operation or delayed operation when facing increasingly complex power grid structures. With the development of signal processing technology, diagnostic methods based on fault feature extraction are widely used. By analyzing the waveforms of fault current, voltage, etc., the fault features are extracted using Fourier transform, wavelet transform and other technologies to achieve accurate diagnosis of short-circuit faults. However, Fourier transform has defects in feature extraction of nonlinear signals; wavelet transform basis functions and decomposition scales need to be selected manually.

[0005] In recent years, artificial intelligence has been increasingly used in substation line short-circuit fault diagnosis. For example, algorithms such as support vector machines, decision trees, and neural networks can train models based on historical fault data to achieve fault diagnosis. These methods can improve diagnostic accuracy while also adapting to complex fault scenarios. However, these methods also have limitations: support vector machines require time to adjust parameters; decision trees tend to ignore the correlation between data; and neural networks require a large amount of data to train models. How to ensure fault diagnosis accuracy while improving processing speed and reducing computational burden is an urgent problem to be solved in the current substation line short-circuit fault diagnosis. Summary of the invention

[0006] Purpose of the invention: In view of the above problems, the purpose of the present invention is to provide a substation short-circuit fault diagnosis method based on an adaptive iterative feature parameter library.

[0007] Technical solution: The substation short-circuit fault diagnosis method based on the adaptive iterative characteristic parameter library of the present invention comprises the following steps:

[0008] Step 1: Use a fault recorder to collect fault signals of substation lines and pre-process the fault signals;

[0009] Step 2: Use the improved Northern Goshawk algorithm to optimize the penalty factor in variational mode decomposition and the number of decomposition layers K, and decompose the preprocessed fault signal into intrinsic mode components using the optimized variational mode decomposition;

[0010] Step 3: Calculate the correlation coefficient between the intrinsic mode component and the fault signal, and use the correlation coefficient to denoise and reconstruct the fault signal;

[0011] Step 4: Use the features of the reconstructed fault signal to build a feature parameter library, and perform adaptive iterative updates on the feature parameter library to select the features with the highest contribution rate;

[0012] Step 5: Build a BiLSTM-Attention network model and train it using the fault signal features. The input of the BiLSTM-Attention network model is the features of the fault signal, and the output is the probability of occurrence of a short-circuit fault in the substation line.

[0013] Step 6: Input the feature with the highest contribution rate into the trained BiLSTM-Attention network model to diagnose substation line short-circuit faults.

[0014] Furthermore, in step 2, the minimum envelope entropy function value is used as the fitness function of the improved northern goshawk algorithm, and the penalty factor of different combinations is substituted each time. The fitness value is calculated by the number of decomposition layers K. When the number of iterations reaches the maximum, the global minimum fitness value and the corresponding parameter combination are saved. and K, which are used as the corresponding parameter values ​​of the optimized variational mode decomposition.

[0015] Further, step 2 specifically includes the following steps:

[0016] Step 21: Randomly initialize the northern goshawk population in the search space. The population matrix is ​​as follows:

[0017] ,

[0018] in, X i Indicates i The location of the northern goshawk, N is the population size of the northern goshawk, M To solve the dimension, x ij For the i The northern goshawkj The original position of the dimension;

[0019] The optimization process includes an exploration phase and a pursuit phase:

[0020] Step 22: Exploration Phase

[0021] The northern goshawk randomly selects prey and attacks it in the first stage, and then finds the area where the optimal solution is located under global search;

[0022] The mathematical model of the exploration stage is:

[0023] ,

[0024] ,

[0025] ,

[0026] In the formula, For the Only the location of the prey, For the The northern goshawk The new position in the dimension, For the The northern goshawk The original position of the dimension, For the The northern goshawk Prey location information, For the Updated location information for the Northern Goshawk. For the The function value of the northern goshawk finding its prey, represents the optimal function value at the current stage, Represents the objective function value of the current stage, is [1, ] is not equal to A random natural number, is a random number in [0,1], Randomly take a value of 1 or 2;

[0027] Step 23: Hunting Phase

[0028] After the exploration, the northern goshawk chases the prey with R as the pursuit radius for local search, and introduces a random factor Optimize the original pursuit radius update formula;

[0029] The mathematical model of pursuit behavior is:

[0030] ,

[0031] ,

[0032] ,

[0033] In the formula, is the current number of iterations, T is the maximum number of iterations, For this stage The objective function value of the eagle, For the The updated location information of the eagle. For the The eagle is in The new location information of the dimension, β for The random factor of

[0034] The variational mode decomposition parameters are updated by combining global search with local search, all objective functions and the current optimal parameter combination are determined, and then the next iteration is entered until the maximum number of iterations is reached and the iteration is stopped. The optimal parameter combination corresponding to the optimal objective function value in the entire iteration process is obtained as the optimal solution of the variational mode decomposition.

[0035] Furthermore, the calculation expression of the correlation coefficient in step 3 is:

[0036] ,

[0037] In the formula, Indicates The correlation coefficient of the IMF components is Indicates the original fault signal, is the mean, represents the decomposed modal components, is the mean, represents mathematical expectation;

[0038] Calculate the standard threshold between the IMF component and the original fault signal based on the correlation coefficient , the calculation expression is:

[0039] ,

[0040] Comparative correlation coefficients With standard threshold The size relationship between The corresponding IMF components, using the retained IMF components Composition Matrix , the matrix As a reconstruction failure signal.

[0041] Furthermore, the characteristics of the fault signal in step 4 include the peak-to-peak value, variance, kurtosis, singular value entropy, energy and energy entropy of the voltage. Any feature is selected to form a basic feature parameter library, and then iteration is performed. In each iteration, any remaining feature is randomly selected to be added to the feature parameter library, and the contribution rate of the principal component in the current feature parameter library is calculated by the principal component analysis method. When the iteration is completed, the feature corresponding to the highest contribution rate is selected as the fault signal feature.

[0042] Furthermore, the BiLSTM-Attention network model in step 5 is composed of BiLSTM and Attention mechanisms. The BiLSTM is regulated by three gating functions: forget gate, input gate, and output gate.

[0043] Forgotten door pair The fault signal characteristic information is selectively filtered and retained until Calculate at every moment, the mathematical model is:

[0044] ,

[0045] In the formula, represents the sigmoid function, represents the input layer input, express Output status at all times, Indicates the current moment, represents the forget gate weight coefficient matrix, represents the forget gate bias vector, , and They are input gate output, tanh function output and unit output respectively. represents vector product, represents the output of the forget gate;

[0046] The input gate returns the new fault signal feature information to the current state and updates the unit state information through the sigmoid function and tanh function. The mathematical model is:

[0047] ,

[0048] In the formula, , and They are input gate output, tanh function output and unit output respectively. represents the input gate weight coefficient matrix, represents the input gate bias vector, represents the unit state weight coefficient matrix, represents the unit state deviation vector;

[0049] The output gate transmits the characteristic information of the fault signal at the current moment through Get unit output , the mathematical model is:

[0050] ,

[0051] In the formula, represents the output gate weight coefficient matrix, represents the output gate bias vector, Represents the output of the output gate;

[0052] Then Input into the Attention mechanism, use the Attention mechanism to filter out key information that is conducive to target identification from a large amount of information, and assign different weights according to the degree of influence of the information on the target to achieve time series prediction. The mathematical model is:

[0053] ,

[0054] In the formula, , Indicates different weights, represents the bias vector, represents the alignment score between the latent vector and the query vector, Represents different weight coefficients, Represents the first Features.

[0055] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0056] The present invention uses truncation preprocessing when processing fault signal data to achieve fault feature retention while reducing calculation complexity; by introducing random factors β The update formula of the hunting radius of the goshawk is optimized to prevent the individual from falling into the local optimum and forming INGO; INGO is used to optimize the number of decomposition layers and penalty coefficients of variational mode decomposition (VMD) to achieve automatic parameter selection, overcoming the drawback of manual parameter selection in traditional methods; the VMD with optimized parameters is used to decompose the fault signal, and the signal denoising and reconstruction are completed through the relationship between the size of the correlation coefficient; an adaptive iterative update feature parameter library that meets the feature quantity requirements is constructed, and the principal component analysis method is used to select the features with the highest contribution rate as the model input to ensure the analysis accuracy and efficiency, which has obvious advantages in the identification of substation line short-circuit faults; a substation line short-circuit fault diagnosis model based on BiLSTM-Attention is proposed, and the deep data feature learning ability of the model is enhanced by using BiLSTM and AM, making the diagnosis process faster and the results more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 Flowchart for improving the Northern Goshawk algorithm;

[0058] Figure 2 This is the structural diagram of the BiLSTM-Attention network model;

[0059] Figure 3 This is the structural diagram of BiLSTM;

[0060] Figure 4 This is the structural diagram of the Attention mechanism;

[0061] Figure 5 It is the structural diagram of LSTM;

[0062] Figure 6 is a flow chart of a short circuit fault diagnosis method;

[0063] Figure 7 It is the fault recording diagram collected by the fault recorder when a certain line fault occurs;

[0064] Figure 8 is the signal diagram after truncated preprocessing;

[0065] Fig. 9 For random operators β Comparison chart of the impact on the performance of the NGO algorithm;

[0066] Fig.10 is the fitness curve of INGO;

[0067] Fig.11 It is the time domain diagram corresponding to the IMF component;

[0068] Fig.12 It is the spectrum diagram corresponding to the IMF component;

[0069] Fig.13 is the specific correlation coefficient of the IMF component;

[0070] Fig.14 Comparison diagram of the original signal and the reconstructed signal. DETAILED DESCRIPTION

[0071] In order to make the objectives, technical solutions and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0072] The substation short-circuit fault diagnosis method based on the adaptive iterative characteristic parameter library described in this embodiment includes the following steps:

[0073] Step 1: Use a fault recorder to collect fault signals of substation lines and pre-process the fault signals;

[0074] Step 2: Use the improved Northern Goshawk algorithm to optimize the penalty factor in variational mode decomposition and the number of decomposition layers K, and decompose the preprocessed fault signal into intrinsic mode components using the optimized variational mode decomposition;

[0075] Step 3: Calculate the correlation coefficient between the intrinsic mode component and the fault signal, and use the correlation coefficient to denoise and reconstruct the fault signal;

[0076] Step 4: Use the features of the reconstructed fault signal to build a feature parameter library, and perform adaptive iterative updates on the feature parameter library to select the features with the highest contribution rate;

[0077] Step 5: Build a BiLSTM-Attention network model and train it using the fault signal features. The input of the BiLSTM-Attention network model is the features of the fault signal, and the output is the probability of occurrence of a short-circuit fault in the substation line.

[0078] Step 6: Input the feature with the highest contribution rate into the trained BiLSTM-Attention network model to diagnose substation line short-circuit faults.

[0079] In step 2, the minimum envelope entropy function value is used as the fitness function of the improved northern goshawk algorithm, and the penalty factor of different combinations is substituted each time. The fitness value is calculated by the number of decomposition layers K. When the number of iterations reaches the maximum, the global minimum fitness value and the corresponding parameter combination are saved. and K, which are used as the corresponding parameter values ​​of the optimized variational mode decomposition.

[0080] like Figure 1 As shown, step 2 specifically includes the following steps:

[0081] Step 21: Set the fault signal length N and the maximum number of iterations T, and randomly initialize the northern goshawk population in the search space. The population matrix is ​​as follows:

[0082] ,

[0083] in, X i Indicates i The location of the northern goshawk, N is the population size of northern goshawks, i.e. the original signal length, M To solve the dimension, x ij For the i The northern goshawk j The original position of the dimension;

[0084] The optimization process includes an exploration phase and a pursuit phase:

[0085] Step 22: Exploration Phase

[0086] The northern goshawk randomly selects prey and attacks it in the first stage, and then finds the area where the optimal solution is located under global search;

[0087] The mathematical model of the exploration stage is:

[0088] ,

[0089] ,

[0090] ,

[0091] In the formula, For the Only the location of the prey, For the The northern goshawk The new position in the dimension, For the The northern goshawk The original position of the dimension, For the The northern goshawk Prey location information, For the The function value of the northern goshawk finding its prey, represents the optimal function value at the current stage, Represents the objective function value of the current stage, is [1, ] is not equal to A random natural number, is a random number in [0,1], Randomly take a value of 1 or 2;

[0092] Step 23: Hunting Phase

[0093] After the exploration, the northern goshawk chases the prey with R as the pursuit radius for local search, and introduces a random factor Optimize the original pursuit radius update formula;

[0094] The mathematical model of pursuit behavior is:

[0095] ,

[0096] ,

[0097] ,

[0098] In the formula, is the current number of iterations, T is the maximum number of iterations, For this stage The objective function value of the eagle, After the update The updated location information of the eagle. For the The eagle is in The new location information of the dimension, β for The random factor of

[0099] The variational mode decomposition parameters are updated by combining global search with local search, all objective functions and the current optimal parameter combination are determined, and then the next iteration is entered until the maximum number of iterations is reached and the iteration is stopped. The optimal parameter combination corresponding to the optimal objective function value in the entire iteration process is obtained as the optimal solution of the variational mode decomposition.

[0100] The calculation expression of the correlation coefficient in step 3 is:

[0101] ,

[0102] In the formula, Indicates The correlation coefficient of the IMF components is Indicates the original fault signal, is the mean, represents the decomposed modal components, is the mean, represents mathematical expectation;

[0103] Calculate the standard threshold between the IMF component and the original fault signal based on the correlation coefficient , the calculation expression is:

[0104] ,

[0105] Comparative correlation coefficients With standard threshold The size relationship between The corresponding IMF components, using the retained IMF components Composition Matrix , the matrix As a reconstruction failure signal.

[0106] The characteristics of the fault signal in step 4 include voltage peak-to-peak value, variance, kurtosis, singular value entropy, energy and energy entropy. Any feature is selected to form a basic feature parameter library, and then it is iterated. In each iteration, any remaining feature is randomly selected to be added to the feature parameter library, and the contribution rate of the principal component in the current feature parameter library is calculated by the principal component analysis method. When the iteration is completed, the feature corresponding to the highest contribution rate is selected as the fault signal feature.

[0107] The BiLSTM-Attention network model in step 5 consists of BiLSTM and Attention mechanisms. Figure 2 This is the structure diagram of the BiLSTM-Attention network model, such as Figure 3 This is the structure diagram of BiLSTM. Figure 4 This is the structural diagram of the Attention Mechanism (AM). The BiLSTM unit state is regulated by three gating functions: the forget gate, input gate, and output gate.

[0108] Forgotten door pair The fault signal characteristic information is selectively filtered and retained until t Calculate at every moment, the mathematical model is:

[0109] ,

[0110] In the formula, represents the sigmoid function, represents the input layer input, express Output status at all times, Indicates the current moment, represents the forget gate weight coefficient matrix, represents the forget gate bias vector, , and They are input gate output, tanh function output and unit output respectively. represents vector product, represents the output of the forget gate;

[0111] The input gate returns the new fault signal feature information to the current state and updates the unit state information through the sigmoid function and tanh function. The mathematical model is:

[0112] ,

[0113] In the formula, , and They are input gate output, tanh function output and unit output respectively. represents the input gate weight coefficient matrix, represents the input gate bias vector, represents the unit state weight coefficient matrix, represents the unit state deviation vector;

[0114] The output gate transmits the characteristic information of the fault signal at the current moment through Get unit output , the mathematical model is:

[0115] ,

[0116] In the formula, represents the output gate weight coefficient matrix, represents the output gate bias vector, Represents the output of the output gate;

[0117] Then Input into the Attention mechanism, use the Attention mechanism to filter out key information that is conducive to target identification from a large amount of information, and assign different weights according to the degree of influence of the information on the target to achieve time series prediction. The mathematical model is:

[0118] ,

[0119] In the formula, , Indicates different weights, represents the bias vector, represents the alignment score between the latent vector and the query vector, Represents different weight coefficients, Represents the first Features.

[0120] The implementation process of the substation short-circuit fault diagnosis method based on the adaptive iterative feature parameter library of the present invention is described in detail through the following examples. The flow chart is as follows: Figure 6 As shown, the following steps are included:

[0121] Step 1: Fault signal acquisition and preprocessing

[0122] A line fault simulation data set is constructed through a digital simulation substation, and the fault signal is collected using a fault recorder and subjected to truncation preprocessing.

[0123] In one embodiment, a typical 220kV substation in a certain city is selected. The main configuration of the substation is as follows: the main transformer adopts a three-phase auto-coupling on-load voltage-changing power transformer with a connection group of YNa0d11; the 220kV partial outgoing line has two loops, using double busbar connection; the 110kV partial outgoing line has two loops, using double busbar connection; the 10kV partial outgoing line has two loops, using single busbar segmented connection; the station transformer capacity is 350kVA. According to the electrical configuration of the substation, a digital dynamic real-time simulation system is used to construct the substation simulation circuit model and the substation virtual secondary protection system. First, the data results of the primary modeling calculation are obtained from the DDRTS electromagnetic transient simulation platform interface, and the basic data is processed by Fourier transform and the real algorithm of the protection device. The voltage and current with time stamp are input into the virtual digital protection device model, and the operation of the virtual protection device in actual production is simulated by combining the startup criteria of the virtual protection device with the logic circuit of the protection device.

[0124] Specifically, analog quantity configuration is performed for line protection. Taking 220kV line as an example, the protection voltage configuration is respectively the analog quantity of three-phase 220kV positive bus voltage transformer, the protection current configuration is the analog quantity of the primary side of the current transformer of this line, and the opposite side protection current configuration is the analog quantity of the primary side of the current transformer of the opposite side of this line; analog quantity configuration is performed for busbar protection, the busbar voltage of I is configured as the analog quantity of the primary side of 220kV positive bus voltage transformer, the busbar voltage of II is configured as the analog quantity of the primary side of 220kV secondary bus voltage transformer, the bus coupling protection current is configured as the analog quantity of the primary side of the bus coupling current transformer, and the protection currents of the four branches are configured as the primary side analog quantities of the current transformers of No. 1 main transformer, No. 2 main transformer, 2598 line on the high voltage side, and 2599 line on the high voltage side; for transformer The protection is configured with analog quantity. The high-voltage side protection voltage is configured as the 220kV positive bus voltage transformer primary side analog quantity, the high-voltage side protection current is configured as the main transformer high-voltage side current transformer primary side analog quantity, the medium-voltage side protection voltage is configured as the 110kV positive bus voltage transformer primary side analog quantity, the medium-voltage side protection current is configured as the main transformer medium-voltage side current transformer primary side analog quantity, the low-voltage side section I branch protection voltage is configured as the 10kV section I bus voltage transformer primary side analog quantity, the low-voltage side section I branch protection current is configured as the low-voltage side section I current transformer primary side analog quantity, the low-voltage side section II branch protection voltage is configured as the 10kV section II bus voltage transformer primary side analog quantity, and the low-voltage side section II branch protection current is configured as the low-voltage side section II current transformer primary side analog quantity.

[0125] On the basis of the above system, the relevant short-circuit fault settings are carried out. When the fault is triggered, the voltage, current and other electrical quantities of the corresponding line will change suddenly first, and then the relay protection devices and other components related to the fault line will analyze and judge the fault characteristic electrical quantities monitored to make the secondary side start to act, and then send a trip signal to the circuit breaker to isolate the corresponding fault line to eliminate the fault, and save the fault recording diagram after the fault occurs. Figure 7 The figure shows the fault waveform collected by a fault recorder at a certain incoming line when a short circuit fault occurs in the station.

[0126] The performance of the proposed method is verified by using the short-circuit fault data of a substation line. The data samples correspond to 11 categories and 2200 groups, including single-phase grounding, two-phase grounding, two-phase phase-to-phase, three-phase grounding and normal state. 1980 groups are used as training sets, and the remaining 220 groups of samples are used as test sets, as shown in Table 1.

[0127] Table 1 Information about the fault types of the test set

[0128]

[0129] When a fault occurs at each node, different numbers of power frequency cycles are intercepted forward and backward respectively, which can reduce the burden of subsequent analysis and calculation while retaining the fault information.

[0130] The results are analyzed by taking a single-phase ground fault (Ag) as an example. The processing methods for other two-phase signals are the same. Based on the time when the fault occurs, 1 power frequency cycle is intercepted forward and 0.5 power frequency cycle is intercepted backward. The truncated signal is used for subsequent analysis, such as Figure 8 shown.

[0131] Step 2: Use the improved northern goshawk algorithm (INGO) to optimize the parameters of variational mode decomposition (VMD), where the parameters include the penalty factor α and the number of decomposition levels K.

[0132] The initial decomposition level of INGO is set to the minimum value =2, maximum value =10, step size is 1; the minimum value of the initial penalty factor is α min =500, minimum value α max =5000, the step size is 1; the population size is 30, the number of iterations is 50; the minimum envelope entropy function value is used as the fitness function. Fig.10 is the INGO-VMD fitness value change curve. Fig.10 It can be seen that when the fitness value reaches the minimum after 37 iterations, it tends to be stable. β The impact on the performance of the Northern Goshawk algorithm is shown in Table 2 and Fig. 9 As shown in the figure, the optimal VMD parameter combination corresponds to [K, α] = [8, 3755].

[0133] Table 2 Random factors β Impact on NGO algorithm performance

[0134]

[0135] Use the optimized VMD to decompose the preprocessed fault signal and obtain IMFs. Input waveform data of different fault categories into the optimized VMD respectively. Fig.11 , Fig.12 The eight IMF components and corresponding spectrum diagrams are obtained after decomposing the A-phase voltage signal.

[0136] Step 3: Calculate the standard threshold and the correlation coefficient between each component and the original signal.

[0137] The correlation coefficient between the IMF component and the original fault signal is calculated. The specific results are shown in Table 3.

[0138] Table 3 Correlation coefficients between each component and the original signal

[0139]

[0140] Calculate threshold value based on correlation coefficient :

[0141] ,

[0142] Fig.13 is the value of the correlation coefficient between each component and the original signal. By comparing the threshold value and the correlation coefficient, the fault signal is denoised and reconstructed. Greater than The components are considered as valid components and retained, otherwise they are considered as noise signal components and discarded. After completing the component selection, the signal is reconstructed, and the similarity of the signals before and after reconstruction is calculated. The closer the correlation coefficient is to 1, the stronger the correlation is, that is, the reconstructed signal retains the fault information while eliminating useless information, and the retained IMF components are formed into a matrix M.

[0143] Calculated threshold is 0.15357, components IMF1, IMF2, and IMF3 are valid components, and other components are removed. The results are as follows Fig.14 The correlation coefficient between the reconstructed signal and the original signal is 0.9968, which proves that the reconstructed signal retains the useful information of the original signal while eliminating useless information.

[0144] Step 4: Process the voltage signals of each phase and construct a basic characteristic parameter library. In this embodiment, the peak-to-peak value and variance are added to the characteristic parameter library, the contribution value is calculated, and then the characteristic parameter library is iteratively updated, the kurtosis is added, and the contribution rate is calculated and sorted each time by the principal component analysis method.

[0145] Then singular value entropy, energy, energy entropy, etc. are added in turn. It can be calculated that when the number of feature parameter libraries is 6, the principal component contribution rate is the highest. When the number exceeds 6, the contribution rate corresponding to the principal component will decrease. Therefore, in this embodiment, the iteration is stopped when the feature parameter library is 6. Calculate the parameters of the reconstructed fault signal, and sort the feature contribution rates by the principal component analysis method, and retain the feature parameter with the highest contribution rate as the model input. Table 4 and Table 5 give the specific contribution rates of each feature when the number of features is 3 and 6, respectively.

[0146] Table 4 Contribution rate of each sub-parameter when the number of features is 3

[0147]

[0148] It can be seen from Table 4 that when there are only 3 features in the feature parameter library, the variance contribution rate is the highest. The feature parameter library is adaptively iteratively expanded, and the number of feature parameters is expanded to 6. The contribution rate of each feature after iteration is shown in Table 5.

[0149] Table 5 Contribution rate of each sub-parameter when the number of features is 6

[0150]

[0151] After analysis, the characteristic parameter with the highest contribution rate is the singular value entropy, which is used as input to train the BiLSTM-Attention network model. Taking Ag as an example, the specific calculation process of the singular value entropy is as follows:

[0152] (1) Reconstruct the matrix Perform singular value decomposition:

[0153] ,

[0154] In the formula, U, is an orthogonal matrix, is a diagonal matrix with diagonal elements is a singular value of the matrix M and satisfies ;

[0155] (2) The singular values Normalization has:

[0156] ,

[0157] In the formula, ,and ;

[0158] (3) Definition of singular value entropy :

[0159] ,

[0160] (4) Singular value entropy Composition Matrix S ,have:

[0161] ,

[0162] The result is a 3×1024-dimensional matrix. The diagonal elements of the matrix are retained and normalized to obtain the singular value matrix :

[0163] ,

[0164] Repeat steps 1 to 6 for the three-phase voltage signals of A, B, and C of other fault types. Calculate the corresponding singular value entropy of each phase of other faults and form a fault feature set. The specific results are shown in Table 6.

[0165] Table 6 Partial fault feature set

[0166]

[0167] Step 5: Establish a BiLSTM-Attention fault diagnosis model.

[0168] Combining BiLSTM with the Attention mechanism can effectively improve the accuracy of short-circuit fault diagnosis of substation-related lines. Figure 5 As shown in the figure, BiLSTM retains the advantages of LSTM while enhancing the learning degree and utilization rate of data by extracting forward and backward information. BiLSTM uses forget gate, input gate, and output gate to adjust the unit state. The purpose of the Attention mechanism is to filter key information that is conducive to target identification from a large amount of information, and to assign different weights according to the degree of influence of the information on the target to achieve fault status prediction.

[0169] Step 6: Use the singular value entropy obtained in step 4 as the input vector, train and test the BiLSTM-Attention fault diagnosis model established in step 5, and output the final line short circuit fault diagnosis result.

[0170] The BiLSTM-Attention model of the present invention is composed of an input layer, a BiLSTM layer, an AM layer and an output layer. BiLSTM extracts data features and inputs the extracted results to the AM layer; the AM layer calculates feature weights to highlight the impact of features on the results; the output layer outputs the final results through the Softmax function. The model improves data utilization and thus improves fault diagnosis accuracy by integrating the attention mechanism.

[0171] like Figure 6The figure shows the specific identification flow chart of the BiLSTM-Attention model for substation line short-circuit faults. The final diagnosis result is presented in a confusion matrix, as shown in Table 7. As can be seen from Table 7, the BiLSTM-Attention model has 1 sample prediction error for each single-phase fault (Ag, Bg, Cg); 1 sample prediction error for each two-phase grounding fault (ABg, BCg); 1 sample prediction error for each two-phase interphase fault (AB, AC); 1 sample prediction error for the normal operation of the line without faults; and no prediction error occurs for three-phase short-circuit faults. The recognition accuracy of the BiLSTM-Attention network model constructed by the present invention for line short-circuit faults is 96.4%, which verifies the effectiveness of the BiLSTM-Attention network model in classifying substation line short-circuit faults.

[0172] Table 7

[0173]

[0174] At the same time, through the accuracy Acc, precision P, recall R and F 1-score The performance of the BiLSTM-Attention network model is further evaluated by other indicators. The indicator calculation formula is as follows:

[0175] ,

[0176] ,

[0177] ,

[0178] ,

[0179] In the formula, Refers to the number of samples that are actually positive and predicted to be positive. Refers to the number of samples that are actually positive and predicted to be negative. Refers to the number of samples that are actually negative and predicted to be positive. Refers to the number of samples that are actually negative and predicted to be negative.

[0180] The specific calculation results of the evaluation indicators are shown in Table 8.

[0181] Table 8 Evaluation index results

[0182]

[0183] As can be seen from Table 8, compared with LSTM, CNN-LSTM and BiLSTM, the BiLSTM-Attention of the present invention has been improved in all four evaluation indicators, indicating that the performance of the BiLSTM-Attention constructed by the present invention is better than that of the other three models in terms of the accuracy of substation line short-circuit fault diagnosis. In terms of processing speed, the data processing time of the proposed model is less than that of the other three models. In summary, it is shown that the BiLSTM-Attention model constructed by the present invention is superior to the comparison model in terms of processing speed and performance indicators, proving the effectiveness and feasibility of the proposed model in substation line short-circuit fault diagnosis.

Claims

1. Substation short-circuit fault diagnosis method based on adaptive iterative feature parameter library, It is characterized in that The steps include: Step 1: Use a fault recorder to collect fault signals of the substation line and perform truncation preprocessing on the fault signals; Step 2: Use the improved Northern Goshawk algorithm to optimize the penalty factor in variational mode decomposition and the number of decomposition layers K, and decompose the preprocessed fault signal into intrinsic mode components using the optimized variational mode decomposition; Step 3: Calculate the correlation coefficient between the intrinsic mode component and the fault signal, and use the correlation coefficient to denoise and reconstruct the fault signal; Step 4: Use the features of the reconstructed fault signal to build a feature parameter library, and perform adaptive iterative updates on the feature parameter library to select the features with the highest contribution rate; Step 5: Build a BiLSTM-Attention network model and train it using fault signal features; The input of the BiLSTM-Attention network model is the characteristics of the fault signal, and the output is the probability of short-circuit fault in the substation line; Step 6: Input the feature with the highest contribution rate into the trained BiLSTM-Attention network model to diagnose the short-circuit fault of the substation line; In step 2, the minimum envelope entropy function value is used as the fitness function of the improved northern goshawk algorithm, and the penalty factor of different combinations is substituted each time. The fitness value is calculated by the number of decomposition layers K. When the number of iterations reaches the maximum, the global minimum fitness value and the corresponding parameter combination are saved. and K, which are taken as the corresponding parameter values ​​of the optimized variational mode decomposition; Step 2 specifically includes the following steps: Step 21: Randomly initialize the northern goshawk population in the search space. The population matrix is ​​as follows: , in, X i Indicates i The location of the northern goshawk, N is the population size of the northern goshawk, M To solve the dimension, x ij For the i The northern goshawk j The original position of the dimension; The optimization process includes an exploration phase and a pursuit phase: Step 22: Exploration Phase The northern goshawk randomly selects prey and attacks it in the first stage, and then finds the area where the optimal solution is located under global search; The mathematical model of the exploration stage is: , , , In the formula, For the Only the location of the prey, For the The northern goshawk The new position in the dimension, For the The northern goshawk The original position of the dimension, For the The northern goshawk Prey location information, For the Updated location information for the Northern Goshawk. For the The function value of the northern goshawk finding its prey, represents the optimal function value at the current stage, Represents the objective function value of the current stage, is [1, ] is not equal to A random natural number, is a random number in [0,1], Randomly take a value of 1 or 2; Step 23: Hunting Phase After the exploration, the northern goshawk chases the prey with R as the pursuit radius for local search, and introduces a random factor Optimize the original pursuit radius update formula; The mathematical model of pursuit behavior is: , , , In the formula, is the current number of iterations, T is the maximum number of iterations, For this stage The objective function value of the eagle, For the The updated location information of the eagle. For the The eagle is in The new location information of the dimension, β for The random factor of The variational mode decomposition parameters are updated by combining global search with local search, all objective functions and the current optimal parameter combination are determined, and then the next iteration is entered until the maximum number of iterations is reached and the iteration is stopped. The optimal parameter combination corresponding to the optimal objective function value in the entire iteration process is obtained as the optimal solution of variational mode decomposition. The characteristics of the fault signal in step 4 include the peak-to-peak value, variance, kurtosis, singular value entropy, energy and energy entropy of the voltage. Any feature is selected to form a basic feature parameter library, and then it is iterated. In each iteration, any remaining feature is randomly selected to be added to the feature parameter library, and the contribution rate of the principal component in the current feature parameter library is calculated by the principal component analysis method. When the iteration is completed, the feature corresponding to the highest contribution rate is selected as the fault signal feature.

2. The substation short circuit fault diagnosis method based on the adaptive iterative characteristic parameter library according to claim 1, It is characterized in that The calculation expression of the correlation coefficient in step 3 is: , In the formula, Indicates The correlation coefficient of the IMF components is Indicates the original fault signal, is the mean, represents the decomposed modal components, is the mean, represents mathematical expectation; Calculate the standard threshold between the IMF component and the original fault signal based on the correlation coefficient , the calculation expression is: , Comparative correlation coefficients With standard threshold The size relationship between The corresponding IMF components, using the retained IMF components Composition Matrix , the matrix As a reconstruction failure signal.

3. The substation short-circuit fault diagnosis method based on the adaptive iterative characteristic parameter library according to claim 2, It is characterized in that The BiLSTM-Attention network model in step 5 consists of BiLSTM and Attention mechanisms. The BiLSTM has three gating functions: forget gate, input gate, and output gate to regulate the unit state. Forgotten door pair The fault signal characteristic information is selectively filtered and retained until Calculate at every moment, the mathematical model is: , In the formula, represents the sigmoid function, represents the input layer input, express Output status at all times, Indicates the current moment, represents the forget gate weight coefficient matrix, represents the forget gate bias vector, , and They are input gate output, tanh function output and unit output respectively. represents vector product, represents the output of the forget gate; The input gate returns the new fault signal feature information to the current state and updates the unit state information through the sigmoid function and tanh function. The mathematical model is: , In the formula, , and They are input gate output, tanh function output and unit output respectively. represents the input gate weight coefficient matrix, represents the input gate bias vector, represents the unit state weight coefficient matrix, represents the unit state deviation vector; The output gate transmits the characteristic information of the fault signal at the current moment through Get unit output , the mathematical model is: , In the formula, represents the output gate weight coefficient matrix, represents the output gate bias vector, Represents the output of the output gate; Then Input into the Attention mechanism, use the Attention mechanism to filter out key information that is conducive to target identification from a large amount of information, and assign different weights according to the degree of influence of the information on the target to achieve time series prediction. The mathematical model is: , In the formula, , Indicates different weights, represents the bias vector, represents the alignment score between the latent vector and the query vector, Represents different weight coefficients, Represents the first Features.