Weak power distribution network fault rapid identification and positioning method

Abstract

The invention relates to a weak power distribution network fault rapid identification and positioning method, and belongs to the field of power distribution network fault identification. According to the method, a fault type identification method based on multi-point time-frequency matrix fusion analysis is provided, an improved adaptive noise complete set empirical mode decomposition method and a Hilbert band-pass filtering algorithm are utilized to construct time-frequency matrixes corresponding to three-phase voltage, current and zero-sequence voltage waveforms at a bus after a fault; performing singular value decomposition, obtaining a singular spectrum and a characteristic index corresponding to the singular spectrum, and inputting the singular spectrum and the characteristic index into a multi-classification SVM model to realize identification of a weak power distribution network fault type; the invention further provides a fault section positioning method based on synchronous waveform feature extraction and matrix analysis, synchronous waveform features at multiple measurement points are extracted from a multi-domain angle, and an analysis object is converted into a random matrix feature value from a fault feature matrix by introducing a random matrix theory. And determining a cluster center point of the fault section in the feature value space by using a historical sample, and further performing online fault positioning to determine the fault section.

Description

Technical Field

[0001] The present invention belongs to the field of distribution network fault identification, and in particular relates to a method for quickly identifying and locating a weak distribution network fault. Background Art

[0002] Fault diagnosis of weak distribution networks faces the following problems: 1) The problem of locating short-circuit fault sections in scenarios where distributed power sources such as small hydropower and photovoltaics are connected, especially when the distribution network cannot achieve full measurement configuration to solve the impact of randomness and intermittency of large-scale distributed power sources and load fluctuations in the area on fault section positioning; 2) The construction of fault characteristic signals of small current grounding faults in multi-source and strong noise scenarios, the correct calibration and effective extraction of main characteristic frequency bands, so as to achieve accurate fault positioning; 3) Accurate identification of fault types. Summary of the invention

[0003] The object of the present invention is to provide a method for quickly identifying and locating faults in a weak distribution network, which method can quickly identify various types of faults in a weak distribution network and locate the fault section.

[0004] To achieve the above object, the technical solution of the present invention is: a method for quickly identifying and locating a weak distribution network fault, comprising:

[0005] A fault type identification method based on multi-point time-frequency matrix fusion analysis is proposed. The time-frequency matrix corresponding to the three-phase voltage, current and zero-sequence voltage waveforms at the bus after the fault is constructed by utilizing the improved adaptive noise complete set empirical mode decomposition method and the Hilbert bandpass filtering algorithm. The time-frequency matrix is ​​subjected to singular value decomposition and the singular spectrum is obtained. The characteristic indicators corresponding to the singular spectrum are further extracted and input into the multi-classification SVM model to realize the identification of the fault type of the weak distribution network.

[0006] In one embodiment of the present invention, the method further includes:

[0007] A fault section location method based on synchronous waveform feature extraction and matrix analysis is proposed. The synchronous waveform features at multiple measurement points are extracted from a multi-domain perspective. The analysis object is transformed from the fault feature matrix into the random matrix eigenvalue by introducing random matrix theory. The historical samples are used to determine the cluster center point of the fault section in the eigenvalue space. During online fault location, the fault section can be determined only based on the eigenvalue space distance matching results.

[0008] In one embodiment of the present invention, the fault type identification method based on multi-point time-frequency matrix fusion analysis specifically includes:

[0009] (1) After a system fault occurs, the recording device at the outlet of the low-voltage side of the main transformer records the fault waveform data; the waveform data of one cycle before and after the fault recorded by the recording device is extracted, including three-phase voltage and three-phase current; the zero-sequence voltage is calculated by Karrenbauer phase mode transformation of the three-phase voltage; the fault waveform is decomposed by ICEEMDAN to obtain k IMF components;

[0010] (2) The time-frequency matrix of each waveform is constructed by using Hilbert transform and bandpass filtering. The instantaneous frequency and amplitude of each IMF component are obtained by Hilbert transform. The bandpass filtering algorithm divides the instantaneous frequency into equal intervals, extracts the waveforms in the same frequency range, and obtains the reconstructed waveforms in the specified frequency range. The reconstructed waveform data are arranged in rows to form the corresponding time-frequency matrix. The number of columns in the time-frequency matrix is ​​equal to the number of sampling points in one cycle, and the number of rows is equal to the number of frequency bands.

[0011] (3) Perform singular value decomposition on the waveform time-frequency matrix, retain the first k-order singular values ​​as the main singular values ​​according to the requirement that the calculated cumulative contribution rate is greater than 90%; calculate the singular spectrum kurtosis K composed of the first k singular values λ 、Singular spectral entropy E λ , singular spectrum mean I λ And the singular spectrum peak factor P λ ; Considering that the singular value dimensions corresponding to different waveforms are different, the singular value statistical indicators are normalized according to the original waveform type to obtain the singular spectrum statistical indicators in the range of 0 to 1;

[0012] (4) The singular spectrum statistical indicators corresponding to each waveform are used as the extracted fault feature vector and input into the multi-classification SVM model to classify the fault type.

[0013] In one embodiment of the present invention, in step (2), considering that the instantaneous frequency of the fault signal is mainly concentrated in the range of 0 to 3 kHz, the frequency bandwidth is set to 300 Hz and the number of frequency bands is set to 10.

[0014] In one embodiment of the present invention, in step (4), the classification SVM model adopts the radial basis RBF kernel function, and the penalty factor c and the kernel parameter γ are determined by optimizing the GS method, the GA method and the PSO method.

[0015] In one embodiment of the present invention, the fault section location method based on synchronous waveform feature extraction and matrix analysis specifically includes:

[0016] (1) Extracting fault feature Z m =[z m,1 ,z m,2 ,···,z m,k], based on the historical fault recording data, calculate the fault characteristics in the time domain, frequency domain and time-frequency domain; use the fusion of ReliefF algorithm and RF algorithm to comprehensively weight the fault characteristics; set the weight threshold w set , retain the comprehensive weight greater than w set k-dimensional fault characteristics;

[0017] (2) Determine the location of the cluster center point corresponding to each segment in the Rspace space; construct a topological matrix for all historical fault samples one by one, and extract eigenvalues ​​using a random matrix algorithm; take the first three-dimensional eigenvalues ​​as the position coordinates of each fault sample in the Rspace space; use the FCM algorithm to search for cluster center points and determine the location of the center point corresponding to each segment in the Rspace space;

[0018] (3) When a single-phase grounding fault is detected in the system, k fault features of the transient zero-mode current in the time domain, frequency domain, and time-frequency domain at each measuring point are calculated on-site, packaged and uploaded to the regional master station; the fault features at each measuring point are normalized in the regional master station and a topological matrix is ​​generated; the topological matrix is ​​transformed to generate a random matrix and the eigenvalues ​​of the random matrix are obtained;

[0019] (4) The first three-dimensional eigenvalues ​​of the random matrix are extracted as its position in the Rspace space, and the distance is matched with the center point of each fault segment in the Rspace space. The segment corresponding to the closest center point is the fault segment.

[0020] In one embodiment of the present invention, the process of using the ReliefF algorithm to screen features is as follows:

[0021] 1) Select any sample X1 from the historical fault samples, the corresponding label of X1 is recorded as R, X1 contains all fault feature quantities Fp (p = 1, 2, ···, h, Fp∈Ω) when the fault occurs in segment R, and h is the total number of fault feature quantities; select k nearest neighbor samples Hj and Mj (j = 1, 2, …, k) from the remaining historical fault samples, where Hj is of the same type as X1, and Mj is of a different type from X1;

[0022] 2) Calculate and update the weight w of each feature Fp one by one according to the following formula B (F P ):

[0023]

[0024] In the formula, Indicates that F after the mth iteration calculation p The weight calculation result is set to 0; P(C) represents the proportion of class C samples in the fault feature set Ω; d(F p ,X 1 ,Hj ) represents sample X 1 With H j In feature F p The distance on the p ,X 1 ,M j ) represents sample X 1 With M j In feature F p The distance on.

[0025] In one embodiment of the present invention, when the ReliefF algorithm is used, the number of iterations m is 20, and the number of nearest neighbor samples K is 10.

[0026] In one embodiment of the present invention, the process of using the RF algorithm to screen features is as follows:

[0027] 1) Using the bootstrap resampling method, q training sample sets and q out-of-bag OOB data sets are generated from the historical fault feature set Ω;

[0028] 2) Using the k-th training sample set, train the decision tree Treek and calculate the classification accuracy Tk of the corresponding k-th out-of-bag OOB data set;

[0029] 3) Apply disturbances to the fault feature quantities Fp (p=1, 2, ···, h) in the out-of-bag OOB data set in turn, and recalculate the classification accuracy Tk,p;

[0030] 4) Let k = k + 1, repeat steps 2) and 3) to calculate Tk and Tk,p;

[0031] 5) Calculate the importance or weight of the fault feature quantity Fp. The formula is as follows:

[0032]

[0033] In order to avoid deviation when using a single method to screen fault features, different preference coefficients are set for the two methods with reference to the weighting results of the ReliefF algorithm and the RF algorithm, and the independent weight results are combined to form a comprehensive weight; the comprehensive weight is a linear combination of the weight calculation results of the two methods, satisfying:

[0034]

[0035] In the formula, w i is the comprehensive weight of the i-th fault feature; β represents the weight preference coefficient of the RF algorithm, and (1-β) is the weight preference coefficient of the ReliefF algorithm; is the weight of the i-th fault feature provided by the RF algorithm, is the weight of the i-th fault feature provided by the ReliefF algorithm;

[0036] In order to determine the value of β, the objective function is established with the goal of minimizing the sum of squares of the deviations between the comprehensive weight and the ReliefF weight and the RF weight:

[0037]

[0038] If O can take the minimum value 0, then the derivative of the above formula can be obtained to get β = 0.5, and the above formula is converted to:

[0039]

[0040] Finally, the comprehensive weight set of each fault characteristic is calculated as W = [w 1 ,w 2 ,···,w n ] T , select the features with larger weights as the objects of subsequent analysis.

[0041] The present invention also provides a computer-readable storage medium on which computer program instructions that can be executed by a processor are stored. When the processor executes the computer program instructions, any of the above-mentioned method steps can be implemented.

[0042] Compared with the prior art, the present invention has the following beneficial effects: The method of the present invention proposes fault type identification based on multi-point time-frequency matrix fusion analysis, quickly identifies various types of faults, and improves fault handling capabilities. It also proposes a method for locating a small current grounding fault section using multi-dimensional fault feature extraction and matrix analysis to avoid the risk of failure of a single fault feature. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a flow chart of the weak distribution network fault type identification according to the present invention.

[0044] Figure 2 Schematic diagram of 10kV distribution network topology and fault location.

[0045] Figure 3 This is the SVM parameter optimization result (GA method).

[0046] Figure 4 This is the multi-classification SVM fault classification result.

[0047] Figure 5 is the influence of noise on the singular spectrum of the time-frequency matrix.

[0048] Figure 6 Calculate the results of fault characteristics for different measurement points.

[0049] Figure 7 It is a fault section location process based on synchronous waveform feature extraction and matrix analysis.

[0050] Figure 8 This is the topology diagram of the 10kV active distribution network.

[0051] Fig. 9 The fault characteristic quantity changes at measurement point 2 when the fault occurs in different sections.

[0052] Fig.10 This is the result of fault feature weight calculation.

[0053] Fig.11 Comparison of the first 6 dimensional eigenvalues ​​of different fault samples.

[0054] Fig.12 It is the location distribution of single-phase grounding fault samples in Rspace. DETAILED DESCRIPTION

[0055] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings.

[0056] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0057] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0058] The present invention provides a method for quickly identifying and locating a weak distribution network fault, comprising:

[0059] A fault type identification method based on multi-point time-frequency matrix fusion analysis is proposed. The time-frequency matrix corresponding to the three-phase voltage, current and zero-sequence voltage waveforms at the bus after the fault is constructed by utilizing the improved adaptive noise complete set empirical mode decomposition method and the Hilbert bandpass filtering algorithm. The time-frequency matrix is ​​subjected to singular value decomposition and the singular spectrum is obtained. The characteristic indicators corresponding to the singular spectrum are further extracted and input into the multi-classification SVM model to realize the identification of the fault type of the weak distribution network.

[0060] A fault section location method based on synchronous waveform feature extraction and matrix analysis is proposed. The synchronous waveform features at multiple measurement points are extracted from a multi-domain perspective. The analysis object is transformed from the fault feature matrix into the random matrix eigenvalue by introducing random matrix theory. The historical samples are used to determine the cluster center point of the fault section in the eigenvalue space. During online fault location, the fault section can be determined only based on the eigenvalue space distance matching results.

[0061] The following is the specific implementation process of the present invention.

[0062] like Figure 1 As shown, this embodiment provides a method for quickly identifying and locating a weak distribution network fault, including:

[0063] 1. Fault type identification based on multi-point time-frequency matrix fusion analysis

[0064] Improving the fault handling capabilities of distribution networks in rural / mountainous areas and ensuring the safety and reliability of weak distribution networks requires, on the one hand, accurate positioning and timely isolation of the fault area, and on the other hand, referring to the fault type, fault cause and other information to formulate reasonable operation, maintenance and inspection plans.

[0065] There are few available fault samples for weak distribution network faults. Taking into account the specific fault phase, there are 10 types of distribution network faults. Reasonable selection of feature quantities that can significantly reflect the differences between various fault types and classifiers suitable for solving small sample problems is the premise for ensuring accurate identification of distribution network fault types. The time-frequency matrix corresponding to the three-phase voltage, current and zero-sequence voltage waveforms at the bus after the fault is constructed using the improved adaptive noise complete set empirical mode decomposition method and the Hilbert bandpass filter algorithm. The time-frequency matrix is ​​subjected to singular value decomposition and the singular spectrum is obtained, and the characteristic indicators corresponding to the singular spectrum are further extracted. The support vector machine (SVM) suitable for solving the small sample problem is selected as the classifier to construct a multi-classification SVM model. The statistical indicators of the singular spectrum of the time-frequency matrix are input into the multi-classification SVM to realize the identification of the distribution network fault type.

[0066] 1.1 Fault type identification process

[0067] like Figure 1 As shown in the figure, the fault type identification process of weak distribution network is as follows:

[0068] (1) After a system fault occurs, the recording device at the outlet of the low-voltage side of the main transformer records the fault waveform data. The waveform data of one cycle before and after the fault recorded by the recording device is extracted, including three-phase voltage and three-phase current. The zero-sequence voltage is calculated by Karrenbauer phase mode transformation of the three-phase voltage. The above fault waveform is decomposed by ICEEMDAN to obtain k IMF components.

[0069] (2) The time-frequency matrix of each waveform is constructed using Hilbert transform and bandpass filtering. The instantaneous frequency and amplitude of each IMF component are obtained through Hilbert transform. The bandpass filtering algorithm divides the instantaneous frequency into equal intervals, extracts the waveforms in the same frequency range, and obtains the reconstructed waveforms in the specified frequency range. The reconstructed waveform data is arranged in rows to form the corresponding time-frequency matrix. The number of columns in the time-frequency matrix is ​​equal to the number of sampling points in one cycle, and the number of rows is equal to the number of frequency bands. Considering that the instantaneous frequency of the fault signal is mainly concentrated in the range of 0 to 3 kHz, the bandwidth is set to 300 Hz and the number of frequency bands is set to 10.

[0070] (3) Perform singular value decomposition on the waveform time-frequency matrix, and retain the first k-order singular values ​​as the main singular values ​​based on the requirement that the calculated cumulative contribution rate is greater than 90%. Calculate the singular spectrum kurtosis K composed of the first k singular values λ 、Singular spectral entropy E λ , singular spectrum mean I λ And the singular spectrum peak factor P λ Considering that the singular value dimensions corresponding to different waveforms are different, the singular value statistical indicators are normalized according to the original waveform type to obtain the singular spectrum statistical indicators in the range of 0 to 1.

[0071] (4) The singular spectrum statistical indicators corresponding to each waveform are used as the extracted fault feature vectors and input into SVM for fault type classification. SVM uses the radial basis function (RBF) kernel function, and the penalty factor c and kernel parameter γ are determined by GS method, GA method and PSO method.

[0072] 1.2 Example Analysis

[0073] Build in PSCAD Figure 2 The 10kV distribution network model is shown in Figure 1. The test samples and training samples are obtained by batch simulation using the PSCAD software called by Python. Considering the factors such as the fault type, initial fault phase angle, transition resistance, and fault phase, a total of 1296 groups of samples are selected for SVM training as shown in Table 1.

[0074] Table 1 Distribution of SVM training samples

[0075]

[0076] The samples where the faults occurred at k10 to k18 were selected as test samples, and the total number of test samples was 1296 groups. Except that the fault location was different from the training samples, the fault scenarios of the test samples were consistent with the training samples in Table 1. The penalty factor c and kernel parameter γ of SVM were optimized using the PSO method, GA method, and GS method. The optimization process of the three algorithms took 234s, 31s, and 105s, respectively, and the classification accuracy rates were 99.18%, 99.38%, and 99.79%, respectively. Although the parameter optimization process of the GS method takes a long time, it has the highest accuracy, so the parameter values ​​determined by the GA method are selected ( Figure 3 ) as the final value of the penalty factor and kernel parameter, namely c = 63.44 and γ = 3.13.

[0077] The final test results are as follows Figure 4 As shown in Table 2, the algorithm has a 100% recognition rate for single-phase grounding faults and two-phase grounding faults; the recognition rates for two-phase short circuits and three-phase short circuits are lower, at 96.30% and 98.77% respectively. On the one hand, when a two-phase AB short circuit occurs, the voltage oscillation process of phase C may also be more intense, causing the corresponding low-order singular value of the time-frequency matrix to be greater than that of the other two phases, resulting in misjudgment. On the other hand, the small number of samples of three-phase short circuit faults may also be the reason why the algorithm's recognition rate for three-phase short circuit faults is lower than that for single-phase grounding faults. However, on the whole, the fault classification accuracy of the algorithm proposed in the present invention is as high as 98.77%, which is at a relatively high level and is less affected by the fault phase, fault location, transition resistance and initial phase angle of the fault.

[0078] Table 2 Test results statistics

[0079]

[0080] The noise in the distribution network is generally 40-50 dB. In order to study the noise resistance of the fault identification algorithm proposed in this invention, Gaussian white noise of 20-50 dB is added to the fault signal to simulate the noise level of the actual distribution network. Taking the zero-sequence voltage as an example, the zero-sequence voltage changes after adding noise with different signal-to-noise ratios as shown in the figure below. Figure 5 As shown in (a), the corresponding time-frequency matrix singular spectrum is Figure 5 (b) is shown. Figure 5It can be seen that the changing trends of the low-order main singular values ​​and singular spectra under different noise levels are basically the same, and the first 6-order singular values ​​basically cover the fault characteristics. The influence of noise on the singular spectrum of the time-frequency matrix is ​​small, and the statistical indicators calculated according to the singular spectrum also change relatively little. The test results under noise interference are shown in Table 3. The test results show that as the signal-to-noise ratio decreases, the recognition accuracy of the algorithm decreases slightly, but the recognition accuracy remains above 97%, which shows that the algorithm proposed in the present invention has good anti-noise performance. It can be seen that the method of selecting singular values ​​and their statistical indicators that are relatively stable to disturbances and noise as fault features for fault classification in the present invention is reasonable and effective.

[0081] Table 3 Test results under noise interference

[0082]

[0083] Distributed generation is increasingly connected to the distribution network, changing the original network flow direction. At the same time, power electronic equipment also brings harmonic interference. In the scenario with distributed generation, it is also an important task to analyze the adaptability of the algorithm. In order to study the impact of DG access on the classification effect of the algorithm, the change Figure 2 The number and position of DG access were adjusted and retested, and the test results are shown in Table 4. The test results show that DG access has little effect on the algorithm recognition effect, and the algorithm recognition accuracy is still at a high level.

[0084] Table 4 Test results under different DG access conditions

[0085]

[0086] Table 5 is the comparison result of different algorithms, among which all algorithms select the time-frequency matrix singular spectrum statistical index proposed in the present invention as input feature. It can be seen from Table 5 that even if there are differences in classifiers, the recognition accuracy of various algorithms is above 95%, that is, they all have high accuracy. The algorithm proposed in the present invention has the highest accuracy, which shows that the multi-level SVM after parameter optimization is more suitable for distribution network fault type identification.

[0087] Table 5 Comparison of recognition accuracy of different algorithms

[0088]

[0089] 2. Fault section location based on synchronous waveform feature extraction and matrix analysis

[0090] The characteristics of small current grounding faults in rural / mountainous distribution networks are weak, the fault scenarios are diverse, the fault characteristics are complex and vary greatly, and the large number of distributed power sources connected has changed the original power flow direction and fault characteristics. The fault section location algorithm using a single steady-state or transient quantity is at risk of failure. With the rapid development of a new generation of measurement technology, whether it is PMU, WMU, intelligent distribution terminal, or transient waveform type fault indicator, it provides rich waveform information for fault section location. Using the synchronized waveform information of multiple measurement nodes to globally mine fault characteristics can effectively make up for the shortcomings of traditional location algorithms. Waveform characteristics can be obtained from the perspectives of time domain, frequency domain, and time-frequency domain. The more fault characteristics are extracted, the higher the accuracy of the measurement device and the computing power of the regional master station are required, and the sensitivity of different fault characteristics to faults also varies. This shows that it is necessary to screen the fault characteristics to a certain extent and extract better fault characteristics to characterize the changing laws of electrical quantities at the measurement points when faults occur in different sections.

[0091] 2.1 Fault feature extraction

[0092] Since single-domain and unique features have blind spots in some fault scenarios (for example, when the lengths of upstream and downstream lines of the fault point are the same), multi-dimensional features of zero-mode current are extracted from the time domain, frequency domain, and time-frequency domain to form the fault feature set Ω. In order to avoid feature redundancy and reduce data transmission pressure, feature dimensionality reduction is performed through fault feature optimization.

[0093] The fault feature set Ω is specifically shown in Table 6, where transient energy is used to describe the energy difference of the fault waveform at different measurement points; the first wave peak ratio is used to reflect the difference in the ratio of the mean value of the fault waveform in the maximum rising edge of the first wave head; the Hu invariant moment is used to reflect the difference in the deformation of the fault waveform; the kurtosis and other features are commonly used in the field of fault diagnosis. Because the centroid frequencies of the zero-sequence current waveforms of the lines on both sides of the fault point are different, the complexity of the amplitude-frequency and phase-frequency curves is also different, so the sample entropy is introduced to quantify the complexity, so the transient centroid frequency, amplitude-frequency sample entropy, phase-frequency sample entropy and other features are selected in the frequency domain; the power spectrum entropy can quantitatively describe the complexity of the energy distribution of zero-mode current in the frequency domain. In the time-frequency domain, wavelet singular entropy is a feature of the fusion of wavelet transform, singular value and information entropy, which is widely used in fault classification, fault line selection and other fields; the band energy entropy is used to reflect the difference in the energy ratio of each frequency band of zero-mode current at different measurement points.

[0094] Table 6 Fault feature set Ω

[0095]

[0096] 2.2 Fault feature screening

[0097] The actual operation conditions of a distribution network are complex, and the fault characteristics are even weaker when the resonant grounding method is adopted for the neutral point. Extracting fault characteristics from the time domain, frequency domain, and time-frequency domain as the basis for fault section location can avoid the risk of misjudgment caused by using a single fault characteristic. However, the more fault characteristics are extracted, the higher the requirements for the accuracy of the measurement device and the computing power of the regional master station, and there are also differences in the sensitivity of different fault characteristics to faults. Table 6 constructs 15 fault characteristics from the perspectives of the time domain, frequency domain, and time-frequency domain. Assuming that all fault characteristics are used for fault section location during each online operation, it will greatly increase the computing pressure of the measurement device. At the same time, for a single fault type, there are some redundant characteristics; the existence of these characteristics may lead to the failure of the fault section location algorithm. Therefore, it is necessary to screen out the optimal characteristics from the above characteristics for algorithm design.

[0098] Figure 6 It is the calculation result after normalizing the fault characteristics at different measurement points when a single-phase grounding fault occurs in a 10 kV distribution network. Among them, PMU5 is located downstream of the fault point, and the rest of the PMUs are located upstream of the fault point. For most fault characteristics, the calculated values on the same side of the fault point are similar, while the calculated values on the opposite side of the fault point vary greatly; some fault characteristics are greatly affected by the distance between the measurement point and the fault point. This example also shows the necessity of screening the original fault characteristic set to extract fault characteristics with good stability and high sensitivity to faults while avoiding the curse of dimensionality, and improving the reliability of the fault section location algorithm.

[0099] Therefore, before using the fault characteristics for fault section location, the fault characteristics are first screened using historical fault data. To avoid bias when using a single method to screen the fault characteristic quantity, a combination of the ReliefF algorithm and the random forest algorithm is used to screen the fault characteristics. As a filtering feature selection method, the ReliefF algorithm has the characteristics of low computational complexity and good general performance, and is suitable for feature weighting and screening of multi-feature and multi-class samples. The process of feature screening by the ReliefF algorithm is as follows:

[0100] (1) Select any sample X1 from the historical fault samples, and the label corresponding to X1 is denoted as R. X1 contains all the fault characteristic quantities Fp (p = 1, 2, ···, h, Fp ∈ Ω) when the fault occurs in section R, where h is the total number of fault characteristic quantities; select k nearest neighbor samples Hj and Mj (j = 1, 2, …, k) from the remaining historical fault samples, where Hj is of the same class as X1 and Mj is of a different class from X1.

[0101] (2) Calculate and update the weight w of each feature Fp one by one according to the following formula B (F P )

[0102]

[0103] In the formula, Indicates that F after the mth iteration calculation p The weight calculation result is set to 0; P(C) represents the proportion of class C samples in the fault feature set Ω; d(F p ,X 1 ,H j ) represents sample X 1 With H j In feature F p The distance on the p ,X 1 ,M j ) represents sample X 1 With M j In feature F p When using the ReliefF algorithm, the nearest neighbor sample number K is 10, which is a good effect. Therefore, in the example analysis part, the number of iterations m is 20 and the number of nearest neighbor samples K is 10.

[0104] The process of feature screening using the Random Forest (RF) algorithm is as follows:

[0105] (1) With the help of the bootstrap resampling method, q training sample sets and q out-of-bag (OOB) data sets are generated from the historical fault feature set Ω.

[0106] (2) Use the k-th training sample set to train the decision tree Treek and calculate the classification accuracy Tk of the corresponding k-th OOB data set.

[0107] (3) Apply perturbations to the fault feature quantities Fp (p=1, 2, ···, h) in the OOB data set in turn, and recalculate the classification accuracy Tk,p.

[0108] (4) Let k = k + 1, repeat steps (2) and (3) to calculate Tk and Tk,p.

[0109] (5) Calculate the importance (weight) of the fault feature value Fp. The formula is as follows:

[0110]

[0111] In order to avoid deviation when using a single method to screen fault features, refer to the weighting results of the ReliefF algorithm and the RF algorithm, set different preference coefficients for the two methods, and combine the independent weight results to form a comprehensive weight. The comprehensive weight is a linear combination of the weight calculation results of the two methods, satisfying:

[0112]

[0113] In the formula, w i is the comprehensive weight of the i-th fault feature; β represents the random forest weight preference coefficient, and (1-β) is the ReliefF weight preference coefficient; is the weight of the i-th fault feature provided by the random forest, is the weight of the i-th fault feature provided by the ReliefF algorithm.

[0114] In order to determine the value of β, the objective function is established with the goal of minimizing the sum of squares of the deviations between the comprehensive weight and the ReliefF weight and the RF weight.

[91] :

[0115]

[0116] If O can take the minimum value 0, then the derivative of the above formula can be obtained to get β = 0.5, and the above formula is converted to:

[0117]

[0118] Finally, the comprehensive weight set of each fault characteristic is calculated as W = [w 1 ,w 2 ,···,w n ] T , select the features with larger weight as the subsequent analysis object. The weight preference coefficient of the present invention is 0.5, and the first 4-dimensional features with larger comprehensive weight are selected as the preferred features.

[0119] 2.3 Fault section location process based on synchronous waveform feature extraction and matrix analysis

[0120] like Figure 7 As shown in the figure, the main idea of ​​small current grounding fault location based on multi-point synchronous waveform and matrix analysis can be summarized as follows: extract the synchronous waveform features at multiple measurement points from a multi-domain perspective, transform the analysis object from the fault feature matrix into the random matrix eigenvalue by introducing the random matrix theory, and use historical samples to determine the cluster center point of the fault section in the eigenvalue space. When locating faults online, the fault section can be determined based on the eigenvalue space distance matching results.

[0121] (1) Extracting fault feature Z m =[z m,1 ,z m,2 ,···,z m,k ] Based on the historical fault recording data, calculate the fault characteristics in the time domain, frequency domain and time-frequency domain. Use the fusion of ReliefF algorithm and RF algorithm to comprehensively weight the above fault characteristics. Set the weight threshold w set , retain the comprehensive weight greater than w set k-dimensional fault features.

[0122] (2) Determine the location of the cluster center point corresponding to each segment in the Rspace space. Construct a topological matrix for all historical fault samples one by one, and extract eigenvalues ​​using a random matrix algorithm. Take the first three-dimensional eigenvalues ​​as the position coordinates of each fault sample in the Rspace space. Use the FCM algorithm to search for cluster center points and determine the location of the center point corresponding to each segment in the Rspace space.

[0123] (3) When a single-phase grounding fault is detected in the system, k fault features of the transient zero-mode current in the time domain, frequency domain, and time-frequency domain at each measuring point are calculated on-site, packaged and uploaded to the regional master station. The fault features at each measuring point are normalized in the regional master station and a topological matrix is ​​generated. The topological matrix is ​​transformed to generate a random matrix and the eigenvalues ​​of the random matrix are obtained.

[0124] (4) Extract the first three-dimensional eigenvalues ​​of the random matrix as its position in the Rspace space, and perform distance matching with the center point of each fault segment in the Rspace space. The segment corresponding to the closest center point is the fault segment.

[0125] 2.4 Algorithm Testing

[0126] Build in PSCAD Figure 8 The 10kV three-outlet distribution network model with distributed generation is shown. Except for line 1 which is an overhead line, the other lines are mixed lines containing cables and overhead lines. The parameters of the line, transformer, load, etc. are set in the same way as above and will not be repeated. DG1 is a photovoltaic, 0.4MW; DG2 is a wind turbine, 0.35MW. The system has a total of 10 measurement points and is artificially divided into 10 sections. A total of 9 fault locations are set in different sections. Each fault location is set with three single-phase grounding faults of phase A, phase B, and phase C. The transition resistance is 1, 100, 300, 500, 1000, and 1500Ω respectively, and the initial phase angle of the fault is 0°, 30°, 90°, and 150°. Using Python to call PSCAD for batch simulation, a total of 648 sets of simulation waveform data are obtained. With the help of Matlab, the waveform data in COMTRADE format is converted into time series data and the algorithm is verified.

[0127] Taking the fault characteristic calculation results at measurement point 2 when single-phase grounding faults (transition resistance is 300Ω) occur in k2~k5 as an example, the difference in fault characteristics at the same measurement point when faults occur in different sections is explained. The characteristic quantity calculation results are presented in the form of a three-dimensional histogram, as shown in Fig. 9As shown. The fault features corresponding to numbers 1 to 15 are: Hu invariant moment, transient energy, kurtosis, skewness, waveform factor, peak factor, pulse factor, margin factor, first wave peak ratio, transient center of gravity frequency, amplitude frequency sample entropy, phase frequency sample entropy, power spectrum entropy, wavelet singular entropy and band energy entropy. Fig. 9 It can be seen that when the fault occurs upstream and downstream of the measurement point, some fault features at the same measurement point have obvious differences, such as transient energy, band energy entropy, etc.; some fault feature quantities have no obvious change pattern, such as waveform factor. Fault features that are more sensitive to changes in the fault section and are less affected by the initial phase angle and transition resistance of the fault are selected as the preferred fault features for the design of subsequent fault section location methods.

[0128] The RF algorithm and the ReliefF algorithm are used to select fault features. The comprehensive weights of each feature calculated based on historical fault samples are as follows: Fig.10 As shown. The fault features with weight values ​​exceeding 0.08 are transient energy, first wave peak ratio, wavelet singular entropy, and band energy entropy, with weights of 0.193, 0.088, 0.128, and 0.108, respectively. The above four feature quantities are selected as elements of the fault feature matrix. When locating the fault section online, it is only necessary to calculate these four fault features on-site and then upload them to the regional master station.

[0129] Considering the different dimensions of various fault characteristic quantities, the characteristic matrix is ​​normalized in blocks. The topological matrix is ​​calculated using the normalized characteristic quantities at all measurement points, and then the corresponding eigenvalues ​​are obtained by the random matrix eigenvalue calculation method. Nine samples are selected from the historical fault samples, corresponding to single-phase grounding faults in sections 1 to 9; the eigenvalue distribution corresponding to the samples is as follows: Fig.11 As shown in the figure, the eigenvalues ​​corresponding to the samples decrease as the dimension increases, and the fault information is mainly reflected in the first 3-dimensional eigenvalues. More importantly, the first 3-dimensional eigenvalues ​​vary greatly when faults occur in different sections. Assuming that the first 3-dimensional eigenvalues ​​corresponding to each sample are the spatial coordinates of the sample in the 3-dimensional space Rspace, the positions of different samples in the Rspace space are obviously different.

[0130] Perform random matrix transformation on the topological matrix corresponding to all historical samples and obtain the random matrix eigenvalues, retaining the first three-dimensional eigenvalues ​​as the position of the sample in the Rspace space. The distribution of each historical sample in the Rspace space is shown in Fig.12 (a). Fig.12 It can be seen that the fault samples of the same fault section show obvious clustering in the Rspace space, and the historical samples of different fault sections are distributed far away in the Rspace space. The cluster center points obtained by the FCM clustering method are as follows: Fig.12 (b) and shown in Table 7.

[0131] Table 7 Cluster center points corresponding to different fault sections

[0132]

[0133] When locating the fault section online, it is necessary to calculate the eigenvalue of the random matrix corresponding to the sample to be tested, so as to determine the position of the sample to be tested in the Rspace space. The distance matching result between the sample to be tested and the center point of each cluster in the Rspace space is used for section positioning, that is, the cluster center point with the smallest spatial distance is found by optimization, and the section represented by the center point is the fault section. 324 groups of samples to be tested were tested, and the algorithm can give accurate positioning results. Table 8 shows the fault section positioning results under some high-resistance grounding fault conditions. Considering that the actual distribution network noise is generally 40dB~50dB, 30dB~50dB Gaussian white noise is added to the sample to be tested in turn to test the noise resistance of the algorithm. Some test results are shown in Table 9. It can be seen from the data in the table that the positioning results of the algorithm proposed in the present invention are accurate under different noise levels, indicating that the fault section positioning algorithm proposed in the present invention has certain noise resistance.

[0134] Table 8 Partial fault section location results (R f =1500Ω)

[0135]

[0136] Table 9 Fault section location results under different noise levels

[0137]

[0138] The present invention studies the identification of distribution network fault types, proposes a distribution network fault type identification method based on the singular spectrum characteristics of the time-frequency matrix and the multi-classification support vector machine model, and conducts detailed simulation tests or field data verification on the above fault type identification method. The results show that the time-frequency matrix constructed by ICEEMDAN decomposition and Hilbert bandpass filtering algorithm covers the time-frequency characteristics of the fault waveform in each sub-band, and contains the time-frequency localization information that characterizes the essential characteristics of the fault waveform. The singular values ​​of the time-frequency matrix are relatively stable and are less affected by disturbances and noise. Using the singular values ​​of the time-frequency matrix to construct fault feature quantities is conducive to improving the robustness of the algorithm.

[0139] Furthermore, various fault features were established from the perspectives of time domain, frequency domain and time-frequency domain. The fault features of all historical samples were calculated, and the ReliefF algorithm and random forest algorithm were integrated to optimize the fault features. The optimized fault features constituted the characteristic matrix of the sample, and then the characteristic matrix was calculated by topological matrix and random matrix transformation, and finally the first 3-dimensional eigenvalues ​​of the random matrix were used to identify the fault section. The effectiveness of the proposed fault section method was verified by PSCAD simulation model test and real test field test.

[0140] The method proposed in the present invention not only avoids the risk of a single fault feature easily causing the failure of the judgment criteria, but also avoids the risk of too many fault features interfering with each other and increasing the computing pressure of the measuring equipment and the regional master station. The results of simulation tests and actual data tests show that the fault section location algorithm proposed in the present invention is not affected by the fault location, transition resistance, fault initial phase angle, and noise. It should be noted that the algorithm belongs to a data-driven model, and the recognition effect depends on historical samples; the richer the available historical samples, the more accurate the center point position corresponding to the section, and the better the fault section recognition effect.

[0141] The present invention also provides a computer-readable storage medium on which computer program instructions that can be executed by a processor are stored. When the processor executes the computer program instructions, any of the above-mentioned method steps can be implemented.

[0142] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0143] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0144] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0145] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0146] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

Claims

1. A method for rapid identification and location of weak distribution network faults, characterized in that: include: A fault type identification method based on multi-point time-frequency matrix fusion analysis is proposed. The time-frequency matrix corresponding to the three-phase voltage, current and zero-sequence voltage waveforms at the bus after the fault is constructed by utilizing the improved adaptive noise complete set empirical mode decomposition method and the Hilbert bandpass filtering algorithm. The time-frequency matrix is ​​subjected to singular value decomposition and the singular spectrum is obtained. The characteristic indicators corresponding to the singular spectrum are further extracted and input into the multi-classification SVM model to realize the identification of the fault type of the weak distribution network.

2. A method for rapid identification and location of weak distribution network faults according to claim 1, characterized in that: Also includes: A fault section location method based on synchronous waveform feature extraction and matrix analysis is proposed. The synchronous waveform features at multiple measurement points are extracted from a multi-domain perspective. The analysis object is transformed from the fault feature matrix into the random matrix eigenvalue by introducing random matrix theory. The historical samples are used to determine the cluster center point of the fault section in the eigenvalue space. During online fault location, the fault section can be determined only based on the eigenvalue space distance matching results.

3. A method for rapid identification and location of weak distribution network faults according to claim 1, characterized in that: The fault type identification method based on multi-point time-frequency matrix fusion analysis specifically includes: (1) After a system fault occurs, the recording device at the outlet of the low-voltage side of the main transformer records the fault waveform data; the waveform data of one cycle before and after the fault recorded by the recording device is extracted, including three-phase voltage and three-phase current; the zero-sequence voltage is calculated by Karrenbauer phase mode transformation of the three-phase voltage; the fault waveform is decomposed by ICEEMDAN to obtain k IMF components; (2) The time-frequency matrix of each waveform is constructed by using Hilbert transform and bandpass filtering. The instantaneous frequency and amplitude of each IMF component are obtained by Hilbert transform. The bandpass filtering algorithm divides the instantaneous frequency into equal intervals, extracts the waveforms in the same frequency range, and obtains the reconstructed waveforms in the specified frequency range. The reconstructed waveform data are arranged in rows to form the corresponding time-frequency matrix. The number of columns in the time-frequency matrix is ​​equal to the number of sampling points in one cycle, and the number of rows is equal to the number of frequency bands. (3) Perform singular value decomposition on the waveform time-frequency matrix, retain the first k-order singular values ​​as the main singular values ​​according to the requirement that the calculated cumulative contribution rate is greater than 90%; calculate the singular spectrum kurtosis K composed of the first k singular values λ 、Singular spectral entropy E λ , singular spectrum mean I λ And the singular spectrum peak factor P λ ; Considering that the singular value dimensions corresponding to different waveforms are different, the singular value statistical indicators are normalized according to the original waveform type to obtain the singular spectrum statistical indicators in the range of 0 to 1; (4) The singular spectrum statistical indicators corresponding to each waveform are used as the extracted fault feature vector and input into the multi-classification SVM model to classify the fault type.

4. A method for rapid identification and location of weak distribution network faults according to claim 3, characterized in that: In step (2), considering that the instantaneous frequency of the fault signal is mainly concentrated in the range of 0 to 3 kHz, the bandwidth is set to 300 Hz and the number of frequency bands is set to 10.

5. A method for rapid identification and location of weak distribution network faults according to claim 3, characterized in that: In step (4), the classification SVM model adopts the radial basis RBF kernel function, and the penalty factor c and the kernel parameter γ are determined by optimizing the GS method, the GA method and the PSO method.

6. A method for rapid identification and location of weak distribution network faults according to claim 1, characterized in that: The fault section location method based on synchronous waveform feature extraction and matrix analysis specifically includes: (1) Extracting fault feature Z m =[z m,1 ,z m,2 ,···,z m,k ], based on the historical fault recording data, calculate the fault characteristics in the time domain, frequency domain and time-frequency domain; use the fusion of ReliefF algorithm and RF algorithm to comprehensively weight the fault characteristics; set the weight threshold w set , retain the comprehensive weight greater than w set k-dimensional fault characteristics; (2) Determine the location of the cluster center point corresponding to each segment in the Rspace space; construct a topological matrix for all historical fault samples one by one, and extract eigenvalues ​​using a random matrix algorithm; take the first three-dimensional eigenvalues ​​as the position coordinates of each fault sample in the Rspace space; use the FCM algorithm to search for cluster center points and determine the location of the center point corresponding to each segment in the Rspace space; (3) When a single-phase grounding fault is detected in the system, k fault features of the transient zero-mode current in the time domain, frequency domain, and time-frequency domain at each measuring point are calculated on-site, packaged and uploaded to the regional master station; the fault features at each measuring point are normalized in the regional master station and a topological matrix is ​​generated; the topological matrix is ​​transformed to generate a random matrix and the eigenvalues ​​of the random matrix are obtained; (4) The first three-dimensional eigenvalues ​​of the random matrix are extracted as its position in the Rspace space, and the distance is matched with the center point of each fault segment in the Rspace space. The segment corresponding to the closest center point is the fault segment.

7. A method for rapid identification and location of weak distribution network faults according to claim 6, characterized in that: The process of feature screening using the ReliefF algorithm is as follows: 1) Select any sample X1 from the historical fault samples, the corresponding label of X1 is recorded as R, X1 contains all fault feature quantities Fp (p = 1, 2, ···, h, Fp∈Ω) when the fault occurs in segment R, and h is the total number of fault feature quantities; select k nearest neighbor samples Hj and Mj (j = 1, 2, …, k) from the remaining historical fault samples, where Hj is of the same type as X1, and Mj is of a different type from X1; 2) Calculate and update the weight w of each feature Fp one by one according to the following formula B (F P ): In the formula, Indicates that F after the mth iteration calculation p The weight calculation result is set to 0; P(C) represents the proportion of class C samples in the fault feature set Ω; d(F p ,X1,H j ) represents the sample X1 and H j In feature F p The distance on the p ,X1,M j ) represents the sample X1 and M j In feature F p The distance on.

8. A method for rapid identification and location of weak distribution network faults according to claim 7, characterized in that: When using the ReliefF algorithm, the number of iterations m is 20 and the number of nearest neighbor samples K is 10.

9. A method for rapid identification and location of weak distribution network faults according to claim 7 or 8, characterized in that: The process of feature screening using the RF algorithm is as follows: 1) Using the bootstrap resampling method, q training sample sets and q out-of-bag OOB data sets are generated from the historical fault feature set Ω; 2) Using the k-th training sample set, train the decision tree Treek and calculate the classification accuracy Tk of the corresponding k-th out-of-bag OOB data set; 3) Apply disturbances to the fault feature quantities Fp (p=1, 2, ···, h) in the out-of-bag OOB data set in turn, and recalculate the classification accuracy Tk,p; 4) Let k = k + 1, repeat steps 2) and 3) to calculate Tk and Tk,p; 5) Calculate the importance or weight of the fault feature quantity Fp. The formula is as follows: In order to avoid deviation when using a single method to screen fault features, different preference coefficients are set for the two methods with reference to the weighting results of the ReliefF algorithm and the RF algorithm, and the independent weight results are combined to form a comprehensive weight; the comprehensive weight is a linear combination of the weight calculation results of the two methods, satisfying: In the formula, w i is the comprehensive weight of the i-th fault feature; β represents the weight preference coefficient of the RF algorithm, and (1-β) is the weight preference coefficient of the ReliefF algorithm; is the weight of the i-th fault feature provided by the RF algorithm, is the weight of the i-th fault feature provided by the ReliefF algorithm; In order to determine the value of β, the objective function is established with the goal of minimizing the sum of squares of the deviations between the comprehensive weight and the ReliefF weight and the RF weight: If O can take the minimum value of 0, then the derivative of the above formula can be obtained to obtain β = 0.5, and the above formula is correspondingly transformed into: Finally, the comprehensive weight set of each fault feature is calculated as W = [w1,w2,···,w n ] T , select the features with larger weights as the objects of subsequent analysis.

10. A computer-readable storage medium having stored thereon computer program instructions that can be executed by a processor, and when the processor executes the computer program instructions, the method steps according to any one of claims 1 to 9 can be implemented.

Images