A multi-port distribution network fault location method based on matrix and deep learning

Abstract

The present invention discloses a multi-port distribution network fault location method based on distance matrix and deep learning, comprising the following steps: installing a traveling wave recorder to collect fault traveling waves, using successive variational mode decomposition (SVMD) and Teager energy operator (TEO) to extract features of the traveling waves, and obtaining the traveling wave velocity and the traveling wave arrival time; deriving a fault end identification method from the perspective of frequency dependence on the traveling wave velocity, and defining an intrinsic distance matrix (IDM), a fault distance matrix (FDM) and a fault section identification matrix (FSIM); analyzing the element characteristics in the FSIM, and determining the estimated position according to the average value of the fault distance; constructing a fault section identification matrix database, and setting the fault distance average value as a label for the fault section identification matrix; building a convolutional neural network CNN model in Pytorch, and inputting FSIM data into the model for training and testing. Through the distribution network fault location method, fault location can be achieved quickly and accurately, and it is feasible and reliable in various fault conditions.

Description

Technical Field

[0001] The present invention relates to the field of electric power automation, and in particular to a multi-port distribution network fault location method based on matrix and deep learning. Background Art

[0002] Distribution lines are an integral part of modern power systems. The widespread access of distributed power sources has changed the power flow direction and fault current distribution of the distribution network, affecting the effectiveness of traditional fault location methods. Effective power system protection methods should be able to detect all faults as quickly as possible. Accurate and timely fault diagnosis is of great significance to the safe operation and power supply reliability of the distribution system.

[0003] The accuracy of the traveling wave (TW) method is affected by the arrival time and velocity of the traveling wave. First, the method of determining the arrival time of the traveling wave has been widely studied. Among the existing traveling wave positioning methods, the main method includes the method of obtaining the arrival time of the traveling wave through discrete wavelet transform (DWT). The disadvantage of this method is that it is difficult to select the mother wavelet and the decomposition level. As for the method of obtaining the arrival time of the traveling wave through the continuous application of empirical mode decomposition (EMD) and Hilbert transform, EMD avoids the difficulty of selecting DWT, but is easily affected by the endpoint effect, which will affect the identification of the wavefront.

[0004] The fault traveling wave is decomposed into K intrinsic mode functions (IMFs) by variational mode decomposition (VMD). Although VMD is not affected by the EMD endpoint effect, it is necessary to know the modal decomposition number K in advance. Overestimation or underestimation of K will have a negative impact on the results.

[0005] Leading to poor adaptability. Summary of the invention

[0006] The purpose of the present invention is to provide a multi-port distribution network fault location method based on distance matrix and deep learning, which can realize rapid fault section identification and precise point location.

[0007] In order to achieve the above purpose, the technical method adopted by the present invention is as follows:

[0008] A multi-port distribution network fault location method based on distance matrix and deep learning, the steps comprising:

[0009] S1. Install a traveling wave recorder at the distribution network terminal to collect fault traveling waves, use successive variational mode decomposition (SVMD) and Teager energy operator (TEO) to extract the characteristics of the traveling waves, and obtain the traveling wave velocity and arrival time;

[0010] S2. From the perspective of frequency-dependent traveling wave velocity, the fault end identification method is derived, and the intrinsic distance matrix (IDM), fault distance matrix (FDM) and fault section identification matrix (FSIM) are defined;

[0011] S3, analyzing the characteristics of the elements in FSIM, and determining the estimated location based on the average value of the fault distance calculated based on the path including the selected reference port and the real fault point;

[0012] S4, construct the distribution network FSIM database, and set the fault distance average value as the label for the fault section identification matrix;

[0013] S5. Build a convolutional neural network (CNN) model in Pytorch, input the FSIM data of the database into the model for training and testing.

[0014] Preferably, in step S1, after collecting the traveling waves, the fault traveling waves collected by each distribution network terminal are converted into line mode components and ground mode components using Karen Bell transformation, wherein the formula of Karen Bell transformation is as follows:

[0015]

[0016] Among them U A , U B and U C They are three-phase voltage or three-phase current respectively, U0 is the ground mode component, also called 0 mode component, which propagates between the conductor and the earth; U α , U β There are two independent line-mode voltage components or current components that propagate between the three-phase conductors.

[0017] Preferably, the specific steps of the successive variational mode decomposition (SVMD) in step S1 are as follows:

[0018] Initialize the modal decomposition number to 0;

[0019] Set the modal decomposition number to increase by one, initialize the model, center frequency, Lagrange multiplier and matrix dimension to 0, and use the value of the matrix dimension plus one as the number of the model, center frequency and Lagrange multiplier;

[0020] Set the matrix dimension to increase by one and update the model, center frequency, and Lagrange multiplier by one;

[0021] Repeat the steps, and when the stopping criteria for updating the parameters are met, output the current mode and its corresponding center frequency;

[0022] When the algorithm end criterion is met, the iterative process terminates and all modes and their corresponding center frequencies are output.

[0023] Preferably, in step S2, the intrinsic distance matrix IDM calculation process is as follows:

[0024]

[0025] Among them, the row elements of the matrix L represent the distance from each distribution network terminal to its adjacent T node, the diagonal elements of the matrix are set to 0, the Mi in the row elements represent the reference port, and the Mi in the column elements represent the remote port.

[0026] Preferably, the elements of the fault distance matrix FDM-d are calculated by the two-end location formula, and FDM-d is defined as follows:

[0027]

[0028] Among them, d MiMj (i=1,2,3,4,j=1,2,3,4) represents the value calculated by the two-end positioning formula, and the diagonal element d of the matrix MiMj Defined as 0.

[0029] Preferably, in step S2, a fault segment identification matrix FSIM-δ is established by calculating the difference between the fault distance matrix FDM and the intrinsic distance matrix IDM, wherein δ=dL.

[0030] Preferably, in step S3, the fault occurs on the branch, when port M2 is selected as the reference port, the path M1-M2 contains the real fault point, and the value d calculated by the positioning formula at both ends is M2M1 Equal to L M2T1 +L fT1 , the path between M2 and the other ports does not contain a real fault point, T1 is a pseudo fault point, and the value d calculated by the positioning formula at both ends M2M3 and d M2M4 Equal to L M2T1 , so the final FSIM-δ is:

[0031]

[0032] Among them, M1 is selected as the reference port to estimate the fault distance. For the fault on the branch containing port Mi, the characteristic of the corresponding element in the matrix δ is expressed as:

[0033] δ(i,:) max ≤0

[0034] δ(:,i) min ≥0;

[0035] Among them, δ(i,:) max represents the maximum value of the elements in the i-th row, δ(:,i) min Represents the minimum value of the elements in the i-th column.

[0036] Preferably, in step S3, when a fault occurs between adjacent T nodes, the corresponding FSIM matrix obtained by analyzing the fault occurring between adjacent T nodes is:

[0037]

[0038] The fault segments between adjacent T nodes, that is, the fault segments adjacent to ports Mi and Mj respectively, the characteristics of the corresponding elements in the matrix δ are expressed as:

[0039] δ(i,:)·δ(j,:)=0;

[0040] Among them, δ(i,:) and δ(j,:) represent the elements in the i-th and j-th rows respectively.

[0041] Preferably, in step S3, when the fault happens to occur at the T node, the FSIM is obtained as:

[0042]

[0043] Among them, all row element values ​​of M3 / M4 are equal to 0. For the fault on the T node adjacent to port Mi, the characteristic of the corresponding element in the matrix δ is expressed as:

[0044] δ(i,:)=0 n ;

[0045] Among them, δ(i,:) represents the element of the i-th row, which represents the zero row vector of n-dimension;

[0046] Preferably, in step S5, the convolutional neural network (CNN) model includes multiple convolutional layers, each of which is followed by an activation layer and a pooling layer. The core function of the three-dimensional convolutional layer is to slide the convolution kernel on the data cube, calculate the convolution of the kernel and the data, and thus extract features.

[0047] Therefore, the present invention adopts a multi-port distribution network fault location method based on distance matrix and deep learning in the above scheme, which has the following beneficial effects:

[0048] (1) The present invention uses successive variational mode decomposition (SVMD) and Teager energy operator (TEO) to extract features from the recorded traveling waves, thus overcoming the difficulties of wavelet transform in selecting mother wavelets and decomposition levels;

[0049] (2) Compared with EMD and VMD, the present invention is less susceptible to the impact of endpoint effects and overcomes the problem of poor adaptability caused by overestimating or underestimating the number of modal decompositions;

[0050] (3) The deep learning method can quickly and accurately locate faults, which is more effective than the computational method.

[0051] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A flow chart of a multi-port distribution network fault location method based on distance matrix and deep learning;

[0053] Figure 2 This is the structural diagram of the five-terminal distribution network. DETAILED DESCRIPTION

[0054] In order to facilitate the understanding of the present invention, the present invention will be described more comprehensively below with reference to the relevant drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and cannot be used to limit the protection scope of the present invention.

[0055] Example

[0056] like Figure 1 As shown, the present invention provides a multi-port distribution network fault location method based on distance matrix and deep learning, and the specific working process is as follows:

[0057] S1: Install a traveling wave recorder to collect fault voltage traveling wave data:

[0058] like Figure 2 As shown, each terminal of the five-terminal distribution network is equipped with a traveling wave recorder. The Karen Bell transformation is used to convert the fault voltage traveling wave collected by each terminal into line mode components and ground mode components, and the line mode components in the high frequency band are selected for positioning. The Karen Bell transformation formula is as follows:

[0059]

[0060] Among them U A , U B and U C They are three-phase voltage (or three-phase current). U0 is the ground mode component, also called the 0-mode component, which propagates between the conductor and the earth. α , U β There are two independent line-mode voltage or current components that propagate between the three-phase conductors.

[0061] S2: The traveling wave velocity and arrival time are extracted using successive variational mode decomposition (SVMD) and Teager energy operator (TEO).

[0062] The steps of Successive Variational Mode Decomposition (SVMD) are as follows:

[0063] 1. Initialize the K value to 0, where K is the modal decomposition number;

[0064] 2. Set K = K + 1 and initialize the parameters and n is 0, where Represents the model, represents the center frequency, represents the Lagrange multiplier;

[0065] 3.n=n+1, update parameters

[0066] 4. Repeat the steps, and when the stopping criteria for updating the parameters are met, output the Kth mode and its corresponding center frequency;

[0067] 5. Repeat steps 2 to 4. When the end criteria of the algorithm are met, the iterative process terminates and all modes and their corresponding center frequencies are output.

[0068] TEO (Teager Energy Operator) is a nonlinear signal operator that can efficiently evaluate the energy changes of transient signals. It is very suitable for detecting the arrival time of traveling waves. For a continuous signal x(t), TEO is defined as follows:

[0069]

[0070] Where ψ(x) represents the instantaneous energy of x, and are the first and second order time derivatives of x(t), respectively.

[0071] S3: Calculate the intrinsic distance matrix (IDM), fault distance matrix (FDM) and fault section identification matrix (FSIM). Given the topological structure of the power transmission network, the distance between the terminal and its adjacent T-node is determined, so the intrinsic distance matrix (IDM) can be obtained:

[0072]

[0073] The row elements of the matrix L represent the distance from each terminal to its adjacent T-node, and the diagonal elements of the matrix are set to 0.

[0074] The elements of the fault distance matrix (FDM) d are calculated using the two-end location formula, and then FDM-d is defined as follows:

[0075]

[0076] Among them, d MiMj (i=1,2,3,4,j=1,2,3,4) represents the value calculated by the two-end positioning formula, and the diagonal element d of the matrix M2M1Defined as 0. Finally, by calculating the difference between the fault distance matrix (FDM) d and the intrinsic distance matrix (IDM) L, the fault segment identification matrix (FSIM) δ can be established.

[0077] δ = dL;

[0078] In the above matrix, Mi in the row element represents the reference port, and Mi in the column element represents the remote port.

[0079] S4: According to different fault locations, analyze the characteristics of the elements in FSIM and calculate the estimated fault location:

[0080] Case 1: The failure occurs on a branch.

[0081] Assume that the fault occurs in Figure 1 At position 1 of branch M1-T1, when port M2 is selected as the reference port, path M1-M2 contains the real fault point. The value d calculated by the two-end positioning formula M2M1 Theoretically equal to L M2T1 +L fT1 , path M2 and the rest of the ports do not contain a real fault point. Therefore, for them, T1 is a pseudo fault point, and the value d calculated by the two-end positioning formula is M2M3 and d M2M4 Theoretically equal to L M2T1 , so the final FSIM-δ is:

[0082]

[0083] It can be seen from the element values ​​in the matrix that the row elements of port M1 are negative or equal to 0, and the column elements are positive or equal to 0, while other ports on the non-faulty branches do not meet this rule. Therefore, M1 can be selected as the reference port to estimate the fault distance. In general, for the fault on the branch containing port Mi, the characteristics of the corresponding elements in the matrix δ can be expressed as:

[0084] δ(i,:) max ≤0

[0085] δ(:,i) min ≥0

[0086] Among them, δ(i,:) max represents the maximum value of the elements in the i-th row, δ(:,i) min It represents the minimum value of the elements in the i-th column. This feature can be used to identify the fault section and estimate the fault location.

[0087] Case 2: When a fault occurs between adjacent T nodes:

[0088] Similarly, if the fault occurs between adjacent T nodes, the corresponding FSIM matrix can be obtained as follows:

[0089]

[0090] By analyzing the elements in the δ matrix, it can be concluded that for the fault segments between adjacent T nodes, that is, the fault segments adjacent to ports Mi and Mj respectively, the characteristics of the corresponding elements in the matrix δ can be expressed as:

[0091] δ(i,:)·δ(j,:)=0

[0092] Among them, δ(i,:) and δ(j,:) represent the elements of the i-th row and the j-th row respectively. When Mi or Mj is selected as the reference port, the dot product of the row element and the column element is 0, which helps to estimate the fault distance.

[0093] Case 3: When the fault happens at node T, the FSIM is:

[0094]

[0095] By analyzing the elements in the δ matrix, it can be concluded that all row element values ​​of M3 / M4 are equal to 0, while the ports of other adjacent non-faulty nodes do not meet this rule. Therefore, M3 / M4 can be selected as the reference port to estimate the fault distance. In general, for a fault on a T node adjacent to port Mi, the characteristics of the corresponding elements in the matrix δ can be expressed as:

[0096] δ(i,:)=0 n

[0097] Among them, δ(i,:) represents the element of the i-th row, which represents the zero row vector of n-dimensional.

[0098] Assuming that the fault occurs 10 km away from point M1, that is, on branch M1-T1, the FSIM matrix obtained according to the distribution network structure and related formulas is:

[0099]

[0100] From the δ element value, it can be seen that the row element of port M1 is not greater than 0 and the column element is not less than 0, which is consistent with the previous analysis. Therefore, the result of fault segment identification is that the fault occurs in branch M1-T1. M1 is selected as the reference port, and the estimated position is 10.051 kilometers with an absolute error of +51 meters. It can be seen that this method can not only locate the fault section, but also accurately locate the fault distance.

[0101] S5: Build a convolutional neural network (CNN) model in Pytorch, construct a distribution network FSIM database, input data into the model for training and testing.

[0102] Construct a convolutional neural network (CNN) model to train the FSIM data generated by the above steps. This network model includes multiple convolutional layers, each of which is followed by an activation layer (usually using the ReLU activation function) and a pooling layer (using maximum pooling or average pooling). The core function of the three-dimensional convolutional layer is to slide the convolution kernel on the data cube, calculate the convolution of the kernel and the data, and thus extract features.

[0103] In order to improve the generalization ability of the network and prevent overfitting, batch normalization layer and dropout layer are introduced into the model. Batch normalization layer is used to adjust the mean and variance of each layer output to ensure the stability of network training. Dropout layer randomly discards a certain proportion of neurons during training to reduce the overfitting of the model to the training data.

[0104] The fault section identification matrix is ​​labeled with the average fault distance, and the data set is divided into a training set and a test set, which is designed for model training. All preprocessed feature cube data are input into the network and trained multiple times. Each iteration calculates the loss function and updates the network weights. In order to evaluate the performance of the model, we test it on an independent validation set and verify the effectiveness and accuracy of the model by calculating the loss rate and the final positioning accuracy.

[0105] In this embodiment, successive variational mode decomposition (SVMD) and Teager energy operator (TEO) are used to extract the traveling wave velocity and the traveling wave arrival time, which overcomes the difficulties of wavelet transform in selecting the mother wavelet and the decomposition level. Compared with EMD and VMD, it is less susceptible to the endpoint effect and overcomes the problem of poor adaptability caused by overestimation or underestimation of the K value.

[0106] A convolutional neural network (CNN) model was built in Pytorch, a distribution network FSIM database was constructed, and data was input into the model for training and testing. The deep learning method can be used to quickly and accurately locate faults, which is more effective than the calculation method.

[0107] Therefore, the present invention adopts the above-mentioned multi-port distribution network fault location method based on distance matrix and deep learning, which can overcome the difficulties of wavelet transform in selecting mother wavelet and decomposition level. Compared with EMD and VMD, it is not easily affected by endpoint effect and overcomes the problem of poor adaptability caused by overestimation or underestimation of K value. In addition, by adopting the deep learning method, only a large amount of historical data is needed to train the model, and fault location can be realized quickly and accurately, which is more effective than the calculation method.

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical method to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A multi-port distribution network fault location method based on distance matrix and deep learning, characterized in that: The following steps are involved: S1. Install a traveling wave recorder at the distribution network terminal to collect fault traveling waves, extract the characteristics of the traveling waves using successive variational mode decomposition (SVMD) and Teager energy operator (TEO), and obtain the traveling wave velocity and arrival time. S2. From the perspective of frequency-dependent traveling wave velocity, the fault end identification method is derived, and the intrinsic distance matrix IDM, fault distance matrix FDM and fault section identification matrix FSIM are defined; S3, analyzing the characteristics of the elements in FSIM, and determining the estimated location based on the average value of the fault distance calculated based on the path including the selected reference port and the real fault point; S4, construct the distribution network FSIM database, and set the fault distance average value as the label for the fault section identification matrix; S5. Build a convolutional neural network (CNN) model in Pytorch, input the FSIM data of the database into the model for training and testing.

2. A multi-port distribution network fault location method based on distance matrix and deep learning according to claim 1, characterized in that: In step S1, after collecting the traveling waves, the fault traveling waves collected by each distribution network terminal are converted into line mode components and ground mode components using the Karen Bell transformation, where the formula of the Karen Bell transformation is as follows: Among them U A , U B and U C They are three-phase voltage or three-phase current respectively, U0 is the ground mode component, also called 0 mode component, which propagates between the conductor and the earth; U α , U β There are two independent line-mode voltage components or current components that propagate between the three-phase conductors.

3. The multi-port distribution network fault location method based on distance matrix and deep learning according to claim 1 is characterized in that: The specific steps of successive variational mode decomposition (SVMD) in step S1 are as follows: Initialize the number of modal decompositions to 0; Set the modal decomposition number to increase by one, initialize the model, center frequency, Lagrange multiplier and matrix dimension to 0, and use the value of the matrix dimension plus one as the number of the model, center frequency and Lagrange multiplier; Set the matrix dimension to increase by one and update the model, center frequency, and Lagrange multiplier by one; Repeat the steps, and when the stopping criteria for updating the parameters are met, output the current mode and its corresponding center frequency; When the algorithm end criterion is met, the iterative process terminates and all modes and their corresponding center frequencies are output.

4. The multi-port distribution network fault location method based on distance matrix and deep learning according to claim 1 is characterized in that: In step S2, the intrinsic distance matrix IDM calculation process is as follows: Among them, the row elements of the matrix L represent the distance from each distribution network terminal to its adjacent T node, the diagonal elements of the matrix are set to 0, the Mi in the row elements represent the reference port, and the Mi in the column elements represent the remote port.

5. The multi-port distribution network fault location method based on distance matrix and deep learning according to claim 1 is characterized in that: The elements of the fault distance matrix FDM-d are calculated using the two-end location formula. FDM-d is defined as follows: Among them, d MiMj (i=1,2,3,4,j=1,2,3,4) represents the value calculated by the two-end positioning formula, and the diagonal element d of the matrix MiMj Defined as 0.

6. The multi-port distribution network fault location method based on distance matrix and deep learning according to claim 1 is characterized in that: By calculating the difference between the fault distance matrix FDM and the intrinsic distance matrix IDM, a fault segment identification matrix FSIM-δ is established, where δ=dL.

7. The multi-port distribution network fault location method based on distance matrix and deep learning according to claim 1 is characterized in that: In step S3, the fault occurs on the branch. When port M2 is selected as the reference port, the path M1-M2 contains the real fault point. The value d calculated by the positioning formula at both ends is M2M1 Equal to L M2T1 +L fT1 , the path between M2 and the other ports does not contain a real fault point, T1 is a pseudo fault point, and the value d calculated by the positioning formula at both ends M2M3 and d M2M4 Equal to L M2T1 , so the final FSIM-δ is: Among them, M1 is selected as the reference port to estimate the fault distance. For the fault on the branch containing port Mi, the characteristic of the corresponding element in the matrix δ is expressed as: δ(i,:) max ≤0 d(:,i) min ≥0; Among them, δ(i,:) max represents the maximum value of the elements in the i-th row, δ(:,i) min Represents the minimum value of the elements in the i-th column.

8. The multi-port distribution network fault location method based on distance matrix and deep learning according to claim 1 is characterized in that: In step S3, when a fault occurs between adjacent T nodes, the corresponding FSIM matrix obtained by analyzing the fault occurring between adjacent T nodes is: The fault segments between adjacent T nodes, that is, the fault segments adjacent to ports Mi and Mj respectively, the characteristics of the corresponding elements in the matrix δ are expressed as: δ(i,:)·δ(j,:)=0; Among them, δ(i,:) and δ(j,:) represent the elements in the i-th and j-th rows respectively.

9. The multi-port distribution network fault location method based on distance matrix and deep learning according to claim 1, characterized in that: In step S3, when the fault happens to occur at node T, the FSIM is obtained as: Among them, M3 / M4 is selected as the reference port to estimate the fault distance. For the fault on the T node adjacent to port Mi, the characteristic of the corresponding element in the matrix δ is expressed as: δ(i,:)=0 n ; Among them, δ(i,:) represents the element of the i-th row, which represents the zero row vector of n-dimensional.

10. The multi-port distribution network fault location method based on distance matrix and deep learning according to claim 1, characterized in that: In step S5, the convolutional neural network (CNN) model includes multiple convolutional layers, each of which is followed by an activation layer and a pooling layer. The core function of the three-dimensional convolutional layer is to slide the convolution kernel on the data cube, calculate the convolution of the kernel and the data, and thus extract features.

Images