A method and device for locating open circuit faults in a dc filter capacitor

Abstract

The application discloses a DC filter capacitor open-circuit fault positioning method and device, wherein the method comprises the following steps: S1, collecting original current data of a DC filter capacitor in a normal state and an open-circuit fault state; S2, extracting an effective value from the original current data, and performing normalization and principal component analysis processing to extract a characteristic vector capable of reflecting the operating state of the DC filter capacitor; and S3, inputting a part of the characteristic vector as a test sample into a DC filter capacitor open-circuit fault positioning model for classification, thereby realizing open-circuit fault positioning of the DC filter capacitor, wherein the DC filter capacitor open-circuit fault positioning model is obtained by inputting another part of the characteristic vector as a training sample into a probabilistic neural network for training.

Description

Technical Field

[0001] This invention relates to the field of power system technology, specifically to a method and apparatus for locating open-circuit faults in DC filter capacitors. Background Technology

[0002] DC filters can filter out characteristic harmonics generated on the DC side of the power grid's commutator converter, reducing their interference with nearby communication systems, and mitigating adverse effects such as overheating of DC equipment and reduced power quality. When some capacitors in a DC filter experience open-circuit faults, the current flowing through the normally connected capacitors in parallel will increase, thus reducing their lifespan. In severe cases, this can lead to an avalanche effect, damaging even more capacitors and jeopardizing the safe and stable operation of the DC filter and even the High Voltage Direct Current (HVDC) transmission system. Therefore, timely capacitor maintenance is essential. In HVDC projects, faults are typically diagnosed by manually measuring the capacitance of each capacitor unit. This method is inefficient, and incomplete capacitor discharge can threaten the safety of maintenance personnel.

[0003] Currently, there is very little research on the location of open-circuit faults in DC filter capacitors. Some scholars have used the changes in amplitude and direction of unbalanced branch current, DC filter start-up current, and the two high-voltage end arm branch currents before and after the fault to determine the specific faulty arm of the high-voltage capacitor. Other scholars have used the branch currents of the two low-voltage end arms and the unbalanced branch current to calculate two relative current values, and used the step change of the relative current values ​​to locate the fault to the specific arm of the high-voltage capacitor. The above methods have achieved good results in locating open-circuit faults in a single arm of the high-voltage capacitor. However, when the same open-circuit fault occurs simultaneously in both arms of the high-voltage capacitor, the above methods do not consider locating the fault to both faulty arms, nor do they consider the location of open-circuit faults in the low-voltage capacitor. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method and device for locating open circuit faults in DC filter capacitors, so as to improve the efficiency of open circuit fault repair and ensure the personal safety of repair personnel.

[0005] To solve the above technical problems, the present invention provides a method for locating open-circuit faults in DC filter capacitors, comprising:

[0006] Step S1: Collect raw current data of the DC filter capacitor under normal and open-circuit fault conditions;

[0007] Step S2: Extract effective values ​​from the original current data, and perform normalization and principal component analysis to extract feature vectors that reflect the operating state of the DC filter capacitor.

[0008] Step S3: A portion of the feature vector is used as a test sample and input into the DC filter capacitor open-circuit fault location model for classification, thereby realizing the open-circuit fault location of the DC filter capacitor. The DC filter capacitor open-circuit fault location model is obtained by training the other portion of the feature vector as a training sample into a probabilistic neural network.

[0009] Furthermore, the open-circuit fault includes open-circuit faults occurring in all four arms of the high-voltage capacitor, open-circuit faults occurring simultaneously in two arms of the high-voltage capacitor, and open-circuit faults occurring in the low-voltage capacitor.

[0010] Furthermore, the raw current data acquired in step S1 includes: the current i measured by the current transformer at the first end of the DC filter. h (t), the bridge arm branch currents i1(t) and i2(t) measured by the current transformers on the left and right sides of the top of the high-voltage capacitor of the DC filter, and the current i measured by the current transformer on the unbalanced branch of the high-voltage capacitor of the DC filter. un (t), the current i measured by the current transformer at the tail end of the DC filter. w (t), the current i measured by the current transformer on the low-voltage reactor branch. l2 (t), the current i measured by the current transformer on the resistor branch. r (t), where t is the sampling time.

[0011] Furthermore, given i1(t), i2(t), and i un (t) The bridge arm branch currents i3(t) and i4(t) on the left and right sides of the bottom of the high-voltage capacitor are calculated as follows: i3(t) = i1(t) - i un (t), i4(t)=i2(t)+i un (t); by i w (t), i l2 (t) and i r (t) Calculate the branch current i of the low-voltage capacitor. c2 (t): i c2 (t)=i w (t)-i l2 (t)-i r (t).

[0012] Further, in step S2, principal component analysis is used to extract the feature vector from the normalized data, specifically including:

[0013] Step S21: Perform a standardization transformation on the normalized 10-dimensional data to obtain the standardized matrix Z;

[0014] Step S22, construct the correlation coefficient matrix M, where m is the number of samples in each dimension of the data:

[0015]

[0016] Step S23: Calculate the eigenvalues ​​λ of the correlation coefficient matrix M. j , where λ1≥λ2≥…≥λ 10 The corresponding feature vector is b j =(b 1j ,b 2j ,…,b 10j ) T , where j = 1, 2, ..., 10;

[0017] Step S24, solve for principal components

[0018] U j =Zb j

[0019] Among them, U j It is the j-th principal component;

[0020] Step S25: Calculate the cumulative variance contribution rate η of the first l principal components. Σ (l) Select the required principal components and extract the feature vectors:

[0021]

[0022] Furthermore, the probabilistic neural network in step S3 consists of an input layer, a hidden layer, a summation layer, and an output layer. During training the probabilistic neural network, the training samples are processed according to the following steps:

[0023] First, the input layer receives the training sample X and simultaneously feeds it to the hidden layer, which has c neurons, representing the dimension of the input training sample X.

[0024] Then, the hidden layer, which is a radial base layer with p neurons, calculates the matching degree between the training sample X and each DC filter capacitor classification category according to the following formula:

[0025]

[0026] Among them, W i The connection weights between the input layer and the hidden layer are set to various training samples, and δ is a smoothing factor.

[0027] Next, the summation layer sums the probabilities of neurons belonging to the same classification category in the hidden layer to obtain the probability density function estimate of the fault type, as shown in the following formula. This layer has k neurons, representing the total number of categories of the DC filter capacitor's operating state:

[0028]

[0029] Where p is the dimension of the training samples, m1 is the number of training samples belonging to category 1, and X 1i This is the i-th training sample in category 1;

[0030] The output layer consists of k competing neurons. In the output of the summation layer, this layer finds the neuron with the highest probability. The output of this neuron is 1, which means that the measured data belongs to the DC filter capacitor operation state type corresponding to this neuron. The outputs of the other neurons are 0.

[0031] Finally, the output layer's result is multiplied by the vector [1,2,3,…,k]', which is multiplied by the classification category value corresponding to each neuron, thereby converting the output layer's result into classification category values:

[0032] y = max(f1,f2,…,f k ).

[0033] The present invention also provides a DC filter capacitor open-circuit fault location device, comprising:

[0034] The acquisition module is used to acquire raw current data of the DC filter capacitor under normal and open-circuit fault conditions.

[0035] The feature vector extraction module is used to extract effective values ​​from the raw current data, and perform normalization and principal component analysis to extract feature vectors that can reflect the operating state of the DC filter capacitor.

[0036] The fault location module is used to input a portion of the feature vector as a test sample into the DC filter capacitor open-circuit fault location model for classification, thereby realizing the open-circuit fault location of the DC filter capacitor. The DC filter capacitor open-circuit fault location model is obtained by inputting another portion of the feature vector as a training sample into a probabilistic neural network for training.

[0037] Furthermore, the open-circuit fault includes open-circuit faults occurring in all four arms of the high-voltage capacitor, the same open-circuit fault occurring simultaneously in two arms of the high-voltage capacitor, and an open-circuit fault occurring in the low-voltage capacitor; the raw current data acquired by the acquisition module includes: the current i measured by the current transformer at the first end of the DC filter. h(t), the bridge arm branch currents i1(t) and i2(t) measured by the current transformers on the left and right sides of the top of the high-voltage capacitor of the DC filter, and the current i measured by the current transformer on the unbalanced branch of the high-voltage capacitor of the DC filter. un (t), the current i measured by the current transformer at the tail end of the DC filter. w (t), the current i measured by the current transformer on the low-voltage reactor branch. l2 (t), the current i measured by the current transformer on the resistor branch. r (t), where t is the sampling time.

[0038] Furthermore, the feature vector extraction module extracts the feature vectors from the normalized data using principal component analysis, specifically including:

[0039] The normalized 10-dimensional data is subjected to a standardization transformation to obtain the standardized matrix Z;

[0040] Construct a correlation coefficient matrix M, where m is the number of samples in each dimension of the data:

[0041]

[0042] Calculate the eigenvalues ​​λ of the correlation coefficient matrix M j , where λ1≥λ2≥…≥λ 10 The corresponding feature vector is b j =(b 1j ,b 2j ,…,b 10j ) T , where j = 1, 2, ..., 10;

[0043] Solve for principal components: U j =Zb j , among which, U j It is the j-th principal component;

[0044] Calculate the cumulative variance contribution rate η of the first l principal components. Σ (l) Select the required principal components and extract the feature vectors:

[0045]

[0046] Furthermore, the probabilistic neural network consists of an input layer, a hidden layer, a summation layer, and an output layer. During training the probabilistic neural network, the training samples are processed according to the following steps:

[0047] First, the input layer receives the training sample X and simultaneously feeds it to the hidden layer, which has c neurons, representing the dimension of the input training sample X.

[0048] Then, the hidden layer, which is a radial base layer with p neurons, calculates the matching degree between the training sample X and each DC filter capacitor classification category according to the following formula:

[0049]

[0050] Among them, W i The connection weights between the input layer and the hidden layer are set to various training samples, and δ is a smoothing factor.

[0051] Next, the summation layer sums the probabilities of neurons belonging to the same classification category in the hidden layer to obtain the probability density function estimate of the fault type, as shown in the following formula. This layer has k neurons, representing the total number of categories of the DC filter capacitor's operating state:

[0052]

[0053] Where p is the dimension of the training samples, m1 is the number of training samples belonging to category 1, and X 1i This is the i-th training sample in category 1;

[0054] The output layer consists of k competing neurons. In the output of the summation layer, this layer finds the neuron with the highest probability. The output of this neuron is 1, which means that the measured data belongs to the DC filter capacitor operation state type corresponding to this neuron. The outputs of the other neurons are 0.

[0055] Finally, the output layer's result is multiplied by the vector [1,2,3,…,k]', which is multiplied by the classification category value corresponding to each neuron, thereby converting the output layer's result into classification category values:

[0056] y = max(f1,f2,…,f k ).

[0057] Implementing this invention has the following beneficial effects: it enables the location of open-circuit faults in DC filter capacitors, improves the efficiency of repairing open-circuit faults in DC filter capacitors, and reduces the possibility of safety accidents for maintenance personnel. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1This is a flowchart illustrating a method for locating open-circuit faults in a DC filter capacitor according to an embodiment of the present invention.

[0060] Figure 2 This is a schematic diagram illustrating the specific process of a DC filter capacitor open-circuit fault location method according to an embodiment of the present invention.

[0061] Figure 3 This is a schematic diagram of a dual-tuned DC filter and an open-circuit fault in an embodiment of the present invention.

[0062] Figure 4 This is a diagram of the probabilistic neural network (PNN) structure in an embodiment of the present invention.

[0063] Figure 5 This is a diagram showing the location results of the test sample in an embodiment of the present invention. Detailed Implementation

[0064] The following description of the embodiments is taken with reference to the accompanying drawings, which illustrate specific embodiments in which the invention can be implemented.

[0065] Therefore, please refer to Figure 1 As shown, Embodiment 1 of the present invention provides a method for locating open-circuit faults in DC filter capacitors, comprising:

[0066] Step S1: Collect raw current data of the DC filter capacitor under normal and open-circuit fault conditions;

[0067] Step S2: Extract effective values ​​from the original current data, and perform normalization and principal component analysis to extract feature vectors that reflect the operating state of the DC filter capacitor.

[0068] Step S3: A portion of the feature vector is used as a test sample and input into the DC filter capacitor open-circuit fault location model for classification, thereby realizing the open-circuit fault location of the DC filter capacitor. The DC filter capacitor open-circuit fault location model is obtained by training the other portion of the feature vector as a training sample into a probabilistic neural network.

[0069] Specifically, please combine Figure 2 and Figure 3 As shown in the figure, this embodiment of the invention takes a dual-tuned DC filter as an example, and its structure and open-circuit fault diagram are as follows. Figure 3As shown, the capacitor adopts an internal fuse structure. In step S1, the raw current data of the DC filter capacitor under normal and open-circuit fault conditions are collected at a sampling frequency f = 5kHz. Open-circuit faults include open-circuit faults in all four arms of the high-voltage capacitor, open-circuit faults in two arms of the high-voltage capacitor simultaneously, and open-circuit faults in the low-voltage capacitor. Therefore, the collected raw current data includes: the current i measured by the current transformer at the beginning of the DC filter. h (t), the bridge arm branch currents i1(t) and i2(t) measured by the current transformers on the left and right sides of the top of the high-voltage capacitor of the DC filter, and the current i measured by the current transformer on the unbalanced branch of the high-voltage capacitor of the DC filter. un (t), the current i measured by the current transformer at the tail end of the DC filter. w (t), the current i measured by the current transformer on the low-voltage reactor branch. l2 (t), the current i measured by the current transformer on the resistor branch. r (t).

[0070] According to Kirchhoff's current law, i1(t), i2(t), and i un The currents i3(t) and i4(t) in the bridge arm branches on the left and right sides of the bottom of the high-voltage capacitor can be calculated using (t):

[0071] i3(t)=i1(t)-i un (t), i4(t)=i2(t)+i un (t)

[0072] by i w (t), i l2 (t) and i r (t) gives the branch current i of the low-voltage capacitor. c2 (t):

[0073] i c2 (t)=i w (t)-i l2 (t)-i r (t)

[0074] Therefore, the collected raw current data has 10 dimensions, where t is the sampling time.

[0075] In step S2, the effective value of the original current is first extracted, and then normalized so that the data is distributed in the interval [0,1]. Finally, Principal Component Analysis (PCA) is used on the data to reduce the data dimensionality and select the necessary principal components to extract the feature vector that can effectively reflect the operating state of the DC filter capacitor.

[0076] Furthermore, the specific process of extracting feature vectors using PCA in step S2 is as follows:

[0077] Step S21: Perform a standardization transformation on the normalized 10-dimensional data to obtain the standardized matrix Z;

[0078] Step S22, construct the correlation coefficient matrix M, where m is the number of samples in each dimension of the data:

[0079]

[0080] Step S23: Calculate the eigenvalues ​​λ of the correlation coefficient matrix M. j , where λ1≥λ2≥…≥λ 10 The corresponding feature vector is b j =(b 1j ,b 2j ,…,b 10j ) T , where j = 1, 2, ..., 10;

[0081] Step S24, solve for principal components

[0082] U j =Zb j

[0083] Among them, U j It is the j-th principal component;

[0084] Step S25: Calculate the cumulative variance contribution rate η of the first l principal components. Σ (l) Select the required principal components and extract the feature vectors; the cumulative variance contribution rate η of the current l principal components. Σ (l) When the percentage is greater than 95%, the first l principal components can reflect most of the information in the original data:

[0085]

[0086] The structure of the probabilistic neural network (PNN) in step S3 is as follows: Figure 4 As shown, the feature vectors used as training samples are input into a probabilistic neural network for training, thereby obtaining a DC filter capacitor open-circuit fault location model. The feature vectors used as test samples are input into the location model for classification, thus realizing the open-circuit fault location of the DC filter capacitor.

[0087] The probabilistic neural network in step S3 consists of four layers, and the specific model is as follows:

[0088] 1) First, the input layer receives the training sample X and simultaneously feeds it to the hidden layer, which has c neurons, representing the dimension of the input training sample X.

[0089] 2) Then, the hidden layer will calculate the matching degree between the training sample X and each DC filter capacitor classification category according to the following formula. This layer is a radial base layer with p neurons, representing the number of training samples X:

[0090]

[0091] Among them, W i The connection weights between the input layer and the hidden layer are set to various training samples, and δ is a smoothing factor.

[0092] 3) Next, the summation layer sums the probabilities of neurons belonging to the same classification category in the hidden layer to obtain the probability density function estimate of the fault type, as shown in the following formula. This layer has k neurons, representing the total number of categories of the DC filter capacitor's operating state:

[0093]

[0094] Where p is the dimension of the training samples, m1 is the number of training samples belonging to category 1, and X 1i This is the i-th training sample in category 1;

[0095] 4) The output layer consists of k competing neurons. In the output of the summation layer, this layer finds the neuron with the highest probability. The output of this neuron is 1, which means that the measured data belongs to the DC filter capacitor operation state type corresponding to this neuron. The output of the other neurons is 0.

[0096] Finally, the output layer result is multiplied by the vector [1,2,3,…,k]', that is, multiplied by the classification category value corresponding to each neuron, thereby converting the output layer result into a classification category value.

[0097] y = max(f1,f2,…,f k ).

[0098] The "training" of a probabilistic neural network involves determining the number of neurons and connection weights based on the input training samples. Once the parameter values ​​in the network are determined, training ends, resulting in a trained localization model. Finally, test samples are fed into the trained localization model to achieve fault classification.

[0099] Corresponding to the DC filter capacitor open-circuit fault location method provided in Embodiment 1 of the present invention, Embodiment 2 of the present invention provides a DC filter capacitor open-circuit fault location device, comprising:

[0100] The acquisition module is used to acquire raw current data of the DC filter capacitor under normal and open-circuit fault conditions.

[0101] The feature vector extraction module is used to extract effective values ​​from the raw current data, and perform normalization and principal component analysis to extract feature vectors that can reflect the operating state of the DC filter capacitor.

[0102] The fault location module is used to input a portion of the feature vector as a test sample into the DC filter capacitor open-circuit fault location model for classification, thereby realizing the open-circuit fault location of the DC filter capacitor. The DC filter capacitor open-circuit fault location model is obtained by inputting another portion of the feature vector as a training sample into a probabilistic neural network for training.

[0103] Furthermore, the open-circuit fault includes open-circuit faults occurring in all four arms of the high-voltage capacitor, the same open-circuit fault occurring simultaneously in two arms of the high-voltage capacitor, and an open-circuit fault occurring in the low-voltage capacitor; the raw current data acquired by the acquisition module includes: the current i measured by the current transformer at the first end of the DC filter. h (t), the bridge arm branch currents i1(t) and i2(t) measured by the current transformers on the left and right sides of the top of the high-voltage capacitor of the DC filter, and the current i measured by the current transformer on the unbalanced branch of the high-voltage capacitor of the DC filter. un (t), the current i measured by the current transformer at the tail end of the DC filter. w (t), the current i measured by the current transformer on the low-voltage reactor branch. l2 (t), the current i measured by the current transformer on the resistor branch. r (t), where t is the sampling time.

[0104] Furthermore, the feature vector extraction module extracts the feature vectors from the normalized data using principal component analysis, specifically including:

[0105] The normalized 10-dimensional data is subjected to a standardization transformation to obtain the standardized matrix Z;

[0106] Construct a correlation coefficient matrix M, where m is the number of samples in each dimension of the data:

[0107]

[0108] Calculate the eigenvalues ​​λ of the correlation coefficient matrix M j , where λ1≥λ2≥…≥λ 10 The corresponding feature vector is b j =(b 1j ,b 2j ,…,b 10j )T , where j = 1, 2, ..., 10;

[0109] Solve for principal components: U j =Zb j , among which, U j It is the j-th principal component;

[0110] Calculate the cumulative variance contribution rate η of the first l principal components. Σ (l) Select the required principal components and extract the feature vectors:

[0111]

[0112] Furthermore, the probabilistic neural network consists of an input layer, a hidden layer, a summation layer, and an output layer. During training the probabilistic neural network, the training samples are processed according to the following steps:

[0113] First, the input layer receives the training sample X and simultaneously feeds it to the hidden layer, which has c neurons, representing the dimension of the input training sample X.

[0114] Then, the hidden layer, which is a radial base layer with p neurons, calculates the matching degree between the training sample X and each DC filter capacitor classification category according to the following formula:

[0115]

[0116] Among them, W i The connection weights between the input layer and the hidden layer are set to various training samples, and δ is a smoothing factor.

[0117] Next, the summation layer sums the probabilities of neurons belonging to the same classification category in the hidden layer to obtain the probability density function estimate of the fault type, as shown in the following formula. This layer has k neurons, representing the total number of categories of the DC filter capacitor's operating state:

[0118]

[0119] Where p is the dimension of the training samples, m1 is the number of training samples belonging to category 1, and X 1i This is the i-th training sample in category 1;

[0120] The output layer consists of k competing neurons. In the output of the summation layer, this layer finds the neuron with the highest probability. The output of this neuron is 1, which means that the measured data belongs to the DC filter capacitor operation state type corresponding to this neuron. The outputs of the other neurons are 0.

[0121] Finally, the output layer's result is multiplied by the vector [1,2,3,…,k]', which is multiplied by the classification category value corresponding to each neuron, thereby converting the output layer's result into classification category values:

[0122] y = max(f1,f2,…,f k ).

[0123] The "training" of a probabilistic neural network involves determining the number of neurons and connection weights based on the input training samples. Once the parameter values ​​in the network are determined, training ends, resulting in a trained localization model. Finally, test samples are fed into the trained localization model to achieve fault classification.

[0124] For the working principle and process of this embodiment, please refer to the description of the aforementioned Embodiment 1 of the present invention, which will not be repeated here.

[0125] This invention takes the structure and parameters of a 12 / 24 dual-tuned DC filter in the Genam HVDC transmission project as an example. The structure is shown in the attached figure. Figure 3 As shown in Table 1, the parameters are as follows. On the ±500kV HVDC transmission system model of the PSCAD / EMTDC simulation platform, open-circuit fault simulations were performed on the four arms of the high-voltage capacitor and the low-voltage capacitor. Taking three fault conditions as examples—two capacitors connected in series and 15 in parallel, three capacitors connected in series and 15 in parallel, and three capacitors connected in series and 18 in parallel—the open-circuit fault was set to occur in the first second, with a fault duration of 1 second. Raw current data of the capacitor under normal and open-circuit fault conditions were collected at a sampling frequency of 5kHz. For each operating state under each fault condition of the DC filter capacitor, 1000 sets of data were randomly selected as samples during the simulation, resulting in a total of 30000 × 10 sets of sample data. The operating states of the DC filter capacitor are shown in Table 2.

[0126] Table 1 Parameters of the Dual-Tune DC Filter

[0127] Tab.1 Double-tuned DC filter parameters

[0128]

[0129] Table 2 Operating Status of DC Filter Capacitors

[0130] Tab.2 DC filter capacitor operating status

[0131]

[0132] First, the effective values ​​of the raw current data were collected. Then, these 10-dimensional variables were normalized to the interval [0,1]. Principal component analysis was then performed on the normalized data, and the top principal components with a cumulative variance contribution rate greater than 95% were selected. For each operating state under each fault condition, 700 sets of data were randomly selected as training samples and input into a probabilistic neural network model to obtain a trained DC filter capacitor open-circuit fault location model, where the SPREAD parameter was set to 1.1. Finally, test samples were fed into the location model, and the location results of the test samples are shown below. Figure 5 As shown in Table 3, the localization results of the test samples indicate that the method proposed in this invention has a good localization effect.

[0133] Table 3. Localization accuracy of test samples

[0134] Tab.3 Correction rate of test sample location

[0135]

[0136] Compared with the prior art, the beneficial effects of the present invention are: to realize the location of open circuit faults in DC filter capacitors, improve the repair efficiency of open circuit faults in DC filter capacitors, and reduce the possibility of safety accidents for maintenance personnel.

[0137] The above description is merely a preferred embodiment of the present invention and should not be construed as limiting the scope of the invention. Therefore, any equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for locating open-circuit faults in DC filter capacitors, characterized in that, include: Step S1: Collect raw current data of the DC filter capacitor under normal and open-circuit fault conditions; Step S2: Extract effective values ​​from the original current data, and perform normalization and principal component analysis to extract feature vectors that reflect the operating state of the DC filter capacitor. Step S3: A portion of the feature vector is used as a test sample and input into the DC filter capacitor open circuit fault location model for classification, thereby realizing the open circuit fault location of the DC filter capacitor. The DC filter capacitor open circuit fault location model is obtained by inputting another portion of the feature vector as a training sample into a probabilistic neural network for training. The open circuit faults include open circuit faults occurring in all four arms of the high-voltage capacitor, open circuit faults occurring simultaneously in two arms of the high-voltage capacitor, and open circuit faults occurring in the low-voltage capacitor. The raw current data acquired in step S1 includes: the current measured by the current transformer at the beginning of the DC filter. i h ( t The bridge arm branch current measured by the current transformers on the left and right sides of the top of the high-voltage capacitor of the DC filter. i 1( t ), i 2( t The current measured by the current transformer on the unbalanced branch of the high-voltage capacitor of the DC filter. i un ( t The current measured by the current transformer at the tail end of the DC filter. i w ( t The current measured by the current transformer on the low-voltage reactor branch. i l2 ( t The current measured by the current transformer on the resistor branch. i r ( t ), t Sampling time; Depend on i 1( t ), i 2( t )and i un ( t The currents of the bridge arm branches on the left and right sides of the bottom of the high-voltage capacitor were calculated. i 3( t ) i 4( t ): i 3( t ) = i 1( t ) - i un ( t ), i 4( t ) = i 2( t ) + i un ( t );Depend on i w ( t ), i l2 ( t )and i r ( t The branch current of the low-voltage capacitor was calculated. i c2 ( t ): i c2 ( t )= i w ( t ) - i l2 ( t )- i r ( t ).

2. The method according to claim 1, characterized in that, In step S2, principal component analysis is used to extract the feature vector from the normalized data, specifically including: Step S21: Perform a standardization transformation on the normalized 10-dimensional data to obtain the standardized matrix. Z ; Step S22, construct the correlation coefficient matrix M ,in m The number of samples per dimension: Step S23, calculate the correlation coefficient matrix. M eigenvalues ,in The corresponding feature vector is ,in j =1, 2,…, 10; Step S24, solve for principal components in, U j For the first j principal component; Step S25, obtain the previous... l Cumulative variance contribution rate of each principal component Select the required principal components and extract the feature vectors: 。 3. The method according to claim 2, characterized in that, The probabilistic neural network in step S3 consists of an input layer, a hidden layer, a summation layer, and an output layer. During training the probabilistic neural network, the training samples are processed according to the following steps: First, the input layer receives training samples. X Simultaneously, it is delivered to the hidden layer, which has c Each neuron represents an input training sample. X dimensionality; Then, the hidden layer will calculate the training samples according to the following formula. X The degree of matching with the classification category of each DC filter capacitor, this layer is a radial base layer, with p Each neuron represents a training sample. X Number of samples: in, W i These are the connection weights between the input layer and the hidden layer, which are set to various training samples. It is a smoothing factor; Next, the summation layer sums the probabilities of neurons belonging to the same classification category in the hidden layer to obtain the probability density function estimate of the fault type, as shown in the following formula. This layer has k Each neuron represents the total number of categories of DC filter capacitor operating states. in, p For the training sample dimensions, m 1 represents the number of training samples belonging to category 1. X 1i For category 1, the first i One training sample; The output layer consists of k This layer consists of competing neurons. In the output of the summation layer, the neuron with the highest probability is found. The output of this neuron is 1, which means that the measured data belongs to the DC filter capacitor operation state type corresponding to this neuron. The output of the other neurons is 0. Finally, multiply the output layer result by the vector [1, 2, 3, …, k This involves multiplying the result by the classification category value corresponding to each neuron, thereby converting the output layer's result into a classification category value. 。 4. A DC filter capacitor open-circuit fault location device, characterized in that, include: The acquisition module is used to acquire raw current data of the DC filter capacitor under normal and open-circuit fault conditions. The feature vector extraction module is used to extract effective values ​​from the raw current data, and perform normalization and principal component analysis to extract feature vectors that can reflect the operating state of the DC filter capacitor. The fault location module is used to input a portion of the feature vector as a test sample into the DC filter capacitor open-circuit fault location model for classification, thereby realizing the open-circuit fault location of the DC filter capacitor. The DC filter capacitor open-circuit fault location model is obtained by inputting another portion of the feature vector as a training sample into a probabilistic neural network for training. The open-circuit faults include open-circuit faults occurring in all four arms of the high-voltage capacitor, open-circuit faults occurring simultaneously in two arms of the high-voltage capacitor, and open-circuit faults occurring in the low-voltage capacitor; the raw current data acquired by the acquisition module includes: the current measured by the current transformer at the first end of the DC filter. i h ( t The bridge arm branch current measured by the current transformers on the left and right sides of the top of the high-voltage capacitor of the DC filter. i 1( t ), i 2( t The current measured by the current transformer on the unbalanced branch of the high-voltage capacitor of the DC filter. i un ( t The current measured by the current transformer at the tail end of the DC filter. i w ( t The current measured by the current transformer on the low-voltage reactor branch. i l2 ( t The current measured by the current transformer on the resistor branch. i r ( t ), t Sampling time; Depend on i 1( t ), i 2( t )and i un ( t The currents of the bridge arm branches on the left and right sides of the bottom of the high-voltage capacitor were calculated. i 3( t ) i 4( t ): i 3( t ) = i 1( t ) - i un ( t ), i 4( t ) = i 2( t ) + i un ( t );Depend on i w ( t ), i l2 ( t )and i r ( t The branch current of the low-voltage capacitor was calculated. i c2 ( t ): i c2 ( t )= i w ( t ) - i l2 ( t )- i r ( t ).

5. The apparatus according to claim 4, characterized in that, The feature vector extraction module extracts the feature vectors from the normalized data using principal component analysis, specifically including: A standardization transformation is performed on the normalized 10-dimensional data to obtain the standardized matrix. Z ; Construct the correlation coefficient matrix M ,in m The number of samples per dimension: Calculate the correlation coefficient matrix M eigenvalues ,in The corresponding feature vector is ,in j =1, 2,…, 10; Solve for principal components: ,in, U j For the first j principal component; Seeking before l Cumulative variance contribution rate of each principal component Select the required principal components and extract the feature vectors: 。 6. The apparatus according to claim 5, characterized in that, The probabilistic neural network consists of an input layer, hidden layers, a summation layer, and an output layer. During training, the training samples are processed according to the following steps: First, the input layer receives training samples. X Simultaneously, it is delivered to the hidden layer, which has c Each neuron represents an input training sample. X dimensionality; Then, the hidden layer will calculate the training samples according to the following formula. X The degree of matching with the classification category of each DC filter capacitor, this layer is a radial base layer, with p Each neuron represents a training sample. X Number of samples: in, W i These are the connection weights between the input layer and the hidden layer, which are set to various training samples. It is a smoothing factor; Next, the summation layer sums the probabilities of neurons belonging to the same classification category in the hidden layer to obtain the probability density function estimate of the fault type, as shown in the following formula. This layer has k Each neuron represents the total number of categories of DC filter capacitor operating states. in, p For the training sample dimensions, m 1 represents the number of training samples belonging to category 1. X 1i For category 1, the first i One training sample; The output layer consists of k This layer consists of competing neurons. In the output of the summation layer, the neuron with the highest probability is found. The output of this neuron is 1, which means that the measured data belongs to the DC filter capacitor operation state type corresponding to this neuron. The output of the other neurons is 0. Finally, multiply the output layer result by the vector [1, 2, 3, …, k This involves multiplying the result by the classification category value corresponding to each neuron, thereby converting the output layer's result into a classification category value. 。

Images