A method and system for locating open circuit faults in a modular multilevel converter sub-module

Abstract

The application discloses a modular multilevel converter sub-module open-circuit fault positioning method and system, and belongs to the technical field of fault positioning, and comprises the following contents: three-phase output voltage and capacitor voltage of each sub-module under various fault conditions of MMC are collected, a training set is constructed, wavelet packet transformation is carried out on the capacitor voltage of the sub-module to extract frequency band energy features, all samples are stacked by rows to obtain a capacitor voltage high-dimensional feature matrix, PCA dimension reduction is carried out on the high-dimensional feature matrix respectively, two sets of independent projection matrices are stored, two types of low-dimensional feature matrices are output, two-stage probability neural networks are built, GA-PSO algorithm is adopted to complete offline training, voltage data is collected during equipment operation, linear dimension reduction is carried out on the voltage data through feature extraction and pre-stored projection matrices, the voltage data is sequentially sent into two-stage networks for hierarchical diagnosis, fault bridge arms and IGBT fault types are identified, and then fault sub-module numbers are output, and open-circuit fault positioning is completed. The application adopts the above method, and realizes hierarchical and accurate positioning of open-circuit faults.

Description

Technical Field

[0001] This invention relates to the technical field of fault location, and in particular to a method and system for locating open-circuit faults in a modular multilevel converter submodule. Background Technology

[0002] Modular multilevel converters (MMCs) are widely used in core power sectors such as flexible DC transmission, renewable energy grid connection, and marine power systems due to their advantages including modular design, high output power quality, strong scalability, and adaptability to medium- and high-voltage, large-capacity scenarios. MMCs consist of numerous cascaded submodules, with the IGBTs within each submodule being critical power devices that operate at high frequencies for extended periods, making them highly susceptible to open-circuit faults. Failure to quickly and accurately identify the type of faulty IGBT, the faulty bridge arm, and the faulty submodule can easily lead to bridge arm voltage imbalance, system harmonic surges, and circulating current distortion. In severe cases, this can cause converter shutdown, threatening the safe and stable operation of the entire power system. Therefore, accurate and efficient submodule open-circuit fault location technology has significant engineering application value.

[0003] Existing hardware-based methods significantly increase system cost and complexity; model-based methods rely heavily on model accuracy for diagnostic performance and exhibit poor robustness under complex operating conditions; while AI-based methods avoid the shortcomings of the former two, their diagnostic results often only output the faulty bridge arm or the type of faulty IGBT, but cannot provide precise information about which specific submodule within that bridge arm the fault occurred in. This drawback forces maintenance personnel to still need to check multiple submodules within the faulty bridge arm one by one, resulting in low localization efficiency. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for locating open-circuit faults in a modular multilevel converter submodule, enabling hierarchical and precise location of faulty bridge arms, IGBT types, and faulty submodules.

[0005] To achieve the above objectives, the present invention provides a method for locating open-circuit faults in a modular multilevel converter submodule, comprising the following steps: S1. Obtain the three-phase output voltage signal and the capacitor voltage signal of each submodule under various fault conditions of MMC, obtain multiple sets of historical samples, and form a training set; S2. Perform wavelet packet transform on the capacitor voltage signal of each sub-module in each group of historical samples to extract the frequency band energy features. Concatenate all the energy feature vectors of the sub-modules in a single group of samples to form the original high-dimensional feature vector. Stack the original high-dimensional feature vectors of all historical samples row by row to form a high-dimensional feature matrix of capacitor voltage. S3. Based on all historical samples, perform PCA dimensionality reduction on the three-phase output voltage signal and capacitor voltage high-dimensional feature matrix respectively, pre-store two sets of independent PCA projection matrices, and obtain the three-phase voltage low-dimensional feature matrix and sub-module capacitor voltage low-dimensional feature matrix corresponding to the historical samples. S4. Construct a two-level probabilistic neural network localization model, and use the GA-PSO collaborative optimization algorithm to optimize the smoothing factor of the two-level probabilistic neural network based on the low-dimensional features of historical samples, and complete the model training. S5. Real-time acquisition of three-phase output voltage signals and capacitor voltage signals of each submodule during the operation of MMC. After feature extraction and dimensionality reduction by pre-stored PCA projection matrix, the signals are sequentially input into the trained two-level probabilistic neural network localization model to identify faulty bridge arm, IGBT type and faulty submodule number in a hierarchical manner, thereby realizing the localization of open circuit faults in submodules.

[0006] Preferably, step S2 specifically includes the following steps: S21. Select db3 as the wavelet packet decomposition basis function for the capacitor voltage signal of a single submodule and perform 3-level wavelet packet decomposition. Extract the coefficient energy of the 8 terminal nodes after decomposition as the energy feature vector of a single submodule. S22. Concatenate the energy feature vectors corresponding to all sub-modules within a single sample to generate the original high-dimensional row vector of the sample. S23. Stack the original high-dimensional feature vectors of all historical samples in the training set row by row to form a high-dimensional feature matrix of capacitor voltage: ; in, This is the high-dimensional characteristic matrix of capacitor voltage. Let Q be the total number of historical samples, and let Q be the original feature dimension of a single sample.

[0007] Preferably, in step S3, performing PCA dimensionality reduction on the high-dimensional feature matrix of the capacitor voltage specifically includes the following: After standardizing the high-dimensional feature matrix of capacitor voltage, the mean of the j-th dimension of the high-dimensional feature matrix of capacitor voltage is: ; The corresponding standard deviation is: ; The standardized eigencomponents of the high-dimensional eigenma matrix of capacitor voltage are: ; Let the standardized high-dimensional feature matrix of capacitor voltage be denoted as... Furthermore, the covariance matrix is ​​constructed from the high-dimensional characteristic matrix of the capacitor voltage: ; Eigenvalue decomposition of the covariance matrix yields: ; No. i The contribution rate of each principal component is: ; forward k The cumulative contribution rate of each principal component is: ; Once the cumulative contribution rate reaches the preset threshold, the top [contribution rate] will be selected. k The principal components constitute a dimension-reduced space, and the corresponding capacitor-voltage PCA projection matrix is ​​written as: ; The low-dimensional feature matrix of the submodule capacitor voltage obtained by the projection matrix is ​​as follows: .

[0008] Preferably, step S4 specifically includes the following steps: S41. Construct a two-level probabilistic neural network localization model. The first-level network takes the three-phase voltage low-dimensional feature matrix as input and is used to output the fault arm and IGBT fault type. The second-level network takes the sub-module capacitor voltage low-dimensional feature matrix as input and is used to output the fault sub-module number. Independent smoothing factors are set for the two-level networks respectively. S42. The GA-PSO collaborative optimization algorithm is used to iteratively optimize the smoothing factor of the two-level network: the smoothing factor is encoded as the particle position to initialize the particle swarm, and the fitness function is constructed with the model prediction error; the particle inertia weight and learning factor are dynamically adjusted, and the particle velocity and position are iteratively updated. Genetic crossover and mutation operations are introduced, and the iteration is continued until the accuracy threshold or the maximum number of iterations is met. The optimal smoothing factor of each of the two-level networks is then output. S43. Assign the optimal smoothing factor to the corresponding probabilistic neural network to complete the training of the two-level probabilistic neural network localization model.

[0009] Preferably, in step S41, the first-level probabilistic neural network and the second-level probabilistic neural network each correspond to an independent fault numbering system, and the specific classification rules are as follows: The first-level probabilistic neural network is used to identify the faulty bridge arm and the faulty IGBT type, and a total of 12 fault numbers are divided; among them, the upper tube T1 of each of the six bridge arms of the MMC is open and corresponds to numbers 1 to 6 in sequence, and the lower tube T2 of each of the six bridge arms is open and corresponds to numbers 7 to 12 in sequence. The second-level probabilistic neural network is used to locate the fault submodule number, and a total of 20 fault numbers are divided; the submodules numbered 1 to 20 in a single bridge arm correspond to fault numbers 1 to 20 respectively; During online diagnosis, the first-level network outputs a global fault number to identify the bridge arm and IGBT type where an open-circuit fault occurs. Then, the low-dimensional feature vector of the capacitor voltage corresponding to the faulty bridge arm is extracted and input into the second-level network, which outputs the fault number corresponding to the faulty submodule within the bridge arm to complete the localization.

[0010] Preferably, step S5 specifically includes the following steps: S51. Real-time acquisition of three-phase output voltage signals and all submodule capacitor voltage signals during the current MMC operation process to obtain a single set of real-time sampling data; S52. Perform the wavelet packet energy extraction process of step S2 on the real-time submodule capacitor voltage, and concatenate them to obtain the original high-dimensional row vector of the real-time capacitor voltage; use the pre-stored capacitor voltage PCA projection matrix to perform linear projection to obtain the low-dimensional feature vector of the real-time submodule capacitor voltage. S53. Organize the real-time three-phase output voltage into an original vector, and use the pre-stored three-phase voltage PCA projection matrix to perform linear projection to obtain the low-dimensional feature vector of the real-time three-phase voltage. S54. Input the real-time three-phase voltage low-dimensional feature vector into the trained first-level probabilistic neural network, and output the faulty bridge arm and the faulty IGBT type. S55. Extract the low-dimensional feature vector of the real-time submodule capacitor voltage corresponding to the faulty bridge arm in step S54, input it into the trained second-level probabilistic neural network, output the faulty submodule number, and complete the accurate location of the submodule open-circuit fault.

[0011] A modular multilevel converter submodule open-circuit fault location system includes a sample acquisition module, a feature extraction module, a PCA dimensionality reduction processing module, a model training module, and an online fault diagnosis module; The sample acquisition module is used to acquire the three-phase output voltage signals and the capacitor voltage signals of each submodule under various fault conditions of the MMC, obtain multiple sets of historical samples and form a training set; and to acquire the three-phase output voltage signals and the capacitor voltage signals of each submodule in real time during the operation of the MMC, and output a single set of real-time sampling data. The feature extraction module, connected to the sample acquisition module, is used to perform wavelet packet transform on the capacitor voltage signals of each sub-module in each group of historical samples and real-time sampling data to extract frequency band energy features. It concatenates the energy feature vectors of all sub-modules in a single sample into the original high-dimensional feature vector, and stacks the original high-dimensional feature vectors of all historical samples row by row to form a high-dimensional feature matrix of capacitor voltage. The PCA dimensionality reduction module, connected to the feature extraction module, performs PCA dimensionality reduction on the high-dimensional feature matrices of the three-phase output voltage signal and capacitor voltage based on all historical samples. It pre-stores two sets of independent PCA projection matrices and outputs the low-dimensional feature matrices of the three-phase voltage and the low-dimensional feature matrices of the sub-module capacitor voltage corresponding to the historical samples. At the same time, it uses the pre-stored PCA projection matrices to linearly project the feature vectors of the real-time sampled data, and outputs the low-dimensional feature vectors of the real-time three-phase voltage and the real-time low-dimensional feature vectors of the sub-module capacitor voltage. The model training module, connected to the PCA dimensionality reduction module, is used to construct a two-level probabilistic neural network localization model. The GA-PSO collaborative optimization algorithm is used to optimize the smoothing factor of the two-level probabilistic neural networks based on the low-dimensional features of historical samples, and complete the offline training of the two-level model. The online fault diagnosis module is connected to the PCA dimensionality reduction module and the model training module, respectively. It is used to receive real-time low-dimensional feature vectors and input them into the trained two-level probabilistic neural network localization model in sequence. It identifies the faulty bridge arm, IGBT type and faulty submodule number in a hierarchical manner, and outputs the submodule open circuit fault localization result.

[0012] Therefore, the present invention employs the above-described method and system for locating open-circuit faults in a modular multilevel converter submodule, which has the following advantages: (1) In this invention, wavelet packet transform is used to extract the multi-band energy features of the capacitor voltage of the sub-module, fully explore the internal fault information of the voltage signal, and solve the problems of insufficient feature representation and easy misjudgment and omission in traditional methods.

[0013] (2) In this invention, independent PCA projection matrices are built for the three-phase output voltage and capacitor voltage respectively to reduce the dimensionality of the dual channels, eliminate redundant features, reduce the amount of calculation, and improve the real-time performance of fault diagnosis.

[0014] (3) In this invention, a two-level probabilistic neural network is set up to identify faults. Combined with the GA-PSO algorithm, the network smoothing factor is automatically optimized. No manual parameter tuning is required. It can simultaneously and accurately distinguish faulty bridge arms, IGBT fault types and faulty sub-module numbers, resulting in stronger positioning accuracy and generalization ability.

[0015] (4) In this invention, through a phased strategy and refined feature processing, the faulty bridge arm, the faulty IGBT type, and the specific number of the faulty submodule in the bridge arm can be accurately located simultaneously, providing more comprehensive and refined diagnostic information, thereby significantly shortening the troubleshooting time and system downtime for maintenance personnel.

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1This is a schematic diagram of the open-circuit fault location method for a modular multilevel converter submodule according to the present invention; Figure 2 A schematic diagram of the data processing flow based on wavelet packet transform and principal component analysis (PCA) provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of a phased fault location process provided in an embodiment of the present invention; Figure 4 The accuracy of the GA-PSO-PNN1 test set provided in the embodiments of the present invention; Figure 5 The GA-PSO-PNN1 confusion matrix provided in this embodiment of the invention; Figure 6 The accuracy of the GA-PSO-PNN2 test set provided in the embodiments of the present invention; Figure 7 The GA-PSO-PNN2 confusion matrix provided in this embodiment of the invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Specific model specifications need to be selected and determined according to the actual specifications of the device, etc. The specific selection calculation method adopts existing technology in the art, and therefore will not be described in detail.

[0019] Example like Figures 1-3 As shown, the present invention provides a method for locating open-circuit faults in a modular multilevel converter submodule, comprising the following steps: S1. Obtain the three-phase output voltage signal and the capacitor voltage signal of each submodule under various fault conditions of MMC, obtain multiple sets of historical samples, and form a training set; S2. Perform wavelet packet transform on the capacitor voltage signal of each sub-module in each group of historical samples to extract the frequency band energy features. Concatenate all the energy feature vectors of the sub-modules in a single group of samples to form the original high-dimensional feature vector. Stack the original high-dimensional feature vectors of all historical samples row by row to form a high-dimensional feature matrix of capacitor voltage. S21. Select db3 as the wavelet packet decomposition basis function for the capacitor voltage signal of a single submodule and perform 3-level wavelet packet decomposition. Extract the coefficient energy of the 8 terminal nodes after decomposition as the energy feature vector of a single submodule. S22. Concatenate the energy feature vectors corresponding to all sub-modules within a single sample to generate the original high-dimensional row vector of the sample. S23. Stack the original high-dimensional feature vectors of all historical samples in the training set row by row to form a high-dimensional feature matrix of capacitor voltage: ; in, This is the high-dimensional characteristic matrix of capacitor voltage. Let Q be the total number of historical samples, and let Q be the original feature dimension of a single sample.

[0020] S3. Based on all historical samples, perform PCA dimensionality reduction on the three-phase output voltage signal and capacitor voltage high-dimensional feature matrix respectively, pre-store two sets of independent PCA projection matrices, and obtain the three-phase voltage low-dimensional feature matrix and sub-module capacitor voltage low-dimensional feature matrix corresponding to the historical samples. Performing PCA dimensionality reduction on the high-dimensional feature matrix of capacitor voltage specifically includes the following: After standardizing the high-dimensional feature matrix of capacitor voltage, the mean of the j-th dimension of the high-dimensional feature matrix of capacitor voltage is: ; The corresponding standard deviation is: ; The standardized eigencomponents of the high-dimensional eigenma matrix of capacitor voltage are: ; Let the standardized high-dimensional feature matrix of capacitor voltage be denoted as... Furthermore, the covariance matrix is ​​constructed from the high-dimensional characteristic matrix of the capacitor voltage: ; Eigenvalue decomposition of the covariance matrix yields: ; No. i The contribution rate of each principal component is: ; forward k The cumulative contribution rate of each principal component is: ; Once the cumulative contribution rate reaches the preset threshold, the top [contribution rate] will be selected. k The principal components constitute a dimension-reduced space, and the corresponding capacitor-voltage PCA projection matrix is ​​written as: ; The low-dimensional feature matrix of the submodule capacitor voltage obtained by the projection matrix is ​​as follows: .

[0021] S4. Construct a two-level probabilistic neural network localization model, and use the GA-PSO collaborative optimization algorithm to optimize the smoothing factor of the two-level probabilistic neural network based on the low-dimensional features of historical samples, and complete the model training. S41. Construct a two-level probabilistic neural network localization model. The first-level network takes the three-phase voltage low-dimensional feature matrix as input and is used to output the fault arm and IGBT fault type. The second-level network takes the sub-module capacitor voltage low-dimensional feature matrix as input and is used to output the fault sub-module number. Independent smoothing factors are set for the two-level networks respectively. The first-level probabilistic neural network and the second-level probabilistic neural network each correspond to an independent fault numbering system, with the specific classification rules as follows: The first-level probabilistic neural network is used to identify the faulty bridge arm and the faulty IGBT type, and a total of 12 fault numbers are divided; among them, the upper tube T1 of each of the six bridge arms of the MMC is open and corresponds to numbers 1 to 6 in sequence, and the lower tube T2 of each of the six bridge arms is open and corresponds to numbers 7 to 12 in sequence. The second-level probabilistic neural network is used to locate the fault submodule number, and a total of 20 fault numbers are divided; the submodules numbered 1 to 20 in a single bridge arm correspond to fault numbers 1 to 20 respectively; During online diagnosis, the first-level network outputs a global fault number to identify the bridge arm and IGBT type where the open-circuit fault occurs; then, the low-dimensional feature vector of the capacitor voltage corresponding to the faulty bridge arm is extracted and input into the second-level network, which outputs the fault number corresponding to the faulty sub-module within the bridge arm to complete the localization. S42. The GA-PSO collaborative optimization algorithm is used to iteratively optimize the smoothing factor of the two-level network: the smoothing factor is encoded as the particle position to initialize the particle swarm, and the fitness function is constructed with the model prediction error; the particle inertia weight and learning factor are dynamically adjusted, and the particle velocity and position are iteratively updated. Genetic crossover and mutation operations are introduced, and the iteration is continued until the accuracy threshold or the maximum number of iterations is met. The optimal smoothing factor of each of the two-level networks is then output. S43. Assign the optimal smoothing factor to the corresponding probabilistic neural network to complete the training of the two-level probabilistic neural network localization model.

[0022] S5. Real-time acquisition of three-phase output voltage signals and capacitor voltage signals of each submodule during the operation of MMC. After feature extraction and dimensionality reduction by pre-stored PCA projection matrix, the signals are sequentially input into the trained two-level probabilistic neural network localization model to identify faulty bridge arm, IGBT type and faulty submodule number in a hierarchical manner, thereby realizing the localization of open circuit faults in submodules. S51. Real-time acquisition of three-phase output voltage signals and all submodule capacitor voltage signals during the current MMC operation process to obtain a single set of real-time sampling data; S52. Perform the wavelet packet energy extraction process of step S2 on the real-time submodule capacitor voltage, and concatenate them to obtain the original high-dimensional row vector of the real-time capacitor voltage; use the pre-stored capacitor voltage PCA projection matrix to perform linear projection to obtain the low-dimensional feature vector of the real-time submodule capacitor voltage. S53. Organize the real-time three-phase output voltage into an original vector, and use the pre-stored three-phase voltage PCA projection matrix to perform linear projection to obtain the low-dimensional feature vector of the real-time three-phase voltage. S54. Input the real-time three-phase voltage low-dimensional feature vector into the trained first-level probabilistic neural network, and output the faulty bridge arm and the faulty IGBT type. S55. Extract the low-dimensional feature vector of the real-time submodule capacitor voltage corresponding to the faulty bridge arm in step S54, input it into the trained second-level probabilistic neural network, output the faulty submodule number, and complete the accurate location of the submodule open-circuit fault.

[0023] A modular multilevel converter submodule open-circuit fault location system includes a sample acquisition module, a feature extraction module, a PCA dimensionality reduction processing module, a model training module, and an online fault diagnosis module; The sample acquisition module is used to acquire the three-phase output voltage signals and the capacitor voltage signals of each submodule under various fault conditions of the MMC, obtain multiple sets of historical samples and form a training set; and to acquire the three-phase output voltage signals and the capacitor voltage signals of each submodule in real time during the operation of the MMC, and output a single set of real-time sampling data. The feature extraction module, connected to the sample acquisition module, is used to perform wavelet packet transform on the capacitor voltage signals of each sub-module in each group of historical samples and real-time sampling data to extract frequency band energy features. It concatenates the energy feature vectors of all sub-modules in a single sample into the original high-dimensional feature vector, and stacks the original high-dimensional feature vectors of all historical samples row by row to form a high-dimensional feature matrix of capacitor voltage. The PCA dimensionality reduction module, connected to the feature extraction module, performs PCA dimensionality reduction on the high-dimensional feature matrices of the three-phase output voltage signal and capacitor voltage based on all historical samples. It pre-stores two sets of independent PCA projection matrices and outputs the low-dimensional feature matrices of the three-phase voltage and the low-dimensional feature matrices of the sub-module capacitor voltage corresponding to the historical samples. At the same time, it uses the pre-stored PCA projection matrices to linearly project the feature vectors of the real-time sampled data, and outputs the low-dimensional feature vectors of the real-time three-phase voltage and the real-time low-dimensional feature vectors of the sub-module capacitor voltage. The model training module, connected to the PCA dimensionality reduction module, is used to construct a two-level probabilistic neural network localization model. The GA-PSO collaborative optimization algorithm is used to optimize the smoothing factor of the two-level probabilistic neural networks based on the low-dimensional features of historical samples, and complete the offline training of the two-level model. The online fault diagnosis module is connected to the PCA dimensionality reduction module and the model training module, respectively. It is used to receive real-time low-dimensional feature vectors and input them into the trained two-level probabilistic neural network localization model in sequence. It identifies the faulty bridge arm, IGBT type and faulty submodule number in a hierarchical manner, and outputs the submodule open circuit fault localization result.

[0024] In this embodiment, the fault location accuracy of GA-PSO-PNN1 and GA-PSO-PNN2 is tested using a test set, such as... Figures 4-5 As shown, Figure 4 and Figure 5 The prediction results and confusion matrix of the GA-PSO-PNN1 model on the test set are presented. The results show that the prediction results of GA-PSO-PNN1 generally match the true labels well, with a recognition accuracy of 97.52% on the test set. Correct recognition results are mainly concentrated on the main diagonal, while off-diagonal elements are reduced, indicating that the model's ability to distinguish various types of faults is improved. A small number of misclassifications still mainly occur between adjacent categories, and there are no obvious cross-interval misclassifications, indicating that the model has good accuracy in the recognition of 12 types of faults.

[0025] like Figures 6-7 As shown, by Figure 6 As can be seen, the prediction results of GA-PSO-PNN2 on the test set are generally highly consistent with the true labels, with only a small number of samples showing bias, and the test set accuracy reaches 96.52%. Figure 7 It can be seen that 3 samples of SM1 were misclassified as SM2, only 1 sample of SM4 was misclassified as SM5, only 1 sample of SM19 was misclassified as SM18, and 2 samples of SM20 were misclassified as SM19. Overall, the number of misclassifications is relatively small, and they are mainly concentrated between adjacent SMs, with no large-scale misclassifications.

[0026] Therefore, this invention employs a modular multilevel converter (MMC) submodule open-circuit fault location method and system. It fully extracts the energy fault features of the submodule capacitor voltage frequency band through wavelet packet transform, and then uses dual-channel independent PCA to reduce the dimensionality of the three-phase voltage and capacitor voltage features to eliminate redundant information. Combined with a two-level probabilistic neural network with GA-PSO collaborative optimization, it achieves hierarchical fault identification. This method can automatically optimize the network smoothing factor, accurately distinguish faulty bridge arms, IGBT fault types, and faulty submodule numbers. It boasts strong feature representation capabilities, high computational efficiency, and good location accuracy, effectively improving the real-time performance and reliability of MMC fault diagnosis.

[0027] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for locating open-circuit faults in a modular multilevel converter submodule, characterized in that: Includes the following steps: S1. Obtain the three-phase output voltage signal and the capacitor voltage signal of each submodule under various fault conditions of MMC, obtain multiple sets of historical samples, and form a training set; S2. Perform wavelet packet transform on the capacitor voltage signal of each sub-module in each group of historical samples to extract the frequency band energy features. Concatenate all the energy feature vectors of the sub-modules in a single group of samples to form the original high-dimensional feature vector. Stack the original high-dimensional feature vectors of all historical samples row by row to form a high-dimensional feature matrix of capacitor voltage. S3. Based on all historical samples, perform PCA dimensionality reduction on the three-phase output voltage signal and capacitor voltage high-dimensional feature matrix respectively, pre-store two sets of independent PCA projection matrices, and obtain the three-phase voltage low-dimensional feature matrix and sub-module capacitor voltage low-dimensional feature matrix corresponding to the historical samples. S4. Construct a two-level probabilistic neural network localization model, and use the GA-PSO collaborative optimization algorithm to optimize the smoothing factor of the two-level probabilistic neural network based on the low-dimensional features of historical samples, and complete the model training. S5. Real-time acquisition of three-phase output voltage signals and capacitor voltage signals of each submodule during the operation of MMC. After feature extraction and dimensionality reduction by pre-stored PCA projection matrix, the signals are sequentially input into the trained two-level probabilistic neural network localization model to identify faulty bridge arm, IGBT type and faulty submodule number in a hierarchical manner, thereby realizing the localization of open circuit faults in submodules.

2. The method for locating open-circuit faults in a modular multilevel converter submodule according to claim 1, characterized in that: Step S2 specifically includes the following steps: S21. Select db3 as the wavelet packet decomposition basis function for the capacitor voltage signal of a single submodule and perform 3-level wavelet packet decomposition. Extract the coefficient energy of the 8 terminal nodes after decomposition as the energy feature vector of a single submodule. S22. Concatenate the energy feature vectors corresponding to all sub-modules within a single sample to generate the original high-dimensional row vector of the sample. S23. Stack the original high-dimensional feature vectors of all historical samples in the training set row by row to form a high-dimensional feature matrix of capacitor voltage: ； in, This is the high-dimensional characteristic matrix of capacitor voltage. Let Q be the total number of historical samples, and let Q be the original feature dimension of a single sample.

3. The method for locating open-circuit faults in a modular multilevel converter submodule according to claim 2, characterized in that: In step S3, performing PCA dimensionality reduction on the high-dimensional feature matrix of the capacitor voltage specifically includes the following: After standardizing the high-dimensional feature matrix of capacitor voltage, the mean of the j-th dimension of the high-dimensional feature matrix of capacitor voltage is: ； The corresponding standard deviation is: ； The standardized eigencomponents of the high-dimensional eigenma matrix of capacitor voltage are: ； Let the standardized high-dimensional feature matrix of capacitor voltage be denoted as... Furthermore, the covariance matrix is ​​constructed from the high-dimensional characteristic matrix of the capacitor voltage: ； Eigenvalue decomposition of the covariance matrix yields: ； No. i The contribution rate of each principal component is: ； forward k The cumulative contribution rate of each principal component is: ； Once the cumulative contribution rate reaches the preset threshold, the top [contribution rate] will be selected. k The principal components constitute a dimension-reduced space, and the corresponding capacitor-voltage PCA projection matrix is ​​written as: ； The low-dimensional feature matrix of the submodule capacitor voltage obtained by the projection matrix is ​​as follows: 。 4. The method for locating open-circuit faults in a modular multilevel converter submodule according to claim 3, characterized in that: Step S4 specifically includes the following steps: S41. Construct a two-level probabilistic neural network localization model. The first-level network takes the three-phase voltage low-dimensional feature matrix as input and is used to output the fault arm and IGBT fault type. The second-level network takes the sub-module capacitor voltage low-dimensional feature matrix as input and is used to output the fault sub-module number. Independent smoothing factors are set for the two-level networks respectively. S42. The GA-PSO collaborative optimization algorithm is used to iteratively optimize the smoothing factor of the two-level network: the smoothing factor is encoded as the particle position to initialize the particle swarm, and the fitness function is constructed with the model prediction error; the particle inertia weight and learning factor are dynamically adjusted, and the particle velocity and position are iteratively updated. Genetic crossover and mutation operations are introduced, and the iteration is continued until the accuracy threshold or the maximum number of iterations is met. The optimal smoothing factor of each of the two-level networks is then output. S43. Assign the optimal smoothing factor to the corresponding probabilistic neural network to complete the training of the two-level probabilistic neural network localization model.

5. The method for locating open-circuit faults in a modular multilevel converter submodule according to claim 4, characterized in that: In step S41, the first-level probabilistic neural network and the second-level probabilistic neural network each correspond to an independent fault numbering system, and the specific classification rules are as follows: The first-level probabilistic neural network is used to identify the faulty bridge arm and the faulty IGBT type, and a total of 12 fault numbers are divided; among them, the upper tube T1 of each of the six bridge arms of the MMC is open and corresponds to numbers 1 to 6 in sequence, and the lower tube T2 of each of the six bridge arms is open and corresponds to numbers 7 to 12 in sequence. The second-level probabilistic neural network is used to locate the fault submodule number, and a total of 20 fault numbers are divided; the submodules numbered 1 to 20 in a single bridge arm correspond to fault numbers 1 to 20 respectively; During online diagnosis, the first-level network outputs a global fault number to identify the bridge arm and IGBT type where an open-circuit fault occurs. Then, the low-dimensional feature vector of the capacitor voltage corresponding to the faulty bridge arm is extracted and input into the second-level network, which outputs the fault number corresponding to the faulty submodule within the bridge arm to complete the localization.

6. The method for locating open-circuit faults in a modular multilevel converter submodule according to claim 5, characterized in that: Step S5 specifically includes the following steps: S51. Real-time acquisition of three-phase output voltage signals and all submodule capacitor voltage signals during the current MMC operation process to obtain a single set of real-time sampling data; S52. Perform the wavelet packet energy extraction process of step S2 on the real-time submodule capacitor voltage, and concatenate them to obtain the original high-dimensional row vector of the real-time capacitor voltage; use the pre-stored capacitor voltage PCA projection matrix to perform linear projection to obtain the low-dimensional feature vector of the real-time submodule capacitor voltage. S53. Organize the real-time three-phase output voltage into an original vector, and use the pre-stored three-phase voltage PCA projection matrix to perform linear projection to obtain the low-dimensional feature vector of the real-time three-phase voltage. S54. Input the real-time three-phase voltage low-dimensional feature vector into the trained first-level probabilistic neural network, and output the faulty bridge arm and the faulty IGBT type. S55. Extract the low-dimensional feature vector of the real-time submodule capacitor voltage corresponding to the faulty bridge arm in step S54, input it into the trained second-level probabilistic neural network, output the faulty submodule number, and complete the accurate location of the submodule open-circuit fault.

7. A modular multilevel converter submodule open-circuit fault location system, used to implement the modular multilevel converter submodule open-circuit fault location method according to any one of claims 1-6, characterized in that: It includes a sample acquisition module, a feature extraction module, a PCA dimensionality reduction module, a model training module, and an online fault diagnosis module; The sample acquisition module is used to acquire the three-phase output voltage signals and the capacitor voltage signals of each submodule under various fault conditions of the MMC, obtain multiple sets of historical samples and form a training set; and to acquire the three-phase output voltage signals and the capacitor voltage signals of each submodule in real time during the operation of the MMC, and output a single set of real-time sampling data. The feature extraction module, connected to the sample acquisition module, is used to perform wavelet packet transform on the capacitor voltage signals of each sub-module in each group of historical samples and real-time sampling data to extract frequency band energy features. It concatenates the energy feature vectors of all sub-modules in a single sample into the original high-dimensional feature vector, and stacks the original high-dimensional feature vectors of all historical samples row by row to form a high-dimensional feature matrix of capacitor voltage. The PCA dimensionality reduction module, connected to the feature extraction module, performs PCA dimensionality reduction on the high-dimensional feature matrices of the three-phase output voltage signal and capacitor voltage based on all historical samples. It pre-stores two sets of independent PCA projection matrices and outputs the low-dimensional feature matrices of the three-phase voltage and the low-dimensional feature matrices of the sub-module capacitor voltage corresponding to the historical samples. At the same time, it uses the pre-stored PCA projection matrices to linearly project the feature vectors of the real-time sampled data, and outputs the low-dimensional feature vectors of the real-time three-phase voltage and the real-time low-dimensional feature vectors of the sub-module capacitor voltage. The model training module, connected to the PCA dimensionality reduction module, is used to construct a two-level probabilistic neural network localization model. The GA-PSO collaborative optimization algorithm is used to optimize the smoothing factor of the two-level probabilistic neural networks based on the low-dimensional features of historical samples, and complete the offline training of the two-level model. The online fault diagnosis module is connected to the PCA dimensionality reduction module and the model training module, respectively. It is used to receive real-time low-dimensional feature vectors and input them into the trained two-level probabilistic neural network localization model in sequence. It identifies the faulty bridge arm, IGBT type and faulty submodule number in a hierarchical manner, and outputs the submodule open circuit fault localization result.

Images