Modular multilevel converter sub-module capacitor fault monitoring method based on improved convolutional neural network

Abstract

The application relates to a kind of modular multilevel converter sub-module capacitor fault monitoring methods based on improved convolutional neural network, capacitor fault monitoring field.The method first extracts the corresponding aging characteristics according to the characteristics of sub-module capacitor voltage under the condition of aging according to the characteristics of sub-module capacitor voltage;Second, the network architecture of convolutional neural network combined with gated recurrent unit (CNN-GRU) is constructed, and the Adam algorithm, batch normalization processing (BN) and other methods are combined to prevent overfitting, and the accuracy of the method is judged by combining the aging capacitor label;Finally, the accuracy is used as the objective function, and the learning rate, regularization coefficient and learning rate drop factor in the CNN-GRU network are optimized by using the whale optimization algorithm (WOA), and the results obtained by the optimization algorithm are brought into the corresponding network, so that a relatively high accuracy can be obtained.

Description

Technical Field

[0001] The present invention belongs to the field of capacitor fault monitoring, and in particular relates to a method for monitoring capacitor faults in a modular multi-level converter submodule based on an improved convolutional neural network. Background Art

[0002] Modular multilevel converters (MMCs) have become a key topology in HVDC transmission technology due to their advantages, including high voltage output waveform quality and modularity. MMC submodules primarily utilize full-bridge and half-bridge structures, both of which primarily consist of insulated gate bipolar transistors (IGBTs) and capacitors. Each load operation of the capacitors in a submodule affects their capacitance and equivalent resistance. Over time, when the capacitors in a module age to 80% of their rated capacitance (when their equivalent resistance is twice their initial value), they become ineffective. Failure to perform regular inspections can lead to a deterioration in the overall power quality of the MMC system, even leading to a major system failure. Therefore, real-time online monitoring of the capacitors in the MMC system submodules is crucial to maintaining the safe and stable operation of the entire MMC system and ensuring power quality.

[0003] Currently, methods for monitoring the status of MMC capacitors include combining hardware monitoring with reference submodules, but these methods raise questions about the reliability of the hardware. Furthermore, in actual MMC systems, the number of submodules on the bridge arms can reach hundreds, making real-time online monitoring of the MMC system impossible. Meanwhile, some researchers have explored monitoring the charge and discharge characteristics of the capacitors themselves and using artificial intelligence algorithms. However, these methods often require consideration of external factors such as noise, temperature, high-frequency harmonics, and systematic and random errors in the sampling process. Summary of the Invention

[0004] The purpose of the present invention is to provide a modular multilevel converter submodule capacitor fault monitoring method based on an improved convolutional neural network to address the problem that aging of submodule capacitors in a modular multilevel converter causes a degradation of submodule voltage quality, which in turn leads to a degradation of output AC voltage quality and even paralysis of the entire system.

[0005] To achieve the above objectives, the technical solution of the present invention is: a method for monitoring capacitor faults in modular multi-level converter submodules based on an improved convolutional neural network, comprising:

[0006] According to the characteristics of the capacitor voltage of the modular multilevel converter submodule, the corresponding aging characteristics are extracted for the submodule capacitor voltage under aging conditions;

[0007] Construct a convolutional neural network combined with a gated recurrent unit (CNN)-GRU network, and use the aging capacitor label to determine the accuracy of the constructed network.

[0008] Taking accuracy as the objective function, the whale optimization algorithm (WOA) is used to optimize the learning rate, regularization coefficient, and learning rate reduction factor in the CNN-GRU network to achieve accurate monitoring of sub-module capacitor aging.

[0009] In one embodiment of the present invention, the method is implemented as follows:

[0010] Step 1: Preliminary data preprocessing: Preprocess the feature data of the modular multilevel converter submodule capacitors to be processed; determine the number of categories, feature dimensions, and sample size of the feature data to be processed, and shuffle the dataset; and divide the dataset into proportional coefficients to determine the corresponding training and test sets.

[0011] Step 2: Perform secondary processing on the preprocessed data: normalize the preprocessed data, flatten the data after processing, and complete the corresponding data format conversion;

[0012] Step 3: Establish a convolutional neural network combined with a gated recurrent unit (CNN)-GRU network: The structure of the convolutional neural network combined with a gated recurrent unit (CNN)-GRU network includes a sequence folding / unfolding layer, an input layer, a convolutional layer, a batch normalization layer, a pooling layer, a GRU layer, a fully connected layer, and an output layer.

[0013] Step 4: Set parameters and build a fitness function: After building a convolutional neural network combined with a gated recurrent unit (CNN-GRU) network, determine that the parameters to be optimized are the learning rate, regularization parameter, and learning rate reduction factor, and build a fitness function.

[0014] Step 5: Use the whale optimization algorithm and set relevant parameters to optimize the convolutional neural network combined with the gated recurrent unit CNN-GRU network.

[0015] In one embodiment of the present invention, in step 2, the characteristic data of the capacitors of the modular multi-level converter submodules include wavelet decomposition characteristics, the maximum value of the differential capacitor voltage, and the periodic integral value of the differential capacitor voltage.

[0016] In one embodiment of the present invention, in step 3, the convolution layer is composed of multiple feature surfaces, each feature surface is composed of multiple neurons, and each neuron of each feature surface is connected to a local area of ​​the feature surface of the previous layer through a convolution kernel; the convolution kernel is a weight matrix; the convolution layer extracts different features of the input through the convolution operation, the first convolution layer extracts low-level features, and the higher convolution layers extract higher-level features. The entire convolution process is according to the following formula:

[0017]

[0018] Where: Represents the output value of the vth neuron in the output feature surface u in the convolutional layer; Represents the output value of the nth neuron in the input feature surface m; ω a(i)b(j) represents the connection weight between the i-th neuron in the input feature plane a and the j-th neuron in the output feature plane b; b i represents the bias of the output feature surface i; f cov represents the activation function.

[0019] In one embodiment of the present invention, in step 3, the batch normalization layer makes the input distribution of each layer of the network more stable by forcibly pulling the distribution of the input value of any neuron in each layer of the neural network back to a standard normal distribution with a mean of 0 and a variance or standard deviation of 1. The corresponding relevant expressions are as follows:

[0020]

[0021] Where: x i Represents the output result of the previous layer (convolutional layer); m is the size of batch normalization; μ n is the mean of the batch number; σ n is the standard deviation after batch data processing; ε is a small value to prevent the denominator from becoming 0; Represents the output result after batch normalization; γ and β are the parameters of the transformation that scales and translates the normalized data. Initially, γ = 1 and β = 0. γ and β can be adjusted through learning.

[0022] In one embodiment of the present invention, in step 3, the input of the pooling layer comes from multiple feature maps in the convolution layer, and the feature maps in the pooling layer and the feature maps in the convolution layer are combined. Figure 1 One-to-one correspondence, the number of feature maps remains unchanged; the neurons in the pooling layer are connected to the local receptive fields in the input layer, so that each neuron only processes a small part of the input data, and the input areas of different neurons do not overlap; pooling includes maximum pooling and average pooling. The maximum pooling formula is as follows:

[0023] m i(x,b) =max (b-1)ω＜t＜bω {o i(x,t)}b∈1,2,...,p

[0024] Among them: i(x,t) represents the t-th neuron of the x-th map in the i-th layer; ω represents the width of the convolution kernel; b represents the b-th pooling kernel, and there are p in total.

[0025] In one embodiment of the present invention, in step 3, the GRU layer includes a reset gate r and an update gate z to control the information flow, so that it can learn to retain important information and discard irrelevant information. Specifically, the reset gate r is used to adjust the combination of the previous memory and the new input, and the update gate is used to determine the amount of the previous memory to be retained. The relevant formula of GRU is as follows:

[0026] z t =σ(W z x t +U z h t-1 )

[0027] r t =σ(W t x t +U t h t- 1)

[0028]

[0029]

[0030] Where: σ(.) represents the activation function; z t Represents the value of the update gate in the current state; r t It means resetting the gate value in the current state; is the cell value, which aims to discard irrelevant historical information and represents the input x t and the past hidden state h t-1 Summary of h t Represents the output of the hidden layer. U z 、W z 、U t 、W t , U and W are trainable parameter matrices, and the model parameters including U and W are shared in each time step.

[0031] In one embodiment of the present invention, in step 3, there are one or more fully connected layers. In the fully connected layer, each neuron is fully connected to all neurons in the previous layer. The output of the last fully connected layer is passed to an output layer, which is classified using Softmax logistic regression. This is called a Softmax layer. The Softmax function corresponds to the following:

[0032]

[0033] Where: As the input vector of the Softmax function, (x0,x1,...,x k ) composition; x i The elements of the input vector of the Softmax function can take any real value; As a standard exponential function, applied to each element of the input vector; the formula is the normalization term, which ensures that all output values ​​of the function sum to 1 and that each value is in the range (0,1) to constitute a valid probability distribution.

[0034] In one embodiment of the present invention, in step 4, the fitness function formula is as follows:

[0035]

[0036] In the fitness function, Represents the accuracy of the training set, T_sim1 represents the result of the training set obtained after training; T_train represents the original result of the training set; length(T_train) represents the number of training sets in this training;

[0037] Overall, The larger it is, the smaller the fitness function is, and the more likely it is to achieve the desired effect.

[0038] In one embodiment of the present invention, in step 5, the whale optimization algorithm includes:

[0039] (1) Surrounding prey: Humpback whales are able to identify the location of prey and surround them. Since the location of the optimal design in the search space is unknown a priori, WOA assumes that the current best candidate solution is the target prey or close to the optimal one. After defining the best search agent, other search agents will try to update their positions to the best search agent, which can be expressed as the following formula:

[0040]

[0041] Where: Represents the intermediate variables of the calculation process; t represents the current iteration, and is the coefficient vector, is the position vector of the best solution obtained so far, is the position vector, || is the absolute value, and · is the element-wise multiplication; where and The expression is as follows:

[0042]

[0043] Where: It means that it decreases linearly from 2 to 0 during the iteration process. The expression is: i represents the current number of iterations, and M is the maximum number of iterations; is a random vector in [0,1];

[0044] (2) Bubble net attack mode: Two mathematical models are established based on the humpback whale bubble net foraging behavior:

[0045] (2.1) Shrinkage and encirclement mechanism: by reducing Specifically, if the random number A is set between [-1, 1], the search position of the new whale group individual is defined as any position between the current whale group individual position and the best whale group individual position.

[0046] (2.2) Spiral Update Mechanism: First, the distance between the whale population position and the target prey position is determined. Then, a spiral equation is created between the positions of the whale and the prey to simulate the spiral movement of the humpback whale. The specific formula is as follows:

[0047]

[0048] Where: represents the distance from the i-th whale to its prey; b is a constant that defines the shape of the logarithmic spiral, l is a random number in [-1,1]; · is element-by-element multiplication;

[0049] Humpback whales swim around their prey in a shrinking circle and simultaneously along a spiral path; to model this simultaneous behavior, a 50% probability is assumed to be chosen between the shrinking surround mechanism or the spiral model to update the whale's position during optimization, as expressed below:

[0050]

[0051] Where: p is a random number between [0,1];

[0052] (3) Prey search mechanism: In addition to the bubble net attack mode, humpback whale groups will randomly search for prey based on each other's positions; using random values ​​greater than 1 or less than -1 To force the search agent to move away from the reference whale and explore more suitable prey, enhance the algorithm's search and predation capabilities, and enable WOA to perform global search. The entire process is expressed as follows:

[0053]

[0054] Where: is a random position vector selected from the current population;

[0055] (4) Setting the upper and lower limits of the target and other related coefficients: After completing the whale optimization algorithm, set the upper and lower limits of the learning rate, regularization coefficient, and learning rate reduction factor. The number of iterations is set to 200. The coefficients of the three are determined by taking the minimum value of the fitness function as the target, and the coefficients are brought into the CNN-GRU network to achieve accurate monitoring of sub-module capacitor aging.

[0056] Compared with the existing technology, the present invention has the following beneficial effects: the present invention provides a modular multi-level converter sub-module capacitance monitoring method based on an improved convolutional neural network, which can improve the accuracy of sub-module capacitance aging monitoring in MMC, and solve a series of problems such as sub-module capacitance aging when MMC operates at a low frequency stage, which in turn leads to a decline in power quality or even system paralysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is the equivalent circuit diagram of the modular multilevel converter.

[0058] Figure 2 This is the network structure diagram of the convolutional neural network-gated recurrent unit (CNN-GRU).

[0059] Figure 3 This is the flow chart of the CNN-GRU neural network.

[0060] Figure 4 This is the overall fault diagnosis flow chart. DETAILED DESCRIPTION

[0061] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0062] The present invention provides a method for monitoring capacitor faults in submodules of a modular multi-level converter based on an improved convolutional neural network, comprising:

[0063] According to the characteristics of the capacitor voltage of the modular multilevel converter submodule, the corresponding aging characteristics are extracted for the submodule capacitor voltage under aging conditions;

[0064] Construct a convolutional neural network combined with a gated recurrent unit (CNN)-GRU network, and use the aging capacitor label to determine the accuracy of the constructed network.

[0065] Taking accuracy as the objective function, the whale optimization algorithm (WOA) is used to optimize the learning rate, regularization coefficient, and learning rate reduction factor in the CNN-GRU network to achieve accurate monitoring of sub-module capacitor aging.

[0066] The following is a specific implementation process of the present invention.

[0067] like Figure 4As shown, the present invention provides a method for monitoring capacitor faults in a modular multi-level converter submodule based on an improved convolutional neural network, comprising the following steps:

[0068] Step 1: Preliminary preprocessing of data. Preprocess the characteristic data of the submodule capacitor to be processed ( Figure 1 The figure shows the equivalent circuit diagram of a modular multilevel converter. (The data here mainly consists of wavelet decomposition features, the maximum value of the differential capacitor voltage, and the periodic integral value of the differential capacitor voltage.) The number of categories, feature dimensions, and samples of the data to be processed are determined, and the dataset is shuffled. The corresponding training and test sets are determined by dividing the dataset using a proportional coefficient.

[0069] Step 2: Perform secondary processing on the pre-processed data. Normalize the pre-processed data, flatten the data (here, flatten the data into 1D data), and complete the corresponding data format conversion.

[0070] Step 3: Establish a neural network model. Figure 2 、 3 As shown in the figure, the network model structure here mainly includes sequence folding / unfolding layer, input layer, convolution layer, batch normalization layer, pooling layer, GRU layer, fully connected layer and output layer. This article mainly describes the convolution layer, batch normalization layer, pooling layer, GRU layer and fully connected layer in the CNN-GRU network.

[0071] Content 1: Convolutional Layer. A convolutional layer consists of multiple feature planes, each of which is composed of multiple neurons. Each neuron is connected to a local region of the feature plane in the previous layer through a convolution kernel. The convolution kernel is a weight matrix. The convolution layer of a CNN extracts different features of the input through convolution operations. The first convolution layer extracts low-level features such as edges, lines, and corners, while higher convolution layers extract higher-level features. Two-dimensional and three-dimensional CNNs can be expanded in this way. The entire convolution process can be described by the following formula:

[0072]

[0073] Where: Represents the output value of the vth neuron in the output feature surface u in the convolutional layer; Represents the output value of the nth neuron of its input feature surface m; ω a(i)b(j) represents the connection weight between the i-th neuron of the input feature surface a and the j-th neuron of the output feature surface b; b i represents the bias of the output feature surface i; f covReLU stands for activation function. Generally, saturated nonlinear functions, such as the Sigmoid function, are used as activation functions. However, unsaturated nonlinear functions can alleviate the exploding / vanishing gradient problem to a certain extent and accelerate convergence. Therefore, the ReLU function is used as the activation function in the convolution part of this CNN structure.

[0074] Content 2: Batch Normalization Layer (BN Layer). Batch normalization is often performed between convolution and ReLU activation functions. By forcibly pulling the distribution of the input value of any neuron in each layer of the neural network back to a standard normal distribution with a mean of 0 and a variance (standard deviation) of 1, the input distribution of each layer of the network is made more stable. The corresponding expression is as follows:

[0075]

[0076]

[0077] Where: x i Represents the output result of the previous layer (convolutional layer); m is the size of batch normalization; μ n is the mean of the batch number; σ n is the standard deviation after batch data processing; ε is a small value to prevent the denominator from becoming 0; represents the output after batch normalization; the last equation represents the scaling and translation transformations performed on the normalized data. γ and β are parameters. Initially, γ = 1 and β = 0, and are then adjusted to appropriate values ​​through learning. Batch normalization addresses the issue of shifting data distribution in intermediate layers during training, preventing vanishing or exploding gradients and accelerating training.

[0078] Content 3: Pooling Layer. The input to the pooling layer comes from multiple feature maps in the convolutional layer. Each feature map in the pooling layer corresponds to a feature map in the convolutional layer, and the number of feature maps remains constant. The neurons in the pooling layer are connected to the local receptive fields in the input layer, so that each neuron processes only a small portion of the input data, and the input areas of different neurons do not overlap. The main purpose of the pooling layer is to reduce the resolution of the feature map, thereby extracting spatially invariant features. It is equivalent to performing secondary feature extraction on the convolutional layer output.

[0079] Generally, pooling includes maximum pooling and average pooling, among which maximum pooling has the best effect and is widely used. Its formula is as follows:

[0080] m i(x,b) =max (b-1)ω＜t＜bω {o i(x,t)}b∈1,2,...,p

[0081] Among them:i(x,t) represents the t-th neuron of the x-th map in the i-th layer; ω represents the width of the convolution kernel; b represents the b-th pooling kernel, and there are p in total.

[0082] Content 4: Gated Recurrent Neural Network Unit Layer (GRU Layer). GRU has a similar network structure to LSTM, but compared to the disadvantage of LSTM's complex internal structure that leads to longer training time, GRU can ensure that the training parameters are reduced while also ensuring prediction accuracy. GRU mainly includes reset gate r and update gate z to control the flow of information. It can learn to retain important information and discard irrelevant information. Specifically, the reset gate r is used to adjust the combination of the previous memory and the new input, and the update gate is used to determine the amount of previous memory to retain. Therefore, GRU can make full use of the useful information of long-chain sequence input for prediction and avoid long-term dependency problems. The relevant formulas for GRU are as follows:

[0083] z t =σ(W z x t +U z h t-1 )

[0084] r t =σ(W t x t +U t h t-1 )

[0085]

[0086] Where: σ(.) represents the activation function, usually the Sigmoid function; z t Represents the value of the update gate in the current state; r t It means resetting the gate value in the current state; is the cell value, which aims to discard irrelevant historical information and represents the input x t and the past hidden state h t-1 Summary of h t represents the output of the hidden layer; U z , W z , U t , W t , U and W are trainable parameter matrices. It should be noted that the model parameters including U and W are shared in each time step, which reduces the total number of training parameters. The specific forms of the Sigmoid function and the tanh function are as follows:

[0087]

[0088] Content 5: Fully connected layer. In the CNN structure, after multiple convolutional layers and pooling layers, one or more fully connected layers are usually connected. In the fully connected layer, each neuron is fully connected to all neurons in the previous layer. Such a design can integrate local information with category distinction capabilities in the convolutional layer or pooling layer. In order to improve the performance of the CNN network, the ReLU function is usually used as the activation function of each neuron in the fully connected layer. The output value of the last fully connected layer will be passed to an output layer, which usually uses Softmax logistic regression for classification. This layer can also be called the Softmax layer. The Softmax function corresponds to the following:

[0089]

[0090] Where: As the input vector of the function, (x0,x1,...,x k ) is composed of; and x i The elements of the input vector of the Softmax function can take any real value; as a standard exponential function, applied to each element of the input vector; while the formula is a normalization term that ensures that all output values ​​of the function sum to 1 and that each value lies in the range (0,1), thus constituting a valid probability distribution.

[0091] Step 4: Set unknown parameters and build a fitness function. After building a rough CNN-GRU neural network, you need to set its parameters. Manual parameter adjustment often requires considering many factors, and the debugging process will be too lengthy. Therefore, in order to quickly determine the corresponding parameters and achieve effective capacitor fault monitoring, the parameters to be optimized this time are the learning rate, regularization parameter, and learning rate reduction factor. The fitness function is based on the accuracy of the training set. The fitness function related to the accuracy of the training set is constructed. The specific formula can be as follows:

[0092]

[0093] In this fitness function, Represents the accuracy of the training set, T_sim1 represents the result of the training set obtained after training; T_train represents the original result of the training set; length(T_train) represents the number of training sets in this training.

[0094] On the whole, The larger it is, the smaller the fitness function is, and the more likely it is to achieve the desired effect.

[0095] Step 5: Use the whale optimization algorithm and set the relevant parameters for optimization. This algorithm imitates the humpback whale's "spiral bubble net" strategy, employing a shrinking encirclement, spiral position updates, and random hunting mechanisms during foraging. It features a simple structure, minimal parameter adjustments, fast convergence, and powerful global optimization capabilities.

[0096] The algorithm mainly includes three major contents: surrounding prey, bubble net attack and searching for prey.

[0097] Content 1: Surrounding Prey. Humpback whales can identify the location of prey and surround them. Since the location of the optimal design in the search space is unknown a priori, the WOA algorithm assumes that the current best candidate solution is the target prey or is close to optimal. After the optimal search agent is defined, the other search agents will attempt to update their positions to match the optimal search agent. This behavior is expressed by the following formula:

[0098]

[0099] Where: Represents the intermediate variables of the calculation process; t represents the current iteration, and is the coefficient vector, is the position vector of the best solution obtained so far, is the position vector, || is the absolute value, and · is the element-by-element multiplication. It is worth mentioning here that if there is a better solution, it should be updated in each iteration. and The expression is as follows:

[0100]

[0101] Where: It decreases linearly from 2 to 0 during the iteration process, and the expression is: i represents the current number of iterations, and M is the maximum number of iterations; is a random vector in [0,1].

[0102] Content 2: Bubble Net Attack Mode. Two mathematical models were established based on the humpback whale bubble net foraging behavior:

[0103] 1. Shrinkage mechanism: This behavior is achieved by reducing This is achieved by adjusting the values ​​in . It should be noted that as a decreases, the range of A also decreases. Specifically, if the random number A is set between [-1, 1], then the search position of the new whale group individual can be defined as any position between the current whale group individual position and the best whale group individual position.

[0104] 2. Spiral Update Mechanism: This method first determines the distance between the whale population location and the target prey location. It then creates a spiral equation between the whale and prey locations to simulate the spiral movement of humpback whales. The specific formula is as follows:

[0105]

[0106] Where: represents the distance from the i-th whale to its prey; b is a constant that defines the shape of the logarithmic spiral, l is a random number in [-1, 1]; · is element-by-element multiplication.

[0107] It is noteworthy that humpback whales swim around their prey in a shrinking circle and along a spiral path at the same time. To model this simultaneous behavior, we assume a 50% probability of choosing between the shrinking and surrounding mechanism or the spiral model to update the whale's position during optimization. The specific expression is as follows:

[0108]

[0109] Where: p is a random number between [0,1].

[0110] Content 3: Prey search mechanism. In addition to the bubble net attack mode, humpback whale groups will randomly search for prey based on each other's positions. Use random values ​​greater than 1 or less than -1 This forces the search agent to move away from the reference whale and explore more suitable prey, thereby enhancing the algorithm's search and predation capabilities and enabling WOA to conduct global search. The expression under this process is as follows:

[0111]

[0112] Where: is a random position vector selected from the current population.

[0113] Content 4: Setting target upper and lower limits and other related coefficients. After completing the above-mentioned whale optimization algorithm, set the upper and lower limits of the learning rate, regularization coefficient, and learning rate reduction factor. The number of iterations is set to 200. The coefficients of the three are determined by minimizing the fitness function as the target. The coefficients are then introduced into the CNN-GRU network to achieve higher accuracy.

[0114] The above are preferred embodiments of the present invention. Any changes made according to the technical solution of the present invention, as long as the resulting functions and effects do not exceed the scope of the technical solution of the present invention, shall fall within the scope of protection of the present invention.

Claims

1. A method for monitoring capacitor faults in modular multi-level converter submodules based on an improved convolutional neural network, characterized in that: include: According to the characteristics of the capacitor voltage of the modular multilevel converter submodule, the corresponding aging characteristics are extracted for the submodule capacitor voltage under aging conditions; Construct a convolutional neural network combined with a gated recurrent unit (CNN)-GRU network, and use the aging capacitor label to determine the accuracy of the constructed network. Taking accuracy as the objective function, the Whale Optimization Algorithm (WOA) is used to optimize the learning rate, regularization coefficient, and learning rate reduction factor in the CNN-GRU network to achieve accurate monitoring of sub-module capacitor aging. The steps to implement this method are as follows: Step 1: Preliminary data preprocessing: Preprocess the feature data of the modular multilevel converter submodule capacitors to be processed; determine the number of categories, feature dimensions, and sample size of the feature data to be processed, and shuffle the dataset; and divide the dataset into proportional coefficients to determine the corresponding training and test sets. Step 2: Perform secondary processing on the preprocessed data: normalize the preprocessed data, flatten the data after processing, and complete the corresponding data format conversion; Step 3: Establish a convolutional neural network combined with a gated recurrent unit (CNN)-GRU network: The structure of the convolutional neural network combined with a gated recurrent unit (CNN)-GRU network includes a sequence folding / unfolding layer, an input layer, a convolutional layer, a batch normalization layer, a pooling layer, a GRU layer, a fully connected layer, and an output layer. Step 4: Set parameters and build a fitness function: After building a convolutional neural network combined with a gated recurrent unit (CNN-GRU) network, determine that the parameters to be optimized are the learning rate, regularization parameter, and learning rate reduction factor, and build a fitness function. Step 5: Use the whale optimization algorithm and set relevant parameters to optimize the convolutional neural network combined with the gated recurrent unit CNN-GRU network; In step 2, the characteristic data of the capacitors of the modular multilevel converter submodules include wavelet decomposition characteristics, the maximum value of the differential capacitor voltage, and the periodic integral value of the differential capacitor voltage; In step 4, the fitness function formula is as follows: In the fitness function, represents the accuracy of the training set, and T_sim1 represents the result of the training set obtained after training; T_train represents the original result of the training set; length(T_train) represents the number of training sets for this training; Overall, The larger it is, the smaller the fitness function is, and the more likely it is to achieve the desired effect; In step 5, the whale optimization algorithm includes: (1) Surrounding prey: Humpback whales are able to identify the location of prey and surround them. Since the location of the optimal design in the search space is unknown a priori, WOA assumes that the current best candidate solution is the target prey or close to the optimal one. After defining the best search agent, other search agents will try to update their positions to the best search agent, which can be expressed as the following formula: Where: Represents the intermediate variables of the calculation process; t represents the current iteration, and is the coefficient vector, is the position vector of the best solution obtained so far, is the position vector, || is the absolute value, and · is the element-wise multiplication; where and The expression is as follows: Where: It means that it decreases linearly from 2 to 0 during the iteration process. The expression is: i represents the current number of iterations, and M is the maximum number of iterations; is a random vector in [0,1]; (2) Bubble net attack mode: Two mathematical models are established based on the humpback whale bubble net foraging behavior: (2.1) Shrinkage and encirclement mechanism: by reducing Specifically, if the random number A is set between [-1, 1], the search position of the new whale group individual is defined as any position between the current whale group individual position and the best whale group individual position. (2.2) Spiral Update Mechanism: First, the distance between the whale population position and the target prey position is determined. Then, a spiral equation is created between the positions of the whale and the prey to simulate the spiral movement of the humpback whale. The specific formula is as follows: Where: represents the distance from the i-th whale to its prey; b is a constant that defines the shape of the logarithmic spiral, l is a random number in [-1,1]; · is element-by-element multiplication; Humpback whales swim around their prey in a shrinking circle and simultaneously along a spiral path; to model this simultaneous behavior, a 50% probability is assumed to be chosen between the shrinking surround mechanism or the spiral model to update the whale's position during optimization, as expressed below: Where: p is a random number between [0,1]; (3) Prey search mechanism: In addition to the bubble net attack mode, humpback whale groups will randomly search for prey based on each other's positions; using random values ​​greater than 1 or less than -1 To force the search agent to move away from the reference whale and explore more suitable prey, enhance the algorithm's search and predation capabilities, and enable WOA to perform global search. The entire process is expressed as follows: Where: is a random position vector selected from the current population; (4) Setting the upper and lower limits of the target and other related coefficients: After completing the whale optimization algorithm, set the upper and lower limits of the learning rate, regularization coefficient, and learning rate reduction factor. The number of iterations is set to 200. The coefficients of the three are determined by taking the minimum value of the fitness function as the target, and the coefficients are brought into the CNN-GRU network to achieve accurate monitoring of sub-module capacitor aging.

2. The method for monitoring capacitor faults in modular multi-level converter submodules based on an improved convolutional neural network according to claim 1, wherein: In step 3, the convolution layer is composed of multiple feature surfaces, each of which is composed of multiple neurons. Each neuron in each feature surface is connected to the local area of ​​the feature surface in the previous layer through a convolution kernel. The convolution kernel is a weight matrix. The convolution layer extracts different features of the input through convolution operations. The first convolution layer extracts low-level features, and the higher convolution layers extract higher-level features. The entire convolution process is based on the following formula: Where: Represents the output value of the vth neuron in the output feature surface u in the convolutional layer; Represents the output value of the nth neuron in the input feature surface m; ω a(i)b(j) represents the connection weight between the i-th neuron in the input feature plane a and the j-th neuron in the output feature plane b; b i represents the bias of the output feature surface i; f cov represents the activation function.

3. The method for monitoring capacitor faults in modular multi-level converter submodules based on an improved convolutional neural network according to claim 1, wherein: In step 3, the batch normalization layer makes the input distribution of each layer of the network more stable by forcibly pulling the distribution of the input value of any neuron in each layer of the neural network back to a standard normal distribution with a mean of 0 and a variance or standard deviation of 1. The corresponding related expressions are as follows: Where: x i represents the output of the convolutional layer; m is the size of batch normalization; μ n is the mean of the batch number; σ n is the standard deviation after batch data processing; ε is a small value that prevents the denominator from becoming zero; Represents the output result after batch normalization; γ and β are the parameters of the transformation that scales and translates the normalized data. Initially, γ = 1 and β = 0. γ and β can be adjusted through learning.

4. The method for monitoring capacitor faults in modular multi-level converter submodules based on an improved convolutional neural network according to claim 1, wherein: In step 3, the input of the pooling layer comes from multiple feature maps in the convolutional layer. The feature maps in the pooling layer correspond one-to-one with the feature maps in the convolutional layer, and the number of feature maps remains unchanged. The neurons in the pooling layer are connected to the local receptive fields in the input layer, so that each neuron only processes a small part of the input data, and the input areas of different neurons do not overlap. Pooling includes maximum pooling and average pooling. The maximum pooling formula is as follows: m i(x,b) =max (b-1)ω＜t＜bω {o i(x,t) }b∈1,2,...,p Among them: i(x,t) represents the t-th neuron of the x-th map in the i-th layer; ω represents the width of the convolution kernel; b represents the b-th pooling kernel, and there are p in total.

5. The method for monitoring capacitor faults in modular multi-level converter submodules based on an improved convolutional neural network according to claim 1, wherein: In step 3, the GRU layer includes a reset gate r and an update gate z to control the information flow, which can learn to retain important information and discard irrelevant information. Specifically, the reset gate r is used to adjust the combination of the previous memory and the new input, and the update gate is used to determine the amount of previous memory to retain. The relevant formula of GRU is as follows: z t =σ(W z x t +U z h t-1 ) r t =σ(W t x t +U t h t-1 ) Where: σ(.) represents the activation function; z t Represents the value of the update gate in the current state; r t It means resetting the gate value in the current state; is the cell value, which aims to discard irrelevant historical information and represents the input x t and the past hidden state h t-1 Summary of h t represents the output of the hidden layer; U z 、W z 、U t 、W t , U and W are trainable parameter matrices, and the model parameters including U and W are shared in each time step.

6. The method for monitoring capacitor faults in modular multi-level converter submodules based on an improved convolutional neural network according to claim 1, wherein: In step 3, there are one or more fully connected layers. In the fully connected layer, each neuron is fully connected to all neurons in the previous layer. The output of the last fully connected layer is passed to an output layer, which is classified using Softmax logistic regression. This is called the Softmax layer. The Softmax function corresponds to the following: Where: As the input vector of the Softmax function, (x0,x1,...,x k ) composition; x i The elements of the input vector of the Softmax function can take any real value; As a standard exponential function, applied to each element of the input vector; the formula is the normalization term, which ensures that all output values ​​of the function sum to 1 and that each value is in the range (0,1) to constitute a valid probability distribution.

Images