Modular multilevel converter capacitor monitoring method based on improved neural network

Abstract

The application discloses a capacitor monitoring method for a modular multilevel converter based on an improved neural network. The capacitor monitoring method comprises the following steps: building a modular multilevel converter, selecting circuit information to be acquired, and performing fault feature extraction, neural network model building and training. The capacitor monitoring method is used for diagnosing capacitor faults in the modular multilevel converter based on an improved neural network model. Genetic algorithm is added to select initial values of parameters on the basis of a traditional neural network, and an analog annealing algorithm is added to optimize the parameters. The method selects switch signals in a sub-module, bridge arm currents on a bridge arm and direct current voltages as data, processes the acquired data, obtains a two-fold fundamental frequency component of the bridge arm current through a filter, extracts a low-frequency part energy feature of the component through a wavelet packet decomposition method, realizes health monitoring of the capacitor of the module, and ensures stable operation of the modular multilevel converter system.

Description

Technical Field

[0001] This invention relates to the field of power electronics fault diagnosis, specifically to a modular multilevel converter capacitor monitoring method based on an improved neural network. Background Technology

[0002] Against the backdrop of energy and environmental issues, new energy power generation and distributed generation have gradually become issues of widespread concern. Due to their inherent characteristics, this type of power generation requires flexible transmission and transformation technology to address its intermittent nature. Modular multilevel converters are often the preferred choice for flexible DC transmission systems, offering advantages such as low stress on switching devices, high voltage and current waveform quality, and flexible output voltage regulation.

[0003] Modular multilevel converters integrate a large number of submodules, each containing one or more module capacitors. As the capacitance increases over time, the decrease in capacitance value significantly affects the reliable operation of the system. Therefore, it is necessary to monitor the health of the module capacitors. Summary of the Invention

[0004] The purpose of this invention is to provide a capacitor monitoring method for modular multilevel converters based on an improved neural network. By extracting features that reflect the fault and incorporating wavelet packet decomposition and signal integration methods, the accuracy of capacitor monitoring is improved and the computational complexity is simplified, based on the existing capacitor fault diagnosis method for modular multilevel converters.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A modular multilevel converter capacitance monitoring method based on an improved neural network, the capacitance monitoring method comprising the following steps:

[0007] Step 1: Build a modular multilevel converter, select the circuit information to be acquired, and monitor the capacitors and corresponding circuit information data.

[0008] Step 2: Extract fault features from the collected data using signal integration and wavelet packet decomposition methods.

[0009] Step 3: Create a dataset and divide it into training, validation, and test sets.

[0010] Step 4: Establish an improved neural network model, determine the number of neurons, loss function, activation function, add a genetic algorithm to optimize the initial values ​​of the neural network parameters, and use simulated annealing algorithm for backpropagation to adjust the parameters.

[0011] Step 5: Train the established model using the training set, evaluate the model using the validation set, and then test the model's accuracy using the test set.

[0012] Furthermore, the steps of the genetic algorithm are as follows:

[0013] Step 1: Chromosomes are encoded in binary. First, determine the length of the chromosome based on the number of neural network parameters, set the population size, maximum number of generations, crossover and mutation probabilities, and generate the population size.

[0014] Step 2: Select the neural network prediction output error as the fitness function and calculate the individual fitness.

[0015] Step 3: Select individuals with high fitness from the population to form a new population. Perform crossover operation on every two chromosomes in the new population according to the crossover probability. The offspring generated will mutate according to the mutation probability.

[0016] Step 4: Repeat steps 2 and 3 until the maximum number of generations is reached to find the chromosome with the highest fitness. Decode the chromosome to obtain the optimal weights and threshold parameters and assign them to the neural network.

[0017] Furthermore, the simulated annealing algorithm sets a temperature parameter T, which increases by 0.9 times with each iteration. During the backpropagation of the neural network, when the loss function increases compared to the previous iteration, a judgment is made. A random number between 0 and 1 is generated and compared with the probability P. If the value is greater than the probability P, the parameter is adjusted in the direction of the increasing gradient of the loss function; otherwise, the parameter is adjusted in the direction of the decreasing gradient of the loss function. The probability function is calculated as follows:

[0018]

[0019] Where loss(n) is the loss function for the nth iteration.

[0020] Furthermore, the circuit information to be selected is information selected specifically for the modular multilevel converter. The circuit information to be selected includes the switching signals of the switching transistors in the submodule, the bridge arm current and DC voltage on the bridge arm.

[0021] Furthermore, the fault feature extraction integrates the switching signal and adds the proportion of the module switching signal on the corresponding bridge arm. The twice-fundamental frequency component of the bridge arm current is obtained through a filter, and the low-frequency energy features of this component are extracted by combining wavelet packet decomposition.

[0022] Furthermore, the wavelet packet decomposition decomposes the second harmonic component of the bridge arm current:

[0023]

[0024]

[0025] Where a is the scaling function, b is the frequency, h and g are the coefficients of the low-pass filter and the high-pass filter, respectively, and l is the position coefficient. The current is decomposed into 6-level wavelet packets using the db4 wavelet to obtain the corresponding wavelet packet coefficient matrix.

[0026] This indicates the degree of fit between the wavelet basis at the node and the input signal in that frequency band. The energy information of the corresponding frequency band for each node is calculated using the following formula:

[0027]

[0028] Where E(i) represents the energy of the i-th node, p i,j Let be the coefficient of the j-th node in the i-th node, and k be the total number of node coefficients in the i-th node.

[0029] Furthermore, the loss function uses the root mean square error, which is the square root of the sum of the squares of the errors corresponding to the predicted data and the original data.

[0030]

[0031] Where y pre y label Let n represent the predicted data and the original data, respectively, and n represents the number of samples for calculating the root mean square error.

[0032] The beneficial effects of this invention are:

[0033] 1. The capacitance monitoring method of this invention incorporates genetic algorithms and simulated annealing strategies during the model training process, thereby improving the accuracy of the model;

[0034] 2. The capacitance monitoring method of the present invention can monitor the health status of capacitors in the circuit online to ensure the stable operation of the modular multilevel converter;

[0035] 3. The capacitance monitoring method of this invention applies intelligent algorithms to the traditional field of fault diagnosis, providing a new diagnostic method for this type of fault problem. Attached Figure Description

[0036] The invention will now be further described with reference to the accompanying drawings.

[0037] Figure 1 This is a flowchart of the modular multilevel converter capacitor fault diagnosis method of the present invention;

[0038] Figure 2 Wavelet waveform diagram using db4 wavelet basis as an example;

[0039] Figure 3 This is a diagram of the neural network structure of the present invention;

[0040] Figure 4This is a graph of activation functions in the neural network of this invention;

[0041] Figure 5 This is a flowchart illustrating the improved neural network of this invention. Detailed Implementation

[0042] 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, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] A modular multilevel converter capacitance monitoring method based on an improved neural network, such as... Figure 1 As shown, the capacitance monitoring method includes the following steps:

[0044] Step 1: Building a modular multilevel converter

[0045] First, a modular multilevel converter is built. Then, the circuit information to be acquired is selected, and the capacitor and corresponding circuit information data are monitored. The circuit information to be acquired includes the switching signals of the switching transistors in the modular multilevel converter sub-modules, the bridge arm current, and the DC bus voltage. The circuit information is collected in one cycle.

[0046] Step 2: Fault Feature Extraction

[0047] Feature extraction is performed on the collected data. First, since the DC bus voltage is relatively stable and basically unchanged, the average value within one cycle is directly calculated.

[0048] The switching signal is processed by integrating the signal according to the acquisition cycle, converting the original pulse signal into a value. Based on the relationship between the capacitance value, the bridge arm current, and the module voltage, and due to the voltage equalization control, the value differs when the capacitance values ​​of different modules are inconsistent. Further removing the center value can further increase the difference. This feature describes the difference between modules on the entire bridge arm. In order to determine which sub-module the calculated capacitance belongs to, the proportion of the integral value of that module in the integral value of the entire bridge arm is added.

[0049] For processing the bridge arm current, a bandpass filter is first applied to obtain the current component of the bridge arm current with respect to twice the fundamental frequency of the reference voltage frequency. Wavelet packet decomposition is then performed on this component to extract the low-frequency energy. The db4 wavelet function is selected as the wavelet basis, and its image is shown below. Figure 2 As shown, the input current signal undergoes a 6-level wavelet packet decomposition. The specific steps are as follows:

[0050]

[0051]

[0052] Where a is the scaling function, b is the frequency, h and g are the coefficients of the low-pass and high-pass filters, respectively, and l is the position coefficient, the corresponding wavelet packet coefficient matrix is ​​obtained, which represents the degree of fit between the wavelet basis and the input signal at the nodes in that frequency band. The energy information of each node in the corresponding frequency band is calculated using the following formula:

[0053]

[0054] Where E(i) represents the energy of the i-th node, p i,j Let be the coefficient of the j-th node in the i-th node, and k be the total number of node coefficients in the i-th node.

[0055] After decentralizing the energy of the first five frequency bands, this characteristic is used as a capacitor fault feature. When the module capacitance changes, the trend of this feature changes significantly.

[0056] Step 3: Creating the dataset

[0057] To split the dataset, add actual capacitance value labels to the data extracted in the second step, randomly shuffle the dataset order, and then divide it into training and validation sets in a 7:3 ratio.

[0058] Step 4: Building the Neural Network

[0059] A regression model using a neural network was constructed and trained using the extracted features. The neural network structure is shown in the attached figure. Figure 3 As shown, the model structure is first determined, consisting of an input layer, hidden layers, and an output layer. Each layer has tens to hundreds of neurons. Neurons within the same layer are not interconnected, while neural networks in adjacent layers are fully connected through weights and thresholds. The model uses a single hidden layer, and the activation function for the hidden layer is the sigmoid function. The formula for the sigmoid function is:

[0060]

[0061] Where S(x) is the output value after activation, and x is the value input to the activation function after the hidden layer. Figure 4 This is the graph of the function.

[0062] The input dimension of the input layer is related to the number of sub-modules in the modular multilevel converter. The input data is the DC bus voltage. Based on the bridge arm current energy extracted in the second step, the integral of the switching signal, and the proportion of the integral of the switching signal, 90 neurons are set in the hidden layer, and the root mean square error is used as the loss function of the model.

[0063]

[0064] Where y pre y label Let n represent the predicted data and the original data, respectively, and n represents the number of samples for calculating the root mean square error.

[0065] Step 5: Training the model

[0066] The improved neural network model training process consists of two parts: the selection of initial values ​​for model parameters and improvements to backpropagation. The overall model training process is as follows: Figure 5 As shown.

[0067] The structure of the improved neural network is optimized by using a genetic algorithm to optimize the initial values ​​of the weights and thresholds. The specific steps for solving the optimal initial values ​​of the neural network are as follows:

[0068] Step 1: First, determine the network structure. Chromosomes are encoded in binary. Calculate the chromosome length based on the number of neural network structure parameters. Randomly generate the initial population size, maximum number of generations, crossover and mutation probabilities, and generate the population size.

[0069] Step 2: Select the neural network prediction output error as the fitness function and calculate the individual fitness;

[0070] Step 3: Select individuals with high fitness from the population to form a new population. Perform crossover operation on every two chromosomes in the new population according to the crossover probability. The offspring generated will mutate according to the mutation probability.

[0071] Step 4: Repeat steps 2 and 3 until the maximum number of generations is reached, find the chromosome with the highest fitness, decode it to obtain the optimal weights and threshold parameters, and assign them to the neural network.

[0072] Next, the model training forward propagation results are used to obtain the predicted output value. The root mean square value of the predicted output value is calculated with the value in the actual sample to obtain the error as the loss function. The loss function generated in each iteration is compared with the loss function of the previous iteration. If the loss decreases, the parameters are optimized in the direction of gradient descent. If the loss increases, simulated annealing is entered to determine whether annealing is necessary.

[0073] In simulated annealing, the loss functions of two iterations are compared. A new temperature parameter T is first set, which decreases linearly with each iteration, decreasing to a factor of 0.9 each time. During the backpropagation of the neural network, when the loss function increases compared to the previous iteration, a judgment is made, and a number randomly generated from 0 to 1 is compared with the annealing probability function P. The probability function is calculated as follows:

[0074]

[0075] Where P is the probability, and loss(n+1) and loss(n) are the loss functions before and after the (n+1)th iteration forward propagation, respectively. When the generated random number is greater than the probability function, the parameters are optimized in the direction of gradient ascent; otherwise, the parameters are optimized in the direction of gradient descent.

[0076] Training ends when the maximum number of iterations is reached or the model error meets the set conditions. The training results of the established model are then validated using a validation set to evaluate the model's generalization ability. Finally, the model accuracy is calculated using a test set.

[0077] Step 4: After training, the model is added to the control of the modular multilevel converter, which enables online monitoring of the health of its module capacitors.

[0078] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0079] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A modular multilevel converter capacitance monitoring method based on an improved neural network, characterized in that, The capacitance monitoring method includes the following steps: Step 1: Build a modular multilevel converter, select the circuit information to be acquired, monitor the capacitors and corresponding circuit information data; Step 2: Extract fault features from the collected data using signal integration and wavelet packet decomposition methods; Step 3: Create the dataset and divide it into training, validation, and test sets; Step 4: Establish an improved neural network model, determine the number of neurons, loss function, activation function, add a genetic algorithm to optimize the initial values ​​of the neural network parameters, and use simulated annealing algorithm for backpropagation to adjust the parameters; Step 5: Train the established model using the training set, evaluate the model using the validation set, and then test the model's accuracy using the test set. The fault feature extraction process involves integrating the switching signal and adding the proportion of the module switching signal on the corresponding bridge arm. The process then uses a filter to obtain the fundamental frequency component of the bridge arm current, and combines wavelet packet decomposition to extract the low-frequency energy features of this component.

2. The modular multilevel converter capacitance monitoring method based on improved neural network according to claim 1, characterized in that, The steps of the genetic algorithm are as follows: Step 1: Chromosomes are encoded in binary. First, determine the length of the chromosome based on the number of neural network parameters, set the population size, maximum number of generations, crossover and mutation probabilities, and generate the population size. Step 2: Select the neural network prediction output error as the fitness function and calculate the individual fitness; Step 3: Select individuals with high fitness from the population to form a new population. Perform crossover operation on every two chromosomes in the new population according to the crossover probability. The offspring generated will mutate according to the mutation probability. Step 4: Repeat steps 2 and 3 until the maximum number of generations is reached to find the chromosome with the highest fitness. Decode the chromosome to obtain the optimal weights and threshold parameters and assign them to the neural network.

3. The modular multilevel converter capacitance monitoring method based on improved neural network according to claim 1, characterized in that, The simulated annealing algorithm sets a temperature parameter T, which increases by 0.9 times with each iteration. During the backpropagation of the neural network, when the loss function increases compared to the previous iteration, a judgment is made. A randomly generated number between 0 and 1 is compared with the probability P. If the number is greater than the probability P, the parameter is adjusted in the direction of the increasing gradient of the loss function; otherwise, the parameter is adjusted in the direction of the decreasing gradient of the loss function. The probability function is calculated as follows: Where loss(n) is the loss function for the nth iteration.

4. The modular multilevel converter capacitance monitoring method based on an improved neural network according to claim 1, characterized in that, The circuit information to be selected is specifically for the modular multilevel converter. The circuit information to be selected includes the switching signals of the switching transistors in the submodule, the bridge arm current and DC voltage on the bridge arm.

5. The modular multilevel converter capacitance monitoring method based on an improved neural network according to claim 1, characterized in that, The wavelet packet decomposition decomposes the second harmonic component of the bridge arm current: Where a is the scaling function, b is the frequency, h and g are the coefficients of the low-pass filter and the high-pass filter, respectively, and l is the position coefficient. The current is decomposed into 6-level wavelet packets using the db4 wavelet to obtain the corresponding wavelet packet coefficient matrix. This indicates the degree of fit between the wavelet basis at the node and the input signal in that frequency band. The energy information of the corresponding frequency band for each node is calculated using the following formula: Where E(i) represents the energy of the i-th node, p i,j Let be the coefficient of the j-th node in the i-th node, and k be the total number of node coefficients in the i-th node.

6. The modular multilevel converter capacitance monitoring method based on improved neural network according to claim 3, characterized in that, The loss function uses the root mean square error, which is the square root of the sum of the squares of the errors corresponding to the predicted data and the original data. Where y pre y label Let n represent the predicted data and the original data, respectively, and n represents the number of samples for calculating the root mean square error.

Images