Parameter optimization-based electric power system disturbance identification method of fusion neural network

Abstract

The invention relates to the technical field of electric power system disturbance identification, in particular to an electric power system disturbance identification method based on a fusion neural network of parameter optimization, which comprises the following steps: acquiring eight types of disturbance frequency data of an electric power system, and optimizing penalty factors and modal feature vector quantity based on variational modal method decomposition by adopting an improved sparrow search algorithm to obtain eight types of disturbance frequency data of the electric power system; obtaining a target parameter group, obtaining modal feature components selected based on a fitness function, optimizing hyper-parameters of a fusion neural network by adopting a Lundao search algorithm, and distributing feature weights of a second time domain index in combination with a multi-head attention mechanism to obtain a time sequence feature signal; and S400, performing disturbance identification on the time sequence characteristic signal in the step S400 by adopting a convolutional neural network-bidirectional gating cycle unit classifier to obtain a disturbance identification result, and remarkably improving the accuracy of disturbance identification through the method.

Description

Technical Field

[0001] The present invention relates to the technical field of power system disturbance identification, and particularly relates to a power system disturbance identification method based on a parameter-optimized fusion neural network. Background Art

[0002] With the improvement of the penetration rate of new energy and the access of a large number of power electronic devices, the trend of power electronics in the power system is becoming increasingly obvious. Its low anti-interference ability and weak supportability increase the risk rate of faults, resulting in frequent power outage accidents. Through research, it is found that a single fault is prone to cause the collapse of the power grid after a chain reaction, leading to large-scale power outage accidents. Disturbance identification is an effective means of fault warning. Wide-area measurement systems and phasor measurement units have the characteristics of synchronization, rapidity, and accuracy. By real-time monitoring the global operating conditions of the system, disturbances can be quickly and accurately identified, and measures can be taken to ensure the normal operation of the system.

[0003] In the prior art, disturbance classification is divided into data-driven and model-driven. With the rapid development of artificial intelligence technology, new ideas and options are provided for analyzing a large amount of PMU measurement data in aspects such as data mining and feature learning, making the disturbance identification develop towards the direction of digitization and intelligence.

[0004] With the increase in the complexity and scale of the power grid, in the face of the upload of a large amount of data, the difficulty of establishing a power grid model based on topological structure and parameters increases, making the disturbance classification driven by the power grid model have certain limitations. The existing power system disturbance identification based on data-driven mainly includes two parts: feature extraction and disturbance classification. However, when variational mode decomposition is used for feature extraction, modal aliasing or information loss is likely to occur, and the process of calculating the optimal parameter combination is lacking. In addition, the training effect of the hyperparameters of the existing deep learning network is poor, resulting in low accuracy of disturbance identification. Summary of the Invention

[0005] The purpose of the present invention is to provide a power system disturbance identification method based on a parameter-optimized fusion neural network to solve the problem of low accuracy of disturbance identification existing in the prior art.

[0006] To achieve the above purpose, the present invention provides a power system disturbance identification method based on a parameter-optimized fusion neural network, including the following steps: S100. Obtain eight types of disturbance frequency data of the power system generated based on simulation software; S200. Optimize the penalty factor and the number of modal feature vectors decomposed based on the variational mode method by using an improved sparrow search algorithm to obtain a target parameter group; S300. Obtain the modal feature components selected based on the fitness function, and extract the second time-domain indicators of each disturbance frequency data in step S100 based on the modal feature components, the target parameter group, and the first time-domain indicators; S400. Optimize the hyperparameters of the fusion neural network using the transit search algorithm. The hyperparameters at least include the learning rate, the convolution kernel size, and the number of neurons, and allocate the feature weights of the second time-domain indicators in step S300 by combining the multi-head attention mechanism to obtain the time-series feature signal; S500. Use a convolutional neural network - bidirectional gated recurrent unit classifier to perform disturbance identification on the time-series feature signal in step S400 to obtain the disturbance identification result.

[0007] In some embodiments of the present application, in S100, obtaining eight types of disturbance frequency data of the power system generated based on the simulation software includes: Use the PowerFactory software to perform simulations based on eight types of disturbance data, and the eight types of disturbance data include three-phase short circuit, single-phase ground fault, generator output reduction, load connection, load disconnection, line tripping, shunt capacitor connection, and shunt capacitor disconnection data; Set the simulation time to 30 s, the simulation step size to 0.02 s, and the trigger disturbance time to 5 s to obtain eight types of disturbance frequency data output by the PowerFactory software simulation.

[0008] In some embodiments of the present application, in S200, using the improved sparrow search algorithm to optimize the penalty factor and the number of modal feature vectors decomposed based on the variational mode decomposition to obtain the target parameter group includes the following steps: S201. Use the osprey optimization algorithm to replace the position update in the sparrow search algorithm during the exploration phase, simulate the process of the osprey detecting and moving towards the prey at a random position, and the calculation formula is: ; Among them, is the updated position of the x-th osprey in the y-th dimension, is the fish selected by the x-th osprey, is the random number range between 0 and 1, is the random number range between 1 and 2; S202. Use the Cauchy mutation operator to make the iteration of the sparrow algorithm jump out of the local optimal solution, and at the same time increase the iteration scale of the sparrow algorithm. The calculation formula of the Cauchy mutation operator is: ; Among them, is the dimension of the Cauchy mutation, is the calculation formula of the fitness function, M is the Cauchy mutation operator, is a local optimal solution. When occurs, the Cauchy mutation operator enables the iteration of the sparrow algorithm to jump out of the local optimal solution and obtain the global optimal solution. S203. Use the chaotic particle mapping iteration algorithm to enable the population to jump out of the local optimal solution and obtain the global optimal solution. The calculation formula of the chaotic particle mapping iteration algorithm is: ; where is the number of the current iteration, is the position after the chaotic particle iterates times, is the position after the chaotic particle iterates times; S204. Select the fitness function and calculate the mutual information entropy. The calculation formula of the mutual information entropy is: ; where is the joint probability distribution, specifically the joint probability that the random variable A takes the value of and the random variable B takes the value of b , is the marginal probability distribution that the random variable A takes the value of , is the marginal probability distribution that the random variable B takes the value of b , is the mutual information entropy; S205. Obtain the permutation entropy used to represent the complexity of the time series. The calculation formula of the permutation entropy is: ; where is the embedding dimension, is the probability of the permutation pattern , is the permutation entropy; S206. Obtain the fitness function based on the mutual information entropy and the permutation entropy. The calculation formula is: ; where is the fitness function, is the permutation entropy, is the mutual information entropy; S207. Obtain the target parameter group selected based on the fitness function.

[0009] In some embodiments of the present application, in S300, extracting the second time-domain index of each disturbance frequency data in step S100 based on the modal feature components, the target parameter group, and the first time-domain index includes the following steps: S301. Select a bandwidth based on the L2 norm of the gradient, and the calculation formula of the constrained variational function is: ; Among them, is the original signal without decomposition, is the partial derivative with respect to the time variable t , is the shorthand for all decomposition modes, is the th decomposition mode, is the shorthand for the central frequencies of all modes, is the th central frequency corresponding to the decomposition mode, is corresponding central frequency exponential term, is the Dirac distribution function, j is the imaginary unit, t is the time variable; S302. Construct an objective function based on the constrained variational function and the misclassification rate, and the calculation formula of the objective function is: ; Among them, is the relaxation variable representing the misclassification metric, is the total number of categories to be classified, α is the regularization parameter, β is the penalty parameter for controlling the misclassification tolerance, δ is the tolerance parameter, is the category index of the specific th category; S303. Use a support vector machine classifier to perform feature extraction based on the objective function, and extract the second time-domain index of each disturbance frequency data in step S100 based on the modal feature components, the target parameter group, and the first time-domain index.

[0010] In some embodiments of the present application, the first time-domain index includes mean, variance, peak value, kurtosis, root mean square value, peak factor, impulse factor, waveform factor, margin factor; The second time-domain index includes mean, peak value, kurtosis, root mean square value, peak factor, waveform factor.

[0011] In some embodiments of the present application, in S400, optimizing the hyperparameters of the fusion neural network by using the transit search algorithm includes: The transit search optimization algorithm is used to simulate the celestial body operation states, which include the galaxy stage, the transit stage, the planet stage, the neighborhood stage, and the development stage. The calculation formula for updating and calculating different celestial body operation states is as follows: ; Among them, c 10 、c 11 is a random number, c 12 is a random vector, is a random vector, is the position of the best star in the exploration stage, is the position of the star in the planet stage, is the vector of the galaxy center; Optimize the hyperparameters of the fusion neural network based on the update calculation results.

[0012] In some embodiments of the present application, in S400, combining the multi-head attention mechanism to allocate the feature weights of the second time-domain index in step S300 includes: Each attention head in the multi-head attention mechanism corresponds to a query vector, a key vector, and a value vector. Set the number of attention heads as i, and the calculation formula for the score value of the i-th attention head is: ; Among them, is , K, V the corresponding random matrix, is the score value of the i-th attention head; Use linear transformation to multiply the score value of each attention head by the random matrix to obtain the calculation formula for the target weight score: ; Among them, is the score value of the n att -th attention head, is the random matrix, is the target weight score; Allocate the feature weights of the second time-domain index in step S300 based on the target weight score.

[0013] In some embodiments of the present application, in S500, using a convolutional neural network - bidirectional gated recurrent unit classifier to perform perturbation identification on the time-series feature signal in step S400 to obtain a perturbation identification result includes the following steps: S501. Use a convolutional filter to extract high-dimensional features in the time-series signal, and the calculation formula for the high-dimensional features is: ; Among them, z is the input data, is the high-dimensional feature, is the filter parameter, k CNN is the size of the convolutional kernel, and b is the output position, c and is the position of the convolutional kernel, is the bias term of the filter parameter, is the activation function; S502. Extract the data feature information in the forward and backward propagations based on the bidirectional gated recurrent unit classifier. The calculation formula of the bidirectional gated recurrent unit classifier is: ; Among them, is the current moment of the input, and are the state memory variables of the previous moment and the current moment respectively, are the update gate, reset gate, candidate hidden state, and output vector of the current moment respectively, are the update gate, reset gate, candidate hidden state, and the weights of the output vector multiplied by the connectivity matrix composed of and respectively, is the sigmoid activation function, is the tanh activation function; S503. Use the Softmax function to convert the original class scores into a probability distribution. The calculation formula for obtaining the predicted probability is: ; Among them, is the original score of the th class perturbation, is the predicted probability of the th class perturbation, N is the total number of original classes; Obtain the perturbation identification result based on the predicted probability result.

[0014] The advantages and beneficial effects of the present invention compared with the prior art are: 1. By improving the sparrow search algorithm to optimize the penalty factor and the number of modal feature vectors of variational mode decomposition, the present invention can decompose the signal more accurately, extract the modal feature components more valuable for perturbation identification, and use the multi-head attention mechanism to allocate weights to the extracted features, which can highlight the important features and suppress the unimportant features, thereby improving the discrimination of the features.

[0015] 2. The classifier of the present invention that fuses a convolutional neural network and a bidirectional gated recurrent unit can fully utilize the ability to extract spatial features and the ability to model time series, thereby more accurately identifying the types of disturbances. By optimizing the hyperparameters of the fusion neural network through the transit search algorithm, the optimal parameter combination, such as the learning rate, the size of the convolutional kernel, and the number of neurons, is calculated, which can make the structure and parameters of the neural network more suitable for the current disturbance data and further improve the accuracy of disturbance identification.

[0016] The technical solution of the present invention will be further described in detail below with reference to the drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flowchart of a power system disturbance identification method based on a fusion neural network with parameter optimization according to the present invention; Figure 2 is a flowchart of feature extraction in an embodiment of the present invention; Figure 3 is a flowchart of the transit search algorithm in an embodiment of the present invention; Figure 4 is a schematic diagram of the fusion neural network in an embodiment of the present invention; Figure 5 is a schematic diagram of the identification results of eight types of disturbances in an embodiment of the present invention; (a) is a confusion matrix without feature selection, and (b) is a confusion matrix after feature selection; Figure 6 is a curve graph of the algorithm accuracy rate and the number of iterations in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] In the description of the present invention, it should be noted that the orientation or positional relationship indicated by the terms "upper", "lower", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings, or the orientation or positional relationship in which the product of the present invention is usually placed during use. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. In the description of the present invention, it should also be noted that unless otherwise clearly specified and defined, the terms "set", "installed", "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected, or indirectly connected through an intermediate medium, and it can be the communication inside two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0019] The embodiments of the present invention will be described in detail below with reference to the drawings.

[0020] As Figure 1 shown, the present invention provides a power system disturbance identification method based on a parameter-optimized fusion neural network, comprising the following steps: S100. Obtain eight types of disturbance frequency data of the power system generated based on simulation software.

[0021] Optionally, use the simulation software PowerFactory to generate disturbance data. The process is specifically as follows: Select a total of 8 disturbance types, namely three-phase short circuit 3-ΦFlt, single-phase ground fault Φ-gFlt, generator power reduction GL, load connection L-on, load disconnection L-off, line tripping LT, shunt capacitor connection SC-on, and shunt capacitor disconnection SC-off, for simulation. The system example is the IEEE 10-machine 39-bus system, the simulation time is 30 s, and the simulation method is shown in Table 1: Table 1 Simulation methods for different disturbances ;

[0022] S200. Optimize the penalty factor and the number of modal feature vectors decomposed by the variational mode decomposition method using an improved sparrow search algorithm to obtain a target parameter group.

[0023] S300. Obtain the modal feature components selected based on the fitness function, and extract the second time-domain index of each disturbance frequency data in step S100 based on the modal feature components, the target parameter group, and the first time-domain index.

[0024] S400. Optimize the hyperparameters of the fusion neural network using the transit search algorithm. The hyperparameters at least include the learning rate, the convolution kernel size, and the number of neurons, and allocate the feature weights of the second time-domain index in step S300 in combination with the multi-head attention mechanism to obtain a time-series feature signal.

[0025] S500. Use a convolutional neural network - bidirectional gated recurrent unit classifier to perform disturbance identification on the time-series feature signal in step S400 to obtain a disturbance identification result.

[0026] The advantages and beneficial effects of the present invention compared with the prior art are as follows: 1. By optimizing the penalty factor and the number of modal feature vectors of the variational mode decomposition using an improved sparrow search algorithm, the present invention can decompose signals more accurately, extract modal feature components that are more valuable for disturbance identification, and use the multi-head attention mechanism to allocate weights to the extracted features, which can highlight important features and suppress unimportant features, thereby improving the discrimination of features.

[0027] 2. By integrating a classifier of a convolutional neural network and a bidirectional gated recurrent unit, the present invention can make full use of the ability to extract spatial features and the ability to model time series, so as to more accurately identify the type of disturbance. By optimizing the hyperparameters of the fusion neural network through the transit search algorithm, the optimal parameter combination, such as the learning rate, the size of the convolutional kernel, and the number of neurons, is calculated, which can make the structure and parameters of the neural network more suitable for the current disturbance data, and further improve the accuracy of disturbance identification.

[0028] In some embodiments of the present application, in S100, obtaining the eight types of disturbance frequency data of the power system generated based on the simulation software includes: Using the PowerFactory software to perform simulations based on the eight types of disturbance data, and the eight types of disturbance data include three-phase short circuit, single-phase ground fault, generator output reduction, load connection, load disconnection, line tripping, shunt capacitor connection, and shunt capacitor disconnection data; Setting the simulation time to 30 s, the simulation step size to 0.02 s, and the trigger disturbance time to 5 s, and obtaining the eight types of disturbance frequency data output by the PowerFactory software simulation.

[0029] In some embodiments of the present application, as Figure 2 shown, in S200, optimizing the penalty factor and the number of modal eigenvectors decomposed based on the variational mode decomposition method by using the improved sparrow search algorithm, and obtaining the target parameter group includes the following steps: S201. Using the osprey optimization algorithm to replace the position update in the sparrow search algorithm during the exploration stage, simulating the process of an osprey detecting a random position and moving towards the prey, and the calculation formula is: ; Among them, is the updated position of the x-th osprey in the y-th dimension, is the fish selected by the x-th osprey, is the random number range between 0 and 1, is the random number range between 1 and 2; S202. Using the Cauchy mutation operator to make the iteration of the sparrow algorithm jump out of the local optimal solution, and at the same time increasing the iteration scale of the sparrow algorithm. The calculation formula of the Cauchy mutation operator is: ; Among them, is the dimension of the Cauchy mutation, is the calculation formula of the fitness function, M is the Cauchy mutation operator, is the local optimal solution. When , the Cauchy mutation operator makes the iteration of the sparrow algorithm jump out of the local optimal solution and obtain the global optimal solution; S203. Use the chaotic particle mapping iteration algorithm to make the population jump out of the local optimal solution and obtain the global optimal solution. The calculation formula of the chaotic particle mapping iteration algorithm is: ; where, is the number of the current iteration, is the position after the chaotic particle iterates times, is the position after the chaotic particle iterates times; S204. Select a fitness function and calculate the mutual information entropy. The calculation formula of the mutual information entropy is: ; where, is the joint probability distribution, specifically the joint probability that the random variable A takes the value of and the random variable B takes the value of b , is the marginal probability distribution that the random variable A takes the value of , is the marginal probability distribution that the random variable B takes the value of b , is the mutual information entropy; S205. Obtain the permutation entropy used to represent the complexity of the time series. The calculation formula of the permutation entropy is: ; where, is the embedding dimension, is the probability of the permutation pattern , is the permutation entropy; S206. Obtain a fitness function based on the mutual information entropy and the permutation entropy. The calculation formula is: ; where, is the fitness function, is the permutation entropy, is the mutual information entropy; S207. Obtain the target parameter group selected based on the fitness function.

[0030] Specifically, since the producers of the sparrow search algorithm rely too much on the update of the previous generation's position, the exploration stage of the osprey optimization algorithm is used to replace the original position update. As the sparrow algorithm iterates, it is difficult to jump out of the local optimal solution. Therefore, a Cauchy mutation operator is introduced, and at the same time, the search scale is expanded. There are two key parameters in variational mode decomposition, the number of decomposed modal characteristic components K and the penalty factor If the value of K is too large, it will cause over-decomposition and generate false modes. If the value of K is too small, the implicit features of the time series cannot be fully extracted. The value affects the bandwidth of the modal component. If it is set small, modal aliasing will occur. If it is set large, local information will be lost. For the accuracy of feature extraction, it is particularly important to determine the target parameter group [K, . Before parameter optimization, it is necessary to determine that the fitness function is the composite index permutation entropy and mutual information entropy.

[0031] In some embodiments of the present application, in S300, based on the modal feature component, the target parameter group, and the first time-domain index, extracting the second time-domain index of each perturbation frequency data in step S100 includes the following steps: S301. Select the bandwidth based on the L2 norm of the gradient, and the calculation formula of the constrained variational function is: ; Among them, among them, is the original signal that has not been decomposed, is the partial derivative with respect to the time variable t of, is the abbreviation of all decomposition modes, is the th decomposition mode, is the abbreviation of the central frequencies of all modes, is the th central frequency corresponding to the decomposition mode, is the corresponding central frequency exponential term, is the Dirac distribution function, j is the imaginary unit, t is the time variable; S302. Construct an objective function based on the constrained variational function and the misclassification rate. The calculation formula of the objective function is: ; Among them, is the slack variable representing the misclassification metric, n 3 is the total number of categories to be classified, α is the regularization parameter, β is the penalty parameter used to control the misclassification tolerance, δ is the tolerance parameter, is the category index of the specific th category; S303. Use a support vector machine classifier to extract features based on the objective function, and extract the second time-domain index of each disturbance frequency data in step S100 based on the modal feature components, the target parameter group, and the first time-domain index.

[0032] In some embodiments of the present application, the first time-domain index includes mean, variance, peak value, kurtosis, root mean square value, peak factor, impulse factor, waveform factor, and margin factor. The second time-domain index includes mean, peak value, kurtosis, root mean square value, peak factor, and waveform factor.

[0033] It should be understood that after the optimal parameter combination [K, of the variational mode decomposition is determined by the improved sparrow algorithm, the optimal penalty factor, the number of modal feature vectors of the decomposition, and the modal feature component corresponding to the minimum fitness value are brought into the variational mode decomposition. According to nine time-domain indexes: mean, variance, peak value, kurtosis, root mean square value, peak factor, impulse factor, waveform factor, and margin factor, a total of nine indexes are used as the feature vectors of the current disturbance signal to perform feature extraction under the current disturbance state. The wrapper method is used for feature selection, and the classification accuracy rate is selected as the evaluation criterion. The classifier is trained using the selected features (or feature subsets), and the overall performance of the feature subset is judged by the output accuracy rate of the classifier. In the experiment, a common classifier, support vector machines (SVM), is selected for training, and finally it is determined that the feature subset is composed of six time-domain indexes: mean, peak value, kurtosis, root mean square value, peak factor, and waveform factor.

[0034] Specifically, by optimizing the penalty factor and the number of modal feature vectors of VMD through the improved sparrow algorithm, the signal can be decomposed more accurately, and the modal feature components more valuable for disturbance identification can be extracted. This avoids the problems of inaccurate signal decomposition and over-decomposition caused by improper parameter selection in traditional methods. Using nine time-domain indexes (mean, variance, peak value, kurtosis, root mean square value, peak factor, impulse factor, waveform factor, and margin factor) as feature vectors comprehensively covers various aspects of information such as the statistical characteristics, transient characteristics, and waveform shape of the signal, providing rich basic data for subsequent feature selection and classification. Through experimental verification, it is finally determined that the feature subset is composed of six time-domain indexes: mean, peak value, kurtosis, root mean square value, peak factor, and waveform factor. These feature indexes show high discrimination and representativeness in the classification of disturbance signals and can effectively reflect the essential characteristics of different disturbance types.

[0035] In some embodiments of the present application, as Figure 3 shown, in S400, the hyperparameters for optimizing the fusion neural network using the transit search algorithm include: The transit search optimization algorithm is used to simulate the operating states of celestial bodies. The operating states of celestial bodies include the galaxy stage, the transit stage, the planet stage, the neighborhood stage, and the development stage. The calculation formula for updating and calculating different celestial body operating states is as follows: ; Among them, c 10 、c 11 is a random number, c 12 is a random vector, is a random vector, is the position of the best star in the exploration stage, is the position of the star in the planet stage, is the vector of the galaxy center; Optimize the hyperparameters of the fusion neural network based on the update calculation results.

[0036] In some embodiments of the present application, in S400, combining the multi-head attention mechanism to assign the feature weights of the second time-domain index in step S300 includes: Each attention head in the multi-head attention mechanism corresponds to a query vector, a key vector, and a value vector. Let the number of attention heads be i, and the calculation formula for the score value of the i-th attention head is: ; Among them, is , K, V the corresponding random matrix, is the score value of the i-th attention head; Use a linear transformation to multiply the score value of each attention head by the random matrix to obtain the calculation formula for the target weight score: ; Among them, is the score value of the i-th attention head, is the random matrix, is the target weight score; Based on the target weight score, assign the feature weights of the second time-domain index in step S300.

[0037] Among them, the accuracy of the hyperparameters of the deep learning network set by the empirical trial and error method is often difficult to reach the best. It is necessary to perform hyperparameter optimization through an optimization algorithm, and a multi-head attention mechanism is added to the fusion neural network model to achieve the goal of automatically assigning weights to the input feature vectors.

[0038] Specifically, the transit search algorithm can comprehensively explore the search space of hyperparameters by simulating the operating states of celestial bodies (including the galaxy stage, transit stage, planet stage, neighborhood stage, and exploitation stage). This multi-stage optimization strategy avoids the problem of being easily trapped in local optima in traditional methods and improves the global search ability.

[0039] In some embodiments of the present application, as Figure 4 shown, in S500, using a convolutional neural network - bidirectional gated recurrent unit classifier to perform perturbation identification on the time series feature signal in step S400 to obtain a perturbation identification result includes the following steps: S501. Extract high-dimensional features from the time series signal using a convolutional filter. The calculation formula for the high-dimensional features is: ; where z is the input data, is the high-dimensional feature, is the filter parameter, k CNN is the size of the convolutional kernel, and b are the output positions, c and are the positions of the convolutional kernel, is the filter parameter bias term, is the activation function; S502. Extract data feature information in the forward and backward propagations based on the bidirectional gated recurrent unit classifier. The calculation formula for the bidirectional gated recurrent unit classifier is: ; where, is the current moment of the input, and are the state memory variables of the previous moment and the current moment respectively, are the update gate, reset gate, candidate hidden state, and output vector at the current moment respectively, are the update gate, reset gate, candidate hidden state, and the output vector multiplied by the weight parameters of the connectivity matrix composed of and respectively, is the sigmoid activation function, is the tanh activation function; S503. Use the Softmax function to convert the original class scores into a probability distribution. The calculation formula for the predicted probability is: ; where, is the original score of the th type of perturbation, is the predicted probability of the th type of disturbance, N where is the total number of original categories;

[0040] Based on the predicted probability results, the disturbance identification result is obtained.

[0041] Specifically, the convolutional neural network extracts high-dimensional non-linear features from the original frequency signal data, the bidirectional gated recurrent unit extracts long-term time series features, and finally the Softmax function is used to convert the original class scores into a probability distribution as the output layer of the model.

[0042] Specifically, high-dimensional features in the time series signal are extracted through convolutional filters, and the CNN can automatically learn local features in the input data. The convolutional operation can capture local patterns in the time series signal, such as the features of disturbances like voltage sags and harmonics. The BiGRU can capture bidirectional dependencies in the time series signal by processing both forward and backward time series information, enabling the model to not only utilize past signal information but also future signal information to better understand the disturbance state at the current moment.

[0042] Combined with a specific scenario of the IEEE 39-node system, the method of the present invention is simulated and tested. The influence of a new energy wind power penetration rate of 40% on disturbance identification is considered in the simulation process. The specific operation is to replace the thermal power units at nodes 30, 32, 33, 36, and 37 with doubly-fed wind turbines of the same capacity. Eight types of disturbances and 960 disturbance events are simulated, including three-phase short circuit (3-ϕFlt), single-phase grounding (ϕ-gFlt), generator power reduction (GL), load connection (L-on), load disconnection (L-off), line tripping (LT), shunt capacitor connection (SC-on), and shunt capacitor disconnection (SC-off), with 120 for each type of disturbance. The training set and test set are constructed in a 7:3 ratio, and the hyperparameters are set as shown in Table 2: Table 2 Hyperparameter settings of the fusion neural network ;

[0043] To verify the reliability of the power system disturbance identification of the present invention, the disturbance identification method of the present invention, including feature extraction and disturbance classification, is used for the simulation system and the actual on-site power system. The comparison diagrams with and without feature selection are as Figure 5 shown. It can be seen that the average classification accuracy of the present invention is 98.3%, and the highest classification accuracy tested using relevant deep learning network models is 97.5%. Therefore, the method provided by the present invention improves the overall classification accuracy. The iterative convergence curves of accuracy and loss are as Figure 6 shown, and the algorithm has good convergence. The present invention can achieve accurate identification of disturbances.

[0044] The beneficial effects of the present invention are as follows: By adopting the above-mentioned power system disturbance identification method based on a parameter-optimized fusion neural network, on the basis of new energy access, a power system disturbance identification method that can achieve feature extraction of frequency signals, parameter optimization of neural network models, and disturbance classification is proposed, improving the overall classification accuracy.

[0045] In this application, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which this application belongs. In case of inconsistency, the meaning stated in this specification or the meaning derived from the content recorded in this specification shall prevail. In addition, the terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0046] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A power system disturbance identification method based on a fusion neural network with parameter optimization, characterized in that Including the following steps: S100. Obtain eight types of disturbance frequency data of the power system generated based on simulation software; S200. Optimize the penalty factor and the number of modal eigenvectors decomposed based on the variational mode decomposition method by using an improved sparrow search algorithm to obtain a target parameter group; S300. Obtain modal eigencomponents selected based on a fitness function, and extract the second time-domain index of each type of disturbance frequency data in step S100 based on the modal eigencomponents, the target parameter group, and the first time-domain index; S400. Optimize the hyperparameters of the fusion neural network by using a transit search algorithm, where the hyperparameters at least include the learning rate, the convolutional kernel size, and the number of neurons, and allocate the feature weights of the second time-domain index in step S300 by combining a multi-head attention mechanism to obtain a time-series feature signal; S500. Use a convolutional neural network-bi-directional gated recurrent unit classifier to perform disturbance identification on the time-series feature signal in step S400 to obtain a disturbance identification result.

2. The power system disturbance identification method based on a fused neural network with parameter optimization according to claim 1, wherein In the said S100, obtaining eight types of disturbance frequency data of the power system generated based on simulation software includes: Using PowerFactory software to perform simulation based on eight types of disturbance data, and the eight types of disturbance data include three-phase short circuit, single-phase ground fault, generator output reduction, load connection, load disconnection, line tripping, shunt capacitor connection, and shunt capacitor disconnection data; Setting the simulation time to 30 s, the simulation step size to 0.02 s, and the triggering disturbance time to 5 s to obtain eight types of disturbance frequency data output by the PowerFactory software simulation.

3. A power system disturbance identification method based on a fusion neural network with parameter optimization according to claim 1, characterized in that In the said S200, optimizing the penalty factor and the number of modal eigenvectors decomposed based on the variational mode decomposition method by using an improved sparrow search algorithm to obtain a target parameter group includes the following steps: S201. Use an osprey optimization algorithm to replace the position update in the sparrow search algorithm in the exploration stage, simulate the process of an osprey detecting a random position and moving towards the prey, and the calculation formula is: ； Among them, is the th fish eagle's updated position in the th iteration, is the current position of the th fish eagle in the th dimension, is the fish selected by the th fish eagle, is a random number between 0 and 1, is a random number between 1 and 2; S202. Use a Cauchy mutation operator to make the iteration of the sparrow algorithm jump out of the local optimal solution, and at the same time increase the iteration scale of the sparrow algorithm. The calculation formula of the Cauchy mutation operator is: ； Among them, is the dimension of Cauchy mutation, is the calculation formula of the fitness function, M is the Cauchy mutation operator, is the local optimal solution. When < 0, the Cauchy mutation operator makes the iteration of the sparrow algorithm jump out of the local optimal solution and obtain the global optimal solution, is the numerical value for the judgment basis; S203. Use a chaotic particle mapping iteration algorithm to make the population jump out of the local optimal solution to obtain the global optimal solution. The calculation formula of the chaotic particle mapping iteration algorithm is: ； Among them, is the number of the current iteration, is the position of the chaotic particle after iterations, is the position of the chaotic particle after iterations; S204. Select a fitness function and calculate the mutual information entropy. The calculation formula of the mutual information entropy is: ； Among them, is the joint probability distribution, specifically the joint probability that the random variable A takes the value of and the random variable B takes the value of b . is the marginal probability distribution that the random variable A takes the value of . is the marginal probability distribution that the random variable B takes the value of b . is the mutual information entropy; S205. Obtain the permutation entropy representing the complexity of the time series. The calculation formula of the permutation entropy is: ； Among them, is the embedding dimension, is the probability of the permutation pattern , is the permutation entropy; S206. Obtain a fitness function based on the mutual information entropy and the permutation entropy. The calculation formula is: ； Among them, is the fitness function, is the permutation entropy, is the mutual information entropy; S207. Obtain a target parameter group selected based on the fitness function.

4. A power system disturbance identification method based on a fusion neural network with parameter optimization according to claim 1, characterized in that, In the said S300, extracting the second time-domain index of each type of disturbance frequency data in step S100 based on the modal eigencomponents, the target parameter group, and the first time-domain index includes the following steps: S301. Select the bandwidth based on the L2 norm of the gradient to obtain the calculation formula of the constrained variational function: ； Among them, is the original signal without decomposition, is the partial derivative with respect to the time variable t , is the abbreviation of all decomposition modes, is the th decomposition mode, is the abbreviation of the central frequencies of all modes, is the central frequency corresponding to the th decomposition mode, is the corresponding central frequency exponential term, is the Dirac distribution function, j is the imaginary unit, t is the time variable; S302. Construct an objective function based on the constrained variational function and the misclassification rate. The calculation formula of the objective function is: ； wherein, is a relaxation variable representing the classification error metric, is the total number of categories to be classified, α is the regularization parameter, β is the penalty parameter for controlling the misclassification tolerance, δ is the tolerance parameter, is for the specific class category index; S303. Use a support vector machine classifier to perform feature extraction based on the objective function, and extract the second time-domain indicators of each disturbance frequency data in step S100 based on the modal feature components, the target parameter group, and the first time-domain indicators.

5. The power system disturbance identification method based on a fused neural network with parameter optimization according to claim 4, wherein The first time-domain indicators include mean, variance, peak value, kurtosis, root mean square value, peak factor, impulse factor, waveform factor, and margin factor. The second time-domain indicators include mean, peak value, kurtosis, root mean square value, peak factor, and waveform factor.

6. A power system disturbance identification method based on a parameter-optimized fusion neural network according to claim 1, characterized in that, In the S400, the hyperparameters of the fusion neural network optimized by the transit search algorithm include: Use the transit search optimization algorithm to simulate the celestial body operation state, which includes the galaxy stage, the transit stage, the planet stage, the neighborhood stage, and the development stage. The calculation formula for updating and calculating different celestial body operation states is: ； Among them, c 10 、c 11 is a random number, c 12 is a random vector, is a random vector, is the position of the best star in the exploration stage, is the position of the star in the planetary stage, is the vector of the galactic center; Optimize the hyperparameters of the fusion neural network based on the update calculation results.

7. A power system disturbance identification method based on a fusion neural network with parameter optimization according to claim 6, characterized in that In the S400, the feature weights of the second time-domain indicators in step S300 are assigned by combining the multi-head attention mechanism, including: In the multi-head attention mechanism, each attention head corresponds to a query vector, a key vector, and a value vector. Setting the number of attention heads to , the calculation formula for the score value of the -th attention head is: ； Among them, is , K, V the corresponding random matrix, is the query vector 、K is the key vector 、V is the value vector, is the score value of the $i$-th attention head; Use a linear transformation to multiply the score value of each attention head by a random matrix. The calculation formula for obtaining the target weight score is: ； Among them, is the score value of the th attention head, is a random matrix, is the target weight score; Assign the feature weights of the second time-domain indicators in step S300 based on the target weight score.

8. A power system disturbance identification method based on a parameter-optimized fusion neural network according to claim 1, characterized in that In the S500, the disturbance identification of the time-series feature signal in step S400 is performed by using a convolutional neural network - bidirectional gated recurrent unit classifier to obtain the disturbance identification result, including the following steps: S501. Use a convolutional filter to extract the high-dimensional features in the time-series signal. The calculation formula for obtaining the high-dimensional features is: ； Among them, z is the input data, is the high-dimensional feature, is the filter parameter, k CNN is the size of the convolution kernel, and b is the output position, c and d is the position of the convolution kernel, is the filter parameter bias term, is the activation function; S502. Extract the data feature information in the forward and backward propagations based on the bidirectional gated recurrent unit classifier. The calculation formula of the bidirectional gated recurrent unit classifier is: ； Among them, is the current moment of the input, and are the state memory variables of the previous moment and the current moment respectively, are the update gate, reset gate, candidate hidden state, and output vector of the current moment respectively, are the weights of the update gate, reset gate, candidate hidden state, and output vector multiplied by the connectivity matrix composed of and respectively, is the sigmoid activation function, is the tanh activation function, is the identity matrix; S503. Use the Softmax function to convert the original class scores into a probability distribution. The calculation formula for obtaining the prediction probability is: ； Among them, is the original score of the th type of perturbation, is the predicted probability of the th type of perturbation, N is the total number of original categories; Obtain the disturbance identification result based on the prediction probability result.

Images