Security domain feature pattern recognition method and system based on capsule neural network

Abstract

The present invention relates to the technical field of dynamic security domains in power systems, and particularly to a security domain feature pattern recognition method and system based on a capsule neural network. The method comprises the steps: sampling data samples of various operating states of a power system by a Halton sampling method to obtain original input data samples of M different power generation combinations; on the basis of the data samples obtained by sampling, selecting, by means of a recursive feature elimination algorithm of a support vector machine, the top N pre-fault features having the greatest impact on an objective function among the data samples, and selecting K active features therefrom; and on the basis of the K selected active features, converting one-dimensional data into two-dimensional data and inputting the two-dimensional data into a capsule neural network, constructing a transient stability discrimination binary classification model, and training in each of a two-dimensional space and a three-dimensional space to generate a dynamic security domain boundary for dividing the feature space into a stable subspace and an unstable subspace. The present invention can provide dispatchers with more efficient, more accurate and more real-time monitoring and decision support.

Description

A security domain feature pattern recognition method and system based on capsule neural network Technical Field

[0001] The present invention belongs to the technical field of dynamic security domains of power systems, and in particular relates to a method and system for identifying security domain characteristic patterns based on capsule neural networks. Background Art

[0002] Deep learning-based feature pattern recognition uses neural networks to automatically extract and analyze features from data to identify patterns or trends. In recent years, it has achieved remarkable results in fields such as medical diagnosis and autonomous driving. However, previous approaches to power system security domain analysis have relied on fitting and analytical methods, as well as machine learning algorithms, but neural network technology has never been incorporated into applications. Hinton, a pioneer of deep learning, proposed a novel capsule neural network (CapsNet) model. Leveraging a dynamic routing algorithm, it exhibits unique features such as spatial hierarchical connectivity and deep feature extraction. CapsNet has demonstrated significant success in identifying power grid operational characteristics and mining stability rules. This approach closely aligns with the specific characteristics of power system security domain analysis, retaining more detailed information about the input objects. This makes it advantageous for solving complex problems involving multiple factors and unclear mechanisms. However, a novel method for power system security domain feature pattern recognition based on CapsNets is currently lacking. Summary of the Invention

[0003] In order to solve the above problems, the present invention starts from the data-driven perspective, taking into account that the essence of pattern recognition is feature classification, which is the process of treating the original information of the identification object, extracting feature vectors according to actual needs, and designing classifiers and implementing classification decisions. Capsule neural networks are classifiers that perform better in terms of feature extraction, classification discrimination, and fitting prediction capabilities. The construction ideas of the dynamic safety domain fitting method have inherent similarities with capsule neural networks. Therefore, a safety domain feature pattern recognition method and system based on capsule neural networks are provided, which successfully fills the gap in the safety domain field and opens up a new direction for the development of the power system safety domain. Research has been carried out from the aspects of sample construction, feature selection, and dynamic safety domain construction to improve computing efficiency and control accuracy. The specific technical solutions are as follows:

[0004] The present invention provides a method for identifying security domain feature patterns based on a capsule neural network, comprising the following steps:

[0005] Step S1: Using the Halton sampling method, data samples of various operating states of the power system are sampled to obtain original input data samples of M different power generation combinations, evenly covering the entire state space of the power grid;

[0006] Step S2: Based on the sampled data samples, the recursive feature elimination algorithm of the support vector machine is used to screen out the top N pre-fault features in the data samples that have the greatest influence on the objective function; and K active features are screened out among them;

[0007] In step S3, based on the K active features selected, the one-dimensional data is converted into two-dimensional data and input into the capsule neural network to construct a transient stability discrimination binary classification model. The model is trained in two-dimensional and three-dimensional spaces to generate a dynamic safety domain boundary that divides the feature space into two subspaces: stable and unstable.

[0008] Preferably, the step S1 specifically includes:

[0009] Adjust the active output of all generators except the balancing machine to fluctuate between 90% and 110%, and adjust the active power of the load according to the principle of power balance;

[0010] The Halton sampling method is used to sample data samples of various operating states of the power system to obtain original input data samples of M different power generation combinations.

[0011] Preferably, the step S1 further includes selecting features, specifically as follows:

[0012] A symmetrical fault is set up in the power system. The original input data samples of M different power generation combinations extracted above are loaded. The active output of the corresponding generator is continuously changed by modifying the tempchange.dat file. PSD-BPA software is called in batches using MATLAB to read the tempcase.out file. A total of m features are obtained, including the active and reactive output of the generator, the active and reactive loads of the load nodes, the active and reactive flows of the transmission lines, and the bus voltage amplitude and phase angle. The stability of the power system is determined by whether the maximum relative power angle difference between the generators is greater than 360° at the end of the simulation.

[0013] Preferably, the objective function of the support vector machine of the recursive feature elimination algorithm of the support vector machine in step S2 is:

[0014] Among them, ω is the weight vector, which represents the normal vector of the hyperplane;

[0015] When the i-th pre-fault feature is deleted, i=1,2,···,N, the objective function changes to: ΔJ(i)≈ω i 2 ;

[0016] Where ΔJ(i) represents the change in the objective function when the i-th pre-fault feature is deleted; ω i represents the weight vector of the i-th pre-fault feature;

[0017] The kernel function used is a linear kernel function, and the sorting coefficient is:

[0018] when When the value of is less than the set threshold, the corresponding pre-fault feature of the i-th fault is deleted.

[0019] Preferably, the transient stability discrimination binary classification model in step S3 is as follows: f(x)=sgn(ω T x+b);

[0020] Among them, x is the sample feature, ω is the classification hyperplane coefficient, and b is the bias term;

[0021] The loss function of the capsule neural network is as follows: L k =T k max(0,m + -||v k ||) 2 +λ(1-T k )max(0,||v k ||-m - ) 2 ;

[0022] Where, L k is the loss of the k-th capsule, T k is a label indicating whether the capsule should be activated (1 means it should be activated, 0 means it should not be activated), v k is the output vector of the kth capsule, m + and m - are normal and negative constant bounds, representing the minimum and maximum thresholds for capsule activation, respectively, and λ k is a scaling factor that adjusts the weight of the negative constant bound.

[0023] Preferably, the capsule neural network in step S3 comprises three network layers, namely a convolutional layer, a main capsule layer and a digital capsule layer;

[0024] The convolution layer has 1 input channel, 256 output channels, a 9x9 kernel size, a stride of 1, and uses the ReLU activation function.

[0025] The number of input channels of the main capsule layer is 256, the number of output channels is 32, the capsule vector dimension is 8, the convolution kernel size is 9x9, and the stride is 2;

[0026] The number of output channels of the digital capsule layer is 2, the capsule vector dimension is 16, the input capsule vector dimension is 8, and the output of the initial capsule layer is processed using a dynamic routing algorithm and a nonlinear activation function.

[0027] The present invention also provides a security domain feature pattern recognition system based on capsule neural network, comprising:

[0028] The data sample construction module is used to sample data samples of various operating states of the power system using the Halton sampling method to obtain original input data samples of M different power generation combinations, evenly covering the entire state space of the power grid;

[0029] The feature screening module is used to screen out the top N pre-fault features that have the greatest impact on the objective function from the data samples obtained through sampling using the recursive feature elimination algorithm of the support vector machine; and screen out the K active features among them;

[0030] The dynamic safety domain boundary classification module is used to convert one-dimensional data into two-dimensional data based on the K active features screened out, and input it into the capsule neural network to build a transient stability discrimination binary classification model. The module is trained in two-dimensional and three-dimensional spaces to generate a dynamic safety domain boundary that divides the feature space into stable and unstable subspaces.

[0031] The present invention also provides a computer-readable storage medium, characterized in that the computer-readable storage medium includes a stored program, wherein when the program is running, the device where the computer-readable storage medium is located is controlled to execute the security domain feature pattern recognition method based on capsule neural network.

[0032] The present invention also provides a processor, characterized in that the processor is used to run a program, wherein the program executes the security domain feature pattern recognition method based on capsule neural network when running.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] First, Halton sampling is performed on the active output of all generators except the balancing machine to obtain a uniform sample covering the space; then, the recursive feature elimination (SVM-RFE) method based on support vector machine is used to select critical pre-fault feature quantities of the system; finally, based on the screened active features, the one-dimensional data is converted into two-dimensional data and input into the capsule neural network. Dynamic safety domain boundaries that can divide the feature space into stable and unstable subspaces are trained in two-dimensional and three-dimensional spaces, respectively. A safety domain feature pattern recognition method based on capsule neural network is proposed, which provides dispatchers with more efficient, accurate and real-time monitoring and decision support. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly describes the drawings required for the specific embodiments or the description of the prior art. Similar elements or parts are generally identified by similar reference numerals throughout the drawings. Elements or parts in the drawings are not necessarily drawn to scale.

[0036] FIG1 is a wiring diagram of the IEEE39 node system used in the implementation of the present invention;

[0037] FIG2 is a diagram showing the Halton sampling results of the active output of generators G35, G36, and G37;

[0038] Figure 3 shows the arrangement diagram of the first 25 features screened by the SVM-RFE method;

[0039] FIG4 is a schematic diagram showing the principle structure of a security domain feature pattern recognition method based on a capsule neural network;

[0040] Figure 5 is a two-dimensional diagram of a dynamic security domain;

[0041] Figure 6 is a three-dimensional diagram of a dynamic security domain. DETAILED DESCRIPTION

[0042] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0043] It will be understood that when used in this specification and the appended claims, the terms “comprises” and “comprising” indicate the presence of described features, integers, steps, operations, elements and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0044] It should also be understood that the terms used in the present specification are only for the purpose of describing particular embodiments and are not intended to limit the present invention. As used in the present specification and the appended claims, the singular forms "a", "an", and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0045] It should be further understood that the term "and / or" used in the present description and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0046] The existing traditional methods of safety domain analysis include fitting methods and analytical methods, or some machine learning algorithms, but no one has introduced and applied neural network technology. This application is based on capsule neural networks and provides a dynamic safety domain feature pattern recognition method, which provides a reliable prediction basis for power grid dispatchers to make emergency plans in the face of emergencies, and also opens up a new idea for the field of safety domain. In order to better introduce the safety domain feature pattern recognition method based on capsule neural networks proposed in the present invention, the following introduces the three links of sample construction, feature selection, and dynamic safety domain construction. First, Halton sampling is performed on the active output of all generators except the balancing machine to obtain a uniform sample covering the space. Then, the recursive feature elimination (SVM-RFE) method based on support vector machines is used to select critical system pre-fault feature quantities. Finally, based on the screened active features, the one-dimensional data is converted into two-dimensional data and input into the capsule neural network, and dynamic safety domain boundaries are constructed in two-dimensional and three-dimensional spaces respectively, which can divide the feature space into two types of stable and unstable subspaces. The details are as follows:

[0047] Example 1:

[0048] The present invention provides a method for identifying safety domain feature patterns based on a capsule neural network, comprising the following steps: Step S1: Using the Halton sampling method, data samples of various operating states of the power system are sampled to obtain raw input data samples for M different power generation combinations, uniformly covering the entire state space of the power grid. Specifically, this method includes adjusting the active output of all generators, except the balancing machine, to fluctuate between 90% and 110%, and adjusting the active power of the load according to the principle of power balance.

[0049] The example used in the present invention is the IEEE39-node standard system, which is a transmission network with a voltage level of 345kV in the New England region of the United States. The system includes 10 generators, 39 busbars, 12 transformers and 34 distributed parameter lines. The system wiring diagram is shown in Figure 1. In the power system, facing the challenge of high-dimensional state space, the traditional random sampling method may be affected by the dimensionality curse, resulting in uneven distribution of sampling points, and may not be able to effectively cover the few extreme cases where transient stability is lost after a fault occurs. In order to solve this problem, the Halton sampling method is introduced. Halton sampling is a deterministic, low-variance sequence sampling method that is particularly suitable for high-dimensional space. By selecting an appropriate prime number as the base, the Halton sequence can cover the entire state space more evenly. Therefore, compared with traditional random sampling, the application of Halton sampling in the power system can provide a more uniform sample distribution, help to more comprehensively explore the various operating states of the power system, effectively improve the uniformity and efficiency of sampling, and provide a more reliable basis for simulating the safety domain feature pattern recognition of various operating conditions of the power system.

[0050] During the sample construction phase, the active output of all generators, excluding the balancing machine, was adjusted to fluctuate between 90% and 110%, and the active power of the load was adjusted according to the principle of power balance. M = 1000 different generation combinations were extracted as the raw input data samples. The Halton sampling results for the active outputs of G35, G36, and G37 in the system are shown in Figure 2. Notably, the sampling results are widely distributed across the entire space, which helps improve the generalization capability of feature pattern recognition in the dynamic safety domain.

[0051] A symmetrical fault is set for the power system. For the IEEE 39-node system, the fault is set to a three-phase short circuit fault on the L4-L5 line, which disappears after 6 cycles.

[0052] Load the original input data samples of M = 1000 different power generation combinations extracted above, continuously change the active output of the corresponding generator by modifying the tempchange.dat file, use MATLAB to batch call PSD-BPA software, read the tempcase.out file, and obtain a total of m = 204 features including the active output and reactive output of the generator (20 features), the active load and reactive load of the load node (38 features), the active power flow and reactive power flow of the transmission line (68 features), and the bus voltage amplitude and phase angle (78 features). The stability of the power system is determined by whether the maximum relative power angle difference between the generators is greater than 360° at the end of the simulation.

[0053] In step S2, these 204 features represent a relatively high-dimensional dataset, and support vector machines typically perform well on high-dimensional data. Recursive feature elimination can determine the relative importance of each feature by continuously training the model and evaluating the contribution of the features, which helps identify the most critical features before system failure. Therefore, based on the sampled data samples, the recursive feature elimination algorithm of the support vector machine is used to filter out the top N pre-fault features in the data sample that have the greatest impact on the objective function, where N is less than m; and K active features are also selected, where K is less than N. The objective function of the support vector machine for the recursive feature elimination algorithm of the support vector machine is:

[0054] Among them, ω is the weight vector, which represents the normal vector of the hyperplane;

[0055] When the i-th pre-fault feature is deleted, i=1,2,···,N, the objective function changes to: ΔJ(i)≈ω i 2 ;

[0056] Where ΔJ(i) represents the change in the objective function when the i-th pre-fault feature is deleted; ωi represents the weight vector of the i-th pre-fault feature;

[0057] The kernel function used is a linear kernel function, and the sorting coefficient is:

[0058] when When the value of is less than the set threshold, the corresponding pre-fault feature of the i-th fault is deleted.

[0059] The recursive feature elimination algorithm SVM-RFE of support vector machine is based on ω i 2 As a sorting criterion, the smaller its value is, the less information it contains and the more likely it will be ranked at the back. It will also be deleted sooner, and then the top N subsets with the greatest influence on the objective function can be selected from the entire set.

[0060] First, two categories were selected from the dataset, and the same number of samples were randomly subsampled to balance the class distribution. Next, the problem was converted to a binary classification task, with stable samples assigned to 0 and unstable samples assigned to 1. Then, using the support vector machine-based recursive feature elimination (SVM-RFE) method, the most critical set of features for support vector machine classification was selected by gradually eliminating features that contributed less to classification performance. To comprehensively evaluate the model's performance, a 10-fold cross-validation experiment was performed. The support vector machine used a linear kernel, the penalty parameter C was set to 1, and the L1QP solver was used. The entire process was repeated 10 times to improve the stability of the results. Finally, the top 25 features were selected by averaging the rankings of the results from all iterations. These features are ranked in descending order of importance, as shown in Figure 3, and the results of the selected features are shown in Table 1.

[0061] Table 1 The top 25 features screened

[0062] In step S3, based on the K active features selected, the one-dimensional data is converted into two-dimensional data and input into the capsule neural network to construct a transient stability discrimination binary classification model. The model is trained in two-dimensional and three-dimensional spaces to generate a dynamic safety domain boundary that divides the feature space into two subspaces: stable and unstable.

[0063] Among the 25 screened features, K=11 are active features, namely the active power flow of line L18-L3, the active power flow of L5-L4, the active power flow of line L6-L5, the active power flow of line L14-L4, the active power flow of line L19-L16, the active power flow of line L8-L5, the active output of generator 34, the active power flow of line L21-L16, the active power flow of line L26-L25, the active power flow of line L17-L16, the active power flow of line L11-L6, and the active power flow of line L7-L6. Since the research on the security domain is based on the active power injection space, these active features are taken and combined with the capsule neural network to realize the construction of the dynamic security domain.

[0064] The safety domain feature pattern recognition method based on capsule neural network is essentially a binary classification problem. The transient stability discrimination binary classification model is as follows: f(x) = sgn(ω T x+b);

[0065] Among them, x is the sample feature, ω is the classification hyperplane coefficient, and b is the bias term;

[0066] The process of training a capsule neural network is to continuously update the weight ω and bias b until a stable ω and b are found to minimize the loss function of the model. The loss function of the capsule neural network is as follows: L k =T k max(0,m + -||v k ||) 2 +λ(1-T k )max(0,||v k ||-m - ) 2 ;

[0067] Where, L k is the loss of the k-th capsule, T k is a label indicating whether the capsule should be activated (1 means it should be activated, 0 means it should not be activated), v k is the output vector of the kth capsule, m + and m - are normal and negative constant bounds, representing the minimum and maximum thresholds for capsule activation, respectively, and λ k Is a scaling factor used to adjust the weight of the negative constant boundary. Take m + =0.9, m - =0.1,λ=0.5;

[0068] The simulation experiment environment uses the PyTorch framework, Windows 11, an AMD Ryzen 7 5800HS Creator Edition CPU (3.20GHz), an NVIDIA GeForce MX450 GPU, and Python 3.7. The learning rate is 10-5, the Adam optimizer is used, and the training and test sets are split in an 8:2 ratio.

[0069] Based on the 11 active features obtained through screening, the one-dimensional data is converted into two-dimensional data and input into the capsule neural network. The capsule neural network consists of three layers: the convolutional layer, the main capsule layer, and the digital capsule layer, as shown in Figure 4. Each layer extracts features and abstractly represents the input data.

[0070] The convolution layer has 1 input channel (grayscale image), 256 output channels, a convolution kernel size of 9x9, a stride of 1, and uses the ReLU activation function.

[0071] The number of input channels of the main capsule layer is 256, the number of output channels is 32, the capsule vector dimension is 8, the convolution kernel size is 9x9, and the stride is 2;

[0072] The number of output channels of the digital capsule layer is 2, the capsule vector dimension is 16, the input capsule vector dimension is 8, and the output of the initial capsule layer is processed using a dynamic routing algorithm and a nonlinear activation function.

[0073] Capsule neural networks can learn classification based on data-driven training, and accurately construct dynamic safety domain boundaries in the hyperparameter space by dividing the feature space into stable and unstable subspaces.

[0074] Taking the active power output of generator 34 and the active power flow of line L14-L4 as examples, a two-dimensional dynamic safety region was constructed, as shown in Figure 5. Taking the active power output of generator 34, the active power flow of line L18-L3, and the active power flow of line L6-L5 as examples, a three-dimensional dynamic safety region was constructed, as shown in Figure 6. It can be seen that the capsule neural network can effectively perform feature pattern recognition in two-dimensional and three-dimensional space, construct dynamic safety region boundaries, and divide the feature space into different subspaces.

[0075] Example 2:

[0076] This embodiment further provides a security domain feature pattern recognition system based on a capsule neural network, including:

[0077] The data sample construction module is used to sample data samples of various operating states of the power system using the Halton sampling method to obtain original input data samples of M different power generation combinations, evenly covering the entire state space of the power grid; the specific principle is shown in the above step S1 and will not be repeated here.

[0078] The feature screening module is used to screen out the top N pre-fault features that have the greatest influence on the objective function from the data samples obtained through sampling, using the recursive feature elimination algorithm of the support vector machine, where N is less than M; and to screen out the K active features therein; the specific principle is described in the above step S2 and will not be repeated here.

[0079] The dynamic safety region boundary classification module is used to convert the one-dimensional data into two-dimensional data based on the K selected active features, and then input it into the capsule neural network to build a transient stability discrimination binary classification model. The module is trained in two-dimensional and three-dimensional space to generate dynamic safety region boundaries that divide the feature space into stable and unstable subspaces. The specific principles are described in step S3 above and will not be repeated here.

[0080] Example 3:

[0081] The present invention also provides a computer-readable storage medium, characterized in that the computer-readable storage medium includes a stored program, wherein when the program is running, the device where the computer-readable storage medium is located is controlled to execute the security domain feature pattern recognition method based on capsule neural network.

[0082] Example 4:

[0083] The present invention also provides a processor, characterized in that the processor is used to run a program, wherein the program executes the security domain feature pattern recognition method based on capsule neural network when running.

[0084] Those skilled in the art will appreciate that the modules of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition of each example has been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0085] In the embodiments provided by the present invention, it should be understood that the division of modules is merely a logical function division, and there may be other division methods in actual implementation, for example, multiple modules can be combined into one module, one module can be split into multiple modules, or some features can be ignored, etc.

[0086] In addition, the functional modules in various embodiments of the present invention may be integrated into a single processing module, or each module may exist physically separately, or two or more modules may be integrated into a single module. The aforementioned integrated modules may be implemented in the form of hardware or software functional modules.

[0087] If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-0nly Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk, etc., various media that can store program code.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and description of the present invention.

Claims

1. A method for identifying security domain feature patterns based on capsule neural network, characterized in that: The following steps are involved: Step S1, using the Halton sampling method to sample data samples of various operating states of the power system, to obtain original input data samples of M different power generation combinations, evenly covering the entire state space of the power grid; Step S2, based on the sampled data samples, a recursive feature elimination algorithm of a support vector machine is used to screen out the top N pre-fault features in the data samples that have the greatest influence on the objective function; And screen the K active features therein; Step S3, based on the screened K active features, convert the one-dimensional data into two-dimensional data and input it into the capsule neural network to build a transient stability discrimination binary classification model, and train in two-dimensional and three-dimensional spaces to generate a dynamic safety domain boundary that divides the feature space into two types of stable and unstable subspaces.

2. According to claim 1, a method for identifying security domain feature patterns based on capsule neural network is characterized in that: The step S1 specifically includes: Adjust the active output of all generators except the balancing machine to fluctuate between 90% and 110%, and adjust the active power of the load according to the principle of power balance; The Halton sampling method is used to sample data samples of various operating states of the power system to obtain original input data samples of M different power generation combinations.

3. According to claim 2, a method for identifying security domain feature patterns based on capsule neural network is characterized in that: The step S1 also includes selecting features, which are as follows: A symmetrical fault is set for the power system, and the original input data samples of the M different power generation combinations extracted above are loaded. The active output of the corresponding generator is continuously changed by modifying the tempchange.dat file. The PSD-BPA software is called in batches using MATLAB to read the tempcase.out file to obtain a total of m features including the active output and reactive output of the generator, the active load and reactive load of the load node, the active and reactive power flows of the transmission line, and the bus voltage amplitude and phase angle. Whether the power system is stable is determined by whether the maximum relative power angle difference between the generators read at the end of the simulation is greater than 360°.

4. According to claim 1, a method for identifying security domain feature patterns based on capsule neural network is characterized in that: The objective function of the support vector machine of the recursive feature elimination algorithm of the support vector machine in step S2 is: Among them, ω is the weight vector, which represents the normal vector of the hyperplane; When the i-th pre-fault feature is deleted, i=1,2,···,N, the objective function changes to: ΔJ(i)≈ω i 2 ; Where ΔJ(i) represents the change of the objective function when the i-th pre-fault feature is deleted; ω i represents the weight vector of the i-th pre-fault feature; The kernel function used is a linear kernel function, and the sorting coefficient is: when When the value of is less than the set threshold, the corresponding pre-fault feature of the i-th fault is deleted.

5. The method for identifying security domain feature patterns based on capsule neural network according to claim 1 is characterized in that: The transient stability discrimination binary classification model in step S3 is specifically as follows: f(x)=sgn(ω T x+b); Among them, x is the sample feature, ω is the classification hyperplane coefficient, and b is the bias term; The loss function of the capsule neural network is as follows: L k =T k max(0,m + -||v k ||) 2 +λ(1-T k )max(0,||v k ||-m - ) 2 ; Where, L k is the loss of the kth capsule, T k is a label indicating whether the capsule should be activated (1 means it should be activated, 0 means it should not be activated), v k is the output vector of the kth capsule, m + and m - are normal and negative constant bounds, representing the minimum and maximum thresholds for capsule activation, respectively, and λ k is a scaling factor that adjusts the weight of the negative constant bounds.

6. The method for identifying security domain feature patterns based on capsule neural network according to claim 1 is characterized in that: The capsule neural network in step S3 comprises three network layers, namely a convolutional layer, a main capsule layer and a digital capsule layer; The number of input channels of the convolution layer is 1, the number of output channels is 256, the convolution kernel size is 9x9, the step size is 1, and the ReLU activation function is used; The number of input channels of the main capsule layer is 256, the number of output channels is 32, the capsule vector dimension is 8, the convolution kernel size is 9x9, and the stride is 2; The number of output channels of the digital capsule layer is 2, the capsule vector dimension is 16, the input capsule vector dimension is 8, and the output of the initial capsule layer is processed using a dynamic routing algorithm and a nonlinear activation function.

7. A security domain feature pattern recognition system based on capsule neural network, characterized in that: include: The data sample construction module is used to sample data samples of various operating states of the power system using the Halton sampling method to obtain original input data samples of M different power generation combinations, evenly covering the entire state space of the power grid; The feature screening module is used to screen out the top N pre-fault features that have the greatest influence on the objective function in the data samples based on the data samples obtained by sampling by using the recursive feature elimination algorithm of the support vector machine; And screen the K active features therein; The dynamic safety domain boundary classification module is used to convert one-dimensional data into two-dimensional data based on the screened K active features and input them into the capsule neural network to build a transient stability discrimination binary classification model. It is trained in two-dimensional and three-dimensional spaces to generate a dynamic safety domain boundary that divides the feature space into two subspaces: stable and unstable.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein when the program is executed, the device where the computer-readable storage medium is located is controlled to execute the security domain feature pattern recognition method based on capsule neural network according to any one of claims 1 to 6.

9. A processor, characterized in that: The processor is used to run a program, wherein the program, when running, executes the security domain feature pattern recognition method based on capsule neural network according to any one of claims 1 to 6.

Images