Boundary scene generation method and device for test verification of power regulation and control system

Abstract

The invention relates to a boundary scene generation method and device for test verification of a power regulation and control system, and the method comprises the steps: obtaining a power system operation mode sample, inputting a convolution denoising auto-encoder model fused with a self-attention mechanism, and extracting depth features capable of reflecting the characteristics of the power system operation mode; carrying out typical operation mode clustering identification on the depth features; calculating the distance between the depth feature of each power system operation mode sample and the depth feature of the typical operation mode identified by clustering, selecting a boundary sample point according to the distance, and decoding the boundary sample point to obtain a boundary scene; performing feature analysis on the boundary scene to obtain a boundary rule, and performing numerical expansion based on the boundary rule to obtain a boundary expansion scene; and introducing a hierarchical clustering algorithm to construct a hierarchical pedigree of a power grid operation scene by using a boundary expansion scene obtained by numerical expansion. According to the method, the typical operation mode conforming to the real physical scene can be extracted, and the generated test verification scene can effectively cover the boundary scene.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power dispatch automation, and particularly relates to a boundary scenario generation method and device for test verification of a power regulation system. BACKGROUND

[0002] Under the background of high penetration of new energy, the dynamic range of the operation state of the new power system is continuously expanding, the operation boundary is increasingly blurred, and strong uncertainty and scenario diversity are presented. Due to the multiple control objects of the new power system, the complexity of the mechanism is improved, and the regulation and control are difficult, which leads to frequent accidents of the regulation and control system at home and abroad, and the influence is large. The actual power grid scene is limited, and the test cost of the new regulation and control rules and strategies in the actual power grid is high, so it is necessary to construct a large number of scenarios to comprehensively test and verify the key applications of the regulation and control system. The large number of scenarios has the problem of long construction time in automatic test verification, and the trend of de-typing of the new power system leads to the fact that the typical scenarios cannot cover all the operating conditions of the new power system. More boundary test verification scenarios are needed as test verification cases, and the scenarios need to be managed, that is, the scenario categories are divided and the correlation between the scenarios is mined. SUMMARY

[0003] The purpose of the application is to solve the problem that the scenario generation method in the prior art cannot automatically extract typical operating modes that conform to the real physical scenario from the large amount of high-dimensional operating data, and the generated test verification scenarios often lack effective coverage of boundary scenarios, and cannot fully test the reliability of the control software under extreme conditions. A boundary scenario generation method and device for test verification of a power regulation system are provided.

[0004] In order to achieve the above purpose, the application has the following technical solutions: In a first aspect, a boundary scenario generation method for test verification of a power regulation system is provided, comprising: obtaining power system operating mode samples, inputting a pre-constructed convolutional denoising autoencoder model fused with a self-attention mechanism, and extracting deep features capable of reflecting the characteristics of the power system operating mode; performing typical operating mode clustering and identification on the deep features capable of reflecting the characteristics of the power system operating mode; calculating the distance between each power system operating mode sample deep feature and the typical operating mode deep feature identified by clustering, selecting a boundary sample point according to the distance, and decoding the boundary sample point to obtain a boundary scenario; performing feature analysis on the boundary scenario to obtain boundary rules, and performing numerical expansion based on the boundary rules to obtain a boundary expanded scenario; using the boundary expanded scenario obtained by numerical expansion, introducing a hierarchical clustering algorithm to construct a hierarchical pedigree of the power grid operating scenario.

[0005] As a preferred scheme, the convolutional denoising self-encoder model fused with the self-attention mechanism is trained separately, and the target of the model training is to restore the original data from the input power system operation mode sample data with noise information.

[0006] As a preferred scheme, in the step of obtaining the power system operation mode sample, inputting the pre-constructed convolutional denoising self-encoder model fused with the self-attention mechanism, and extracting the deep features capable of reflecting the characteristics of the power system operation mode, the power system operation mode sample is composed of the load active power, new energy and traditional energy active power output of different regions at the same time as the characteristic variables. The sample is divided according to the region, and any operation mode sample of the pre-processed operation mode sample set is P m , and the expression is as follows:

[0007] In the formula, i represents the region number, w represents the operation mode combination corresponding to the wind power; pv represents the operation mode combination corresponding to the photovoltaic; d represents the operation mode combination corresponding to the load; h represents the operation mode combination corresponding to the hydropower; th represents the operation mode combination corresponding to the thermal power; The loss function of the convolutional denoising self-encoder model fused with the self-attention mechanism during training is L rec The mean square error (MSE) of the unnormalized reconstructed output data and the normalized original input data, and the calculation expression is as follows: .

[0008] As a preferred scheme, in the step of performing typical operation mode clustering and identification on the deep features capable of reflecting the characteristics of the power system operation mode, the optimal cluster number is selected by using an evaluation index system combining technical and business indexes; The contour coefficient is selected as the evaluation index on the technical level, and the evaluation index measures the rationality of the sample attribution in the clustering result by combining the cohesion and separation degrees of the sample; on the business level, the business index is defined according to the scene generation requirement, and the clustering algorithm based on the evaluation index defined by the scene generation requirement is used to determine the optimal cluster number for feature clustering.

[0009] As a preferred approach, in the step of clustering and identifying typical operating modes based on deep features that reflect the characteristics of power system operation modes, three typical indicators are defined from the perspectives of new energy sources, load, and power flow direction. The maximum active power variance of new energy sources within a cluster, the maximum active power variance of loads within a cluster, and the minimum consistency rate of active power direction at cross-sections within a cluster are used as evaluation indicators for clustering performance. The calculation expressions are as follows: The silhouette coefficient of the clustering algorithm for technical indicators is calculated using the following formula:

[0010] In the formula, a ( i ) indicates a sample i The average Euclidean distance to all other samples in the same cluster reflects the cluster density; the smaller the value, the more concentrated the points in the cluster. b ( i ) indicates a sample i The minimum average Euclidean distance to all samples in other clusters reflects the inter-cluster separation; the larger the value, the more distinct the separation from other clusters. Calculate the maximum active power variance of new energy within the business indicator category using the following formula:

[0011] In the formula, S 2 n,max This represents the variance of the maximum active power output of new energy sources across all categories, reflecting the maximum volatility of active power output of new energy sources. The smaller the value, the smaller the overall intra-category active power volatility of new energy sources across all categories. Calculate the maximum variance of the load within the business indicator class using the following formula:

[0012] In the formula, S 2 d,max This represents the maximum variance of the load across all categories, reflecting the maximum variability of the load. The smaller the value, the smaller the overall intra-category load variability across all categories. The minimum consistency rate of active power direction within the same business indicator category is calculated using the following formula:

[0013] In the formula, num z,all num represents the total number of samples in the z-th type of operating mode cluster. z,lcon This represents the number of samples in the z-th type of operation mode cluster whose active power direction at the operation mode section is consistent with the class center. The min() function represents the minimum value among the Z categories, reflecting the minimum active power direction consistency rate of the cross-sections across all categories. The larger the value, the higher the overall intra-category active power direction consistency rate across all categories.

[0014] As a preferred approach, in the step of clustering and identifying typical operating modes based on deep features that reflect the characteristics of power system operation modes, the two indicators, the maximum active power variance of renewable energy sources within a cluster and the maximum load variance within a cluster, are preprocessed by taking their reciprocals. Then, the four indicators—the clustering algorithm silhouette coefficient, the maximum active power variance of renewable energy sources within a cluster, the maximum load variance within a cluster, and the minimum consistency rate of active power direction at the cross-section within a cluster—are normalized and arithmetically averaged to obtain a comprehensive average index. The K value corresponding to the optimal comprehensive average index is selected as the final number of clusters. The following steps are used to determine the physically meaningful typical operating modes: Calculate the cluster center of each cluster, and then for each cluster, calculate the distance from all sample points within the cluster to the cluster center. Select the closest real historical sample point as the deep feature representing the operating mode of the corresponding cluster. The deep features are decoded to obtain the original, unnormalized data, which is then denormalized to obtain a representative sample of the operating mode. The distance metric used is Euclidean distance.

[0015] As a preferred embodiment, in the steps of calculating the distance between the depth features of each power system operation mode sample and the depth features of the typical operation mode identified by clustering, selecting boundary sample points based on the distance, and decoding the boundary sample points to obtain the boundary scene, the Euclidean distance from all sample points to the cluster center of their respective categories is calculated, and the samples are sorted from largest to smallest. Several sample points are selected based on a set distance criterion, and the selected sample points are initially defined as boundary sample points under the corresponding typical operation mode; or, a distance threshold is set according to the business context, and boundary sample points are selected according to the set distance threshold; the selected boundary sample points are decoded and inversely normalized to obtain the boundary scene.

[0016] As a preferred approach, in the step of performing feature analysis on the boundary scene to obtain boundary rules, and then performing numerical expansion based on the boundary rules to obtain the boundary extended scene, the feature analysis adopts a combination of decision tree algorithm and rule extraction method. The obtained boundary scene is used as a sample with negative label, and the other scenes are used as samples with positive label and put into the decision tree model for binary classification training. The rule extraction method is used to obtain the judgment rule of the boundary scene and convert it into if-then rule. The judgment rule of the boundary scene directly gives the key feature combination that constitutes the boundary scene. In the boundary extended scene generation stage, more extreme combinations are constructed to become extreme extended scenes.

[0017] As a preferred embodiment, the step of using the boundary expansion scenario obtained by numerical expansion and introducing a hierarchical clustering algorithm to construct a hierarchical spectrum of power grid operation scenarios includes: based on the initial clustering results, using a hierarchical clustering algorithm to merge similar initial clusters to form a hierarchical tree diagram, revealing the correlation and potential evolution paths between different scenarios; In the tree diagram, the horizontal axis represents the initial classification scene category number, and the vertical axis represents the relative distance, intuitively expressing the distance when merging categories; the lower the vertical axis value, the earlier the operation mode is merged, and the higher the vertical axis value, the later the operation mode is merged. By combining the generated boundary scenarios, boundary extension scenarios, and typical operating modes, a test and verification scenario library is constructed to test and verify the control software under extreme conditions and boundary performance.

[0018] Secondly, a boundary scenario generation system for power control system testing and verification is provided, including: The deep feature extraction module is used to obtain power system operation mode samples, input a pre-built convolutional denoising autoencoder model with self-attention mechanism, and extract deep features that can reflect the characteristics of power system operation mode. The typical operation mode clustering and identification module is used to perform typical operation mode clustering and identification on deep features that can reflect the characteristics of power system operation modes; The boundary scene acquisition module is used to calculate the distance between the depth features of the sample of each power system operation mode and the depth features of the typical operation mode identified by clustering, select boundary sample points according to the distance, and decode the boundary sample points to obtain the boundary scene; The boundary expansion scene acquisition module is used to perform feature analysis on the boundary scene to obtain boundary rules, and to perform numerical expansion based on the boundary rules to obtain the boundary expansion scene. The hierarchical spectrum construction module is used to construct a hierarchical spectrum of the power grid operation scenario by using the boundary expansion scenario obtained by numerical expansion and introducing a hierarchical clustering algorithm.

[0019] As a preferred approach, the deep feature extraction module acquires power system operation mode samples and inputs them into a pre-built convolutional denoising autoencoder model with a self-attention mechanism. When extracting deep features that can reflect the characteristics of power system operation mode, the convolutional denoising autoencoder model with a self-attention mechanism is trained separately. The goal of model training is to recover the original data from the input power system operation mode sample data with added noise information.

[0020] As a preferred approach, the deep feature extraction module acquires power system operation mode samples and inputs a pre-built convolutional denoising autoencoder model with a self-attention mechanism to extract deep features that can reflect the characteristics of power system operation mode. The power system operation mode samples are composed of load active power, new energy and traditional energy active power output at the same historical moment in different regions as feature variables. The samples are divided into regions, and any sample of a particular operating mode after preprocessing is... P m The expression is as follows:

[0021] In the formula, i Indicates the area code. w This indicates the combination of operating modes corresponding to wind power; pv This indicates the combination of operating modes corresponding to photovoltaics; d This indicates the combination of operating modes corresponding to the load; h This indicates the combination of operating modes corresponding to hydropower; th This indicates the combination of operating modes corresponding to thermal power plants; The loss function of the convolutional denoising autoencoder model that incorporates a self-attention mechanism during training is... L rec The mean square error (MSE) between the reconstructed output data (without inverse normalization) and the normalized original input data is calculated using the following expression: .

[0022] As a preferred option, the clustering identification module in the typical operation mode uses an evaluation index system that combines technical and business indicators to select the optimal number of clusters; From a technical perspective, the silhouette coefficient is chosen as the evaluation metric. This metric measures the rationality of sample attribution in the clustering results by combining the cohesion and separation of the samples. From a business perspective, custom business metrics are generated based on scenario requirements, including feature clustering based on clustering algorithms that determine the optimal number of clusters using evaluation metrics defined by scenario requirements.

[0023] As a preferred approach, the typical operation mode clustering identification module defines three typical indicators from the perspectives of new energy sources, load, and power flow direction. It uses the maximum active power variance of new energy sources within a cluster, the maximum active power variance of loads within a cluster, and the minimum consistency rate of active power direction at cross-sections within a cluster as evaluation indicators for clustering performance. The calculation expressions are as follows: The silhouette coefficient of the clustering algorithm for technical indicators is calculated using the following formula:

[0024] In the formula, a ( i) indicates a sample i The average Euclidean distance to all other samples in the same cluster reflects the cluster density; the smaller the value, the more concentrated the points in the cluster. b ( i ) indicates a sample i The minimum average Euclidean distance to all samples in other clusters reflects the inter-cluster separation; the larger the value, the more distinct the separation from other clusters. Calculate the maximum active power variance of new energy within the business indicator category using the following formula:

[0025] In the formula, S 2 n,max This represents the variance of the maximum active power output of new energy sources across all categories, reflecting the maximum volatility of active power output of new energy sources. The smaller the value, the smaller the overall intra-category active power volatility of new energy sources across all categories. Calculate the maximum variance of the load within the business indicator class using the following formula:

[0026] In the formula, S 2 d,max This represents the maximum variance of the load across all categories, reflecting the maximum variability of the load. The smaller the value, the smaller the overall intra-category load variability across all categories. The minimum consistency rate of active power direction within the same business indicator category is calculated using the following formula:

[0027] In the formula, num z,all num represents the total number of samples in the z-th type of operating mode cluster. z,lcon This represents the number of samples in the z-th type of operation mode cluster whose active power direction at the operation mode section is consistent with the class center. The min() function represents the minimum value among the Z categories, reflecting the minimum active power direction consistency rate of the cross-sections across all categories. The larger the value, the higher the overall intra-category active power direction consistency rate across all categories.

[0028] As a preferred approach, the typical operation mode clustering identification module preprocesses two indicators—the maximum active power variance of new energy sources within a cluster and the maximum load variance within a cluster—by taking their reciprocals. Then, it normalizes four indicators—the clustering algorithm silhouette coefficient, the maximum active power variance of new energy sources within a cluster, the maximum load variance within a cluster, and the minimum consistency rate of active power direction at cross-sections within a cluster—and calculates an arithmetic average index. The K value corresponding to the optimal comprehensive average index is selected as the final number of clusters. The following steps are used to determine physically meaningful typical operation modes: Calculate the cluster center of each cluster, and then for each cluster, calculate the distance from all sample points within the cluster to the cluster center. Select the closest real historical sample point as the deep feature representing the operating mode of the corresponding cluster. The deep features are decoded to obtain the original, unnormalized data, which is then denormalized to obtain a representative sample of the operating mode. The distance metric used is Euclidean distance.

[0029] As a preferred approach, the boundary scene acquisition module calculates the Euclidean distance from all sample points to the corresponding cluster center of their respective categories, sorts them from largest to smallest distance, selects several sample points based on a set distance criterion, and preliminarily defines the selected sample points as boundary sample points under the corresponding typical operating mode; alternatively, a distance threshold is set according to the business scenario, and boundary sample points are selected according to the set distance threshold; the selected boundary sample points are decoded and inversely normalized to obtain the boundary scene.

[0030] As a preferred embodiment, the boundary expansion scene acquisition module performs feature analysis on the boundary scene to obtain boundary rules. When expanding the boundary scene based on the numerical expansion of the boundary rules, the feature analysis adopts a combination of decision tree algorithm and rule extraction method. The obtained boundary scene is used as a sample with negative label, and the other scenes are used as samples with positive label and put into the decision tree model for binary classification training. The rule extraction method is used to obtain the judgment rule of the boundary scene and convert it into if-then rule. The judgment rule of the boundary scene directly gives the key feature combination that constitutes the boundary scene. In the boundary expansion scene generation stage, more extreme combinations are constructed to become extreme expansion scenes.

[0031] As a preferred embodiment, the hierarchical genealogy construction module, based on the initial clustering results, uses a hierarchical clustering algorithm to merge similar initial clusters to form a hierarchical tree diagram, revealing the correlation and potential evolutionary paths between different scenarios; In the tree diagram, the horizontal axis represents the initial classification scene category number, and the vertical axis represents the relative distance, intuitively expressing the distance when merging categories; the lower the vertical axis value, the earlier the operation mode is merged, and the higher the vertical axis value, the later the operation mode is merged. By combining the generated boundary scenarios, boundary extension scenarios, and typical operating modes, a test and verification scenario library is constructed to test and verify the control software under extreme conditions and boundary performance.

[0032] Thirdly, an electronic device is provided, including a processor and a memory, wherein the processor is used to execute a computer program stored in the memory to implement the boundary scenario generation method for experimental verification of power control systems.

[0033] Fourthly, a computer-readable storage medium is provided, the computer-readable storage medium storing at least one instruction, which, when executed by a processor, implements the boundary scenario generation method for experimental verification of power control systems.

[0034] Compared with the prior art, the first aspect of the present invention has at least the following beneficial effects: This invention provides a boundary scene generation method for power control system testing and verification. In the feature extraction stage, it employs a convolutional denoising autoencoder model with a self-attention mechanism for representation learning, extracting deep features of the operating mode. Compared to traditional autoencoder models, this model improves performance in learning essential features and produces denser features that reflect spatiotemporal correlations. The model learns robust features, even with minor perturbations in the original data, and outputs a dense feature vector reflecting the operating mode characteristics, providing deep features for subsequent clustering analysis. In the clustering analysis stage, typical operating modes are determined by associating and decoding specific, historically existing sample points, overcoming the drawback of traditional clustering algorithms where cluster centers are often abstract mathematical points. This ensures that each representative operating mode corresponds to an actual operating state that is understandable and traceable to power grid dispatchers, possessing clear physical meaning and engineering interpretability. This not only enhances the intuitiveness and credibility of the model output results, but also enables analyses based on typical modes (such as power flow calculations) to be directly applied to actual power grid simulations, thereby improving the engineering practical value of the research results.

[0035] This invention proposes a boundary scenario generation method for power control system testing and verification that covers extreme operating conditions. By calculating the distance between samples and cluster centers, boundary sample points in the feature space are accurately located based on this distance and decoded to form boundary scenarios. This ensures that the generated boundary scenarios are not only statistical outliers but also physically real critical conditions near safe operating boundaries. Furthermore, to cover more extreme conditions in the testing and verification, the boundary scenarios undergo feature analysis, rule extraction, and numerical extension to obtain extended boundary scenarios. Using these extended boundary scenarios for the testing and verification of control software can more effectively expose the software's handling capabilities and potential defects under extreme and critical conditions, thereby significantly improving the software's robustness and reliability.

[0036] This invention introduces a hierarchical clustering algorithm into the boundary scenario generation method for power grid operation scenarios, constructing a hierarchical spectrum of power grid operation scenarios. This method brings two core benefits: first, it enhances the clustering effect, revealing more stable and macroscopic categories of operation modes through category merging; second, it clearly depicts the correlation and evolution path between different categories of scenarios. This provides an intuitive perspective on the evolution of the power grid from one scenario mode to another (e.g., from a "new energy-dominated" mode to a "traditional energy-dominated" mode), which helps in the construction of a hierarchical scenario library for control software testing and verification and proactively identifies systemic evolutionary risks.

[0037] It is understood that the beneficial effects of the second to fourth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 Flowchart of the boundary scenario generation method for power control system test verification according to an embodiment of the present invention; Figure 2 Architecture diagram of a convolutional denoising autoencoder model incorporating a self-attention mechanism according to an embodiment of the present invention; Figure 3 Example diagram of clustering identification in a typical operating mode with physical practical significance in the embodiments of the present invention; Figure 4 Example diagram of boundary scene generation in this embodiment of the invention; Figure 5 Example diagram of hierarchical clustering in this invention; Figure 6 Example diagram of hierarchical management of boundary scenes according to an embodiment of the present invention; Figure 7 This invention provides a schematic diagram illustrating the construction of a hierarchical control software test and verification scenario library for a specific province, city, or region. Detailed Implementation

[0040] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0041] Key applications of power control systems mainly include advanced applications of the power system energy management system (EMS) and automatic control software. The advanced applications of EMS mainly include core modules such as state estimation and static security analysis, while the automatic control software mainly includes automatic generation control and automatic voltage control software. When the system detects certain specific conditions (such as frequency drop, voltage anomaly, or equipment failure), these software programs will immediately trigger preset control logic, issue control commands, and adjust the generator, load, or grid structure to restore the system to steady state.

[0042] Boundary scenarios refer to scenarios formed by decoding real sample points located at the boundary of the characteristic space of typical operating modes, extracted from the deep features of historical power grid operation data. These scenarios themselves are safe in historical operation, but they represent the critical state of the system's safety boundary under a certain operating mode.

[0043] This invention proposes a hierarchical clustering boundary scene generation method for key application testing and verification in power control systems. First, a convolutional denoising autoencoder model incorporating a self-attention mechanism is constructed. Historical power grid operation data may contain noise and other disturbances during collection due to inherent errors in measuring devices, packet loss during transmission, and other factors. This denoising autoencoder is trained separately, with the goal of recovering the original data from the noise-injected input. The aim is to extract deep features that characterize the essence of the operation mode from high-dimensional, complex power grid operation data, fully capturing the complex spatiotemporal correlations between variables. Based on the extracted deep features, a clustering algorithm that automatically determines the optimal number of clusters based on evaluation metrics is used for feature clustering. Subsequently, the actual historical sample points closest to the cluster centers of each cluster are selected as representatives of that class, and they are decoded to obtain typical power grid operation modes. The distances from all sample points to their respective cluster centers are calculated and sorted in descending order of distance. The group of sample points furthest apart in each category is selected and initially defined as the boundary sample points under this typical operating mode. These are decoded to determine the valid "boundary scenarios," and feature analysis and numerical expansion are performed on these boundary scenarios to form extended boundary scenarios. A hierarchical clustering algorithm is then used to merge similar initial clusters. The extended boundary scenarios, boundary scenarios, and physically interpretable typical operating modes are combined to form a multi-layered, comprehensive scenario library for experimental verification.

[0044] Please see Figure 1 The boundary scenario generation method for power control system testing and verification according to embodiments of the present invention mainly includes: S1. Obtain power system operation mode samples, input them into a pre-built convolutional denoising autoencoder model with self-attention mechanism, and extract deep features that can reflect the characteristics of power system operation mode; S2. Clustering and identifying typical operating modes based on deep features that can reflect the characteristics of power system operation modes; S3. Calculate the distance between the depth features of each power system operation mode sample and the depth features of the typical operation mode identified by clustering, select boundary sample points based on the distance, and decode the boundary sample points to obtain the boundary scene; S4. Perform feature analysis on the boundary scene to obtain boundary rules, and perform numerical extension based on the boundary rules to obtain the boundary extended scene; S5. Using the boundary extension scenario obtained by numerical extension, a hierarchical clustering algorithm is introduced to construct a hierarchical spectrum of the power grid operation scenario.

[0045] In one possible implementation, the convolutional denoising autoencoder model architecture incorporating a self-attention mechanism according to embodiments of the present invention is as follows: Figure 2As shown, the convolutional denoising autoencoder model with self-attention mechanism described in step S1 is trained separately. The goal of the model training is to recover the original data from the input power system operation mode sample data with added noise information.

[0046] For the operation mode sample, the power system operation mode is the operating state of the power system determined by a combination of power system elements such as generator output, line power flow, and node load. The active power output of load, renewable energy, and traditional energy at the same historical moment in a province or region is selected as feature variables to construct the operation mode sample. The sample is divided according to region, and any operation mode sample after preprocessing is... P m The expression is as follows:

[0047] In the formula, i Indicates the area code. w This indicates the combination of operating modes corresponding to wind power; pv This indicates the combination of operating modes corresponding to photovoltaics; d This indicates the combination of operating modes corresponding to the load; h This indicates the combination of operating modes corresponding to hydropower; th This indicates the combination of operating modes corresponding to thermal power plants.

[0048] This invention, based on a denoising auto-encoder (DAE) model, intelligently extracts the complex spatial coupling relationships among operating variables of high-proportion renewable energy power grids, and preprocesses the operating mode sequence. P =[ P 1 T , P 2 T , ..., P m T ] T This serves as input data for the algorithm. Noise is added to the original data, and the algorithm learns to recover the original data, thus preserving as many deep features as possible that reflect the characteristics of the operation, providing deep features for subsequent clustering algorithm layers.

[0049] The loss function of the convolutional denoising autoencoder model incorporating self-attention mechanism in this invention during training. L rec The mean square error (MSE) between the reconstructed output data (without inverse normalization) and the normalized original input data is calculated using the following expression: .

[0050] In one possible implementation, please refer toFigure 3 Based on the deep features extracted in step S1, a custom method for evaluating clustering performance using business indicators is adopted. This method selects the optimal number of clusters using an evaluation indicator system that combines technical and business indicators. Technically, the silhouette coefficient is selected as the evaluation indicator. This indicator measures the rationality of sample attribution in the clustering results by combining the cohesion and separation of samples. At the business level, custom business indicators are defined according to scenario generation requirements. Specifically, feature clustering is performed based on a clustering algorithm that determines the optimal number of clusters using evaluation indicators defined by scenario generation requirements. This invention defines three typical indicators from three perspectives: new energy, load, and power flow direction. To ensure that each category satisfies the requirements of small active power fluctuations in new energy sources, small load fluctuations, and high consistency rate of power flow direction within the cluster, the maximum active power variance of new energy sources within the cluster, the maximum load variance within the cluster, and the minimum consistency rate of active power direction at cross-sections within the cluster are used instead of the average measurement method. Subsequently, the actual historical sample points closest to the cluster centers of each cluster are selected as representatives of that category and decoded to obtain a physically meaningful typical power grid operation mode. First, the optimal number of clusters K is determined. A range of K values ​​is preset. For each candidate K value, a clustering algorithm (K-Means) is run to divide the data into K clusters. Then, for each K, a clustering performance evaluation index is calculated, including the index and its corresponding formula as follows: The silhouette coefficient of the clustering algorithm for technical indicators is calculated using the following formula:

[0051] In the formula, a ( i ) indicates a sample i The average Euclidean distance to all other samples in the same cluster reflects the cluster density; the smaller the value, the more concentrated the points in the cluster. b ( i ) indicates a sample i The minimum average Euclidean distance to all samples in other clusters reflects the inter-cluster separation; the larger the value, the more distinct the separation from other clusters. Calculate the maximum active power variance of new energy within the business indicator category using the following formula:

[0052] In the formula, S 2 n,max This represents the variance of the maximum active power output of new energy sources across all categories, reflecting the maximum volatility of active power output of new energy sources. The smaller the value, the smaller the overall intra-category active power volatility of new energy sources across all categories. Calculate the maximum variance of the load within the business indicator class using the following formula:

[0053] In the formula, S 2d,max This represents the maximum variance of the load across all categories, reflecting the maximum variability of the load. The smaller the value, the smaller the overall intra-category load variability across all categories. The minimum consistency rate of active power direction within the same business indicator category is calculated using the following formula:

[0054] In the formula, num z,all num represents the total number of samples in the z-th type of operating mode cluster. z,lcon This represents the number of samples in the z-th type of operation mode cluster whose active power direction at the operation mode section is consistent with the class center. The min() function represents the minimum value among the Z categories, reflecting the minimum active power direction consistency rate of the cross-sections across all categories. The larger the value, the higher the overall intra-category active power direction consistency rate across all categories.

[0055] Furthermore, in step S2 of this embodiment of the invention, the two indicators, the maximum active power variance of new energy sources within a class and the maximum load variance within a class, are preprocessed by taking their reciprocals. Then, the four indicators, namely the clustering algorithm contour coefficient, the maximum active power variance of new energy sources within a class, the maximum load variance within a class, and the minimum consistency rate of active power direction of cross sections within a class, are normalized and arithmetically averaged to obtain a comprehensive average index. The K value corresponding to the optimal comprehensive average index is selected as the final number of clusters. Secondly, the following steps are used to determine the typical operating modes with physical significance: Calculate the cluster center of each cluster, and then for each cluster, calculate the distance from all sample points within the cluster to the cluster center. Select the closest real historical sample point as the deep feature representing the operating mode of the corresponding cluster. The deep features are decoded using the decoder structure trained in step S1 to obtain the recovered original data that has not been denormalized. After denormalization, a representative sample of the operating mode is obtained. The distance metric used is Euclidean distance.

[0056] In one possible implementation, please refer to Figure 4 Step S3 calculates the Euclidean distance from all sample points to the cluster center of their respective categories, sorts them in descending order of distance, selects the sample points with the farthest distance in each category, and preliminarily defines the selected sample points as the boundary sample points under the corresponding typical operating mode; or, sets a distance threshold according to the business scenario, and selects boundary sample points according to the set distance threshold; decodes the selected boundary sample points through the decoder structure trained in step S1, and performs inverse normalization processing to obtain the boundary scene.

[0057] In one possible implementation, step S4 involves boundary scene feature analysis and boundary expansion scene generation. This step primarily involves performing feature analysis on the obtained boundary scenes to derive boundary rules, i.e., analyzing the reasons for these sample boundaries, and extending the boundary scenes with business numerical values ​​to generate more extreme boundary expansion scenes. This supports the comprehensive experimental verification of key control software using a scenario use case library containing different extreme levels. In the feature analysis stage, a combination of decision tree algorithm and rule extraction method is used. First, the boundary scenes obtained in step S3 are treated as samples with negative labels, and the remaining scenes are treated as samples with positive labels. These are then fed into the decision tree model for binary classification training. The rule extraction method is then used to obtain the judgment rules for the boundary scenes, which are then converted into if-then rules, as shown in the example below:

[0058] These rules directly provide the key feature combinations that constitute the boundary scenario. In the boundary expansion scenario generation stage, more extreme combinations are constructed, such as increasing one of the business numerical variables by a certain proportion to create an extreme expansion scenario.

[0059] In one possible implementation, step S5 of this embodiment of the invention further employs a hierarchical clustering algorithm to merge similar initial clusters based on the initial clustering results. The core idea of ​​hierarchical clustering is to form a hierarchical nested tree (dendrogram) by continuously merging or splitting based on the distance between sample points. This step aims to enhance the clustering effect and construct a hierarchical scene hierarchy, thereby revealing the correlation and potential evolutionary paths between different scenes. An example of hierarchical clustering is shown below. Figure 5 As shown, the horizontal axis represents the initial classification scenario category number, and the vertical axis represents the relative distance, intuitively representing the distance during category merging. The vertical axis of hierarchical clustering constructs a clear hierarchical structure. Low vertical axis values ​​represent very similar, subdivided operating modes (such as the "high new energy - light load" mode at different time points). These are merged first. High vertical axis values ​​represent more differentiated, macroscopic categories of operating modes (such as "new energy-dominated" and "traditional energy-dominated"), which are merged later. Hierarchical clustering methods can construct hierarchical tree diagrams to hierarchically manage boundary scenarios, such as... Figure 6As shown. The boundary scenarios and boundary extension scenarios generated in the above steps are combined with physically interpretable typical operating modes to form a multi-layered, comprehensive scenario library for experimental verification, specifically used for comprehensive experimental verification of the control software under extreme conditions and boundary performance. For the set of operating states of the entire region or a specific province / city, the set of all operating states in that region is first used as the root node. Then, using the boundary scenario generation method, it is subdivided into three categories as the second level: boundary scenarios, boundary extension scenarios, and normal scenarios. Next, hierarchical clustering is used to generate hierarchical experimental verification scenarios for that province / city under the boundary scenario, boundary extension scenario, and normal scenario types. A specific example of constructing a hierarchical control software experimental verification scenario library for a specific province / city or region is shown below. Figure 7 As shown.

[0060] This invention presents a boundary scene generation method for power control system testing and verification. First, at the deep feature extraction layer, a convolutional denoising autoencoder with integrated self-attention mechanism is used to mine and capture local spatiotemporal patterns and learn robust, essential deep-level features from power grid operation data of a province, city, or region. Then, at the clustering layer, a custom method using business evaluation metrics for clustering effectiveness is employed to adaptively determine the optimal number of clusters for the clustering algorithm, accurately dividing the dimensionality-reduced features. This method associates real data points with cluster centers, giving the acquired typical power grid operation modes clear physical meaning and significantly improving the model's interpretability. Based on this, the model measures the distance between samples and cluster centers, selecting the farthest boundary sample points and decoding them to form effective boundary scenes. Feature analysis and numerical expansion of the boundary scenes are then performed to form more extreme extended boundary scenes. Furthermore, the framework introduces hierarchical clustering technology to integrate the initial clustering results, not only optimizing clustering quality but also clearly revealing the potential correlations and evolution paths between different operation mode categories.

[0061] This invention proposes an optimized denoising autoencoder model for representation learning, enabling the learning of deeper robust characteristics of historical operating modes. It also proposes a method for generating physically realistic typical operating modes for testing and verification of control software. Typical operating modes are obtained by decoding real historical sample points with clear physical meaning, constructing a test and verification benchmark that highly reflects actual power grid operation. This ensures that the test and verification cases not only cover various mainstream operating states but also possess strong engineering interpretability, improving the practical value and credibility of the test and verification results.

[0062] This invention proposes a boundary scenario generation method for experimental verification of control software. Boundary scenarios are generated by calculating the embedding distance with cluster centers, and further, extended boundary scenarios are generated. This effectively extends the experimental verification scope from normal operating conditions to the critical boundary of system safe operation, and further to more extreme states. It proactively derives the system's safety boundaries and vulnerabilities from massive historical data, enabling proactive prevention before accidents occur. It can accurately expose the performance shortcomings of control software under extreme and critical conditions, significantly improving the depth of experimental verification and the robustness of the software.

[0063] This invention proposes a hierarchical clustering-based scenario library construction method for experimental verification of control software. By introducing a hierarchical clustering-based operational mode hierarchy, it clearly reveals the correlations and evolutionary patterns among different operational modes. This further strengthens the hierarchical management of experimental verification scenarios for control software, enabling experimental verification design to evolve from isolated scenario verification to a comprehensive examination of the system's evolutionary path. This provides crucial support for evaluating the adaptability and cascading failure risks of software during operational mode transitions.

[0064] Another embodiment of the present invention proposes a boundary scenario generation system for experimental verification of power control systems, comprising: The deep feature extraction module is used to obtain power system operation mode samples, input a pre-built convolutional denoising autoencoder model with self-attention mechanism, and extract deep features that can reflect the characteristics of power system operation mode. The typical operation mode clustering and identification module is used to perform typical operation mode clustering and identification on deep features that can reflect the characteristics of power system operation modes; The boundary scene acquisition module is used to calculate the distance between the depth features of the sample of each power system operation mode and the depth features of the typical operation mode identified by clustering, select boundary sample points according to the distance, and decode the boundary sample points to obtain the boundary scene; The boundary expansion scene acquisition module is used to perform feature analysis on the boundary scene to obtain boundary rules, and to perform numerical expansion based on the boundary rules to obtain the boundary expansion scene. The hierarchical spectrum construction module is used to construct a hierarchical spectrum of the power grid operation scenario by using the boundary expansion scenario obtained by numerical expansion and introducing a hierarchical clustering algorithm.

[0065] In one possible implementation, the deep feature extraction module acquires power system operation mode samples and inputs them into a pre-built convolutional denoising autoencoder model with a self-attention mechanism. When extracting deep features that can reflect the characteristics of power system operation mode, the convolutional denoising autoencoder model with a self-attention mechanism is trained separately. The goal of the model training is to recover the original data from the input power system operation mode sample data with added noise information.

[0066] In one possible implementation, the deep feature extraction module acquires power system operation mode samples, inputs a pre-built convolutional denoising autoencoder model with a self-attention mechanism, and extracts deep features that can reflect the characteristics of power system operation mode. The power system operation mode samples are composed of active power output of load, new energy and traditional energy at the same historical moment in different regions as feature variables. The samples are divided into regions, and any sample of a particular operating mode after preprocessing is... P m The expression is as follows:

[0067] In the formula, i Indicates the area code. w This indicates the combination of operating modes corresponding to wind power; pv This indicates the combination of operating modes corresponding to photovoltaics; d This indicates the combination of operating modes corresponding to the load; h This indicates the combination of operating modes corresponding to hydropower; th This indicates the combination of operating modes corresponding to thermal power plants; The loss function of the convolutional denoising autoencoder model that incorporates a self-attention mechanism during training is... L rec The mean square error (MSE) between the reconstructed output data (without inverse normalization) and the normalized original input data is calculated using the following expression: .

[0068] In one possible implementation, the typical operation mode clustering identification module uses an evaluation index system that combines technical and business indicators to select the optimal number of clusters. From a technical perspective, the silhouette coefficient is chosen as the evaluation metric. This metric measures the rationality of sample attribution in the clustering results by combining the cohesion and separation of the samples. From a business perspective, custom business metrics are generated based on scenario requirements, including feature clustering based on clustering algorithms that determine the optimal number of clusters using evaluation metrics defined by scenario requirements.

[0069] In one possible implementation, the typical operation mode clustering identification module defines three typical indicators from the perspectives of new energy sources, load, and power flow direction. It uses the maximum active power variance of new energy sources within a cluster, the maximum active power variance of loads within a cluster, and the minimum consistency rate of active power direction at cross-sections within a cluster as evaluation indicators for clustering performance. The calculation expressions are as follows: The silhouette coefficient of the clustering algorithm for technical indicators is calculated using the following formula:

[0070] In the formula, a ( i ) indicates a sample i The average Euclidean distance to all other samples in the same cluster reflects the cluster density; the smaller the value, the more concentrated the points in the cluster. b ( i ) indicates a sample i The minimum average Euclidean distance to all samples in other clusters reflects the inter-cluster separation; the larger the value, the more distinct the separation from other clusters. Calculate the maximum active power variance of new energy within the business indicator category using the following formula:

[0071] In the formula, S 2 n,max This represents the variance of the maximum active power output of new energy sources across all categories, reflecting the maximum volatility of active power output of new energy sources. The smaller the value, the smaller the overall intra-category active power volatility of new energy sources across all categories. Calculate the maximum variance of the load within the business indicator class using the following formula:

[0072] In the formula, S 2 d,max This represents the maximum variance of the load across all categories, reflecting the maximum variability of the load. The smaller the value, the smaller the overall intra-category load variability across all categories. The minimum consistency rate of active power direction within the same business indicator category is calculated using the following formula:

[0073] In the formula, num z,all num represents the total number of samples in the z-th type of operating mode cluster. z,lcon This represents the number of samples in the z-th type of operation mode cluster whose active power direction at the operation mode section is consistent with the class center. The min() function represents the minimum value among the Z categories, reflecting the minimum active power direction consistency rate of the cross-sections across all categories. The larger the value, the higher the overall intra-category active power direction consistency rate across all categories.

[0074] In one possible implementation, the typical operation mode clustering identification module preprocesses two indicators—the maximum active power variance of new energy sources within a cluster and the maximum load variance within a cluster—by taking their reciprocals. Then, it normalizes four indicators—the clustering algorithm silhouette coefficient, the maximum active power variance of new energy sources within a cluster, the maximum load variance within a cluster, and the minimum consistency rate of active power direction at cross-sections within a cluster—and calculates an arithmetic average index. The K value corresponding to the optimal comprehensive average index is selected as the final number of clusters. The following steps are used to determine physically meaningful typical operation modes: Calculate the cluster center of each cluster, and then for each cluster, calculate the distance from all sample points within the cluster to the cluster center. Select the closest real historical sample point as the deep feature representing the operating mode of the corresponding cluster. The deep features are decoded to obtain the original, unnormalized data, which is then denormalized to obtain a representative sample of the operating mode. The distance metric used is Euclidean distance.

[0075] In one possible implementation, the boundary scene acquisition module calculates the Euclidean distance from all sample points to the corresponding cluster center of their respective categories, sorts them from largest to smallest distance, selects several sample points based on a set distance criterion, and preliminarily defines the selected sample points as boundary sample points under the corresponding typical operating mode; or, a distance threshold is set according to the business scenario, and boundary sample points are selected according to the set distance threshold; the selected boundary sample points are decoded and denormalized to obtain the boundary scene.

[0076] In one possible implementation, the boundary expansion scene acquisition module performs feature analysis on the boundary scene to obtain boundary rules. When expanding the boundary scene based on the numerical expansion of the boundary rules, the feature analysis adopts a combination of decision tree algorithm and rule extraction method. The obtained boundary scene is used as a sample with negative label, and the remaining scenes are used as samples with positive label and put into the decision tree model for binary classification training. The rule extraction method is used to obtain the judgment rule of the boundary scene and convert it into if-then rule. The judgment rule of the boundary scene directly gives the key feature combination that constitutes the boundary scene. In the boundary expansion scene generation stage, more extreme combinations are constructed to become extreme expansion scenes.

[0077] In one possible implementation, the hierarchical genealogy construction module, based on the initial clustering results, uses a hierarchical clustering algorithm to merge similar initial clusters to form a hierarchical tree diagram, revealing the correlation and potential evolutionary paths between different scenarios; In the tree diagram, the horizontal axis represents the initial classification scene category number, and the vertical axis represents the relative distance, intuitively expressing the distance when merging categories; the lower the vertical axis value, the earlier the operation mode is merged, and the higher the vertical axis value, the later the operation mode is merged. By combining the generated boundary scenarios, boundary extension scenarios, and typical operating modes, a test and verification scenario library is constructed to test and verify the control software under extreme conditions and boundary performance.

[0078] Another embodiment of the present invention also proposes an electronic device, including a processor and a memory, wherein the processor is used to execute a computer program stored in the memory to implement the boundary scenario generation method for experimental verification of power control systems.

[0079] Another embodiment of the present invention also proposes a computer-readable storage medium storing at least one instruction, which, when executed by a processor, implements the boundary scenario generation method for experimental verification of power control systems.

[0080] The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable storage medium can include any entity or device capable of carrying the computer program code, a medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory, a random access memory, an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals. For ease of explanation, the above content only shows the parts related to the embodiments of the present invention; for specific technical details not disclosed, please refer to the method section of the embodiments of the present invention. This computer-readable storage medium is non-transitory and can be stored in storage devices formed by various electronic devices, enabling the execution process described in the method of the embodiments of the present invention.

[0081] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0082] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0083] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0084] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for generating boundary scenarios for power regulation system test verification, characterized in that, The method comprises the following steps: obtaining power system operation mode samples, inputting a pre-constructed convolutional denoising self-encoder model with a fusion self-attention mechanism, and extracting deep features capable of reflecting the characteristics of the power system operation mode; performing typical operation mode clustering identification on the deep features capable of reflecting the characteristics of the power system operation mode; calculating the distance between the deep features of each power system operation mode sample and the deep features of the typical operation mode identified through clustering, selecting boundary sample points according to the distance, and decoding the boundary sample points to obtain boundary scenarios; performing feature analysis on the boundary scenarios to obtain boundary rules, and performing numerical expansion based on the boundary rules to obtain boundary expansion scenarios; using the boundary expansion scenarios obtained through numerical expansion, introducing a hierarchical clustering algorithm to construct a hierarchical pedigree of power grid operation scenarios.

2. The method of claim 1, wherein the boundary scenario is generated based on a power regulation system test. The convolutional denoising self-encoder model with a fusion self-attention mechanism is trained separately, and the target of the model training is to recover the original data from the input power system operation mode sample data with noise information.

3. The method of claim 2, wherein the boundary scenario is generated based on a power flow regulation system test verification. In the step of obtaining power system operation mode samples, inputting a pre-constructed convolutional denoising self-encoder model with a fusion self-attention mechanism, and extracting deep features capable of reflecting the characteristics of the power system operation mode, the power system operation mode samples are composed of the active power of the load, the active power output of new energy and traditional energy at the same time in different regions as characteristic variables. ​ The sample is divided according to regions, and any one running mode sample after preprocessing of the running mode sample set is P m , and the expression is as follows: In the formula, i represents the region number, w represents the operation mode combination corresponding to wind power; pv represents the operation mode combination corresponding to photovoltaic power; d represents the operation mode combination corresponding to load; h represents the operation mode combination corresponding to hydropower; th represents the operation mode combination corresponding to thermal power; The convolutional denoising autoencoder model fused with self-attention mechanism has a loss function L rec The mean square error (MSE) of the non-reverse normalized reconstruction output data and the normalized original input data is calculated as follows: 。 4. The method of claim 1, wherein the boundary scenario generation method for the power regulation system test verification is characterized by, In the step of performing typical operation mode clustering identification on the deep features capable of reflecting the characteristics of the power system operation mode, an evaluation index system combining technical and business indicators is used to select the optimal clustering number. On the technical level, the silhouette coefficient is selected as the evaluation index, which measures the rationality of sample attribution in the clustering result by combining the cohesion and separation degrees of the samples. On the business level, business indicators are defined according to the scenario generation requirements, including a clustering algorithm based on evaluation indicators defined by scenario generation requirements to determine the optimal cluster number for feature clustering.

5. The method of claim 4, wherein the boundary scenario is generated based on a power flow regulation system test verification. In the step of performing typical operation mode clustering identification on the deep features capable of reflecting the characteristics of the power system operation mode, three typical indicators are defined from the perspectives of new energy, load and power flow direction, and the intra-class new energy maximum active power variance, intra-class load maximum variance and intra-class section active power direction minimum consistency rate are used as clustering effect evaluation indicators, and the calculation expressions are as follows: The technical index clustering algorithm silhouette coefficient is calculated as follows: In the formula, a ( i ) indicates a sample i The average Euclidean distance to all other samples in the same cluster reflects the cluster density; the smaller the value, the more concentrated the points in the cluster. b ( i ) indicates a sample i The minimum average Euclidean distance to all samples in other clusters reflects the inter-cluster separation; the larger the value, the more distinct the separation from other clusters. The business index intra-class new energy maximum active power variance is calculated as follows: In the formula, S 2 n,max The variance of the maximum active power of new energy in all categories is represented, which reflects the maximum volatility of the active power of new energy. The smaller the value is, the smaller the in-class new energy active power fluctuation degree of the whole of all categories is. The business index intra-class load maximum variance is calculated as follows: In the formula, S 2 d,max represents the maximum variance of load in all categories, reflecting the maximum volatility of load, the smaller the value, the smaller the degree of load fluctuation within the category as a whole. The business index intra-class section active power direction minimum consistency rate is calculated as follows: In the formula, num z,all is the total number of samples of the zth operating mode cluster; num z,lcon is the number of samples of the zth operating mode cluster in which the operating mode cross-section active power direction is consistent with the cluster center. The min() function represents the minimum value in Z categories, reflecting the minimum value of the section active power direction consistency rate in all categories. The larger the value, the higher the intra-class section active power direction consistency rate of the overall category.

6. The method of claim 5, wherein the boundary scenario is generated based on a power flow regulation system test. In the step of identifying the typical operation mode cluster capable of reflecting the deep features of the power system operation mode characteristics, the maximum active power variance of the new energy within the class and the maximum load variance within the class are preprocessed by taking the reciprocal, then the comprehensive average index is obtained by normalizing and arithmetically averaging the four indexes of the cluster algorithm profile coefficient, the maximum active power variance of the new energy within the class, the maximum load variance within the class and the minimum consistency rate of the active power direction of the section within the class, and the K value corresponding to the optimal comprehensive average index is selected as the final cluster number; the following steps are adopted to determine the typical operation mode with physical significance: The cluster center of each cluster is calculated, then for each cluster, the distance of all sample points within the corresponding cluster to the cluster center of the cluster is calculated, and the nearest real historical sample point is selected as the deep feature of the representative operation mode of the corresponding cluster; The deep features are decoded to obtain the restored original data which is not subjected to the inverse normalization, and the representative operation mode sample is obtained through the inverse normalization processing, wherein the Euclidean distance measurement is adopted for distance measurement.

7. The method of claim 1, wherein the boundary scenario generation method for power regulation system test verification is characterized by, In the step of calculating the distance between the deep features of each power system operation mode sample and the deep features of the typical operation mode identified through clustering, selecting the boundary sample point according to the distance, and decoding the boundary sample point to obtain the boundary scenario, the Euclidean distances of all sample points to the cluster center of the corresponding category are calculated, and are sorted in descending order of distance, and a plurality of sample points are selected according to the set distance size judgment basis, and the selected sample points are initially defined as the boundary sample points under the corresponding typical operation mode; alternatively, the distance threshold is set according to the business context, and the boundary sample points are selected according to the set distance threshold; the selected boundary sample points are decoded and subjected to inverse normalization processing to obtain the boundary scenario.

8. The method of claim 1, wherein the boundary scenario generation method for power regulation system test verification is characterized by, In the step of performing feature analysis on the boundary scenario to obtain the boundary rule, and performing numerical expansion based on the boundary rule to obtain the boundary expansion scenario, the feature analysis adopts the combination of the decision tree algorithm and the rule extraction method, the obtained boundary scenario is taken as a sample with a negative label, the remaining scenarios are taken as samples with positive labels and are put into the decision tree model for binary classification training, the boundary scenario determination rule is obtained through the rule extraction method and is converted into an if-then rule, and the boundary scenario determination rule directly gives the key feature combination constituting the boundary scenario, in the boundary expansion scenario generation stage, a more extreme combination is constructed to become an extreme expansion scenario.

9. The method of claim 1, wherein the boundary scenario generation method for power regulation system test verification is characterized by, The step of constructing the hierarchical pedigree of the power grid operation scenario by using the boundary expansion scenario obtained through numerical expansion includes the following steps: on the basis of the initial clustering result, the hierarchical clustering algorithm is adopted to merge similar initial clusters to form a hierarchical tree diagram, and the correlation and potential evolution path between different scenarios are revealed; The horizontal coordinate in the tree diagram represents the initial classification scenario category number, and the vertical coordinate represents the relative distance, which directly expresses the distance when the categories are merged; the operation mode with a lower vertical coordinate value is merged earlier, and the operation mode with a higher vertical coordinate value is merged later; By combining the generated boundary scene, boundary expansion scene and typical operation mode, a test verification scene library is formed for testing and verifying the extreme working condition and boundary performance of the regulation software. 10.A boundary scenario generation system for power regulation system test verification, characterized in that, Comprise: The deep feature extraction module is used for acquiring power system operation mode samples, inputting a pre-constructed convolutional denoising autoencoder model with fusion self-attention mechanism, and extracting deep features capable of reflecting power system operation mode characteristics; The typical operation mode clustering identification module is used for performing typical operation mode clustering identification on the deep features capable of reflecting power system operation mode characteristics; The boundary scene acquisition module is used for calculating the distance between the deep features of each power system operation mode sample and the deep features of the clustered and identified typical operation mode, selecting a boundary sample point according to the distance, and decoding the boundary sample point to obtain a boundary scene; The boundary expansion scene acquisition module is used for performing feature analysis on the boundary scene to obtain boundary rules, and performing numerical expansion based on the boundary rules to obtain a boundary expansion scene; The hierarchical pedigree construction module is used for introducing a hierarchical clustering algorithm to construct a hierarchical pedigree of power grid operation scenes by using the boundary expansion scene obtained by numerical expansion.

11. The boundary scenario generation system for power regulation system test verification of claim 10, wherein, When the deep feature extraction module acquires power system operation mode samples, inputs a pre-constructed convolutional denoising autoencoder model with fusion self-attention mechanism, and extracts deep features capable of reflecting power system operation mode characteristics, the convolutional denoising autoencoder model with fusion self-attention mechanism is trained separately, and the target of model training is to restore the original data from the input power system operation mode sample data with noise information.

12. The boundary scenario generation system for power regulation system test verification of claim 11, wherein, When the deep feature extraction module acquires power system operation mode samples, inputs a pre-constructed convolutional denoising autoencoder model with fusion self-attention mechanism, and extracts deep features capable of reflecting power system operation mode characteristics, the power system operation mode samples are composed of active power of load, active power output of new energy and traditional energy at the same time in different regions as feature variables; The sample is divided according to regions, and any one running mode sample after preprocessing of the running mode sample set is P m , and the expression is as follows: In the formula, i represents the region number, w represents the operation mode combination corresponding to wind power; pv represents the operation mode combination corresponding to photovoltaic power; d represents the operation mode combination corresponding to load; h represents the operation mode combination corresponding to hydropower; th represents the operation mode combination corresponding to thermal power; The convolutional denoising autoencoder model fused with self-attention mechanism has a loss function L rec The mean square error (MSE) of the non-inverse normalized reconstruction output data and the normalized original input data is calculated as follows: 。 13. The boundary scenario generation system for power regulation system test verification of claim 10, wherein, The typical operation mode clustering identification module selects the optimal cluster number by using an evaluation index system combining technical and business indexes; On the technical level, the silhouette coefficient is selected as the evaluation index, which measures the rationality of sample attribution in the clustering result by combining the cohesion and separation degrees of the samples. On the business level, the business indexes are defined according to the scene generation requirements, including the clustering algorithm based on the evaluation index defined by the scene generation requirements to determine the optimal cluster number for feature clustering.

14. The boundary scenario generation system for power regulation system test verification of claim 10, wherein, The typical operation mode clustering identification module defines three typical indexes from the perspectives of new energy, load and power flow direction, and uses the maximum active power variance of new energy within the class, the maximum variance of load within the class and the minimum consistency rate of active power direction within the section as the clustering effect evaluation indexes, and the calculation expressions are as follows: The technical index clustering algorithm silhouette coefficient is calculated as follows: wherein, a ( i ) denotes the average Euclidean distance of the sample i to all other samples within the same cluster, reflecting the tightness of the cluster, the smaller the value, the more concentrated the points within the cluster; b ( i ) denotes the minimum value of the average Euclidean distance of the sample i to all samples in other clusters, reflecting the separation degree between clusters, the larger the value, the more obvious the distinction from other clusters; The business index maximum active power variance of new energy within the class is calculated as follows: In the formula, S 2 n,max The variance of the maximum active power of new energy in all categories is represented, which reflects the maximum volatility of the active power of new energy. The smaller the value is, the smaller the in-class new energy active power fluctuation degree of the whole of all categories is. The business index maximum variance of load within the class is calculated as follows: In the formula, S 2 d,max represents the maximum variance of load in all categories, reflecting the maximum volatility of load, the smaller the value, the smaller the degree of load fluctuation within the category as a whole. The business index minimum consistency rate of active power direction within the section is calculated as follows: In the formula, num z,all is the total number of samples of the zth operating mode cluster; num z,lcon is the number of samples of the zth operating mode cluster in which the operating mode cross-section active power direction is consistent with the cluster center. The min() function represents the minimum value in the Z categories, reflecting the minimum value of the cross-section active power direction consistency rate in all categories. The greater the value, the higher the overall intra-class cross-section active power direction consistency rate.

15. The boundary scenario generation system for power regulation system test verification of claim 14, wherein, The typical operation mode clustering identification module pre-processes the intra-class new energy maximum active power variance and the intra-class load maximum variance by taking the reciprocal, then normalizes and arithmetically averages the four indexes of the cluster algorithm profile coefficient, the intra-class new energy maximum active power variance, the intra-class load maximum variance, and the intra-class cross-section active power direction minimum consistency rate to obtain a comprehensive average index, and selects the K value corresponding to the optimal comprehensive average index as the final clustering number; the following steps are used to determine the typical operation mode with physical significance: The cluster center of each cluster is calculated, and then for each cluster, the distance of all sample points in the corresponding cluster to the cluster center of the cluster is calculated, and the nearest real historical sample point is selected as the deep feature of the corresponding cluster representative operation mode; The deep feature is decoded to obtain the restored original data which has not been subjected to inverse normalization, and the representative operation mode sample is obtained through inverse normalization processing, wherein the distance measurement adopts Euclidean distance measurement.

16. The boundary scenario generation system for power regulation system test verification of claim 10, wherein, The boundary scene acquisition module calculates the Euclidean distance of all sample points to the corresponding cluster center, and sorts them from large to small, and selects a number of sample points as the boundary sample points under the corresponding typical operation mode according to the set distance size determination basis; or, according to the business context, a distance threshold is set, and the boundary sample points are selected according to the set distance threshold; the selected boundary sample points are decoded and subjected to inverse normalization processing to obtain the boundary scene.

17. The system of claim 10, wherein the system is configured to generate boundary scenarios for testing and validating power regulation systems. The boundary expansion scene acquisition module analyzes the features of the boundary scene to obtain the boundary rule, and expands the boundary rule value to obtain the boundary expansion scene, wherein the feature analysis adopts a combination of the decision tree algorithm and the rule extraction method, the obtained boundary scene is taken as a sample with a negative label, the remaining scenes are taken as samples with positive labels and put into a decision tree model for binary classification training, the boundary scene determination rule is obtained through the rule extraction method and converted into an if-then rule, and the boundary scene determination rule directly gives the key feature combination constituting the boundary scene. In the boundary expansion scene generation stage, more extreme combinations are constructed to become extreme expansion scenes.

18. The system of claim 10, wherein the system is configured to generate boundary scenarios for testing and validating power regulation systems. The hierarchical pedigree construction module adopts a hierarchical clustering algorithm to merge similar initial clusters based on the initial clustering result, forms a hierarchical tree diagram, and reveals the correlation and potential evolution path between different scenes; The horizontal coordinate in the tree diagram represents the initial classification scene category number, and the vertical coordinate represents the relative distance, which intuitively expresses the distance when the categories are merged; The operation mode with a lower vertical coordinate value is merged first, and the operation mode with a higher vertical coordinate value is merged later; The generated boundary scene, boundary expansion scene, and typical operation mode are combined to form a test verification scene library, which is used for testing and verifying the extreme working conditions and boundary performance of the regulation software.

19. An electronic device, comprising: The computer readable storage medium stores at least one instruction, and the at least one instruction is executed by the processor to implement the boundary scenario generation method for the power regulation system test verification according to any one of claims 1 to 9.

20. A computer-readable storage medium, characterized in that, The computer readable storage medium stores at least one instruction, and the at least one instruction is executed by the processor to implement the boundary scenario generation method for the power regulation system test verification according to any one of claims 1 to 9.

Images