A method and system for generating power grid critical stability scenarios based on CGAN

Abstract

The present invention proposes a method and system for generating power grid critical stability scenarios based on CGAN, which belongs to the technical field of power grid critical stability scenario generation, including: generating a power grid operation benchmark scenario that takes into account the uncertainty on both the source and load sides; screening key power grid features based on the generated power grid operation benchmark scenario; analyzing the impact of key features on power grid safety margin based on a feedforward neural network, and calculating the sensitivity of the screened key power grid features; dynamically adjusting key parameters of the power grid based on the sensitivity, so that the power grid system gradually approaches the critical stability operation point, and then generating a critical stability operation scenario; constructing a critical stability scenario generation model based on a deep conditional generative adversarial network; designing a comprehensive optimization strategy based on the constructed critical stability scenario generation model; and using the deep conditional generative adversarial network technology to generate a required number of power grid critical stability scenarios based on the comprehensive optimization strategy.

Description

Technical Field

[0001] The present invention belongs to the technical field of power grid critical stability scenario generation, and in particular relates to a method and system for generating power grid critical stability scenarios based on CGAN. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] The power system has maintained safe and stable operation for a long time. During most periods, the system is within the normal operating range, and there are very few cases of safety and stability issues such as overload and instability. This has led to serious imbalances in historical operating scenario data. If historical operating scenario data is directly used as data samples for artificial intelligence model training, the sample library will be overloaded with similar samples and lack diversity, which will seriously restrict the effectiveness of pattern mining and control strategy formulation, and may even cause the model to misjudge critical states. Taking an extreme case as an example, when there are only 10 unstable samples in 1,000 training samples, the algorithm only needs to evaluate all samples as stable to achieve an evaluation accuracy of 99%. Such evaluation results obviously lack practical significance and application value.

[0004] Therefore, it is of great research significance to study high-quality generation methods for critical stability scenarios of power grids, solve the problem of imbalanced training samples, and provide methodological guidance and data support for key areas such as accurate characterization of power system stability boundaries, construction of security defense systems, and formulation of strategies to deal with extreme working conditions.

[0005] With the large-scale grid integration of renewable energy sources such as wind power and photovoltaics, and the widespread adoption of high-voltage direct current (HVDC) transmission technology, the proportion of traditional conventional generating units has dropped significantly, and the grid topology and dynamic operating characteristics have become more complex than ever before. Faced with the challenge of significantly increased randomness and volatility in the power system, power system scenario generation technology has emerged and rapidly developed. The core of this technology is to generate large-scale and representative power system operating scenarios through sampling methods, thereby accurately characterizing the random fluctuation characteristics and complex dynamic behavior of the power system under the background of high-proportion integration of renewable energy. This provides a solid theoretical foundation and technical support for key areas such as power system security assessment and source-grid-load coordination optimization.

[0006] Traditional power system scenario generation methods primarily parameterize the uncertainty of renewable energy sources based on probabilistic models such as Gaussian and Beta distributions. However, these methods struggle to describe the correlation of renewable energy output. When the application environment is highly complex, their parameter fitting errors are significant, reducing the accuracy and reliability of scenario generation.

[0007] With the development of artificial intelligence technology, generative learning methods, such as generative adversarial networks (GANs), have shown significant advantages in the field of power system scenario generation. GANs consist of two adversarial neural networks, a generator and a discriminator. Through iterative game training, they achieve precise mapping from random noise space to target data distribution. They can directly extract realistic data distribution characteristics from massive amounts of data without constructing a complex parameter hypothesis system. As an advanced variant of GANs, deep conditional generative adversarial networks (CGANs) achieve precise control of generated sample features by integrating conditional information during the generation process. CGANs perform well in power system scenario generation and can significantly improve the quality of operational scenario generation.

[0008] However, existing scenario generation methods based on generative adversarial networks still face two key challenges: first, they fail to capture the scene features under extreme conditions, making it difficult to accurately generate critical stable operation scenarios for power grids; second, the inherent correlation between extreme scenario features and the model network structure is not studied in depth, which restricts the generalization ability of the model. Summary of the Invention

[0009] To overcome the shortcomings of the above-mentioned existing technologies, the present invention provides a method for generating critical stability scenarios of power grids based on CGAN. This method can effectively construct diversified operating scenario data of the safety and stability edge on the basis of normal operating scenarios, so that the samples can cover various operating scenarios that may occur in the power grid to the greatest extent, thereby meeting the demand for data sample diversity for intelligent analysis and decision-making of the power grid, and providing a reliable data foundation for the accurate identification of the stability boundary of the power system and the formulation of security defense strategies.

[0010] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0011] In a first aspect, a method for generating power grid critical stability scenarios based on CGAN is disclosed, comprising:

[0012] Generate a grid operation benchmark scenario that takes into account uncertainties on both the source and load sides;

[0013] Screen key grid characteristics based on the generated grid operation benchmark scenario;

[0014] The impact of key features on grid security margin is analyzed based on feedforward neural network, and the sensitivity of the selected key features of the grid is calculated;

[0015] Dynamically adjust key grid parameters based on sensitivity to gradually approach the critical stable operation point, and then generate a critical stable operation scenario;

[0016] Construct a critically stable scene generation model based on deep conditional generative adversarial networks;

[0017] Design a comprehensive optimization strategy based on the constructed critical stability scene generation model, including: designing the loss function, self-attention mechanism, residual connection structure, and incremental feedback of generated samples;

[0018] Based on a comprehensive optimization strategy, a deep conditional generative adversarial network technique is used to generate the required number of grid critical stability scenarios.

[0019] As a further technical solution, the steps for generating a grid operation benchmark scenario considering the uncertainty on both the source and load sides are as follows:

[0020] Based on Dvine Copula, a dual-period model of the output of multiple renewable energy stations was established, and a set of renewable energy output scenarios was generated based on the established model.

[0021] Normal distribution is used to describe the uncertainty on the load side, with the benchmark load as the mean and the standard deviation set as a specific proportion of the benchmark load to generate a set of load scenarios;

[0022] Combining source and load scenarios based on the set of new energy output scenarios and load scenarios, and performing scenario reduction;

[0023] Combined with the generator dispatch plan, the generator power output is determined and the power flow is calculated to solve the voltage and power distribution of the power grid and generate a complete set of steady-state operation scenarios.

[0024] As a further technical solution, the steps of screening key features of the power grid are specifically as follows:

[0025] Collect node power and voltage characteristic data under different steady-state operation scenarios, calculate dynamic safety margin indicators through time-domain simulation, and construct an original data set corresponding to the characteristic data and safety margin indicators;

[0026] For each dynamic safety margin indicator, calculate the UIC value between it and each candidate feature;

[0027] Arrange all features related to each margin indicator in descending order based on UIC value to determine the degree of influence of the feature on the specific safety margin indicator;

[0028] The cross-validation technique is used to determine the optimal UIC threshold, and the key feature set of the power grid that has a significant impact on the safety margin index is screened based on this threshold.

[0029] As a further technical solution, the sensitivity of the selected key features of the power grid is calculated. The specific steps are as follows:

[0030] The screened key features of the power grid are input into a feedforward neural network, and the output of the feedforward neural network is a dynamic safety margin indicator;

[0031] Sensitivity is defined as the partial derivative of the safety margin indicator with respect to the input feature, and is used to measure the intensity and direction of the change in the dynamic safety margin after a small perturbation of the key feature under the current operating mode.

[0032] The sensitivities of all key features are organized into gradient vector forms.

[0033] As a further technical solution, the key parameter adjustment steps based on sensitivity are as follows:

[0034] Based on feedforward neural network, the automatic differentiation method is used to calculate the local sensitivity of key features to safety margin indicators;

[0035] According to the sensitivity information under the current operation mode, key parameters are adjusted along the negative gradient direction;

[0036] Input the updated key parameters into the feedforward neural network and output the updated safety margin value;

[0037] In order to determine whether the system has approached the critical stable state, the convergence of the safety margin index is set as the termination condition; when the system has reached the critical stable operating point, the iterative process is terminated;

[0038] The final set of key parameters is used as the input of the power flow calculation to obtain the complete power flow results of the system, and time domain simulation verification is performed to generate the critical stable operation scenario.

[0039] As a further technical solution, the steps for generating power grid critical stability scenarios based on deep conditional generative adversarial networks are as follows:

[0040] Extract the hidden features of the existing critical stability scene based on SDAE and construct the conditional input information;

[0041] The random noise vector and conditional input information are used as the input of the generator, and a deep network structure with a self-attention mechanism and residual connection is used to generate a critical stable scene feature set.

[0042] Based on the input condition information, the discriminator performs discriminative training on real samples and generated samples, guiding the generator to generate diverse and reasonable operating scenarios while approximating the real data distribution, thereby establishing a stable adversarial optimization mechanism.

[0043] For samples that are judged as true by the discriminator, the dynamic safety margin index is calculated using time domain simulation to verify whether the critical stability conditions are met;

[0044] The verified scenes are included in the training set as new samples for incremental training of the discriminator and SDAE, enhancing the model's ability to extract critical stable features.

[0045] In the second aspect, a CGAN-based power grid critical stability scenario generation system is disclosed, comprising:

[0046] The power grid operation benchmark scenario generation module is configured to: generate a power grid operation benchmark scenario that takes into account uncertainties on both the source and load sides;

[0047] The grid key feature screening module is configured to: screen the grid key features based on the generated grid operation benchmark scenario;

[0048] The sensitivity calculation module is configured to: analyze the impact of key features on the power grid safety margin based on a feedforward neural network, and calculate the sensitivity of the selected key features of the power grid;

[0049] A grid critical stability scenario generation module is configured to: dynamically adjust key grid parameters based on sensitivity so that the grid system gradually approaches a critical stability operation point, and then generate a critical stability operation scenario;

[0050] Construct a critically stable scene generation model based on deep conditional generative adversarial networks;

[0051] Design a comprehensive optimization strategy based on the constructed critical stability scene generation model, including: designing the loss function, self-attention mechanism, residual connection structure, and incremental feedback of generated samples;

[0052] Based on a comprehensive optimization strategy, a deep conditional generative adversarial network technique is used to generate the required number of grid critical stability scenarios.

[0053] One or more of the above technical solutions have the following beneficial effects:

[0054] The technical solution of this invention considers both source and load uncertainties and utilizes scenario generation and reduction methods to generate a set of representative operating scenarios. For a specific set of anticipated faults, the uniform information coefficient (UIC) feature filtering method is used to screen key features that influence the grid's dynamic safety margin indicators, taking into account power angle stability and transient voltage safety margin indicators. A sensitivity calculation method based on a feedforward neural network and automatic differentiation (AD) technology rapidly calculates the sensitivity of the grid's key steady-state operating characteristics to the safety margin indicators. Based on this sensitivity information, key grid parameters are iteratively adjusted to gradually approach the critical stability operating point. A grid critical stability scenario generation model is constructed based on a deep conditional generative adversarial network. A stacked denoising autoencoder (SDAE) is used to extract hidden features from existing critical stability scenarios as conditional input information, guiding the generator to generate a large number of critical stability scenarios. Based on the generative adversarial framework, a sample incremental feedback mechanism is designed to implement the iterative optimization mode of "generation → verification → enhancement → regeneration". This enables the generator to continuously learn the characteristic laws of the critical stability state during continuous training, improves the quality and reliability of critical stability scenario generation, and solves the problem of scarcity of critical stability operation scenarios in the power grid.

[0055] The extreme operating conditions of the power system correspond to the critical stability operating points of the power grid. Insufficient capture of critical stability scene features:

[0056] The sub-technical solution of this embodiment adopts a stacked denoising autoencoder (SDAE) to extract hidden layer features as the conditional input of a deep conditional generative adversarial network (CGAN).

[0057] The technical solution of the present invention incorporates the verified scenes as new samples into the training set for incremental training of the discriminator and SDAE, enhances the model's ability to extract critical stable features, and realizes the iterative optimization mode of "generation → verification → enhancement → regeneration".

[0058] To address the lack of depth in the intrinsic correlation between critical stable scene features and the model network structure, the present invention introduces a self-attention mechanism and a residual connection structure into the generator network structure. The self-attention mechanism calculates the correlation between each node feature; the residual connection structure prevents the attenuation of key features in deep networks.

[0059] The technical solution of the present invention is based on the feature filtering method of unified information coefficient, which can evaluate the influence of operating characteristics on the system dynamic safety margin index under specific anticipated faults, effectively identify the key features that significantly affect the stability of the system, and reduce the computational complexity of generating critical stability scenarios.

[0060] The present invention combines feedforward neural networks with automatic differentiation technology to achieve rapid calculation of the sensitivity of key steady-state operating characteristics of the power grid to dynamic safety margin indicators. Through exponential decay step size control and upper and lower boundary constraint mechanisms, key parameters are adjusted along the negative gradient direction, so that the system gradually reaches the critical stable operating point, improving the efficiency and accuracy of critical stability scenario generation.

[0061] The present invention combines a stacked denoising autoencoder with a deep conditional generative adversarial network, and uses the stacked denoising autoencoder to extract the hidden layer features of existing critical stability scenes as conditional input information, thereby enhancing the directional learning ability of the generative adversarial network and enabling it to accurately capture the characteristic distribution laws of critical stability scenes.

[0062] The present invention improves and innovates the generator architecture of deep conditional generative adversarial networks, introduces a self-attention mechanism and a residual connection structure, enables the model to adaptively learn the correlation strength between the operating characteristics of each node, and improves the accuracy of the generated scenarios in reflecting the overall operating characteristics of the system.

[0063] Advantages of additional aspects of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0065] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0066] Figure 2 This is a schematic diagram of the structure of D vine Copula according to embodiment of the present invention;

[0067] Figure 3 It is a schematic diagram of the dual-period spatial correlation model;

[0068] Figure 4 Flowchart for tuning strategy for key parameters based on sensitivity;

[0069] Figure 5 This is the SDAE structure diagram;

[0070] Figure 6 Generate a model diagram for the CGAN critical stability scenario;

[0071] Figure 7 Schematic diagram of the generator network structure;

[0072] Figure 8 Flowchart of the method for generating power grid critical stability scenarios based on deep conditional generative adversarial networks. DETAILED DESCRIPTION

[0073] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0074] It should be noted that the terms used herein are for describing particular embodiments only and are not intended to be limiting of exemplary embodiments according to the present invention.

[0075] In the absence of conflict, the embodiments of the present invention and the features thereof may be combined with each other.

[0076] CGAN: The abbreviation of conditional generative adversarial networks in English, Chinese: Deep Conditional Generative Adversarial Network.

[0077] Example 1

[0078] In view of the fact that there are few critical stable operation scenarios in the power grid, which leads to serious imbalance in historical operation scenario data and insufficient diversity in the sample library, and then causes the model to inaccurately identify the stability boundary, poor generalization ability, and large errors in the safety and stability assessment results, which cannot meet the reliable operation requirements of the new power system, a method and system for generating critical stability scenarios of the power grid based on deep conditional generative adversarial networks are proposed.

[0079] See also Figure 1 As shown, this embodiment discloses a method for generating a power grid critical stability scenario based on CGAN, including:

[0080] Step 1: Considering the uncertainty on both the source and load sides, multiple grid operation benchmark scenarios are generated using scenario generation and reduction methods.

[0081] Step 2: Based on the different dynamic safety margin indicators under different types of anticipated fault scenarios, the key features that affect the dynamic safety margin indicators of the power grid are screened based on the unified information coefficient feature filtering method. In this step, key features such as the voltage and power of a key node are referred to as key features.

[0082] Step 3: Based on the sensitivity calculation method combining feedforward neural network and automatic differentiation technology, the sensitivity of key characteristics of the power grid to the dynamic safety margin index is quickly calculated.

[0083] Step 4: Based on the sensitivity information, gradually and iteratively adjust the key characteristics of the power grid so that the system gradually approaches the critical stable operating point.

[0084] Step 5: Build a grid critical stability operation scenario generation model based on a deep conditional generative adversarial network. Extract the hidden features of existing critical stability scenarios as conditional input information to guide the generator to generate a large number of critical stability scenarios.

[0085] In step 5 above, hidden features of critically stable scenes are extracted using a stacked denoising autoencoder. First, noise is added to the input data. Then, multiple autoencoders are trained layer by layer and stacked to form a deep network. Finally, the output of the intermediate hidden layer is extracted as the feature representation of the data.

[0086] In step 1, in order to generate a grid operation benchmark scenario, a dual-period model of multiple renewable energy station outputs based on D-vine Copula is first constructed.

[0087] The Copula theory in D-vine Copula was proposed by Sklar, whose core Sklar theorem reveals the essential structure of multivariate probability distribution. This theorem proves that: n The joint distribution of random variables can be decomposed into these n Each variable has its own marginal distribution function and a copula function. The marginal distribution function describes the individual probability behavior of each variable, while the copula function characterizes the nonlinear dependency structure between variables. Traditional multivariate copulas struggle to accurately capture high-dimensional dependency structures, and therefore have limitations in modeling the joint probability distribution of high-dimensional random variables. Vine Copula addresses these limitations by combining multiple binary copulas, providing a more flexible framework for modeling the correlation between multivariate random variables.

[0088] Commonly used vine copula structures include C-vine and D-vine structures. The C-vine structure is suitable for scenarios with a dominant variable that is highly correlated with other variables, while the D-vine structure is suitable for scenarios where variables generally show moderate or weak correlation and there is no obvious dominant variable. Since the power output of multiple new energy stations usually does not have an obvious dominant relationship, the D-vine copula structure is more suitable for building a new energy station power model. Figure 2 As shown in Figure 3, the dual-period output model of multiple renewable energy stations based on D-vine Copula can describe the spatial correlation of the output of multiple renewable energy stations.

[0089] The D-vine Copula model is used to model the correlation between the outputs of multiple renewable energy stations. The steps to generate renewable energy output scenarios are as follows:

[0090] 1-1) Based on the historical output data of multiple renewable energy stations, the kernel density estimation method is used to fit k The marginal probability distribution of the output of each new energy station.

[0091] The output data of renewable energy stations generally does not conform to the normal distribution or other classical distribution models, but instead exhibits complex nonlinear characteristics and uncertainty. Therefore, using empirical distribution functions to construct the marginal distribution of renewable energy station output can more realistically characterize its data characteristics.

[0092] X i For the i The output of new energy stations is a random variable, x i is the observed value of the random variable, N his is the number of historical samples, x his i , j For the i New energy stations j historical data observations, then i The empirical distribution function expression of the output of each new energy station is as follows:

[0093] (1)

[0094] Where, I [.] It is an indicator function. If the condition in the brackets is met, the function takes the value of 1, otherwise it is 0.

[0095] Based on the empirical distribution function, the cumulative probability distribution function value corresponding to the historical output data of each new energy station is calculated, that is, u i,j = F i ( x i,j ), and then estimate the parameters of the Copula function.

[0096] 1-2) Fit the parameters of each commonly used copula function based on the maximum likelihood estimation method, and calculate the Euclidean distance between the copula function and the empirical copula function.

[0097] Specifically, the Copula function parameter estimation based on the maximum likelihood method is:

[0098] Powered by two new energy stations X 1. X 2 as an example, the joint likelihood function of the Copula function can be expressed as follows:

[0099] (2)

[0100] Where, c θ is the density function of Copula, θis the parameter to be estimated, u his i , j = F i ( x his i , j ).

[0101] To simplify the calculation, the logarithm is usually taken to obtain the log-likelihood function:

[0102] (3)

[0103] 1-3) Using the minimum Euclidean distance as the criterion for selecting the optimal Copula model, the optimal Copula function between each pair of variables is selected according to the D-vine Copula structure to construct the D-vine Copula model.

[0104] 1-4) Based on the Monte Carlo sampling method, the D-vine Copula model is sampled and the obtained n Group data ( u 1i , u 2i ,…, u ki ), i = 1,2,…, n . u represents a random variable X The cumulative probability distribution function value of the observation, such as u 1= F 1( x 1), F 1 represents a random variable x The cumulative distribution function of 1.

[0105] 1-5)Yes( u 1i , u 2i ,…, u ki ) to perform inverse transformation F k -1 ( u k ), and get the corresponding new energy output data ( P 1i , P 2i ,…, P ki ), thus forming n The new energy output scenario that complies with the D-vine Copula distribution.

[0106] Assume that there are a total ofk New energy stations, generating n A scene. P mi ( m =1,2,…, k ; i =1,2,…, n ) indicates the i In the scene m A renewable energy output scenario can be understood as a set of output data for all renewable energy stations in a power grid calculation example.

[0107] In the above steps, the D-vine Copula model can accurately capture the complex nonlinear correlation structure between multiple renewable energy sites. It is particularly suitable for processing the output relationship between renewable energy sites such as wind power and photovoltaic power plants without obvious dominant variables, thereby generating scenario data that is more in line with the actual output characteristics of multiple renewable energy sources.

[0108] In the process of modeling the output correlation of multiple renewable energy sites, a dual-period spatial correlation model for daytime and nighttime was established to address the differences in the operating characteristics of wind power and photovoltaic power. This dual-period spatial correlation model captures the different correlation characteristics of photovoltaic-dominated and wind power-dominated periods, respectively, thereby generating richer and more diverse scenario data that aligns with the output characteristics of renewable energy.

[0109] During the daytime, wind power and photovoltaic power outputs both show significant spatial autocorrelation characteristics. At the same time, due to the coupling of atmospheric turbulence and cloud shielding effects, there are cross-site cross-correlation characteristics between wind and solar heterogeneous energy sources. Therefore, a daytime spatial correlation model is established to comprehensively consider the spatial correlation of wind-wind, solar-solar and wind-solar outputs. During the nighttime, the photovoltaic system shows zero output characteristics. At this time, only the output correlation between multiple wind farms needs to be considered. The daytime and nighttime correlation models are as follows: Figure 3 shown.

[0110] Generating grid benchmark scenarios that account for uncertainty on both the source and load sides is an important step in generating critical stability scenarios. High-quality benchmark scenarios can truly reflect the diverse operating states of the system, providing a rich foundation for critical stability scenario generation and improving their representativeness and coverage.

[0111] The steps for generating the grid benchmark scenario considering the uncertainty on both the source and load sides are as follows:

[0112] (1) Using D-vine Copula to describe the uncertainty of the source side, a dual-period model of multiple renewable energy station output is established to generate a set of renewable energy output scenarios, specifically: n The new energy output scenario that complies with the D-vine Copula distribution.

[0113] Based on the D-vine Copula, daytime and nighttime models were established. The daytime model considers all wind and solar power stations. Since the photovoltaic system has zero output at night, the nighttime model only considers the output correlation between multiple wind farms, and photovoltaic stations are not involved in the modeling.

[0114] (2) The normal distribution is used to describe the uncertainty on the load side, with the reference load as the mean and the standard deviation set as a specific proportion of the reference load to generate a set of load scenarios.

[0115] (3) A quantile-based statistical characteristic matching method is used to combine source-load scenarios, that is, the combination of the new energy output scenario set and the load scenario set. The K-medoids clustering algorithm is used to reduce the scenarios. The input is a large number of source-load scenario combinations generated previously, and the output is a moderate number of representative source-load scenario combinations.

[0116] (4) Determine the generator power output in conjunction with the generator dispatch plan, use the Newton method or PQ decomposition method to perform power flow calculations, solve the voltage and power distribution of the power grid, and generate a complete set of steady-state operation scenarios. A steady-state operation scenario can be represented as a combination of power and voltage data of all nodes in the power grid calculation example; it can also be more simply represented as a combination of active power and reactive power of all PQ nodes, a combination of active power and voltage amplitude of all PV nodes, and the voltage amplitude and voltage phase angle of the balancing node. A steady-state operation scenario set refers to multiple steady-state operation scenarios that can reflect different load levels and different power generation outputs of the power grid.

[0117] In this implementation example, the generation of grid benchmark scenarios that take into account the uncertainties on both the source and load sides can reflect the diverse operating states of the system, provide rich basic samples for the generation of critical stability scenarios, improve the representativeness and coverage of critical stability scenarios, and enable the samples to cover various operating scenarios that may occur in the power grid to the greatest extent possible.

[0118] In step 2, the key features of the power grid are screened based on the UIC feature filtering method.

[0119] For a variety of anticipated fault scenarios, time-domain simulation is used to calculate dynamic safety margin indicators for the power system. For example, the critical clearing time (CCT) is used as the power angle stability margin indicator, and the transient voltage offset is used as the voltage safety margin indicator. The UIC feature filtering method is used to identify and filter key characteristic variables that have a significant impact on the power grid safety margin from the high-dimensional feature space. The UIC calculation method is as follows:

[0120] For two sets of eigenvectors X = [ x 1,…, x n ]and Y = [ y1,…, y n ], n is the sequence length, and the mutual information coefficient calculation model can be written as

[0121] (4)

[0122] Where, I MI ( X ; Y )for X and Y Mutual information coefficient between ; P ( x , y )for x and y The joint probability density function of p ( x )and p ( y ) are respectively x and y The marginal probability density function of .

[0123] According to the unified classification method, X and Y Evenly divide into several segments, which can be written as

[0124] (5)

[0125] Where, l x and l y They are X and Y The partition unit length; x max and x min The eigenvectors are X The maximum and minimum values ​​of ; y max and y min The eigenvectors are X The maximum and minimum values ​​of ; a and b They are X and Y The number of segments; n 0.6 Represents the partition grid size, usually taken as 0.6 power of the data volume.

[0126] According to the unified partitioning method, the calculation model of UIC can be written as

[0127] (6)

[0128] Where, I UIC ( X ; Y )for X and Y The unified information coefficient between a , b )for a and b The minimum value in .

[0129] The steps for screening key grid features based on the UIC feature filtering method are as follows:

[0130] 2-1) Collect and generate characteristic data such as node power and voltage under different steady-state operation scenarios, calculate dynamic safety margin indicators through time-domain simulation, and construct an original data set corresponding to the characteristic data and safety margin indicators.

[0131] The specific calculation of dynamic safety margin indicators can be based on power system time domain simulation toolkits, such as STEPS (developed by Shandong University), PSS / E, PSASP, etc.

[0132] 2-2) For each dynamic safety margin indicator, the UIC value between it and each candidate feature is calculated according to the method proposed in this invention, specifically as formula (6).

[0133] 2-3) Arrange all features related to each margin indicator in descending order based on the UIC value to determine the degree of influence of the feature on the specific safety margin indicator.

[0134] 2-4) Use cross-validation technology to determine the optimal UIC threshold. Based on this threshold, select the key grid feature set that has a significant impact on the safety margin index, specifically the voltage and power of certain key nodes.

[0135] 2-5) Use the selected key features to re-establish the power grid security margin assessment model and evaluate the effectiveness and reliability of the feature screening results.

[0136] The process of establishing the power grid safety margin assessment model described above can be based on machine learning methods such as random forests or artificial neural networks (ANNs). The combination of existing features and safety margin values ​​is used as a sample set, which is then divided into training and test sets. First, a prediction model is established using all features as input and safety margin values ​​as output. Metrics such as mean absolute error (MAE) and root mean square error (RMSE) are used to evaluate model accuracy. Model accuracy is further evaluated using the selected key features as input and safety margin values ​​as output. If the accuracy of the latter model is similar to that of the former, the feature selection results are validated.

[0137] In the above steps, the screening of key grid features based on the UIC feature filtering method can accurately measure the impact of each feature on the safety margin index and quickly identify the most representative key features from high-dimensional features, thereby reducing model complexity and improving the accuracy of scenario generation.

[0138] In the specific step three, the key feature sensitivity is quickly calculated based on the feedforward neural network and automatic differentiation technology.

[0139] In power system dynamic security assessments, sensitivity typically refers to the responsiveness of a specific system dynamic safety margin indicator to input characteristic parameters (such as node voltage and power). During the generation of critical stability scenarios, by calculating the local sensitivity of key characteristics to safety margin indicators, we can determine the key characteristic parameters that most significantly impact system safety under the current operating mode and the adjustment direction for each parameter. Iterative parameter adjustment based on local sensitivity can gradually bring the system closer to the critical stability operating point, enabling targeted generation of critical stability scenarios.

[0140] In order to realize the sensitivity calculation of key steady-state operation characteristics of the power grid to dynamic safety margin indicators, the present invention proposes a sensitivity calculation method based on feedforward neural network and automatic differentiation technology to achieve accurate calculation of local sensitivity.

[0141] The present invention uses a feedforward neural network as a modeling tool, with the input being the selected key characteristic variables and the output being the dynamic safety margin index. The structure of the feedforward neural network is set as follows:

[0142] Input layer dimension: equal to the number of key features n ;

[0143] Hidden layer: adopts a fully connected structure and the activation function is ReLU;

[0144] The number of neurons in each layer is determined by grid search or layer-by-layer optimization;

[0145] Output layer: single node output, used for regression prediction of dynamic safety margin indicators;

[0146] Loss function: mean-square error (MSE) between the model's predicted value and the true value;

[0147] Optimizer: Adam, with adaptive learning rate decay.

[0148] After training is completed, the network model can be expressed as

[0149] (7)

[0150] Where, X Represents key feature input; yIndicates the safety margin indicator output; θ represents the set of neural network parameters.

[0151] In practical applications, sensitivity is defined as the partial derivative of the safety margin indicator with respect to the input feature, which is used to measure the intensity and direction of the change in the dynamic safety margin after the key feature is subjected to a small disturbance under the current operating mode.

[0152] For the i Key features x i , whose sensitivity is defined as

[0153] (8)

[0154] Where, X Represents the input vector composed of all key features; f ( X ) indicates that the neural network is inputting the current feature X The predicted output of is the dynamic safety margin index of the system; S i Indicates the i Key features x i Sensitivity to safety margin indicators.

[0155] Further organize the sensitivity of all key features into gradient vector form

[0156] (9)

[0157] This gradient can be calculated using automatic differentiation. The core idea of ​​automatic differentiation is to use the chain rule to decompose complex functions into a sequence of elementary operations and efficiently compute derivatives by tracing the connections between these operations. In feedforward neural networks, automatic differentiation computes the model output through forward propagation and solves the gradient through backpropagation combined with the chain rule. The computational complexity of backpropagation is similar to that of forward propagation, allowing it to maintain computational efficiency even in high-dimensional input spaces.

[0158] In step 4, the strategy is adjusted based on the key parameters of sensitivity.

[0159] Based on the rapid calculation of sensitivity, this paper proposes a key parameter adjustment strategy based on sensitivity. The basic process is as follows: Figure 4 This method dynamically adjusts the key parameters of the power grid in an iterative manner, so that the system gradually approaches the critical stable operating point.

[0160] The key parameter adjustment steps based on sensitivity are:

[0161] 4-1) Based on the trained feedforward neural network, the automatic differentiation method is used to calculate the local sensitivity of the key features to the safety margin indicator, see formula (9).

[0162] 4-2) To reduce the safety margin index value, adjust the key parameters along the negative gradient direction according to the sensitivity information under the current operating mode.

[0163] During adjustment: After using the automatic differentiation method to calculate the sensitivity of each key feature to the system safety margin, adjust the parameters in the direction of reducing the safety margin, that is, reduce the parameters with positive sensitivity and increase the parameters with negative sensitivity. After the adjustment is completed, input the updated key parameters into the feedforward neural network and output the updated safety margin value. Then repeat step 4-1 to recalculate the sensitivity and adjust the key parameters again according to the new sensitivity. Repeat this cycle until the system gradually reaches the critical stable operating point. The flow chart is as follows Figure 4 shown.

[0164] 4-3) To determine whether the system has reached a critical stability state, the convergence of the safety margin indicator is set as the termination condition. When the prediction results show that the system's current safety margin indicator is sufficiently close to 0, the operating mode is considered to have reached the critical stability boundary, and the iteration process terminates.

[0165] 4-4) Input the final set of key parameters into the power flow calculation program to obtain the complete power flow results of the system, and perform time domain simulation verification to generate the critical stable operation scenario.

[0166] This sensitivity-based key parameter adjustment step repeatedly calculates the local sensitivity of key features to the safety margin indicator, identifying the degree and direction of each parameter's impact on system stability under the current operating state. Based on this, the parameter values ​​are gradually adjusted along a negative gradient, ensuring that the system's safety margin score gradually decreases to zero, even as the system gradually reaches the critical stability operating point. This method effectively controls the directionality and convergence of the adjustment process, efficiently generating physically feasible grid operation scenarios that are close to the critical stability operating point.

[0167] In the process of adjusting key parameters, in order to avoid system divergence or oscillation caused by excessive parameter adjustment, the present invention introduces a step size factor to control the adjustment amplitude of each round, thereby improving the stability and controllability of the iterative process. The parameter update expression is:

[0168] (10)

[0169] Where, X k Indicates the key parameter set of the current iteration round; S k represents the corresponding sensitivity vector; α kIndicates the step length factor. The present invention uses exponential decay to adjust the step length.

[0170] (11)

[0171] Where, α k Indicates the k The step size factor of the iteration is used to control the adjustment range of the key parameters of the current round; α 0 indicates the initial step size factor; Represents the exponential decay coefficient, which is used to control the decay rate of the iterative step size.

[0172] To ensure that the updated parameter combination is within the physically feasible range, the intermediate results need to be X * k +1 adds operation constraints, mainly including upper and lower limits of node voltage and source-load power. The final adjustment value is determined by the limit method, and the expression is

[0173] (12)

[0174] Where, X min and X max Respectively represent the minimum and maximum values ​​allowed for each key parameter, which are determined by the equipment limit or scheduling rules. This mechanism can effectively avoid parameter over-limit and ensure that the operating mode obtained after each round of adjustment is physically feasible and engineering practical. At the same time, in order to ensure that the system meets the power balance between power generation and load, the present invention adds a power balance correction link. The specific approach is: after each key parameter adjustment, calculate the difference between the total power generation power and the load power, and select generator nodes with low sensitivity and regulation capabilities, adjust the power output according to their capacity ratio to compensate for the power difference, and ensure that the updated parameters meet the source-load power balance constraint.

[0175] In step five, a critically stable scene generation model is obtained based on a deep conditional generative adversarial network.

[0176] In order to achieve efficient generation of critical stable operation scenarios, this paper introduces CGAN as an extended model. The basic idea of ​​CGAN is to generate a critical stable operation scenario under given external condition information. u SDAE has a multi-layer stacking structure and noise suppression mechanism, which can extract hierarchical feature representation of data, so that CGAN can obtain conditional information with more discriminative and anti-interference capabilities. This paper uses SDAE to extract the hidden layer features of the existing critical stable operation scene as the conditional information of CGAN. u , its model structure is as follows Figure 5 shown.

[0177] in l Indicates the number of layers of SDAE; h i Indicates the i Hidden layer features extracted by the layer encoding function; z i Represents the reconstruction output of the autoencoder, which is used to calculate the reconstruction error and guide the optimization of network parameters, so that the encoder can learn high-quality feature representation.

[0178] At the model input, all variables must be normalized before being sent to the network to adapt to the numerical range and activation function characteristics of the neural network. z and condition information u is the input and output scene set X̃ = G ( z , u ); The discriminator receives the pre-built critical stable operation scene set or the scene set generated by the generator and combines it with the conditional information u Combined, the true and false probabilities are output to guide the iterative optimization of the generator.

[0179] SDAE is used to extract latent features from existing critically stable operating scenarios. The specific process is as follows: SDAE extracts latent features from the data through a multi-layer autoencoder structure. Its core mechanism is to add noise to the input data during training, then force the network to reconstruct the original noise-free data. This "denoising" process enables the model to learn more robust feature representations that not only capture the primary structure of the data but also resist interference from input variations. Each layer in SDAE is pre-trained as a separate denoising autoencoder. These layers are then stacked to form a deep network. Finally, the entire network is fine-tuned to optimize feature extraction, making the extracted latent features more robust and generalizable.

[0180] CGAN critical stable scene generation model such as Figure 6 shown.

[0181] Then we design a comprehensive optimization strategy for deep conditional generative adversarial networks.

[0182] To achieve targeted generation of critical stable operation scenarios for power grids, this paper, based on the completed CGAN structure, designs a comprehensive optimization strategy based on the data characteristics of power grid operation scenarios. The strategy specifically includes: loss function design, self-attention mechanism, residual connection structure, and incremental feedback of generated samples.

[0183] About loss function design:

[0184] In order to guide the generator to accurately capture the distribution of key features and approach the critical stability boundary of power grid operation, this example adopts the objective function of the standard conditional generative adversarial network as the optimization basis.

[0185] (13)

[0186] Where, G For the generator; D is the discriminator; z represents a random noise vector; u It represents the condition information obtained after the critical stable operation scenario set is encoded by SDAE.

[0187] Taking into account the actual operation requirements of the power system, the present invention introduces boundary constraint loss and source-load power balance constraint loss based on adversarial learning. Boundary constraints are used to limit the generated node voltage, source-load power and other parameters to be within the physically feasible range, and source-load power balance constraints are used to maintain the power balance between power generation and load in the generation scenario. The loss function of the generator is

[0188] (14)

[0189] Where, L adv To counter the loss, we push the generated samples to deceive the discriminator; L phys It is the boundary constraint loss, which is used to penalize node parameters that exceed the actual operating range; L balance The source-load power balance constraint loss is used to limit the deviation of the source-load power in the generated samples; λ is the weighting coefficient, which is used to adjust the trade-off between the three. L adv Using the cross entropy loss form:

[0190] (15)

[0191] (16)

[0192] (17)

[0193] Where, x i max and x i min Respectively i The maximum and minimum values ​​allowed for each operating parameter. P̂ i The generator generates the i Active power of each power generation node; P̂ j The generator generates the j The active power of each load node.

[0194] The goal of the discriminator is to accurately distinguish between real samples and generated samples, and its loss function is

[0195] (18)

[0196] Self-Attention Mechanism:

[0197] Because the physical devices within the power system are connected by wires to form a network topology, they must strictly adhere to power flow constraint equations. This leads to significant and complex correlations between system operating characteristics, such as the coupling relationship between node voltage and adjacent branch power flow, generator output and load power, and other parameters. Traditional fully connected neural network structures are often unable to fully capture such complex correlations. To enhance the ability to perceive the correlations between nodes, this paper introduces a self-attention mechanism into the generator architecture, enabling the model to adaptively learn the strength of the correlations between nodes and achieve collaborative modeling of global features. This mechanism dynamically calculates the degree of correlation between features and assigns differentiated weights to different features, ensuring that the model comprehensively considers the influence of other related parameters when generating each parameter, thereby more accurately reflecting the overall operating characteristics and physical constraints of the power system.

[0198] This paper introduces a local incremental update strategy in the later stages of model training to fine-tune and optimize the self-attention module. After the generator converges, its backbone network parameters are frozen, retaining only the self-attention module for iterative updates. This further enhances the modeling capability of node-to-node correlation while maintaining global feature stability.

[0199] Residual Connection Structure: In power system operation scenario modeling, some steady-state features may have strong global dominance. For example, the voltages of certain key nodes and the output of large generators may have a global impact on system characteristics. This information may gradually decay during propagation through deep neural networks due to gradient dissipation or activation function saturation, causing the generator to ignore the core supporting structure of the system. To address this, a residual connection structure is introduced in the generator to directly transmit the initial low-level features to the deep network through cross-layer connections, alleviating the problem of feature decay.

[0200] Incremental feedback of generated samples: In order to further utilize the high-quality scene information generated by the generator, the present invention designs a generation sample incremental feedback mechanism based on simulation verification. When some of the generated samples are verified to be true by the discriminator, that is, D ( x ∣ u)→1, and after time domain simulation verification, it is determined that its dynamic safety margin index approaches 0, which can be considered to have true critical stability characteristics.

[0201] For the discriminator, such samples are appended to the training set as "new true samples" to update the discriminant's discriminative ability; for SDAE, the features of such samples are introduced as incremental data, and SDAE is incrementally trained to optimize its hidden layer feature expression ability, thereby improving the encoding quality of critical stable scene features.

[0202] This mechanism forms an iterative pattern of "generation → verification → enhancement → regeneration" during the training process, which enables the generator to continuously learn the characteristic laws of critical stable states during continuous training, and ultimately achieve high-precision critical stable scene generation. The network structure diagram of the generator is shown in the figure below. Figure 7 shown.

[0203] Power grid critical stability scenario generation process: The flow chart of the power grid critical stability scenario generation method based on deep conditional generative adversarial network is as follows: Figure 8 As shown in Figure 3, the process integrates the incremental update of the model to form an iteration of "generate → verify → enhance → regenerate", achieving high-quality expansion of critical stability scenarios.

[0204] The steps for generating power grid critical stability scenarios based on deep conditional generative adversarial networks are as follows:

[0205] 5-1) Based on SDAE, the hidden features of the existing critical stable operation scenario set are extracted to construct conditional input information.

[0206] 5-2) Using random noise vectors and conditional input information as input to the generator, a deep network structure with self-attention mechanism and residual connection is used to generate a critical stable scene feature set.

[0207] In this step, the input data first includes hidden layer feature information extracted from an existing critically stable scene using SDAE, as well as a random noise vector. Before entering the generator, this data needs to be normalized to ensure that the numerical ranges of data in different dimensions are consistent and meet the numerical stability requirements of deep network training. After completion of this processing, the normalized noise vector is concatenated with the encoded conditional information and serves as the final input to the generator, driving the generation of critically stable scenes.

[0208] The self-attention mechanism and residual connection structure mentioned in this paper are not preprocessing operations on the input data, but rather network structure designs used to enhance the internal modeling capabilities of the generator. The self-attention mechanism is used to learn the global dependencies between various operating characteristics of the power grid, while the residual connection structure is used to retain key information in the input features and prevent information decay in deep networks. These mechanisms are designed to improve the rationality and accuracy of the generated results and are independent of the initial processing of the input data.

[0209] 5-3) Based on the input condition information, the discriminator performs discrimination training on real samples and generated samples, guiding the generator to generate diverse and reasonable operating scenarios while approximating the real data distribution, thereby establishing a stable adversarial optimization mechanism.

[0210] 5-3) For samples that are judged as true by the discriminator, time domain simulation is used to calculate their dynamic safety margin indicators and simulate and verify whether they meet the critical stability conditions.

[0211] 5-4) The verified scenes are included as new samples in the training set for incremental training of the discriminator and SDAE, strengthening the model's ability to extract critical stable features and realizing the iterative optimization model of "generation → verification → enhancement → regeneration".

[0212] The technical solution of the present invention takes into account the uncertainty of both the source and the load, and adopts the quantile-based statistical characteristic matching and K-medoids clustering method to construct a representative set of grid benchmark operation scenarios;

[0213] Based on the unified information coefficient feature filtering method, for a specific set of anticipated faults, the impact of operating characteristics on different dynamic safety margin indicators is evaluated to screen out key features that have a significant impact on safety margins.

[0214] By combining feedforward neural networks with automatic differentiation techniques, the sensitivity of key steady-state operating characteristics of the power grid to dynamic safety margin indicators can be quickly calculated.

[0215] A stacked denoising autoencoder is used to extract the hidden features of existing critical stability scenarios as conditional input information. This is then combined with a self-attention mechanism and a residual connection structure to guide the generator network to generate a large number of critical stability scenarios for power grids.

[0216] The high-quality scenarios generated by the generator are verified through time-domain simulation. The verified critical stability scenarios are used as new true samples for incremental training of the discriminator and stacked denoising autoencoder, forming an iterative optimization model of "generation → verification → enhancement → regeneration". Ultimately, a large number of critical stability scenarios for the power grid can be obtained.

[0217] Example 2

[0218] The purpose of this embodiment is to provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the program.

[0219] Example 3

[0220] The purpose of this embodiment is to provide a computer-readable storage medium.

[0221] A computer-readable storage medium stores a computer program, which, when executed by a processor, performs the steps of the above method.

[0222] Example 4

[0223] The purpose of this embodiment is to provide a CGAN-based power grid critical stability scenario generation system, including:

[0224] The power grid operation benchmark scenario generation module is configured to: generate a power grid operation benchmark scenario that takes into account uncertainties on both the source and load sides;

[0225] The grid key feature screening module is configured to: screen the grid key features based on the generated grid operation benchmark scenario;

[0226] The sensitivity calculation module is configured to: analyze the impact of key features on the power grid safety margin based on a feedforward neural network, and calculate the sensitivity of the selected key features of the power grid;

[0227] A grid critical stability scenario generation module is configured to: dynamically adjust key grid parameters based on sensitivity so that the grid system gradually approaches a critical stability operation point, and then generate a critical stability operation scenario;

[0228] Construct a critically stable scene generation model based on deep conditional generative adversarial networks;

[0229] Design a comprehensive optimization strategy based on the constructed critical stability scene generation model, including: designing the loss function, self-attention mechanism, residual connection structure, and incremental feedback of generated samples;

[0230] Based on a comprehensive optimization strategy, a deep conditional generative adversarial network technique is used to generate the required number of grid critical stability scenarios.

[0231] Based on the D-vine Copula theory, a multi-renewable energy station output correlation model is constructed, and a large amount of renewable energy output data is generated as input for constructing a grid operation benchmark scenario.

[0232] Considering the uncertainty on both the source and load sides, the quantile-based statistical characteristic matching and K-medoids clustering method are used to construct a grid benchmark operation scenario set, providing a diverse set of basic scenarios for the generation of critical stability scenarios.

[0233] A key feature screening method for power grids is proposed based on the unified information coefficient. For a specific set of anticipated faults, the key features that significantly affect system stability are identified by evaluating the impact of operating characteristics on different dynamic safety margin indicators.

[0234] Taking the dynamic safety margin index as the target output of model training, a regression prediction model based on a feedforward neural network is constructed to learn the nonlinear mapping relationship between key features and safety margin indicators. The local sensitivity is quickly calculated through automatic differentiation technology to achieve a quantitative description of the impact of key features on safety margin changes.

[0235] Based on the calculated sensitivity information, a parameter iterative update strategy is constructed. Combining exponential decay step size control with upper and lower boundary constraints, key parameters are fine-tuned along the negative gradient direction, so that the system gradually approaches the critical stable operating point.

[0236] SDAE is used to extract the hidden features of existing critical stability scenarios as conditional input information, and a deep conditional generative adversarial network model is constructed to generate a large number of critical stability operation scenarios of power grids.

[0237] Under the framework of generative adversarial network, a composite loss function is designed that integrates adversarial loss, boundary constraint loss and source-load power balance constraint loss. The adversarial loss is used to guide the generated samples to deceive the discriminator, the boundary constraint loss is used to penalize node parameters that exceed the allowable range of the equipment, and the source-load power balance constraint loss is used to maintain the power balance between power generation and load in the generated scenario.

[0238] In order to fully capture the correlation between the various operating characteristics of the power system, the generative adversarial network model introduces a self-attention mechanism. By calculating the correlation between the operating characteristics, weights are assigned to different features. This enables the model to comprehensively consider the influence of other features when generating each feature, thereby more accurately reflecting the overall operating characteristics of the power system.

[0239] The Generative Adversarial Network model incorporates an incremental feedback mechanism for generated samples, implementing an iterative model of "generation → verification → enhancement → regeneration." For the discriminator, the "new true samples" generated by the generator are verified through time-domain simulation. Once verified, they are added to the training set to enhance discriminative capabilities. For the SDAE, the features of these samples are introduced as incremental data, allowing for incremental training of the SDAE to optimize its hidden layer feature representation capabilities.

[0240] Example 5

[0241] The purpose of this embodiment is to provide a computer program product containing instructions, which, when running on a computer, enables the computer to execute the methods and functions involved in any of the above embodiments.

[0242] The steps involved in the apparatus of the above embodiment correspond to those of the method embodiment 1. For detailed implementation, please refer to the relevant description of embodiment 1. The term "computer-readable storage medium" should be understood to mean a single medium or multiple media containing one or more instruction sets; it should also be understood to include any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and causing the processor to perform any method of the present invention.

[0243] Those skilled in the art will appreciate that the modules or steps of the present invention described above can be implemented using a general-purpose computer device. Alternatively, they can be implemented using program code executable by a computing device, which can then be stored in a storage device and executed by the computing device. Alternatively, they can be fabricated into separate integrated circuit modules, or multiple modules or steps can be fabricated into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0244] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.

Claims

1. A method for generating power grid critical stability scenarios based on CGAN, characterized by: include: Generate a grid operation benchmark scenario that takes into account uncertainties on both the source and load sides; Screen key grid characteristics based on the generated grid operation benchmark scenario; The impact of key features on grid security margin is analyzed based on feedforward neural network, and the sensitivity of the selected key features of the grid is calculated; Dynamically adjust key grid parameters based on sensitivity to gradually approach the critical stable operation point, and then generate a critical stable operation scenario; Construct a critically stable scene generation model based on deep conditional generative adversarial networks; Design a comprehensive optimization strategy based on the constructed critical stability scene generation model, including: designing the loss function, self-attention mechanism, residual connection structure, and incremental feedback of generated samples; Based on a comprehensive optimization strategy, a deep conditional generative adversarial network technique is used to generate the required number of grid critical stability scenarios.

2. The method for generating a power grid critical stability scenario based on CGAN according to claim 1, wherein: The steps for generating the grid operation benchmark scenario considering the uncertainty on both the source and load sides are as follows: Based on Dvine Copula, a dual-period model of the output of multiple renewable energy stations was established, and a set of renewable energy output scenarios was generated based on the established model. Normal distribution is used to describe the uncertainty on the load side, with the benchmark load as the mean and the standard deviation set as a specific proportion of the benchmark load to generate a set of load scenarios; Combining source and load scenarios based on the set of new energy output scenarios and load scenarios, and performing scenario reduction; Combined with the generator dispatch plan, the generator power output is determined and the power flow is calculated to solve the voltage and power distribution of the power grid and generate a complete set of steady-state operation scenarios.

3. The method for generating a power grid critical stability scenario based on CGAN according to claim 1, wherein: The steps of screening key features of the power grid are specifically as follows: Collect node power and voltage characteristic data under different steady-state operation scenarios, calculate dynamic safety margin indicators through time-domain simulation, and construct an original data set corresponding to the characteristic data and safety margin indicators; For each dynamic safety margin indicator, calculate the UIC value between it and each candidate feature; Arrange all features related to each margin indicator in descending order based on UIC value to determine the degree of influence of the feature on the specific safety margin indicator; The cross-validation technique is used to determine the optimal UIC threshold, and the key feature set of the power grid that has a significant impact on the safety margin index is screened based on this threshold.

4. The method for generating a power grid critical stability scenario based on CGAN according to claim 1, wherein: Calculate the sensitivity of the selected key features of the power grid. The specific steps are as follows: The screened key features of the power grid are input into a feedforward neural network, and the output of the feedforward neural network is a dynamic safety margin indicator; Sensitivity is defined as the partial derivative of the safety margin indicator with respect to the input feature, and is used to measure the intensity and direction of the change in the dynamic safety margin after a small perturbation of the key feature under the current operating mode. The sensitivities of all key features are organized into gradient vector forms.

5. The method for generating a power grid critical stability scenario based on CGAN according to claim 1, wherein: The key parameter adjustment steps based on sensitivity are: Based on feedforward neural network, the automatic differentiation method is used to calculate the local sensitivity of key features to safety margin indicators; According to the sensitivity information under the current operation mode, key parameters are adjusted along the negative gradient direction; Input the updated key parameters into the feedforward neural network and output the updated safety margin value; In order to determine whether the system has approached the critical stable state, the convergence of the safety margin index is set as the termination condition; The iterative process terminates when the system has reached the critical stable operating point; The final set of key parameters is used for power flow calculation to obtain the complete power flow results of the system, and time domain simulation verification is performed to generate a critical stable operation scenario.

6. The method for generating a power grid critical stability scenario based on CGAN according to claim 1, wherein: The steps for generating power grid critical stability scenarios based on deep conditional generative adversarial networks are as follows: Based on SDAE, the latent features of the existing critical stable operation scenario set are extracted to construct conditional input information; The random noise vector and conditional input information are used as the input of the generator, and a deep network structure with a self-attention mechanism and residual connection is used to generate a critical stable scene feature set. Based on the input condition information, the discriminator performs discriminative training on real samples and generated samples, guiding the generator to generate diverse and reasonable operating scenarios while approximating the real data distribution, thereby establishing a stable adversarial optimization mechanism. For samples that are judged as true by the discriminator, the dynamic safety margin index is calculated using time domain simulation to verify whether the critical stability conditions are met; The verified scenes are included in the training set as new samples for incremental training of the discriminator and SDAE, enhancing the model's ability to extract critical stable features.

7. A CGAN-based power grid critical stability scenario generation system, characterized by: include: The power grid operation benchmark scenario generation module is configured to: generate a power grid operation benchmark scenario that takes into account uncertainties on both the source and load sides; The grid key feature screening module is configured to: screen the grid key features based on the generated grid operation benchmark scenario; The sensitivity calculation module is configured to: analyze the impact of key features on the power grid safety margin based on a feedforward neural network, and calculate the sensitivity of the selected key features of the power grid; A grid critical stability scenario generation module is configured to: dynamically adjust key grid parameters based on sensitivity so that the grid system gradually approaches a critical stability operation point, and then generate a critical stability operation scenario; Construct a critically stable scene generation model based on deep conditional generative adversarial networks; Design a comprehensive optimization strategy based on the constructed critical stability scene generation model, including: designing the loss function, self-attention mechanism, residual connection structure, and incremental feedback of generated samples; Based on a comprehensive optimization strategy, a deep conditional generative adversarial network technique is used to generate the required number of grid critical stability scenarios.

8. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the method described in any one of claims 1 to 6 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method described in any one of claims 1 to 6 are performed.

Images