A method for adaptive risk assessment of a stability control system strategy

Abstract

The application discloses a kind of stable control system strategy adaptability risk assessment method, it is related to power system safety and stability control technical field.The method includes: processing new energy historical output data and establishing time sequence correlation model, generating partition scene set;Accordingly generate multiple risk working condition sample, determine key feature vector set by transient simulation, feature extraction and dimension reduction;Scene clustering reduction is obtained to strategy scene set containing probability weight;Enumerate fault chain sequence, execute time sequence simulation to three-layer stable control strategy;Calculate multiple risk indexes to construct multidimensional matrix;Combined with new energy power time-varying quantile fluctuation band, the risk value and occurrence probability are identified and classified using fuzzy comprehensive judgment, and the comprehensive evaluation conclusion is output.The application can accurately depict the time sequence fluctuation characteristics of new energy, realize stable control strategy full-link assessment and risk form differentiation control, and provide support for power grid safety decision-making.

Description

Technical Field

[0001] This invention relates to the field of power system safety and stability control technology, and in particular to a method for risk assessment of the adaptability of stability control system strategies. This method is applicable to the scientific assessment of the risks of power grid stability control system strategies under the background of high-proportion renewable energy integration, providing technical support for the safe and stable operation of power systems under large-scale renewable energy grid connection conditions. Background Technology

[0002] With the large-scale grid connection of new energy sources such as wind and solar power, the operating characteristics of the power system have undergone profound changes. The output of new energy sources exhibits significant randomness, volatility, and intermittency, posing unprecedented challenges to the stable operation of the power grid. As the last line of defense for power grid security, the adaptability and reliability of the stability control system (safety and stability control system) directly determine the grid's safety level under disturbances. Given the high proportion of new energy integration, how to scientifically assess the risks faced by the stability control system strategy has become a core issue in the field of power system security analysis.

[0003] Existing methods for handling uncertainties in new energy sources mainly fall into two categories: First, deterministic equivalence methods based on scenario analysis, which replace continuous probability distributions with typical power output scenarios (such as large-scale generation, small-scale generation, and rated output) and then conduct independent stability simulation analyses for each scenario; second, probabilistic power flow analysis methods based on Monte Carlo simulations, which generate system operating state samples through extensive random sampling, statistically analyze the probability distribution characteristics of system state variables, and then assess the system's safety margin. These methods have achieved certain engineering application results in new energy scenarios with low to medium penetration rates.

[0004] Taking probabilistic power flow analysis as an example, its basic principle is as follows: First, a probability distribution model (such as Weibull distribution or Beta distribution) is established for the original input quantities of new energy sources, such as wind speed and solar intensity. Then, the power conversion curve is used to map this model to the probability distribution of new energy output. Subsequently, Monte Carlo sampling or semi-invariant methods are used to solve for the probability distribution of state quantities such as system node voltage and line power flow. Finally, indicators such as the probability of exceeding limits and the expected power shortage are used to characterize the system safety risk. The core advantage of this method is that it can systematically consider the randomness of the input quantities and describe the system risk in probabilistic language.

[0005] However, the aforementioned existing technologies have the following key drawbacks: First, both traditional scenario-based methods and probabilistic power flow methods ignore the temporal correlation of renewable energy output and cannot accurately describe the dynamic fluctuation characteristics of renewable energy power on the time axis. As a result, the generated scenario sets are difficult to reflect the impact of extreme dynamic behaviors such as power ramp-up and reversal on the adaptability of the stability control strategy within continuous time periods. Second, existing risk assessment methods are mostly limited to a single dimension (such as voltage stability or frequency stability) or a single fault mode. They lack the ability to model the entire link of the stability control strategy action process under conditions of successive and widespread faults. Furthermore, they cannot distinguish the differentiated impact of different risk forms, such as high probability of high loss and low probability of high loss, on system safety decisions. This results in coarse-grained assessment conclusions, which are difficult to provide effective support for the refined optimization of stability control strategies. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method for risk assessment of the adaptability of a stability control system strategy. This method integrates time-series Monte Carlo simulation and quantile regression techniques to accurately describe the time-series correlation and extreme fluctuation characteristics of new energy output. It models the full-link response of the stability control strategy through fault chain development search and strategy action simulation, and uses fuzzy comprehensive evaluation technology to distinguish different risk forms, providing a scientific decision-making basis for the adaptive optimization of the stability control system strategy.

[0007] To achieve the above objectives, the technical solution provided by the present invention is as follows: A method for risk assessment of the adaptability of a stability control system strategy, the method comprising: The historical output time series data of new energy power plants are processed to establish a time-series correlation structure model and generate a partitioned scenario set covering different confidence levels; Based on the partitioned scenario set, multi-risk working condition samples are generated, and simulation analysis is carried out for each working condition. After feature extraction and dimensionality reduction, the key feature vector set is determined. Based on the key feature vector set, the working conditions are clustered and reduced to obtain a set of stability control system strategy scenarios containing probability weights; Based on the aforementioned stability control system strategy scenario set, the fault chain sequence under each representative scenario is enumerated, and the three-layer stability control strategy is subjected to time-series simulation, outputting the fault evolution sequence and the timing record of the stability control strategy action; Based on the fault evolution sequence, multiple risk indicators are calculated to form a multidimensional risk indicator matrix. Using the multidimensional risk index matrix and the time-varying quantile fluctuation band of new energy power as input, the fuzzy comprehensive evaluation method is used to jointly identify and classify the risk value and the probability of occurrence, and output the comprehensive risk assessment conclusion of the stability control system strategy.

[0008] Furthermore, the historical output time-series data of new energy power plants are processed to establish a time-series correlation structure model, generating a set of partitioned scenarios covering different confidence levels, including: Using the preprocessed historical power output time series data of new energy power plants as input, the marginal probability distribution of new energy power output is fitted for different time periods. The Weibull distribution is used for wind power and the Beta distribution is used for photovoltaic power to obtain the set of probability distribution parameters for each time period. Based on the set of probability distribution parameters, the Teng Copula method is used to model the nonlinear temporal correlation between power outputs in adjacent time periods, resulting in a temporal correlation structure model. Using the aforementioned time-series correlation structure model as a framework, the Monte Carlo method is employed for time-series sampling to generate new energy output time-series scenarios, forming an original scenario sample library; The original scene sample library is input into the quantile regression model, and multiple conditional quantile functions covering low, medium and high confidence levels are fitted for the output distribution at each time point to obtain a set of conditional quantile functions; Based on the set of conditional quantile functions, the regression coefficient vector of each quantile is solved by minimizing the quantile loss function; Based on the quantile regression coefficient vector, the quantile curves are arranged along the time axis to form the time-varying quantile fluctuation band of the new energy power, and the partition scenario sets corresponding to low output, medium output and high output are defined.

[0009] Furthermore, the preprocessing includes sequentially performing missing value imputation, outlier removal, and normalization on the historical output time series data of the new energy power stations; the quantile levels of the conditional quantile functions are 5%, 10%, 25%, 50%, 75%, 90%, and 95%, respectively, and the independent variables of each conditional quantile function include historical output lag terms and meteorological prediction variables; the number of new energy output time series scenarios is no less than 10,000.

[0010] Furthermore, based on the aforementioned partitioned scenario set, multi-risk working condition samples are generated, and simulation analysis is performed on each working condition. Through feature extraction and dimensionality reduction, a set of key feature vectors is determined, including: Using the output quantile intervals of each new energy power station in the partitioned scenario set at each time period as input, each output quantile interval is mapped to the coordinate axis of a high-dimensional uniform sampling space to obtain a high-dimensional uniform sampling space. Based on the high-dimensional uniform sampling space, Latin hypercube sampling is used to generate a high-dimensional sample point set to ensure that there is exactly one and only one sample in each sub-interval. The high-dimensional sample point set is mapped back to the actual physical space of new energy output, and combined with maintenance plans, load levels and tie line power flow to form a multi-risk operating condition sample set; Using the multi-risk operating condition sample set as input, transient simulation is performed on each operating condition instance. The mean, standard deviation, peak value, and rate of change of each time-domain response sequence are extracted to obtain a statistical feature set. A fast Fourier transform is applied to the time-domain response sequence to extract the frequency and damping ratio of the dominant oscillation mode, resulting in a frequency domain feature set. The electrical distance and betweenness centrality of each node are calculated based on the current network topology to obtain a topological feature set. The margin difference between the preset action criteria of the stability control device and the current system state variables is extracted to obtain a stability control action condition feature set. After concatenating the statistical feature set, frequency domain feature set, topological feature set, and stability control action condition feature set, dimensionality reduction is performed using principal component analysis or an autoencoder. Principal components with a cumulative variance contribution rate of not less than 95% are retained to determine the key feature vector set.

[0011] Furthermore, when generating the high-dimensional sample point set using Latin hypercube sampling, the number of sample points is between 500 and 2000.

[0012] Furthermore, based on the key feature vector set, the operating scenarios are clustered and reduced to obtain a set of stability control system strategy scenarios including probability weights, including: Using the various feature dimensions in the key feature vector set as input, the analytic hierarchy process (AHP) combined with expert knowledge is used to assign weight coefficients to each feature dimension. Among them, the stability control action condition feature and the safety margin feature are assigned the first weight, and the remaining features are assigned the second weight. The first weight is higher than the second weight, thus obtaining the weight coefficient set of each feature dimension. Based on the set of weight coefficients for each feature dimension, a distance matrix between scenes is constructed using weighted Mahalanobis distance; Based on the inter-scene distance matrix, the K-medoids clustering algorithm is used to reduce the number of scenes to 20 to 50 representative scenes. The number of clusters is determined by the silhouette coefficient and the Davies-Bouldin index. Each cluster uses its medoid as a representative scene, and the probability of occurrence of all sample points in the cluster is accumulated and assigned to the representative scene to obtain the candidate stability control system strategy scene set. Using the candidate stability control system strategy scenario set and the original scenario set as input, the Wasserstein distance is used to verify the probability distribution deviation of the scenario sets before and after reduction. If the deviation exceeds a preset threshold, the number of clusters is increased and the above clustering steps are repeated until the accuracy requirement is met, and the verified stability control system strategy scenario set is output.

[0013] Furthermore, based on the aforementioned stability control system strategy scenario set, the fault chain sequence under each representative scenario is enumerated, and a timing simulation is performed on the three-layer stability control strategy to output the fault evolution sequence and the timing record of the stability control strategy actions, including: Using the power flow distribution, node voltage, and frequency of each representative scenario in the stability control system strategy scenario set as input, calculate the line power flow over-limit margin, voltage stability margin, frequency deviation, and power flow transfer sensitivity matrix. Construct the vulnerability index of each component by weighted summation to obtain a list of key components. Based on the list of key components, an algorithm combining depth-first search and probabilistic pruning is adopted to enumerate N-1, N-2 and finite-depth Nk fault chain sequences. In each expansion step, the vulnerability index of each component is updated, and components exceeding the threshold are selected as candidates for the next level of faults. When the cumulative probability is lower than the minimum significant probability, pruning is performed to obtain a set of fault chain sequences. Using the fault chain sequence set as input, timing simulation is performed on the three-layer stability control strategy: emergency control layer, recovery control layer, and optimization control layer. The first-layer emergency control performs machine tripping and load shedding within 0 to 500 milliseconds after the fault. The second-layer recovery control performs low-frequency and low-voltage load shedding and automatic reclosing within 500 milliseconds to 5 minutes. The third-layer optimization control performs load transfer and network reconfiguration after more than 5 minutes. Transient simulation tools are used to record the action time, action amount, and system state after the action of each layer, and the fault evolution sequence and stability control strategy action timing record are output.

[0014] Furthermore, based on the fault evolution sequence, multiple risk indicators are calculated to form a multidimensional risk indicator matrix, including: Using the fault evolution sequence as input, calculate the transient stability margin exceedance probability, frequency safety risk index, voltage stability risk index, expected value of load shedding, expected value of generator shedding, expected value of power outage loss, and risk index of clustered successive faults, and combine the above indicators to form the multidimensional risk index matrix.

[0015] Furthermore, using the multidimensional risk index matrix and the time-varying quantile fluctuation band of the new energy power as input, a fuzzy comprehensive evaluation method is employed to jointly identify and classify the risk magnitude and probability of occurrence, outputting a comprehensive risk assessment conclusion for the stability control system strategy, including: Using the multidimensional risk index matrix and the time-varying quantile fluctuation band of new energy power as input, the confidence interval of new energy power is divided into extreme interval, high-risk interval and normal interval. With the multidimensional risk index value as the vertical axis, a two-dimensional risk assessment matrix is ​​constructed. The matrix elements represent the comprehensive quantitative value of various risks occurring in each confidence interval. Based on the two-dimensional risk assessment matrix, a trapezoidal fuzzy membership function is used to map the risk magnitude dimension to four fuzzy language sets: low, medium, high, and extremely high. A sigmoid membership function is used to map the probability dimension to three fuzzy sets: low probability, medium probability, and high probability, thus obtaining the risk membership vector for each scenario. Based on the aforementioned risk membership vector, a fuzzy comprehensive evaluation method is used to determine the comprehensive risk level of each scenario through the maximum membership principle or the weighted average method. Based on the comprehensive risk level, the risk patterns of each scenario are divided into four categories: high loss with high probability, high loss with low probability, low loss with high probability, and low loss with low probability. For high loss with high probability scenarios, an emergency plan will be activated immediately; for high loss with low probability scenarios, a special emergency plan will be formulated; for low loss with high probability scenarios, they will be included in daily monitoring; and for low loss with low probability scenarios, the level of regular monitoring will be maintained. The comprehensive risk assessment conclusion of the stability control system strategy, which includes risk level, risk pattern classification, and handling recommendations, will be output.

[0016] Furthermore, the method also includes: Using grid node and branch data as input, a dynamic graph data structure is constructed. The graph structure data is obtained by taking the renewable energy output, load level, voltage amplitude and phase angle of the nodes as node features. Based on the graph structure data, a graph convolutional network or a graph attention network is used for message passing and feature aggregation to obtain node embedding representations. Using the node embedding representation as input and risk index values ​​from historical simulation data as supervision labels, a risk assessment network model is obtained through training. In the online assessment phase, real-time operating data is mapped into graph data and then input into the risk assessment network model to directly output predicted values ​​of various risk indicators, thereby achieving rapid online risk assessment for large-scale power grids.

[0017] Furthermore, the method also includes: The stability control strategy parameters are modeled as the action space of a reinforcement learning agent, with the power grid operation state as the state space and the weighted negative value of the comprehensive risk index as the reward function, thus obtaining the reinforcement learning problem framework. Based on the aforementioned reinforcement learning problem framework, the agent is trained in an offline simulation environment using a near-end policy optimization algorithm, enabling it to learn the optimal combination of stability control policy parameters under different new energy output scenarios, thereby obtaining the agent policy model. Using the online predicted confidence interval of new energy power and the intelligent agent strategy model as input, the system outputs real-time suggestions for adjusting the stability control strategy to adapt to the current operating state.

[0018] The beneficial effects of this invention are: 1. This invention integrates time-series Monte Carlo simulation and quantile regression techniques to construct a new energy power fluctuation band with time-varying statistical characteristics. It can effectively capture the time-series correlation and extreme fluctuation characteristics of new energy output, and significantly improve the accuracy of new energy scenario modeling.

[0019] 2. This invention uses Latin hypercube sampling to achieve uniform and efficient coverage of multiple risk conditions in a high-dimensional quantile scenario space, which is significantly better than the representativeness and computational efficiency of traditional random sampling methods, and can fully cover the new energy output space with a smaller sample size.

[0020] 3. This invention proposes a multi-dimensional nonlinear feature extraction method that integrates system operation mode and stability control action conditions. It combines clustering algorithm to achieve scene reduction and typical representation, which greatly reduces the scene scale while retaining key risk information, and balances assessment accuracy and computational feasibility.

[0021] 4. This invention explicitly embeds the uncertainty of new energy power and the hierarchical mechanism of defense strategy into the fault chain search framework, realizing the systematic traversal and simulation of the successive fault paths, and providing technical support for the full-link evaluation of the stability control system strategy.

[0022] 5. This invention constructs a two-dimensional fuzzy evaluation matrix that integrates risk indicator values ​​and occurrence probabilities, and innovatively distinguishes different risk forms such as high loss with high probability and high loss with low probability, providing refined decision support for the optimization of stability control strategies and realizing accurate assessment and differentiated management of power system safety risks. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the adaptive risk assessment method for a stability control system according to the present invention. Figure 2 A flowchart illustrating the process of processing historical power output time series data of new energy power plants and generating partitioned scenario sets. Detailed Implementation

[0024] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0025] It should be understood that the specific embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit the scope of the invention.

[0026] like Figure 1 As shown, the present invention provides a method for risk assessment of the adaptability of a stability control system strategy, comprising the following steps: Step S1: Process the historical output time series data of new energy power plants, establish a time-series correlation structure model, and generate a partitioned scenario set covering different confidence levels.

[0027] like Figure 2 As shown, step S1 specifically includes: Step S1.1: Using the preprocessed historical power output time series data of the new energy power station as input, fit the marginal probability distribution of new energy power output for different time periods. The Weibull distribution is used for wind power and the Beta distribution is used for photovoltaic power to obtain the probability distribution parameter set for each time period.

[0028] In this embodiment, preprocessing includes sequentially performing missing value imputation, outlier removal, and normalization on the historical output time series data of new energy power plants. Missing value imputation uses linear interpolation or the average value of nearby times to fill in missing points in the historical data; outlier removal uses the 3σ criterion or box plot method to identify and process abnormal output points; normalization divides the output data of each power plant by its rated capacity to convert it into a normalized value in the range of 0 to 1.

[0029] For wind farms, the Weibull distribution is used to describe the probability distribution characteristics of their power output: in, This represents the wind power output value. For scale parameters, For shape parameters, a Beta distribution is used to describe the probability distribution characteristics of power output for photovoltaic power plants. in, For photovoltaic power output, and For distribution parameters, Let be the Beta function. Using the maximum likelihood estimation method, the parameters of the above distribution are fitted to historical power output data for different time periods (such as 24 hours in a day or different seasons) to obtain the set of probability distribution parameters for each time period.

[0030] Step S1.2: Based on the set of probability distribution parameters, the Teng Copula method is used to model the nonlinear temporal correlation between power outputs in adjacent time periods to obtain a temporal correlation structure model.

[0031] Vine Copula is a method for constructing high-dimensional joint distributions, capable of flexibly describing complex dependency structures between variables. In this step, a nonlinear time-series correlation model of renewable energy output in adjacent time periods is established using C-Vine or D-Vine structures, constructed as follows: in, It is distributed on the periphery. For a bivariate Copula density function, and These are the corresponding marginal distribution functions.

[0032] Step S1.3: Using the aforementioned time-series correlation structure model as a framework, the Monte Carlo method is used to perform time-series sampling to generate new energy output time-series scenarios and form an original scenario sample library.

[0033] Based on the Copula model constructed in step S1.2, a large number of random samples conforming to the temporal correlation structure are generated using the Monte Carlo method. The specific sampling process includes: (1) Draw a vector of relevant uniform random variables u=(u1,u2,...,u2) from the uniform distribution U(0,1). n ); (2) Use the inverse transformation method to convert the uniform random variable into a random vector that follows the joint distribution of Copula; (3) The above random vector is converted into a new energy output time series with physical meaning by using the inverse function of the marginal distribution.

[0034] In this embodiment, the number of generated new energy output time-series scenarios is no less than 10,000 to ensure the stability of statistical characteristics and complete coverage of the probability space.

[0035] Step S1.4: Input the original scene sample library into the quantile regression model, and fit multiple conditional quantile functions covering low, medium and high confidence levels for the output distribution at each time point to obtain a set of conditional quantile functions.

[0036] Quantile regression models can describe the relationship between different quantiles of the dependent variable and the independent variable. Compared to traditional regression, which only focuses on the mean, quantile regression provides more comprehensive distribution information. (This refers to the output of new energy sources at each time t.) Establish the following conditional quantile function: in, Quantile level, The covariate vector (containing historical output lag terms) , (such as meteorological forecast variables) For the solution to be found Quantile regression coefficient vector.

[0037] In this embodiment, the quantile levels of the conditional quantile function are selected as 5%, 10%, 25%, 50%, 75%, 90%, and 95%, respectively, covering low, medium, and high confidence levels. Step S1.5: Based on the conditional quantile function set, solve for the regression coefficient vector of each quantile by minimizing the quantile loss function. For each quantile level The corresponding regression coefficient vector is obtained by minimizing the following quantile loss function: in, , This is an indicator function.

[0038] Step S1.6: Based on the quantile regression coefficient vector, arrange the quantile curves according to the time axis to form the time-varying quantile fluctuation band of the new energy power, and delineate the partition scenario set corresponding to low output, medium output and high output.

[0039] Substituting the regression coefficient vector β_τ obtained from the preceding steps into the conditional quantile function yields the prediction curves corresponding to each quantile τ. These curves are then arranged along the time axis to form a time-varying quantile fluctuation band for renewable energy power. Based on this fluctuation band, the renewable energy output space is divided into multiple zones: Low output range: between the 5% and 25% quantiles Medium output range: between the 25th and 75th percentiles High output range: between the 75th and 95th percentiles Extreme output range: [0%, 5%]∪[95%, 100%] quantiles These partitions together constitute a set of partitioned scenarios covering different confidence levels, providing a foundation for subsequent screening of risky operating conditions.

[0040] Step S2: Based on the partitioned scenario set, generate multi-risk working condition samples, conduct simulation analysis for each working condition, and determine the key feature vector set through feature extraction and dimensionality reduction.

[0041] Step S2 specifically includes: Step S2.1: Using the output quantile intervals of each new energy power station in the partitioned scenario set at each time period as input, map each output quantile interval to the coordinate axis of the high-dimensional uniform sampling space to obtain the high-dimensional uniform sampling space.

[0042] Assume the system has Each of the new energy power stations has a new energy power station. The output of each time period was divided into Each quantile interval forms a A high-dimensional sampling space is used, where each dimension corresponds to the output range of a power station during a specific time period. The quantile intervals for each dimension are then defined. The coordinate axes [0,1] of the high-dimensional uniform sampling space are mapped, and the coordinate transformation relationship is established.

[0043] Step S2.2: Based on the high-dimensional uniform sampling space, use Latin hypercube sampling to generate a high-dimensional sample point set, ensuring that there is exactly one and only one sample in each sub-interval.

[0044] Latin hypercube sampling (LHS) is a stratified sampling method that can significantly reduce the required sample size while ensuring sample representativeness. The specific steps are as follows: (1) Divide each dimension evenly into N parts (in this embodiment, the number of sample points N is between 500 and 2000). (2) In each dimension, randomly select a point from each sub-interval to ensure that there is exactly one sample in each interval; (3) Randomly combine sample points of each dimension to form N high-dimensional sample points, which constitute a high-dimensional sample point set.

[0045] Step S2.3: Map the high-dimensional sample point set back to the actual physical space of new energy output, and combine it with maintenance plans, load levels and tie line power flow to form a multi-risk operating condition sample set.

[0046] Mapping sample points in a high-dimensional uniform sampling space back to actual new energy output values: in, For station At any moment The output value, This is the quantile function corresponding to the marginal distribution. The coordinates of the sampling point in this dimension.

[0047] The obtained renewable energy output scenarios are combined with the system's current maintenance plans (such as the maintenance status of important transmission lines), predicted load levels, and tie line planned power flow to form a complete multi-risk operating condition sample set. Each sample represents a power system operating condition with clear physical meaning.

[0048] Step S2.4: Using the multi-risk operating condition sample set as input, perform transient simulation for each operating condition instance, extract the mean, standard deviation, peak value, and rate of change of each time-domain response sequence to obtain a statistical feature set; apply a fast Fourier transform to the time-domain response sequence to extract the frequency and damping ratio of the dominant oscillation mode to obtain a frequency domain feature set; calculate the electrical distance and betweenness centrality of each node based on the current network topology to obtain a topological feature set; extract the margin difference between the preset action criteria of the stability control device and the current system state quantity to obtain a stability control action condition feature set. For each operating condition instance, power system transient simulation (including electromechanical transients or electromagnetic transients) is performed to obtain the time-domain response sequences of key nodes, such as voltage amplitude, phase angle, and frequency, and the following multi-dimensional features are extracted: (1) Statistical Feature Set: Calculate the statistics for each time-domain response sequence, including: - Mean: - Standard deviation: - Peak value: - Rate of change: (2) Frequency domain feature set: for time domain sequences Apply a Fast Fourier Transform (FFT) to obtain its spectrum. Extract from it: Frequency of dominant oscillation mode (The frequency corresponding to the maximum power spectral density) Damping ratio of dominant mode (Obtained through half-power bandwidth method or Prony analysis) (3) Topological feature set: Based on the current network topology, calculate: Electrical distance between nodes (Shortest path length considering branch impedance) Node betweenness centrality (Representation Node) (Importance in the network) (4) Stabilization Action Condition Feature Set: Extract the margin difference between the preset action criterion threshold of the stabilization device and the current system state quantity, such as: Power deficit margin: Frequency margin: Voltage over-limit margin: These margin differences reflect the sensitivity to the triggering of stabilization strategies.

[0049] Step S2.5: After concatenating the statistical feature set, frequency domain feature set, topological feature set, and stability control action condition feature set, perform dimensionality reduction using principal component analysis or an autoencoder, retaining principal components with a cumulative variance contribution rate of not less than 95%, and determine the key feature vector set.

[0050] The above-mentioned multiple features are concatenated to form a high-dimensional feature vector. Dimensionality reduction can be achieved using principal component analysis (PCA) or an autoencoder. Principal Component Analysis (PCA) Dimensionality Reduction: (1) Calculate the characteristic covariance matrix (2) Perform eigenvalue decomposition on the covariance matrix: (3) Select the eigenvalue with the largest value eigenvectors, satisfying the cumulative variance contribution rate (4) Construct the projection matrix The dimensionality-reduced feature vectors are obtained. Alternatively, an autoencoder can be used for nonlinear dimensionality reduction: (1) Design encoder network and decoder network (2) Optimize the reconstruction loss: (3) Use the trained encoder to extract high-dimensional features Mapping to a low-dimensional representation Finally, the set of key feature vectors after dimensionality reduction is obtained. This is used for subsequent scene clustering and reduction.

[0051] Step S3: Based on the key feature vector set, cluster and reduce the working conditions to obtain a set of stability control system strategy scenarios containing probability weights.

[0052] Step S3 specifically includes: Step S3.1: Using the various feature dimensions in the key feature vector set as input, the analytic hierarchy process (AHP) combined with expert knowledge is used to assign weight coefficients to each feature dimension. The stability control action condition feature and safety margin feature are assigned first weights, and the remaining features are assigned second weights. The first weight is higher than the second weight, thus obtaining the weight coefficient set for each feature dimension.

[0053] The Analytic Hierarchy Process (AHP) is a structured decision-making method that establishes a judgment matrix by having experts compare the importance of indicators in pairs, and then calculates the eigenvectors to obtain the weights of each indicator. In this embodiment, features are divided into two categories based on their importance to the risk of the stability control strategy: First weighted features: stability control operation condition features and safety margin features (such as frequency over-limit margin and voltage stability margin). Second weighting features: other statistical features, frequency domain features, and topological features Assume the weight of the first weighted feature is The weight of the second weighted feature is ,and , Within each feature category, the weights of each specific dimension are further refined using the AHP method, ultimately yielding a complete set of feature dimension weight coefficients. .

[0054] Step S3.2: Based on the set of weight coefficients for each feature dimension, construct the inter-scene distance matrix using weighted Mahalanobis distance.

[0055] Mahalanobis distance takes into account the correlation between features and can handle features of different scales and correlations. For feature vectors and Its weighted Mahalanobis distance is defined as: in, This is a weighted diagonal matrix, whose diagonal elements are the weight coefficients determined in step S3.1; Let be the characteristic covariance matrix.

[0056] Calculate the distance between all scene pairs and construct the complete scene distance matrix. ,in .

[0057] Step S3.3: Based on the distance matrix between the scenes, the number of scenes is reduced to 20 to 50 representative scenes using the K-medoids clustering algorithm. The number of clusters is determined by the silhouette coefficient and the Davies-Bouldin index. Each cluster is represented by its medoid, and the probability of occurrence of all sample points in the cluster is accumulated and assigned to the representative scene to obtain the candidate stability control system strategy scene set.

[0058] K-medoids clustering algorithms (such as PAM, Partitioning Around Medoids) are a variant of K-means that select actual data points as cluster centers, making them more robust to outliers. The optimization objective of K-medoids is: in, For the first Clusters, This is the medoid of the cluster (one of the actual data points).

[0059] To determine the optimal number of clusters This embodiment uses two clustering evaluation metrics: - Silhouette Coefficient: Measures intra-cluster compactness and inter-cluster separation. - Davies-Bouldin index: assesses the compactness and separability of clustering results. calculate Among the various values ​​from 10 to 60 corresponding to the above indicators, the one with the largest silhouette coefficient and the smallest Davies-Bouldin index was selected. The value is used as the final cluster number.

[0060] After clustering, the probability weights of the sample points within each cluster are accumulated and assigned to the medoid (representative scenario) of that cluster to ensure... This forms a set of candidate stability control system strategy scenarios.

[0061] Step S3.4: Using the candidate stability control system strategy scenario set and the original scenario set as input, the Wasserstein distance is used to verify the probability distribution deviation of the scenario sets before and after reduction. If the deviation exceeds a preset threshold, the number of clusters is increased and the above clustering steps are repeated until the accuracy requirements are met, and the verified stability control system strategy scenario set is output.

[0062] The Wasserstein distance (also known as Earth Mover's Distance) measures the difference between two probability distributions and is suitable for evaluating the consistency of distributions before and after scene reduction. For the original scene set P and the reduced scene set Q, the Wasserstein distance is calculated as follows: in, It is the set of all possible joint distributions, with marginal distributions P and Q.

[0063] In practical calculations, the following linear programming problem can be solved to approximate the Wasserstein distance: If the calculated Wasserstein distance Exceeding the preset threshold If the value is 0.1, then increase the cluster number K and repeat step S3.3 until... The final verified set of stability control system strategy scenarios is output.

[0064] Step S4: Combining the stability control system strategy scenario set, enumerate the fault chain sequence under each representative scenario, perform timing simulation on the three-layer stability control strategy, and output the fault evolution sequence and the timing record of stability control strategy actions.

[0065] Step S4 specifically includes: Step S4.1: Using the power flow distribution, node voltage and frequency of each representative scenario in the stability control system strategy scenario set as input, calculate the line power flow over-limit margin, voltage stability margin, frequency deviation and power flow transfer sensitivity matrix, construct the vulnerability index of each component by weighted summation, and obtain the list of key components.

[0066] For each representative scenario, the following calculations are performed: (1) Line power flow margin: in, For the line The current trend Its rated capacity.

[0067] (2) Voltage stability margin: Calculated by the continuous power flow method (CPF) or the minimum singularity method.

[0068] (3) Frequency deviation: in, For system frequency, The nominal frequency is 50Hz.

[0069] (4) Power Transfer Sensitivity Matrix: Calculate the PTDF (Power Transfer Distribution Factor) matrix to describe the redistribution of power flow on other branches after a fault in a branch.

[0070] Based on the above indicators, the vulnerability index of structural components is as follows: in, These are the weighting coefficients for each indicator.

[0071] According to the vulnerability index Sort, Filter Components exceeding the threshold are compiled into a list of critical components.

[0072] Step S4.2: Based on the list of key components, an algorithm combining depth-first search and probabilistic pruning is used to enumerate N-1, N-2 and finite-depth Nk fault chain sequences. In each expansion step, the vulnerability index of each component is updated, and components exceeding the threshold are selected as candidates for the next level of faults. When the cumulative probability is lower than the minimum significance probability, pruning is performed to obtain a set of fault chain sequences.

[0073] An algorithm combining depth-first search (DFS) and probabilistic pruning is used to enumerate the fault chain sequence: (1) Select components from the list of critical components in sequence as the initial faults to form the N-1 fault starting point; (2) For each N-1 fault, calculate the system state after the fault and update the vulnerability index of each component; (3) Screening updated vulnerability indices Exceeding the threshold The components are considered as candidates for the next level of failure; (4) Select one component from the candidates to cause a failure, forming an N-2 failure. Repeat steps (2) and (3) to form N-3 and N-4 failures until the preset maximum depth is reached (e.g., 5). (5) Assign conditional probabilities to each fault chain. Calculate the cumulative probability ; (6) When the cumulative probability Below the minimum significance probability When the value is 0.0001, pruning is performed to stop the further expansion of the fault chain.

[0074] Through the above process, a fault chain sequence set is obtained for each representative scenario, which includes multiple fault chains and their corresponding probabilities.

[0075] Step S4.3: Using the fault chain sequence set as input, perform timing simulation on the three-layer stability control strategy of emergency control layer, recovery control layer and optimization control layer: The first layer emergency control performs machine tripping and load shedding within 0 to 500 milliseconds after the fault; the second layer recovery control performs low-frequency and low-voltage load shedding and automatic reclosing within 500 milliseconds to 5 minutes; the third layer optimization control performs load transfer and network reconfiguration after more than 5 minutes; use transient simulation tools to record the action time, action amount and system state after action of each layer, and output the fault evolution sequence and stability control strategy action timing record.

[0076] For each fault chain sequence, simulate the timing action process of the three-layer strategy of the stability control system: (1) First-level emergency control (0-500ms): Emergency control measures triggered immediately after a malfunction occurs. This includes emergency unit disconnection (rapid unit disconnection). Emergency load shearing (rapid load shearing) Direct power cut-off response time requirement is within hundreds of milliseconds. (2) Second-level recovery control (500ms-5min): Restorative measures implemented after emergency control. Low-frequency load shedding device activated (staged load shedding). Low voltage shearing device activated The automatic reclosing device has activated (attempting to restore the tripped circuit). Response time ranges from hundreds of milliseconds to several minutes. (3) Third-level optimization control (>5min): Optimization and adjustment measures after the system has initially stabilized Load transfer (restoring power to areas with disconnected loads via alternative pathways) Network reconstruction (adjusting network topology to restore optimal operation). Response time is more than several minutes For each layer of control policy, record the following information: Action trigger time t_act Action quantity ΔP (machine cut-off quantity or load cut-off quantity) System state after action (node ​​voltage V, frequency f, phase angle θ, etc.) Through the above timing simulation, a complete fault evolution sequence and timing record of the stability control strategy action are output.

[0077] Step S5: Based on the fault evolution sequence, calculate multiple risk indicators to form a multidimensional risk indicator matrix.

[0078] Using the fault evolution sequence output from step S4 and the timing record of the stability control strategy actions as input, calculate the following risk indicators: (1) Probability of exceeding the transient stability margin limit : Calculate the probability of transient stability-related indicators exceeding limits under each fault chain: ,in, For the set of all failure chains, For indicator functions, For the first The transient stability margin of a fault chain, This is the safety threshold.

[0079] (2) Frequency security risk indicators : ,in, For the first The lowest point of system frequency in a fault chain, For frequency safety thresholds (e.g., 49Hz). This is the frequency recovery time.

[0080] (3) Voltage stability risk indicators : ,in, For the system node set, For nodes voltage, x This refers to the voltage safety threshold (e.g., 0.9 per unit). (4) Expected value of load shedding : ,in, For the first The load shedding power after a fault chain triggers the stabilization action.

[0081] (5) Expected value of cutting volume : ,in, For the first The power output of the machine after a fault chain triggers the stabilization action.

[0082] (6) Expected value of power outage loss : ,in, The corresponding expected power shortage (MWh).

[0083] (7) Risk indicators for clustered successive failures : ,in, For lengths exceeding a preset threshold (e.g.) A set of fault chains (level 1 and above), For the first The severity index of a fault chain. This is the cascade depth weighting coefficient.

[0084] Combining the above indicators forms a multidimensional risk indicator matrix: Step S6: Using the multidimensional risk index matrix and the time-varying quantile fluctuation band of new energy power as input, the fuzzy comprehensive evaluation method is used to jointly identify and classify the risk value and the probability of occurrence, and output the comprehensive risk assessment conclusion of the stability control system strategy.

[0085] Step S6 specifically includes: Step S6.1: Using the multidimensional risk index matrix and the time-varying quantile fluctuation band of new energy power as input, divide the confidence interval of new energy power into extreme interval, high-risk interval and normal interval. With the multidimensional risk index value as the vertical axis, construct a two-dimensional risk assessment matrix. The matrix elements represent the comprehensive quantitative value of various risks occurring in each confidence interval.

[0086] The confidence intervals for new energy power are divided into three categories: Extreme range: [0%, 5%] ∪ [95%, 100%] High-risk range: [5%, 25%] ∪ [75%, 95%] Normal range: [25%, 75%] Constructing a two-dimensional risk assessment matrix The rows correspond to different confidence intervals, the columns correspond to different risk indicators, and the matrix elements... Indicates the first Within the confidence interval, the first The quantitative value of the risk-like indicator is calculated using the following formula: in, For the first A set of scenarios within a confidence interval For the scene probability weights, For the scene The corresponding number Risk index value.

[0087] Step S6.2: Based on the two-dimensional risk assessment matrix, a trapezoidal fuzzy membership function is used to map the risk magnitude dimension to four fuzzy language sets: low, medium, high, and extremely high. A sigmoid membership function is used to map the probability dimension to three fuzzy sets: low probability, medium probability, and high probability, to obtain the risk membership vector for each scenario.

[0088] (1) Fuzzy membership function for risk metrics: for each type of risk indicator Define four fuzzy language sets {low, medium, high, and extremely high}, and use a trapezoidal fuzzy membership function: - Low risk ( Membership function: - Medium risk ( Membership function: - High risk ( Membership function: - Very High Risk (VH) Membership Function: The parameters a, b, c, and d are determined based on the statistical distribution characteristics of the risk indicators, and can be set as follows: in and These represent the mean and standard deviation of the risk indicator across all scenarios.

[0089] (2) Fuzzy membership function for probability dimension: Define three fuzzy sets {low probability, medium probability, high probability}, and use a Sigmoid membership function: - Low probability (LP) membership function: - Medium probability (MP) membership function: - High probability (HP) membership function: The parameters are determined based on historical statistical probability distributions.

[0090] For each scenario k, calculate its membership vector along the two dimensions of risk value and probability: Step S6.3: Based on the risk membership vector, the comprehensive risk level of each scenario is determined by using the fuzzy comprehensive evaluation method through the maximum membership principle or the weighted average method.

[0091] A risk assessment matrix is ​​constructed using the fuzzy comprehensive evaluation method: We obtain a 4×3 matrix, representing the membership degree of scenario k under different risk values ​​and probability combinations.

[0092] The final risk level is determined using either the maximum membership principle or the weighted average method. - Maximum membership principle: Select the risk value and probability combination corresponding to the largest element in R_k. - Weighted average method: Assign weights to different risk levels (e.g., VH=4, H=3, M=2, L=1; HP=3, MP=2, LP=1) and calculate the weighted average.

[0093] Step S6.4: Based on the comprehensive risk level, classify the risk patterns of each scenario into four categories: high loss and high probability, high loss and low probability, low loss and high probability, and low loss and low probability. For high loss and high probability scenarios, immediately activate the emergency plan; for high loss and low probability scenarios, formulate a special emergency plan; for low loss and high probability scenarios, include them in daily monitoring; for low loss and low probability scenarios, maintain the level of routine monitoring; and output the comprehensive risk assessment conclusion of the stability control system strategy, which includes the risk level, risk pattern classification, and handling suggestions.

[0094] Based on the combination of risk value membership degree and probability membership degree, the risk forms in various scenarios are divided into four categories: (1) High Risk, High Probability (HH): The risk value is high or extremely high, and the probability is high. Response strategy: Immediately activate the emergency plan and revise the stabilization strategy. (2) High risk, low probability (HL): The risk value is high or extremely high, but the probability is low or medium. Response strategy: Include in the key prevention and control list, and formulate a special emergency response plan. (3) Low Risk, High Probability (LH): The risk level is low or medium, but the probability is high. Response strategy: Incorporate into routine monitoring and pay attention to cumulative effects. (4) Low Risk, Low Probability (LL): The risk level is low or medium, and the probability is low or medium. Response strategy: Maintain routine monitoring levels The final output includes a comprehensive risk assessment conclusion for the stability control system strategy, containing the following: Scenario description: Power output level of new energy sources and system operation characteristics Risk Level: Low, Medium, High, and Extremely High Risk Level Determination Risk types are classified into four categories: HH, HL, LH, and LL. Recommendations: Specific response measures for different risk profiles In another embodiment of the present invention, a power grid topology-sensing risk prediction method based on graph neural networks is also included: (1) Using the data of power grid nodes and branches as input, a dynamic graph data structure is constructed. The graph structure data is obtained by taking the renewable energy output, load level, voltage amplitude and phase angle of the nodes as node features. (2) Based on the graph structure data, a graph convolutional network or a graph attention network is used to perform message passing and feature aggregation to obtain node embedding representations; (3) Using the node embedding representation as input, and using the risk index values ​​in the historical simulation data as supervision labels, a risk assessment network model is obtained; in the online assessment stage, after mapping the real-time running data into graph data, the data is input into the risk assessment network model, and the predicted values ​​of various risk indicators are directly output to realize rapid online risk assessment of large-scale power grids.

[0095] Specifically, the graph neural network model is constructed as follows: (1) Graph data representation: The power grid is represented as a graph. ,in It is a set of nodes (including buses, generators, etc.). This is an edge set (representing a path). Each node... Having feature vectors This includes information such as new energy output, load level, voltage amplitude, and phase angle. (2) Definition of graph convolutional layer: in, For nodes In the Layer feature representation, For nodes The neighborhood group, The normalization constant is For learnable weight matrix, is the activation function. (3) Graph attention mechanism: , in, Attention weights represent the node's attention weights. For nodes The importance of the nodes. (4) Stacking of multi-layer graph neural networks: By stacking multiple layers of graph convolution or graph attention layers, node information is aggregated layer by layer to finally obtain the embedded representation of the nodes. (5) Prediction layer: The node embedded representation is input into the fully connected layer to predict the values ​​of various risk indicators.

[0096] This method reduces assessment time from minutes to seconds, enabling rapid online risk assessment for large-scale power grids.

[0097] In another embodiment of the present invention, a reinforcement learning-based adaptive optimization method for stability control strategy is also included: (1) the stability control strategy parameters are modeled as the action space of a reinforcement learning agent, the power grid operating state is used as the state space, and the weighted negative value of the comprehensive risk index is used as the reward function to obtain a reinforcement learning problem framework; (2) based on the reinforcement learning problem framework, the agent is trained in an offline simulation environment using a near-end strategy optimization algorithm to learn the optimal combination of stability control strategy parameters under different new energy output scenarios to obtain an agent strategy model; (3) using the online predicted new energy power confidence interval and the agent strategy model as input, a stability control strategy adjustment suggestion adapted to the current operating state is output in real time.

[0098] Specifically, the reinforcement learning framework is designed as follows: (1) State space : Includes key characteristics of the power grid's operating status, such as new energy output scenarios, system power flow distribution, key node voltages, frequencies, etc. (2) Action space : Includes key parameters of the stabilization strategy, such as the cut-off threshold, the load cut-off threshold, and the action criterion threshold. (3) Reward function Defined as the weighted negative value of the comprehensive risk index, i.e. The lower the risk, the higher the reward. (4) Policy network: A multilayer perceptron or recurrent neural network is used, with input state features and output probability distribution of action parameters. (5) Optimization algorithm: The proximal policy optimization (PPO) algorithm is used, which includes the following steps: a. Using the current policy b. Collect trajectory data. estimating the advantage function. c. Optimize the policy parameters by maximizing the objective function: in Importance sampling ratio By using a reinforcement learning framework, adaptive optimization of the stability control strategy parameters can be achieved, effectively compensating for the lack of adaptability of traditional offline tuning strategies.

[0099] By using a reinforcement learning framework, adaptive optimization of the stability control strategy parameters can be achieved, effectively compensating for the lack of adaptability of traditional offline tuning strategies.

[0100] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any person skilled in the art can make many possible variations and modifications to the technical solutions of the present invention, or modify them into equivalent embodiments, without departing from the scope of the present invention. Therefore, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention, without departing from the content of the present invention, shall still fall within the protection scope of the present invention.

[0101] This invention discloses a method for adaptive risk assessment of a stability control system strategy, comprising: processing historical output time-series data of renewable energy power plants, establishing a time-series correlation structure model, and generating a set of partitioned scenarios covering different confidence levels; generating multi-risk operating condition samples based on the partitioned scenario sets, conducting simulation analysis for each operating condition, and determining a key feature vector set through feature extraction and dimensionality reduction; clustering and reducing the operating condition scenarios based on the key feature vector set to obtain a set of stability control system strategy scenarios containing probability weights; enumerating fault chain sequences under each representative scenario, performing time-series simulation on a three-layer stability control strategy, and outputting fault evolution sequences and time-series records of stability control strategy actions; calculating multiple risk indicators to form a multi-dimensional risk indicator matrix; and using a fuzzy comprehensive evaluation method to jointly identify and classify risk values ​​and occurrence probabilities, outputting a comprehensive risk assessment conclusion for the stability control system strategy. This invention can accurately characterize the time-series correlation and extreme fluctuation characteristics of renewable energy output, perform full-link assessment of stability control strategies, and achieve accurate risk identification and classification.

Claims

1. A method for assessing the adaptive risk of a stability control system strategy, characterized in that, The method includes: The historical output time series data of new energy power plants are processed to establish a time-series correlation structure model and generate a set of partitioned scenarios covering different confidence levels; Based on the partitioned scenario set, multi-risk working condition samples are generated, and simulation analysis is carried out for each working condition. After feature extraction and dimensionality reduction, the key feature vector set is determined. Based on the key feature vector set, the working conditions are clustered and reduced to obtain a set of stability control system strategy scenarios containing probability weights; Based on the aforementioned stability control system strategy scenario set, the fault chain sequence under each representative scenario is enumerated, and the three-layer stability control strategy is subjected to time-series simulation, outputting the fault evolution sequence and the timing record of the stability control strategy action; Based on the fault evolution sequence, multiple risk indicators are calculated to form a multidimensional risk indicator matrix. Using the multidimensional risk index matrix and the time-varying quantile fluctuation band of new energy power as input, the fuzzy comprehensive evaluation method is used to jointly identify and classify the risk value and the probability of occurrence, and output the comprehensive risk assessment conclusion of the stability control system strategy.

2. The method for assessing the adaptive risk of a stability control system strategy according to claim 1, characterized in that, The historical output time series data of new energy power plants are processed to establish a time-series correlation structure model, generating a set of partitioned scenarios covering different confidence levels, including: Using the preprocessed historical power output time series data of new energy power plants as input, the marginal probability distribution of new energy power output is fitted for different time periods. The Weibull distribution is used for wind power and the Beta distribution is used for photovoltaic power to obtain the set of probability distribution parameters for each time period. Based on the set of probability distribution parameters, the Teng Copula method is used to model the nonlinear temporal correlation between power outputs in adjacent time periods, resulting in a temporal correlation structure model. Using the aforementioned time-series correlation structure model as a framework, the Monte Carlo method is employed for time-series sampling to generate new energy output time-series scenarios, forming an original scenario sample library; The original scene sample library is input into the quantile regression model, and multiple conditional quantile functions covering low, medium and high confidence levels are fitted for the output distribution at each time point to obtain a set of conditional quantile functions; Based on the set of conditional quantile functions, the regression coefficient vector of each quantile is solved by minimizing the quantile loss function; Based on the quantile regression coefficient vector, the quantile curves are arranged along the time axis to form the time-varying quantile fluctuation band of the new energy power, and the partition scenario sets corresponding to low output, medium output and high output are defined.

3. The method for assessing the adaptive risk of a stability control system strategy according to claim 2, characterized in that, The preprocessing includes sequentially performing missing value imputation, outlier removal, and normalization on the historical power output time series data of the new energy power stations; the quantile levels of the conditional quantile functions are 5%, 10%, 25%, 50%, 75%, 90%, and 95%, respectively, and the independent variables of each conditional quantile function include historical power output lag terms and meteorological prediction variables; the number of new energy power output time series scenarios is no less than 10,000.

4. The method for assessing the adaptive risk of a stability control system strategy according to claim 1, characterized in that, Based on the aforementioned partitioned scenario set, multi-risk working condition samples are generated. Simulation analysis is performed on each working condition. After feature extraction and dimensionality reduction, a set of key feature vectors is determined, including: Using the output quantile intervals of each new energy power station in the partitioned scenario set at each time period as input, each output quantile interval is mapped to the coordinate axis of a high-dimensional uniform sampling space to obtain a high-dimensional uniform sampling space. Based on the high-dimensional uniform sampling space, Latin hypercube sampling is used to generate a high-dimensional sample point set to ensure that there is exactly one and only one sample in each sub-interval. The high-dimensional sample point set is mapped back to the actual physical space of new energy output, and combined with maintenance plans, load levels and tie line power flow to form a multi-risk operating condition sample set; Using the multi-risk operating condition sample set as input, transient simulation is performed on each operating condition instance. The mean, standard deviation, peak value, and rate of change of each time-domain response sequence are extracted to obtain a statistical feature set. A fast Fourier transform is applied to the time-domain response sequence to extract the frequency and damping ratio of the dominant oscillation mode, resulting in a frequency domain feature set. The electrical distance and betweenness centrality of each node are calculated based on the current network topology to obtain a topological feature set. The margin difference between the preset action criteria of the stability control device and the current system state variables is extracted to obtain a stability control action condition feature set. After concatenating the statistical feature set, frequency domain feature set, topological feature set, and stability control action condition feature set, dimensionality reduction is performed using principal component analysis or an autoencoder. Principal components with a cumulative variance contribution rate of not less than 95% are retained to determine the key feature vector set.

5. The method for assessing the adaptive risk of a stability control system strategy according to claim 4, characterized in that, When generating the high-dimensional sample point set using Latin hypercube sampling, the number of sample points is between 500 and 2000.

6. The method for assessing the adaptive risk of a stability control system strategy according to claim 1, characterized in that, Based on the aforementioned key feature vector set, the operating scenarios are clustered and reduced to obtain a set of stability control system strategy scenarios containing probability weights, including: Using the various feature dimensions in the key feature vector set as input, the analytic hierarchy process (AHP) combined with expert knowledge is used to assign weight coefficients to each feature dimension. Among them, the stability control action condition feature and the safety margin feature are assigned the first weight, and the remaining features are assigned the second weight. The first weight is higher than the second weight, thus obtaining the weight coefficient set of each feature dimension. Based on the set of weight coefficients for each feature dimension, a distance matrix between scenes is constructed using weighted Mahalanobis distance; Based on the inter-scene distance matrix, the K-medoids clustering algorithm is used to reduce the number of scenes to 20 to 50 representative scenes. The number of clusters is determined by the silhouette coefficient and the Davies-Bouldin index. Each cluster uses its medoid as a representative scene, and the probability of occurrence of all sample points in the cluster is accumulated and assigned to the representative scene to obtain the candidate stability control system strategy scene set. Using the candidate stability control system strategy scenario set and the original scenario set as input, the Wasserstein distance is used to verify the probability distribution deviation of the scenario sets before and after reduction. If the deviation exceeds a preset threshold, the number of clusters is increased and the above clustering steps are repeated until the accuracy requirement is met, and the verified stability control system strategy scenario set is output.

7. The method for assessing the adaptive risk of a stability control system strategy according to claim 1, characterized in that, Based on the aforementioned stability control system strategy scenario set, the fault chain sequence under each representative scenario is enumerated. Timing simulation is performed on the three-layer stability control strategy, outputting the fault evolution sequence and the timing record of the stability control strategy actions, including: Using the power flow distribution, node voltage, and frequency of each representative scenario in the stability control system strategy scenario set as input, calculate the line power flow over-limit margin, voltage stability margin, frequency deviation, and power flow transfer sensitivity matrix. Construct the vulnerability index of each component by weighted summation to obtain a list of key components. Based on the list of key components, an algorithm combining depth-first search and probabilistic pruning is adopted to enumerate N-1, N-2 and finite-depth Nk fault chain sequences. In each expansion step, the vulnerability index of each component is updated, and components exceeding the threshold are selected as candidates for the next level of faults. When the cumulative probability is lower than the minimum significant probability, pruning is performed to obtain a set of fault chain sequences. Using the fault chain sequence set as input, timing simulation is performed on the three-layer stability control strategy: emergency control layer, recovery control layer, and optimization control layer. The first-layer emergency control performs machine tripping and load shedding within 0 to 500 milliseconds after the fault. The second-layer recovery control performs low-frequency and low-voltage load shedding and automatic reclosing within 500 milliseconds to 5 minutes. The third-layer optimization control performs load transfer and network reconfiguration after more than 5 minutes. Transient simulation tools are used to record the action time, action amount, and system state after the action of each layer, and the fault evolution sequence and stability control strategy action timing record are output.

8. The method for assessing the adaptive risk of a stability control system strategy according to claim 1, characterized in that, Based on the aforementioned fault evolution sequence, multiple risk indicators are calculated to form a multidimensional risk indicator matrix, including: Using the fault evolution sequence as input, calculate the transient stability margin exceedance probability, frequency safety risk index, voltage stability risk index, expected value of load shedding, expected value of generator shedding, expected value of power outage loss, and risk index of clustered successive faults, and combine the above indicators to form the multidimensional risk index matrix.

9. The method for assessing the adaptive risk of a stability control system strategy according to claim 1, characterized in that, Using the multidimensional risk index matrix and the time-varying quantile fluctuation band of the new energy power as input, a fuzzy comprehensive evaluation method is used to jointly identify and classify the risk magnitude and probability of occurrence, and output a comprehensive risk assessment conclusion for the stability control system strategy, including: Using the multidimensional risk index matrix and the time-varying quantile fluctuation band of new energy power as input, the confidence interval of new energy power is divided into extreme interval, high-risk interval and normal interval. With the multidimensional risk index value as the vertical axis, a two-dimensional risk assessment matrix is ​​constructed. The matrix elements represent the comprehensive quantitative value of various risks occurring in each confidence interval. Based on the two-dimensional risk assessment matrix, a trapezoidal fuzzy membership function is used to map the risk magnitude dimension to four fuzzy language sets: low, medium, high, and extremely high. A sigmoid membership function is used to map the probability dimension to three fuzzy sets: low probability, medium probability, and high probability, thus obtaining the risk membership vector for each scenario. Based on the aforementioned risk membership vector, a fuzzy comprehensive evaluation method is used to determine the comprehensive risk level of each scenario through the maximum membership principle or the weighted average method. Based on the comprehensive risk level, the risk patterns of each scenario are divided into four categories: high loss with high probability, high loss with low probability, low loss with high probability, and low loss with low probability. For high loss with high probability scenarios, an emergency plan will be activated immediately; for high loss with low probability scenarios, a special emergency plan will be formulated; for low loss with high probability scenarios, they will be included in daily monitoring; and for low loss with low probability scenarios, the level of regular monitoring will be maintained. The comprehensive risk assessment conclusion of the stability control system strategy, which includes risk level, risk pattern classification, and handling recommendations, will be output.

10. The method for assessing the adaptive risk of a stability control system strategy according to claim 1, characterized in that, Also includes: Using grid node and branch data as input, a dynamic graph data structure is constructed. The graph structure data is obtained by taking the renewable energy output, load level, voltage amplitude and phase angle of the nodes as node features. Based on the graph structure data, a graph convolutional network or a graph attention network is used for message passing and feature aggregation to obtain node embedding representations. Using the node embedding representation as input and risk index values ​​from historical simulation data as supervision labels, a risk assessment network model is obtained through training. During the online assessment phase, real-time operational data is mapped into graph data and then input into the risk assessment network model to directly output predicted values ​​of various risk indicators, thereby enabling rapid online risk assessment of large-scale power grids.

11. The method for assessing the adaptive risk of a stability control system strategy according to claim 1, characterized in that, Also includes: The stability control strategy parameters are modeled as the action space of a reinforcement learning agent, with the power grid operation state as the state space and the weighted negative value of the comprehensive risk index as the reward function, thus obtaining the reinforcement learning problem framework. Based on the aforementioned reinforcement learning problem framework, the agent is trained in an offline simulation environment using a near-end policy optimization algorithm, enabling it to learn the optimal combination of stability control policy parameters under different new energy output scenarios, thereby obtaining the agent policy model. Using the online predicted confidence interval of new energy power and the intelligent agent strategy model as input, the system outputs real-time suggestions for adjusting the stability control strategy to adapt to the current operating state.

Images