Power system evaluation method and system based on topology enhanced GCN classification regression collaboration

Abstract

The application discloses a power system evaluation method and system based on topological enhancement GCN classification regression cooperation, and belongs to the technical field of intelligent power evaluation. The method comprises an offline training stage and an online evaluation stage. In the offline stage, an optimal load reduction optimization model is built to obtain two types of labels, i.e., classification and regression, and node features are constructed by fusing node electrical parameters and multi-dimensional topological center features, a weighted adjacency matrix is generated by combining branch electrical parameters, and a topological enhancement graph data is formed. A graph-level classification network and a multi-task node regression network are built, the classification adopts a weighted cross-entropy to relieve the sample imbalance problem, the regression introduces an MSE to match a KKT multi-constraint composite loss to constrain the prediction to meet the physical rules of the power grid. In the online stage, the classification network is used to screen no-cut load samples in advance, only fault samples are sent into the regression network to calculate the cut load, and according to the prediction results, LOLP and EDNS reliability indexes are calculated. The application describes the coupling relationship between the fault network topology and the branch electrical parameters, and improves the reliability calculation efficiency and the index calculation precision.

Description

Technical Field

[0001] This invention relates to the field of technology, and specifically to a power system evaluation method and system based on topology-enhanced GCN classification and regression synergy. Background Technology

[0002] With the large-scale grid connection of new energy sources such as wind power and photovoltaics in new power systems, the randomness of power output on the power source side, the temporal fluctuation of the load side, and the randomness of grid faults have significantly increased. Power system reliability assessment has become a core supporting technology for grid planning, dispatch risk control, and grid transformation. The traditional mainstream reliability assessment adopts a technical route of Monte Carlo simulation combined with optimal load reduction optimization. This approach solves the load shedding amount by enumerating massive random fault scenarios and constructing DC optimal load reduction models for each scenario. Key reliability indicators such as LOLP and EDNS are statistically analyzed based on massive samples. However, with the continuous expansion of provincial and inter-regional interconnected power grids, the number of system buses, branches, and units often exceeds a thousand. Repeatedly calling commercial optimization solvers under hundreds of thousands of simulation samples will result in huge computing power overhead. In scenarios such as online risk warning and rapid comparison of multiple planning schemes, the computational latency cannot meet the timeliness requirements of engineering. To address this, the industry has gradually shifted to the acceleration approach of machine learning to replace optimization solutions. Intelligent models are trained using input-label datasets generated by offline simulations, and neural network inference is used to replace optimization calculations in the online stage. This has become the mainstream research and development direction in the field of power reliability in the past decade.

[0003] Domestic and international academic communities have conducted extensive basic research on accelerating reliability evaluation using convolutional neural networks, resulting in two representative and classic academic achievements. Internationally, Kamruzzaman et al. published a paper titled "A convolutional neural network-based approach to composite power system reliability evaluation" in the *International Journal of Electrical Power & Energy Systems* (SCI journal, DOI: 10.1016 / j.ijepes.2021.107468) in 2022. This paper was the first to introduce a single-structure CNN into the reliability evaluation of a composite power system. It used a regularized matrix to flatten the grid data as model input, relied on CNN feature extraction to determine system failure states, and reduced computational load by filtering failure samples and then solving the optimization model. This approach verified the feasibility of using deep learning to replace some optimization solutions. However, the model input is only a flattened numerical matrix, which cannot naturally adapt to the topological connections of grid nodes and branches, and cannot quantify the impact of changes in the grid structure after a fault on the power supply gap. In the domestic field, Shao Chengcheng's team published "A Reliability Assessment Method for New Energy Power Systems Based on Classification-Regression Convolutional Neural Networks" in the Proceedings of the Chinese Society for Electrical Engineering (EI, CSCD Core Journal, Vol. 44, No. 23) in 2024. Based on the aforementioned research, they proposed a collaborative architecture of CNN classification + segmented regression: first, the CNN classification model is used to screen samples without load shedding and directly set the load to zero; then, CNN regression is used only for samples with power outages due to faults to predict the load shedding amount, which greatly reduces the frequency of online optimization calls. This is a benchmark academic solution for accelerating reliability assessment using classification-regression in China. However, both papers use Euclidean regular tensors as model inputs, which can only flatten the power grid topology and electrical parameters. They cannot explicitly characterize the physical connection between the bus and the branch, or the strength of the differential electrical coupling between different lines. The model cannot effectively capture information on topological changes such as the outage of key branches and the instability of hub nodes after a fault. At the same time, the model loss only uses the MSE error between the predicted value and the actual load shedding. The training process does not bind power balance, upper and lower limits of unit output, branch capacity and other power physical constraints. The regression output is prone to infeasible solutions that violate the law of power grid operation. In the end, the calculation error of reliability indicators such as EDNS is too large, which limits the engineering implementation.

[0004] The domestic invention patent that has been implemented based on the above-mentioned academic achievements is CN115564310A (A Reliability Assessment Method for New Energy Power Systems Based on Convolutional Neural Networks). Its technical solution completely inherits the CNN classification-regression architecture of the aforementioned paper in the "Chinese Journal of Electrical Engineering": it stores the electrical data of the entire system in a fixed-size two-dimensional matrix, builds a two-layer convolutional + fully connected classification network and a multi-layer convolutional regression network, trains offline based on samples generated by Monte Carlo, first classifies and screens out safe operating conditions online, and then regresses to estimate the load shedding amount, so as to realize the rapid calculation of reliability indicators. Compared to traditional Monte Carlo successive optimization schemes, this patent effectively shortens computation time and has verified its acceleration effect in small- and medium-scale test power grids. However, the patented scheme still has three inherent technical shortcomings that cannot be avoided: First, the data input is a regular matrix, and the unstructured power grid topology of buses and branches is forcibly normalized. The topology information caused by grid reconstruction and branch disconnection after a fault is lost, and the importance characteristics of key nodes and key transmission channels cannot be embedded in the model input. Second, the model training is based solely on label-supervised loss and does not introduce the KKT physical optimality constraints corresponding to the optimal load shedding model. When predicting load shedding, results that violate the laws of electrical physics, such as unit output exceeding limits, branch power flow exceeding limits, and system power imbalance, are often observed, resulting in insufficient reliability of the indicators. Third, the model cannot define edge weights based on the differences in line reactance and branch availability. The electrical coupling strength of various branches cannot be distinguished by the model, and the fault impact of high-impedance tie lines and main transmission lines is quantified equally, further amplifying the prediction bias. This makes it difficult to adapt to the complex online reliability assessment scenario of large power grids with a high proportion of new energy access.

[0005] Existing academic papers and patents all adopt the technical approach of CNN and regularized data. However, due to the inherent defects of Euclidean data representation and training without physical constraints, the topology adaptation capability and the physical feasibility of prediction results are difficult to meet the practical engineering needs in the scenario of large-scale power grids with high penetration of new energy. At present, existing power reliability-related patents related to graph convolution mostly focus on distribution network fault location, grid reconfiguration, and equipment status assessment. Most of them only achieve single fault identification or grid optimization, and do not have a complete technical chain of classification-regression collaboration and KKT physical constraints specifically for reliability assessment. They lack an integrated assessment scheme that takes into account both the refined topology representation and power operation constraints. Summary of the Invention

[0006] Based on the above-mentioned technical problems, this application discloses a power system evaluation method and system based on topology-enhanced GCN classification and regression collaboration; the power system evaluation method based on topology-enhanced GCN classification and regression collaboration includes an offline training phase and an online evaluation phase;

[0007] The offline training phase includes:

[0008] S1. Collect basic parameters of the power system and generate a system status sample containing component operating status, load data, and wind power output data through Monte Carlo sampling. ;

[0009] S2, for each system state sample Construct and solve the DC optimal load shedding model, and generate load shedding classification labels based on the solution results. Regression labels with shear load ;

[0010] S3. Constructing topology enhancement graph data based on system failure topology. The The node feature matrix, Weighted adjacency matrix for branch parameters;

[0011] S4. Utilizing graph data A graph-level GCN classification model is trained, which is processed sequentially through two graph convolutional layers, a global average pooling layer, and three fully connected layers to output the probability of load shedding.

[0012] S5. Utilizing graph data A node-level GCN regression model is trained, which is configured with a multi-task output head. The training uses a joint loss function that integrates MSE loss and KKT-constrained residual loss. ;

[0013] The online assessment phase includes:

[0014] S6. For the power system state samples to be evaluated, construct the topology augmentation graph data as in step S3, input it into the trained GCN classification model, and distinguish whether the samples have experienced load shedding according to the preset classification threshold.

[0015] S7. Input only the samples that are determined to have experienced load shedding into the trained node-level GCN regression model. The regression model outputs the load shedding amount of each bus node. Based on the discrimination results of all samples and the predicted values ​​of load shedding, the power system reliability indices LOLP and EDNS are statistically calculated.

[0016] Preferably, in S2, the total system load shedding is obtained by summing the optimal load shedding of all buses in a single sample, using the following formula: In the formula: For the first The system's overall optimal load shedding amount corresponding to each sample; It is the set of all bus nodes in the power grid; For the first Busbar under each sample The optimal load shedding amount is obtained through optimization.

[0017] Set non-negative threshold The following formula is used to classify 0-1 binary labels:

[0018]

[0019] In the formula: For the first The cut load classification label for each sample; This is a preset non-negative threshold used to eliminate numerical calculation errors; a label value of 1 indicates that the corresponding system sample has a shear load, and a label value of 0 indicates that the corresponding system sample has no shear load.

[0020] Preferably, S3 constructs the node feature matrix. At that time, single busbar node Feature vectors: ;

[0021] In the formula: For the first Nodes under sample topology eigenvectors; For nodes The original load; For nodes The maximum available output when connected to a conventional generating unit; For nodes Available output from connected wind turbine units; , , , The nodes in the fault topology are listed in order. The corresponding degree centrality, clustering coefficient, betweenness centrality, and improved proximity centrality are calculated. The four types of centrality values ​​are solved according to the post-fault topology, and the node feature matrix is ​​formed by concatenating all node feature vectors.

[0022] Preferably, S3 constructs a weighted adjacency matrix. Formula for calculating matrix elements:

[0023]

[0024] In the formula: For the first Nodes in the adjacency matrix of samples With nodes Weighting coefficients between them; It is the set of all physical branches of the system; branch road The available status indicator, with a value of 1 representing that the branch is intact and available, and a value of 0 representing that the branch is faulty and out of service; branch road The absolute value of reactance; To avoid using extremely small positive numbers with a denominator of zero; the node feature matrix and the weighted adjacency matrix are combined to form single-sample topology enhancement graph data.

[0025] The preferred formula for S4 single-layer graph convolution operation is as follows: ;

[0026] In the formula: , The first Layer convolution input, the first Convolutional layers output feature matrices; , They are identity matrices of the same order; For matrix The corresponding degree matrix; , The first Convolutional layers can be trained to generate weight matrices and bias vectors; It is a ReLU nonlinear activation function;

[0027] The formula for global average pooling after two convolutional layers is as follows: ;

[0028] In the formula: This refers to the graph-level embedding features corresponding to the entire power grid diagram; This represents the total number of nodes in the power grid. The node after two layers of convolution operations Feature data;

[0029] Formula for iterative calculation of a three-layer fully connected layer: ;

[0030] In the formula: , These are the feature vectors of the preceding and following layers of the fully connected layer; , These are the weight and bias parameters for the fully connected layer; the Sigmoid function is used at the end of the fully connected layer to output the probability of load shedding.

[0031] Preferably, the classification model is trained using a binary cross-entropy loss function with class weights. The loss calculation formula is as follows:

[0032]

[0033] In the formula: This represents the overall loss value of the classification model; The total number of samples used in model training; These are the class weight coefficients used to balance sample imbalance; The first output of the model The predicted probability of a sample experiencing load shedding is used to update all network parameters of the classification model in reverse.

[0034] Preferably, the regression model of S5 is set with a multi-task output head to synchronously output node load shedding, unit output, wind curtailment, and optimization dual variables; ;

[0035] In the formula: This represents the total loss of the regression model; The mean squared error loss between the predicted load shedding value and the actual label; The weighted penalty coefficient for the KKT loss term; This is a composite constraint loss constructed based on the optimal model KKT conditions;

[0036] Mean squared error loss:

[0037]

[0038] In the formula: The number of effective samples used in training the regression model; For the model to predict the first Nodes under each sample Shear load; For the first The total load of each sample system is the true label;

[0039]

[0040] In the formula: For the first The original feasibility loss corresponding to each sample; For the first The dual feasibility loss corresponding to each sample; For the first The complementary relaxation loss corresponding to each sample; For the first The stationary condition loss corresponds to each sample; the four sub-losses are constructed according to the KKT optimal conditions corresponding to the optimal load reduction model, without separately limiting the expressions of the four sub-losses.

[0041] Preferred penalty parameters The DCV calculation formula is determined based on the average normalized constraint violation rate of the validation set: ;

[0042] In the formula: To verify the violation of the average normalization constraint; The total number of samples in the validation set; For the first The normalization constraint of the validation samples violated the index; multiple candidate sets were traversed. The value corresponding to the minimum DCV value is selected as the final penalty parameter.

[0043] Preferably, during the online evaluation phase, after the samples to be tested are classified and screened, regression prediction is only enabled for samples with reduced load. The reliability index is calculated using the following formula:

[0044]

[0045] In the formula: This represents the probability of insufficient power. For the expected power shortage; To assess the total number of samples to be tested online; For the first Binary discriminant for each sample shear load; For the first The total system load shedding is predicted from a sample.

[0046] A power system assessment system based on topology-enhanced GCN classification and regression synergy includes:

[0047] The data acquisition module collects various basic power data, including busbars, branches, conventional generator sets, wind turbines, loads, component failure rates, and line reactance, and standardizes all data dimensions and storage formats.

[0048] The status sampling module uses Monte Carlo sampling to generate system status samples containing component commissioning status, node load, and available wind power output based on component failure probability, load fluctuation range, and wind power output probability distribution.

[0049] The label generation module builds a DC optimal load reduction model for each system state sample and completes the solution calculation. Based on the optimal solution, it calculates the total load shedding amount and the binary load shedding classification label.

[0050] The graph data construction module, based on the post-fault power grid topology, calculates the node electrical characteristics and four types of topological centrality parameters to form a node feature vector and concatenates them to obtain a node feature matrix. It then combines the branch operating parameters and reactance values ​​to calculate a weighted adjacency matrix and generates topology-enhanced graph data.

[0051] The classification model training module constructs a graph-level classification network consisting of two layers of symmetric normalized graph convolutional layers, a global average pooling layer, and three fully connected layers connected in sequence. Iterative optimization of all network parameters is achieved using a binary cross-entropy loss function with class weights.

[0052] The regression model training module builds a node-level regression network with four layers of shared graph convolutional layers and multi-branch multi-task output heads. The network training adopts a composite loss function consisting of a mean squared error term and four KKT constraint loss terms, and the loss weighting coefficients are determined based on the constraint violation index of the validation set.

[0053] The online evaluation and calculation module receives the power grid data to be evaluated and converts it into topology enhancement graph data. It then calls the trained classification network and selectively enables the regression network to determine the load shedding status and predict the load shedding amount. Based on the prediction statistics, it calculates two reliability indicators, LOLP and EDNS.

[0054] Compared with the prior art, the technical solution of this application has the following technical effects:

[0055] This invention establishes a step-by-step computational architecture that combines classification and regression. During online reliability calculations, the classification model can pre-identify system conditions where load shedding is not expected, and regression calculations are only performed on samples where load shedding is determined to have occurred. This setup eliminates numerous unnecessary numerical solution steps, simplifies the overall evaluation process, reduces the overall data processing workload, and makes the statistical calculation of power system reliability indicators more concise and streamlined, adapting to the routine evaluation needs of large-scale system samples.

[0056] This invention integrates multiple types of electrical parameters of nodes and four types of topology center indicators when constructing graph data, along with a weighted adjacency matrix generated based on line parameters. These features can fully reconstruct the actual topology of the power grid and the electrical coupling relationships of branches after a fault occurs, enabling the graph convolutional model to fully capture the differences in node status and network connection characteristics, enriching the dimensions of the model's input information and improving the model's ability to identify features under different fault conditions.

[0057] This invention designs a layered, concatenated graph convolutional and fully connected network structure. Features are extracted layer by layer according to a standardized convolution calculation formula, and the model parameters are optimized using a weighted cross-entropy loss method to address the sample imbalance problem. The ordered network layers can progressively extract deep features of the power grid, and the weight settings mitigate model bias caused by class imbalance, stabilizing the accuracy of operating condition discrimination in the classification stage and providing a reliable preliminary screening basis for subsequent regression calculations.

[0058] This invention introduces a composite loss combining mean squared error and KKT multivariate constraints during the regression training phase. The regression network synchronously outputs load shedding and various auxiliary physical variables. Model training is guided by optimality constraints, ensuring that the constrained prediction results closely align with various system operating boundaries. This reduces the probability of predicted values ​​violating physical rules such as power balance and equipment output limits, improving the rationality of load shedding prediction results and ensuring stable and reliable reliability index calculations.

[0059] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0060] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0062] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows:

[0063] Figure 1 This is a flowchart illustrating the overall implementation of the topology-enhanced GCN reliability assessment method of this invention.

[0064] Figure 2 This is a schematic diagram of the hierarchical structure of the graph-level GCN classification model of the present invention;

[0065] Figure 3 This is a block diagram of the multi-task node GCN regression network and loss structure of the present invention;

[0066] Figure 4 This is a diagram showing the functional modules of the power reliability assessment system of the present invention. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0068] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0069] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0070] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0071] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0072] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0073] Example 1

[0074] This embodiment mainly describes a power system evaluation method based on topology-enhanced GCN classification and regression synergy, such as... Figure 1 As shown, it includes an offline training phase and an online evaluation phase;

[0075] The offline training phase includes:

[0076] S1. Collect basic parameters of the power system and generate a system status sample containing component operating status, load data, and wind power output data through Monte Carlo sampling. ;

[0077] S2, for each system state sample Construct and solve the DC optimal load shedding model, and generate load shedding classification labels based on the solution results. Regression labels with shear load ;

[0078] S3. Constructing topology enhancement graph data based on system failure topology. The The node feature matrix, Weighted adjacency matrix for branch parameters;

[0079] S4. Utilizing graph data A graph-level GCN classification model is trained, which is processed sequentially through two graph convolutional layers, a global average pooling layer, and three fully connected layers to output the probability of load shedding.

[0080] S5. Utilizing graph data A node-level GCN regression model is trained, which is configured with a multi-task output head. The training uses a joint loss function that integrates MSE loss and KKT-constrained residual loss. ;

[0081] The online assessment phase includes:

[0082] S6. For the power system state samples to be evaluated, construct topology augmentation graph data as in step S3, input it into the trained GCN classification model, and distinguish whether the samples have experienced load shedding based on the preset classification threshold.

[0083] S7. Input only the samples that are determined to have experienced load shedding into the trained node-level GCN regression model. The regression model outputs the load shedding amount of each bus node. Based on the discrimination results of all samples and the predicted value of load shedding, the power system reliability index is statistically calculated.

[0084] Furthermore, in specific implementations, the classification model and regression model are not general algorithms detached from the power system scenario. Instead, they take the generator's available output, load level, wind power's available output, degree centrality, clustering coefficient, betweenness, and proximity centrality as inputs, and output load shedding judgment results and load shedding prediction results that serve the calculation of reliability indicators, forming a dedicated technical processing flow adapted to power system reliability assessment.

[0085] To ensure consistent definitions throughout the text, the following agreement is made in advance: It is the collection consisting of all the busbars of the power grid. It is a collection consisting of all transmission branches in the entire network. This represents the entire collection of conventional generator sets in the system. Represents the entire collection of wind turbine units; It is connected to the bus. A subset of conventional units, It is connected to the bus. A subset of wind turbines; Used to mark the first A set of randomly generated power grid operating condition samples.

[0086] Offline Training S1: Basic Data Acquisition and System State Sample Generation

[0087] This stage serves as the data source for the entire algorithm, comprehensively collecting measured and design parameters of the power system. Specific parameters cover various types of information, including rated power consumption parameters for each bus, reactance, impedance, and ultimate transmission capacity of each transmission branch, rated maximum output of each generator unit, single-unit installed capacity of wind farms, historical measured load time-series data, historical wind power output data, and the annual average fault probability of generators and transmission lines. After unifying the data format and units of all parameters, they serve as the underlying data source for subsequent fault random sampling and optimization modeling. The basic sampling method used in this step is Monte Carlo random sampling. Optional alternatives include: importance sampling, Latin hypercube sampling, quasi-Monte Carlo, scene clustering, fault-oriented screening, and generation of measured historical scenes. Importance sampling focuses on amplifying the proportion of low-probability severe faults. Latin hypercube sampling stratifies and evenly divides the random variable intervals. Quasi-Monte Carlo relies on low-discrepancy sequences to optimize sample clustering. Scene clustering compresses the massive number of samples. Fault screening focuses on high-risk component combinations. Historical scenes replicate real working conditions based on measured data. The output data structure of all alternative methods is completely consistent with the original sampling.

[0088] In the component failure determination stage, generators and transmission lines in the power system are components that can be randomly shut down. In engineering, a random number method is used to determine whether a single component is faulty under the current operating conditions based on the component failure rate. A decimal number is randomly generated in the range of 0 to 1, and this value is compared with the component failure probability to determine the component's operating status.

[0089] The core judgment rule for generating faulty grid structures is to divide the operating conditions of components into two categories: available or faulty, by comparing the magnitude of random values ​​and the inherent failure rate of components.

[0090]

[0091] in Representative components Randomly generated sample values ​​within the (0,1) interval; Representative components The average annual failure probability obtained from long-term statistics; This indicates that the components are intact and functioning normally when connected to the power grid. This indicates that a component has failed and has been disconnected from the power grid. Special record of the A sample of conventional units The operational status, Record number Transmission branch under each sample The operational status.

[0092] After completing the component fault sampling, for the load side, the real-time power load of each bus is randomly generated according to the historical load fluctuation range; for the wind power side, the available active power of a single wind turbine is randomly generated according to the wind power output probability distribution function. Finally, all information on component status, bus load, and wind power output is summarized to form a complete system operating condition sample.

[0093] The unit status, branch status, bus load, and available turbine output obtained from the previous step-by-step sampling are all integrated and stored in a standardized aggregate format, containing all data from a single sample.

[0094]

[0095] in For the first In each sample, the busbar The real-time load size; For the first Wind turbine units in a sample The available active power output.

[0096] Offline Training S2: Optimal Load Reduction Model Construction and Label Generation

[0097] For each set of generated system operating condition samples, an optimal load shedding optimization mathematical model is constructed. The physical meaning of the model is to find the combination of unit output, wind curtailment, and bus load shedding that minimizes the total load shedding value of the entire network, while satisfying all grid operation constraints. Optional alternatives: The basic optimization constraint model uses DC power flow, which can be replaced with two types of convex optimization models: linearized AC power flow and decoupled linear AC power flow. After replacement, the optimization objective remains unchanged; only the power flow constraint equations are rewritten according to the model type, and the corresponding KKT constraint expressions are adaptively adjusted synchronously.

[0098] The variables that need to be optimized in the model are divided into three categories: active power output of conventional generator sets, load shedding power of each bus, and wind curtailment power of wind farms.

[0099] The optimization objective function, which aims to minimize the cumulative load shedding value of all buses in the entire system, represents the optimization direction of the entire optimization problem. ,in The busbar to be solved during the optimization process Shear load variable.

[0100] The system-wide active power balance constraint states that the sum of the output of all generator units in the system and the output of wind power that has not been discarded equals the total electricity consumption remaining after deducting the load shedding from the total grid load. This is the most fundamental power conservation rule of the power system. ,in For conventional units Actual output; For wind turbine The amount of wind power curtailed.

[0101] The generator set output is subject to upper and lower boundary constraints. The maximum output of the unit is limited by the rated capacity of the equipment. The output of a faulty unit is directly reduced to zero. Therefore, the actual output cannot exceed the product of the available capacity and the rated capacity. ,in It is a conventional unit The rated maximum active power output specified by the manufacturer.

[0102] Due to the single busbar load shedding constraint, in this project, load shedding can only remove the existing load and cannot remove power exceeding the current power consumption of the busbar. Therefore, the upper limit of load shedding is the real-time load of the busbar. ;

[0103] Wind curtailment is a constraint; curtailment essentially means forfeiting the usable output of a wind turbine. Therefore, the amount of curtailed wind cannot exceed the current generating capacity of the turbine. ;

[0104] Based on the power transmission distribution factor, the power flow of a single branch is calculated from the net power of the bus, and the formula for calculating the power flow of a branch is established. ;

[0105] Transmission branch capacity constraints: the actual power flow of a branch cannot exceed its rated transmission limit; a faulty branch has no power flow transmission capacity. ;

[0106] in branch road The actual trend; branch road Maximum conveying capacity; For the bus under the faulty topology to the side road The power transmission distribution factor.

[0107] The power balance formula for a single busbar node sums up the output of all power sources connected to the busbar, deducts its own load and shed load, and obtains the net injected power of the busbar. ,in Representative busbar The net injected power at the node.

[0108] By optimizing the solver to solve the entire set of constraint equations, the optimal shear load value for each bus is obtained, and the summation yields the optimal total shear load for the entire system, which serves as the true value label for neural network regression.

[0109] The optimal load shedding values ​​of all buses are summed to obtain the overall optimal load shedding value of the system for a single operating condition, which is used as the baseline label for regression training. ,in After optimizing the solution, the busbar The optimal load shedding value.

[0110] The optimized solution may contain extremely small floating-point calculation errors. A non-negative threshold approaching 0 is set to filter these errors. Binary classification labels of 0 and 1 are used based on the total load shedding. ,in It is the smallest non-negative threshold used to shield numerical errors; This indicates that a load shedding fault exists in this operating condition. This indicates that the system has sufficient power supply and no load shedding.

[0111] Offline Training S3: Topology Augmentation Graph Data Build

[0112] This step is the core innovation of this invention, transforming traditional flat power grid data into topological graph data adapted to graph neural networks. The graph data consists of two parts: a node feature matrix and a weighted adjacency matrix. Node features can be optionally replaced: in addition to the four basic centrality coefficients, one or more of the following can be added: feature vector centrality, PageRank, electrical betweenness, current betweenness, voltage sensitivity, and power flow sensitivity. The added features are also calculated based on the fault topology and electrical parameters and then incorporated into the feature vector.

[0113] The feature vector of a single bus node simultaneously integrates real-time electrical parameters and four types of topological features derived from the faulty network structure.

[0114] The single-node load, aggregated power output, and four types of topology center coefficients are arranged in sequence and spliced ​​together to form a complete feature vector of the node. All features together constitute a single row of data in the node feature matrix.

[0115]

[0116] in busbar The total available maximum power of the connected units; busbar The power output of the collector fan is available; In order, they are degree, clustering, betweenness, and improved proximity center coefficient.

[0117] The degree center is used to quantify the proportion of the number of directly connected branches of a single node in the entire network, and intuitively reflects the density of the local network structure connection of the node.

[0118]

[0119] in This is the total number of system buses; For nodes in the faulty network structure The number of directly connected branches.

[0120] The clustering coefficient measures the density of connections between neighboring nodes, reflecting the clustering characteristics of the local network around the node. If a node has fewer than two connected branches, it cannot form a closed loop, and the coefficient is set to zero. ,in For nodes The number of branches connecting all adjacent vertices.

[0121] The Betweenness Center (TNC) measures the percentage of nodes that fall on the shortest transmission path between any two points in the entire network, quantifying the node's pivotal criticality within the network's transmission channels. ;in For nodes ,node The total number of all shortest paths between them; for arrive Nodes along the path The number of paths.

[0122] The improved proximity center characterizes the connectivity between a node and the rest of the network based on the shortest electrical distance between nodes. This factor is not included in the summation when there is no connecting path between isolated nodes. ;

[0123] By accumulating the branch reactances, the path with the shortest electrical length among all feasible paths between two points is found, thus obtaining the shortest electrical distance. ,in branch road Absolute value of reactance; Representative node , The set of all feasible transmission paths between them.

[0124] Adjacency matrix can be replaced: In addition to reactance, adjacency weights can also be calculated using single or multiple parameters such as impedance, admittance, line capacity, power flow sensitivity, and branch failure rate.

[0125] The coupling weight between nodes is calculated based on the branch's on / off state and selected electrical parameters. When a branch is disconnected due to a fault, there is no electrical connection between the two points, and the weight is directly assigned to 0. in A very small constant is set to prevent the denominator from being zero; This indicates that the branch has exited due to a fault, and the corresponding coupling weight is reduced to zero.

[0126] Offline training S4: Training a graph-level GCN classification model, such as... Figure 2 , Figure 3 As shown;

[0127] The overall structure of the classification network is fixed as two layers of symmetric normalized GCN, global average pooling, and three fully connected layers connected in sequence. Optional replacement options: The basic GCN can replace GraphSAGE, GAT, GIN, ChebNet, message passing network, edge-conditional GCN, spatiotemporal graph, and attention graph network. After replacement, the graph input remains unchanged, only the feature aggregation rules change, and the output remains the load shedding probability.

[0128] The original adjacency matrix is ​​superimposed with an identity matrix to supplement node self-loops, and then a normalized degree matrix is ​​constructed to perform convolution operations.

[0129] We employ a convolution calculation formula that combines self-loop enhancement with symmetric normalization to balance the feature weights of high-degree and low-degree nodes, thus preventing the features of highly connected nodes from excessively dominating the extraction results.

[0130]

[0131] in , They are identity matrices of the same order; yes The corresponding degree matrix; It is the first Layer input feature matrix; These are the learnable weights and biases of the convolutional layer, respectively. It is the ReLU nonlinear activation function.

[0132] After two convolutional operations, global mean compression is performed on all node features to condense the massive node information of the entire graph into a single vector, thereby achieving global feature representation of the entire power grid. ,in The node after two layers of convolution operations Feature data; This is the graph embedding feature corresponding to the entire power grid.

[0133] The three fully connected layers perform non-linear transformations on the graph embedding features layer by layer, continuously mapping the feature dimensions. The final layer incorporates a sigmoid function to output the probability. ,in These are the feature vectors before and after the fully connected layer; These are the weights and bias parameters of the fully connected layer, and the Sigmoid output at the end of the network. That is, the sample load shedding probability.

[0134] Threshold Optional Replacement: Classification Threshold The constant is not fixed and can be adaptively optimized based on recall rate, false positive rate, indicator error, missed cut cost, and verification comprehensive indicator.

[0135] We employ weighted binary cross-entropy to construct a classification loss, addressing the sample imbalance problem caused by the number of unloaded samples far exceeding the number of faulty samples. We then use this loss to inversely update all network parameters. ,in The total number of all samples participating in the training; The category balancing weight is obtained by dividing the number of unloaded samples by the number of faulty samples.

[0136] Offline Training S5: Node-level GCN Regression Model Training

[0137] The regression model is configured with a four-layer shared convolutional structure and a multi-head output structure. Multiple task heads simultaneously output load shedding, unit output, wind curtailment, and the dual variables required by KKT. KKT losses can be replaced: the original four-term composite loss can be replaced with augmented Lagrange, feasibility projection, exterior point penalty, original dual residual, and constraint violation loss. These can be selected individually or in a weighted combination. All replacement terms are constructed based on the physical constraints of optimal load reduction.

[0138] To facilitate a unified construction of the KKT loss, the original multi-constraint optimization model is reorganized into a standard convex optimization mathematical form. ,in A vector consisting of all optimization variables; Corresponding to the system power balance equation constraints, Corresponding to upper and lower bound inequality constraints for various types of equipment; This is the coefficient vector of the objective function.

[0139] The overall regression loss consists of a weighted sum of the mean square error of the load shedding prediction fit and the KKT physical constraint loss, with weighting coefficients... Used to adjust the proportion of constraint loss in total loss. ;

[0140] Mean squared error (MSE) measures the numerical difference between the total load predicted by the model and the optimized true label; a smaller MSE indicates higher fitting accuracy. ,in The number of effective samples used in regression training; For the model-predicted bus Shear load value.

[0141] The KKT loss is obtained by adding four types of residuals: primal feasibility, dual feasibility, complementary relaxation, and stationary point condition. These four terms correspond to the four optimal criteria for convex optimization. ;

[0142] The original feasibility loss is determined by ensuring that the output variables of the constrained model satisfy the boundary conditions of equality and inequality equations. Values ​​exceeding these boundaries will result in residuals that are included in the loss. ;

[0143] Dual feasibility loss is determined by the constraint that the dual variable corresponding to the inequality cannot be negative; negative values ​​constitute a penalty loss. ;

[0144] Complementary relaxation loss occurs when the product of the inequality constraints and the corresponding dual variables at the optimal solution approaches zero; values ​​deviating from zero constitute the loss. ;

[0145] Where ⊙ represents Hadama's element-wise multiplication operation; These are the original variables, the dual variables of equality, and the dual variables of inequality predicted by the model.

[0146] Stationary condition loss satisfies the balance between the objective gradient and the constraint gradient; deviation from the optimal gradient results in loss. ,in These are the Jacobian matrices corresponding to equality constraints and inequality constraints, respectively.

[0147] By traversing different The final penalty system is determined by selecting the optimal value from the average constraint of the validation set. ,in The total number of samples in the validation set; The comprehensive violation index is the result of normalizing all constraints for a single validation sample.

[0148] Online assessment follows the S3 rule to convert the on-site data into topology data, which is then fed into the trained classification network to determine the operating condition type. Indicator expansion and replacement: In addition to LOLP and EDNS, EENS, LOLE, LOLF, load distribution, branch overload probability, and multi-dimensional risk indicators can be calculated, all based on predicted data statistics.

[0149] By comparing the model output load shedding probability with a preset threshold, a 0-1 format operating condition discrimination label is generated. ,in Predict the load shedding probability for online samples; Enable regression calculation; otherwise, the system load will be directly assigned a value of 0.

[0150] By summarizing the predicted load shedding for each bus based on only the load shedding samples, the total predicted power shortage load for a single-sample system can be obtained. ;

[0151] By traversing all online samples and conducting statistical analysis, the probability of insufficient power (LOLP) is calculated, representing the average probability of load shedding under random operating conditions. ;

[0152] By traversing all online sample statistics, the expected power shortage (EDNS) is calculated, representing the average power shortage scale of the system. ,in To assess the total number of samples to be tested online.

[0153] This embodiment fully restores the electrical and topological characteristics of the faulty network structure through topology enhancement graph data. It eliminates a large number of invalid optimization solutions by using classification pre-screening, and accurately extracts power grid characteristics by using hierarchical graph network. In the regression stage, it corrects the prediction results through multiple types of KKT physical constraints, effectively avoiding the problem of outputs that violate the power operation law. While ensuring the accuracy of reliability index calculation, it reduces computing power overhead. It also broadens the application scenarios through multi-dimensional equivalent replacement, and can flexibly adapt to the rapid assessment of power grid reliability for various power supply structures.

[0154] Example 2 describes in detail a power system evaluation system based on topology-enhanced GCN classification and regression collaboration, such as... Figure 4 As shown, it specifically includes:

[0155] The data acquisition module is used to acquire bus, branch, conventional generator, wind turbine, load, line parameters, component failure rate and historical operating data, and to process the data format and units in a unified manner.

[0156] The state sampling module is used to generate system state samples based on component failure rate, load fluctuation and available wind power output, and form sample variables such as conventional generator available state, branch available state, node load and available wind power output.

[0157] The label generation module is used to build and solve the optimal load shedding model, generate classification labels indicating whether load shedding has occurred, and regression labels for the amount of load shedding at nodes.

[0158] The graph data construction module is used to generate node feature matrices and weighted adjacency matrices based on node electrical characteristics, node topological centrality characteristics, and line parameters, so that the samples can express post-fault topology, electrical coupling, and node importance information in graph structure form.

[0159] The classification model training module is used to train the graph-level GCN classification model and output the judgment result of whether the state to be evaluated has experienced load shedding.

[0160] The regression model training module is used to train a node-level GCN regression model with KKT joint loss and outputs the node load shedding, conventional generator output, wind curtailment, and auxiliary variables required for constructing physical constraint residuals.

[0161] The online evaluation module is used to call the trained classification and regression models to quickly predict the load shedding amount of the samples to be evaluated, and calculate the reliability index based on the prediction results.

[0162] The system can be deployed on computer equipment, servers, power dispatching auxiliary analysis platforms, or cloud computing platforms. This invention can also be implemented as an electronic device, comprising a processor and a memory, wherein the memory stores a computer program, which, when executed by the processor, implements the above-described method steps.

[0163] The present invention can also be implemented as a computer-readable storage medium or a computer program product, wherein the computer program stored thereon or carried thereon is executed by a processor to implement the above method steps.

[0164] The system in this embodiment completes data collection, sample generation, model labeling, graph data construction and model training in modules. It optimizes prediction accuracy based on topological information and physical constraints, reduces invalid calculations through classification and screening, eliminates the need for repeated solution optimization of the model, and can efficiently calculate multiple reliability indicators to meet the rapid reliability assessment needs of different power grid operating conditions.

[0165] Example 3 uses the RTS-79 system. This system comprises 24 bus nodes, 32 conventional generators, and 33 branches, with a total installed capacity of 3405 MW and a peak load of 2850 MW. Wind turbines with a capacity of 150 MW are connected to buses 1, 2, 6, 7, and 13. To simulate different wind power penetration rates, some conventional generators are replaced with wind turbines of equal capacity while introducing wind power, keeping the total installed capacity of the system constant. This example is only used to illustrate the implementation and technical effects of the invention and does not constitute a limitation on the system scale, number of nodes, number of branches, or wind power connection location.

[0166] In this embodiment, the load level can be sampled based on a 5% fluctuation range above and below the peak load, and the available wind power output can be sampled using a probability distribution with the predicted output as the mean and a preset variance parameter to determine the fluctuation range. Python is used to implement graph neural network training and inference, and Matlab is used to call the Gurobi optimization solver to solve the optimal load reduction model and generate training labels. The dataset contains 100,000 samples, with 20% used as the test set and 10% as the validation set. Training can use the Adam optimizer with 400 training epochs, a learning rate of 0.0005, and a batch size of 128. In actual engineering, the above training parameters can be adjusted according to the sample size, system size, and validation set performance. The following methods were compared in testing: the traditional solution method (using the Gurobi solver, i.e., the baseline method (M1), the method based on a CNN classification-regression model (M2), the method based on a GCN classification-regression model (M3), and the method using a GCN classification-regression model with KKT joint loss (M4). The reliability evaluation results of the four methods are shown in Table 1.

[0167] Table 1 Reliability Assessment Results

[0168] M1: Traditional optimization method (benchmark) 0.0731 - 13.67 - 582.68 M2: CNN Classification-Regression 0.0743 1.64% 15.13 10.60% 17.01 M3: GCN Classification - Regression 0.0722 1.23% 13.32 2.56% 24.62 M4: Method of the present invention 0.0722 1.23% 13.51 1.17% 25.43

[0169] As can be seen from the example results, compared with traditional optimization methods, the computation time of the method of the present invention is reduced from 582.68 s to 25.43 s, and the efficiency is improved by approximately 22.9 times; compared with the CNN classification-regression method, the relative error of EDNS is reduced from 10.60% to 1.17%. These results indicate that topology enhancement graph structure data is beneficial for capturing grid topology changes caused by component faults, and KKT joint loss helps reduce physical constraint violations and improve the accuracy of load shedding prediction. The above numerical results are used to illustrate the effects of the present invention in specific embodiments and do not limit the scope of protection of the present invention.

[0170] This embodiment has been verified by actual testing. Compared with the traditional sample-by-sample optimization algorithm, this solution significantly reduces the simulation calculation time, reduces the load shedding prediction deviation based on topological characteristics and KKT physical constraints, significantly improves the accuracy of reliability index calculation, and allows for the smooth implementation of various equivalent replacement structures. It can meet the engineering requirements for online reliability calculation of new energy grid-connected power grids.

[0171] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A power system evaluation method based on topology-enhanced GCN classification and regression synergy, characterized in that, Includes an offline training phase and an online evaluation phase; The offline training phase includes: S1. Collect basic parameters of the power system and generate a system status sample containing component operating status, load data, and wind power output data through Monte Carlo sampling. ; S2, for each system state sample Construct and solve the DC optimal load shedding model, and generate load shedding classification labels based on the solution results. Regression labels with shear load ; S3. Constructing topology enhancement graph data based on system failure topology. The The node feature matrix, Weighted adjacency matrix for branch parameters; S4. Utilizing graph data A graph-level GCN classification model is trained, which is processed sequentially through two graph convolutional layers, a global average pooling layer, and three fully connected layers to output the probability of load shedding. S5. Utilizing graph data A node-level GCN regression model is trained, which is configured with a multi-task output head. The training uses a joint loss function that integrates MSE loss and KKT-constrained residual loss. ; The online assessment phase includes: S6. For the power system state samples to be evaluated, construct the topology augmentation graph data as in step S3, input it into the trained GCN classification model, and distinguish whether the samples have experienced load shedding according to the preset classification threshold. S7. Input only the samples that are determined to have experienced load shedding into the trained node-level GCN regression model. The regression model outputs the load shedding amount of each bus node. Based on the discrimination results of all samples and the predicted values ​​of load shedding, the power system reliability indices LOLP and EDNS are statistically calculated.

2. The power system evaluation method based on topology-enhanced GCN classification and regression synergy as described in claim 1, characterized in that, The total system load shedding is obtained by summing the optimal load shedding of all buses in S2 from a single sample. The formula is as follows: In the formula: For the first The system's overall optimal load shedding amount corresponding to each sample; It is the set consisting of all bus nodes of the power grid; For the first Busbar under each sample The optimal load shedding amount is obtained through optimization. Set non-negative threshold The following formula is used to classify 0-1 binary labels: In the formula: For the first The cut load classification label for each sample; This is a preset non-negative threshold used to eliminate numerical calculation errors; a label value of 1 indicates that the corresponding system sample has a shear load, and a label value of 0 indicates that the corresponding system sample has no shear load.

3. The power system evaluation method based on topology-enhanced GCN classification and regression synergy as described in claim 2, characterized in that, S3 constructs the node feature matrix At that time, single busbar node Feature vectors: ; In the formula: For the first Nodes under sample topology eigenvectors; For nodes The original load; For nodes The maximum available output when connected to a conventional generating unit; For nodes Available output from connected wind turbine units; , , , The nodes in the fault topology are listed in order. The corresponding degree centrality, clustering coefficient, betweenness centrality, and improved proximity centrality are calculated. The four types of centrality values ​​are solved according to the post-fault topology, and the node feature matrix is ​​formed by concatenating all node feature vectors.

4. The power system evaluation method based on topology-enhanced GCN classification and regression synergy as described in claim 3, characterized in that, S3 constructs a weighted adjacency matrix. Formula for calculating matrix elements: In the formula: For the first Nodes in the adjacency matrix of samples With nodes Weighting coefficients between them; It is the set of all physical branches of the system; branch road The available status indicator, with a value of 1 representing that the branch is intact and available, and a value of 0 representing that the branch is faulty and out of service; branch road The absolute value of reactance; To avoid extremely small positive numbers with a denominator of zero; The node feature matrix and the weighted adjacency matrix are combined to form single-sample topology enhancement graph data.

5. The power system evaluation method based on topology-enhanced GCN classification and regression synergy according to claim 4, characterized in that, S4 single-layer graph convolution operation formula: ; In the formula: , The first Layer convolution input, the first Convolutional layers output feature matrices; , They are identity matrices of the same order; For matrix The corresponding degree matrix; , The first Convolutional layers can be trained to generate weight matrices and bias vectors; It is a ReLU nonlinear activation function; The formula for global average pooling after two convolutional layers is as follows: ; In the formula: This refers to the graph-level embedding features corresponding to the entire power grid diagram; This represents the total number of nodes in the power grid. The node after two layers of convolution operations Feature data; Formula for iterative calculation of a three-layer fully connected layer: ; In the formula: , These are the feature vectors of the preceding and following layers of the fully connected layer; , These are the weight and bias parameters for the fully connected layer; the Sigmoid function is used at the end of the fully connected layer to output the probability of load shedding.

6. The power system evaluation method based on topology-enhanced GCN classification and regression synergy according to claim 5, characterized in that, The classification model is trained using a binary cross-entropy loss function with class weights. The loss calculation formula is as follows: In the formula: This represents the overall loss value of the classification model. The total number of samples used in model training; These are the class weight coefficients used to balance sample imbalance; The first output of the model The predicted probability of a sample experiencing load shedding is used to update all network parameters of the classification model in reverse.

7. The power system evaluation method based on topology-enhanced GCN classification and regression synergy as described in claim 6, characterized in that, The S5 regression model is set up with a multi-task output head to synchronously output node load shedding, unit output, wind curtailment, and optimization dual variables. ; In the formula: This represents the total loss of the regression model; The mean squared error loss between the predicted load shedding value and the actual label; The weighted penalty coefficient for the KKT loss term; This is a composite constraint loss constructed based on the optimal model KKT conditions; Mean squared error loss: In the formula: The number of effective samples used in training the regression model; For the model to predict the first Nodes under each sample Shear load; For the first The total load of each sample system is the true label; In the formula: For the first The original feasibility loss corresponding to each sample; For the first The dual feasibility loss corresponding to each sample; For the first The complementary relaxation loss corresponding to each sample; For the first The stationary condition loss corresponds to each sample; the four sub-losses are constructed according to the KKT optimal conditions corresponding to the optimal load reduction model, without separately limiting the expressions of the four sub-losses.

8. The power system evaluation method based on topology-enhanced GCN classification and regression synergy according to claim 7, characterized in that, Penalty parameters The DCV calculation formula is determined based on the average normalized constraint violation rate of the validation set: ; In the formula: To verify the set average normalization constraint violation rate; The total number of samples in the validation set; For the first The normalization constraint of the validation samples violated the index; multiple candidate sets were traversed. The value corresponding to the minimum DCV value is selected as the final penalty parameter.

9. The power system evaluation method based on topology-enhanced GCN classification and regression synergy according to claim 8, characterized in that, During the online evaluation phase, after the samples to be tested are classified and screened, regression prediction is only enabled for samples with unloaded loads. The reliability index is calculated using the following formula: In the formula: This represents the probability of insufficient power. For the expected power shortage; To assess the total number of samples to be tested online; For the first Binary discriminant for each sample shear load; For the first The total system load shedding is predicted from a sample.

10. A power system evaluation system based on topology-enhanced GCN classification and regression synergy, characterized in that, include: The data acquisition module collects various basic power data, including busbars, branches, conventional generator sets, wind turbines, loads, component failure rates, and line reactance, and standardizes all data dimensions and storage formats. The status sampling module uses Monte Carlo sampling to generate system status samples containing component commissioning status, node load, and available wind power output based on component failure probability, load fluctuation range, and wind power output probability distribution. The label generation module builds a DC optimal load reduction model for each system state sample and completes the solution calculation. Based on the optimal solution, it calculates the total load shedding amount and the binary load shedding classification label. The graph data construction module, based on the post-fault power grid topology, calculates the node electrical characteristics and four types of topological centrality parameters to form a node feature vector and concatenates them to obtain a node feature matrix. It then combines the branch operating parameters and reactance values ​​to calculate a weighted adjacency matrix and generates topology-enhanced graph data. The classification model training module constructs a graph-level classification network consisting of two layers of symmetric normalized graph convolutional layers, a global average pooling layer, and three fully connected layers connected in sequence. Iterative optimization of all network parameters is achieved using a binary cross-entropy loss function with class weights. The regression model training module builds a node-level regression network with four layers of shared graph convolutional layers and multi-branch multi-task output heads. The network training adopts a composite loss function consisting of a mean squared error term and four KKT constraint loss terms, and the loss weighting coefficients are determined based on the constraint violation index of the validation set. The online evaluation and calculation module receives the power grid data to be evaluated and converts it into topology enhancement graph data. It then calls the trained classification network and selectively enables the regression network to determine the load shedding status and predict the load shedding amount. Based on the prediction statistics, it calculates two reliability indicators, LOLP and EDNS.

Images