Fast evaluation method and system for high-proportion new energy access sending-end power grid

Abstract

The present application belongs to the technical field of power system safety and stability analysis, and discloses a kind of fast evaluation method and system for high proportion new energy access sending terminal power grid.The method obtains operating mode data, power grid topology structure, equipment parameters and fault parameters, calculates physical baseline evaluation result according to corresponding simplified physical model calling stability index to be evaluated, and constructs the above data into graph structure data containing power grid operating state information and fault scene information, inputs corresponding pre-training graph neural network model, obtains prediction residual value for physical baseline evaluation result, and adds the two to obtain second stability evaluation result.For critical clearance time evaluation, flow consistency check can also be performed on the decoded pre-fault steady-state physical state, and the physical baseline evaluation result is used when the check fails.This scheme realizes residual correction for fault scenarios while preserving the physical baseline, and reduces the risk of abnormal model representation participating in result correction.

Description

Technical Field

[0001] This invention belongs to the field of power system security and stability analysis technology, specifically relating to a rapid assessment method and system for high-proportion renewable energy access to the sending-end power grid. Background Technology

[0002] As the proportion of new energy sources such as wind power and photovoltaics connected to the sending-end power grid increases, the number of power electronic interface devices in the power grid also increases. System inertia, reactive power support characteristics, and dynamic response characteristics after faults change with operating modes. Under different combinations of new energy output, load levels, power grid topology, and fault scenarios, it is necessary to evaluate stability indicators such as transient power angle stability, frequency stability, and voltage stability. With the increased number of combinations of operating modes and fault conditions, stability assessments need to handle more scenarios simultaneously, placing demands on the computation time of a single assessment and the consistency of assessment results across different scenarios.

[0003] High-precision stability analysis typically requires solving the dynamic processes of the power grid for specific operating modes and fault scenarios, and its calculation process involves many equipment states and dynamic equations. When it is necessary to continuously evaluate multiple operating modes, fault locations, and fault types, high-precision calculations need to be repeatedly performed for different scenarios. The computational resources and processing time increase with the number of scenarios to be evaluated, which limits the application of this type of method when a large number of operating scenarios need to be quickly screened.

[0004] To reduce computational load, simplified physical models corresponding to the respective stability indices can be used to calculate the stability indices. These simplified physical models are calculated based on some key grid conditions and equipment parameters. The model assumptions and state simplifications cause discrepancies between the calculated stability indices and the results of high-precision stability analysis. Furthermore, these discrepancies vary with operating modes, grid topology, equipment parameters, and fault types and locations. Relying solely on fixed error correction parameters makes it difficult to characterize the changing relationships of these discrepancies under different operating scenarios.

[0005] When using data-driven models to predict stability indices, the model input needs to reflect the operating states, network connectivity, and fault conditions that affect the stability results. For bus faults, line faults, and faults at different locations within lines in the sending-end power grid, if the correspondence between fault type, fault location, and fault components and the power grid structure is not included in the model input, it will be difficult to form distinguishable model inputs for different fault scenarios under the same operating mode. For stability indices such as critical clearing time, which are related to specific fault scenarios, this issue will affect the model's representation of differences in fault scenarios.

[0006] Furthermore, the data-driven model forms internal hidden state features based on training samples, and its prediction results depend on the extent to which the training samples cover the current operating mode, topology, and fault scenarios. When the internal state representation formed by the model in the current scenario to be evaluated deviates from the basic physical relationship of the power grid, if there is a lack of verification and result selection mechanisms corresponding to the physical equations of the power grid, the correction amount generated by the model may still participate in the calculation of the final stability assessment result.

[0007] Therefore, based on the rapid calculation of simplified physical models, it is necessary to establish corresponding error learning mechanisms for different stability indices, so that the data-driven model can learn the residuals of the simplified physical model relative to the high-precision stability analysis for the same operating mode, power grid topology, equipment parameters and fault scenarios. At the same time, it is necessary to establish a correspondence between fault scenarios and graph structure data, and provide corresponding physical consistency check methods for the physical state representation formed by the model under the current scenario to be evaluated. Summary of the Invention

[0008] Purpose of the invention: The purpose of this invention is to address the shortcomings of existing technologies by providing a rapid evaluation method and system for high-proportion renewable energy access to the power grid. This method establishes a simplified physical model and a residual prediction graph neural network for the stability index to be evaluated. The stability evaluation result obtained from the simplified physical model is used as the physical baseline evaluation result. The graph neural network is then used to predict the residual of the physical baseline evaluation result relative to the high-precision stability evaluation result, thereby correcting the evaluation deviation while maintaining the physical baseline.

[0009] Technical Solution: The primary objective of this invention is to provide a rapid evaluation method for high-proportion renewable energy grid connections, comprising: acquiring operational mode data, grid topology, equipment parameters, and fault parameters of the grid to be evaluated, and determining a stability index to be evaluated; determining a simplified physical model corresponding to the stability index to be evaluated; calculating a first stability evaluation result using the simplified physical model based on the operational mode data, grid topology, equipment parameters, and fault parameters, and using the first stability evaluation result as a physical baseline evaluation result; determining a pre-trained graph neural network model corresponding to the stability index to be evaluated; constructing the operational mode data, grid topology, equipment parameters, and fault parameters into graph structure data containing grid operational status information and fault scenario information; inputting the graph structure data into the graph neural network model to obtain a prediction residual value for the physical baseline evaluation result; and adding the physical baseline evaluation result to the prediction residual value to obtain a second stability evaluation result.

[0010] Another aspect of the present invention provides a rapid evaluation system for high-proportion renewable energy grid integration, including a physical model calculation module, a graph neural network inference module, and a result fusion module;

[0011] The physical model calculation module is used to determine the simplified physical model corresponding to the stability index to be evaluated based on the stability index to be evaluated, and to calculate the first stability evaluation result based on the operation mode data, grid topology, equipment parameters and fault parameters of the power grid to be evaluated, and to use the first stability evaluation result as the physical baseline evaluation result.

[0012] The graph neural network inference module is used to determine a pre-trained graph neural network model corresponding to the stability index to be evaluated based on the stability index to be evaluated, construct the operation mode data, power grid topology, equipment parameters and fault parameters into graph structure data containing power grid operation status information and fault scenario information, and input the graph structure data into the graph neural network model to obtain the predicted residual value for the physical baseline evaluation result;

[0013] The result fusion module is used to add the physical baseline evaluation result to the predicted residual value to obtain the second stability evaluation result.

[0014] Beneficial effects: This invention forms a physical baseline assessment result using a simplified physical model corresponding to the stability index to be evaluated, and uses graph structure data containing power grid operating status information and fault scenario information to predict the residual of the physical baseline assessment result, so that rapid physical calculation, scenario difference characterization and residual correction can be performed in a coordinated manner under the same stability index.

[0015] (1) Reduce the mapping complexity required for graph neural networks to directly learn the absolute values ​​of stability indices. This invention determines a simplified physical model based on the stability index to be evaluated. The simplified physical model obtains the physical baseline evaluation result based on the operating mode data, power grid topology, equipment parameters, and fault parameters. Then, the graph neural network model outputs the predicted residual value for the physical baseline evaluation result. The object learned by the graph neural network changes from the absolute value of the stability index to the deviation between the high-precision stability analysis result and the physical baseline evaluation result. Finally, the predicted residual value is applied to the physical baseline evaluation result through addition. This allows the simplified physical model to handle the basic stability calculation and the graph neural network to handle the scenario-related deviation correction, thereby reducing the situation where the stability evaluation result is directly generated by the data-driven model.

[0016] (2) Enables different fault scenarios to form corresponding input differences in graph structure data. Fault scenario information includes at least one of the following: fault type, fault location, fault impedance, fault component identification, and post-fault topology state corresponding to the preset fault clearing scheme. Bus faults are written into the fault characteristics of the node corresponding to the fault bus, line faults are written into the fault characteristics of the edge corresponding to the fault line, and internal line faults can also be represented by dividing the fault line and setting virtual fault nodes to indicate the specific fault location. Thus, physical baseline assessment and graph neural network residual prediction can be performed for the same operating mode, network structure, and fault scenario, avoiding the problem that different fault locations are difficult to distinguish by the model when they only have the same static power grid graph input; for the CCT residual prediction model, the fault clearing time or CCT value to be predicted is excluded from the graph structure input, and the information source used for residual prediction is kept separate from the label to be predicted.

[0017] (3) Establish corresponding residual mapping relationships for stability indices with different dimensions and physical meanings. Training sample sets are established for different types of stability indices to be evaluated, and corresponding graph neural network models are trained. The difference between the true stability evaluation value obtained from the corresponding high-precision stability analysis and the physical baseline evaluation result obtained from the corresponding simplified physical model is used as the true residual value. For CCT, the lowest frequency point, or other corresponding stability indices, each model processes its corresponding physical baseline and residual label separately. Therefore, during model training, it is not necessary to mix residuals of different dimensions in the same output target, and the model can select the corresponding residual prediction relationship based on the stability index to be evaluated when calling the model.

[0018] (4) Applying the physical relationships of the power grid to the training process of the residual model sharing hidden state features. The graph neural network model forms node hidden state features through the graph attention network backbone network. The graph-level aggregation layer and residual prediction layer output predicted residual values ​​based on the node hidden state features, and the physical quantity decoder outputs physical state variables based on at least some of the node hidden state features; during training, according to Simultaneously, data loss and physical loss determined by the residuals of the AC power flow equations are introduced. The prediction residual branch and the physical quantity decoding branch are therefore affected by the same shared hidden state features. The AC power flow relationship can participate in the model parameter update through the physical loss. The internal feature representation used to form the prediction residual value is simultaneously constrained by the training sample error and the physical relationship of the power grid.

[0019] (5) Provide a model representation credibility screening for the current scenario to be evaluated for CCT residual correction. When the stability index to be evaluated is CCT, the graph neural network model generates CCT prediction residual values ​​and node voltage amplitudes and node phase angles corresponding to the steady-state operation state before the fault based on the shared hidden state features of the same scenario to be evaluated. Combined with the node active power injection, node reactive power injection, grid topology and equipment parameters before the fault, the power flow consistency residual of the operating state before the fault is determined. When the power flow consistency residual meets the preset conditions, the second stability evaluation result enters the output stage; when the power flow consistency residual does not meet the preset conditions, the prediction residual value is discarded and the physical baseline evaluation result is adopted. When the physical state representation inside the model is abnormal, the corresponding CCT residual correction can be blocked before the final output is formed.

[0020] (6) Reduce the redundant investment of high-precision stability analysis resources in samples with low information content. During the training of the graph neural network model, the scene to be labeled is determined based on at least one of the following: model prediction variance, the divergence of prediction results among multiple independently trained graph neural network models, the proximity of the prediction results to the stability boundary, and the degree of difference between the unlabeled scene and the existing training samples. Then, high-precision stability analysis is performed on the scene to be labeled and added to the training sample set. The addition of training samples can therefore focus on scenes that the model currently finds difficult to judge, are close to the stability boundary, or have large differences from existing samples, so that the newly added high-precision analysis results can be directly entered into the subsequent residual model training without applying the same labeling priority to the unlabeled scene.

[0021] (7) A continuous evaluation link is formed from physical baseline calculation, fault scenario residual correction to anomaly correction blocking. A simplified physical model alone can obtain the stability index baseline, but it is difficult to correct its model bias for different operating modes and fault scenarios; a residual graph neural network alone can learn scenario-related biases, but its output still depends on the internal feature representation formed by the model for the current scenario; a pre-fault steady-state power flow check alone can determine the consistency between the decoded physical state and the basic power flow relationship, but it cannot generate the deviation correction amount of the stability index. In this invention, a physical baseline is first formed by a simplified physical model corresponding to the stability index to be evaluated, and then the power grid operating state information and fault scenario information of the same scenario are input into the corresponding graph neural network to generate the prediction residual. At the same time, physical state variables for power flow consistency check are obtained from the shared hidden state features. The check result further determines whether the prediction residual value participates in the final result output, thereby forming a fast stability evaluation link that connects physical baseline calculation, residual correction and correction result selection. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall technical architecture of the rapid evaluation system according to an embodiment of the present invention.

[0023] Figure 2 This is a schematic diagram of the offline residual graph neural network training process according to an embodiment of the present invention.

[0024] Figure 3 This is a schematic diagram of the training structure and loss calculation of the physical knowledge embedding graph neural network according to an embodiment of the present invention.

[0025] Figure 4 This is a schematic diagram of the online rapid evaluation and pre-fault steady-state power flow consistency check process according to an embodiment of the present invention. Detailed Implementation

[0026] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the embodiments described.

[0027] This invention addresses the problem of rapid stability assessment for grids with a high proportion of renewable energy connected to the sending-end. It calculates the physical baseline assessment result using a simplified physical model corresponding to the stability index to be assessed, and then uses a pre-trained GNN model 150 for the stability index to predict the residual between the physical baseline assessment result and the high-precision stability assessment result. A second stability assessment result is formed by adding the predicted residual value to the physical baseline assessment result. For the assessment of the critical clearing time (CCT), the shared hidden state features formed by the GNN model 150 can be used to decode the physical state variables corresponding to the steady-state operating state before the fault occurs, and a steady-state power flow consistency check before the fault is performed to determine whether the predicted residual value should be used for correction.

[0028] Example 1: Refer to Figure 1 This embodiment provides a rapid evaluation system for high-proportion renewable energy access to the power grid, the system including an offline training module 100 and an online evaluation module 200.

[0029] The offline training module 100 receives static power grid data 10, historical dynamic data 20, and physical equations and model rules from the physical knowledge base 30. It then constructs training data, generates true residual labels, and trains corresponding GNN models 150 for different types of stability indices. The static power grid data 10 describes the relatively fixed parameters of the power grid structure and equipment, including at least one of the following: power grid topology, line parameters, transformer parameters, generator parameters, and parameters of renewable energy access equipment. The historical dynamic data 20 describes at least one of the following under different operating modes: generator output, load power, renewable energy output, node voltage, node phase angle, and equipment operating status. The physical knowledge base 30 stores physical equations or model rules used to simplify physical model calculations and embed physical knowledge into training.

[0030] The offline training module 100 trains the corresponding residual GNNs according to the stability indices to be evaluated. For the critical clearing time (CCT), the lowest frequency point, and the static voltage stability index, training data is established using the corresponding stability evaluation ground truth, simplified physical model, and residual labels, thereby obtaining the GNN model 150 corresponding to the respective stability index.

[0031] After training, the GNN model 150 corresponding to the indicators is deployed to the online evaluation module 200. The online evaluation module 200 receives the scenario 210 to be evaluated. The scenario 210 includes the current operating mode data of the power grid to be evaluated, the power grid topology, equipment parameters, fault parameters, and the stability indicators to be evaluated. The online evaluation module 200 calls the corresponding simplified physical model and the GNN model 150 corresponding to the indicators according to the stability indicators to be evaluated, obtains the physical baseline evaluation results and prediction residual values ​​respectively, and forms a stability evaluation result based on the two.

[0032] In a further implementation, the online evaluation module 200 also performs a physical consistency check on the physical state representation formed by the GNN model 150 corresponding to the index for the current scenario to be evaluated, and determines whether to use the predicted residual value for correction based on the check result, and finally outputs the final stability evaluation result 220.

[0033] Corresponding to the aforementioned method, the offline training module 100 is used to obtain the pre-trained index-corresponding GNN model 150 required for online evaluation; the online evaluation module 200 is used to perform data acquisition of the scenario to be evaluated, physical baseline evaluation, graph structure data construction, prediction residual calculation, result fusion, and physical consistency check and result selection in a further embodiment.

[0034] Example 2: Refer to Figure 2 This embodiment illustrates the offline training process of the GNN model 150 corresponding to the index.

[0035] Training step T100: Determine the training scenario and stability metrics

[0036] Training scenario 110 is acquired, and the stability index corresponding to the current training is determined through the stability index selection module 115. Training scenario 110 consists of power grid operation mode, power grid topology, equipment parameters, and fault parameters. Training scenario 110 can be formed based on power grid static data 10 and historical dynamic data 20, and covers different load levels, new energy output levels, system inertia, fault types, and fault locations.

[0037] Training sample sets are established for different stability indices, including: a CCT training sample set for the critical clearing time (CCT), a frequency stability training sample set for the lowest frequency point, and a corresponding static voltage stability training sample set for the static voltage stability index.

[0038] The stability index selection module 115 determines a stability analysis method, a simplified physical model, and a corresponding GNN training task that are compatible with the current training objective.

[0039] Training step T200: Generate stability evaluation ground truth and physical baseline evaluation results

[0040] High-precision stability analysis and simplified physics model calculation are performed on the same training scenario 110. The stability analysis module 120 performs stability analysis based on the training scenario 110 to obtain the stability evaluation true value 121. The simplified physics model engine 130 uses a simplified physics model corresponding to the current stability index to calculate the same training scenario 110 to obtain the physics baseline evaluation result 131.

[0041] For CCT evaluation, the stability analysis module 120 obtains the true value of CCT through high-precision time-domain stability analysis, and the simplified physical model engine 130 calculates the CCT physical baseline evaluation results using the transient energy function (TEF) model or the extended equal area criterion (EEAC) model.

[0042] For frequency stability assessment, the stability analysis module 120 obtains the true value of the lowest frequency point based on the high-precision dynamic analysis results, and the simplified physical model engine 130 uses the frequency response analysis model to obtain the physical baseline assessment result of the lowest frequency point.

[0043] For static voltage stability assessment, the stability analysis module 120 obtains the true value of the static voltage stability index through corresponding high-precision voltage stability analysis, and the simplified physical model engine 130 uses the PV curve model to calculate the physical baseline assessment results of the same type of static voltage stability index.

[0044] The stability evaluation ground value 121 and the physical baseline evaluation result 131 in the same training task above correspond to the same stability index.

[0045] Training step T300: Generate true residual labels

[0046] The residual calculation module 141 receives the stability assessment truth value 121 and the physical baseline assessment result 131, and determines the true residual label 142 according to the difference between the two.

[0047] The true residual value satisfies: ;

[0048] In the formula, This represents the true residual value. This represents the true value 121 of the stability assessment obtained by the stability analysis module 120. This represents the physics baseline evaluation result 131 obtained by the simplified physics model engine 130.

[0049] The true residual value is used as the true residual label corresponding to the current training scenario 142.

[0050] When CCT is used as the stability index, the true residual label 142 is the difference between the true value of CCT and the evaluation result of the physical baseline of CCT; when the lowest frequency point is used as the stability index, the true residual label 142 is the difference between the true value of the lowest frequency point and the evaluation result of the physical baseline of the lowest frequency point; when the static voltage stability index is used as the training target, the true residual label 142 is the difference between the true value of the corresponding static voltage stability index and the evaluation result of the same type of physical baseline.

[0051] Training step T400: Construct graph-structured data containing fault scenario information

[0052] The fault scenario graph construction module 112 constructs graph structure data 143 based on the operating mode data, power grid topology, equipment parameters, and fault parameters in training scenario 110. Nodes in graph structure data 143 represent at least one of the following: busbars, generators, loads, and renewable energy access equipment; edges represent at least one of the following: transmission lines, transformers, and switching equipment. Node features include at least one of the following: node type, voltage level, voltage amplitude, node phase angle, active power generation, reactive power generation, active load, reactive load, renewable energy active output, renewable energy reactive output, unit parameters, and equipment operating status. Edge features include at least one of the following: component type, line resistance, line reactance, line susceptance, transformer turns ratio, phase shift angle, and switching status.

[0053] For discrete categorical features, categorical encoding can be used to form the input features of a graph neural network; for continuous features, normalization or standardization can be performed based on the training data.

[0054] Fault parameters are encoded into graph structure data 143 to form different graph inputs for different fault scenarios. For bus faults, a fault identifier is written into the node features of the corresponding faulty bus, along with at least one fault information, including fault type and fault impedance. For line faults, a fault identifier is written into the edge features of the corresponding faulty line, along with at least one fault information, including fault type, fault impedance, and fault location within the line.

[0055] For a fault occurring inside the line, in one implementation, the faulty line can be divided into at least two sides according to the fault location, and a virtual fault node can be set between the at least two sides to represent the fault location inside the line through the virtual fault node.

[0056] The post-fault topology state corresponding to the preset fault clearing scheme can also be written into the graph structure data 143 as fault scenario information. For the GNN model 150 with CCT as the prediction target, the actual fault clearing time or CCT value of the object to be predicted is not written into the graph structure data 143.

[0057] Training step T500: Training the GNN model corresponding to the stability metric.

[0058] The physical knowledge embedding GNN trainer 140 receives graph structure data 143, real residual labels 142, and physical equations in the physical knowledge base 30, and trains the GNN model 150 according to the model rules corresponding to the current stability index.

[0059] During training, the GNN model 150 takes graph structure data 143 as input, uses the real residual labels 142 as residual supervision information, and updates its parameters by combining the data loss and physical loss formed by the physical equations provided by the physical knowledge base 30.

[0060] After training, a GNN model 150 corresponding to the current stability metric is obtained. The above process is repeated for different types of stability metrics to obtain the corresponding GNN model 150 for each metric.

[0061] Example 3: Physics Knowledge Embedding Training and Loss Calculation

[0062] Reference Figure 3 This embodiment further illustrates the internal processing of the physical knowledge embedded GNN trainer 140 training the GNN model 150.

[0063] The graph structure data 143 is first input into the GAT backbone network 151. The GAT backbone network 151 processes the association information between nodes and edges through multi-layer graph message passing, forming the shared hidden state features 152 corresponding to each node. The shared hidden state features 152 are then fed into the residual prediction branch and the physical quantity decoding branch, respectively.

[0064] In the residual prediction branch, the graph-level aggregation layer 153 aggregates the shared hidden state features 152 of each node to form a graph-level representation that characterizes the current complete training scene. The graph-level aggregation layer 153 can employ average aggregation, summation aggregation, or attention aggregation methods.

[0065] The residual prediction layer 154 outputs the predicted residual value 411 based on the graph-level representation formed by the graph-level aggregation layer 153. The data loss calculation module 420 receives the predicted residual value 411 and the true residual label 142, and calculates the data loss 421 based on the difference between the two.

[0066] In one implementation, data loss 421 is calculated using mean squared error:

[0067]

[0068] in, The number of training samples used in the current loss calculation. For the first The predicted residual value for each training sample is 411. For the first The true residual labels corresponding to each training sample are 142. This indicates a data loss of 421.

[0069] In the physical quantity decoding branch, the physical quantity decoder 155 receives at least partially shared hidden state features 152 and decodes them into a pre-fault physical state 412 corresponding to physical constraints. In the implementation using CCT as the stability index, the pre-fault physical state 412 includes at least the node voltage amplitude and node phase angle corresponding to the steady-state operation state before the fault occurs. Simultaneously, pre-fault operating data is extracted from the current training scenario to form pre-fault operating mode data 413, and the pre-fault grid topology and equipment parameters are extracted to form pre-fault topology / equipment parameters 414.

[0070] The AC power flow residual calculation module 430 calculates the AC power flow equation residuals based on the physical state 412 before the fault, the operating mode data 413 before the fault, and the topology / equipment parameters 414 before the fault, and determines the physical losses 431 based on the AC power flow equation residuals.

[0071] Therefore, the predicted residual value 411 and the physical state 412 before the fault both come from the shared hidden state feature 152 formed by the GAT backbone network 151, and the physical loss 431 can act on the shared feature representation that forms the predicted residual value 411 during the training process.

[0072] The data loss 421 and the physical loss 431 are weighted together to form the total loss 441:

[0073]

[0074] in, The total loss is 441. This indicates a data loss of 421. This indicates a physical loss of 431. This represents the weighting factor corresponding to the physical loss.

[0075] like Figure 3 As shown, weighting factors A dynamic adjustment method can be adopted. In the initial stage of training, [the following will be used]. Set to a relatively small value so that parameter updates are primarily based on the data loss between the true residual label 142 and the predicted residual value 411; increase according to a preset adjustment strategy as training progresses. This allows the physical loss 431 to gradually participate in the model parameter update.

[0076] The total loss 441 is input for backpropagation / parameter update 442. Backpropagation / parameter update 442 determines the joint gradient based on the total loss 441 and updates the trainable parameters of the GAT backbone network 151, graph-level aggregation layer 153, residual prediction layer 154, and physical quantity decoder 155.

[0077] After iterative training, a GNN model 150 with corresponding stability index is obtained. In the implementation of the GNN model 150 using a graph attention network, after completing online inference, the attention weights generated by the GAT backbone network 151 can be extracted to characterize the relative influence of different nodes and edges in the graph structure data 143 on the current model output, and the corresponding nodes or edges can be used as auxiliary analysis information for the evaluation results.

[0078] Example 4: Reference Figure 4 This embodiment provides a detailed explanation of the rapid evaluation method.

[0079] Step S100: Obtain the scenario 210 to be evaluated and determine the stability index to be evaluated.

[0080] Obtain the scenario 210 to be evaluated and determine the stability index to be evaluated. The scenario 210 to be evaluated includes at least the operating mode data, grid topology, equipment parameters, and fault parameters of the power grid to be evaluated, and determine the current stability index to be evaluated. The operating mode data includes at least one of the following: generator active power output, generator reactive power output, load active power, load reactive power, renewable energy active power output, renewable energy reactive power output, node voltage, node phase angle, and equipment operating status. The fault parameters include at least one of the following: fault type, fault location, fault impedance, fault component identification, and the post-fault topology state corresponding to the preset fault clearing scheme. Determine the corresponding simplified physical model and the GNN model 150 corresponding to the index based on the current stability index to be evaluated.

[0081] Step S200: Calculate the physical baseline assessment result 242.

[0082] The data corresponding to scenario 210 to be evaluated is input into the simplified physical model engine 130. The simplified physical model engine 130 adopts the corresponding simplified physical model according to the stability index to be evaluated, and calculates the first stability evaluation result based on the operating mode data, power grid topology, equipment parameters and fault parameters. The first stability evaluation result is used as the physical baseline evaluation result 242.

[0083] When using CCT as the stability metric to be evaluated, the simplified physical model engine 130 calculates the CCT physical baseline evaluation result 242 corresponding to the current fault scenario using either the TEF model or the EEAC model. In the minimum frequency point evaluation, the simplified physical model engine 130 uses a frequency response analysis model to obtain the minimum frequency point physical baseline evaluation result 242. In the static voltage stability evaluation, the simplified physical model engine 130 uses a PV curve model to obtain the corresponding static voltage stability metric physical baseline evaluation result 242.

[0084] Step S300: Construct an online fault scenario diagram and obtain the predicted residual value 241.

[0085] The online fault scenario graph builder 230 constructs graph structure data based on the operating mode data, power grid topology, equipment parameters, and fault parameters in the scenario to be evaluated 210, and inputs it into the GNN model 150 corresponding to the current stability index to be evaluated. The encoding method of the online fault scenario graph builder 230 for bus faults, line faults, and internal line faults is consistent with the encoding rules of the fault scenario graph construction module 112 in embodiment 2 during the training phase, so that the index-corresponding GNN model 150 receives graph structure features with corresponding semantics during the training and online evaluation phases. The index-corresponding GNN model 150 performs forward inference on the online graph structure data and outputs the predicted residual value 241 through the residual prediction branch. When the current stability index to be evaluated is CCT, the actual fault clearing time or the predicted CCT value is not used as the input of the GNN model 150.

[0086] In one implementation, the physical baseline calculation of the simplified physical model engine 130 is executed in parallel with the residual inference of the online fault scenario graph builder 230 and the GNN model 150, and the two calculation channels generate physical baseline evaluation results 242 and predicted residual values ​​241, respectively.

[0087] Step S400: Form the second stability assessment result 251.

[0088] The result fusion module 250 receives the predicted residual value 241 and the physical baseline evaluation result 242, and adds them together to obtain the second stability evaluation result 251.

[0089] For general stability indices, the relationship is expressed as follows:

[0090]

[0091] in, This indicates the second stability assessment result 251. This indicates the physical baseline assessment result 242. This indicates a predicted residual value of 241.

[0092] For CCT assessment, the relationship is expressed as:

[0093]

[0094] in, Indicates the evaluation results of candidate CCTs. This indicates the CCT physical baseline assessment result 242. This indicates that the CCT prediction residual value is 241.

[0095] In an implementation without further physical consistency checks, the second stability assessment result 251 can be output as the second stability assessment result.

[0096] Example 5: This example is a further implementation method after step S400 in Example 4, referring to... Figure 4 Please provide an explanation.

[0097] When the stability index to be evaluated is CCT, the corresponding GNN model 150 is a GNN model trained on the CCT residuals. GNN model 150 forms shared hidden state features based on the scenario 210 to be evaluated. The residual prediction branch obtains the predicted residual value 241 based on the shared hidden state features, and the physical quantity decoding branch obtains the pre-fault physical state 243 based on at least some of the same shared hidden state features. The pre-fault physical state 243 includes at least the node voltage amplitude corresponding to the steady-state operating state before the fault occurred. Phase angle of nodes .

[0098] Step S500: Determine the node injection power 262 before the fault.

[0099] Based on the pre-fault operation mode in scenario 210 to be evaluated, determine the active power injection and reactive power injection of each node, and form the pre-fault node injection power 262.

[0100] For nodes ,have:

[0101]

[0102]

[0103] in, and Representing nodes respectively The active and reactive power outputs of a conventional generator. and Representing nodes respectively The active and reactive power outputs of the new energy access equipment and Representing nodes respectively Active and reactive loads, and The power injected into the node before the fault is 262.

[0104] Step S600: Construct the pre-fault node admittance matrix 263.

[0105] Extract the pre-fault topology / equipment parameters corresponding to scenario 210 to be evaluated, and construct the pre-fault node admittance matrix 263 based on the power grid topology, line parameters, transformer parameters, and switch status before the fault occurs:

[0106]

[0107] in, This represents the admittance matrix of the node before the fault, 263. Let be the real part of the nodal admittance matrix. Let be the imaginary part of the nodal admittance matrix. The imaginary unit is used here. The power grid topology used to construct the pre-fault node admittance matrix 263 is the power grid topology before the fault occurs. For the fault scenario graph data used in residual prediction, it may include the post-fault topology state corresponding to the preset fault clearing scheme; the post-fault topology state is used to describe the fault scenario to be evaluated, and is not used to construct the pre-fault node admittance matrix 263 in this step.

[0108] Step S700: Perform a pre-fault steady-state power flow consistency check.

[0109] The pre-fault steady-state power flow consistency check module 260 receives the pre-fault physical state 243, the pre-fault node injected power 262, and the pre-fault node admittance matrix 263. Based on the node voltage amplitude and node phase angle obtained by decoding the GNN model 150, it calculates the corresponding AC power flow values ​​for each node.

[0110] node The calculated value of active power is:

[0111]

[0112] node The calculated reactive power value is:

[0113]

[0114] in, and These represent the node numbers respectively. and These represent the real and imaginary parts of the corresponding elements in the admittance matrix 263 of the node before the fault, respectively.

[0115] Further determine the nodes Active residual and reactive residual:

[0116]

[0117]

[0118] Based on system power reference value The active and reactive power residuals are normalized, and the power flow consistency residuals of the pre-fault operating state are determined based on the normalized residuals of all nodes. :

[0119]

[0120] in, Indicates the system power reference value. It represents the power flow consistency residual of the operating state before the failure.

[0121] The pre-fault steady-state power flow consistency check module 260 will Consistency threshold with preset power flow Compare the preset power flow consistency threshold. The threshold can be determined based on the verification scenario, which is independent of the training samples. In one implementation, the corresponding power flow consistency residuals are calculated for multiple verification scenarios, and a preset power flow consistency threshold is determined based on the statistical distribution of the obtained power flow consistency residuals. .

[0122] Step S800: Select the final stability assessment result 220 based on the inspection results.

[0123] When the following conditions are met: When the steady-state power flow consistency check before the fault is passed, the result selection module 255 selects input A, namely the second stability evaluation result 251, and uses it as the final stability evaluation result 220.

[0124] When the following conditions are met: When the steady-state power flow consistency check fails before the fault, the result selection module 255 selects input B, namely the physical baseline evaluation result 242, does not use the correction result formed by the predicted residual value 241, and uses the physical baseline evaluation result 242 as the final stability evaluation result 220, while generating alarm information 270.

[0125] like Figure 4 As shown, in addition to being input to the result fusion module 250 to form the second stability evaluation result 251, the physical baseline evaluation result 242 is also input to the result selection module 255 via channel B so that the physical baseline evaluation result 242 can be directly selected when the steady-state power flow consistency check fails before the fault.

[0126] In this embodiment, the pre-fault steady-state power flow consistency check is used to check the reliability of the shared hidden state features formed by the GNN model 150 for the current scenario 210 to the physical representation of the pre-fault power grid operating state. Both the predicted residual value 241 and the pre-fault physical state 243 are formed based on the shared hidden state features. Therefore, when the decoded pre-fault physical state 243 cannot satisfy the AC power flow relationship corresponding to the pre-fault operating mode and pre-fault network topology, the result selection module 255 stops using the predicted residual value 241 obtained based on the same shared hidden state features for correction.

[0127] The pre-fault steady-state power flow consistency check corresponds to the physical relationship check of the steady-state operating state before the fault occurs, and its result serves as the basis for determining whether to adopt the predicted residual value 241.

[0128] Example 6: Refer to Figure 1 and Figure 4 This embodiment describes the various modules in the system and their correspondence with the method steps in Embodiments 4 and 5.

[0129] The online evaluation module 200 includes a simplified physical model engine 130 for implementing physical model calculation functions, an online fault scenario graph builder 230 for implementing graph neural network inference functions, an index-corresponding GNN model 150, a result fusion module 250, and a pre-fault steady-state power flow consistency check module 260 and a result selection module 255, which are set in a further embodiment.

[0130] The simplified physical model engine 130 corresponds to execution step S200, which is used to determine the corresponding simplified physical model based on the stability index to be evaluated, and to obtain the physical baseline evaluation result 242 based on the operating mode data, power grid topology, equipment parameters and fault parameters in the scenario to be evaluated 210.

[0131] The online fault scenario graph builder 230 and the index-corresponding GNN model 150 jointly implement the function of the graph neural network inference module, corresponding to the execution step S300. The online fault scenario graph builder 230 is responsible for converting the scenario 210 to be evaluated into a graph structure input containing power grid operating status information and fault scenario information. The index-corresponding GNN model 150 is responsible for outputting the prediction residual value 241, and further outputting the physical state 243 before the fault in the CCT implementation.

[0132] The result fusion module 250, corresponding to step S400, adds the physical baseline evaluation result 242 to the predicted residual value 241 to form the second stability evaluation result 251. The pre-fault steady-state power flow consistency check module 260 implements the physical consistency check module function in the system, corresponding to the physical consistency check processing in steps S500 to S700. It obtains the power flow consistency residual of the pre-fault operating state based on the pre-fault physical state 243, the pre-fault node injected power 262, and the pre-fault node admittance matrix 263, and determines whether the power flow consistency residual meets the preset conditions.

[0133] The result selection module 255 receives the second stability assessment result 251 and the physical baseline assessment result 242, and performs step S800 based on the judgment result of the pre-fault steady-state power flow consistency check module 260 to determine the final stability assessment result 220 from the two inputs.

[0134] Therefore, in the system embodiment, physical model calculation, graph neural network inference, result fusion, and further physical consistency check and result selection correspond to physical baseline calculation, graph structure construction and residual inference, result fusion, pre-fault steady-state power flow consistency check, and final result selection in the method embodiment.

[0135] Example 7: Based on the training process of Examples 2 and 3, an active learning approach can be used to select training scenarios that require further stability analysis.

[0136] In one implementation, the current metric-corresponding GNN model 150 is used to perform multiple inferences on the unlabeled scene, and the model prediction variance is determined based on the variance of the multiple predicted residual values. In another implementation, multiple independently trained GNN models are used to predict the same unlabeled scene, and the model divergence is determined based on the differences between the multiple prediction results. The labeling priority of the unlabeled scene can also be evaluated based on the closeness of the prediction results to the current stable boundary, or the degree of difference between the unlabeled scene and existing training samples.

[0137] The scene to be labeled is determined based on at least one of the following: model prediction variance, the degree of divergence of prediction results among multiple independently trained GNN models, the degree of proximity of the prediction results to the stability boundary, and the degree of difference between the unlabeled scene and the existing training samples. The determined scene to be labeled is input into the stability analysis module 120 to obtain the corresponding stability evaluation ground truth 121. After forming the true residual label 142 according to the method of Example 2, it is added to the training sample set of the corresponding stability index for subsequent iteration training of the GNN model 150 corresponding to the index.

[0138] In another training implementation, the parameters of the GNN model already trained on a power grid with similar topology and equipment type can be used as the initial parameters for training the target power grid, and the corresponding model parameters can be trained again based on the training samples of the target power grid. This training method is used to reduce the data processing required for the target power grid model to start training from random initial parameters, without changing the online evaluation process of adding the physical baseline evaluation results and the prediction residual values.

[0139] In the above embodiments, CCT residual prediction, pre-fault steady-state power flow consistency check, and physical baseline result selection under abnormal conditions constitute a specific implementation form; for frequency stability index or static voltage stability index, the corresponding GNN model 150 is trained according to the corresponding stability index, and the physical baseline evaluation result is obtained by using the simplified physical model corresponding to the stability index. The basic residual correction process is consistent with the process in Embodiment 4.

[0140] As described above, although the invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the invention itself. Various changes in form and detail may be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A rapid evaluation method for high-proportion renewable energy grid connections, characterized in that, include: Obtain data on the operating mode, topology, equipment parameters, and fault parameters of the power grid to be evaluated, and determine the stability indicators to be evaluated; Based on the stability index to be evaluated, a simplified physical model corresponding to the stability index to be evaluated is determined. Based on the operating mode data, power grid topology, equipment parameters and fault parameters, the first stability evaluation result is calculated using the simplified physical model, and the first stability evaluation result is used as the physical baseline evaluation result. Based on the stability index to be evaluated, a pre-trained graph neural network model corresponding to the stability index to be evaluated is determined. The operation mode data, power grid topology, equipment parameters and fault parameters are constructed into graph structure data containing power grid operation status information and fault scenario information. The graph structure data is then input into the graph neural network model to obtain the predicted residual value for the physical baseline evaluation result. The physical baseline assessment result is added to the predicted residual value to obtain the second stability assessment result.

2. The rapid evaluation method for high-proportion renewable energy grid integration according to claim 1, characterized in that, Training sample sets are established for different types of stability indices to be evaluated, and corresponding graph neural network models are trained. For each training scenario, the true value of stability evaluation is obtained by using high-precision stability analysis adapted to the corresponding stability index, and the physical baseline evaluation result is obtained by using a simplified physical model adapted to the corresponding stability index. The difference between the true value of stability evaluation and the physical baseline evaluation result is determined as the true residual value and used as the residual label of the graph neural network model, so that different graph neural network models learn the residual mapping relationship of the corresponding stability index.

3. The rapid evaluation method for high-proportion renewable energy grid integration according to claim 1, characterized in that, The fault scenario information includes at least one of the following: fault type, fault location, fault impedance, fault component identifier, and post-fault topology state corresponding to the preset fault clearing scheme. Among them, bus faults are written into the graph structure data through the fault characteristics of the node corresponding to the fault bus, and line faults are written into the graph structure data through the fault characteristics of the edge corresponding to the fault line. For faults located inside the line, the corresponding fault line is divided into at least two edges according to the fault location inside the line, and a virtual fault node is set between the at least two edges to represent the fault location.

4. The rapid evaluation method for high-proportion renewable energy grid integration according to claim 1, characterized in that, When the stability index to be evaluated is the critical clearing time (CCT), the simplified physical model is the transient energy function (TEF) model or the extended equal area criterion (EEAC) model, the graph neural network model is a CCT residual prediction model trained on the CCT residual, and the graph structure data of the CCT residual prediction model includes at least one of the fault location, fault type, fault impedance, and post-fault topology state corresponding to the preset fault clearing scheme, and the graph structure data does not include the fault clearing time or CCT value as the object to be predicted.

5. The rapid evaluation method for high-proportion renewable energy grid integration according to claim 1, characterized in that, The graph neural network model includes a graph attention network backbone, a graph-level aggregation layer, a residual prediction layer, and a physical quantity decoder. The graph attention network backbone performs graph message passing based on the graph structure data to form node hidden state features. The graph-level aggregation layer forms a graph-level representation based on the node hidden state features. The residual prediction layer outputs the predicted residual value based on the graph-level representation. The physical quantity decoder outputs physical state variables for physical constraint calculation based on at least some of the node hidden state features.

6. The rapid evaluation method for high-proportion renewable energy grid integration according to claim 5, characterized in that, When training the graph neural network model, data loss is determined based on the difference between the predicted residual value and the actual residual value of the corresponding training sample. The residual of the AC power flow equation is calculated and the physical loss is determined based on the physical state variables output by the physical quantity decoder, the operating mode data of the corresponding training scenario, the power grid topology, and equipment parameters. Determine the total loss and update the parameters of the graph neural network model based on the total loss, wherein, Indicates the total loss. Indicates data loss. Indicates physical loss. The weighting factor represents the physical loss.

7. The rapid evaluation method for high-proportion renewable energy grid integration according to claim 1, characterized in that, When the stability index to be evaluated is the Critical Clearing Time (CCT), the graph neural network model forms a shared hidden state feature based on the same evaluation scenario to generate CCT prediction residuals and physical state variables. The physical state variables include at least the node voltage amplitude and node phase angle corresponding to the steady-state operating state before the fault. The node admittance matrix before the fault is constructed based on the node active power injection, node reactive power injection, grid topology, and equipment parameters before the fault. The power flow consistency residual of the operating state before the fault is determined based on the node voltage amplitude, node phase angle, node injection power, and the node admittance matrix before the fault. When the power flow consistency residual meets the preset conditions, the second stability evaluation result is used as the output result. When the power flow consistency residual does not meet the preset conditions, the CCT prediction residual is discarded and the physical baseline evaluation result is used as the output result.

8. The rapid evaluation method for high-proportion renewable energy grid integration according to claim 2, characterized in that, During the training process of the graph neural network model, unlabeled scenes are predicted, and scenes to be labeled are determined based on at least one of the following: model prediction variance, the degree of divergence of prediction results among multiple independently trained graph neural network models, the degree of proximity of prediction results to the stability boundary, and the degree of difference between unlabeled scenes and existing training samples. The scenes to be labeled are input into high-precision stability analysis to obtain the corresponding stability evaluation ground value, and the scenes that have obtained the stability evaluation ground value are added to the training sample set to iteratively train the corresponding graph neural network model.

9. A rapid evaluation system for high-proportion renewable energy grid integration, characterized in that, include: The physical model calculation module is used to determine a simplified physical model corresponding to the stability index to be evaluated, and to calculate a first stability evaluation result based on the operation mode data, grid topology, equipment parameters, and fault parameters of the power grid to be evaluated, and to use the first stability evaluation result as the physical baseline evaluation result; the graph neural network inference module is used to determine a pre-trained graph neural network model corresponding to the stability index to be evaluated, to construct the operation mode data, grid topology, equipment parameters, and fault parameters into graph structure data containing grid operation status information and fault scenario information, and to input the graph structure data into the graph neural network model to obtain the prediction residual value for the physical baseline evaluation result; The result fusion module is used to add the physical baseline evaluation result to the predicted residual value to obtain the second stability evaluation result.

10. The rapid evaluation system for high-proportion renewable energy grid integration according to claim 9, characterized in that, It also includes a physical consistency check module. When the stability index to be evaluated is the critical clearing time (CCT), the graph neural network inference module is further used to decode the node voltage amplitude and node phase angle corresponding to the steady-state operating state before the fault occurs based on the shared hidden state features formed for the same scenario to be evaluated. The physical consistency check module is used to determine the power flow consistency residual of the operating state before the fault occurs based on the node voltage amplitude, node phase angle, node active power injection, node reactive power injection, grid topology and equipment parameters before the fault occurs, and output the second stability evaluation result when the power flow consistency residual meets the preset conditions, and discard the predicted residual value and output the physical baseline evaluation result when the power flow consistency residual does not meet the preset conditions.

Images