Transient voltage security margin estimation method and device based on graph space-time network learning
By employing a graph-temporal network learning method, typical accident scenarios are automatically identified and safety margins are estimated, solving the deployment challenge of transient voltage safety domains in large-scale power grids and achieving high-precision safety margin assessment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HUNAN UNIV
- Filing Date
- 2026-03-17
- Publication Date
- 2026-05-15
AI Technical Summary
In existing technologies, transient voltage safety domain methods are difficult to deploy in large-scale power grids, especially in cases of unknown fault scenarios where the accuracy of safety margin estimation is insufficient, and their reliance on a set of anticipated incidents leads to poor applicability.
A graph-based spatiotemporal network learning approach is adopted, which constructs a model through graph convolutional networks to perform transient temporal simulation, shapelet clustering learning, and safety domain construction, and automatically identifies typical accident scenarios and estimates safety margins.
It achieves high-precision safety margin estimation in unknown accident scenarios, overcomes the limitations of traditional methods such as computational intensity and reliance on expected accident sets, and provides reliable safety margin assessment.
Smart Images
Figure CN121886384B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of transient voltage safety assessment, and in particular to a method and apparatus for estimating transient voltage safety margin based on graph spatiotemporal network learning. Background Technology
[0002] With the expansion of the power system and the increase in the adoption rate of renewable energy, the risk of short-term voltage instability caused by large disturbances in heavily loaded receiving areas has become quite prominent, posing a significant threat to the safe and stable operation of the power system.
[0003] Traditional transient voltage safety assessment methods typically evaluate system voltage safety by examining the system operating point in the pre-fault injection space for each predefined incident and determining whether the system remains stable. However, this point-by-point process is computationally intensive, making it difficult to implement in real-time operating environments. Currently, the dynamic safety domain method has advantages in transient voltage safety assessment research. The transient voltage safety domain characterizes the feasible operating region in the power injection space where the system maintains voltage safety after a specific disturbance. Methods based on transient voltage safety domains show great potential in assisting system operators to monitor voltage safety status and make timely preventative control decisions. However, in existing technologies, the transient voltage safety domain method remains difficult to deploy in large-scale power grids. Specific reasons include: traditional transient voltage safety domain construction requires establishing an independent safety domain model for each anticipated incident, which becomes computationally difficult as operating conditions and fault types in actual power systems become increasingly diverse; furthermore, when actual faults fall outside the anticipated incident set, safety assessment must rely on manually selecting the most similar transient voltage safety domain, which may lead to a significant decrease in the accuracy of safety margin estimation. In summary, existing technologies suffer from several drawbacks, including over-reliance on the set of anticipated incidents in constructing the transient voltage safety domain, poor applicability in the face of unknown scenarios, and large errors in safety margin estimation.
[0004] Therefore, a new technical solution is urgently needed to address the technical problem of how to automatically identify typical accident scenarios and estimate transient voltage safety margins. Summary of the Invention
[0005] This invention provides a transient voltage safety margin estimation method and apparatus based on graph spatiotemporal network learning, which solves the technical problem of how to automatically identify typical accident scenarios and perform transient voltage safety margin estimation.
[0006] To achieve the above objectives, this invention provides a transient voltage safety margin estimation method based on graph spatiotemporal network learning, comprising:
[0007] Transient time-domain simulation of the target power grid is performed based on a preset fault set to obtain a sample set containing pre-selected parameters for each node; shapelet clustering learning is performed based on the sample set to obtain a set of scenario parameters categorized based on typical scenarios; active power critical point search is performed based on the set of scenario parameters to obtain the transient voltage safety domain for each typical scenario; shapelet search for safety margin regression prediction is performed based on the transient voltage safety domain to obtain the shapelet for each typical scenario.
[0008] The first model is constructed based on graph convolutional networks. The first model is trained according to the shapelet and scene parameter set of each typical scenario to obtain the second model of each typical scenario. The second model of the corresponding typical scenario is selected according to the actual pre-selected parameters of the target power grid to estimate the safety margin.
[0009] Preferably, transient time-domain simulation of the target power grid is performed based on a preset fault set to obtain a sample set containing pre-selected parameters for each node, including:
[0010] For the number of nodes The power grid, using the active power injection combination of heavily loaded nodes as variables, is analyzed based on a preset fault set. Subtransient time-domain simulation generation A sample set is obtained by taking samples; the samples include pre-selected parameters for each node in a transient time-domain simulation; the pre-selected parameters include voltage amplitude, power angle, active power response and reactive power response.
[0011] Preferably, shapelet clustering learning is performed on the sample set to obtain a set of scene parameters based on typical scene classification, including:
[0012] Based on the sample set Two samples are selected as initial cluster centers. The distance between each remaining sample and each cluster center is calculated, and the remaining samples are assigned to the cluster corresponding to the nearest cluster center, resulting in two distance data subsets. The separation coefficient expression is obtained based on the mean and standard deviation of the two distance data subsets.
[0013] Maximizing the separation coefficient expression is used as the objective of shapelet clustering learning. Pre-selected parameters for each node are used as the original input, and a u-shapelet transformation is employed to transform the sample set. The distance dataset is obtained by mapping from the original temporal space to the distance feature space. Based on the distance dataset Clustering is performed using agglomerative hierarchical clustering methods, selecting preset clustering levels. The clustering results yielded... A typical scenario; based on A typical scenario will use the sample set The scene parameters are divided into segments to obtain the scene parameter set.
[0014] Preferably, the sample set is transformed using a u-shapelet transformation. The distance dataset is obtained by mapping from the original temporal space to the distance feature space. include:
[0015] Set the number of nodes in the target power grid The product of the number of types of pre-selected parameters is Then the sample set exist One dimension; for dimensional sample set , define the first Dimensional Data medium length is The shapelet is , Next The time series of each time series sample is , 1 ≤ ≤ n, where the nth Sub-time series ,but and The Euclidean distance between them includes:
[0016] ;
[0017] With the goal of minimizing the Euclidean distance, from the time series Extract the subsequence with the smallest distance value from the sample set. Transform into a distance dataset Distance to dataset The Middle Dimensional Data .
[0018] Preferably, based on the distance dataset Clustering using agglomerative hierarchical clustering methods includes:
[0019] Distance dataset In Each sample is clustered as an independent cluster, and the squared Euclidean distance between samples is used as a measure of dissimilarity between any two samples. After one clustering, the sum of squared deviations of each cluster is obtained based on the dissimilarity measure, and the distance dataset is obtained based on the sum of squared deviations of each cluster. The sum of squared deviations; attempt to aggregate any two clusters and select the distance dataset before and after aggregation. Aggregation is performed using the aggregation method with the smallest increment of the sum of squared deviations, until all clusters are aggregated into one cluster.
[0020] Preferably, based on the scenario parameter set, an active power critical point search is performed to obtain the transient voltage safety domain for each typical scenario, including:
[0021] The target power grid exists. Each heavily loaded node performs the first processing step under various typical scenarios. The first processing step includes:
[0022] With preset 3D active power injection vector As a baseline value, the stability of the target power grid is determined through time-domain simulation. When the target power grid is stable, the active power injection vector is increased; when the target power grid is unstable, the active power injection vector is decreased. When the stability of the target power grid changes after two consecutive increases or decreases in the active power injection vector, the latest active power injection vector is recorded as the first vector. The average of the baseline value and the first vector is calculated to obtain the second vector. Using the second vector as the new baseline value, the stability of the target power grid is determined again through time-domain simulation. This process is iterated until the interval between the latest active power injection vector and the previous active power injection vector is less than a preset threshold when the stability of the target power grid changes. The loop ends, and the average of the latest active power injection vector and the previous active power injection vector is calculated to obtain the critical sample. A preset number of baseline values with different combinations of values are set for time-domain simulation to obtain critical samples. When the number of obtained critical samples is greater than... When the acquisition stops, the critical sample set is obtained;
[0023] Based on the critical sample sets of each typical scenario, the least squares method is used to fit the samples in the power space to obtain the hyperplane expression of the transient voltage safety domain for each typical scenario.
[0024] Preferably, a shapelet search for safety margin regression prediction is performed based on the transient voltage safety domain, resulting in shapelets for various typical scenarios, including:
[0025] When performing a shapelet search on a single typical scenario, the shapelet is defined as existing in the typical scenario. For each sample, calculate The distance between the active power injection vector corresponding to each sample and the hyperplane expression of the transient voltage safety domain for the corresponding typical scenario in the power space is used as the actual safety margin value of each sample; the sample set is defined. The Middle Dimensional Data Next The actual safety margin of each time series sample is [value]. By approximating the boundary of the transient voltage safety domain in the distance feature space as a hyperplane, a multivariate linear safety margin estimation model is established, expressed as:
[0026] ;
[0027] in, and These are the bias coefficient and the weight coefficients for each dimension, respectively. and Obtain a compact vector ;
[0028] The objective function for safety margin regression prediction is constructed as follows:
[0029] ;
[0030] in, For regularization parameters;
[0031] With the goal of minimizing the objective function and estimation error, the sample set... Perform a shapelet search to obtain the shapelet for a single typical scenario; perform a shapelet search for each typical scenario to obtain the shapelet for each typical scenario.
[0032] Preferably, the first model includes graph convolutional layers and system layers;
[0033] Graph convolutional layers are used to extract graph structure features of the target power grid, including:
[0034] Normalized Laplace matrix :
[0035] ;
[0036] ;
[0037] in, Adjacency matrix Diagonal matrix corresponding to node degree; adjacency matrix The node admittance matrix of the target power grid is used; It is the identity matrix; It is a diagonal matrix composed of eigenvalues; The eigenvector matrix is sorted by its eigenvalues; This indicates the transpose;
[0038] Chebyshev polynomial approximation graph convolution kernels are used to perform local graph convolution operations; based on the shapelets of each typical scenario, the time-series data of voltage amplitude, active power response and reactive power response input to the first model are transformed into distance feature data, and nonlinear features are modeled by ReLU activation function to obtain voltage amplitude data, active power response data and reactive power response data after graph convolution layer processing;
[0039] The system layer is used to output a safety margin estimate based on the voltage amplitude data, active power response data, and reactive power response data processed by the graph convolutional layer, including:
[0040] The voltage amplitude data, active power response data, and reactive power response data after graph convolution layer processing are sequentially convolved and weighted to obtain the features of each node. Based on the features of each node and the preset node weight vector, the safety margin estimate is obtained through linear mapping.
[0041] Preferably, the first model is trained based on the shapelet and scene parameter set of each typical scenario, and the second model for each typical scenario is obtained as follows:
[0042] According to the scenario, the shapelet of each typical scenario and the pre-selected parameters corresponding to the scenario in the scenario parameter set are input into the first model. The root mean square error of the safety margin estimation is used as the loss function to train each typical scenario separately until the error of the safety margin estimation meets the preset requirements, and the second model of each typical scenario is obtained.
[0043] The present invention also provides a transient voltage safety margin estimation device based on graph spatiotemporal network learning, which is used in the method of the present invention. The device includes a first module, a second module, a third module and a fourth module.
[0044] The first module is used to perform transient time-domain simulation of the target power grid based on a preset fault set, and obtain a sample set containing pre-selected parameters for each node; shapelet clustering learning is performed based on the sample set to obtain a set of scenario parameters based on typical scenario classification;
[0045] The second module is used to perform active power critical point search based on the scenario parameter set to obtain the transient voltage safety domain for each typical scenario; and to perform shapelet search for safety margin regression prediction based on the transient voltage safety domain to obtain the shapelet for each typical scenario.
[0046] The third module is used to build the first model based on the graph convolutional network. The first model is trained according to the shapelet and scene parameter set of each typical scene to obtain the second model of each typical scene.
[0047] The fourth module is used to select a second model corresponding to a typical scenario based on the actual pre-selected parameters of the target power grid for safety margin estimation.
[0048] The present invention has the following beneficial effects:
[0049] This invention presents a transient voltage safety margin estimation method based on graph spatiotemporal network learning. By automatically classifying typical scenarios and selecting corresponding safety margin estimation models, it achieves accurate safety margin estimation. The method employs a hierarchical accident clustering approach based on time-series trajectories, demonstrating good generalization ability even in unknown accident scenarios. This effectively overcomes the limitations of traditional safety margin estimation, which relies on fault information, thus providing reliable safety margin assessments even under unknown disturbances. Furthermore, the method combines shapelet-guided temporal feature extraction with a graph spatiotemporal network model that considers topology, effectively learning the spatiotemporal characteristics of the power system and achieving high-precision safety margin estimation. Analysis of learnable parameters in the model effectively identifies critical vulnerable nodes affecting system safety.
[0050] The transient voltage safety margin estimation device based on graph spatiotemporal network learning of the present invention, when used in the method of the present invention, has the same beneficial effects as the method of the present invention.
[0051] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description
[0052] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:
[0053] Figure 1 This is a schematic diagram of the method flow of a preferred embodiment of the present invention. Detailed Implementation
[0054] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.
[0055] See Figure 1 In a preferred embodiment of the present invention, a transient voltage safety margin estimation method based on graph spatiotemporal network learning is provided, comprising:
[0056] F1. Perform transient time-domain simulation on the target power grid based on the preset fault set to obtain a sample set containing the pre-selected parameters of each node.
[0057] In a preferred embodiment of the present invention, F1 specifically includes:
[0058] For the number of nodes The power grid, using the active power injection combination of heavily loaded nodes as variables, is analyzed based on a preset fault set. Subtransient time-domain simulation generation A sample set is obtained by taking samples; the samples include pre-selected parameters for each node in a transient time-domain simulation; the pre-selected parameters include voltage amplitude, power angle, active power response, and reactive power response. The dimension of the sample set is then determined. .
[0059] F2. Perform shapelet clustering learning based on the sample set to obtain a set of scene parameters based on typical scene classification.
[0060] In a preferred embodiment of the present invention, F2 specifically includes:
[0061] Based on the sample set Two samples are selected as initial cluster centers. The distances between the remaining samples and each cluster center are calculated, and the remaining samples are assigned to the clusters corresponding to the nearest cluster centers, resulting in two distance data subsets. The separation coefficient expression is obtained based on the mean and standard deviation of the two distance data subsets, including:
[0062] Define the distance data subsets as follows and The separation coefficient is then expressed as:
[0063] ;
[0064] in, and They are respectively and The mean, and They are respectively and The standard deviation.
[0065] Maximizing the separation coefficient expression is used as the objective of shapelet clustering learning. Pre-selected parameters for each node are used as the original input, and a u-shapelet transformation is employed to transform the sample set. Mapping from the original temporal space to The distance dataset is obtained by using a dimensional distance feature space. Based on the distance dataset Clustering is performed using agglomerative hierarchical clustering methods, selecting preset clustering levels. The clustering results yielded... A typical scenario; based on A typical scenario will use the sample set The scene parameters are divided into segments to obtain the scene parameter set.
[0066] In a preferred embodiment of the present invention, a u-shapelet transformation is used to transform the sample set. Mapping from the original temporal space to The distance dataset is obtained by using a dimensional distance feature space. include:
[0067] Set the number of nodes in the target power grid The product of the number of types of pre-selected parameters is Then the sample set exist One dimension; for dimensional sample set , define the first Dimensional Data medium length is The shapelet is , Next The time series of each time series sample is , 1 ≤ ≤ , of which Sub-time series ,but and The Euclidean distance between them includes:
[0068] ;
[0069] With the goal of minimizing the Euclidean distance, from the time series Extract the value with the minimum distance (i.e., the minimum distance value) The sample set is selected from the subsequences that best match the shape. Transform into a distance dataset Distance to dataset The Middle Dimensional Data .
[0070] In a preferred embodiment of the present invention, based on the distance dataset Clustering using agglomerative hierarchical clustering methods includes:
[0071] Distance dataset In Each sample is clustered as an independent cluster, and the squared Euclidean distance between samples is used as a measure of dissimilarity between any two samples. After one clustering, the sum of squared deviations of each cluster is obtained based on the dissimilarity measure, and the distance dataset is obtained based on the sum of squared deviations of each cluster. The sum of squared deviations; attempt to aggregate any two clusters and select the distance dataset before and after aggregation. Aggregation is performed using the aggregation method that minimizes the increment of the sum of squared deviations, until all clusters are aggregated into one cluster. Where:
[0072] Sample data points and The measure of dissimilarity between them is expressed as:
[0073] ;
[0074] Define the prior knowledge obtained before a certain iteration of learning There are ... Then the first one The sum of squared deviations of each cluster is expressed as:
[0075] ;
[0076] in, For the first The cluster centers of each cluster; For the first The number of samples in each cluster; For the first The first cluster in the cluster One sample;
[0077] At this time, the distance from the dataset Sum of squared deviations Represented as:
[0078] ;
[0079] F3. Based on the scenario parameter set, perform an active power critical point search to obtain the transient voltage safety domain for each typical scenario.
[0080] In a preferred embodiment of the present invention, F3 specifically includes:
[0081] The target power grid exists. Each heavily loaded node performs the first processing step under various typical scenarios. The first processing step includes:
[0082] With preset 3D active power injection vector As a baseline value, the stability of the target power grid is determined through time-domain simulation. When the target power grid is stable, the active power injection vector is increased; when the target power grid is unstable, the active power injection vector is decreased. When the stability of the target power grid changes after two consecutive increases or decreases in the active power injection vector, the latest active power injection vector is recorded as the first vector. The average of the baseline value and the first vector is calculated to obtain the second vector. Using the second vector as the new baseline value, the stability of the target power grid is determined again through time-domain simulation. This process is iterated until the interval between the latest active power injection vector and the previous active power injection vector is less than a preset threshold when the stability of the target power grid changes. The loop ends, and the average of the latest active power injection vector and the previous active power injection vector is calculated to obtain the critical sample. A preset number of baseline values with different combinations of values are set for time-domain simulation to obtain critical samples. When the number of obtained critical samples is greater than... When the acquisition stops, the critical sample set is obtained;
[0083] Based on the critical sample sets of each typical scenario, and combined with the least squares method, sample fitting is performed in the power space to obtain the hyperplane expression of the transient voltage safety domain for each typical scenario, specifically including:
[0084] ;
[0085] in, For hyperplane coefficients; For bias terms; The active power injection values for heavily loaded nodes.
[0086] F4. Based on the transient voltage safety domain, perform shapelet search for safety margin regression prediction to obtain shapelets for each typical scenario.
[0087] In a preferred embodiment of the present invention, F4 specifically includes:
[0088] When performing a shapelet search on a single typical scenario, the shapelet is defined as existing in the typical scenario. For each sample, calculate The distance between the active power injection vector corresponding to each sample and the hyperplane expression of the transient voltage safety domain for the corresponding typical scenario in the power space is used as the actual safety margin value of each sample; the sample set is defined. The Middle Dimensional Data Next The actual safety margin of each time series sample is [value]. By approximating the boundary of the transient voltage safety domain in the distance feature space as a hyperplane, a multivariate linear safety margin estimation model is established, expressed as:
[0089] ;
[0090] in, and These are the bias coefficient and the weight coefficients for each dimension, respectively. and Obtain a compact vector ;
[0091] The objective function for safety margin regression prediction is constructed as follows:
[0092] ;
[0093] in, For regularization parameters;
[0094] With the goal of minimizing the objective function and estimation error, the sample set... Perform a shapelet search to obtain the shapelet for a single typical scenario; perform a shapelet search for each typical scenario to obtain the shapelet for each typical scenario.
[0095] F5. Construct the first model based on the graph convolutional network, train the first model according to the shapelet and scene parameter set of each typical scene, and obtain the second model of each typical scene.
[0096] In a preferred embodiment of the present invention, the first model includes a graph convolutional layer and a system layer;
[0097] Graph convolutional layers are used to extract graph structure features of the target power grid, including:
[0098] Normalized Laplace matrix :
[0099] ;
[0100] ;
[0101] in, Adjacency matrix Diagonal matrix corresponding to node degree; adjacency matrix The node admittance matrix of the target power grid is used; It is the identity matrix; It is a diagonal matrix composed of eigenvalues; The eigenvector matrix is sorted by its eigenvalues; This indicates the transpose;
[0102] In a preferred embodiment of the present invention, to reduce computational complexity, a Chebyshev polynomial approximation graph convolution kernel is used for local graph convolution operations. Specifically, a Chebyshev polynomial approximation graph convolution kernel is used. To implement local graph convolution operations, including:
[0103] ;
[0104] ;
[0105] ;
[0106] in, These are the learnable parameters in the convolution kernel; Here is the Chebyshev filter matrix; It is the largest eigenvalue; Let be the Chebyshev order.
[0107] Based on the shapelets of each typical scenario, the time-series data of voltage amplitude, active power response, and reactive power response input to the first model are transformed into distance feature data, which are represented as follows: , and Subsequently, spatial correlation modeling of these features is performed using graph convolutional layers to obtain voltage amplitude data after graph convolutional layer processing. Active power response data and reactive power response data Finally, the ReLU activation function is used to model the non-linear features. Graph convolution operations can be described as follows:
[0108] ;
[0109] ;
[0110] ;
[0111] in, , and The parameters learned by the graph convolutional layer; Represents graph convolution operations; Represents the Chebyshev filter matrix; This is the normalized Laplace operator.
[0112] , , It contains the features of each node in the graph convolutional layer after processing according to voltage, active and reactive channels.
[0113] The system layer is used to output a safety margin estimate based on the voltage amplitude data, active power response data, and reactive power response data processed by the graph convolutional layer, including:
[0114] The voltage amplitude data, active power response data, and reactive power response data after graph convolution layer processing are sequentially convolved and weighted summed at each node to obtain the features of each node. ,include:
[0115] ;
[0116] in, , , They are respectively , , The Middle Each node is a feature processed by a graph convolutional layer;
[0117] Based on the characteristics of each node and a pre-defined node weight vector, a safety margin estimate is obtained through linear mapping. ,include:
[0118] ;
[0119] in, , It is the learned node weight vector; It is the bias vector; This indicates the transpose operation.
[0120] In a preferred embodiment of the present invention, by introducing node weights, the degree of contribution of different nodes to the safety margin estimation can be reflected.
[0121] In a preferred embodiment of the present invention, training a first model based on the shapelet and scene parameter set of each typical scene to obtain a second model for each typical scene includes:
[0122] According to the scenario, the shapelet of each typical scenario and the pre-selected parameters corresponding to the scenario in the scenario parameter set are input into the first model. The root mean square error of the safety margin estimation is used as the loss function to train each typical scenario separately until the error of the safety margin estimation meets the preset requirements, and the second model of each typical scenario is obtained.
[0123] F6. Based on the actual pre-selected parameters of the target power grid, select the second model corresponding to the typical scenario to estimate the safety margin.
[0124] In the real-time prediction phase, second models for N typical scenarios are obtained. For unknown accident samples, a u-shapelet transformation is first used to map them to a distance feature space, and their similarity is compared with each typical scenario to assign them to the closest typical scenario. Subsequently, the corresponding second model is used to quickly calculate the distance from the sample to the safety domain boundary, thereby achieving a safety margin estimate for unknown events.
[0125] This invention presents a transient voltage safety margin estimation method based on graph spatiotemporal network learning. By automatically classifying typical scenarios and selecting corresponding safety margin estimation models, it achieves accurate safety margin estimation. The method employs a hierarchical accident clustering approach based on time-series trajectories, demonstrating good generalization ability even in unknown accident scenarios. This effectively overcomes the limitations of traditional safety margin estimation, which relies on fault information, thus providing reliable safety margin assessments even under unknown disturbances. Furthermore, the method combines shapelet-guided temporal feature extraction with a graph spatiotemporal network model that considers topology, effectively learning the spatiotemporal characteristics of the power system and achieving high-precision safety margin estimation. Analysis of learnable parameters in the model effectively identifies critical vulnerable nodes affecting system safety.
[0126] In a preferred embodiment of the present invention, a transient voltage safety margin estimation device based on graph spatiotemporal network learning is also provided for use in the method of the present invention. The device includes a first module, a second module, a third module, and a fourth module.
[0127] The first module is used to perform transient time-domain simulation of the target power grid based on a preset fault set, and obtain a sample set containing pre-selected parameters for each node; shapelet clustering learning is performed based on the sample set to obtain a set of scenario parameters based on typical scenario classification;
[0128] The second module is used to perform active power critical point search based on the scenario parameter set to obtain the transient voltage safety domain for each typical scenario; and to perform shapelet search for safety margin regression prediction based on the transient voltage safety domain to obtain the shapelet for each typical scenario.
[0129] The third module is used to build the first model based on the graph convolutional network. The first model is trained according to the shapelet and scene parameter set of each typical scene to obtain the second model of each typical scene.
[0130] The fourth module is used to select a second model corresponding to a typical scenario based on the actual pre-selected parameters of the target power grid for safety margin estimation.
[0131] The transient voltage safety margin estimation device based on graph spatiotemporal network learning of the present invention, when used in the method of the present invention, has the same beneficial effects as the method of the present invention.
[0132] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A transient voltage safety margin estimation method based on graph spatiotemporal network learning, characterized in that, include: Transient time-domain simulation of the target power grid is performed based on a preset fault set to obtain a sample set containing pre-selected parameters for each node. S ; According to the sample set S Shapelet clustering learning is performed to obtain a set of scene parameters based on typical scene classification; Based on the scenario parameter set, an active power critical point search is performed to obtain the transient voltage safety domain for each typical scenario: The target power grid exists. m Each heavily loaded node performs a first process under various typical scenarios, the first process including: With preset m 'Dynamic active power injection vector' p Using 0 as the base value, the stability of the target power grid is determined through time-domain simulation. When the target power grid is stable, the active power injection vector is increased; when the target power grid is unstable, the active power injection vector is decreased. When the stability of the target power grid changes after two consecutive increases or decreases in the active power injection vector, the latest active power injection vector is recorded as the first vector. The average of the base value and the first vector is calculated to obtain the second vector. Using the second vector as the new base value, the stability of the target power grid is determined again through time-domain simulation. This process is iterated until the interval between the latest active power injection vector and the previous active power injection vector is less than a preset threshold when the stability of the target power grid changes. The loop ends, and the average of the latest active power injection vector and the previous active power injection vector is calculated to obtain critical samples. A preset number of base values with different combinations are set for time-domain simulation to obtain critical samples. When the number of obtained critical samples is greater than 2... m When the acquisition stops, the critical sample set is obtained; m This represents the number of nodes in the power grid. Based on the critical sample sets of each typical scenario, the least squares method is used to perform sample fitting in the power space to obtain the hyperplane expression of the transient voltage safety domain for each typical scenario. Based on the transient voltage safety domain, a shapelet search is performed for safety margin regression prediction to obtain shapelets for each typical scenario: When performing a shapelet search on a single typical scenario, the shapelet is defined as existing in the typical scenario. n 'samples, calculate n The distance between the active power injection vector corresponding to each sample and the hyperplane expression of the transient voltage safety domain for the corresponding typical scenario in the power space is used as the actual safety margin value of each sample; the sample set is defined as follows. S The Middle k Dimensional Data S k Next p The actual safety margin of each time series sample is [value]. y p By approximating the boundary of the transient voltage safety domain in the distance feature space as a hyperplane, a multivariate linear safety margin estimation model is established, expressed as: ； in, W 0 and W k These are the bias coefficient and the weight coefficients for each dimension, respectively. W 0 and W k Obtain a compact vector W ={ W 0, W 1, W 2,…, W d }; d The number of dimensions; D k,p For distance dataset D The k Dimensional Data D k = { D k,1 , D k,2 ,…, D k,n The first in} p 1 element; distance to dataset D To use u-shapelet transformation to transform the sample set S Obtained by mapping from the original temporal space to the distance feature space; The objective function for safety margin regression prediction is constructed as follows: ； in, λ For regularization parameters; n The number of samples; For sample set S The k Dimensional Data S k medium length is shapelet; With the objective function and estimation error as the goal, the sample set is... S Perform shapelet search to obtain the shapelet for a single typical scenario; perform shapelet search for each typical scenario to obtain the shapelet for each typical scenario. A first model is constructed based on a graph convolutional network. The first model is trained according to the shapelet of each typical scenario and the scenario parameter set to obtain a second model for each typical scenario. The second model corresponding to the typical scenario is selected according to the actual pre-selected parameters of the target power grid for safety margin estimation.
2. The transient voltage safety margin estimation method based on graph spatiotemporal network learning according to claim 1, characterized in that, Transient time-domain simulation of the target power grid is performed based on a preset fault set to obtain a sample set containing pre-selected parameters for each node. S include: For the number of nodes m The power grid, using the active power injection combination of heavily loaded nodes as variables, is analyzed based on a preset fault set. n Subtransient time-domain simulation generation n The sample set is obtained from a sample of samples. S The sample includes pre-selected parameters for each node in a transient time-domain simulation; the pre-selected parameters include voltage amplitude, power angle, active power response, and reactive power response.
3. The transient voltage safety margin estimation method based on graph spatiotemporal network learning according to claim 2, characterized in that, According to the sample set S Shapelet clustering learning is performed to obtain a set of scene parameters based on typical scene classification, including: According to the sample set S Two samples are selected as initial cluster centers. The distance between each remaining sample and each cluster center is calculated, and the remaining samples are assigned to the cluster corresponding to the nearest cluster center, resulting in two distance data subsets. The separation coefficient expression is obtained based on the mean and standard deviation of the two distance data subsets. Maximizing the separation coefficient expression is used as the objective of shapelet clustering learning. Pre-selected parameters for each node are used as the original input, and a u-shapelet transformation is employed to transform the sample set. S The distance dataset is obtained by mapping from the original temporal space to the distance feature space. D According to the distance dataset D Clustering is performed using agglomerative hierarchical clustering methods, selecting preset clustering levels. N The clustering results yielded... N A typical scenario; according to the above N A typical scenario will use the sample set S The scene parameter set is obtained by dividing the scene into its components.
4. The transient voltage safety margin estimation method based on graph spatiotemporal network learning according to claim 3, characterized in that, The sample set is transformed using a u-shapelet transformation. S The distance dataset is obtained by mapping from the original temporal space to the distance feature space. D include: Set the number of nodes in the target power grid m The product of the number of types of the preselected parameters is d Then the sample set S exist d One dimension; for d The sample set of dimensional S , define the first k Dimensional Data S k medium length is The shapelet is , S k Next p The time series of each time series sample is X p , 1 ≤ p ≤ n , of which j Sub-time series ,but and The Euclidean distance between them includes: ； With the objective of minimizing the Euclidean distance, from the time series X p Extract the subsequence with the smallest distance value from the sample set. S Transform into a distance dataset D = { D 1, D 2,…, D d }, distance to dataset D The Middle k Dimensional Data D k = { D k,1 , D k,2 ,…, D k,n } 5. The transient voltage safety margin estimation method based on graph spatiotemporal network learning according to claim 4, characterized in that, According to the distance dataset D Clustering using agglomerative hierarchical clustering methods includes: The distance dataset D In n Each sample is clustered as an independent cluster, and the squared Euclidean distance between samples is used as a measure of dissimilarity between any two samples. After one clustering, the sum of squared deviations of each cluster is obtained based on the dissimilarity measure, and the distance dataset is obtained based on the sum of squared deviations of each cluster. D The sum of squared deviations; attempt to aggregate any two clusters, and select the distance dataset before and after aggregation. D Aggregation is performed using the aggregation method with the smallest increment of the sum of squared deviations, until all clusters are aggregated into one cluster.
6. The transient voltage safety margin estimation method based on graph spatiotemporal network learning according to claim 5, characterized in that, The first model includes graph convolutional layers and system layers; The graph convolutional layer is used to extract the graph structure features of the target power grid, including: Normalized Laplace matrix L : ； ； in, D Adjacency matrix A Diagonal matrix corresponding to node degree; adjacency matrix A The node admittance matrix of the target power grid is used; I n Δ is the identity matrix; Δ is a diagonal matrix composed of eigenvalues. U The eigenvector matrix is sorted by its eigenvalues; T This indicates the transpose; Chebyshev polynomial approximation graph convolution kernels are used to perform local graph convolution operations; based on the shapelets of each typical scenario, the time-series data of voltage amplitude, active power response and reactive power response input to the first model are transformed into distance feature data, and nonlinear features are modeled by ReLU activation function to obtain voltage amplitude data, active power response data and reactive power response data after graph convolution layer processing; The system layer is used to output a safety margin estimate based on the voltage amplitude data, active power response data, and reactive power response data processed by the graph convolutional layer, including: The voltage amplitude data, active power response data, and reactive power response data processed by the graph convolutional layer are sequentially convolved and weighted to obtain the features of each node. Based on the features of each node and the preset node weight vector, a safety margin estimate is obtained through linear mapping.
7. The transient voltage safety margin estimation method based on graph spatiotemporal network learning according to claim 6, characterized in that, The first model is trained based on the shapelet of each typical scenario and the scenario parameter set to obtain the second model for each typical scenario, including: According to the scenario, the shapelet of each typical scenario and the pre-selected parameters corresponding to the scenario in the scenario parameter set are input into the first model. The root mean square error of the safety margin estimation is used as the loss function to train each typical scenario separately until the error of the safety margin estimation meets the preset requirements, and the second model of each typical scenario is obtained.
8. A transient voltage safety margin estimation device based on graph spatiotemporal network learning, used in the method described in any one of claims 1 to 7, characterized in that, The device includes a first module, a second module, a third module, and a fourth module; The first module is used to perform transient time-domain simulation of the target power grid based on a preset fault set to obtain a sample set containing pre-selected parameters for each node; and to perform shapelet clustering learning based on the sample set to obtain a set of scenario parameters based on typical scenario classification. The second module is used to perform active power critical point search based on the scenario parameter set to obtain the transient voltage safety domain for each typical scenario; Based on the transient voltage safety domain, a shapelet search is performed for safety margin regression prediction to obtain shapelets for each typical scenario. The third module is used to construct a first model based on a graph convolutional network, train the first model according to the shapelet of each typical scene and the scene parameter set, and obtain a second model for each typical scene. The fourth module is used to select a second model corresponding to a typical scenario based on the actual pre-selected parameters of the target power grid to estimate the safety margin.