Electric power system transient stability discrimination method and system based on graph isomorphic network

Abstract

The invention discloses an electric power system transient stability discrimination method and system based on a graph isomorphic network, and the method comprises the steps: building a training sample set of a graph isomorphic network model based on the operation condition of an electric power system; determining node features of transient stability discrimination of the power system, determining numerical values of the node features based on the training sample set, and preprocessing the numerical values of the node features; training a graph isomorphic network model based on the numerical values of the preprocessed node features to obtain a trained graph isomorphic network model; and judging the transient stability of the power system based on the trained graph isomorphic network model. The invention provides a method and a system for discriminating transient stability of a power system based on a graph isomorphic network, and the method and the system achieve the precise and efficient discrimination of the transient stability of a complex power grid through the powerful topological feature extraction capability of the GIN and a multi-layer perceptron aggregation mode, and provide reliable decision support for the security defense of the power grid.

Description

Technical Field

[0001] This invention relates to the field of power system transient stability discrimination technology, and more specifically, to a power system transient stability discrimination method and system based on graph isomorphic networks. Background Technology

[0002] With the continuous expansion of power grid scale and increasingly complex operation modes, transient stability assessment faces severe challenges: the dynamic response after system disturbances exhibits strong nonlinear characteristics, and phenomena such as power angle swing and voltage fluctuation are coupled with each other, making it difficult for traditional single-dimensional criteria to accurately identify stability boundaries; at the same time, online safety analysis needs to handle massive fault scenarios, placing higher demands on the real-time performance and generalization of assessment methods. Rapid and accurate assessment of transient stability is crucial for initiating emergency control measures and preventing faults from escalating into cascading trips or even major power outages.

[0003] Existing transient stability assessment methods are mainly divided into three categories: time-domain simulation methods, direct methods, and machine learning methods. Time-domain simulation methods simulate the dynamic evolution of the system by solving high-order differential algebraic equations. Although they have high accuracy, the computational cost increases exponentially with the system size, making it difficult to meet the real-time requirements of online pre-decision making in modern power grids. Direct methods, such as the extended equal-area criterion, rely on the precise analytical expression of the transient energy function. They are insufficient in characterizing the stability boundary under complex fault scenarios, and the results are often conservative. Traditional machine learning methods (such as support vector machines and random forests) are limited by manual feature engineering and have difficulty extracting deep nonlinear transient features from high-dimensional power grid data. Although deep learning methods such as convolutional neural networks and long short-term memory networks have self-learning capabilities, they require the power grid topology to be forcibly normalized into a grid structure, which destroys the connection semantics between nodes, resulting in distortion of the feature space and making it unable to adapt to the inherent graph structure characteristics of the power grid.

[0004] Graph Neural Networks (GNNs) offer a new paradigm for transient stability assessment by fusing topological connectivity and node state variables through message passing mechanisms. However, mainstream GNN models have significant drawbacks: Graph Convolutional Networks (GCNs) use an average aggregation method of neighbor features, which can lead to overly smooth node features and loss of topological details; Graph Sampling and Aggregation Networks (GraphSAGE) rely on node sampling strategies, which can easily ignore weakly connected nodes, resulting in a significant decrease in generalization performance when the power grid topology changes abruptly; more critically, the existing feature fusion mechanisms of GNNs cannot effectively capture the key features of the critical state of transient stability, and their robustness in complex disturbance scenarios is insufficient, limiting their engineering applications in real power grids. Summary of the Invention

[0005] The present invention provides a method and system for determining transient stability of power systems based on graph isomorphic networks, in order to solve the problem of how to accurately determine the transient stability state of complex power grids based on graph isomorphic networks.

[0006] To address the aforementioned problems, this invention provides a method for determining the transient stability of a power system based on graph isomorphic networks. The method includes:

[0007] Based on the operating conditions of the power system, a training sample set for a graph isomorphic network model is established.

[0008] The node characteristics for determining transient stability of the power system are determined, the numerical values ​​of the node characteristics are determined based on the training sample set, and the numerical values ​​of the node characteristics are preprocessed.

[0009] The graph isomorphic network model is trained based on the numerical values ​​of the preprocessed node features to obtain the trained graph isomorphic network model.

[0010] Based on a trained graph isomorphic network model, transient stability of the power system is determined.

[0011] Preferably, a training sample set for the graph isomorphic network model is established based on the operating conditions of the power system, including:

[0012] By adjusting the operating conditions of the power system, multiple operating conditions can be generated;

[0013] By randomly disconnecting one or more lines to adjust the topology of the power system, the operating conditions of the power system are regenerated based on the adjusted topology.

[0014] A training sample set for a graph isomorphic network model is established based on the regenerated operating conditions and the set fault parameters.

[0015] Preferably, the node characteristics include: voltage amplitude, voltage phase angle, average useful power flux of the node, average reactive power flux of the node, fault duration, and fault location.

[0016] Preferably, the node average useful power flux is defined as follows:

[0017]

[0018] Wherein, the input power P ji This indicates that active power flows from node j to node i, and the output power P ik This indicates that active power flows from node i to node k, N in N represents the set of neighboring nodes as input. out Indicates the set of output nodes;

[0019] The definition of the average reactive power flux at a node is:

[0020]

[0021] Wherein, the input power Qji This indicates that reactive power flows from node j to node i, with an output power Q. ik This indicates that reactive power flows from node i to node k.

[0022] Preferably, the method further includes establishing a graph isomorphic network model:

[0023] The node features H and the adjacency matrix A are input into the graph isomorphic network model for feature extraction. The feature vector of node i in the k-th layer of the graph isomorphic network model is obtained. Updated to:

[0024]

[0025] Among them, ò (k) For the learnable perturbation parameters of the k-th layer, Let be the feature vector of node i in the (k-1)th layer, and N(i) be the set of neighbors of node i. MLP is the sum of the feature vectors of all neighboring nodes of node i in the (k-1)th layer. (k) For a multilayer perceptron with layer k, which contains linear layers and ReLU layers:

[0026] MLP(z) = W1ReLU(W2z + b1) + b2

[0027] Where W1 and W2 are learnable matrices, b1 and b2 are bias terms, and ReLU is the activation function.

[0028] Preferably, the method further includes performing max pooling on the graph isomorphic network model to obtain the graph-level representation h. G :

[0029]

[0030] Among them, H (L) ∈R N×d R represents the node features obtained after passing through L layers. N×d Let represent an N×d dimensional matrix, and MAX denotes max pooling.

[0031] Preferably, the method further includes using graph-level representation h G The input is fed into the classifier's hidden layer using a multilayer perceptron (MLP) classification head for prediction.

[0032] h′ G =ReLU(h G W3+b3)

[0033]

[0034] in, To measure the probability distribution, W3, W outLet b3 be a learnable matrix. out This is a bias term.

[0035] Preferably, the optimization objective for training the graph isomorphic network model is to minimize the prediction probability distribution. Cross-entropy loss between the true label y and the actual label y:

[0036]

[0037] Among them, y c It is the c-th component of the one-hot encoded vector y. The probability distribution output by the graph isomorphic network model The c-th component.

[0038] Based on another aspect of the present invention, the present invention provides a power system transient stability discrimination system based on graph isomorphic networks, the system comprising:

[0039] Establish a training sample set for a graph isomorphic network model based on the operating conditions of the power system.

[0040] A determining unit is used to determine the node features for power system transient stability discrimination, determine the values ​​of the node features based on the training sample set, and preprocess the values ​​of the node features.

[0041] The training unit is used to train the graph isomorphic network model based on the numerical values ​​of the preprocessed node features to obtain the trained graph isomorphic network model.

[0042] The result unit is used to determine the transient stability of the power system based on the trained graph isomorphic network model.

[0043] Preferably, a training sample set for the graph isomorphic network model is established based on the operating conditions of the power system, including:

[0044] By adjusting the operating conditions of the power system, multiple operating conditions can be generated;

[0045] By randomly disconnecting one or more lines to adjust the topology of the power system, the operating conditions of the power system are regenerated based on the adjusted topology.

[0046] A training sample set for a graph isomorphic network model is established based on the regenerated operating conditions and the set fault parameters.

[0047] Preferably, the node characteristics include: voltage amplitude, voltage phase angle, average useful power flux of the node, average reactive power flux of the node, fault duration, and fault location.

[0048] Preferably, the node average useful power flux is defined as follows:

[0049]

[0050] Wherein, the input power P ji This indicates that active power flows from node j to node i, and the output power P ik This indicates that active power flows from node i to node k, N in N represents the set of neighboring nodes as input. out Indicates the set of output nodes;

[0051] The definition of the average reactive power flux at a node is:

[0052]

[0053] Wherein, the input power Q ji This indicates that reactive power flows from node j to node i, with an output power Q. ik This indicates that reactive power flows from node i to node k.

[0054] Preferably, the method further includes establishing a graph isomorphic network model:

[0055] The node features H and the adjacency matrix A are input into the graph isomorphic network model for feature extraction. The feature vector of node i in the k-th layer of the graph isomorphic network model is obtained. Updated to:

[0056]

[0057] Among them, ò (k) For the learnable perturbation parameters of the k-th layer, Let be the feature vector of node i in the (k-1)th layer, and N(i) be the set of neighbors of node i. MLP is the sum of the feature vectors of all neighboring nodes of node i in the (k-1)th layer. (k) For a multilayer perceptron with layer k, which contains linear layers and ReLU layers:

[0058] MLP(z) = W1ReLU(W2z + b1) + b2

[0059] Where W1 and W2 are learnable matrices, b1 and b2 are bias terms, and ReLU is the activation function.

[0060] Preferably, the method further includes performing max pooling on the graph isomorphic network model to obtain the graph-level representation h. G :

[0061]

[0062] Among them, H (L) ∈RN×d R represents the node features obtained after passing through L layers. N×d Let represent an N×d dimensional matrix, and MAX denotes max pooling.

[0063] Preferably, the method further includes using graph-level representation h G The input is fed into the classifier's hidden layer using a multilayer perceptron (MLP) classification head for prediction.

[0064] h′ G =ReLU(h G W3+b3)

[0065]

[0066] in, To measure the probability distribution, W3, W out Let b3 be a learnable matrix. out This is a bias term.

[0067] Preferably, the optimization objective for training the graph isomorphic network model is to minimize the prediction probability distribution. Cross-entropy loss between the true label y and the actual label y:

[0068]

[0069] Among them, y c It is the c-th component of the one-hot encoded vector y. The probability distribution output by the graph isomorphic network model The c-th component.

[0070] According to another aspect of the present invention, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a power system transient stability discrimination method based on graph isomorphic networks.

[0071] According to another aspect of the present invention, the present invention provides an electronic device, comprising:

[0072] The aforementioned computer-readable storage medium; and

[0073] One or more processors for executing a program in the computer-readable storage medium.

[0074] This invention provides a method and system for determining transient stability in power systems based on graph isomorphic networks. The method includes: establishing a training sample set for a graph isomorphic network model based on the operating conditions of the power system; determining the node features for determining transient stability, determining the numerical values ​​of the node features based on the training sample set, and preprocessing the numerical values ​​of the node features; training the graph isomorphic network model based on the preprocessed node feature values ​​to obtain a trained graph isomorphic network model; and determining the transient stability of the power system based on the trained graph isomorphic network model. This invention provides a method and system for determining transient stability in power systems based on graph isomorphic networks, utilizing the powerful topological feature extraction capability of GIN and the multilayer perceptron aggregation method to achieve accurate and efficient determination of the transient stability state of complex power grids, providing reliable decision support for power grid security defense. Attached Figure Description

[0075] Exemplary embodiments of the present invention can be more fully understood by referring to the following figures:

[0076] Figure 1 This is a flowchart of a power system transient stability determination method based on a graph isomorphic network according to a preferred embodiment of the present invention.

[0077] Figure 2 The above is a flowchart of the overall process for transient stability determination based on GIN according to a preferred embodiment of the present invention.

[0078] Figure 3 This is a schematic diagram of the training results of a graph isomorphic network model according to a preferred embodiment of the present invention;

[0079] Figure 4 This is a schematic diagram of the confusion matrix for performance evaluation of a graph isomorphic network model according to a preferred embodiment of the present invention;

[0080] Figure 5 A schematic diagram comparing and analyzing Graph Convolutional Networks (GCN), Graph Attention Networks (GAT), GraphSAGE, and the Graph Isomorphic Networks (GIN) proposed in this paper according to preferred embodiments of the present invention; and

[0081] Figure 6 This is a structural diagram of a power system transient stability discrimination system based on a graph isomorphic network according to a preferred embodiment of the present invention. Detailed Implementation

[0082] Exemplary embodiments of the invention will now be described with reference to the accompanying drawings. However, the invention may be embodied in many different forms and is not limited to the embodiments described herein. These embodiments are provided to fully and completely disclose the invention and to fully convey its scope to those skilled in the art. The terminology used in the exemplary embodiments illustrated in the drawings is not intended to limit the invention. In the drawings, the same units / elements are referred to by the same reference numerals.

[0083] Unless otherwise stated, the terms used herein (including technical terms) have their common meaning as understood by one of ordinary skill in the art. Furthermore, it is understood that terms defined in commonly used dictionaries should be understood to have a meaning consistent with the context of their relevant field, and not to be interpreted as having an idealized or overly formal meaning.

[0084] To address the issues of traditional deep learning models' inability to adapt to the characteristics of power grid graph structures and the oversmoothing and poor generalization of mainstream GNN models in existing transient stability discrimination methods, this invention provides a power system transient stability discrimination method based on graph isomorphic networks. By leveraging the powerful topological feature extraction capability of GIN and the aggregation method of multilayer perceptrons, it achieves accurate and efficient discrimination of transient stability states in complex power grids, providing reliable decision support for power grid security defense.

[0085] like Figure 1 As shown, this invention provides a method for determining the transient stability of a power system based on graph isomorphic networks. The method includes:

[0086] Step 101: Based on the operating conditions of the power system, establish a training sample set for the graph isomorphic network model;

[0087] Preferably, a training sample set for the graph isomorphic network model is established based on the operating conditions of the power system, including:

[0088] By adjusting the operating conditions of the power system, multiple operating conditions can be generated;

[0089] By randomly disconnecting one or more lines to adjust the topology of the power system, the operating conditions of the power system are regenerated based on the adjusted topology.

[0090] A training sample set for a graph isomorphic network model is established based on the regenerated operating conditions and the set fault parameters.

[0091] like Figure 2 This outlines the overall process for GIN-based transient stability assessment, including data generation, feature selection and data processing, the GIN principle, and classification prediction.

[0092] This invention is based on the PSASP simulation platform and constructs a model training sample set on the IEEE 39-node system. The dataset construction process is as follows: First, basic samples are generated by changing the system operating conditions: the power of all generators and loads varies within the range of 80%-120% of their rated values ​​in 10% increments, resulting in 5 baseline operating points. Based on this, a generator power is randomly selected and varied within the range of 70%-100%, or a load power is randomly varied within the range of 70%-130%, retaining one baseline operating condition, generating a total of 16 disturbance operating points. Subsequently, network topology changes are introduced to enhance data diversity: one or two lines are randomly disconnected before performing all the above changes, with specific configurations detailed in the table below. The aforementioned operating conditions are regenerated for each new topology. After ensuring that the power flow convergence of all operating conditions is achieved, fault simulation is performed to collect data: the fault is set as a three-phase short circuit, occurring on all 46 lines of the system at positions 10%, 50%, and 90% from the i-th side of line, with durations of 0.1s, 0.2s, 0.3s, and 0.5s, respectively. For each fault simulation, electrical characteristics and fault information obtained from short-circuit experiments are collected as input features, and an adjacency matrix is ​​constructed by treating buses as nodes and lines and generators as edges. Simultaneously, transient stability tests are conducted under the same conditions, and the results are used as sample labels. Finally, all generated samples are rigorously validated, and invalid samples caused by power flow non-convergence, simulation failure, or data anomalies are removed to form the final valid dataset.

[0093]

[0094] Step 102: Determine the node characteristics for power system transient stability assessment, determine the numerical values ​​of the node characteristics based on the training sample set, and preprocess the numerical values ​​of the node characteristics.

[0095] Preferably, the node characteristics include: voltage amplitude, voltage phase angle, average useful power flux of the node, average reactive power flux of the node, fault duration, and fault location.

[0096] Preferably, the node average useful power flux is defined as follows:

[0097]

[0098] Wherein, the input power P ji This indicates that active power flows from node j to node i, and the output power P ik This indicates that active power flows from node i to node k, N in N represents the set of neighboring nodes as input. out Indicates the set of output nodes;

[0099] The definition of the average reactive power flux at a node is:

[0100]

[0101] Wherein, the input power Q ji This indicates that reactive power flows from node j to node i, with an output power Q. ik This indicates that reactive power flows from node i to node k.

[0102] This invention considers the dynamic characteristics of the system and the physical mechanisms of faults, selecting six key node characteristics: voltage amplitude, voltage phase angle, average active power flux at the node, average reactive power flux at the node, fault duration, and fault location. The average active power flux at the node is defined as follows:

[0103]

[0104] Wherein, the input power P ji This indicates that active power flows from node j to node i, and the output power P ik This indicates that active power flows from node i to node k, N in N represents the set of neighboring nodes as input. out Represents the set of output nodes, and the average reactive power flux per node. Similarly, the definition is as follows:

[0105]

[0106] Wherein, the input power Q ji This indicates that reactive power flows from node j to node i, with an output power Q. ik This indicates that reactive power flows from node i to node k. These features correlate electrical state with transient response, effectively utilizing grid characteristics and significantly improving the model's generalization ability.

[0107] Step 103: Train the graph isomorphic network model based on the numerical values ​​of the preprocessed node features to obtain the trained graph isomorphic network model.

[0108] Step 104: Determine the transient stability of the power system based on the trained graph isomorphic network model.

[0109] Preferably, the method further includes establishing a graph isomorphic network model:

[0110] The node features H and the adjacency matrix A are input into the graph isomorphic network model for feature extraction. The feature vector of node i in the k-th layer of the graph isomorphic network model is obtained. Updated to:

[0111]

[0112] Among them, ò (k) For the learnable perturbation parameters of the k-th layer, Let be the feature vector of node i in the (k-1)th layer, and N(i) be the set of neighbors of node i. MLP is the sum of the feature vectors of all neighboring nodes of node i in the (k-1)th layer. (k) For a multilayer perceptron with layer k, which contains linear layers and ReLU layers:

[0113] MLP(z) = W1ReLU(W2z + b1) + b2

[0114] Where W1 and W2 are learnable matrices, b1 and b2 are bias terms, and ReLU is the activation function.

[0115] Preferably, the method further includes performing max pooling on the graph isomorphic network model to obtain the graph-level representation h. G :

[0116]

[0117] Among them, H (L) ∈R N×d R represents the node features obtained after passing through L layers. N×d Let represent an N×d dimensional matrix, and MAX denotes max pooling.

[0118] Preferably, the method further includes using graph-level representation h G The input is fed into the classifier's hidden layer using a multilayer perceptron (MLP) classification head for prediction.

[0119] h′ G =ReLU(h G W3+b3)

[0120]

[0121] in, To measure the probability distribution, W3, W out Let b3 be a learnable matrix. out This is a bias term.

[0122] Preferably, the optimization objective for training the graph isomorphic network model is to minimize the prediction probability distribution. Cross-entropy loss between the true label y and the actual label y:

[0123]

[0124] Among them, y c It is the c-th component of the one-hot encoded vector y. The probability distribution output by the graph isomorphic network model The c-th component.

[0125] The graph isomorphic network model of this invention takes the graph data G=(A,H) after graph sample construction and feature selection as input. The core architecture adopts a GIN encoder stack to learn the higher-order topological representation of nodes, and finally achieves transiently stable classification. The specific steps are as follows:

[0126] 1) Input the obtained node features H and adjacency matrix A into the GIN model for feature extraction, and extract the feature vector of node i in the k-th layer. Updated to:

[0127]

[0128] In the formula, ò (k) This represents the learnable perturbation parameters of the k-th layer. Let N(i) represent the feature vector of node i in the (k-1)th layer, and let N(i) represent the set of neighbors of node i. MLP represents the sum of the feature vectors of all neighboring nodes of node i at layer k-1. (k) This represents a multilayer perceptron with layer k containing linear and ReLU layers:

[0129] MLP(z) = W1ReLU(W2z + b1) + b2

[0130] In the formula, W1 and W2 are both learnable matrices, b1 and b2 are bias terms, and ReLU is the activation function.

[0131] 2) Unlike traditional average pooling, this paper uses max pooling to obtain the transient stability discrimination boundary, resulting in a graph-level representation h. G :

[0132]

[0133] In the formula, H (L) ∈R N×d R represents the node features obtained after passing through L layers. N×d Representing an N×d dimensional matrix, we can obtain d hid dimensional vector representation h G .

[0134] 3) The resulting graph-level representation h G The MLP classification head, input into the hidden layer of the classifier, is used for prediction.

[0135] h′ G =ReLU(h G W3+b3)

[0136]

[0137] In the formula, It measures the probability distribution, W3, Wout It is a learnable matrix, b3, b out This is a bias term.

[0138] 4) The optimization objective of model training is to minimize the predicted probability distribution. Cross-entropy loss between the true label y and the actual label y:

[0139]

[0140] In the formula, y c It is the c-th component of the one-hot encoded vector y. It is the probability distribution of the model output. The c-th component, L is the loss value of the final output.

[0141] The prediction model constructed in this invention includes the following key parameters and configurations: the node features input to the model adopt the 6-dimensional features described above; the model architecture uses an encoder consisting of an alternating stack of one layer of GIN and one layer of exponential linear unit (ELU) activation functions, for a total of three stacked layers, with the number of hidden layer channels set to 64, 256, and 1024 respectively, to achieve progressive feature extraction; to effectively identify transient stability boundaries, the graph pooling layer uses max pooling instead of traditional summation pooling; the classifier contains a hidden layer with a dimension of 512 and an output layer with a dimension of 2; during training, the learning rate is set to 0.001, the weight decay coefficient is 0.0005, the batch size is 64, and the maximum number of training epochs is 300; simultaneously, when the validation set loss does not decrease for 10 consecutive epochs, the learning rate will be multiplied by a decay factor of 0.7. The specific parameter configurations of the model are detailed in the table below:

[0142]

[0143] Experiments were conducted using a dataset generated by the IEEE 39-node system, containing 88,970 samples, of which 66,305 were stable and 22,665 were unstable. These samples were divided into training and validation sets in an 8:2 ratio. The results of feeding this dataset into the GIN model are as follows: Figure 3 As shown.

[0144] During model training, the accuracy and loss on the training and validation sets change as follows: Figure 3As shown in the figure, in the early stages of training, both the accuracy and loss of the training set increased rapidly, indicating that the model efficiently learned the data features. As training progressed, the accuracy of the training set eventually reached 98.79%, and the loss approached zero, demonstrating the model's strong fitting ability to the training data. The accuracy of the validation set generally showed an upward trend, but the rate of increase gradually slowed down, eventually stabilizing after 150 training epochs, with a peak value of 98.01%. The validation set loss remained at a relatively stable low level amidst fluctuations. Given that the accuracy of both the training and validation sets stabilized above 95% after 150 epochs, and the validation set reached its peak accuracy at this point, the model with the best performance on the validation set (accuracy 98.01%) was selected as the final discrimination model. Its performance evaluation confusion matrix is ​​shown in the figure. Figure 4 As shown.

[0145] The confusion matrix constructed based on the model with the highest accuracy on the validation set allows for a deeper analysis of the transient stability discrimination performance. For stable states labeled 1, the model correctly identified 13017 samples and misidentified 244. For unstable states labeled 0, it correctly identified 4424 samples and misidentified 109. The distribution of the number of identified samples shows that the model's ability to identify stable states is more prominent, with a relatively lower misclassification rate, indicating that the model captures the features of stable conditions more comprehensively. Although the number of misclassified unstable states is lower, there is still room for optimization. Overall, the confusion matrix validates the discrimination efficiency of the model with the highest accuracy on the validation set and also clarifies the direction for subsequent model optimization. Further optimization can be achieved by using reinforcement learning on the features of unstable state samples to improve the accuracy of transient stability discrimination, especially for unstable conditions, and to further optimize the model's performance in transient stability discrimination tasks.

[0146] The present invention has high discrimination accuracy: on the IEEE 39-node system dataset, the model classification accuracy reaches 98.01%, and the specificity (instability detection capability) is 0.5-0.8 percentage points better than mainstream GNN models such as GCN, GAT, and GraphSAGE, effectively reducing the risk of instability and missed detection;

[0147] The present invention has strong topology adaptability: it accurately identifies subtle differences in power grid topology through multilayer perceptron aggregation, and under the topology change scenario of single / double line disconnection, the fluctuation of various indicators is less than 0.3%, and the generalization ability is significantly better than the traditional GNN model.

[0148] The present invention has superior computational efficiency: it does not rely on complex physical models for solving, achieves millisecond-level discrimination based on data-driven methods, and its computation speed far exceeds that of time-domain simulation methods, thus meeting the real-time requirements of online security analysis.

[0149] The present invention has outstanding feature learning capabilities: it uses MLP to replace the simple aggregation operation of traditional GNN, which enhances the fitting ability of nonlinear transient features and can effectively capture the key dynamic response features of the system after disturbance.

[0150] To systematically evaluate the performance of the model in the task of power system transient stability discrimination, this invention selects Graph Convolutional Networks (GCN), Graph Attention Networks (GAT), GraphSAGE, and the proposed Graph Isomorphic Network (GIN) for comparative analysis, such as... Figure 5 As shown. Evaluation metrics include recall, accuracy, specificity, and Matthews correlation coefficient (MCC). Performance comparisons for each model are as follows. Figure 5 As shown in the figure. Experimental results show that GIN performs best in all three core metrics: Accuracy, Recall, and MCC, leading other models by 0.5 to 0.8 percentage points, highlighting its high discrimination accuracy for stable state samples. Crucially, in Specificity, a key metric for measuring unstable state identification, GIN significantly outperforms all comparison models with an excellent score of 98.01%, leading the second-best model (GraphSAGE) by 0.7 percentage points, fully validating GIN's ability to discriminate key unstable state samples.

[0151] These results fully demonstrate that the GIN-based model proposed in this invention exhibits superior overall performance in transient stability discrimination tasks. Its significant advantage in specificity, in particular, indicates an extremely low rate of false negatives for instability states. This is crucial for power system safety, because compared to the unnecessary increased control costs resulting from misclassifying stable conditions as unstable, failing to identify true instability states in a timely manner could trigger severe cascading failures or even major blackouts. The dual advantages of the GIN model in recall and specificity ensure that it can effectively capture stable patterns and, more importantly, accurately identify potential instability risks as a critical safety defense, thus providing more reliable and engineering-value-added stability discrimination support in complex and ever-changing power system transient scenarios.

[0152] This invention proposes a sample generation method based on three-dimensional disturbances of "topology-power-fault": by coordinating the adjustment of power grid topology, operating power parameters and fault settings, a high-quality sample set covering multiple scenarios is generated, providing data support for the model's generalization ability;

[0153] This invention employs a GIN feature encoder based on multilayer perceptron aggregation: it utilizes multilayer perceptrons to achieve nonlinear mapping of node neighbor features, replacing the simple average aggregation of traditional GNNs, avoiding oversmoothing problems, and improving the accuracy of topological feature extraction.

[0154] This invention proposes a global graph representation learning strategy based on max pooling: by extracting key transient boundary information from node features through max pooling, the ability to identify unstable states is enhanced and the risk of missed detection is reduced.

[0155] This invention presents a complete process for power system transient stability assessment based on GIN: integrating sample generation, feature encoding, pooling aggregation, and classification prediction to form a technical solution that can be directly applied to online assessment of transient stability in complex power grids.

[0156] Figure 6 This is a structural diagram of a power system transient stability discrimination system based on a graph isomorphic network according to a preferred embodiment of the present invention.

[0157] like Figure 6 As shown, this invention provides a power system transient stability discrimination system based on graph isomorphic networks. The system includes:

[0158] Unit 601 is established to create a training sample set for a graph isomorphic network model based on the operating conditions of the power system.

[0159] The determining unit 602 is used to determine the node characteristics for power system transient stability discrimination, determine the values ​​of the node characteristics based on the training sample set, and preprocess the values ​​of the node characteristics.

[0160] Training unit 603 is used to train the graph isomorphic network model based on the numerical values ​​of the preprocessed node features to obtain the trained graph isomorphic network model.

[0161] Result unit 604 is used to determine the transient stability of a power system based on a trained graph isomorphic network model.

[0162] Preferably, a training sample set for the graph isomorphic network model is established based on the operating conditions of the power system, including:

[0163] By adjusting the operating conditions of the power system, multiple operating conditions can be generated;

[0164] By randomly disconnecting one or more lines to adjust the topology of the power system, the operating conditions of the power system are regenerated based on the adjusted topology.

[0165] A training sample set for a graph isomorphic network model is established based on the regenerated operating conditions and the set fault parameters.

[0166] Preferably, the node characteristics include: voltage amplitude, voltage phase angle, average useful power flux of the node, average reactive power flux of the node, fault duration, and fault location.

[0167] Preferably, the node average useful power flux is defined as follows:

[0168]

[0169] Wherein, the input power Pji This indicates that active power flows from node j to node i, and the output power P ik This indicates that active power flows from node i to node k, N in N represents the set of neighboring nodes as input. out Indicates the set of output nodes;

[0170] The definition of the average reactive power flux at a node is:

[0171]

[0172] Wherein, the input power Q ji This indicates that reactive power flows from node j to node i, with an output power Q. ik This indicates that reactive power flows from node i to node k.

[0173] Preferably, the method further includes establishing a graph isomorphic network model:

[0174] The node features H and the adjacency matrix A are input into the graph isomorphic network model for feature extraction. The feature vector of node i in the k-th layer of the graph isomorphic network model is obtained. Updated to:

[0175]

[0176] Among them, ò (k) For the learnable perturbation parameters of the k-th layer, Let be the feature vector of node i in the (k-1)th layer, and N(i) be the set of neighbors of node i. MLP is the sum of the feature vectors of all neighboring nodes of node i in the (k-1)th layer. (k) For a multilayer perceptron with layer k, which contains linear layers and ReLU layers:

[0177] MLP(z) = W1ReLU(W2z + b1) + b2

[0178] Where W1 and W2 are learnable matrices, b1 and b2 are bias terms, and ReLU is the activation function.

[0179] Preferably, the method further includes performing max pooling on the graph isomorphic network model to obtain the graph-level representation h. G :

[0180]

[0181] Among them, H (L) ∈R N×d R represents the node features obtained after passing through L layers. N×d Let represent an N×d dimensional matrix, and MAX denotes max pooling.

[0182] Preferably, the method further includes using graph-level representation h G The input is fed into the classifier's hidden layer using a multilayer perceptron (MLP) classification head for prediction.

[0183] h′ G =ReLU(h G W3+b3)

[0184]

[0185] in, To measure the probability distribution, W3, W out Let b3 be a learnable matrix. out This is a bias term.

[0186] Preferably, the optimization objective for training the graph isomorphic network model is to minimize the prediction probability distribution. Cross-entropy loss between the true label y and the actual label y:

[0187]

[0188] Among them, y c It is the c-th component of the one-hot encoded vector y. The probability distribution output by the graph isomorphic network model The c-th component.

[0189] The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a power system transient stability determination method based on a graph isomorphic network.

[0190] This invention provides an electronic device, comprising:

[0191] The aforementioned computer-readable storage medium; and

[0192] One or more processors for executing a program in a computer-readable storage medium.

[0193] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0194] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0195] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0196] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0197] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0198] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0199] The invention has been described with reference to a few embodiments. However, as will be known to those skilled in the art, and as defined in the appended claims, other embodiments besides those disclosed above fall equivalently within the scope of the invention.

[0200] Generally, all terms used in the claims are to be interpreted according to their ordinary meaning in the art, unless otherwise expressly defined herein. All references to “a / the / the [device, component, etc.]” ​​are openly interpreted as at least one instance of said device, component, etc., unless otherwise expressly stated. The steps of any method disclosed herein need not be performed in the exact order disclosed unless explicitly stated otherwise.

Claims

1. A method for determining transient stability of a power system based on graph isomorphic networks, the method comprising: Based on the operating conditions of the power system, a training sample set for a graph isomorphic network model is established. Determine the node characteristics for power system transient stability assessment, determine the numerical values ​​of the node characteristics based on the training sample set, and preprocess the numerical values ​​of the node characteristics. The graph isomorphic network model is trained based on the numerical values ​​of the preprocessed node features to obtain the trained graph isomorphic network model. The transient stability of the power system is determined based on the trained graph isomorphic network model.

2. The method according to claim 1, wherein establishing the training sample set for the graph isomorphic network model based on the power system operating conditions includes: By adjusting the operating conditions of the power system, multiple operating conditions can be generated; By randomly disconnecting one or more lines to adjust the topology of the power system, the operating conditions of the power system are regenerated based on the adjusted topology. A training sample set for a graph isomorphic network model is established based on the regenerated operating conditions and the set fault parameters.

3. The method according to claim 1, wherein the node features include: Voltage amplitude, voltage phase angle, average useful power flux at nodes, average reactive power flux at nodes, fault duration, and fault location.

4. The method according to claim 3, wherein the average useful power flux of the node is defined as: in, Input power P ji This indicates that active power flows from node j to node i, and the output power P ik This indicates that active power flows from node i to node k, N in N represents the set of neighboring nodes. out Indicates the set of output nodes; The definition of the average reactive power flux at the node is: Wherein, the input power Q ji This indicates that reactive power flows from node j to node i, with an output power Q. ik This indicates that reactive power flows from node i to node k.

5. The method according to claim 1, further comprising establishing a graph isomorphic network model: The node features H and the adjacency matrix A are input into the graph isomorphic network model for feature extraction. The feature vector of node i in the k-th layer of the graph isomorphic network model is obtained. Updated to: in, ò (k) For the learnable perturbation parameters of the k-th layer, Let be the feature vector of node i in the (k-1)th layer, and N(i) be the set of neighbors of node i. MLP is the sum of the feature vectors of all neighboring nodes of node i in the (k-1)th layer. (k) For a multilayer perceptron with layer k, which contains linear layers and ReLU layers: MLP(z) = W1ReLU(W2z + b1) + b2 Where W1 and W2 are learnable matrices, b1 and b2 are bias terms, and ReLU is the activation function.

6. The method according to claim 5, further comprising performing max pooling on the graph isomorphic network model to obtain a graph-level representation h. G : in, H (L) ∈R N×d R represents the node features obtained after passing through L layers. N×d Let represent an N×d dimensional matrix, and MAX denotes max pooling.

7. The method of claim 6, further comprising using the graph hierarchy representation h G The input is fed into the classifier's hidden layer using a multilayer perceptron (MLP) classification head for prediction. h′ G =ReLU(h G W3+b3) in, To measure the probability distribution, W3, W out Let b3 be a learnable matrix. out This is a bias term.

8. The method according to claim 6, wherein the optimization objective of training the graph isomorphic network model is to minimize the prediction probability distribution. Cross-entropy loss between the true label y and the actual label y: in, y c It is the c-th component of the one-hot encoded vector y. The probability distribution output by the graph isomorphic network model The c-th component.

9. A power system transient stability determination system based on graph isomorphic networks, the system comprising: Establish a training sample set for a graph isomorphic network model based on the operating conditions of the power system. A determining unit is used to determine the node features for power system transient stability discrimination, determine the values ​​of the node features based on the training sample set, and preprocess the values ​​of the node features. The training unit is used to train the graph isomorphic network model based on the numerical values ​​of the preprocessed node features to obtain the trained graph isomorphic network model. The result unit is used to determine the transient stability of the power system based on the trained graph isomorphic network model.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1-8.

11. An electronic device, characterized in that, include: The computer-readable storage medium as described in claim 10; as well as One or more processors for executing a program in the computer-readable storage medium.

Images