A power grid weak node identification method, system, device and medium based on a graph neural network
By constructing a graph neural network model that combines the power grid topology and electrical characteristics, the problem of inaccurate identification of weak nodes in the power grid in the existing technology is solved, and the accurate identification and evaluation of weak nodes in the power grid is realized, thereby improving the intelligent decision-making ability for safe operation of the power grid.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGXI POWER GRID CORP
- Filing Date
- 2026-05-06
- Publication Date
- 2026-06-12
AI Technical Summary
Existing technologies cannot effectively embed physical constraints, the prediction results lack physical feasibility, and the ability to integrate multi-dimensional operational data is insufficient, resulting in inaccurate identification of weak nodes in the power grid.
A graph neural network model is constructed, which combines the topology, electrical characteristics and operation data of the power grid. A loss function is generated through graph structure data and node feature set. An optimization algorithm is used to train the model and output a comprehensive vulnerability index of the nodes.
It enables accurate identification of weak nodes in the power grid, improves the intelligent decision-making capability for power grid security risk prevention and control, ensures the stability and generalization of model training, and provides scientific basis to support power grid planning and operation and maintenance.
Smart Images
Figure CN122196840A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power grid weak node identification technology, specifically to a power grid weak node identification method, system, device, and medium based on graph neural networks. Background Technology
[0002] With the advancement of new power system construction, the integration of high proportions of renewable energy and the complex interactions of diverse loads have significantly altered the performance of the power grid. Identifying weak nodes in the power grid is crucial for ensuring the safe and stable operation of the system. Traditional methods for identifying weak nodes mainly rely on static topological indices from complex network theory, such as node degree, betweenness centrality, and average path length, or on electrical quantity analysis methods based on power flow calculations, such as sensitivity analysis and electrical betweenness calculation. However, these traditional methods have many shortcomings. Complex network theory ignores the physical constraints and real-time operating status of the power grid, failing to accurately reflect the actual vulnerable nodes under fault disturbances. Traditional electrical quantity analysis methods only consider some physical characteristics of the power grid, resulting in computational complexity, poor real-time performance, and an inability to integrate multi-dimensional operational data.
[0003] In recent years, the development of artificial intelligence technology has provided new methods for identifying weak nodes in power grids. Graph neural networks are among the most advanced networks for processing non-Euclidean data. Graph neural networks abstract the power grid into a graph structure, and through node feature learning and neighbor information aggregation, they can uncover complex relationships within the power grid. However, graph neural network-based methods still have many shortcomings. Most models merely abstract the power grid into a simple topological graph, failing to embed physical constraints, and the prediction results lack physical feasibility. Summary of the Invention
[0004] In view of the above-mentioned existing problems, the present invention provides a method, system, device and medium for identifying weak nodes in power grids based on graph neural networks, in order to solve the problems of the failure to effectively embed physical constraints, the lack of physical feasibility of prediction results and the insufficient ability to fuse multi-dimensional operational data in the prior art.
[0005] To address the aforementioned technical problems, a method for identifying weak nodes in a power grid based on graph neural networks is proposed, including: The process involves constructing graph structure data of the target power grid, calculating and generating a set of node features and node vulnerability labels for model learning based on power grid operation scenario data, constructing and initializing a graph neural network model, using the graph structure data and node feature set as input, and configuring a loss function to measure prediction deviation. An optimization algorithm is then used to train the graph neural network model, using the node vulnerability labels as the target, and iteratively optimizing the loss function to update the model parameters to obtain the trained model. The trained model is then deployed and applied, inputting the power grid data to be identified, and outputting a comprehensive vulnerability index reflecting the vulnerability of each node.
[0006] As a preferred embodiment of the graph neural network-based method for identifying weak nodes in a power grid as described in this invention, the graph structure data for constructing the target power grid includes taking the busbars in the power grid as graph nodes, the transmission branches as connecting edges between nodes, and generating an adjacency matrix representing the connection status of all nodes in the network.
[0007] As a preferred embodiment of the graph neural network-based method for identifying weak nodes in a power grid, the generation of a node feature set includes comprehensively calculating the topological connectivity characteristics, basic electrical parameters, and real-time operating status of the power grid to form a multi-dimensional feature vector for each node. The generated node vulnerability labels include composite labels constructed by quantifying the structural criticality and actual load-bearing capacity of nodes in power transmission.
[0008] As a preferred embodiment of the power grid weak node identification method based on graph neural network described in this invention, the construction and initialization of the graph neural network model includes building a model architecture that includes a feature aggregation layer and a feature transformation layer. The feature aggregation layer receives the graph structure data of the adjacency matrix and the feature vectors of all nodes as input, and aggregates its own features and the feature information of all neighboring nodes for each node according to the connection relationship defined by the adjacency matrix to obtain the local fusion features of each node. The feature transformation layer receives the fused features of each node from the feature aggregation layer. Through a fully connected network with a nonlinear activation function, it performs nonlinear transformation and dimension mapping on the fused features, and outputs the predicted label of each node, namely the predicted node hub index and node capability index.
[0009] As a preferred embodiment of the graph neural network-based method for identifying weak nodes in a power grid, the configuration of the loss function for measuring prediction deviation includes selecting the mean squared error function as the objective function for model training, and quantifying the overall deviation of the model prediction by calculating the average of the sum of squares of the differences between the predicted labels and the true labels of all nodes.
[0010] As a preferred embodiment of the graph neural network-based method for identifying weak nodes in a power grid, the method for training the graph neural network model includes: training with an optimization algorithm, inputting training data containing an adjacency matrix and node feature vectors into the model, performing forward propagation calculations, and obtaining the predicted labels of each node. The loss value between the predicted label and the preset true label is calculated by using the configured mean squared error loss function, and the gradient of the loss value with respect to all weights and bias parameters in the model is solved by using the backpropagation algorithm. Based on historical gradient information, the exponential moving average of the first moment and the exponential moving average of the second moment of the gradient are updated respectively, and numerical corrections are performed on the updated first moment and second moment. Based on the corrected moment estimation, the learning step size of each parameter is adaptively determined, and the model weights and bias parameters are updated. The current process is repeated until the model training converges, and the trained graph neural network model is obtained.
[0011] As a preferred embodiment of the graph neural network-based power grid weak node identification method described in this invention, the output of the comprehensive weak index reflecting the vulnerability of each node includes applying the trained model to a new power grid operation scenario, inputting the adjacency matrix of the power grid and the multi-dimensional feature vector of the nodes in the current scenario, and outputting the predicted value of the hub index and the predicted value of the capability index of each node after forward propagation. The predicted values of each node are linearly combined to calculate the comprehensive vulnerability index of the current node. The comprehensive vulnerability index of all nodes constitutes the node vulnerability assessment result of the entire power grid. The linear combination includes the following: for any node, the comprehensive weakness index is the weighted sum of the predicted values of the hub index and the predicted values of the capability index, and the weight coefficients of the predicted values of the hub index and the capability index are the same.
[0012] The beneficial effects of this preferred technical solution are as follows: by combining the power grid topology, electrical characteristics and operating data to construct a graph neural network model, it is possible to identify and evaluate weak nodes in the power grid, thereby improving the intelligent decision-making ability for power grid safety risk prevention and control.
[0013] As a preferred embodiment of the power grid weak node identification system based on graph neural network described in this invention, it is characterized by including a power grid topology modeling module, a node feature and label construction module, a graph neural network model module, and a model training and evaluation module.
[0014] The power grid topology modeling module is used to transform the actual power grid structure into a data form suitable for graph neural network processing, abstracting the bus in the power grid as graph nodes and the transmission lines as edges between nodes, and describing the topological connection relationship of the power grid by generating an adjacency matrix.
[0015] The node feature and label construction module is used to extract the topological features, electrical features and operating status features of nodes from the power grid operation data to form a nine-dimensional feature vector. Based on the power grid structure and power flow distribution, it calculates the hub index reflecting the importance of the node structure and the capacity index reflecting the weakness of power carrying capacity, which together constitute the supervision label for model training.
[0016] The graph neural network model module is used to aggregate the feature information of neighboring nodes using graph convolutional layers, capture the local structural correlation of the power grid, and perform nonlinear transformation and dimensional mapping on the aggregated features through fully connected layers to output the predicted values of the node's hub index and capability index, thereby obtaining an end-to-end identification framework.
[0017] The model training and evaluation module is used to train the model using an adaptive moment estimation algorithm through an iterative process of forward propagation, loss calculation, backpropagation, and parameter update. After training, the model outputs the comprehensive weakness index of each node based on the input power grid data.
[0018] A computer device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of a method for identifying weak nodes in a power grid based on a graph neural network.
[0019] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a method for identifying weak nodes in a power grid based on a graph neural network.
[0020] The beneficial effects of this invention are as follows: This invention combines the construction of a graph neural network model with the power grid topology, electrical characteristics, and operational data to identify weak nodes in the power grid. By integrating multi-dimensional features such as node degree, betweenness centrality, average path length, node electrical betweenness, and node voltage importance, and using node hub index and node capability index as composite labels, it reflects the structural importance and functional vulnerability of nodes in the power grid. A standardized preprocessing method is used to eliminate the influence of different feature dimensions, ensuring the stability and convergence of model training. The combined use of the mean squared error loss function and the Adam optimization algorithm optimizes model parameters, avoids overfitting, improves the model's generalization ability, and provides technical assurance for the safe operation of the power grid. By using node hub index and node capability index, the vulnerability of nodes is comprehensively evaluated. The trained model can quickly predict weak points in the power grid, providing a scientific basis for power grid planning, operation and maintenance, and risk prevention and control, thus achieving safe and stable operation of the power grid. Attached Figure Description
[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0022] Figure 1The above is a flowchart of a method for identifying weak nodes in a power grid based on a graph neural network, which is provided as an embodiment of the present invention.
[0023] Figure 2 This is a schematic diagram of the graph neural network model training process for a graph neural network-based method for identifying weak nodes in a power grid, provided as an embodiment of the present invention.
[0024] Figure 3 This is a schematic diagram of the graph neural network model structure of a power grid weak node identification method based on graph neural network provided in an embodiment of the present invention.
[0025] Figure 4 This diagram illustrates the change in training loss value during the training process of a graph neural network model for identifying weak nodes in a power grid, as provided in an embodiment of the present invention.
[0026] Figure 5 This is a schematic diagram illustrating the prediction of the comprehensive weakness index of each node in the IEEE 30-node system using a graph neural network-based method for identifying weak nodes in a power grid, as provided in one embodiment of the present invention.
[0027] Figure 6 The present invention provides a system scheme flowchart for a power grid weak node identification system based on graph neural network as an embodiment of the present invention.
[0028] in, Figure 2 'e' represents the training epoch of the graph neural network model. Detailed Implementation
[0029] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.
[0030] Example 1, referring to Figure 1 As an embodiment of the present invention, a method for identifying weak nodes in a power grid based on a graph neural network is provided, comprising: S100: Construct graph structure data of the target power grid, and calculate and generate a set of node features and node weakness labels for model learning based on power grid operation scenario data.
[0031] S200: Construct and initialize the graph neural network model, taking graph structure data and node feature set as input, and configure the loss function to measure prediction bias.
[0032] S300: The graph neural network model is trained using an optimization algorithm. Taking the node vulnerability label as the target, the model parameters are updated by iteratively optimizing the loss function to obtain the trained model. The trained model is then deployed and applied. The data of the power grid to be identified is input, and the comprehensive vulnerability index reflecting the vulnerability of each node is output.
[0033] It should be noted that this invention combines graph convolutional networks with fully connected neural networks to aggregate neighbor node information and collect complex nonlinear relationships in the power grid, making the identification of weak nodes more accurate. Furthermore, it considers the topological location of nodes and incorporates the actual operating status and electrical attributes of nodes, making the identification results more consistent with the actual operation of the power grid.
[0034] Example 2, refer to Figure 1 This is a second embodiment of the present invention, which provides a method for identifying weak nodes in a power grid based on a graph neural network, including: In step S100, the construction of the graph structure data for the target power grid includes constructing the power grid adjacency matrix, specifically including steps S101~S104: S101: Construct a graph neural network model by treating the busbars and transmission branches in the power grid as nodes of the graph neural network.
[0035] S102: Abstract the power grid as a graph structure to construct a graph neural network model. Treat the bus in the power grid as nodes of the graph neural network and the transmission branches as connecting edges between nodes to construct the power grid adjacency matrix.
[0036] S103: Constructing the node adjacency matrix includes an N×N matrix A, where the elements of the node adjacency matrix A are defined as follows: Where N is the total number of power grid nodes, i.e., the total number of buses in the diagram. Let be the element in the i-th row and j-th column of the adjacency matrix A. and Let i and j be the i-th and j-th nodes.
[0037] S104: The method for constructing a graph neural network model is to construct a graph convolutional network and a full neural network, and connect them together. The formula for the graph convolutional network is expressed as: in, Let be the feature matrix of the nodes in the l-th layer of the graph neural network. Let be the feature matrix of the nodes in the (l+1)th layer (i.e., the next layer) of the graph neural network. For activation function, For matrix The degree matrix, For adjacency matrices with added self-connections, It is an N-dimensional identity matrix. Let be the trainable weight matrix of the l-th layer.
[0038] The formula for a fully neural network is expressed as: in, This is the output vector of the entire neural network. This is the trainable weight matrix for all layers of the neural network. The input vector for the entire neural network layer. is the trainable bias vector for the entire neural network layer.
[0039] Furthermore, in step S100, the node feature set includes node degree, betweenness centrality, average path length, node electrical betweenness, node voltage importance, power supply active output, power supply reactive output, active load, and reactive load; the node weakness label includes node hub index and node capability index.
[0040] Furthermore, in step S100, calculating and generating the node feature set and node weakness labels includes steps S111~S114: S111: Using power grid output, node load, and grid topology as data sources, calculate the characteristics and labels of each node and standardize each sample's characteristics and labels. Each node's characteristics and labels include nine features and two labels. The calculation process for the first to fifth features is as follows: in, For nodes The degree, Let be the element in the i-th row and j-th column of the adjacency matrix A. For nodes betweenness centrality, and In the network, except for nodes The other pair of source and target node numbers, This represents the total number of shortest paths from source node s to target node t. To find the shortest path from s to t that passes through the nodes The number of paths, For nodes Average path length, For nodes To the node The shortest path electrical distance between them For nodes Average electrical quantity For nodes The connected first The current in the branch, For nodes The importance of voltage, For nodes Maximum voltage under multiple operating scenarios For nodes Minimum voltage under multiple operating scenarios To take the reciprocal.
[0041] S112: Obtain nodes directly or through state estimation The four operating parameters, namely the fifth to ninth characteristics, are: active power output, reactive power output, active load, and reactive load.
[0042] Finally, the feature vector of each node is represented as: The features of all nodes constitute a feature matrix, which is represented as follows: .
[0043] in, For nodes The power supply has active power output. For nodes The power supply has no reactive power output. For nodes Active load, For nodes reactive load, For nodes eigenvectors.
[0044] S113: The calculation formulas for node hub indicators and node capability indicators are expressed as follows: in, Node hub indicators quantify the criticality of the topology. As a node capability indicator, it quantifies the operational vulnerability. This is the node number of the node where the generator is located. The node number of the node where the load is located. This is the set of nodes where the power source is located. The set of nodes where the load resides. This represents the minimum number of impedance paths between node r, where the power source is located, and node k, where the load is located. This indicates whether the minimum impedance path between node r and node k passes through a node. If it does, set it to 1; otherwise, set it to 0. This represents the active power flowing between node r and node k.
[0045] It should be noted that the node metrics for each node are expressed as follows: The node index matrix of all nodes is represented as follows: ; in, For the first Node metrics for each node This is a transpose operation.
[0046] S114: Standardize the features and labels of each node, as expressed by the formula: in, These are the original values before standardization. The mean of the original values. The standard deviation of the original values. This is the standardized output value.
[0047] In step S200, the construction and initialization of the graph neural network model includes the construction of a graph neural network model consisting of two parts: graph convolutional layers and fully connected layers connected sequentially, specifically including steps S201 to S203: S201: The graph convolutional layer uses an adjacency matrix with self-connections and introduces an identity matrix to ensure that each node retains its initial features during information aggregation.
[0048] To ensure the stability of the calculation process and overcome the influence of differences in the number of node connections, the adjacency matrix needs to be symmetrically normalized, and the diagonal elements of the degree matrix correspond to the adjusted number of connections for each node.
[0049] The normalized structural information is combined with the feature matrix of the current layer node and the trainable weight parameters, and after linear transformation, a non-linear activation function is used to perform feature mapping on the node. Each node generates a new feature representation that merges the local neighborhood structural relationships.
[0050] S202: After obtaining node features containing network structure information through graph convolutional layers, the feature vector of the current layer is input into the fully connected neural network.
[0051] A fully connected neural network consists of cascaded transformation layers. Each layer performs a linear weighted summation of the input features and adds a bias, and then uses a non-linear activation function to generate the output. The weights and biases are learnable parameters.
[0052] The output layer of the fully connected neural network generates a two-dimensional vector for each node, which corresponds to the quantitative prediction value of the criticality of the power grid structure and the quantitative prediction value of the weakness of the power carrying capacity of the current node, respectively.
[0053] S203: Configuring the loss function to measure prediction bias includes selecting the mean squared error loss function. The mean squared error loss function effectively penalizes large differences between predicted and true values, driving the model to fit continuous targets more accurately. L2 regularization is used to prevent overfitting. The graph neural network model parameters are optimized with the goal of minimizing the loss. The formula is expressed as: in, The value of the loss function. For the first The true value of each node, For the first The predicted value of each node.
[0054] In step S300, obtaining the trained model includes steps S301-S302: S301: Based on the labels in the dataset, the Adam algorithm is selected as the optimization algorithm. The labels of the nodes obtained through forward propagation are then used for backpropagation optimization. The formula for the Adam algorithm is as follows: in, For model parameter indexing, This represents the current iteration number. The loss function is relative to the q-th The gradient of each parameter, Let q be the first moment of the gradient of the q-th parameter. Let q be the second moment of the gradient of the q-th parameter. This is the first-order moment estimate after bias correction for the q-th parameter. Let be the attenuation coefficient of the first moment of the gradient, set to 0.9. Let be the attenuation coefficient of the second moment of the gradient, set to 0.999. Let q be the model parameter to be optimized. This is the bias-corrected second-order moment estimate for the q-th parameter. The learning rate is set to 0.0005. Let be an extremely small constant, denoted as 1e. -8This is used to prevent the denominator from being zero, where the decay coefficient is a default value verified based on deep learning practices.
[0055] S302: Dynamic decay scheduling based on the metric platform continuously monitors the loss function on the validation set. When the metric no longer improves after 10 consecutive training epochs, the learning rate decay is triggered. The monitoring value is set to min, meaning the smaller the loss function, the better. The decay factor is 0.5, meaning that when the metric no longer improves after 10 consecutive training epochs, the learning rate is multiplied by the current factor for decay. The decay tolerance value is 10, meaning that when the metric no longer improves after 10 consecutive training epochs, the learning rate is reduced.
[0056] It should be noted that setting the decay factor to 0.5 prevents the learning rate from dropping drastically, which is beneficial for the model to converge to a precise optimum. It also helps to narrow the search range without causing the model to get stuck in local suboptimal areas or stagnate during training due to rapid changes in step size. If the patience value is too short (2 or 3), decay may be triggered during a temporary small performance fluctuation, causing the model to converge slowly or get stuck in a suboptimal solution. If the patience value is too long (20 or 30), when the model clearly reaches a plateau, it will perform a large amount of redundant training with an excessively high learning rate, resulting in wasted time and effort and a high risk of overfitting. Setting the patience value to 10 avoids both excessively short and excessively long values.
[0057] Furthermore, in step S300, the output of the comprehensive vulnerability index reflecting the vulnerability of each node includes steps S311~S312: S311: Save the final model obtained from training, and obtain the node hub index and node capability index of each node based on the features and edge index of the input network. To ensure the objectivity of the evaluation results and the simplicity of the calculation model, the weights of the node hub index and node capability index are both set to 0.5.
[0058] The weight design is mainly to avoid subjective bias and to conduct a balanced evaluation of the hub attributes and capability attributes of nodes, so as to obtain the comprehensive weakness index of each node in the network.
[0059] S312: Input the edge index of the network. Combine the actual power grid topology and the active power flow direction between nodes to construct the edge index matrix. The adjacency matrix is E×2-dimensional, where E is the number of branches.
[0060] In this matrix, the first column represents the node number of the active power flow inflow to the branch, and the second column represents the node number of the active power flow outflow to the branch. The formula for calculating the comprehensive weakness index of each node in the network is as follows: in, For the first The comprehensive weakness index of each node is sorted in descending order based on the calculated comprehensive weakness index values of each node. The larger the value, the higher the comprehensive weakness of the node, and thus the more likely it is to be identified as a weak node of the power grid in the current operating scenario.
[0061] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
[0062] Example 3, referring to Figure 2 - Figure 5 This is the third embodiment of the present invention, which provides a method for identifying weak nodes in a power grid based on a graph neural network. To verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.
[0063] Based on the IEEE 30-node standard case, and using MATLAB software, the active and reactive power demand of the load and the active power output of the generator were set to fluctuate within the range of 70% to 130% of the IEEE 30-node standard case. 10,000 converged power flow calculation results were obtained, and the features and labels of each converged data point were calculated as a database for training the model. The standard case was selected as the data for validating the model.
[0064] After preparing the data, the model was trained using the graph neural network-based weak node identification method for power grids provided in this invention. The specific process is as follows: Figure 2 As shown.
[0065] In this embodiment, the training parameters are set according to Table 1. After forward propagation, the predicted labels are obtained. The predicted labels and the labels in the dataset are input into the loss function for calculation. The Adam optimizer is used to optimize the feature vector through backpropagation. After optimization, the weight matrix is updated. The above steps are repeated. The training parameters are shown in Table 1.
[0066] like Figure 4 As shown, the loss function value decreases rapidly in the first 50 iterations during training, and then tends to stabilize from the 100th iteration. Finally, the iterations are repeated 200 times, and the final trained model is saved.
[0067] Calculate the features of the IEEE 30-node standard example, and use these features as input to the trained model to obtain the following result: Figure 5 The overall weakness indicators of each node are shown.
[0068] Table 1 Training parameters
[0069] Example 4, refer to Figure 6 This is the fourth embodiment of the present invention. This embodiment provides a power grid weak node identification system based on graph neural network, including a power grid topology modeling module, a node feature and label construction module, a graph neural network model module, and a model training and evaluation module.
[0070] The power grid topology modeling module is used to transform the existing power grid structure into a power grid model suitable for graph neural network data processing, transform power grid buses into graph nodes, transmission lines into edges between nodes, and generate an adjacency matrix to represent the power grid topology relationship.
[0071] The node feature and label construction module is used to extract the topological features, electrical features and operating status features of nodes from the power grid operation data to generate a nine-dimensional feature vector, and to calculate the hub index reflecting the importance of the node structure and the capacity index reflecting the weakness of power carrying capacity based on the power grid structure and power flow distribution.
[0072] The graph neural network model module is used to aggregate the feature information of neighboring nodes using graph convolutional layers, capture the local structural correlation of the power grid, and perform nonlinear transformation and dimensional mapping on the aggregated features through fully connected layers to output the predicted values of the node's hub index and capability index, thereby obtaining an end-to-end identification framework.
[0073] The model training and evaluation module is used to train the model using an adaptive moment estimation algorithm through an iterative process of forward propagation, loss calculation, backpropagation, and parameter update. After training, the model outputs the comprehensive weakness index of each node based on the input power grid data.
[0074] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
[0075] Example 5, the fifth embodiment of the present invention, differs from the previous four embodiments in that: If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0076] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.
[0077] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.
[0078] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.
Claims
1. A method for identifying weak nodes in a power grid based on graph neural networks, characterized in that: include, Construct graph structure data of the target power grid, and calculate and generate a set of node features and node weakness labels for model learning based on power grid operation scenario data; Build and initialize a graph neural network model, using graph structure data and a set of node features as input, and configure a loss function to measure prediction bias; An optimization algorithm is used to train a graph neural network model. The model parameters are updated by iteratively optimizing the loss function to obtain the trained model. The trained model is then deployed and applied. The data of the power grid to be identified is input, and the comprehensive weakness index reflecting the vulnerability of each node is output.
2. The method for identifying weak nodes in a power grid based on a graph neural network as described in claim 1, characterized in that: The graph structure data for constructing the target power grid includes using the busbars in the power grid as graph nodes, the transmission branches as connecting edges between nodes, and generating an adjacency matrix that represents the connection status of all nodes in the network.
3. The method for identifying weak nodes in a power grid based on a graph neural network as described in claim 2, characterized in that: The generation of node feature sets includes comprehensively calculating the topological connection characteristics, basic electrical parameters and real-time operating status of the power grid to form a multi-dimensional feature vector for each node; The generated node vulnerability labels include composite labels constructed by quantifying the structural criticality and actual load-bearing capacity of nodes in power transmission.
4. The method for identifying weak nodes in a power grid based on a graph neural network as described in claim 3, characterized in that: The construction and initialization of the graph neural network model includes building a model architecture that includes a feature aggregation layer and a feature transformation layer. The feature aggregation layer receives the graph structure data of the adjacency matrix and the feature vectors of all nodes as input, and aggregates its own features and the feature information of all neighboring nodes for each node according to the connection relationship defined by the adjacency matrix to obtain the local fusion features of each node. The feature transformation layer receives the fused features of each node from the feature aggregation layer. Through a fully connected network with a nonlinear activation function, it performs nonlinear transformation and dimension mapping on the fused features, and outputs the predicted label of each node, namely the predicted node hub index and node capability index.
5. The method for identifying weak nodes in a power grid based on a graph neural network as described in claim 4, characterized in that: The configuration of the loss function to measure prediction bias includes selecting the mean squared error function as the objective function for model training, and quantifying the overall bias of model prediction by calculating the average of the sum of squares of the differences between the predicted labels and the true labels of all nodes.
6. The method for identifying weak nodes in a power grid based on a graph neural network as described in claim 5, characterized in that: The training of the graph neural network model includes training with an optimization algorithm, inputting training data containing adjacency matrices and node feature vectors into the model, performing forward propagation calculations, and obtaining the predicted labels of each node. The loss value between the predicted label and the preset true label is calculated by using the configured mean squared error loss function, and the gradient of the loss value with respect to all weights and bias parameters in the model is solved by using the backpropagation algorithm. Based on historical gradient information, the first and second moments exponential moving averages of the gradient are updated respectively, and numerical corrections are performed on the updated first and second moments. Based on the corrected moment estimation, the learning step size of each parameter is adaptively determined, and the model weights and bias parameters are updated. The current process is repeated until the model training converges, and the trained graph neural network model is obtained.
7. The method for identifying weak nodes in a power grid based on a graph neural network as described in claim 6, characterized in that: The output, which reflects the comprehensive vulnerability index of each node, includes applying the trained model to a new power grid operation scenario, inputting the adjacency matrix of the power grid and the multi-dimensional feature vector of the nodes in the current scenario, and outputting the predicted value of the hub index and the predicted value of the capability index of each node after forward propagation. The predicted values of each node are linearly combined to calculate the comprehensive vulnerability index of the current node. The comprehensive vulnerability index of all nodes constitutes the node vulnerability assessment result of the entire power grid. The linear combination involves, for any node, the comprehensive weakness index being equal to the weighted sum of the predicted values of the hub index and the capability index, where the weight coefficients of the predicted values of the hub index and the capability index are set to the same value.
8. A power grid weak node identification system based on graph neural networks, employing the power grid weak node identification method based on graph neural networks as described in any one of claims 1 to 7, characterized in that, It includes a power grid topology modeling module, a node feature and label construction module, a graph neural network model module, and a model training and evaluation module; The power grid topology modeling module is used to transform the actual power grid structure into a data form suitable for graph neural network processing, abstract the bus in the power grid as graph nodes, abstract the transmission lines as edges between nodes, and describe the topological connection relationship of the power grid by generating an adjacency matrix. The node feature and label construction module is used to extract the topological features, electrical features and operating status features of nodes from the power grid operation data to form a nine-dimensional feature vector. Based on the power grid structure and power flow distribution, it calculates the hub index reflecting the importance of the node structure and the capacity index reflecting the weakness of power carrying capacity, which together constitute the supervision label for model training. The graph neural network model module is used to aggregate the feature information of neighboring nodes using graph convolutional layers, capture the local structural correlation of the power grid, and perform nonlinear transformation and dimensional mapping on the aggregated features through fully connected layers to output the predicted values of the node's hub index and capability index, thereby obtaining an end-to-end identification framework. The model training and evaluation module is used to train the model using an adaptive moment estimation algorithm through an iterative process of forward propagation, loss calculation, backpropagation, and parameter update. After training, the model outputs the comprehensive weakness index of each node based on the input power grid data.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the power grid weak node identification method based on graph neural networks as described in any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the power grid weak node identification method based on graph neural network as described in any one of claims 1 to 7.