A method and apparatus for identifying power grid sensitive nodes based on graph attention networks

Abstract

This invention discloses a method and apparatus for identifying sensitive nodes in a power grid based on a graph attention network. The method includes: obtaining a parameter set of the power grid at the current moment based on graph theory; the parameter set includes an adjacency matrix and a set of node power feature vectors; obtaining a first node embedding based on the parameter set and a graph attention network; correcting the first node embedding by fusing the first vector set from the previous moment with a moving average to obtain a first vector set at the current moment; obtaining a second node embedding based on the first vector set and the parameter set at the current moment and a graph attention network; correcting the second node embedding by fusing the second vector set from the previous moment with a moving average to obtain a second vector set at the current moment; and obtaining the quantification result and ranking of the node sensitivity at the current moment based on the second vector set at the current moment and a fully connected neural network.

Description

Technical Field

[0001] This invention relates to the field of novel power system transmission network technology, and in particular to a method and apparatus for identifying sensitive nodes in a power grid based on graph attention networks. Background Technology

[0002] Traditional power grids are centered on centralized power sources such as thermal power, hydropower, and nuclear power, with load adjustments occurring seasonally and maintaining topological stability. However, distributed renewable energy sources, primarily in the form of photovoltaics, wind power, and energy storage, are more susceptible to environmental influences, characterized by small capacity, wide dispersion, and intermittent output. This necessitates frequent switching of distributed power sources on and off the grid to prevent voltage exceedances and backfeeding. As renewable energy penetration increases, the power grid system becomes increasingly complex, random, and volatile, significantly increasing the risk of cascading failures. Monitoring sensitive nodes can promptly detect early signs of faults and interrupt their propagation, which is crucial for ensuring stable grid operation.

[0003] See Figure 1 and Figure 2 The current assessment of critical and sensitive nodes in the power grid is mainly based on graph theory. First, the equipment and lines in the power grid are abstracted as nodes and edges in a graph (e.g., ...). Figure 1 The bus nodes 1 to 9 in the abstract are: Figure 2 The methods, from V1 to V9, consider the power flow characteristics and topology of the power grid. Two approaches are used: 1) Model-driven methods, such as the PageRank algorithm in complex network theory, use the actual impedance of lines as edge weights in the graph. By calculating indicators such as electrical betweenness, power flow betweenness, and node voltage safety margin, the importance of nodes is determined and ranked. 2) Data- and model-driven methods first evaluate and cluster power grid operation simulation data to obtain prior knowledge. Then, a graph convolutional neural network is used to capture the spatial dependencies of nodes and identify sensitive nodes under the current operating conditions of the power grid. Both evaluation methods are based on static topology snapshot analysis of the power grid, i.e., calculating the importance of nodes based on power grid operation data and topology at a fixed point in time. These methods do not fully consider the frequent, small-scale dynamic changes in the power grid topology over time and space after the integration of distributed renewable energy sources.

[0004] Therefore, a new technical solution is urgently needed to address the technical problem of how to quickly and accurately identify sensitive nodes in the power grid. Summary of the Invention

[0005] This invention provides a method and apparatus for identifying sensitive nodes in a power grid based on graph attention networks, in order to solve the technical problem of how to quickly and accurately identify sensitive nodes in a power grid.

[0006] To achieve the above objectives, this invention provides a method for identifying power grid sensitive nodes based on graph attention networks, comprising: The parameter set of the power grid at this moment is obtained based on graph theory. The parameter set includes the adjacency matrix and the set of node power feature vectors. The first node embedding is obtained by combining the parameter set with a graph attention network. The first node embedding is corrected by fusing the first vector set of the previous moment with a moving average to obtain the first vector set at this moment. The second node embedding is obtained by combining the first vector set and parameter set at this time with the graph attention network; the second node embedding is corrected by fusing the second vector set from the previous time step with a moving average to obtain the second vector set at this time step; the quantification result and ranking of the node sensitivity at this time step are obtained by combining the second vector set at this time step with a fully connected neural network.

[0007] Preferably, the embedding of the first node obtained by combining the parameter set with the graph attention network includes: The parameter set is input into the graph attention network and subjected to a first process to obtain the first node embedding; the first process includes: After unifying the dimensions of the power feature vectors of each node by using a weight matrix, the attention coefficients between each node and all neighboring nodes are calculated by combining the adjacency matrix to obtain the initial attention coefficients. The initial attention coefficients are corrected based on the edge weights to obtain the corrected attention coefficients; the corrected attention coefficients are then normalized using the Softmax function to obtain the first attention weights. The first attention weight is weighted and fused with the features of the neighboring nodes, then processed by a nonlinear function, and a preset residual term is superimposed to output the node embedding; When performing the first processing based on the parameter set, the preset residual term includes the power feature vector of the node itself.

[0008] Preferably, the attention coefficients between each node and all its neighboring nodes are calculated to obtain the initial attention coefficients, including: For nodes compute nodes For nodes Attention coefficient , Represents a node The set of neighboring nodes, attention coefficient Represented as: ; in, The graph attention parameter vector is represented by T; T represents the transpose operation. Indicates feature splicing; Represents the linear mapping matrix of weights; and They are respectively Time Node and nodes The power characteristic vector.

[0009] Preferably, the initial attention coefficients are corrected based on the edge weights to obtain corrected attention coefficients; the corrected attention coefficients are then normalized using the Softmax function to obtain the first attention weights, including: The initial attention coefficients are corrected based on the edge weights to obtain the corrected attention coefficients. : ; in, for Time Side The weight value; The corrected attention coefficients are normalized using the Softmax function to obtain the first attention weights. : ; in, It is an exponential function; Belonging to Nodes; For nodes For nodes Attention coefficient.

[0010] Preferably, the first attention weights are weighted and fused with the features of neighboring nodes, then processed by a nonlinear function, and a preset residual term is superimposed. The output node embedding includes: ; in, This is the node embedding output after superimposing preset residual terms; for Time Node The preset residual term; For nodes For nodes Attention weights; for Time Node The power characteristic vector.

[0011] Preferably, the embedding of the first node is corrected by fusing the first vector set from the previous time step using a moving average, resulting in the first vector set for the current time step, which includes: The following adjustments were made to each node: ; in, for The current time node The first vector; for The previous time node The first vector is initialized to 0 when t is 0; for The current time node The first node is embedded; This is a correction factor; The first vector set at this moment is obtained based on the first vector of each node at this moment.

[0012] Preferred correction coefficient include: ; in, for The sum of edge weights in a graph-based power grid at any given time. for The sum of edge weights in a graph-based power grid at any given time.

[0013] Preferably, the second node embedding is obtained by combining the first vector set and parameter set at this moment with the graph attention network, including: The first vector set is added to the node power feature vector set to obtain the optimized node power feature vector set; the optimized node power feature vector set and the adjacency matrix are input into the graph attention network for the first processing to obtain the second node embedding; when performing the first processing based on the optimized node power feature vector set and the adjacency matrix, the preset residual term includes the first vector set.

[0014] Preferably, the second node embedding is corrected by fusing the second vector set from the previous time step using a moving average, resulting in the second vector set for the current time step, which includes: The following adjustments were made to each node: ; in, for The current time node The second vector; for The previous time node The second vector is initialized to 0 when t is 0; for The current time node The second node embedding; The second vector set at this moment is obtained based on the second vector of each node at this moment.

[0015] The present invention also provides a power grid sensitive node identification device based on graph attention network, which is used in the method of the present invention. The device includes a first module, a second module, a third module and a fourth module. The first module is used to obtain the parameter set of the power grid at this moment based on graph theory; the parameter set includes the adjacency matrix and the set of node power feature vectors; The second module is used to obtain the first node embedding based on the parameter set and the graph attention network; the first node embedding is corrected by fusing the first vector set of the previous time step with a moving average to obtain the first vector set of the current time step. The third module is used to obtain the second node embedding by combining the first vector set and parameter set at this time with the graph attention network; the second node embedding is corrected by fusing the second vector set at the previous time through moving average to obtain the second vector set at this time. The fourth module is used to obtain the quantification and ranking of the node sensitivity at this time by combining the second vector set at this time with the fully connected neural network.

[0016] The present invention has the following beneficial effects: This invention presents a power grid sensitive node identification method based on graph attention networks. By introducing a graph attention mechanism with edge weights, the graph neural network can selectively fuse actual physical parameters such as power system line resistance and real-time power flow values. This accurately captures the differentiated interaction relationships between different types of nodes, such as hub substations and distributed renewable energy sources, and lines operating in different environments. This avoids the neglect of line physical characteristics by traditional unweighted graph models, improving the physical correlation and accuracy of node feature representation. By continuously fusing node embeddings from the historical operating states of the power system, the static graph is expanded into a dynamic graph, accurately reflecting the dynamic evolution characteristics of nodes over time and adapting to changes in temporal operating states such as power system load fluctuations, equipment switching, and fault transients. Simultaneously, the rolling update mechanism preserves the temporal dependencies of historical operating states, overcoming the limitation of static graph models in failing to depict node state migration over time, and ensuring the continuity and integrity of node features in the temporal dimension. This invention's method tracks the spatiotemporal migration of sensitive nodes using temporal topology data and real-time power grid operating data, enabling rapid and dynamic identification of sensitive nodes in the power grid.

[0017] The power grid sensitive node identification device based on graph attention network of the present invention, used in the method of the present invention, has the same beneficial effects as the method of the present invention.

[0018] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0019] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the power grid topology according to a preferred embodiment of the present invention.

[0020] Figure 2 This is a graph-based power grid diagram according to a preferred embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram of the method flow of a preferred embodiment of the present invention.

[0022] Figure 4 This is a schematic diagram illustrating the principle of the attention mechanism that considers edge weights in a preferred embodiment of the present invention. Detailed Implementation

[0023] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0024] See Figures 3 to 4 In a preferred embodiment of the present invention, a method for identifying power grid sensitive nodes based on graph attention networks is provided, characterized by comprising: F1. Obtain the parameter set of the power grid at this moment based on graph theory; the parameter set includes the adjacency matrix and the set of node power feature vectors.

[0025] In a preferred embodiment of the present invention, the power system network is abstracted into a structure containing graph theory. A graph with 1 node ,in Represents the set of bus nodes, with weighted edges. The set of paths between nodes can be represented by an adjacency matrix. This indicates that the elements in the adjacency matrix Representative node and nodes Are there any directly connected edges between them? , Including nodes and nodes Electrical properties such as line impedance, line admittance, and current carrying capacity between them. This is the attribute feature vector of the bus node, whose components are the node category, voltage, current, active power, reactive power, voltage safety margin, etc., which can be collected in real time through equipment such as synchronous phasor measurement devices. The power grid operating state at each moment can be viewed as a graph. Snapshot at the corresponding moment .

[0026] Get Time-based power grid model The values ​​of each element contained therein, , and They are The set of nodes, edges, and attributes of the power grid topology at any given time; the actual data to be obtained includes: Get feedback Adjacency matrix of the connection relationships between nodes in the time-limited network ,by Taking a power grid with a node as an example, that is, at this time The dimension is , expressions and Values ​​of each element The expressions include: ; ; in, for Time Node and nodes The lines between The normalized values ​​of electrical properties such as impedance, admittance, and upper limit of allowable current.

[0027] Obtain the feature vectors of all nodes (in The node's sequence number. , The values ​​of each element in the node feature dimension are normalized. The feature vectors are stacked vertically to form Input feature matrix at time step ( ).

[0028] F2. Obtain the first node embedding based on the parameter set and the graph attention network; correct the first node embedding by fusing the first vector set from the previous time step with a moving average to obtain the first vector set at the current time step.

[0029] In a preferred embodiment of the present invention, obtaining the first node embedding based on the parameter set combined with the graph attention network includes: The parameter set is input into a graph attention network and subjected to a first process to obtain a first node embedding; in a preferred embodiment of the present invention, the parameter set is processed by a first graph attention network layer and the first node embedding is output; the first process includes: After unifying the dimensions of the power feature vectors of each node by using a weight matrix, the attention coefficients between each node and all neighboring nodes are calculated by combining the adjacency matrix to obtain the initial attention coefficients. The initial attention coefficients are corrected based on the edge weights to obtain the corrected attention coefficients; the corrected attention coefficients are then normalized using the Softmax function to obtain the first attention weights. The first attention weight is weighted and fused with the features of the neighboring nodes, then processed by a nonlinear function, and a preset residual term is superimposed to output the node embedding; When performing the first processing based on the parameter set, the preset residual term includes the power feature vector of the node itself.

[0030] In a preferred embodiment of the present invention, unifying the dimensionality of the power feature vectors of each node through a weight matrix includes: Through the weight matrix Linear mapping, increase and unify Node feature matrix at time step Dimensions of node features: ; in, for The dimension of the unified node feature at any time is The feature matrix after that.

[0031] In a preferred embodiment of the present invention, the attention coefficients of each node and all its neighboring nodes are calculated to obtain the initial attention coefficients, including: For nodes compute nodes For nodes Attention coefficient , Represents a node The set of neighboring nodes, attention coefficient Represented as: ; in, The graph attention parameter vector is represented by T; T represents the transpose operation. Indicates feature splicing; Represents the linear mapping matrix of weights; and They are respectively Time Node and nodes The power characteristic vector.

[0032] In a preferred embodiment of the present invention, the initial attention coefficients are corrected based on the edge weights to obtain corrected attention coefficients; the corrected attention coefficients are then normalized using the Softmax function to obtain the first attention weights, including: The initial attention coefficients are corrected based on the edge weights to obtain the corrected attention coefficients. : ; in, for Time Side The weight value; The corrected attention coefficients are normalized using the Softmax function to obtain the first attention weights. : ; in, It is an exponential function; Belonging to Nodes; For nodes For nodes The attention coefficient. The numerator is the attention coefficient modified by mapping it to a positive number through an exponential function and amplifying the differences between attention coefficients. The denominator is the sum of the attention coefficients at each node. All adjacent nodes centered on the center (including nodes) (It itself) calculates the corresponding exponential function And sum them.

[0033] In a preferred embodiment of the present invention, the first attention weight is weighted and fused with the features of neighboring nodes, then processed by a nonlinear function, and a preset residual term is superimposed. The output node embedding includes: ; in, This is the node embedding output after superimposing preset residual terms; for Time Node The preset residual term; For nodes For nodes Attention weights; for Time Node The power characteristic vector; It is a nonlinear function.

[0034] By introducing residual terms, gradients can be more smoothly backpropagated to the input layer, avoiding gradient vanishing and improving the stability and convergence speed of model training.

[0035] In a preferred embodiment of the present invention, the embedding of the first node is corrected by fusing the first vector set from the previous time step using a moving average, resulting in the first vector set for the current time step, which includes: The following adjustments were made to each node: ; in, for The current time node The first vector; for The previous time node The first vector is initialized to 0 when t is 0; for The current time node The first node is embedded; This is a correction factor; The first vector set at this moment is obtained based on the first vector of each node at this moment.

[0036] In a preferred embodiment of the present invention, the correction coefficient include: ; in, for The sum of edge weights in a graph-based power grid at any given time. for The sum of edge weights in a graph-based power grid at any given time.

[0037] In a preferred embodiment of the present invention, a correction coefficient is introduced to extend the model from a static GNN to a dynamic GNN. Reflects from Time's up Time map The changes in the edges.

[0038] F3. Based on the first vector set and parameter set at this moment, the second node embedding is obtained by combining the graph attention network; the second node embedding is corrected by fusing the second vector set from the previous moment through moving average, and the second vector set at this moment is obtained.

[0039] In a preferred embodiment of the present invention, the second node embedding is obtained by combining the first vector set and parameter set at the current moment with a graph attention network. Specifically, this involves processing the first vector set and parameter set through a second graph attention network layer and outputting the second node embedding; including: The first vector set is added to the node power feature vector set to obtain the optimized node power feature vector set; the optimized node power feature vector set and the adjacency matrix are input into the graph attention network for the first processing to obtain the second node embedding; when performing the first processing based on the optimized node power feature vector set and the adjacency matrix, the preset residual term includes the first vector set.

[0040] In a preferred embodiment of the present invention, the embedding of the second node is corrected by fusing the second vector set from the previous time step using a moving average, resulting in the second vector set for the current time step, which includes: The following adjustments were made to each node: ; in, for The current time node The second vector; for The previous time node The second vector is initialized to 0 when t is 0; for The current time node The second node embedding; The second vector set at this moment is obtained based on the second vector of each node at this moment.

[0041] F4. Based on the second vector set at this time step and combined with the fully connected neural network, obtain the quantification result and ranking of the node sensitivity at this time step. Specifically, this includes: The second vector set is input into the multilayer perceptron, and the quantification result of the node sensitivity is obtained through the Softmax function, while retaining the sorting with a preset number of bits m; the upper limit of the preset number of bits m is the minimum number of nodes in the power grid under graph theory.

[0042] In subsequent actual operation, the quantitative results and ranking of real-time node sensitivity are continuously obtained according to the method of the present invention, providing data support for the assessment of key nodes in the power grid, improving the accuracy and timeliness of the assessment of key nodes in the power grid, and enhancing the safe operation guarantee capability of the new power system under dynamic topology.

[0043] This invention presents a power grid sensitive node identification method based on graph attention networks. By introducing a graph attention mechanism with edge weights, the graph neural network can selectively fuse actual physical parameters such as power system line resistance and real-time power flow values. This accurately captures the differentiated interaction relationships between different types of nodes, such as hub substations and distributed renewable energy sources, and lines operating in different environments. This avoids the neglect of line physical characteristics by traditional unweighted graph models, improving the physical correlation and accuracy of node feature representation. By continuously fusing node embeddings from the historical operating states of the power system, the static graph is expanded into a dynamic graph, accurately reflecting the dynamic evolution characteristics of nodes over time and adapting to changes in temporal operating states such as power system load fluctuations, equipment switching, and fault transients. Simultaneously, the rolling update mechanism preserves the temporal dependencies of historical operating states, overcoming the limitation of static graph models in failing to depict node state migration over time, and ensuring the continuity and integrity of node features in the temporal dimension. This invention's method tracks the spatiotemporal migration of sensitive nodes using temporal topology data and real-time power grid operating data, enabling rapid and dynamic identification of sensitive nodes in the power grid.

[0044] In a preferred embodiment of the present invention, a power grid sensitive node identification device based on graph attention network is also provided for use with the method of the present invention. The device includes a first module, a second module, a third module and a fourth module. The first module is used to obtain the parameter set of the power grid at this moment based on graph theory; the parameter set includes the adjacency matrix and the set of node power feature vectors; The second module is used to obtain the first node embedding based on the parameter set and the graph attention network; the first node embedding is corrected by fusing the first vector set of the previous time step with a moving average to obtain the first vector set of the current time step. The third module is used to obtain the second node embedding by combining the first vector set and parameter set at this time with the graph attention network; the second node embedding is corrected by fusing the second vector set at the previous time through moving average to obtain the second vector set at this time. The fourth module is used to obtain the quantification and ranking of the node sensitivity at this time by combining the second vector set at this time with the fully connected neural network.

[0045] The power grid sensitive node identification device based on graph attention network of the present invention, used in the method of the present invention, has the same beneficial effects as the method of the present invention.

[0046] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for power grid sensitive node identification based on graph attention network, characterized in that, include: The parameter set of the power grid at this moment is obtained based on graph theory; the parameter set includes an adjacency matrix and a set of node power feature vectors. The first node embedding is obtained by combining the parameter set with the graph attention network; the first node embedding is corrected by fusing the first vector set of the previous time step with a moving average to obtain the first vector set of the current time step. The second node embedding is obtained by combining the first vector set and the parameter set at this moment with a graph attention network; the second node embedding is corrected by fusing the second vector set from the previous moment with a moving average to obtain the second vector set at this moment; the quantification result and ranking of the node sensitivity at this moment are obtained by combining the second vector set at this moment with a fully connected neural network.

2. The power grid sensitive node identification method based on a graph attention network according to claim 1, characterized in that, The first node embedding is obtained by combining the parameter set with the graph attention network, including: The parameter set is input into the graph attention network and subjected to a first process to obtain the first node embedding; the first process includes: After unifying the dimensions of the power feature vectors of each node by using a weight matrix, the attention coefficients between each node and all neighboring nodes are calculated by combining the adjacency matrix to obtain the initial attention coefficients. The initial attention coefficients are corrected based on the edge weights to obtain corrected attention coefficients; the corrected attention coefficients are then normalized using the Softmax function to obtain the first attention weights. The first attention weight is weighted and fused with the features of the neighboring nodes, then processed by a nonlinear function, and a preset residual term is superimposed to output the node embedding. When performing the first processing based on the parameter set, the preset residual term includes the power feature vector of the node itself.

3. The power grid sensitive node identification method based on graph attention network according to claim 2, characterized in that, Calculate the attention coefficients between each node and all its neighboring nodes to obtain the initial attention coefficients, including: For nodes compute nodes For nodes Attention coefficient , Represents a node The set of neighboring nodes, the attention coefficient Represented as: ； in, The graph attention parameter vector is represented by T; T represents the transpose operation. Indicates feature splicing; Represents the linear mapping matrix of weights; and They are respectively Time Node and nodes The power characteristic vector.

4. The power grid sensitive node identification method based on graph attention network according to claim 3, characterized in that, The initial attention coefficients are corrected based on the edge weights to obtain the corrected attention coefficients; The corrected attention coefficients are normalized using the Softmax function to obtain the first attention weights, which include: The initial attention coefficients are corrected based on the edge weights to obtain the corrected attention coefficients. : ； in, for Time Side The weight value; The corrected attention coefficients are normalized using the Softmax function to obtain the first attention weights. : ； in, It is an exponential function; Belonging to Nodes; For nodes For nodes Attention coefficient.

5. The power grid sensitive node identification method based on graph attention network according to claim 4, characterized in that, The first attention weights are weighted and fused with the features of neighboring nodes, then processed by a nonlinear function, and a preset residual term is superimposed. The output node embedding includes: ； in, This is the node embedding output after superimposing preset residual terms; for Time Node The preset residual term; For nodes For nodes Attention weights; for Time Node The power characteristic vector.

6. The power grid sensitive node identification method based on graph attention network according to claim 5, characterized in that, The embedding of the first node is corrected by fusing the first vector set from the previous time step using a moving average, resulting in the first vector set for the current time step, which includes: The following adjustments were made to each node: ； in, for The current time node The first vector; for The previous time node The first vector is initialized to 0 when t is 0; for The current time node The first node is embedded; This is a correction factor; The first vector set at this moment is obtained based on the first vector of each node at this moment.

7. The power grid sensitive node identification method based on graph attention network according to claim 6, characterized in that, The correction coefficient include: ； in, for The sum of edge weights in a graph-based power grid at any given time. for The sum of edge weights in a graph-based power grid at any given time.

8. The power grid sensitive node identification method based on graph attention network according to claim 7, characterized in that, The second node embedding is obtained by combining the first vector set and the parameter set at this moment with the graph attention network, including: The first vector set is added to the node power feature vector set to obtain an optimized node power feature vector set; the optimized node power feature vector set and the adjacency matrix are input into the graph attention network for the first processing to obtain the second node embedding; when performing the first processing based on the optimized node power feature vector set and the adjacency matrix, the preset residual term includes the first vector set.

9. The method for identifying power grid sensitive nodes based on graph attention networks according to claim 8, characterized in that, The embedding of the second node is corrected by fusing the second vector set from the previous time step using a moving average, resulting in the second vector set for the current time step, which includes: The following adjustments were made to each node: ； in, for The current time node The second vector; for The previous time node The second vector is initialized to 0 when t is 0; for The current time node The second node embedding; The second vector set at this moment is obtained based on the second vector of each node at this moment.

10. A power grid sensitive node identification device based on graph attention network, used in the method according to any one of claims 1 to 9, characterized in that, The device includes a first module, a second module, a third module, and a fourth module; The first module is used to obtain the parameter set of the power grid at the current moment based on graph theory; the parameter set includes an adjacency matrix and a set of node power feature vectors; The second module is used to obtain the first node embedding based on the parameter set and the graph attention network; and to correct the first node embedding by fusing the first vector set of the previous time step with a moving average to obtain the first vector set of the current time step. The third module is used to obtain the second node embedding by combining the first vector set at the current moment and the parameter set with the graph attention network; and to correct the second node embedding by fusing the second vector set at the previous moment through moving average to obtain the second vector set at the current moment. The fourth module is used to obtain the quantification result and ranking of the node sensitivity at this moment by combining the second vector set at this moment with the fully connected neural network.

Images