Self-adaptive power grid weak node detection method based on dynamic implicit topology reasoning

Abstract

The invention discloses an adaptive power grid weak node detection method based on dynamic implicit topology reasoning, and the method comprises the steps: obtaining power grid network topology data, and obtaining graph data # imgabs0 #; a nonlinear transformation layer is added before GAT, a parameter weight matrix of the nonlinear transformation layer is decoupled into a two-factor implicit adjacency tensor, and node high-order features and dynamic association information are aggregated; inputting node high-order features obtained through a nonlinear transformation layer into GAT, updating node representation through multi-layer message transmission, and generating an initial soft clustering probability matrix; constructing a joint optimization loss function, and performing iterative optimization training on the GAT network added with the nonlinear transformation layer; and calculating a node membership entropy value, and judging a weak node according to the entropy value. According to the method, the problems of node judgment omission and transient risk level misjudgment caused by insufficient dynamic topology tracking capability and poor cross-regional weak node identification effect caused by node membership fuzziness constraint deficiency in the prior art are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system security protection, and in particular relates to an adaptive power grid weak node detection method based on dynamic implicit topology reasoning. Background Art

[0002] With the continuous development of power systems and the widespread integration of renewable energy sources, power grid structures are becoming increasingly complex. Grid failures not only severely disrupt people's daily lives and production processes but can also trigger widespread power outages, resulting in significant economic losses. During grid operation, the weaker the nodes in the topology, the more significant their impact on the overall system stability and the greater the likelihood of triggering system-level collapse under sudden disturbances. Therefore, accurately identifying weak nodes in the grid during normal operation is of great theoretical and practical significance.

[0003] However, with the rapid development and large-scale access of new energy, the topology of the power network has shown dynamic and time-varying characteristics, and the complexity of electrical correlations between nodes has increased significantly. Most existing weak node detection methods are based on static topology analysis, which makes it difficult to capture the high-order interactions and time-series evolution characteristics between nodes, and cannot effectively capture the potential risk characteristics in dynamic operation scenarios. They have the following defects: (1) Traditional static detection methods rely on fixed topology and steady-state data, while the access of new energy leads to frequent dynamic changes in the grid topology and significant complication of transient characteristics (such as voltage instability and frequency fluctuations). Such methods cannot synchronously track dynamic topological changes, nor can they capture the changing characteristics of risk propagation in complex transient processes, resulting in missed weak node detection and misjudgment of transient risk levels under specific operating conditions; (2) Traditional methods use a single indicator (such as voltage sensitivity) to rank vulnerabilities, ignoring the amplification effect of the community structure in the power grid on node risk propagation. When dividing subnetworks and detecting communities, they directly rely on the original feature space, making it difficult to capture the complex high-order interactions and cross-regional electrical coupling relationships between nodes. In addition, the existing community division method does not consider the ambiguity of node membership, that is, under dynamic topology, a node may belong to multiple communities at the same time, and its membership probability distribution is chaotic, resulting in the missed detection of cross-community weak nodes (such as "bridge nodes"); (3) The existing method does not constrain the uncertainty of node membership during the community division process, and the node membership entropy value is not included in the optimization target, resulting in the model being unable to distinguish between real cross-community nodes and instantaneous fuzzy nodes caused by noise, reducing the reliability of weak node detection. Summary of the Invention

[0004] The purpose of the present invention is to provide an adaptive power grid weak node detection method based on dynamic implicit topology reasoning, which solves the problems in the prior art of insufficient dynamic topology tracking capability leading to missed nodes and misjudgment of transient risk levels, as well as poor cross-region weak node identification effect caused by the lack of fuzzy constraints on node affiliation.

[0005] The technical solution adopted by the present invention is an adaptive power grid weak node detection method based on dynamic implicit topology reasoning, which specifically includes the following steps: Step 1: Obtain the power grid network topology data and construct an adjacency matrix using the edge list , Use node attributes to build feature matrix , get graph data ; Step 2: Dynamic implicit topology reasoning: Add a nonlinear transformation layer before the GAT and decouple the nonlinear transformation layer parameter weight matrix into a two-factor implicit adjacency tensor to aggregate node high-order features and dynamic association information; Step 3: Community partitioning: Input the node high-order features obtained by the nonlinear transformation layer into GAT, update the node representation through multi-layer message passing, and generate the initial soft clustering probability matrix; Step 4: Construct a joint optimization loss function and perform iterative optimization training on the GAT network with the added nonlinear transformation layer; Step 5: Calculate the node membership entropy value and determine the weak nodes based on the entropy value.

[0006] The present invention is also characterized in that: In step 2, a nonlinear transformation layer is added before GAT. The nonlinear transformation layer is a fully connected layer. For the feature matrix of the input graph data Processing is performed, which is expressed as: (1) Where, is a high-order feature, is the activation function, is the weight matrix, is the bias vector; Weight Matrix It is a learnable implicit structure generation basis, which is further decoupled into: (2) Where U and V encode the potential association between nodes and the correlation of feature dimensions, respectively, forming a two-factor implicit adjacency tensor.

[0007] Step 3 specifically includes the following steps: Step 3.1: Input the high-order feature Y into GAT and update the node representation through multi-layer message passing, which is expressed as: (3) Where, Indicates the The node characteristics of the layer, ; Step 3.2: Normalize the final node features through the softmax function and output the soft clustering assignment matrix, which is expressed as: (4) In the formula, the soft clustering assignment matrix is ​​generated , n is the total number of nodes, k is the number of potential clusters, the soft clustering assignment matrix Elements in Representation node i Belong to cluster k probability; is the clustering weight matrix, Represents the final node features obtained.

[0008] Construct a joint optimization loss function, expressed as: (5) Where, represents the joint optimization loss function, represents the modularity optimization item, Indicates the module resolution control item, represents the entropy regularization term.

[0009] Modularity optimization items ,make , for modularity Matrix representation is as follows: (6) Where, is the cluster assignment matrix, Representation node Belong to the same community, Representation node Not belonging to the same community, That is, the sum of the elements in B that belong to the same sub-region.

[0010] Module resolution control , which is expressed as follows: (7) Where, is the F-form, n is the total number of nodes, and k is the number of potential clusters.

[0011] Entropy regularization term , which is expressed as follows: (8) Where, The membership entropy value of each node is calculated as follows: (9).

[0012] Step 5 is as follows: The membership entropy value of each node is calculated according to formula (9), and the entropy values ​​are sorted from large to small, and the top 5% to 10% of nodes are determined to be weak nodes.

[0013] The beneficial effects of the present invention are: The present invention is an adaptive power grid weak node detection method based on dynamic implicit topology reasoning. A nonlinear feature encoder is embedded in community detection. Through implicit topology reasoning, the temporal evolution characteristics of the new energy power grid are captured. The potential associations of power grid nodes are explicitly encoded through a two-factor implicit adjacency tensor. At the same time, the entropy weight model is used to quantify the node vulnerability level. This effectively overcomes the problems of traditional static detection methods in missing key vulnerable nodes and misjudging transient risk levels in time-varying topology scenarios, providing a decision-making basis for power grid security protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a flow chart of the adaptive power grid weak node detection method based on dynamic implicit topology reasoning of the present invention; Figure 2 It is a schematic diagram of the network structure of adding a fully connected layer to the multi-layer neural network in the adaptive power grid weak node detection method based on dynamic implicit topology reasoning of the present invention. DETAILED DESCRIPTION

[0015] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0016] The present invention is based on the adaptive power grid weak node detection method of dynamic implicit topology reasoning, such as Figure 1 As shown, the specific steps include: Step 1: Obtain the power grid network topology data and construct an adjacency matrix using the edge list , Use node attributes to build feature matrix , get graph data . Feature Matrix Including dynamic parameters such as voltage, power, and new energy output.

[0017] Step 2: Dynamic implicit topology reasoning: Add a nonlinear transformation layer before GAT and decouple the nonlinear transformation layer parameter weight matrix into a two-factor implicit adjacency tensor to aggregate node high-order features and dynamic association information.

[0018] The Graph Attention Network (GAT) has a linear property, meaning that information propagates across neighbors as a linear combination. To enhance nonlinear modeling capabilities, a nonlinear transformation layer is added before the GAT. By introducing a two-factor implicit adjacency tensor, the weight parameter matrix of the original graph structure is decoupled into two low-rank matrices, U and 𝑉. This allows for joint modeling of the dynamic topological changes of the power grid and the high-order semantic associations between nodes. This process maps features into a higher-dimensional latent space, significantly enhancing the distinguishability of node features and the latent semantic space, providing a richer feature representation for subsequent community delineation and weak node detection.

[0019] Add a nonlinear transformation layer before GAT, and the nonlinear transformation layer is a fully connected layer, such as Figure 2 As shown; for the feature matrix of the input graph data Processing is performed, which is expressed as: (1) Where, is a high-order feature, is the activation function, is the weight matrix, is the bias vector; Weight Matrix It is a learnable implicit structure generation basis, which is further decoupled into: (2) Where U and V encode the potential association between nodes and the correlation of feature dimensions, respectively, forming a two-factor implicit adjacency tensor. This decomposition enables the model to explicitly learn the implicit structure of the graph, thereby better utilizing the global topological information of the graph.

[0020] Step 3: Community partitioning: The high-order features of the nodes obtained through the nonlinear transformation layer are input into GAT, the node representation is updated through multi-layer message passing, and the initial soft clustering probability matrix is ​​generated.

[0021] The specific steps include: Step 3.1: Input the high-order feature Y into GAT and update the node representation through multi-layer message passing, which is expressed as: (3) Where, Indicates the The node characteristics of the layer, ; Step 3.2: Normalize the final node features through the softmax function and output the soft clustering assignment matrix, which is expressed as: (4) In the formula, the soft clustering assignment matrix is ​​generated , n is the total number of nodes, k is the number of potential clusters (usually set to the number of nodes), the soft clustering assignment matrix Elements in Representation node i Belong to cluster k probability; is the clustering weight matrix, Represents the final node features obtained.

[0022] Step 4: Construct a joint optimization loss function and perform iterative optimization training on the GAT network with the added nonlinear transformation layer.

[0023] The constructed joint optimization function includes modularity optimization term, module resolution control term and entropy regularization term, which is expressed as follows: (5) Where, represents the joint optimization loss function, represents the modularity optimization item, Indicates the module resolution control item, represents the entropy regularization term. The modularity optimization term is used to detect community structure, the resolution suppression term is used to prevent large modules from merging surrounding small modules, and the entropy regularization term is used to reduce the uncertainty of node membership.

[0024] Specifically, the modularity optimization term ,make , for modularity Matrix representation is as follows: (6) Where, is the cluster assignment matrix, Representation node Belong to the same community, Representation node Not belonging to the same community, That is, the sum of the elements in B that belong to the same sub-region.

[0025] To constrain the size of the module, the cluster allocation matrix Constraints are made to obtain module resolution control items , which is expressed as follows: (7) Where, is the F paradigm, n is the total number of nodes, and k is the number of potential clusters. This term is used to prevent large communities from merging into small communities and to constrain the size of the community.

[0026] To minimize the uncertainty of node membership, the entropy regularization term is set , which is expressed as follows: (8) Where, The membership entropy value of each node is calculated as follows: (9).

[0027] Entropy regularization term The degree of confusion in node membership distribution is reduced.

[0028] During training, we set the learning rate to 0.001, set an appropriate dropout rate, train for 1000 epochs, and update the network parameters after each epoch.

[0029] Step 5: Calculate the node membership entropy value and determine the weak nodes based on the entropy value.

[0030] According to the output of the soft clustering assignment matrix of the optimized GAT network with the added nonlinear transformation layer, the membership entropy value of each node is calculated according to formula (9), and the entropy values ​​are sorted from large to small, and the top 5% to 10% of the nodes are determined to be weak nodes.

[0031] Entropy A larger entropy value indicates that the node is often located at the intersection of different communities. It has connections to multiple communities, but none of them are strong, meaning it acts as a "bridge" or "buffer" within the network. Based on the entropy ranking, the top 5% of nodes are considered weak nodes. Failure of such nodes can disrupt connections between multiple subnetworks, hindering the flow of information and energy between local areas of the power grid. Therefore, they are considered potential weak nodes.

[0032] The method of the present invention is an entropy constraint-driven community partition optimization. It constructs an entropy model based on the node's membership information in the community, and uses the uncertainty of the node community's affiliation to measure its vulnerability. In the community partition stage, in addition to designing a joint modularity and resolution suppression term as a loss function, an entropy regularization term is added to constrain the size of the entropy value, reduce the uncertainty of the node membership, jointly optimize the clustering distribution and entropy value of the node, and judge that the node with a higher entropy value is a weak node in the power network. It may assume functions such as cross-community connection and edge support in the power network. Once a failure occurs, it may cause collaborative instability among multiple communities. Based on the entropy weight, this method dynamically evaluates the probability of multi-community instability caused by node failure, outputs a risk level ranking, realizes the collaborative modeling between the power grid topology evolution characteristics and node vulnerability, and improves the accuracy and adaptability of dynamic power network weak node detection.

[0033] Example 1 This embodiment provides an adaptive power grid weak node detection method based on dynamic implicit topology reasoning, which specifically includes the following steps: Step 1: Obtain the power grid network topology data and construct an adjacency matrix using the edge list , Use node attributes to build feature matrix , get graph data ; Step 2: Dynamic implicit topology reasoning: Add a nonlinear transformation layer before the GAT and decouple the nonlinear transformation layer parameter weight matrix into a two-factor implicit adjacency tensor to aggregate node high-order features and dynamic association information; Step 3: Community partitioning: Input the node high-order features obtained by the nonlinear transformation layer into GAT, update the node representation through multi-layer message passing, and generate the initial soft clustering probability matrix; Step 4: Construct a joint optimization loss function and perform iterative optimization training on the GAT network with the added nonlinear transformation layer; Step 5: Calculate the node membership entropy value and determine the weak nodes based on the entropy value.

[0034] Example 2 On the basis of Example 1, in step 2, a nonlinear transformation layer is added before GAT. The nonlinear transformation layer is a fully connected layer. For the feature matrix of the input graph data Processing is performed, which is expressed as: (1) Where, is a high-order feature, is the activation function, is the weight matrix, is the bias vector; Weight Matrix It is a learnable implicit structure generation basis, which is further decoupled into: (2) Where U and V encode the potential association between nodes and the correlation of feature dimensions, respectively, forming a two-factor implicit adjacency tensor.

[0035] Example 3 Based on Example 2, step 3 specifically includes the following steps: Step 3.1: Input the high-order feature Y into GAT and update the node representation through multi-layer message passing, which is expressed as: (3) Where, Indicates the The node characteristics of the layer, ; Step 3.2: Normalize the final node features through the softmax function and output the soft clustering assignment matrix, which is expressed as: (4) In the formula, the soft clustering assignment matrix is ​​generated , n is the total number of nodes, k is the number of potential clusters, the soft clustering assignment matrix Elements in Representation node i Belong to cluster k probability; is the clustering weight matrix, Represents the final node features obtained.

[0036] Example 4 Based on Example 3, a joint optimization loss function is constructed, which is expressed as: (5) Where, represents the joint optimization loss function, represents the modularity optimization item, Indicates the module resolution control item, represents the entropy regularization term.

[0037] Example 5 Based on Example 4, the modularity optimization item ,make , for modularity Matrix representation is as follows: (6) Where, is the cluster assignment matrix, Representation node Belong to the same community, Representation node Not belonging to the same community, That is, the sum of the elements in B that belong to the same sub-region.

[0038] Module resolution control , which is expressed as follows: (7) Where, is the F-form, n is the total number of nodes, and k is the number of potential clusters.

[0039] Entropy regularization term , which is expressed as follows: (8) Where, The membership entropy value of each node is calculated as follows: (9).

[0040] Example 6 Based on Example 5, step 5 is specifically as follows: The membership entropy value of each node is calculated according to formula (9), and the entropy values ​​are sorted from large to small, and the top 5% to 10% of nodes are determined to be weak nodes.

Claims

1. An adaptive power grid weak node detection method based on dynamic implicit topology reasoning, characterized by: The specific steps include: Step 1: Obtain the power grid network topology data and construct an adjacency matrix using the edge list , Use node attributes to build feature matrix , get graph data ; Step 2: Dynamic implicit topology reasoning: Add a nonlinear transformation layer before the GAT and decouple the nonlinear transformation layer parameter weight matrix into a two-factor implicit adjacency tensor to aggregate node high-order features and dynamic association information; Step 3: Community partitioning: Input the node high-order features obtained by the nonlinear transformation layer into GAT, update the node representation through multi-layer message passing, and generate the initial soft clustering probability matrix; Step 4: Construct a joint optimization loss function and perform iterative optimization training on the GAT network with the added nonlinear transformation layer; Step 5: Calculate the node membership entropy value and determine the weak nodes based on the entropy value.

2. The adaptive power grid weak node detection method based on dynamic implicit topology reasoning according to claim 1 is characterized in that: In step 2, a nonlinear transformation layer is added before GAT. The nonlinear transformation layer is a fully connected layer. For the feature matrix of the input graph data Processing is performed, which is expressed as: （1） Where, is a high-order feature, is the activation function, is the weight matrix, is the bias vector; Weight Matrix It is a learnable implicit structure generation basis, which is further decoupled into: （2） Where U and V encode the potential association between nodes and the correlation of feature dimensions, respectively, forming a two-factor implicit adjacency tensor.

3. The adaptive power grid weak node detection method based on dynamic implicit topology reasoning according to claim 2 is characterized in that: Step 3 specifically includes the following steps: Step 3.1: Input the high-order feature Y into GAT and update the node representation through multi-layer message passing, which is expressed as: （3） Where, Indicates the The node characteristics of the layer, ; Step 3.2: Normalize the final node features through the softmax function and output the soft clustering assignment matrix, which is expressed as: （4） In the formula, the soft clustering assignment matrix is ​​generated , n is the total number of nodes, k is the number of potential clusters, the soft clustering assignment matrix Elements in Representation node i Belong to cluster k probability; is the clustering weight matrix, Represents the final node features obtained.

4. The adaptive power grid weak node detection method based on dynamic implicit topology reasoning according to claim 3 is characterized in that: Construct a joint optimization loss function, expressed as: （5） Where, represents the joint optimization loss function, represents the modularity optimization item, Indicates the module resolution control item, represents the entropy regularization term.

5. The adaptive power grid weak node detection method based on dynamic implicit topology reasoning according to claim 4 is characterized in that: Modularity optimization items ,make , for modularity Matrix representation is as follows: （6） Where, is the cluster assignment matrix, Representation node Belong to the same community, Representation node Not belonging to the same community, That is, the sum of the elements in B that belong to the same sub-region.

6. The adaptive power grid weak node detection method based on dynamic implicit topology reasoning according to claim 5 is characterized in that: Module resolution control , which is expressed as follows: （7） Where, is the F-form, n is the total number of nodes, and k is the number of potential clusters.

7. The adaptive power grid weak node detection method based on dynamic implicit topology reasoning according to claim 6 is characterized in that: Entropy regularization term , which is expressed as follows: （8） Where, The membership entropy value of each node is calculated as follows: （9）。 8. The adaptive power grid weak node detection method based on dynamic implicit topology reasoning according to claim 7 is characterized in that: Step 5 is as follows: The membership entropy value of each node is calculated according to formula (9), and the entropy values ​​are sorted from large to small, and the top 5% to 10% of nodes are determined to be weak nodes.

Images