Generative information physical power system topology construction method

Abstract

The invention relates to the technical field of power system planning and modeling, in particular to a generative information physical power system topology construction method, which comprises the following steps: representing a power system and a communication network as a double-layer network structure, and establishing an input network model; performing serialization generation on the topological structure by adopting a recurrent neural network module, and outputting a generative adjacency matrix; mapping an input adjacent matrix and an attribute set of all nodes to a hidden space by using a variational auto-encoder module, and outputting generative node attributes; establishing a joint loss function, comprehensively considering an adjacent matrix reconstruction error, a hidden space regularization item and a network feature distribution difference item, and performing joint optimization on a variational auto-encoder module; and fusing and iterating the generative adjacency matrix and the generative node attributes through a network feature reconstruction module to obtain generative edge attributes meeting constraint conditions. Through the construction method, the synthetic topology which is consistent with a real power system in statistical characteristics but does not leak sensitive information can be obtained.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system planning and modeling, in particular to a generative information-physical power system topology construction method. BACKGROUND

[0002] With the increasing digitalization of modern power systems, power systems gradually evolve into typical network-physical coupled complex systems, i.e. cyber-physical power systems (CPS). In such systems, power transmission networks and communication networks are tightly coupled, and their operating characteristics and security are deeply influenced by network structure. In order to carry out reliable operation analysis and security assessment, it is necessary to have a network model consistent with the characteristics of the real system. However, the topology structure and parameter information of the real power system often belong to sensitive data, and direct disclosure may cause system security risks.

[0003] Currently, the research on generative power networks mainly focuses on the construction of statistical characteristics of power topology, such as degree distribution, average path length, network density and other complex network parameters. However, such methods often ignore the inherent properties of nodes and edges, such as communication link bandwidth, transmission line capacity and other key characteristics, making it difficult for the generated network to fully reflect the operating properties of the real system. At the same time, existing researches pay more attention to the modeling of the physical layer of power systems, but lack consideration of the coupling characteristics between the communication layer and the physical layer.

[0004] Existing literature proposes to use complex network theory to generate statistically realistic power topology, but its method is limited to structural matching, lacking modeling ability for node function type (such as generator, load, zero injection node) and edge attribute distribution, thus having great limitations in system security simulation and attack defense research. Especially in the background of increasing network attack and information leakage risk, the existing method is difficult to meet the dual needs of "hiding sensitive information" and "maintaining system characteristics".

[0005] In the prior art, a Chinese patent document with the publication number CN120262400A and the publication date of July 4, 2025 is proposed. The technical solution disclosed in this patent document is as follows: A power distribution network grid topology identification method based on an improved graph neural network, the steps include: collecting real-time electrical quantity data of nodes and edges of the power distribution network, and modeling the power distribution network graph structure; constructing node, edge and super-edge feature matrices with real-time data collection, inputting the improved graph neural network topology identification model, dynamically weighting and adjusting the edge weight through the graph attention network, and splicing and fusing the local topology features and global topology features to generate a predicted adjacency matrix; when the power distribution network changes dynamically, locate the change area through the adjacency matrix difference, and extract the subgraph for incremental update; based on the final adjacency matrix, construct the topology structure of the power distribution network, and output the physical position of the node and the connection relationship of the edge.

[0006] The above technical solution can only generate supplementary topology based on the existing topology according to the topology features, and cannot generate a new complete network topology, so it cannot realize the generation of the generated power system topology. Moreover, this technical solution realizes generation based on detailed data of the existing topology, so it cannot guarantee the data security of the input network topology, and there is a possibility of leaking critical facility data. This technical solution only describes the application method for the power distribution network, and cannot realize the generation of different types of networks, so it cannot generate a complete cyber-physical power system.

[0007] In the prior art, a Chinese patent document with the publication number CN118037143A and the publication date of May 14, 2024 is proposed. The technical solution disclosed in this patent document is as follows: A power distribution network state representation method based on graph mask reconstruction, comprising the following steps: S1: modeling the topology structure of the power distribution network as a heterogeneous graph; S2: reconstructing the path through the path mask predictor; S3: reconstructing the node attribute through the node attribute mask predictor; S4: optimizing the model.

[0008] The mask reconstruction mechanism of the above technical solution mainly aims at the recovery of local topology and node attributes, and lacks the modeling ability of global statistical features of the network. Moreover, this method cannot generate a new network structure while keeping the consistency of node types, edge attributes and degree distribution, so it is difficult to be used for safe modeling of the generated topology of the power system. In addition, since the mask prediction model relies on a large number of labeled samples for supervised training, the generalization performance is reduced under the condition of data scarcity or limitation. SUMMARY

[0009] To solve the above technical problems, the application provides a generative information physical power system topology construction method, which uses a generative framework combining a recurrent neural network and a variational autoencoder, can learn the feature distribution of a real power system, directly generate a complete information physical power system topology and node attributes, and obtain a synthetic topology consistent with the real power system in statistical characteristics but not leaking sensitive information.

[0010] The application is implemented by adopting the following technical scheme: A generative information physical power system topology construction method, comprising the following steps: Step S1. Representing the power system and the communication network as a double-layer network structure, extracting the adjacency matrix, node attributes and edge attributes respectively, and establishing an input network model ; wherein, A represents the adjacency matrix; X represents the attribute set of all nodes, E represents the attribute set of all edges; Step S2. Using a recurrent neural network module to serialize the topology structure and output a generative adjacency matrix ; using a variational autoencoder module to map the input adjacency matrix A and the attribute set of all nodes X to a hidden space, generate the type feature of the node and the network feature distribution of the node, and output the generative node attribute ; Step S3. Establishing a joint loss function, comprehensively considering the adjacency matrix reconstruction error, the hidden space regularization term and the network feature distribution difference term, and jointly optimizing the variational autoencoder module; Step S4. Fusing the generative adjacency matrix and the generative node attribute through a network feature reconstruction module, iteratively updating and deriving the edge attribute, obtaining the generative edge attribute satisfying the constraint condition, and obtaining the generative information physical power system network.

[0011] In the input network model, , ; wherein, is the type feature of the node, is the network feature of the node; represents the capacity of the edge; the network feature of the node is defined as the sum of the capacities of the adjacent edges: , wherein, represents the edge set connected with the node i .

[0012] Step S1 includes the following steps: Step S 11 The communication network includes a set of substation nodes and a set of communication links in the communication layer, and the power system includes a set of substation nodes and a set of power transmission lines in the physical layer. All nodes and edges in the power system and the communication network are numbered respectively, and a cross-layer mapping relationship is established so that each substation node in the physical layer is associated with a corresponding substation node in the communication layer, thus constructing a two-layer network structure of power-communication coupling. Step S 12 Based on step S 11 The established two-layer network structure is used to extract its adjacency matrix, node attributes, and edge attributes to construct the input network model. .

[0013] The recurrent neural network module includes node-level recurrent neural networks and edge-level recurrent neural networks that work together in a nested manner. A complete network connection structure is gradually constructed through sequential recursion, ultimately yielding a generative adjacency matrix. .

[0014] The sequential recursion specifically refers to the following: the node-level recurrent neural network is responsible for maintaining the global state of the graph during the sequential node generation process. Whenever a new node is generated, the edge-level recurrent neural network is called to predict the connection relationship between the node and all previous nodes. The edge-level recurrent neural network generates the existence or non-existence of edges step by step under the condition of the current hidden state of the node-level recurrent neural network, and determines the adjacency vector of the new node. Subsequently, the node-level recurrent neural network updates its hidden state according to the output of the edge-level recurrent neural network, providing contextual information for the generation of the next node.

[0015] The encoder of the variational autoencoder module uses an adjacency matrix. A and the set of attributes of all nodes X As input, a two-layer graph convolutional network is used to project the input into the latent space. Z Hidden Space Z Regularized by an isotropic Gaussian distribution; the decoder of the variational autoencoder module is derived from the latent space. Z Adjacency matrix obtained by inner product sampling As input, a two-layer graph convolutional network is used to output generative node attributes. .

[0016] A two-layer graph convolutional network is used to project the input into the latent space. Z Specifically refers to: ; In the formula, Represents encoder, is the latent space variable; the latent space is regularized by an isotropic Gaussian prior distribution: Output generation formula node attribute Specifically: wherein, is the decoder, is the adjacency matrix obtained by inner product sampling of the latent space Z , , , and are the type features of the generative node and the network features of the generative node, respectively.

[0017] The graph convolution network in the decoder of the variational autoencoder module is: , wherein, , are trainable parameters, , is a symmetric normalized adjacency matrix, is the degree matrix of .

[0018] Further comprising: converting the node attribute vector X in the attribute set x of the node into a probability distribution , using the Kullback-Leibler divergence to measure the difference between the network feature distribution of the real network and the network feature distribution of the generative network: , wherein, is the network feature distribution in the real network, is the feature distribution in the generative network.

[0019] The joint loss function is: , wherein, is the joint loss, is the adjacency matrix reconstruction error, is the latent space regularization term, is the network feature distribution difference term, is the KL divergence between and two distributions.

[0020] The step S4 specifically comprises the following steps: Step S 41 ​​​The recurrent neural network module and the variational autoencoder module output a generative adjacency matrix. and generative node attributes Then, the edge attributes Initialize to 0; Step S 42 For generative node attributes Sort in descending order; Step S 43 Based on generative adjacency matrix Arrange the nodes in descending order of degree; Step S 44 . Arrange the attributes of the generative nodes Assign values ​​to the sorted nodes in sequence; Step S 45 Locate the node with the minimum degree value; Step S 46 Find the adjacent edges of the node with the minimum degree value, and calculate the edge attributes of the adjacent edges based on the remaining node attribute values ​​of that node. And update the attribute values ​​of the remaining nodes. ; Step S 47 Repeat step S. 45 Step S 46 until all edge attributes >0; Generates the final edge attributes. .

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The construction method proposed in this invention utilizes a generative framework, Graph-CPS, which combines a recurrent neural network module, a variational autoencoder module, and a network feature reconstruction module. This framework can learn the statistical feature distribution of real power systems under unsupervised conditions and automatically generate cyber-physical power system topologies that satisfy multidimensional attribute constraints and structural consistency. This framework can directly generate complete network structures and node attributes, effectively maintaining a high degree of consistency between the generated topology and the real system in key statistical indicators such as node type distribution, edge attribute distribution, average path length, degree distribution, and network density, thereby ensuring the authenticity and engineering usability of the generated results.

[0022] Meanwhile, this generation framework possesses strong data privacy protection capabilities. Because the generation process does not rely on the explicit exposure of the original data, and the output is a synthetic topology rather than a direct mapping of the real system, coupled with the fact that power systems are highly sensitive to even minor differences in topology and parameters, attackers find it difficult to accurately correlate the generated results with the actual system, thus effectively preventing the leakage of sensitive information.

[0023] Furthermore, this invention achieves the collaborative generation of topology and node attributes by introducing network feature distribution constraints and a joint loss function optimization mechanism.

[0024] 2. This invention, through a two-layer network structure design, comprehensively considers the "partial one-to-one" coupling relationship between the communication layer and the power physical layer, which can truly reflect the interaction characteristics between the network and physical systems, realize collaborative modeling of the network layer and the physical layer, and improve the structural completeness of the generated model.

[0025] 3. The construction method of the present invention has good scalability and can be applied to power systems and communication networks of different scales and types, realizing cross-system and cross-scenario topology generation and simulation applications.

[0026] 4. In this invention, the recurrent neural network module includes a node-level recurrent neural network and an edge-level recurrent neural network that work together in a nested manner. Through the "node-edge" two-layer recursive structure, the graph recurrent neural network can effectively capture the global topological features and local connection patterns of the graph while maintaining the dependency of the generation sequence, thereby realizing the progressive generation of complex network structures.

[0027] 5. Compared to mask reconstruction or probabilistic inference methods that rely on a large number of labeled samples, this invention employs a joint architecture of graph neural networks and variational autoencoders to achieve unsupervised learning, significantly reducing the dependence on the scale of training data. By introducing a feature distribution regularization mechanism, the interpretability of the model and the stability of the generation process are improved, maintaining high fidelity and robustness even in scenarios with scarce training data, effectively overcoming the limitations of existing methods in preserving global statistical properties and data dependence.

[0028] 6. This invention utilizes a joint loss function to optimize the variational autoencoder model. Based on the basic variational autoencoder loss function, this invention incorporates the probability distribution of node attributes of the input and output networks into the loss function, which ensures that the model accurately learns the node attribute features of the input network.

[0029] 7. This invention achieves deep integration of node attributes and adjacency matrix in the generated topology through a network feature reconstruction module. Compared with traditional methods that rely on preset rules or table lookup assignment, this method simulates the large-capacity operation characteristics of hub nodes in real power systems through a heuristic mechanism that corresponds to high attribute values ​​for nodes with high height numbers, thereby enhancing the realism of the generated results in structure-function mapping.

[0030] Furthermore, the network feature reconstruction module ensures that all edge attributes satisfy fundamental physical constraints such as non-zero properties through an iterative optimization mechanism, effectively avoiding the generation of invalid or infeasible topologies. The final generated network maintains a high degree of consistency with the real system in terms of topology, node type distribution, and edge attribute statistical characteristics. Attached Figure Description

[0031] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments, wherein: Figure 1 This is a schematic diagram of the Graph-CPS generative framework of the present invention; Figure 2 This is a diagram illustrating a partial one-to-one correspondence between the power system and the communication network in this invention. Figure 3 This is a flowchart illustrating the network feature reconstruction module in this invention. Figure 4 This is a schematic diagram of the IEEE 39-node generative power network in this invention; Figure 5 This is a schematic diagram of the IEEE 39-node generative communication network in this invention; Figure 6 This is a schematic diagram of the IEEE 39-node generative cyber-physical power system in this invention; Figure 7 This is a schematic diagram of the node attributes of the generative power system in this invention; Figure 8 This is a schematic diagram of node attributes in the generative communication system of this invention. Detailed Implementation

[0032] Example 1 As a basic embodiment of the present invention, the present invention includes a generative cyber-physical power system topology construction method, comprising the following steps: Step S1. Represent the power system and communication network as a two-layer network structure, extract the adjacency matrix, node attributes, and edge attributes respectively, and establish the input network model. ;in, A Represents the adjacency matrix; X Represents the set of attributes for all nodes. E This represents the set of attributes for all edges.

[0033] Step S2. Use a recurrent neural network module to serialize and generate the topology, and output a generative adjacency matrix. The variational autoencoder module is used to process the input adjacency matrix. A and the set of attributes of all nodes X Mapping to the latent space, the type features of the generated nodes and the distribution of network features of the nodes are used to output generative node attributes. .

[0034] Step S3. Establish a joint loss function, taking into account the adjacency matrix reconstruction error, the latent space regularization term, and the network feature distribution difference term, and jointly optimize the variational autoencoder module.

[0035] Step S4. The generative adjacency matrix is ​​reconstructed using the network feature reconstruction module. With generative node attributes By fusing, iteratively updating, and deriving edge attributes, generative edge attributes that satisfy the constraints are obtained. To obtain a generative cyber-physical power system network.

[0036] Example 2 As a preferred embodiment of the present invention, the present invention includes a generative cyber-physical power system topology construction method, comprising the following steps: Step S1. Represent the power system and communication network as a two-layer network structure, extract the adjacency matrix, node attributes, and edge attributes respectively, and establish the input network model. .in, A Represents the adjacency matrix; X Represents the set of attributes for all nodes. E This represents the set of attributes for all edges. Specifically, the input network model... The method for establishing it includes the following steps: Step S 11 The communication network includes a set of substation nodes and a set of communication links in the communication layer, and the power system includes a set of substation nodes and a set of power transmission lines in the physical layer. All nodes and edges in the power system and communication network are numbered, and a cross-layer mapping relationship is established, so that each substation node in the physical layer is associated with a corresponding substation node in the communication layer, thus constructing a two-layer network structure coupled with power communication.

[0037] Step S 12 Based on step S 11 The established two-layer network structure is used to extract its adjacency matrix, node attributes, and edge attributes to construct the input network model. .

[0038] Step S2. Use a recurrent neural network module to serialize and generate the topology, and output a generative adjacency matrix. The variational autoencoder module is used to process the input adjacency matrix. A and the set of attributes of all nodes X Mapping to the latent space, the type features of the generated nodes and the distribution of network features of the nodes are used to output generative node attributes. .

[0039] Step S3. Establish a joint loss function, taking into account the adjacency matrix reconstruction error, the latent space regularization term, and the network feature distribution difference term, and jointly optimize the variational autoencoder module.

[0040] Step S4. The generative adjacency matrix is ​​reconstructed using the network feature reconstruction module. With generative node attributes By fusing, iteratively updating, and deriving edge attributes, generative edge attributes that satisfy the constraints are obtained. This leads to the generation of a cyber-physical power system. Specifically, the network feature reconstruction module assigns node attributes and reconstructs edge attributes. The core idea is to assign larger attribute values ​​to nodes with higher degrees. Through iterative updates, until all edge attributes satisfy non-zero constraints, a complete generative cyber-physical power system is finally obtained.

[0041] Example 3 In another preferred embodiment of the present invention, the present invention includes a generative cyber-physical power system topology construction method, comprising the following steps: Step S1. Represent the power system and communication network as a two-layer network structure, extract the adjacency matrix, node attributes, and edge attributes respectively, and establish the input network model. .in, A Represents the adjacency matrix; X Represents the set of attributes for all nodes. E This represents the set of attributes for all edges.

[0042] Step S2. Use a recurrent neural network module to serialize and generate the topology, and output a generative adjacency matrix. The recurrent neural network module includes node-level recurrent neural networks and edge-level recurrent neural networks that work together in a nested manner, and gradually builds a complete network connection structure through sequential recursion.

[0043] The input adjacency matrix is ​​processed using a variational autoencoder module. A and the set of attributes of all nodes X Mapping to the latent space, the type features of the generated nodes and the distribution of network features of the nodes are used to output generative node attributes. Specifically, the encoder of the variational autoencoder module uses an adjacency matrix. A and the set of attributes of all nodes X As input, a two-layer graph convolutional network is used to project the input into the latent space. Z Hidden Space Z Regularized by an isotropic Gaussian distribution. The decoder of the variational autoencoder module is derived from the latent space. Z Adjacency matrix obtained by inner product sampling As input, a two-layer graph convolutional network is used to output generative node attributes. .

[0044] Step S3. Establish a joint loss function, taking into account the adjacency matrix reconstruction error, the latent space regularization term, and the network feature distribution difference term, and jointly optimize the variational autoencoder module.

[0045] Specifically, the joint loss function is: ; In the formula, For joint losses, For the adjacency matrix reconstruction error, For latent space regularization terms, This represents the difference in network feature distribution. yes and The KL divergence between two distributions.

[0046] Step S4. The generative adjacency matrix is ​​reconstructed using the network feature reconstruction module. With generative node attributes By fusing, iteratively updating, and deriving edge attributes, generative edge attributes that satisfy the constraints are obtained. To obtain a generative cyber-physical power system network.

[0047] Example 4 In another preferred embodiment of the present invention, the present invention includes a generative cyber-physical power system topology construction method, comprising the following steps: Step S1. Represent the power system and communication network as a two-layer network structure, extract the adjacency matrix, node attributes, and edge attributes respectively, and establish the input network model. Specifically, it includes the following steps: Step S 11 The communication network includes a set of substation nodes and a set of communication links in the communication layer, and the power system includes a set of substation nodes and a set of power transmission lines in the physical layer. Specifically, the communication network can be represented as follows: The power system can be represented as .in, , , represents the set of substation nodes in the communication layer; , , representing the set of substation nodes in the physical layer; , Represents a set of communication links; , This represents a set of power transmission lines.

[0048] All nodes and edges in the power system and communication network are numbered, and cross-layer mapping relationships are established. Specifically, in this embodiment, referring to a typical substation communication structure, numerical protection relays, merging units, and process units communicate within the substation via a local area network (LAN); these devices then connect to the wide area network (WAN) through a routing gateway and relay communication node in the substation, and subsequently connect to the control center. Therefore, the interdependence of the cyber-physical power system is defined as a "partial one-to-one" correspondence, meaning that each physical layer substation node is associated with a communication layer substation node (i.e., a routing gateway), but not all communication nodes are connected to physical substation nodes, as shown in the appendix to the specification. Figure 2 As shown, a two-layer network structure coupling power and communication is constructed.

[0049] Step S 12 Based on step S 11 The established two-layer network structure is used to extract its adjacency matrix, node attributes, and edge attributes to construct the input network model. .in, A Represents the adjacency matrix; , represents the set of attributes of all nodes; , represents the set of attributes for all edges. Where, For the type characteristics of nodes, These are the network characteristics of the nodes.

[0050] In this embodiment, three different types of nodes in the power system are considered: generator nodes, load nodes, and zero-injection nodes, denoted as follows: In the communication model, all nodes are considered as substation routers.

[0051] Network characteristics of nodes Defined as the sum of the capacities of its adjacent edges: , in, Represents nodes i The set of connected edges. This represents the capacity of an edge, such as the capacity of a transmission line in a power system, or the bandwidth of a link in a communication layer.

[0052] Refer to the instruction manual appendix Figure 1 This embodiment establishes a generative framework Graph-CPS, including a recurrent neural network module, a variational autoencoder module, and a network feature reconstruction module. The recurrent neural network module and the variational autoencoder module are used to generate the topology and network features of the cyber-physical power system, respectively; the network feature reconstruction module integrates the generated results to form a new generative network. Specifically, it includes the following steps: Step S2. Use a recurrent neural network module to serialize and generate the topology, and output a generative adjacency matrix. .

[0053] The recurrent neural network module includes node-level recurrent neural networks (RNNs) and edge-level recurrent neural networks (RNNs) that work together in a nested manner. A complete network connection structure is gradually constructed through sequential recursion, ultimately yielding a generative adjacency matrix. Both the node-level recurrent neural network and the edge-level recurrent neural network include state transition functions and output functions. , , in, This represents the encoding of the graph generated at the current time step. Let be the adjacency vector from the previous time step. Represents a node o The distribution of binary adjacency vectors. (Function) and It can be any neural network structure.

[0054] The sequential recursion specifically refers to the following: the node-level recurrent neural network is responsible for maintaining the global state of the graph during the sequential node generation process. Whenever a new node is generated, the edge-level recurrent neural network is called to predict the connection relationship between the node and all previous nodes. The edge-level recurrent neural network generates the existence or non-existence of edges step by step under the condition of the current hidden state of the node-level recurrent neural network, and determines the adjacency vector of the new node. Subsequently, the node-level recurrent neural network updates its hidden state according to the output of the edge-level recurrent neural network, providing contextual information for the generation of the next node.

[0055] The input adjacency matrix is ​​processed using a variational autoencoder module. A and the set of attributes of all nodes X Mapping to the latent space, the type features of the generated nodes and the distribution of network features of the nodes are used to output generative node attributes. .

[0056] Specifically, the variational autoencoder module includes an encoder and a decoder. The encoder uses an adjacency matrix... A and the set of attributes of all nodes X As input, a two-layer graph convolutional network (GCN) is used to project the input into the latent space. : , In the formula, Represents encoder, Let these be latent space variables; the latent space is regularized by an isotropic Gaussian prior distribution: .

[0057] The decoder of the variational autoencoder module also employs a two-layer graph convolutional network (GCN), using the adjacency matrix obtained by sampling the inner product of the latent space Z. As input, output generative node attributes : , , In the formula, It's a decoder. , , These are the type characteristics of generative nodes and the network characteristics of generative nodes, respectively.

[0058] The graph convolutional network in the decoder of the variational autoencoder module is as follows: , In the formula, , For trainable parameters, , For a symmetric normalized adjacency matrix, for The degree matrix.

[0059] Step S3. Establish a joint loss function, taking into account the adjacency matrix reconstruction error, the latent space regularization term, and the network feature distribution difference term, and jointly optimize the variational autoencoder module.

[0060] To comprehensively capture global network characteristics, this embodiment uses the node's attribute set. X Node attribute vectors x Convert to probability distribution When comparing the feature distributions of real and generative networks, the Kullback-Leibler divergence is used to measure their differences. , In the formula, It is the distribution of network characteristics in real networks. It is the feature distribution in generative networks.

[0061] Therefore, the joint loss function is: ; In the formula, For joint losses, For the adjacency matrix reconstruction error, For latent space regularization terms, This represents the difference in network feature distribution. yes and The KL divergence between two distributions.

[0062] Step S4. The generative adjacency matrix is ​​reconstructed using the network feature reconstruction module. With generative node attributes By fusing, iteratively updating, and deriving edge attributes, generative edge attributes that satisfy the constraints are obtained. This yields a generative cyber-physical power system network. Specifically, the generative adjacency matrix is ​​output from the recurrent neural network module and the variational autoencoder module. and generative node attributes Then, the network feature reconstruction module assigns node attributes and reconstructs edge attributes. The core idea is to assign larger attribute values ​​to nodes with higher degrees. Through iterative updates, this continues until all edge attributes satisfy non-zero constraints, ultimately yielding a complete generative cyber-physical power system. (Refer to the appendix of the instruction manual.) Figure 3 Specifically, it includes the following steps: Step S 41 The recurrent neural network module and the variational autoencoder module output a generative adjacency matrix. and generative node attributes Then, the edge attributes Initialize to 0.

[0063] Step S 42 For generative node attributes Sort in descending order.

[0064] Step S 43 Based on generative adjacency matrix Sort by degree in descending order , Arrange them. Among them, and These are the sets of nodes in the information layer and the physical layer, respectively.

[0065] Step S 44 . Arrange the attributes of the generative nodes Assign values ​​to in sequence , .

[0066] Step S 45 Locate the node with the minimum degree value.

[0067] Step S 46 Find the adjacent edges of the node with the minimum degree value, and obtain the set of adjacent edges of that node. ,That And define These are the remaining node attribute values ​​of that node (attribute values ​​not assigned to any edge), at the start of the algorithm. Calculate the edge attributes of adjacent edges based on the remaining node attribute values. And update the attribute values ​​of the remaining nodes. : .

[0068] Step S 47 Repeat step S. 45 Step S 46 until all edge attributes >0; Generates the final edge attributes. .

[0069] To verify the effectiveness and scalability of the generative cyber-physical power system topology construction method proposed in this invention, the following explanation is based on typical power system and communication network examples.

[0070] Experimental subjects: At the physical layer, the IEEE 39-node standard test system, the 380kV / 400kV transmission systems of Italy and Germany, and the European continental power grid were selected as sample networks. At the communication layer, the Jiangsu Provincial Power Grid communication network and the verified IEEE 39-node and 118-node communication systems were selected as sample networks. The scale of these networks was expanded from 18 nodes to 1225 nodes, covering both standard test systems and actual power grid examples.

[0071] Implementation process: The physical layer and communication layer networks mentioned above are input into the generative framework Graph-CPS of this invention. The topology is generated by the recurrent neural network module, the node attributes and feature distribution are generated by the variational autoencoder module, and finally the complete generative network is obtained by the network feature reconstruction module.

[0072] Various statistical metrics, including average node degree, average shortest path length, network diameter, network density, and node betweenness centrality (mean and maximum), are compared between real and generative networks. These metrics characterize the global structural properties of the network. Simultaneously, the mean and variance of node features are compared to verify the consistency of attribute distribution in the generative network.

[0073] Results and Analysis: Experimental results show that the network generated by the method of this invention has little difference from the real network in the above-mentioned indicators, accurately maintains the structural characteristics and node attribute distribution of the input network, and exhibits good scalability and applicability in power systems and communication networks of different scales and types. Table 1 below shows the statistical comparison between the real system and the generative system. From a topological perspective, the quality of the generated network is evaluated using multiple complex network parameters such as average node degree, average shortest path length, network diameter, network density, average node betweenness, and maximum node betweenness. These parameters reflect the global structural characteristics of the network. From the perspective of network features, the generation quality is evaluated by comparing the normalized feature mean and variance. As can be seen from Table 1, the differences between all generated parameters and the original network are small. Therefore, it can be proved that the generative framework Graph-CPS has good scalability and accuracy, and can maintain feature consistency under different types (power and communication networks) and different scales of input networks.

[0074] Table 1 Fault Types and Fault Symptom Numbers

[0075] In Table 1, For the number of nodes, Let be the number of edges. The average node degree. The average shortest path length, For network diameter, For network density, For average node betweenness centrality, For maximum node betweenness centrality, To normalize the mean of the generative node attributes, To normalize the variance of generative node attributes, (*): the number in parentheses represents the difference between the generative parameters and the true parameters.

[0076] For further explanation, the IEEE 39-node system is used as an example below, and the generated results are shown in the appendix to the specification. Figure 4 ~Instruction manual included Figure 6 As shown: In the physical layer generation results, there are 15 load nodes, 10 generator nodes, and 14 zero-injection nodes, a distribution close to that of the real system (17 loads, 10 generators, and 12 zero-injection nodes). In the communication layer generation results, the network maintains a tree structure similar to that of the real network, verifying the ability of this method to learn the overall structural features.

[0077] In the comparison of node attributes, as shown in the instruction manual... Figure 7 Included with instruction manual Figure 8As shown, the normalized mean of the node features in the IEEE 39-node system is 0.239, while the result generated by this invention is 0.287; the normalized mean of the real communication system is 0.418, while the generated result is 0.436. The differences in variance between the two are 0.008 and 0.003, respectively, which are relatively small.

[0078] The above examples demonstrate that this invention can generate generative networks that are highly consistent with real systems in terms of statistical characteristics and attribute distribution without disclosing sensitive information about the real power grid, proving the reliability and practicality of the method. This method can be widely applied to power system planning and design, operation simulation, and network security research.

[0079] In summary, any other corresponding modifications made by those skilled in the art after reading this invention document, without requiring creative mental effort, based on the technical solutions and concepts of this invention, are all within the scope of protection of this invention.

Claims

1. A generative information-physical power system topology construction method, characterized in that: The method comprises the following steps: Step S1. Represent the power system and the communication network as a two-layer network structure, extract the adjacency matrix, node attributes and edge attributes respectively, and establish an input network model ; wherein, A represents the adjacency matrix; X represents the attribute set of all nodes, E represents the attribute set of all edges; Step S2. Adopting a recurrent neural network module to serialize the topological structure generation, outputting the generated adjacency matrix ; An input adjacency matrix is mapped to a latent space by a variational autoencoder module A and a set of attributes of all nodes X to generate type features of nodes and network feature distribution of nodes, and output generated node attributes ; Step S3. A joint loss function is established to comprehensively consider the adjacency matrix reconstruction error, the hidden space regularization term and the network feature distribution difference term, and the variational autoencoder module is jointly optimized; Step S4. Reconstructing the generative adjacency matrix through the network feature reconstruction module with the generative node attributes Fusion, iterative update and derivation of edge attributes, and obtaining generative edge attributes satisfying the constraint conditions , and obtaining a generative cyber-physical power system network.

2. The method of claim 1, wherein: In an input network model, , ; wherein, is a type feature of a node, is a network feature of a node; denotes a capacity of an edge; a network feature of a node is defined as the sum of the capacities of its adjacent edges: ， wherein, represents a set of edges connected to the node i .

3. The method of claim 1, wherein: The step S1 comprises the following steps: Step S 11 . The communication network includes a set of substation nodes and a set of communication links in a communication layer, and the power system includes a set of substation nodes and a set of power transmission lines in a physical layer; all nodes and edges in the power system and the communication network are numbered respectively, and a cross-layer mapping relationship is established, so that each substation node in the physical layer is associated with a corresponding substation node in the communication layer, and a power-communication coupled double-layer network structure is constructed; Step S 12 . Based on step S 11 The double-layer network structure is established, the adjacency matrix, node attributes and edge attributes are extracted, and an input network model is constructed .

4. The method of claim 1, wherein: The recurrent neural network module comprises a node-level recurrent neural network and an edge-level recurrent neural network which work in a nested manner, and a complete network connection structure is gradually constructed in a sequential recursive manner, and finally a generative adjacency matrix is obtained .

5. The method of claim 4, wherein: The sequential recursion specifically refers to: the node-level recurrent neural network is responsible for maintaining the global state of the graph in the sequential node generation process, and the edge-level recurrent neural network is called to predict the connection relationship between the new node and all previous nodes whenever a new node is generated; the edge-level recurrent neural network gradually generates the existence or nonexistence of the edge under the condition of the current hidden state of the node-level recurrent neural network to determine the adjacency vector of the new node; then, the node-level recurrent neural network updates its hidden state according to the output result of the edge-level recurrent neural network to provide context information for the generation of the next node.

6. The method of claim 1 or 5, wherein: The encoder of the variational autoencoder module takes as input the adjacency matrix A and the set of attributes of all nodes X and projects the input into the latent space using a two-layer graph convolutional network Z , the latent space Z is regularized by an isotropic Gaussian distribution; the decoder of the variational autoencoder module takes as input the adjacency matrix Z sampled from the inner product of the latent space and outputs generative node attributes using a two-layer graph convolutional network .

7. The method of claim 6, wherein: projecting the input to a latent space using a two-layer graph convolutional network Z Specifically, ； wherein represents an encoder, is a latent space variable; the latent space is regularized by a isotropic Gaussian prior: ; Output generation node properties Specifically refers to: ， ， In the formula, is a decoder, is a latent space Z is an adjacency matrix obtained by inner product sampling, , , are the type features of the generative node and the network features of the generative node, respectively.

8. The method of claim 7, wherein: The graph convolution network in the decoder of the variational autoencoder module is: ， wherein , are trainable parameters, , is a symmetric normalized adjacency matrix, is the degree matrix of .

9. The method of claim 1, wherein: Further comprising: The set of attributes of a node X Node attribute vectors x Convert to probability distribution The Kullback-Leibler divergence is used to measure the difference in network feature distributions between real and generative networks. ， wherein is the distribution of network features in the real network, is the distribution of features in the generative network.

10. The method of claim 9, wherein: The joint loss function is: ； wherein is the joint loss, is the adjacency matrix reconstruction error, is the latent space regularization term, is the network feature distribution difference term, is and the KL divergence between two distributions.

11. The method of claim 1, wherein: The step S4 specifically comprises the following steps: Step S 41 . In the recurrent neural network module and the variational autoencoder module output a generative adjacency matrix and generative node attributes After that, the edge attribute is initialized to 0; Step S 42 . The generated node attribute is arranged in descending order; Step S 43 . Based on the generated adjacency matrix , arrange the nodes in descending order of degrees; Step S 44 . The arranged generated formula node attributes are assigned to the arranged nodes in turn; Step S 45 . Position the node with the smallest degree value; Step S 46 . find the adjacent edge of the node with the smallest degree value, calculate the edge attribute of the adjacent edge based on the residual node attribute value of the node , and update the residual node attribute value ; Step S 47 . Repeat step S 45 ~Step S 46 , until all edge attributes > 0; generate final edge attributes .

Images