A method of machine learning training a topology generator and a topology discriminator model

Abstract

The application discloses a kind of machine learning training topology generator and the method for generating topological discriminator model, comprising the following steps: with matrix pair form characterizing real topological data, and pre-processing;Topology generator is constructed, and topological pseudo-matrix pair data is generated;Topological discriminator is constructed to identify the source of topological data;The topological generator and the topological discriminator are based on the adversarial training of generative adversarial network;Low-voltage area topological matrix pair data is generated, and is converted into visual topological graph.The application realizes the coevolution of topological generator and topological discriminator by constructing topological generator and topological discriminator and through the adversarial training of both, and then enables topological generator to automatically generate a large number of topological data in accordance with real station area statistical distribution and engineering rationality from simple input, with the technical effects of high generation quality, high automation efficiency, guaranteeing physical rationality, solving data scarcity problem and strong flexibility and adaptability.

Description

Technical Field

[0001] This invention relates to the field of power system distribution network technology, and in particular to a method for training topology generator and topology discriminator models using machine learning. Background Technology

[0002] In the planning, design, upgrading, and digital modeling of distribution networks (especially low-voltage distribution areas), the accuracy of the topology is fundamental for advanced applications such as line loss calculation, fault location, power supply reliability analysis, and distributed power source integration. Currently, obtaining the topology of low-voltage distribution areas mainly relies on two methods: The first is manual on-site surveying and drawing input, where maintenance personnel verify line connections on-site and then input the data into a graphical system. This method is highly accurate but time-consuming, labor-intensive, costly, and difficult to update dynamically, failing to meet the needs of large-scale, rapidly changing distribution area management. The second method is topology identification based on measurement data. This method utilizes time-series data such as voltage, current, and phase collected by smart meters, and then infers the electrical connection relationships between user nodes through correlation analysis and clustering algorithms. This method requires less manual labor but is highly dependent on the quality, density, and synchronization of the measurement data. In practical applications, issues such as missing data, noise interference, three-phase imbalance, and asynchronous measurements often lead to errors or ambiguities in the identified topology, especially in distribution areas with complex branch levels, where accuracy is difficult to guarantee.

[0003] Existing technologies mainly face a contradiction between "accuracy" and "efficiency": while manual on-site surveys and drawing input methods can ensure topological accuracy, they are inefficient and cannot meet the needs of rapid modeling of massive transformer areas; while topological identification methods based on measurement data can improve efficiency and reduce costs, topological accuracy is severely constrained by data quality. The essence of topological identification methods based on measurement data is to infer a deterministic structure from imperfect observation data. When the available data quality is low or the information is incomplete, the solution to this "inverse problem" has inherent uncertainties and multiple solutions. The model is prone to getting trapped in local optima, generating unreasonable or physically infeasible topologies, which are difficult to meet engineering application standards.

[0004] It is understood that the above statements only provide background information related to the present invention and do not necessarily constitute prior art. Summary of the Invention

[0005] Based on the aforementioned technical problems, the purpose of this invention is to provide a method for training topology generator and topology discriminator models using machine learning. By allowing a topology generator and a topology discriminator to learn together in an adversarial manner, the topology generator can automatically generate a large amount of topology data that conforms to the statistical distribution and engineering rationality of real transformer substations from simple inputs. This generated topology data can be used as a candidate set for planning assistance schemes and data completion, or as training data to enhance topology-related algorithms, thereby improving the automation level and robustness of low-voltage transformer substation digitization work.

[0006] To achieve the above objectives, this invention proposes a method for training topology generator and topology discriminator models using machine learning, comprising the following steps: Real-world topological data is represented in matrix pairs and preprocessed. Build a topology generator to generate topological pseudo-matrix pairs of data; Construct a topology discriminator to identify the source of topology data; The topology generator and the topology discriminator are trained adversarially based on generative adversarial networks. Generate the low-voltage transformer area topology matrix data and convert it into a visual topology map.

[0007] Furthermore, the method for representing real topological data and performing preprocessing includes the following steps: Multiple sets of real low-voltage transformer area topology data verified by manual survey were collected to construct a real topology dataset, which was then uniformly converted into matrix pairs to form real topology matrix pair data. The data is cleaned using the real topology matrix, and invalid samples containing loops, isolated nodes, or abnormal electrical distances are removed to complete the standardization preprocessing.

[0008] Furthermore, the method for constructing a real topological dataset and uniformly converting it into matrix pair form includes the following steps: The topology of a single low-voltage transformer area in the constructed real topology dataset is abstracted into a topology graph structure: in, V It is a set of nodes, covering core nodes such as transformers, branch boxes, and user meters; E Let be the set of edges, representing the physical connection relationships between nodes via wires; Transform the topological graph structure into a normalized matrix pair form: adjacency matrix. and node feature matrix ; in, N The total number of nodes. F The adjacency matrix represents the node feature dimension. In the middle, if nodei and j There is a physical connection. Otherwise, it is 0; the node feature matrix In the diagram, each row corresponds to a feature vector of a node. The feature dimensions of the feature vector include, but are not limited to, node type, three-phase phase attributes, equipment rated capacity, historical average / peak load, and power supply radius. The node type includes transformers, branch points, and user nodes.

[0009] Furthermore, the topology generator generates a pseudo-topology matrix for the output layer from the input layer's random noise vector and conditional constraint vector: In the formula, z It is a random noise vector, sampled from a standard normal distribution. ; For a pair of pseudo-topological matrices, where, This is the node adjacency matrix. , where is the node feature matrix, and , respectively, corresponds to the pseudo-adjacency matrix and the pseudo-node feature matrix; C The conditional constraint vector contains engineering constraint metadata of the target topology. The engineering constraint metadata includes, but is not limited to, the total number of nodes, transformer location index, number of topology levels, and three-phase load ratio. The number of topology levels includes the trunk, branches, and user layers.

[0010] Furthermore, the network architecture of the topology generator adopts a graph neural network encoding and decoding structure, and the method for constructing the topology generator includes the following steps: The encoder processes the conditional constraint vector and random noise vector to generate global topological prior features; The node feature matrix is ​​generated layer by layer by generating a graph network, and the node adjacency matrix is ​​generated based on the correlation of node features to ensure the structural coherence of the generated topology.

[0011] Furthermore, the input layer of the topology generator is given a random noise vector to generate unconstrained and diverse topologies, which are then filtered and selected according to engineering rules to meet the requirements; the method for constructing the topology generator includes the following steps: The encoder processes random noise vectors to generate unconstrained global topological prior features. The global topological prior features are generated by filtering and selecting samples that meet the requirements from the unconstrained global topological prior features using engineering rules. The node feature matrix is ​​generated layer by layer by generating a graph network, and the node adjacency matrix is ​​generated based on the correlation of node features to ensure the structural coherence of the generated topology.

[0012] Furthermore, the topology discriminator is used to determine whether the source of the input topology data is real topology matrix pairs or pseudo topology matrix pairs generated by the topology generator; The input to the topology discriminator is a pair of topology matrices, and the output of the topology discriminator is a scalar. , representing the probability that the topological matrix pair is a true topological matrix pair in the data, scalar The closer the value is to 1, the higher the confidence level for determining it as the logarithm of the true topological matrix; in, Given the input adjacency matrix, The input feature matrix is ​​denoted as .

[0013] Furthermore, the topology discriminator adopts a fusion architecture of graph convolutional network and graph attention network. The bottom layer of the topology discriminator adopts a graph convolutional network architecture to extract local structural features of topology matrix pairs; the bottom layer of the topology discriminator adopts a graph attention network architecture to weighted aggregate global topology features through an attention mechanism; the top fully connected layer of the topology discriminator outputs the authenticity probability and introduces engineering rule constraints to enhance the ability to identify physically unreasonable topologies.

[0014] Furthermore, the generative adversarial network optimizes the objective function through a minimax game to achieve adversarial training between the topology generator and the topology discriminator. The minimax game optimization objective function is: In the formula, G is the topology generator, and D is the topology discriminator. To generate models, The core expectation operator.

[0015] Furthermore, the training method for the topology generator and the topology discriminator includes the following steps: Fix the topology generator and train the topology discriminator: input the real topology matrix pair data sample and the pseudo topology matrix pair data sample, optimize the topology discriminator so that the output of the real topology matrix pair data sample approaches 1 and the output of the pseudo topology matrix pair data sample approaches 0; Fix the topology discriminator and train the topology generator: optimize the topology generator so that the pseudo-topology matrix generated by the topology generator approaches 1 for the output of the data sample after being discriminated by the topology discriminator. The topology generator and the topology discriminator are trained alternately and iteratively until the minimax game optimization objective function converges.

[0016] Compared with the prior art, the present invention has the following advantages: This invention represents topological data in matrix pair form, which is compatible with the optimal state of graph neural networks, thereby improving the adaptability and computational efficiency of graph neural networks to topological data. By constructing a topology generator and a topology discriminator and training them adversarially, the topology generator and discriminator achieve co-evolution, enabling the topology generator to automatically generate a large amount of topological data that conforms to the statistical distribution and engineering rationality of real transformer substations from simple inputs. This generated topological data can be used as a candidate set for planning assistance schemes and data completion, or as training data to enhance topology-related algorithms, thereby improving the automation level and robustness of low-voltage transformer substation digitization. In addition, this invention also has the technical effects of high generation quality, high automation efficiency, ensuring physical rationality, solving the problem of data scarcity, and strong flexibility and adaptability. Attached Figure Description

[0017] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the drawings in the following description are one embodiment of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort: Figure 1 This is a flowchart illustrating a method for training a topology generator and a topology discriminator model using machine learning, as provided in an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of the topology graph structure and topology matrix pair data provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] It should be noted that, in this document, the terms "comprising," "including," "having," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Unless otherwise specified, an element defined by the phrase "comprising..." or "including..." does not exclude the presence of additional elements in the process, method, article, or terminal device that includes said element.

[0021] It should be noted that the accompanying drawings are all in a very simplified form and use non-precise ratios, and are only used to facilitate and clearly illustrate the purpose of the embodiments of the present invention.

[0022] This invention provides a method for training a topology generator and a topology discriminator model using machine learning. By constructing a topology generator and a topology discriminator and training them adversarially, the topology generator learns and captures the inherent statistical distribution patterns of real low-voltage transformer area topology data, ultimately achieving intelligent generation of highly realistic low-voltage transformer area topology data.

[0023] like Figure 1 As shown, the present invention provides a method for training a topology generator and a topology discriminator model using machine learning, comprising the following steps: S1. Represent the real topological data in the form of matrix pairs and perform preprocessing.

[0024] Abstracting the topology of a single low-voltage distribution area into a topology graph structure. ,in, V It is a set of nodes, covering core nodes such as transformers, branch boxes, and user meters; E The edge set represents the physical connections between nodes via wires. To adapt to neural network processing paradigms, the topology graph is transformed into a standardized matrix pair form: the adjacency matrix. and node feature matrix , N The total number of nodes. F For node feature dimensions.

[0025] The adjacency matrix In the middle, if node i and j There is a physical connection. Otherwise, it is 0. This invention addresses the radial topology characteristics of low-voltage distribution areas by employing undirected graph modeling; therefore, the adjacency matrix... It is a symmetric matrix. The node feature matrix In the diagram, each row corresponds to a feature vector of a node. The feature dimensions include, but are not limited to, node type, three-phase phase attributes, equipment rated capacity, historical average / peak load, and power supply radius. Node types include transformers, branch points, and user nodes. For nodes with missing features, one-hot encoding or industry default values ​​are used to fill in the missing features. Figure 2 The diagram shown is a schematic diagram of the topological matrix pair data provided in this embodiment.

[0026] Step S1 includes the following steps: S1.1 Collect multiple sets of real low-voltage transformer area topology data verified by manual survey, construct a real topology dataset, and uniformly convert it into matrix pair form to form real topology matrix pair data; S1.2 Clean the data using the real topology matrix, removing invalid samples containing loops, isolated nodes, or abnormal electrical distances, and complete the standardization preprocessing.

[0027] In other embodiments, the representation of the real topology data includes, but is not limited to, edge list form and association matrix form.

[0028] S2. Construct a topology generator to generate topological pseudo-matrix pairs of data.

[0029] The core function of the topology generator is to learn the mapping from the input space to the topological matrix space. In this embodiment, the topology generator generates pseudo-topological matrix pairs from the input layer's random noise vector and conditional constraint vector as the output layer's data. In the formula, z It is a random noise vector, sampled from a standard normal distribution. , C The condition constraint vector; For a pair of pseudo-topological matrices, where, This is the node adjacency matrix. is the node feature matrix, corresponding to the pseudo-adjacency matrix and the pseudo-node feature matrix, respectively.

[0030] Condition constraint vector C The inclusion of engineering constraint metadata for the target topology upgrades the generation process from "unconstrained random generation" to "directional and controllable generation," meeting customized planning requirements. This engineering constraint metadata includes, but is not limited to, the total number of nodes, transformer location indexes, the number of topology levels, and the three-phase load ratio. The number of topology levels includes the trunk, branches, and user layers.

[0031] The network architecture of the topology generator adopts a graph neural network encoding and decoding structure. The method for constructing the topology generator includes the following steps: S2.1. Generate global topological prior features by processing the conditional constraint vector and random noise vector through the encoder; S2.2 Generating node feature matrices layer by layer using a graph generation network. And generate a node adjacency matrix based on the correlation of node features. This ensures the structural consistency of the generated topology.

[0032] In another embodiment, the input layer of the topology generator receives only random noise vectors to generate unconstrained and diverse topologies, which are then filtered and selected according to engineering rules to meet the requirements. The method for constructing the topology generator includes the following steps: S2.1. The random noise vector is processed by the encoder to generate unconstrained global topological prior features; S2.2. Filter and select samples that meet the requirements from the unconstrained global topological prior features through engineering rules to generate global topological prior features; S2.3. Generate node feature matrices layer by layer using a graph generation network. And generate a node adjacency matrix based on the correlation of node features. This ensures the structural consistency of the generated topology.

[0033] In another embodiment, the network architecture of the topology generator may adopt, but is not limited to, variational autoencoder structures and normalized flow structures.

[0034] S3. Construct a topology discriminator to identify the source of topology data.

[0035] The core objective of the topology discriminator is to distinguish the source of the input topology data, i.e., whether the topology data originates from genuine topology matrix pairs or pseudo-topology matrix pairs generated by the topology generator. The input to the topology discriminator is a pair of topology matrices. The output of the topology discriminator is a scalar. Characterize the topological matrix pair The probability of the true topological matrix pair of data, scalar The closer the value is to 1, the higher the confidence level for determining it as the logarithm of the true topological matrix.

[0036] The topology discriminator employs a fusion architecture of graph convolutional networks and graph attention networks, overcoming the limitation of traditional discriminators that only focus on node features. The bottom layer of the topology discriminator uses a graph convolutional network architecture to extract local structural features of topological matrix pairs, such as node degree distribution and neighborhood connection patterns. The bottom layer of the topology discriminator uses a graph attention network architecture to weighted aggregate global topological features through an attention mechanism. These global topological features include connectivity, clustering coefficients, and average path length. The top fully connected layer of the topology discriminator outputs the probability of authenticity and introduces engineering rule constraints to enhance the ability to identify physically unreasonable topologies. These engineering rule constraints include radial structure verification and loop-free verification.

[0037] S4. The topology generator and the topology discriminator are trained adversarially based on generative adversarial networks.

[0038] Following the standard alternating training paradigm of generative adversarial networks (GANs), the objective function is optimized through a minimax game to achieve the co-evolution of the topology generator and the topology discriminator. The minimax game optimization objective function is: In the formula, G is the topology generator, and D is the topology discriminator. To generate models, The core expectation operator.

[0039] The training method for the topology generator and the topology discriminator includes the following steps: S4.1 Fix the topology generator and train the topology discriminator: Input the real topology matrix pair data sample and the pseudo topology matrix pair data sample, optimize the topology discriminator so that the output of the real topology matrix pair data sample approaches 1 and the output of the pseudo topology matrix pair data sample approaches 0; S4.2 Fix the topology discriminator and train the topology generator: optimize the topology generator so that the output of the pseudo-topology matrix generated by the topology generator after the data sample is discriminated by the topology discriminator approaches 1; The pseudo-topology matrix generated by the topology generator approaches 1 after the data sample is judged by the topology discriminator, meaning that the pseudo-topology matrix generated by the topology generator can "fool" the discriminator for the data sample. S4.3 The topology generator and the topology discriminator are trained alternately and iteratively until the minimax game optimization objective function converges.

[0040] When the minimax game optimization objective function converges, the accuracy of the topology discriminator stabilizes at around 50%, and it cannot distinguish between true and false topology matrix pairs.

[0041] Preferably, the following optimization strategies are introduced during the training process of the topology generator and the topology discriminator: For the generated node adjacency matrix Symmetric constraints are applied to ensure the properties of the undirected graph; gradient penalties are added to the generative adversarial network to avoid mode collapse and improve the diversity of generated topologies; engineering rule penalties, such as loop penalties and isolated node penalties, are incorporated into the loss function of the topology discriminator to guide the generator to learn physically reasonable topologies.

[0042] S5. Generate the low-voltage transformer area topology matrix data and convert it into a visual topology map.

[0043] After the topology generator is trained, the target condition constraint vector C (or default random conditions) and the sampling noise vector z can be input to quickly generate a low-voltage transformer area topology matrix pair data that meets the constraints. After matrix-graph structure transformation, a visualized topology map is output.

[0044] This invention supports the following core application scenarios: Transformer Area Planning Assistance: Provides multiple alternative wiring schemes that conform to engineering practices for new or renovated transformer areas, reducing the workload of manual drawing; Data Augmentation: Generates a large amount of high-quality pseudo-topology data to supplement the deficiencies of real datasets and improve the generalization ability of downstream topology identification models; Topology Completion or Repair: Generates reasonable completion suggestions for existing transformer area topologies with missing or damaged data, and repairs problems such as isolated nodes and disconnected branches; Algorithm Testing: Provides diverse test topology sets for algorithms such as line loss calculation and fault location to verify the robustness of the algorithms.

[0045] In this embodiment, the training environment for the topology generator and the topology discriminator is Python 3.8 + PyTorch 2.0, and the hardware is adapted to NVIDIA RTX 3090 and above GPUs; the topology visualization is implemented using NetworkX + Matplotlib, and the generated topology data can be exported as SVG, CAD and other engineering standard formats, and directly connected to the power distribution network digital platform.

[0046] The method for training topology generator and topology discriminator models using machine learning proposed in this invention has been verified by field tests and possesses the following core technical effects: High generation quality: The divergence of the graph theory statistical indicators (node ​​degree distribution, path length distribution, clustering coefficient) between the generated topology and the real topology is reduced by more than 60%, and the visual features and engineering features are highly consistent.

[0047] High automation efficiency: After the model is trained, the generation time for a single set of topology data is ≤1 second, which is more than 3 orders of magnitude more efficient than manual surveying and mapping, and is suitable for the rapid modeling needs of massive transformer areas.

[0048] Addressing the issue of data scarcity: For scenarios with small or zero sample transformer areas, the generated pseudo-topology data can improve the accuracy of downstream topology identification models by 15%-30%.

[0049] Ensuring physical rationality: Through the engineering rule constraints of the discriminator, the probability of generating unreasonable structures such as topologies containing loops and isolated subnets is reduced by more than 90%, which is far superior to traditional data back-inference methods.

[0050] High flexibility and adaptability: By adjusting the condition vector C, topologies of different scales (50-500 nodes) and different structures (single transformer / multiple branches) can be generated to meet the planning needs of different urban and rural scenarios.

[0051] Although the present invention has been described in detail through the preferred embodiments above, it should be understood that the above description should not be considered as a limitation of the present invention. Various modifications and substitutions to the present invention will be apparent to those skilled in the art after reading the above description. Therefore, the scope of protection of the present invention should be defined by the appended claims.

Claims

1. A method for training topology generator and topology discriminator models using machine learning, characterized in that, Includes the following steps: Real-world topological data is represented in matrix pairs and preprocessed. Build a topology generator to generate topological pseudo-matrix pairs of data; Construct a topology discriminator to identify the source of topology data; The topology generator and the topology discriminator are trained adversarially based on generative adversarial networks. Generate the low-voltage transformer area topology matrix data and convert it into a visual topology map.

2. The method for training a topology generator and a topology discriminator model using machine learning as described in claim 1, characterized in that, The method for representing real topological data and preprocessing it includes the following steps: Multiple sets of real low-voltage transformer area topology data verified by manual survey were collected to construct a real topology dataset, which was then uniformly converted into matrix pairs to form real topology matrix pair data. The data is cleaned using the real topology matrix, and invalid samples containing loops, isolated nodes, or abnormal electrical distances are removed to complete the standardization preprocessing.

3. The method for training a topology generator and topology discriminator model using machine learning as described in claim 2, characterized in that, The method for constructing a real topology dataset and uniformly converting it into matrix pair form includes the following steps: The topology of a single low-voltage transformer area in the constructed real topology dataset is abstracted into a topology graph structure: in, V It is a set of nodes, covering core nodes such as transformers, branch boxes, and user meters; E Let be the set of edges, representing the physical connection relationships between nodes via wires; Transform the topological graph structure into a normalized matrix pair form: adjacency matrix. and node feature matrix ; in, N The total number of nodes. F The adjacency matrix represents the node feature dimension. In the middle, if node i and j There is a physical connection. Otherwise, it is 0; the node feature matrix In the diagram, each row corresponds to a feature vector of a node. The feature dimensions of the feature vector include, but are not limited to, node type, three-phase phase attributes, equipment rated capacity, historical average / peak load, and power supply radius. The node type includes transformers, branch points, and user nodes.

4. The method for training a topology generator and topology discriminator model using machine learning as described in claim 1, characterized in that, The topology generator generates a pseudo-topology matrix for the output layer from the input layer's random noise vector and conditional constraint vector. In the formula, z It is a random noise vector, sampled from a standard normal distribution. ; For a pair of pseudo-topological matrices, where, This is the node adjacency matrix. , where is the node feature matrix, and , respectively, corresponds to the pseudo-adjacency matrix and the pseudo-node feature matrix; C The conditional constraint vector contains engineering constraint metadata of the target topology. The engineering constraint metadata includes, but is not limited to, the total number of nodes, transformer location index, number of topology levels, and three-phase load ratio. The number of topology levels includes the trunk, branches, and user layers.

5. The method for training a topology generator and a topology discriminator model using machine learning as described in claim 4, characterized in that, The network architecture of the topology generator adopts a graph neural network encoding and decoding structure, and the method for constructing the topology generator includes the following steps: The encoder processes the conditional constraint vector and random noise vector to generate global topological prior features; The node feature matrix is ​​generated layer by layer by generating a graph network, and the node adjacency matrix is ​​generated based on the correlation of node features to ensure the structural coherence of the generated topology.

6. The method for training a topology generator and a topology discriminator model using machine learning as described in claim 1, characterized in that, The topology generator takes a random noise vector as input to its input layer, generates unconstrained and diverse topologies, and then filters out suitable samples using engineering rules. The method for constructing the topology generator includes the following steps: The encoder processes random noise vectors to generate unconstrained global topological prior features. The global topological prior features are generated by filtering and selecting samples that meet the requirements from the unconstrained global topological prior features using engineering rules. The node feature matrix is ​​generated layer by layer by generating a graph network, and the node adjacency matrix is ​​generated based on the correlation of node features to ensure the structural coherence of the generated topology.

7. The method for training a topology generator and a topology discriminator model using machine learning as described in claim 1, characterized in that, The topology discriminator is used to determine whether the source of the input topology data is real topology matrix pairs or pseudo topology matrix pairs generated by the topology generator. The input to the topology discriminator is a pair of topology matrices, and the output of the topology discriminator is a scalar. , representing the probability that the topological matrix pair is a true topological matrix pair in the data, scalar The closer the value is to 1, the higher the confidence level for determining it as the logarithm of the true topological matrix; in, Given the input adjacency matrix, The input feature matrix.

8. The method for training a topology generator and a topology discriminator model using machine learning as described in claim 1, characterized in that, The topology discriminator adopts a fusion architecture of graph convolutional network and graph attention network. The bottom layer of the topology discriminator adopts a graph convolutional network architecture to extract local structural features of topology matrix pairs; the bottom layer of the topology discriminator adopts a graph attention network architecture to weighted aggregate global topology features through an attention mechanism. The top fully connected layer of the topology discriminator outputs the authenticity probability and introduces engineering rule constraints to enhance the ability to identify physically unreasonable topologies.

9. The method for training a topology generator and a topology discriminator model using machine learning as described in claim 7, characterized in that, The generative adversarial network optimizes the objective function through a minimax game, enabling adversarial training between the topology generator and the topology discriminator. The minimax game optimization objective function is: In the formula, G is the topology generator, and D is the topology discriminator. To generate models, The core expectation operator.

10. The method for training a topology generator and a topology discriminator model using machine learning as described in claim 9, characterized in that, The training method for the topology generator and the topology discriminator includes the following steps: Fix the topology generator and train the topology discriminator: input the real topology matrix pair data sample and the pseudo topology matrix pair data sample, optimize the topology discriminator so that the output of the real topology matrix pair data sample approaches 1 and the output of the pseudo topology matrix pair data sample approaches 0; Fix the topology discriminator and train the topology generator: optimize the topology generator so that the pseudo-topology matrix generated by the topology generator approaches 1 for the output of the data sample after being discriminated by the topology discriminator. The topology generator and the topology discriminator are trained alternately and iteratively until the minimax game optimization objective function converges.

Images