Power-traffic coupling network vulnerability identification method and system

Abstract

The invention discloses an electric power-traffic coupling network vulnerability identification method and system. The method comprises the following steps: S1, coupling an electric power network and a traffic network to form a graph model; the graph model comprises a node set and an edge set; constructing an adjacent matrix and a node feature matrix based on the node set and the edge set; s2, inputting the adjacent matrix and the node feature matrix into a graph convolutional neural network model, performing feature extraction model training through a semi-supervised learning mechanism, and outputting an updated node feature matrix; and S3, learning an optimal strategy by using a deep Q network, and identifying a fragile node sequence through interaction with the environment. According to the method, a traditional single network evaluation mode is broken through, accurate quantification of the cross-network coupling effect is achieved, complex topological information can be captured, intelligent dynamic recognition is achieved, dependence on artificial experience is reduced, decision visualization can be supported, the system practicability is high, and the method can be applied to power-traffic coupling network catastrophe scenes.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of intelligent power distribution network optimal operation, and particularly relates to a power-traffic coupled network vulnerability identification method and system. BACKGROUND

[0002] With the deep integration of global energy internet and intelligent transportation system, as the key infrastructure of modern society, the coupling and complexity of power network and traffic network are growing exponentially. The large-scale expansion of intelligent power distribution network, the three-dimensional development of traffic network, and the interaction and penetration of multi-energy carriers not only improve the system coordination efficiency, but also make the operation characteristics of power-traffic coupled network present highly nonlinear characteristics. The frequent occurrence of extreme natural disasters such as typhoon and heavy rain, combined with equipment aging and cross-network coupling, makes the failure mechanism of coupled network change from single component failure to complex mode of multi-disaster induction and cross-domain cascading evolution. The traditional vulnerability node identification method based on single network (e.g., only power network or only traffic network) cannot accurately depict the dynamic mapping relationship of vulnerability nodes under the condition of high coupling of power and traffic networks.

[0003] In the face of complex failure problems of power-traffic coupled network in disaster scenarios, the traditional vulnerability node evaluation system is facing multiple technical bottlenecks. Specifically, 1) unable to effectively handle cross-network coupling effect: the traditional method is based on single network (e.g., only power network or only traffic network) for vulnerability evaluation, ignoring the interaction and cascading failure mechanism between power and traffic networks, resulting in inaccurate evaluation results; 2) insufficient topological characterization ability: the traditional graph structure-based topological identification method cannot express the high-dimensional complex correlation and high-dimensional topology of heterogeneous networks (e.g., the heterogeneity of power nodes and traffic nodes in attributes and functions), leading to the need for manual experience in vulnerability node identification and ignoring cross-network coupling failure; 3) insufficient intelligence and adaptability: existing methods mostly use static analysis or simple heuristic rules, which cannot dynamically learn network state changes, lack sequence decision-making ability, and the single fault location method makes it difficult for the system to achieve intelligent evaluation and accurate positioning of vulnerability nodes in disaster scenarios, seriously restricting the safe operation ability of key infrastructure under extreme disasters.

[0004] The statements herein merely provide background technology related to the present application, and do not necessarily constitute prior art. SUMMARY

[0005] The purpose of the present application is to provide a power-traffic coupled network vulnerability identification method and system that integrates graph convolution and reinforcement learning, to realize intelligent, accurate and dynamic identification of vulnerability nodes in power-traffic coupled network, while fully considering the cross-network coupling effect and the complexity of network topology.

[0006] In order to achieve the above object, the present application provides a power-traffic coupling network vulnerability identification method, comprising: S1, coupling the power network and the traffic network to form a graph model: (1) In the formula: V represents a node set, which is a set of all power grid nodes and traffic network nodes, E represents an edge set, which is a set of connection relationships between nodes; an adjacency matrix and a node feature matrix are constructed based on the node set and the edge set; S2, inputting the adjacency matrix and the node feature matrix into a graph convolutional neural network model, training a feature extraction model through a semi-supervised learning mechanism, and outputting an updated node feature matrix; S3, learning an optimal strategy by using a deep Q network, and identifying a vulnerable node sequence by interacting with the environment.

[0007] Optionally, in S1, the adjacency matrix adopts a block matrix form, comprising: a power grid intra-edge adjacency matrix for describing the connection relationship between power nodes; a traffic network intra-edge adjacency matrix for describing the connection relationship between traffic nodes; an inter-network coupling edge adjacency matrix for describing the connection relationship between power nodes and traffic nodes.

[0008] Optionally, in the graph convolutional neural network model, the feature conversion formula of the first layer to the l +1 layer is: l (4) In the formula: is the node feature matrix of the l layer, is the feature dimension, , is an adjacency matrix with a self-loop, I is an identity matrix, which makes the node's own attribute participate in convolution, is a corresponding degree matrix; is the weight matrix of the l layer; is a bias term; is an activation function.

[0009] Optionally, in S2, the feature extraction model training adopts a semi-supervised loss function, and the semi-supervised loss function comprises a cross-entropy loss of labeled nodes and an L2 regularization term; the semi-supervised loss function is defined as: (5)​ wherein: is a set of labeled nodes; C is the number of classification categories; is a set of nodes v with true labels; is the model prediction probability; is the regularization coefficient; is all the learnable parameters of the model.

[0010] Optionally, in S2, the feature extraction model training comprises: Initialization: randomly initialize the weight matrix and bias term of each layer, set hyperparameters, the hyperparameters include at least one of the number of graph convolution layers, feature dimension, regularization coefficient; Forward propagation: input the node feature matrix and adjacency matrix into the model, calculate the feature of each layer by formula (4), and finally obtain the model prediction probability; Loss calculation: calculate the semi-supervised loss by formula (5); Back propagation: update the model parameters by gradient descent method to minimize the loss function; Iteration termination: when the loss converges or reaches the maximum number of iterations, output the node feature matrix and the classification probability matrix : (6) wherein: is the output layer weight matrix; is the output layer bias term; ensure that the probability sum of each row is 1.

[0011] Optionally, step S3 comprises: modeling the fragile node identification problem as a Markov decision process, training with a deep Q network and an environment defined by the Markov decision process, learning an optimal node evaluation strategy, and thus identifying a fragile node sequence; wherein the state space is constructed based on the node feature matrix output by step S2, the action space is the selection action on the node to be evaluated, and the reward function is designed based on the degree of attenuation of the overall performance of the power-traffic coupled network after the selected node fails.

[0012] Optionally, in step S3, the node vulnerability quantitative evaluation index is defined as the relative attenuation rate of the coupled network system performance after the node fails, and the formula is: (7) wherein: is the node v contribution to the vulnerability of the power grid; is the node v contribution to the vulnerability of the traffic network; is the nodev initiated cross-network cascading vulnerability; are weight coefficients, respectively reflecting the relative importance of the power grid, the traffic network and the coupling effect, and satisfying .

[0013] Optionally, the method further comprises: Step S4, vulnerability visualization: marking and / or displaying the vulnerable nodes identified in step S3 in a visualization interface.

[0014] The application also provides a power-traffic coupling network vulnerability identification system, comprising: a data acquisition and processing module for acquiring raw data of the power system and the traffic system and performing data cleaning and standardization; a topology generation module for abstracting the power-traffic coupling network into a graph model, generating an adjacency matrix and a node feature matrix; a feature extraction module comprising a graph convolution network for extracting multi-level feature embeddings of nodes; a vulnerability identification module comprising a deep Q network for modeling the identification of vulnerable nodes as a Markov decision process and outputting a sequence of vulnerable nodes.

[0015] Optionally, the system further comprises a visualization module for displaying the results of the identification of vulnerable nodes.

[0016] Compared with the prior art, the application has the following beneficial effects: 1) Breakthrough traditional single network evaluation mode, realize accurate quantification of cross-network coupling effect. Through the block adjacency matrix and the cascade item in the vulnerability index, the interaction between the power and traffic networks is explicitly modeled, improving the comprehensiveness and accuracy of vulnerability identification, especially in extreme disaster scenarios, avoiding the problem of ignoring cross-network cascading failure in traditional methods.

[0017] 2) Realize intelligent dynamic identification, reduce dependence on artificial experience. Provide a semi-supervised learning mechanism of GCN, only a small amount of labels are needed; combine the MDP framework of DQN, so that the system can learn the optimal strategy through interaction; the identification process is more adaptive and efficient, and can adapt to network dynamic changes, overcoming the limitations of traditional methods relying on expert knowledge and static analysis.

[0018] 3) Improve feature extraction capability, can capture complex topological information. The multi-layer convolution operation of GCN can aggregate node neighborhood information to generate node embeddings rich in topological context; thus, it can effectively handle the heterogeneity of heterogeneous nodes (such as substations and intersections), capture complex topological information, and make the vulnerability identification more consistent with the actual network structure.

[0019] 4) Further, decision visualization can be supported, and the system is practical. Through the visualization module, a geographic information system can be integrated, and the distribution of vulnerable nodes can be intuitively displayed; thus, intuitive insights can be provided for operation and maintenance personnel, and it is convenient to quickly develop protection measures, and the resilience management efficiency of the power-traffic coupled system is improved. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 FIG. 1 is a structural schematic diagram of a power-traffic coupled network vulnerability identification system according to the present application.

[0021] Figure 2 FIG. 2 is a flowchart of a power-traffic coupled network vulnerability identification method according to the present application.

[0022] Figure 3 FIG. 3 is a flowchart of a fault node intelligent identification method based on a hypergraph (graph model) and DQN (Deep Q-Network, DQN) fusion according to the present application.

[0023] Figure 4 FIG. 4 is an IEEE-33 node topology diagram of an embodiment.

[0024] Figure 5 FIG. 5 is a traffic network topology diagram of an embodiment.

[0025] Figure 6 FIG. 6 is a training reward curve diagram of an embodiment.

[0026] Figure 7 FIG. 7 is a vulnerable node identification result of an embodiment. DETAILED DESCRIPTION

[0027] The power-traffic coupled network vulnerability identification method and system according to the present application will be further described in detail below in combination with the accompanying drawings and specific embodiments. The advantages and features of the present application will be more apparent according to the following description. It should be noted that the accompanying drawings are greatly simplified and all use non-precise proportions, only for the purpose of facilitating and clarifying the purpose of assisting the description of the embodiments of the present application. In order to make the purpose, features and advantages of the present application more apparent and easy to understand, please refer to the accompanying drawings. It should be noted that the structure, proportion, size, etc. shown in the drawings attached to the present specification are only used to cooperate with the content disclosed in the specification, so as to be understood and read by those skilled in the art, and do not have technical significance, any modification of structure, change of proportion relationship or adjustment of size, without affecting the effect and purpose that can be achieved by the present application, should still fall within the scope of the technical content disclosed by the present application.

[0028] In order to solve the problems in the prior art, such as inability to effectively handle the coupling effect across the network, insufficient topological characterization capability, inability to dynamically learn the network state change, lack of sequence decision-making capability, and difficulty in accurately positioning the vulnerable nodes in the catastrophic scenario, the inventors have carried out in-depth research on the power-traffic coupled network vulnerability identification method combining graph convolution and reinforcement learning, and proposed a combination of graph convolutional neural network (GCN) and reinforcement learning (DQN) to construct an end-to-end vulnerability identification framework: the graph convolutional neural network (GCN) is responsible for feature extraction, and the reinforcement learning (DQN) is responsible for sequence decision-making. Then, through the data acquisition, topology generation, feature extraction, vulnerability identification and visualization modules, a closed-loop system is integrated.

[0029] As shown in Figure 1 A power-traffic coupled network vulnerability identification system 100 according to the present application includes a data acquisition and processing module 10, a topology generation module 20, a feature extraction module 30, a vulnerability identification module 40, and optionally, a visualization module 50.

[0030] The data acquisition and processing module 10 is used to acquire the original data of the power system and the traffic system, and to perform data cleaning and standardization.

[0031] In some embodiments, the power system data and the traffic network related data of a specified area are collected. The collected data is cleaned, preprocessed and standardized, abnormal values and error data are removed, and data from different sources are converted into a unified format and scale.

[0032] The topology generation module 20 is used to abstract the power-traffic coupled network into a graph model, and to generate an adjacency matrix and a node feature matrix.

[0033] In some embodiments, based on the data acquired by the data acquisition and processing module 10, the power-traffic coupled network is abstracted into a unified graph model, the connection relationship between the power grid, the traffic network and the two networks is described through the adjacency matrix block form, and the node feature matrix is generated to describe the attribute characteristics of each node, so as to realize complete modeling and characterization of the topology structure of the power-traffic coupled network.

[0034] The feature extraction module 30 includes a graph convolutional network, which is used to extract multi-level feature embedding of nodes.

[0035] In some embodiments, based on the network topology structure generated by the topology generation module 20, the GCN and the semi-supervised learning mechanism are used to realize deep extraction of the node topology features and attribute features of the power-traffic coupled network. After iterative training, the node embedding matrix and the classification probability matrix containing multi-level features are finally output, which provides structured feature input for subsequent intelligent evaluation of vulnerable nodes.

[0036] The fragile recognition module 40 comprises a deep Q network for modeling the fragile node recognition as a Markov decision process and outputting a fragile node sequence.

[0037] In some embodiments, the fragile node recognition problem is modeled as a Markov decision process based on the node features output by the feature extraction module 30, optimized by DQN training, and an optimal recognition strategy is formed in iterative learning by using a neural network to approximate the Q value function, so as to accurately locate the fragile node set that has the greatest impact on system safety.

[0038] The visualization module 50 is used to display the fragile node recognition results.

[0039] In some embodiments, the data and analysis results output by the fragile recognition module 40 are integrated to generate an intuitive and easy-to-understand evaluation analysis report. Through various visualization means such as geographic information system maps, charts, and graphs, the running state of the power-traffic coupled network and the distribution of fragile nodes and other information are displayed, so that users can intuitively and quickly understand the running state and fault information of the power-traffic coupled system, and provide a scientific basis for decision-making.

[0040] As shown in Figure 2 Fig. 1 is a flowchart of a power-traffic coupled network fragile point recognition method according to an embodiment of the present application, which comprises: S1, coupling the power network and the traffic network to form a graph model: (1) In formula (1): V represents a node set, which is a set of all power grid nodes and traffic network nodes, E represents an edge set, which is a set of connection relationships between nodes; based on the node set and the edge set, an adjacency matrix and a node feature matrix are constructed.

[0041] The adjacency matrix adopts a block matrix form, which comprises: a power grid intra-edge adjacency matrix for describing the connection relationship between power nodes; a traffic network intra-edge adjacency matrix for describing the connection relationship between traffic nodes; a network inter-coupling edge adjacency matrix for describing the connection relationship between power nodes and traffic nodes.

[0042] S2, inputting the adjacency matrix and the node feature matrix into a graph convolutional neural network model, training the feature extraction model through a semi-supervised learning mechanism, and outputting an updated node feature matrix.

[0043] Furthermore, the topology, node attributes, and inter-network dependencies of the power-transportation coupled network are encoded using a graph convolutional network (GCN) to generate node embeddings containing multi-level features, and output node feature matrices and classification probability matrices.

[0044] S3 utilizes a deep Q-network to learn the optimal strategy and identifies vulnerable node sequences through interaction with the environment.

[0045] The identification of vulnerable nodes is transformed into a sequential decision problem. Based on feature embedding, Deep Q-Network (DQN) is used to learn the optimal identification strategy through dynamic interaction with the network environment, and finally locate vulnerable nodes in the power-transportation coupled network.

[0046] The node vulnerability quantification assessment index is defined as the relative degradation rate of the coupled network system performance after node failure. The node vulnerability quantification assessment index includes a weighted comprehensive assessment of power grid performance, transportation network performance, and cross-network cascading vulnerability.

[0047] The deep Q-network adopts a target network and an online network structure, and optimizes the Q-value function through empirical replay and gradient descent.

[0048] The following detailed description is provided in conjunction with specific examples.

[0049] Example like Figure 3 As shown, the flowchart of the intelligent fault node identification method based on the fusion of hypergraph and DQN (Deep Q-Network, DQN) of the present invention includes two parts: graph feature extraction and DQN training process.

[0050] Graph feature extraction methods include: Step 1: Propose an architecture for vulnerability identification in power-transportation coupled networks that integrates graph convolution and reinforcement learning. Specifically, this includes: Step 1.1 Feature extraction of coupled network based on graph convolution: The topology, node attributes and inter-network dependencies of the power-transportation coupled network are encoded by a graph convolutional neural network (GCN) to generate node embeddings containing multi-level features; Step 1.2 Vulnerable Node Evaluation Method Based on DQN (Deep Q-Network): Vulnerable node identification is transformed into a sequential decision problem. Based on feature embedding, the deep Q-network learns the optimal identification strategy through dynamic interaction with the network environment, and finally locates the vulnerable nodes in the power-transportation coupled network.

[0051] Step 2: Establish a coupling network feature extraction method based on graph convolution. Specifically, it includes: Step 2.1, power-traffic coupling network graph structured representation The power-traffic coupling network is a complex heterogeneous system formed by the power system and the traffic system through specific nodes and interaction relationships, which can be abstracted as a unified graph model, i.e. the coupling forms the following graph model: (1) In the formula: V represents the node set, which is the set of all power grid nodes and traffic network nodes, E represents the edge set, which is the set of connection relationships between nodes.

[0052] Node set is an abstraction of all physical entities in the coupling network, covering the core elements of the power grid and the traffic network, and the two types of nodes (power grid nodes and traffic network nodes) have significant heterogeneity in function and attribute. can be divided into power grid node subset (including substations, charging piles, power transmission line endpoints, etc.) and traffic network node subset (including intersections, road segments, parking lots, etc.), which satisfies: (2) Based on the node set and edge set, the adjacency matrix and node feature matrix are constructed, as follows: To more simply describe the connection relationship between all nodes, the adjacency relationship of the network edge and the coupling edge can be integrated into a unified adjacency matrix , where is the total number of nodes in the coupling network, if there is an adjacency relationship between nodes, then , otherwise , i, j represent adjacent nodes. The adjacency matrix can simultaneously reflect the connection relationship within the power grid, within the traffic network and between the two networks through the block form, realizing the complete characterization of the topology of the entire coupling network. The adjacency matrix A adopts a block matrix form: (3) In the formula: is the power grid internal edge adjacency matrix, describing the connection relationship between power nodes; is the traffic network internal edge adjacency matrix, describing the connection relationship between traffic nodes; is the network inter-coupling edge adjacency matrix, describing the connection relationship between power nodes and traffic nodes; is the transpose matrix of , ensuring the symmetry of the adjacency matrix.

[0053] Node feature matrix X Defined as In the formula: d The feature dimension is the number of feature metrics for each node, and |V| represents the total number of nodes in the coupled network. The matrix elements are... Indicates the first v The node of the first k Item attribute value.

[0054] Step 2.2 Feature extraction model based on graph convolution Based on the power-transportation coupled network graph model proposed in step 2.1, a semi-supervised graph convolutional model is used to extract node features and node classification is achieved using a small amount of label information. The model construction needs to combine the graph structure features of the coupled network, capture the local topology and attribute information of nodes through graph convolutional layers, and optimize the model parameters using a semi-supervised loss function.

[0055] (1) Graph convolution model that captures the graph structure features of coupled networks Node feature matrix X With the integrated adjacency matrix A The input model is used to design an improved graph convolutional layer to address the heterogeneity of the power-transportation coupled network. This layer captures the features of different node types through weight sharing and type differentiation. l layer to the first l The feature transformation formula for layer +1 is: (4) In the formula: For the first l The node feature matrix of the layer For feature dimension, . To add a self-loop adjacency matrix, I It is an identity matrix, which allows the node's own attributes to participate in the convolution. for The corresponding degree matrix; For the first l Layer weight matrix; For bias terms; For activation functions, ReLU is typically used.

[0056] (2) Semi-supervised learning mechanism The model employs a semi-supervised loss function, optimizing the overall parameters through a small number of key nodes with known vulnerability levels. The loss function is defined as follows: (5) In the formula: A collection of labeled nodes; C The number of categories; For nodesv the true label of the node; the predicted probability of the model; the regularization coefficient; all the learnable parameters of the model; the first term is the cross-entropy loss (only for labeled nodes), and the second term is the L2 regularization (to prevent overfitting).

[0057] After the graph convolution operation, the output node feature matrix and the classification probability matrix are obtained: (6) where: is the output layer weight matrix; is the output layer bias term; to ensure that the probability sum of each row is 1.

[0058] (3) Feature extraction model training process 1) Initialization: randomly initialize the weight matrix and bias term of each layer, set the number of graph convolution layers, feature dimension, regularization coefficient, and other hyperparameters; 2) Forward propagation: input the node attribute matrix X and the adjacency matrix A into the model, calculate the features of each layer through equation (4), and finally obtain the predicted probability.

[0059] 3) Loss calculation: calculate the semi-supervised loss using equation (5), only considering the prediction error of labeled nodes.

[0060] 4) Backpropagation: update the model parameters through gradient descent method to minimize the loss function.

[0061] 5) Iteration termination: when the loss converges or reaches the maximum number of iterations, output the final node features and classification results.

[0062] Step 3: Construct a DQN-based vulnerable node evaluation method. Specifically, it includes: Step 3.1 Vulnerability quantification evaluation index Node vulnerability is defined as the relative attenuation rate of the coupled network system performance after the failure of the node v , and the formula is: (7) where: is the vulnerability contribution of node v to the power grid; is the vulnerability contribution of node v to the transportation network; is the cross-network cascading vulnerability caused by node v ; are weight coefficients, respectively, reflecting the relative importance of power grid, traffic network and coupling effect, satisfying .

[0063] When a node fails, if v is a power node, the performance degradation of power grid is jointly determined by the loss of power supply capacity and the decline of topology connectivity; if v is a traffic node, the performance degradation of traffic network is determined by the loss of traffic capacity and the decline of road network accessibility; if v is a coupling node, the cascading vulnerability quantifies the cross-network cascading reaction triggered by the node failure through the coupling edge, and the calculation formula is as follows (8) (9) (10) In the formula, is the total normal power supply of power grid (unit: MW); is the total power supply after the failure of node v ; is the normal global efficiency of power grid, reflecting the topology connectivity, defined as the average reciprocal of the shortest path length of all node pairs; is the global efficiency after the failure of node v ; is the weight, balancing the influence of functional loss and topology damage. is the total normal traffic capacity of traffic network (unit: vehicle / hour); is the total traffic capacity after the failure of node v ; is the normal global efficiency of traffic network; is the global efficiency after the failure of node v ; is the weight. is the set of coupling nodes; is the adjacency matrix element; is the coupling edge weight, reflecting the relationship between power supply and demand intensity; is the performance degradation coefficient of node u affected by the failure of node v , with a value range of [0, 1], and 1 represents complete failure.

[0064] Step 3.2 Markov decision process Markov decision process (MDP) is a mathematical framework for modeling sequential decision problems in a stochastic environment, whose core is that the future state of the system only depends on the current state and action, and is independent of the historical state.

[0065] Node representation based on feature extraction Hyperedge representation After that, we construct the fragile node identification problem as a Markov Decision Process (MDP), whose input includes node representation, hyperedge representation and current network topology state. The construction process of MDP is as follows: 1) State space : the information of the external environment where the agent is located.

[0066] The constructed state space needs to accurately describe the state of dynamic coding disaster propagation.

[0067] (11) wherein: is the node propagation feature, is the hyperedge dynamic feature, is the real-time load vector.

[0068] 2) Action space : the action to be taken by the agent after perceiving the external environment.

[0069] When the action is implemented, the agent needs to execute node destruction and update the network topology.

[0070] (12) wherein: is the destroyed node of discrete action.

[0071] 3) Reward function : the feedback value obtained by the agent after perceiving the external environment and taking action.

[0072] The reward is designed according to the cascading failure consequences caused by node damage.

[0073] (13) Step 3.3 DQN training process After constructing the MDP, the DQN network is trained, and the agent updates the network according to experience, finally learns the strategy of destroying nodes and identifies the fragile nodes.

[0074] DQN is an improvement of the traditional deep Q network algorithm. DQN uses a neural network to approximate the Q function, and the input state s outputs the Q value of all actions.

[0075] (14) wherein: is the neural network parameter.

[0076] The objective of DQN is to make the Q value output by the Q network as close as possible to the ideal target Q value, and the objective function thereof is: (15) In the formula: is the parameter of the target network The loss function thereof is: (16) To verify the effectiveness of the method proposed in the application, this embodiment constructs a power-traffic deep coupling simulation environment, including an IEEE-33 node power distribution network system and a 4x5 node traffic network. The IEEE-33 node topology is shown in Figure 4 . The traffic network topology is shown in Figure 5 . All simulations in the application are verified in a server configuration of AMD Ryzen 9 7700X, NVIDIA RTX4070, RAM 8GB, and a simulation software Matlab2024b environment.

[0077] Figure 6 The reward value of each round of training of the proposed algorithm is shown. As can be seen from the figure, when the average reward rises to about -250 at 0 to 200 rounds, the error range is wide, reflecting that the model is in the early stage of policy exploration, the reward fluctuates greatly and is insufficiently stable; when the average reward is at 200-700 rounds, the average reward enters a rapid and relatively stable rising period, gradually approaching -5 from about -250; after 700 rounds, the error range converges significantly, indicating that the learning effect of the model on the optimal strategy improves rapidly, the reward fluctuation degree is greatly reduced, the average reward is basically maintained at a level close to -5, and there is no obvious rise, and the error range further converges to a very small interval, indicating that the model has learned a relatively optimal strategy and the reward stability has reached a high level. The average reward rises rapidly in the early stage, gradually slows down and tends to be stable in the later stage, and the error range continuously decreases, indicating that the learning effect of the model on the strategy gradually improves in the training process, the reward fluctuation degree decreases and the stability enhances.

[0078] The fragile node identification result is shown in Figure 7 . There are three fragile nodes P in the simulation result, respectively located at nodes numbered 5, 9 and 26. When the coupled network is impacted, these power nodes are critical interaction or transmission nodes in the network topology and the traffic network, and have high vulnerability. If failure occurs, it will have a significant impact on the overall operation of the power-traffic coupled network and the cross-network cascading effect.

[0079] In summary, the power-traffic coupling network vulnerability identification method architecture combining fusion graph convolution and reinforcement learning provided by the application constructs a full-chain analysis framework, breaks through the traditional single network vulnerability node evaluation mode, and provides a systematic logical architecture for node vulnerability quantization analysis of a coupling network in a catastrophic scenario. Further, the application combines graph theory and semi-supervised learning mechanism to depict multiple types of nodes and complex correlations of the power-traffic network, establishes the vulnerable node identification problem as an MDP decision process, and combines DQN to construct a multi-objective reward mechanism to drive the agent to dynamically identify vulnerable nodes in the hypergraph topology space, break through the dependence on experience of traditional methods and the neglect of cross-network coupling failure, and realize intelligent evaluation and accurate positioning of vulnerable nodes.

[0080] It should be noted that in this document, the terms "comprise", "comprise", or any other variant thereof are intended to cover non-exclusive inclusion, so that processes, methods, articles or devices that include a series of elements not only include those elements, but also include other elements not explicitly listed, or include elements inherent to such processes, methods, articles or devices. Without more limitations, the element defined by the statement "comprises a" does not exclude the presence of additional identical elements in the process, method, article or device that includes the element.

[0081] In the description of the application, it should be understood that the orientation or positional relationship indicated by the terms "center", "height", "thickness", "upper", "lower", "vertical", "horizontal", "top", "bottom", "inner", "outer", "axial", "radial", "circumferential" and the like is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the application and simplifying the description, and does not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the application. In the description of the application, unless otherwise specified, the meaning of "a plurality of" is two or more.

[0082] In the description of the application, unless otherwise specified and limited, the terms "mounting", "connection", "connection", "fixing" should be understood broadly, for example, it can be fixed connection, or detachable connection, or integral; it can be mechanical connection, or electrical connection; it can be directly connected, or indirectly connected through an intermediate medium; it can be the internal communication of two elements or the interaction relationship between two elements. For those skilled in the art, the specific meaning of the above terms in the application can be understood according to the specific circumstances.

[0083] In the present application, unless specifically stated and limited otherwise, the "on" or "under" of a first feature to a second feature can include that the first and second features are in direct contact, or that the first and second features are not in direct contact but are in contact through another feature between them. Moreover, the "on", "above" and "on top of" of a first feature to a second feature includes that the first feature is directly above and obliquely above the second feature, or only means that the first feature is horizontally higher than the second feature. The "under", "below" and "underneath" of a first feature to a second feature includes that the first feature is directly below and obliquely below the second feature, or only means that the first feature is horizontally lower than the second feature.

[0084] Although the present application has been described in detail by the foregoing preferred embodiments, it should be recognized that the foregoing description is by way of example only and that various modifications and substitutions can be made by those skilled in the art without departing from the scope of the present application. Accordingly, the scope of the present application should be limited only by the appended claims.

Claims

1. A method for identifying vulnerable points of a power-transport coupled network, characterized in that, The method comprises the following steps: S1, coupling the power network and the traffic network to form a graph model: (1) In formula (1): V denotes a node set, which is a set of all power grid nodes and traffic network nodes, E denotes an edge set, which is a set of connection relationships between nodes; an adjacency matrix and a node feature matrix are constructed based on the node set and the edge set; S2, inputting the adjacency matrix and the node feature matrix into a graph convolutional neural network model, training a feature extraction model through a semi-supervised learning mechanism, and outputting an updated node feature matrix; S3, learning an optimal strategy by using a deep Q network, and identifying a vulnerable node sequence by interacting with an environment.

2. The method of claim 1, wherein the method further comprises: identifying a set of critical nodes in the power-transportation coupled network; and identifying a set of critical links in the power-transportation coupled network. In S1, the adjacency matrix adopts a block matrix form, comprising: an intra-edge adjacency matrix of the power network, used to describe the connection relationship between power nodes; an intra-edge adjacency matrix of the traffic network, used to describe the connection relationship between traffic nodes; an inter-network coupling edge adjacency matrix, used to describe the connection relationship between power nodes and traffic nodes.

3. The method of claim 1, wherein, In the graph convolutional neural network model, the feature conversion formula of the first layer to the first l layer to the first l +1 layer is as follows: （4） In the formula: For the first l The node feature matrix of the layer For feature dimension, , To add a self-loop adjacency matrix, I Given an identity matrix, the node's own attributes are used in the convolution. for The corresponding degree matrix; For the first l Layer weight matrix; For bias terms; This is the activation function.

4. The method of claim 3, wherein, In S2, the feature extraction model training adopts a semi-supervised loss function, wherein the semi-supervised loss function comprises a cross-entropy loss of labeled nodes and an L2 regularization term; and the semi-supervised loss function is defined as: （5） where: is the set of labeled nodes; C is the number of classification classes; is the set of nodes v is the true label of the node; is the model prediction probability; is the regularization coefficient; is all the learnable parameters of the model.

5. The method for identifying vulnerabilities in a power-transportation coupled network as described in claim 4, characterized in that, In S2, the feature extraction model training comprises: initialization: randomly initializing weight matrices and bias terms of each layer, and setting hyperparameters, wherein the hyperparameters comprise at least one of the number of graph convolutional layers, the feature dimension, and the regularization coefficient; forward propagation: inputting the node feature matrix and the adjacency matrix into the model, calculating features of each layer through formula (4), and finally obtaining a model prediction probability; loss calculation: calculating a semi-supervised loss by using formula (5); back propagation: updating model parameters by using a gradient descent method, and minimizing the loss function; Iteration termination: output node feature matrix when loss converges or reaches maximum number of iterations and classification probability matrix : （6） where: is the output layer weight matrix; is the output layer bias term; ensuring that the probability sum for each row is 1.

6. The method of claim 1, wherein, Step S3 comprises: modeling the vulnerable node identification problem as a Markov decision process, interacting with an environment defined by the Markov decision process by using a deep Q network, learning an optimal node evaluation strategy, and thus identifying a vulnerable node sequence; wherein a state space is constructed based on the node feature matrix output by step S2, an action space is a selection action on a node to be evaluated, and a reward function is designed based on the degree of attenuation of the overall performance of the power-traffic coupled network after the selected node fails.

7. The method of claim 6, wherein, In step S3, a node vulnerability quantitative evaluation index is defined as a relative attenuation rate of the performance of the coupled network system after the node fails, and the formula is: （7） In the formula: is a node v contribution to the vulnerability of the power grid; is a node v contribution to the vulnerability of the traffic network; is a node v induced cross-network cascading vulnerability; is a weight coefficient, respectively reflecting the relative importance of the power grid, the traffic network and the coupling effect, satisfying .

8. The method of claim 1, wherein, Further comprising: Step S4, vulnerable point visualization: marking and / or displaying the vulnerable nodes identified in step S3 in a visualization interface.

9. A power-traffic coupled network vulnerability identification system for implementing the power-traffic coupled network vulnerability identification method according to any one of claims 1-8, characterized in that, The system comprises: a data acquisition and processing module, configured to acquire original data of a power system and a traffic system, and perform data cleaning and standardization; a topology generation module, configured to abstract the power-traffic coupled network into a graph model, and generate an adjacency matrix and a node feature matrix; a feature extraction module, comprising a graph convolutional network, configured to extract multi-level feature embeddings of nodes; a vulnerability identification module, comprising a deep Q network, configured to model the vulnerable node identification as a Markov decision process and output a vulnerable node sequence.

10. The power-traffic coupled network vulnerability identification system of claim 9, wherein, Further comprising: a visualization module, configured to display the vulnerable node identification result.

Images