Electric power traffic coupling network sentry node identification method and system

Abstract

The invention discloses an electric power traffic coupling network sentry node identification method and system, and the method comprises the steps: carrying out the value quantification and screening of collected electric power traffic multi-source heterogeneous data based on an optimized gradient value, and outputting high-value data; the method comprises the following steps: constructing a unified hypergraph model of an electric power traffic coupling network, defining electric power nodes, traffic nodes and charging station nodes as a heterogeneous node set, constructing multiple types of hyperedges according to physical connection, flow similarity and a cross-network supply-demand relationship, and extracting a node high-order topological feature embedding matrix by using a hypergraph convolutional neural network; based on a reinforcement learning algorithm, a sentinel node identification problem is modeled as a Markov decision process, and a sentinel node set enabling a normalized mean square error function to be minimum is dynamically searched and output. According to the method, in a disaster environment with high data noise and limited computing resources, the local key nodes are utilized to accurately invert the guard node deployment of the whole-network macroscopic state.

Description

Technical Field

[0001] This invention belongs to the field of smart grid and transportation network integration technology, and particularly relates to a method and system for identifying sentinel nodes in a power-transportation coupled network. Background Technology

[0002] With the deep integration of the energy internet and intelligent transportation systems, the coupling between power systems and transportation systems is becoming increasingly complex. On the one hand, the power grid and transportation network have established a close physical connection through charging facilities. The cross-grid movement of electric vehicles means that the spatiotemporal distribution of traffic flow directly determines the fluctuation characteristics of charging load. This "vehicle-electricity" interaction makes the operation of one network highly dependent on the other. On the other hand, in disaster scenarios such as natural disasters or sudden attacks, the coupling relationship between the two networks changes from "cooperation" to "risk propagation." Communication interruptions and sensor failures can lead to high noise, labeling errors, and feature redundancy in the collected power flow and traffic flow data, making traditional situational awareness methods based on high-quality data difficult to apply.

[0003] Current monitoring and control of such coupled networks remain in an "idealized mode": existing network modeling methods are mostly based on simple graph structures, making it difficult to depict the "one-to-many" high-order coupling relationship between charging stations and multiple road segments and distribution nodes; moreover, state assessment relies on the entire network's data, ignoring the extreme scarcity of computing resources in disaster environments. In extreme disaster scenarios, traditional monitoring strategies reveal even more significant shortcomings: due to the lack of a value screening mechanism for low-quality data, noisy data input leads to extremely large model prediction biases; and the computational complexity of traditional key node identification methods (such as Shapley values) increases exponentially with network size, resulting in computational lag in emergency scenarios requiring millisecond-level responses, preventing decision centers from timely grasping the overall network situation. This perception blind spot caused by low data quality and slow computational efficiency necessitates the development of a lightweight identification strategy based on sentinel nodes in complex coupled environments to utilize local high-value information for accurate inversion of the entire network's state.

[0004] The statements herein provide only background information in relation to this invention and do not necessarily constitute prior art. Summary of the Invention

[0005] The main objective of this invention is to overcome the shortcomings and deficiencies of the prior art and propose a method and system for identifying sentinel nodes in a power-transportation coupled network under disaster scenarios. This method can accurately invert the deployment of sentinel nodes in the overall network macroscopic state by utilizing local key nodes in disaster environments with high data noise and limited computing resources.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution: A method for identifying sentinel nodes in a power-transport coupled network includes: Step S1: Based on the optimized gradient value, the collected multi-source heterogeneous data of power and transportation are valued and filtered to output high-value data.

[0007] Step S2: Construct a unified hypergraph model for the power-transportation coupled network, which defines power nodes, transportation nodes and charging station nodes as a heterogeneous set of nodes. Based on physical connections, traffic similarity and cross-network supply and demand relationships, construct multiple types of hyperedges, and use hypergraph convolutional neural networks to extract high-order topological features of nodes and embed them into the matrix.

[0008] Step S3: Based on the reinforcement learning algorithm, the sentinel node identification problem is modeled as a Markov decision process, and the set of sentinel nodes that minimizes the normalized mean square error function is dynamically searched and output.

[0009] Optionally, step S1 includes: Step S1.1: Using a strategy that combines Monte Carlo approximation with single-sample gradient descent, calculate the optimized gradient value of each data point in the multi-source heterogeneous data of power and transportation.

[0010] Step S1.2: Compare the optimized gradient value with the set noise filtering threshold and redundancy filtering threshold respectively to filter the multi-source heterogeneous data of power and transportation and obtain high-value data.

[0011] Step S1.3: Using the high-value data, construct power grid topology maps respectively. Transportation network map Map showing the distribution of charging facilities .

[0012] The multi-source heterogeneous data on power and transportation includes power grid flow data, transportation network traffic data, and charging network operation data in the early stages of the disaster.

[0013] Optionally, step S1.1 includes: The i-th data point in the multi-source heterogeneous data of power and transportation is calculated using the following formula. Optimization gradient value , In the formula, M The cumulative number of Monte Carlo samplings is used to ensure the statistical significance of the value assessment; For the first m In the random permutation, located at the data point Previous data subset; This is the model utility function.

[0014] Step S1.2 includes: setting a noise filtering threshold. and redundant filtering thresholds : like Determine the data point Data with incorrect labels or maliciously injected data will be removed. like Determine the data point Redundant features that have no significant impact on model performance are removed.

[0015] Step S1.3 includes: Using the high-value data retained after filtering, power grid flow data, traffic network flow data, and charging network operation data are extracted respectively; the node attributes and line connection relationships in the power grid flow data are analyzed to construct a power grid topology map. ; Analyze the road segment connectivity and geographical location information in the traffic network traffic data to construct a traffic network map. ; Analyze the spatial coordinates and equipment configuration information of charging stations in the charging network operation data to construct a charging facility distribution map. .

[0016] Optionally, step S2 includes: Step S2.1: Using the aforementioned power grid topology diagram The aforementioned transportation network diagram and the distribution map of the charging facilities Based on this, a unified hypergraph model of the power-transportation coupled network is constructed. In the formula, V Represents a set of heterogeneous nodes. E Denotes the set of superedges. H This represents the correlation matrix of the unified hypergraph model of the power-transportation coupled network. X This represents the node feature matrix.

[0017] Among them, the set of heterogeneous nodes Includes: power node set Charging station node set and traffic node set .

[0018] Power node set This includes: generator nodes, substation nodes, and ordinary load nodes.

[0019] Charging station node set As a coupling hub.

[0020] Traffic node set This includes: road intersections and road segment endpoints.

[0021] Hyperedge set The following three types of hyperedges can represent higher-order one-to-many relationships: Power topology superedge Traffic similarity hyperedge and cross-network coupling hyperedge ; Traffic similarity hyperedge In n This represents the total number of traffic similarity hyperedges constructed. Indicates the first Traffic similarity superedge.

[0022] The correlation matrix of the unified hypergraph model of the power-transportation coupled network This is used to mathematically represent the relationship between heterogeneous nodes in a set of heterogeneous nodes and hyperedges in a set of hyperedges, where... Represents the set of real numbers; Represents a set of heterogeneous nodes The total number of nodes in; Represents the set of superedges The total number of superedges in the array.

[0023] Step S2.2: Construct a node feature matrix of physical properties and operating states of heterogeneous nodes. ,in, Represents the set of real numbers. The total number of all heterogeneous nodes. For node feature dimensions.

[0024] Step S2.3: Use a hypergraph convolutional neural network to simulate the propagation process of disaster impact in the power and transportation coupled network. Through the alternating propagation mechanism of heterogeneous node to hyperedge aggregation and hyperedge to heterogeneous node update, feature learning is performed to extract the high-order topological features of heterogeneous nodes and output the node high-order topological feature embedding matrix.

[0025] Optionally, step S2.2 includes: constructing normalized feature vectors for the three types of heterogeneous node sets in the heterogeneous node set, and concatenating them according to the node index to form a global node feature matrix, thus forming the node feature matrix. .

[0026] Step S2.2.1: Construct the feature vector of the power node: For any power node Its eigenvector is defined as In the formula, The node load factor represents the proportion of the current power load to the rated capacity, reflecting the operating margin of the power node. The upper limit of power supply capacity characterizes the physical power supply or transmission limit of a power node and determines its support potential in disaster recovery. Voltage amplitude represents the voltage stability of power nodes and is a key indicator for determining whether a power system exceeds its limits.

[0027] Step S2.2.2: Construct the feature vector of the charging station node: For any charging station node As a coupling hub connecting power and transportation, its feature vector is defined as follows. In the formula, The number of charging piles represents the static service capacity base of the charging station node; This provides real-time facility utilization, reflecting the current level of activity and queuing congestion risk at charging station nodes. The real-time total charging load not only represents its power extraction demand from the power network, but also indirectly reflects the intensity of traffic flow convergence at this charging station node.

[0028] Step S2.2.3: Construct traffic node feature vectors: For any traffic node Its eigenvector is defined as In the formula, The road grade coefficient quantifies the road type and determines the upper limit of traffic capacity and the benchmark for traffic speed at traffic nodes. Cross-sectional traffic flow reflects the number of vehicles passing through the traffic node per unit time and characterizes the dynamic distribution of traffic flow. The Road Congestion Index is calculated based on the ratio of average vehicle speed to free-flow vehicle speed, representing traffic efficiency and vehicle congestion.

[0029] Step S2.2.4 Stack all the feature vectors of power nodes, charging station nodes, and traffic nodes according to the global index order of the corresponding heterogeneous nodes in the unified hypergraph model of the power-transport coupling network to form the initial node feature matrix. X .

[0030] Step S2.3 includes: converting the node feature matrix X and the correlation matrix H Used as input to a hypergraph convolutional neural network for feature updates: Step S2.3.1, Heterogeneous node to hyperedge aggregation: The information of each heterogeneous node is determined based on the correlation matrix. They converge at their superedge, forming a process l The hyperedge features output after processing by a layer hypergraph convolutional neural network This is used to simulate the impact of local heterogeneous node states on hyperedges.

[0031] In the formula: This is a hyperdiagonal matrix used for normalization; It is an incidence matrix. For the process l After processing by a hypergraph convolutional neural network, the output node feature matrix, if heterogeneous nodes v If it belongs to a hyperedge, then the incidence matrix The corresponding element in the array is 1 if it is 1, otherwise it is 0; For the process l The output of the hypergraph convolutional neural network is a trainable feature transformation weight matrix.

[0032] Step S2.3.2, Update the superedge to the heterogeneous node: Aggregated hyperedge features The updated node feature matrix is ​​then passed back to all heterogeneous nodes contained in the hyperedge. To simulate the feedback and adjustment of the hyperedge to the state of the internal heterogeneous nodes: In the formula: It is a diagonal matrix of node degrees; This is the hyperedge weight matrix, used to characterize the differences in importance of different types of hyperedges in feature propagation; It is a non-linear activation function.

[0033] Through multiple layers of convolution stacking, the final output is a high-order topological feature embedding matrix of nodes. In the formula: Indicates the process The final output of the hypergraph convolutional neural network is the node feature matrix.

[0034] Optionally, step S3 includes: Step S3.1: Construct a Markov decision process for the interaction between the intelligent agent and the unified hypergraph model of the power transportation coupled network.

[0035] Step S3.2: The agent uses a trained deep Q-network model to solve the Markov decision process and outputs the set of sentinel nodes that minimizes the normalized mean square error function.

[0036] Optionally, step S3.1 includes: Step S3.1.1, State Space: Embedded by the higher-order topological features of the nodes and current decision-making steps t The set of sentinel nodes below mask vector Jointly composed of state vectors The set of sentinel nodes mask vector It is a one-dimensional Boolean vector. Let be the total number of heterogeneous nodes in the entire network, where is the th in the mask vector. The nth element being 1 indicates that the nth element is 1. Each heterogeneous node belongs to the current decision step. t The set of sentinel nodes below The first in the mask vector If any element is 0, it indicates that the heterogeneous node is a non-sentinel node, and the decision is made through the current step. t The mask vector of the sentinel node set below The intelligent agent can perceive the deployment location of heterogeneous nodes in the network topology and the current observation scheme in real time.

[0037] Step S3.1.2, Motion Space: Define the action as node replacement: in the decision step Based on the current strategy, the agent decides to select from the current set of sentinel nodes. Remove the sentinel node with the lowest contribution. And select the heterogeneous node with the highest potential value from the remaining non-sentinel nodes. Join to form a new set of sentinel nodes. .

[0038] Step S3.1.3, Reward Function: Define rewards To represent the negative increment of the approximation error, a regularization term based on the Fisher information matrix is ​​introduced: In the formula: the first term This indicates encouragement to reduce error. Indicates the previous decision step Normalized mean square error function for the sentinel node set; Indicates the current decision step t The normalized mean square error function of the sentinel node set; the second term is the regularization constraint. For coefficients, These are the diagonal elements of the Fisher information matrix, used to penalize parameters important to historical tasks. Dramatic changes This represents the fixed optimal first position after the historical task training is completed. Each network weight parameter, This indicates the th in the currently training deep Q-network. Each network weight parameter.

[0039] Diagonal elements of the Fisher information matrix This reflects the first in the deep Q-network model Weight parameters The importance of historical missions In the formula, For the agent in state The policy probability of the action defined in the action space is selected below. This serves as an experience replay pool for storing the data distribution of historical tasks. This represents the average of the sampling results over the historical data distribution. This represents the probability distribution of actions in a given state; Represents action variables; This represents the set of weight parameters for a deep Q-network model.

[0040] Step S3.2 includes: Step S3.2.1: The deep Q-network architecture of the deep Q-network model includes: Input layer, used to input state vectors .

[0041] The feature fusion layer combines the state vector The embedded features and mask features are concatenated or weighted through an attention mechanism to extract the comprehensive feature representation of heterogeneous nodes under the current topology coverage, and the fused state feature representation is output.

[0042] The hidden layer, connected to the feature fusion layer, is used to perform nonlinear transformation on the fused state feature representation, extract deep topological and state association patterns, and output the hidden feature representation.

[0043] The output layer is used to perform linear mapping on the latent feature representation, output the predicted Q value of the node replacement action, and select the optimal action based on the predicted Q value to update the current set of sentinel nodes until the set of sentinel nodes with the smallest approximate error is found and output.

[0044] Step S3.2.2: Construct the loss function for the deep Q-network architecture. : Introduce a target network with the same structure but lagging weight parameter updates into the loss function calculation and weight parameter update process of the deep Q-network architecture. The network weight parameters of the target network are denoted as... .

[0045] Define the loss function for a deep Q-network architecture. Predicted Q-values ​​for the current deep Q-network architecture With the target Q value Mean square error between: In the formula, Indicates from the experience replay pool A transition sample quadruple randomly sampled from the middle The calculated average loss value, Represents the state vector at the current moment; Indicates an action; Indicates a reward; This represents the state vector that the user transitions to at the next moment after performing an action. The weight parameters represent the deep Q-network architecture; the target Q-value. Calculations based on Bellman equations: , It is the negative increment of the approximation error. As a discount factor, This refers to finding the optimal action that yields the greatest expected return. This represents the predicted action value based on the action performed in the state vector at the next time step. This represents the candidate actions that may be executed under the state vector at the next time step.

[0046] Transfer Sample Experience Replay Pool This is used to store the transition sample quadruples generated by the agent's interaction with the current hypergraph environment, for training the deep Q-network architecture; the transition sample quadruples are denoted as... In the formula, Indicates the current decision step The state vector; This represents the specific actions of the intelligent agent; This represents the immediate reward after the agent performs an action; Indicates the next decision step The state vector.

[0047] Step S3.2.3: Train the constructed deep Q-network architecture: First, from the experience replay pool Randomly select a small batch of transfer sample quadruplets Then the current state vector in the transition sample quadruple is... Input a deep Q-network architecture and compute the actions to be performed. The corresponding predicted Q value, and utilizing the instant reward. and the state vector of the next decision step The target Q-value obtained based on the Bellman equation and the target network is calculated, and then the mean squared error loss between the two is obtained; then, the backpropagation algorithm is used to obtain the loss function. Relative to the current network weight parameters gradient Finally, an adaptive moment estimation optimizer, such as Adam, is used to adjust the network weight parameters according to the gradient direction. Iterative updates are performed to gradually optimize the action value function.

[0048] An adaptive moment estimation optimizer is used to optimize the current network weight parameters. Update: In the formula, For learning rate, The loss function is expressed with respect to the network weight parameters. gradient operator, This represents the loss function of a deep Q-network architecture.

[0049] When the loss function After convergence or reaching the preset number of training iterations, fix the network weight parameters. The training of the deep Q-network architecture was completed.

[0050] On the other hand, the present invention also provides a sentinel node identification system for a power transportation coupled network, comprising: The data value assessment module is used to quantify and filter the value of collected multi-source heterogeneous data on power and transportation based on optimized gradient values, and output high-value data.

[0051] The Unified Hypergraph Modeling Module is used to construct a unified hypergraph model of the power and transportation coupled network and extract high-order topological features of nodes into the embedding matrix.

[0052] The sentinel node identification module is used to model the sentinel node identification problem as a Markov decision process. It uses a trained deep Q-network model to solve the Markov decision process and outputs the set of sentinel nodes that minimizes the approximation error.

[0053] Optionally, the data value assessment module specifically adopts a strategy combining Monte Carlo approximation and single-sample gradient descent to calculate the optimized gradient value of the data points; based on the comparison between the optimized gradient value and the set noise filtering threshold and redundancy filtering threshold, the multi-source heterogeneous data of power and transportation is screened to obtain high-value data.

[0054] In another aspect, the present invention also provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, it implements the method described above.

[0055] Compared with the prior art, the present invention has at least one of the following technical effects: This invention proposes a data value assessment model based on optimized gradient values. By simulating the model parameter update process, the marginal contribution of data points is calculated, reducing the time complexity of data value assessment from exponential to linear. This model can quickly screen high-value data in the high-noise environment at the beginning of a disaster, providing a lightweight and high-quality input foundation for subsequent modeling.

[0056] This invention constructs a unified hypergraph model for the power-transportation coupled network, and uses a hypergraph convolutional neural network to learn features of heterogeneous nodes. It effectively represents the "one-to-many" high-order coupling relationship that is common in the power-transportation-charging network, and breaks through the limitation of traditional graph models that are difficult to capture cross-network interaction features, thus realizing in-depth mining of the topology of the coupled system.

[0057] This invention proposes a sentinel node identification strategy based on reinforcement learning, which models node selection as a Markov decision process. Through continuous interaction between the agent and the environment, the optimal combination of observations is dynamically searched. This can accurately invert the macroscopic dynamics of the entire network with very few local observation nodes, significantly reducing monitoring costs and communication pressure in disaster scenarios.

[0058] This invention constructs a sentinel node identification system for power-transportation coupled networks in disaster scenarios. By integrating three major functional modules—data value assessment, unified hypergraph modeling, and sentinel node identification—it forms a closed-loop process from multi-source heterogeneous data cleaning and complex topology reconstruction to intelligent decision-making at key nodes. This achieves close collaboration and automated operation of each algorithm component, providing an efficient and practical integrated solution for emergency monitoring of power-transportation coupled systems in disaster environments. Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating a method for identifying sentinel nodes in a power-transportation coupled network according to an embodiment of the present invention. Figure 2 This is a framework diagram of a sentinel node identification system for a power-transportation coupled network provided in an embodiment of the present invention; Figure 3 This is a framework diagram of the deep Q-network architecture of a deep Q-network model provided in an embodiment of the present invention. Detailed Implementation

[0060] The following detailed description, in conjunction with the accompanying drawings and specific embodiments, provides a further detailed explanation of the sentinel node identification method and system for a power-transportation coupled network proposed in this invention. The advantages and features of this invention will become clearer from the following description. It should be noted that the accompanying drawings are in a very simplified form and use non-precise proportions, used only to facilitate and clearly illustrate the embodiments of this invention. Please refer to the accompanying drawings to make the objectives, features, and advantages of this invention more apparent and understandable. It should be understood that the structures, proportions, sizes, etc., depicted in the accompanying drawings are only for illustrative purposes to aid those skilled in the art and are not intended to limit the implementation conditions of this invention. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to the size, without affecting the effects and objectives achieved by this invention, should still fall within the scope of the technical content disclosed in this invention.

[0061] like Figure 1 As shown, this embodiment provides a method for identifying sentinel nodes in a power-transportation coupled network, including: Step S1: Based on the optimized gradient value, the collected multi-source heterogeneous data of power and transportation are valued and filtered to output high-value data.

[0062] In this embodiment, the multi-source heterogeneous data of power and transportation includes power grid flow data, transportation network traffic data, and charging network operation data collected in the early stages of a disaster.

[0063] By quantifying the value of multi-source heterogeneous data in power and transportation, we can solve the problems of high noise, labeling errors, and feature redundancy in power grid flow data, transportation network traffic data, and charging network operation data collected in the early stages of a disaster.

[0064] Step S2: Construct a unified hypergraph model for the power-transportation coupled network, which defines power nodes, transportation nodes and charging station nodes as a heterogeneous set of nodes. Based on physical connections, traffic similarity and cross-network supply and demand relationships, construct multiple types of hyperedges, and use hypergraph convolutional neural networks to extract high-order topological features of nodes and embed them into the matrix.

[0065] Step S3: Based on the reinforcement learning algorithm, the sentinel node identification problem is modeled as a Markov decision process, and the set of sentinel nodes that minimizes the normalized mean square error function is dynamically searched and output.

[0066] This embodiment can accurately invert the overall macroscopic state of the network by utilizing local key nodes in a disaster environment with high data noise and limited computing resources.

[0067] In this embodiment, step S1 includes: step S1.1, using a strategy that combines Monte Carlo approximation with single-sample gradient descent to calculate the optimized gradient value of each data point in the multi-source heterogeneous data of power and transportation, so as to realize the calculation of data value quantification.

[0068] In this embodiment, step S1.1 includes: defining data points for multi-source heterogeneous data of power and transportation. The value of this method is discussed. To overcome the shortcomings of traditional Shapley values, where computational complexity increases exponentially with data volume and fails to meet the real-time requirements of disaster emergency response, this embodiment employs a strategy combining Monte Carlo approximation and single-sample gradient descent. Specifically, data points... The value of a data point is defined as its performance in multiple randomized training iterations. After adding to the training set, the model utility function Average marginal contribution, data points After adding to the training set, the model utility function The average marginal contribution is also a data point Optimization gradient value : The following formula is used to calculate the i-th data point in the multi-source heterogeneous data of power and transportation. Optimization gradient value : (1) In the formula, M The cumulative number of Monte Carlo samplings is used to ensure the statistical significance of the value assessment; For the first m In the random permutation, located at the data point The previous subset of data.

[0069] Calculate the model utility function in formula (1) above. In this embodiment, the time-consuming full model retraining for each data subset is avoided. Instead, a baseline neural network model is first constructed, with its network weight parameters set as follows: Define the loss function To predict the mean squared error, a single-step gradient update strategy is then employed to quickly estimate the data impact. First, based on the current subset of data... B Calculate the loss function of the benchmark neural network model Regarding the current network weight parameters gradient , Represents the loss function Regarding network weight parameters The gradient operator.

[0070] Secondly, the optimal learning rate is determined through linear search or a pre-set strategy. This maximizes the impact of a single network weight parameter update on the performance of the benchmark neural network model. (2) In the formula, This represents the network weight parameters of the updated baseline neural network model.

[0071] Finally, in the independent validation set Calculate the updated network weight parameters The model performance metrics of the baseline neural network model (such as prediction accuracy or the reciprocal of load forecast accuracy) are used as a subset of the data. Model utility function utility value : (3) This step reduces the time complexity of data value assessment from exponential to linear, adapting to the rapid response needs in disaster scenarios.

[0072] Step S1.2: Compare the optimized gradient value with the set noise filtering threshold and redundancy filtering threshold respectively to filter the multi-source heterogeneous data of power and transportation and obtain high-value data.

[0073] In this embodiment, step S1.2 includes: setting a noise filtering threshold. and redundant filtering thresholds : like Determine the data point Data with incorrect labels or maliciously injected data will be removed.

[0074] like Determine the data point Redundant features that have no significant impact on model performance are removed.

[0075] Step S1.3: Using the high-value data retained after filtering in step S1.2, construct a power grid topology map, a transportation network map, and a charging facility distribution map, respectively, as a lightweight input basis for subsequent hypergraph modeling. The specific construction process includes: (1) Constructing the power grid topology diagram Extract power flow data from high-value data, analyze the node attributes and line connection relationships in the power flow data, and construct a power grid topology map. Specifically, the electrical attributes of generator nodes, substation nodes, and ordinary load nodes are analyzed and obtained. Using the actual physical feeder and transformer connections of the distribution network as topology edges, a power grid topology diagram reflecting the electrical constraints of the system is constructed. ; (2) Constructing a transportation network map Extract traffic network flow data from high-value data, analyze the road segment connectivity and geographical location information in the traffic network flow data, and construct a traffic network map. Specifically, road intersections and road segment endpoints are extracted as traffic nodes, and the actual physical road segments connecting each traffic node are extracted as topological edges. Combined with geographical location information such as road segment level and length, a traffic network map reflecting the spatial distribution of traffic flow is constructed. ; (3) Construct a distribution map of charging facilities Extract charging network operation data from high-value data, analyze the spatial coordinates and equipment configuration information of charging stations in the charging network operation data, and construct a charging facility distribution map. Specifically, the static configuration information such as the geospatial coordinates and the number of charging piles of each charging station is analyzed. A distribution map is constructed with charging stations as nodes, and the spatial proximity correspondence between them and the surrounding power nodes and transportation nodes is recorded, which provides a foundation for the subsequent construction of cross-network coupling hyperedges.

[0076] In this embodiment, step S2 includes: Step S2.1: Using the aforementioned power grid topology diagram The aforementioned transportation network diagram and the distribution map of the charging facilities Based on this, a unified hypergraph model of the power-transportation coupled network is constructed, that is, a unified hypergraph model describing the multi-dimensional power-transportation-charging system is constructed.

[0077] The unified hypergraph model of the power-transportation coupled network is represented by the following formula: (4) In the formula, Represents a set of heterogeneous nodes. Denotes the set of superedges. H This represents the correlation matrix of the unified hypergraph model of the power-transportation coupled network. X Represents the node feature matrix; where the set of heterogeneous nodes is... It includes the following three types of heterogeneous nodes: power node set Charging station node set and traffic node set .

[0078] Among them, the set of power nodes Includes: generator nodes, substation nodes, and general load nodes. Charging station node set. As a coupling hub. A set of traffic nodes. This includes: road intersections and road segment endpoints.

[0079] Hyperedge set The following three types of hyperedges can represent higher-order one-to-many relationships: Power topology superedge Based on the physical connection relationship of the distribution network, all power nodes connected to the same feeder, the same transformer or the same bus are aggregated into a superedge, which represents the physical constraint propagation relationship of electrical quantities.

[0080] Traffic similarity hyperedge In the formula, n This represents the total number of traffic similarity hyperedges constructed. Indicates the first Traffic similarity superedge.

[0081] In this embodiment, the calculation of any two traffic nodes is based on the spatiotemporal correlation of traffic flow. and Historical flow vector L2 Norm distance : (5) In the formula, Indicates traffic nodes Historical traffic data vector, Indicates traffic nodes Historical traffic data vector; if ( (For the set similarity threshold), then traffic nodes are considered to be... and If there are homogeneous traffic patterns, they are aggregated into a traffic similarity hyperedge.

[0082] Cross-network coupling hyperedge Using charging stations as the core link, a heterogeneous hyperedge is constructed, which includes "charging station nodes - power distribution nodes that supply power to them - traffic segment nodes within their coverage area", to characterize the cross-network interaction mechanism of energy flow (electricity) and traffic flow (vehicles) driven by charging behavior.

[0083] In this embodiment, step S2.1 further includes defining a hypergraph association matrix of a unified hypergraph model, which is used to mathematically represent the attribution relationship between heterogeneous nodes in a set of heterogeneous nodes and hyperedges in a set of hyperedges.

[0084] The correlation matrix of the unified hypergraph model : ,in, Represents the set of real numbers; Represents a set of heterogeneous nodes The total number of nodes in; Represents the set of superedges The total number of superedges in the array.

[0085] (6) In the formula, Representation of the correlation matrix The elements within the matrix, Represents any heterogeneous node, Represents any type of hyperedge (power, transportation, or coupling hyperedge).

[0086] Step S2.2: Construct a node feature matrix that can uniformly describe the physical properties and operating states of heterogeneous nodes. ,in, Represents the node feature matrix, Represents the set of real numbers. The total number of all heterogeneous nodes. For node feature dimensions.

[0087] Step S2.2 includes: constructing normalized feature vectors for the three types of heterogeneous node sets in the heterogeneous node set, and concatenating them according to the node index to form a global feature matrix, that is, forming the node feature matrix, in order to solve the problem of inconsistent dimensions of physical quantities in multiple networks.

[0088] First, feature vectors are constructed for the three types of nodes: power, transportation, and charging stations. Then, they are stacked vertically according to the index order of the nodes in the hypergraph to form a unified global node feature matrix for subsequent calculations.

[0089] Step S2.2.1: Construct the feature vector of the power node: For any power node (Including substations, distribution transformers, and load nodes), define its feature vector. : (7) In the formula, The node load factor represents the proportion of the current power load to the rated capacity, reflecting the operating margin of the power node. The upper limit of power supply capacity characterizes the physical power supply or transmission limit of a power node and determines its support potential in disaster recovery. Voltage amplitude represents the voltage stability of power nodes and is a key indicator for determining whether a power system exceeds its limits.

[0090] Step S2.2.2: Construct the feature vector of the charging station node: For any charging station node As a coupling hub connecting power and transportation, its feature vector is defined. : (8) In the formula, The number of charging piles represents the static service capacity base of the charging station node; This provides real-time facility utilization, reflecting the current level of activity and queuing congestion risk at charging station nodes. The real-time total charging load not only represents its power extraction demand from the power network, but also indirectly reflects the intensity of traffic flow convergence at this charging station node.

[0091] Step S2.2.3: Construct traffic node feature vectors: For any traffic node Define its eigenvectors : (9) In the formula, The road grade coefficient quantifies the road type and determines the upper limit of traffic capacity and the benchmark for traffic speed at traffic nodes. Cross-sectional traffic flow reflects the number of vehicles passing through the traffic node per unit time and characterizes the dynamic distribution of traffic flow. The Road Congestion Index is calculated based on the ratio of average vehicle speed to free-flow vehicle speed, representing traffic efficiency and vehicle congestion.

[0092] Step S2.2.4 Stack all the feature vectors of power nodes, charging station nodes, and traffic nodes according to the global index order of the corresponding heterogeneous nodes in the unified hypergraph of the power-transportation coupling network to form the initial node feature matrix. X The initial node feature matrix can be specifically represented by the following formula. ， The node feature matrix X It not only preserves the unique properties of each subnetwork, but also provides standardized input data for subsequent Hypergraph Convolutional Neural Networks (HGCN) to mine higher-order cross-network associations through a unified vector space mapping.

[0093] Step S2.3: Use a hypergraph convolutional neural network to simulate the propagation process of disaster impact in the power and transportation coupled network. Through the alternating propagation mechanism of heterogeneous node to hyperedge aggregation and hyperedge to heterogeneous node update, feature learning is performed to extract the high-order topological features of heterogeneous nodes and output the node high-order topological feature embedding matrix.

[0094] In this embodiment, step S2.3 includes: processing the node feature matrix. X and the correlation matrix H It is used as input to a hypergraph convolutional neural network for feature updates; Step S2.3.1, Heterogeneous node to hyperedge aggregation: To achieve normalization of feature propagation, heterogeneous node degree matrices are defined respectively. and hypermarginality matrix Heterogeneous node degree matrix diagonal elements and hypermarginality matrix diagonal elements The calculations are as follows: (10) In the formula, For super-edge The weight can be set to 1 by default, or it can be adaptively assigned according to the cross-network coupling strength.

[0095] The information of each heterogeneous node is determined based on the correlation matrix. They converge at their superedge, forming a process l The hyperedge features output after processing by a layer hypergraph convolutional neural network To simulate the impact of local heterogeneous node states on hyperedges (global regions): (11) In the formula: It is a hyperdiagonal matrix whose elements are equal to the sum of the elements in the corresponding columns of the incidence matrix, and is used for normalization; It is an incidence matrix. For the process l The node feature matrix output after processing by a hypergraph convolutional neural network, if heterogeneous nodes Belongs to superedge Then the corresponding elements of the correlation matrix ,otherwise ; For the process l The output of the hypergraph convolutional neural network is a trainable feature transformation weight matrix.

[0096] Step S2.3.2, Update the superedge to the heterogeneous node: Aggregated hyperedge features The updated node feature matrix is ​​then passed back to all heterogeneous nodes contained in the hyperedge. This is to simulate the feedback and regulation of the hyperedge (region state) on the state of the internal heterogeneous nodes: (12) In the formula: It is a diagonal matrix of node degrees; This is the hyperedge weight matrix, used to characterize the differences in importance of different types of hyperedges (power, transportation, coupling) in feature propagation; It is a non-linear activation function.

[0097] Through multiple layers of convolution stacking, the final output is a high-order topological feature embedding matrix of nodes. In the formula: Indicates the process The final output of a hypergraph convolutional neural network is the node feature matrix. That is, the feature matrix of the nodes in the output layer (i.e., the 1st layer). The final heterogeneous node feature embedding matrix is ​​generated by (layer). The final output node high-order topology feature embedding matrix deeply integrates the physical constraints of the power grid, the spatiotemporal patterns of the transportation network, and the coupling relationship between the two.

[0098] Step S3 includes Step S3.1: Construct a Markov decision process for the interaction between the intelligent agent and the unified hypergraph model of the power transportation coupled network.

[0099] Understandably, in the Markov decision-making process, the unified hypergraph model of the power-transportation coupled network is modeled as the external environment for agent interaction: agents observe the state of the unified hypergraph model of the power-transportation coupled network and perform actions on the unified hypergraph model of the power-transportation coupled network, while the unified hypergraph model of the power-transportation coupled network feeds back reward signals.

[0100] Step S3.2: The agent uses a trained deep Q-network model to solve the Markov decision process and outputs the set of sentinel nodes that minimizes the approximation error.

[0101] Specifically, before executing step S3.1, the macroscopic state and approximate error objective function of the power-transportation coupled network are defined: Define the true macroscopic state (true average state) of the entire network under the equilibrium state of disaster evolution. This is the true macroscopic state. It reflects the overall operational level of the power-transportation coupled network. Its calculation formula is the average of the state values ​​of all heterogeneous nodes: (13) In the formula: This represents the total number of heterogeneous nodes. Indicates the first i The state values ​​of heterogeneous nodes.

[0102] Definition by n A set of sentinel nodes The observed approximate state, i.e., the sentinel observation state. : (14) In the formula: Indicates the first j The state values ​​observed by each sentinel node.

[0103] In this embodiment, to ensure that the sentinel node maintains the accuracy of its observations across different disaster severity levels and different stages of dynamic evolution, a normalized mean square error function is defined. This is the core objective of optimizing the set of sentinel nodes.

[0104] The normalized mean square error function Measured in One simulated disaster scenario (corresponding to different control parameters) D The cumulative deviation between observed and true values ​​of sentinel nodes under conditions such as fault propagation rate or coupling strength: (15) In the formula: Indicates the first A simulated disaster scenario; and Each of the following is the first The actual average state and sentinel observation state under a simulated disaster scenario.

[0105] In this embodiment, step S3.1 includes: Step S3.1.1, State Space : Embedded by the higher-order topological features of the nodes and current decision-making steps t The set of sentinel nodes below mask vector Jointly composed of state vectors The set of sentinel nodes mask vector It is a one-dimensional Boolean vector. Let be the total number of heterogeneous nodes in the entire network, where is the th in the mask vector. The element being 1 indicates that the th element is 1. Each heterogeneous node belongs to the current decision step. t The set of sentinel nodes below The first in the mask vector If any element is 0, it indicates that the heterogeneous node is a non-sentinel node, based on the current decision step. t The mask vector of the sentinel node set below The agent can perceive the deployment location of heterogeneous nodes in the network topology and the current observation plan in real time. This allows the agent to not only know the current observation plan, but also to utilize the importance and location characteristics of the embedded sensing nodes in the network topology.

[0106] Step S3.1.2, Motion Space : Define the action as node replacement (Swap): in the decision step Based on the current strategy, the agent decides to select from the current set of sentinel nodes. Remove the sentinel node with the lowest contribution. And select the heterogeneous node with the highest potential value from the remaining non-sentinel nodes. Join to form a new set of sentinel nodes. .

[0107] It is understandable that the state vector mentioned above... mask vector in It is precisely the set of sentinel nodes The binary mask representation. The agent observes the state vector. This will allow us to determine which nodes currently belong to the sentinel node set. .

[0108] Step S3.1.3, Reward Function : To guide the agent to quickly find the set of sentinel nodes with the smallest approximate error, a reward is defined. This represents the negative increment of the approximation error; Meanwhile, to prevent the agent from forgetting old patterns when adapting to new disaster modes (a continuous learning problem), a regularization term based on the Fisher information matrix is ​​introduced: (16) In the formula: the first term This indicates encouragement to reduce error. Indicates the previous decision step Sentinel node set The corresponding normalized mean square error function; Indicates the current decision step t Sentinel node set The corresponding normalized mean square error function; the second term is the regularization constraint. For coefficients, These are the diagonal elements of the Fisher information matrix, used to penalize parameters important to historical tasks. Dramatic changes This represents the fixed optimal first position after the historical task training is completed. Each network weight parameter, This indicates the th in the currently trained Deep Q-Network (DQN) model. Each network weight parameter.

[0109] Diagonal elements of the Fisher information matrix Reflects the i-th network weight parameter The importance of historical tasks is calculated based on the gradient of the log-likelihood function: (17) In the formula: For the agent in state (It is understandable that the state vector) It refers to the specific state, here. Selecting an action under a general symbol for a state The strategy probability, This serves as an experience replay pool for storing the data distribution of historical tasks. Expressing expectations, Representation strategy;a Indicates an action; The set of weight parameters for a deep Q-network model. Specifically refers to the first in the set of weight parameters Each scalar weight parameter This is the parameter index, with values ​​ranging from 1 to the total number of network weight parameters.

[0110] In this embodiment, step S3.2 includes: the deep Q-network model uses a deep neural network to fit the action value function, and through experience replay and target network mechanism, efficiently searches for the optimal set of sentinel nodes in the huge combinatorial solution space.

[0111] Step S3.2.1: Construct a deep neural network Used to approximate the action value function, where These are the network weight parameters.

[0112] like Figure 3 As shown, the deep Q-network architecture of the deep Q-network model includes: Input layer, used to input state vectors .

[0113] The feature fusion layer combines the state vector The embedded features and mask features are concatenated or weighted through an attention mechanism to extract the comprehensive feature representation of heterogeneous nodes under the current topology coverage, and the fused state feature vector is output.

[0114] Understandably, topology coverage refers to the range of nodes that the currently selected set of sentinel nodes can radiate or monitor in the unified hypergraph model of the power-transportation coupled network through physical connections (such as hyperedges). Specifically, due to the high-order association characteristics of the unified hypergraph model of the power-transportation coupled network, a small number of sentinel nodes can "cover" and perceive the state of their neighboring nodes by sharing hyperedges. The purpose of the feature fusion layer is precisely to capture this network-wide perception capability determined by the current sentinel layout.

[0115] The input layer is used to receive and parse the state vector. Extract the hypergraph embedding matrix. With the current sentinel node mask vector , which serves as the raw input data for subsequent neural networks; The feature fusion layer, as the first core computational layer of the network, is connected to the input layer. This layer embeds the flattened hypergraph into a matrix. With mask vector The data is concatenated, followed by linear mapping and activation, to reduce the dimensionality of the state information and extract deep features, outputting a comprehensive feature representation. Taking the splicing and fusion method as an example, its calculation process is represented by the following formula: (18) In the formula, This represents a vector concatenation operation. For flattening operation, and For the trainable weights and biases of the input layer, This is the activation function.

[0116] The hidden layer, connected to the feature fusion layer, is used to perform nonlinear transformations on the fused state feature vectors, extract deep topological and state association patterns, and output latent feature representations.

[0117] The hidden layer includes a fully connected layer, a normalized layer, and a non-linear activation function (e.g., a ReLU function). The fully connected layer is used to perform a linear weighted mapping on the input features (in this embodiment, the input features are the fused state feature vectors) to project the high-dimensional feature space onto the action value space.

[0118] Understandably, fully connected layers can be placed between sub-layers within a hidden layer and at the end of the hidden layer.

[0119] The normalization layer is placed between the fully connected layer and the non-linear activation function. The normalization layer standardizes the output data of the fully connected layer (making its mean 0 and variance 1) to prevent oscillations or divergence in the predicted Q-values ​​of potential "node replacement" actions due to internal data distribution shifts, thereby improving the stability of deep Q-network model training.

[0120] Non-linear activation function (ReLU function).

[0121] (19) In the formula: x This represents the input value of the activation function.

[0122] This is used to endow deep Q-network architectures with the ability to fit complex nonlinear topological relationships.

[0123] The output layer is used to perform linear mapping on the latent feature representation, output the predicted Q value of the node replacement action, and select the optimal action based on the predicted Q value to update the current set of sentinel nodes until the set of sentinel nodes with the smallest approximate error is found and output.

[0124] The dimension of the output layer corresponds to the size of the action space. Each element in the network output vector corresponds to a specific "node replacement" action. The predicted Q value is the expected reduction in cumulative approximation error after performing the replacement action.

[0125] Step S3.2.2: Construct the loss function for the deep Q-network architecture. : To address the non-stationarity of the target network during training, a target network with the same structure but lagging parameter updates is introduced. The network weights of the target network are denoted as follows: Specifically, this involves introducing a target network with the same structure but lagging weight parameter updates into the loss function calculation and network weight parameter update process of a deep Q-network model or architecture. In the initial stage of the algorithm, an additional target network with the exact same structure as the current deep Q-network is instantiated. This target Q-value is then used to calculate the loss function. y When using the target network for forward propagation, its network weight parameters are maintained. The network weight parameters of the current depth Q-network structure are fixed, and are only changed after a certain number of training steps. Copy it to it.

[0126] Define the loss function for a deep Q-network architecture. Predicted Q-values ​​for the current deep Q-network architecture With the target Q value Mean squared error (MSE) between: (20) In the formula, Indicates from the experience replay pool A transition sample quadruple randomly sampled from the middle The calculated average loss value, Represents the state vector at the current moment; This indicates the action taken by the agent in the current state; This represents the immediate reward value obtained after performing an action; This indicates the state transitioned to at the next moment after the action is performed; This represents the network weight parameters of a deep Q-network architecture.

[0127] Target Q value Calculations based on the Bellman Equation: (twenty one) In the formula: It is the negative increment of the approximation error. As a discount factor, This indicates an operation that takes the maximum value of all possible actions at the next moment; This represents the value estimate of the state-action pair at the next time step, calculated by the target network. It indicates the action that may be taken in the next moment.

[0128] Transfer Sample Experience Replay Pool Specifically used to store the quadruple set of transfer samples generated by the interaction between the agent and the hypergraph environment. , to be used for training the deep Q-network architecture; where, Indicates the current time step The state vector; Indicates time step The action chosen by the agent; Indicates time step The immediate reward value provided by the environment after an action is performed; Indicates the next time step The state vector.

[0129] Step S3.2.3: Train the constructed deep Q-network architecture: From the transferred sample experience replay pool Transfer sample quadruples from mini-batch random sampling Training is then performed. This mechanism breaks down the correlation between time series samples, improving data utilization.

[0130] The loss function is calculated based on the sampled transfer samples. Regarding the current network weight parameters gradient ; The Adaptive Moment Estimation Optimizer (Adam) is used to optimize the current network weight parameters. Update: (twenty two) In the formula, For learning rate, The loss function is expressed with respect to the network weight parameters. gradient operator, This represents the loss function of a deep Q-network architecture.

[0131] Strategy iteration and output: During the training phase, the following methods were adopted: - Greedy strategy ( -greedy) to make action selections to balance exploration and exploitation.

[0132] This represents the probability of exploration. Specifically, it indicates the likelihood that the agent / deep Q-network architecture will randomly try new actions rather than choosing what it currently considers the best action. As the number of training epochs increases, Linear decay allows the agent / deep Q-network architecture to gradually shift from random exploration to optimal policy execution. In this project, A dynamic linear decay strategy is adopted, with the specific value typically set between 1.0 and 0.1. Setting a high initial value of 1.0 ensures sufficient exploration in complex topologies under disaster scenarios, while linear decay during training ensures that the agent / deep Q-network architecture eventually converges to the optimal decision logic.

[0133] When the loss function After convergence or reaching the preset number of training iterations, fix the network weight parameters. The deep Q-network architecture is trained, and its output predicts the action sequence with the largest Q-value, corresponding to the optimal sentinel node adjustment strategy. The final output is the normalized mean squared error function that maximizes the network's overall performance. Minimal set of sentinel nodes .

[0134] like Figure 2 As shown, this embodiment also provides a sentinel node identification system for a power-transportation coupled network, used to perform the method described above, including: The data value assessment module 101 is used to quantify and filter the value of the collected multi-source heterogeneous data of power and transportation based on the optimized gradient value, and output high-value data. The unified hypergraph modeling module 102 is used to construct a unified hypergraph model of the power-transportation coupled network and extract the high-order topological features of nodes into the embedding matrix. The sentinel node identification module 103 is used to model the sentinel node identification problem as a Markov decision process, and to solve the Markov decision process using a trained deep Q-network model, outputting the set of sentinel nodes that minimizes the approximation error.

[0135] In this embodiment, the data value assessment module 101 specifically adopts a strategy combining Monte Carlo approximation and single-sample gradient descent to calculate the optimized gradient value of the data points; based on the comparison between the optimized gradient value and the set noise filtering threshold and redundancy filtering threshold, the multi-source heterogeneous data of power and transportation is screened to obtain high-value data.

[0136] The data value assessment module 101 is primarily responsible for the preprocessing and value quantification of multi-source heterogeneous data in the early stages of a disaster. It receives collected power grid flow data, transportation network traffic data, and charging network operation data. Using a data value assessment model based on optimized gradient values, it calculates the marginal contribution of data points through Monte Carlo approximation and a single-step gradient update strategy. This module sorts the data in descending order based on the calculated value and, combined with preset noise and redundancy filtering thresholds, removes high-noise and low-value data, providing a lightweight, high-quality data foundation for subsequent modeling.

[0137] The unified hypergraph modeling module 102 is primarily responsible for constructing a topological model capable of representing complex coupling relationships. It defines power nodes, transportation nodes, and charging station nodes as a heterogeneous node set and constructs three types of hyperedges that incorporate power topology, transportation similarity, and cross-network coupling relationships, forming a unified hypergraph model. Subsequently, the unified hypergraph modeling module 102 utilizes a hypergraph convolutional neural network, through an alternating propagation mechanism of node-to-hyperedge aggregation and hyperedge-to-node updates, to extract high-order topological feature embeddings of nodes that integrate physical constraints and spatiotemporal patterns.

[0138] The sentinel node identification module 103 is primarily responsible for intelligently determining the optimal set of observation nodes. It models the sentinel node selection problem as a Markov decision process, constructing a reward function with the objective of minimizing the overall network state inversion error. This module incorporates an intelligent agent, employing a deep Q-network algorithm to interact with the hypergraph environment. Based on the current network state and mask vector, it dynamically executes "node replacement" actions, ultimately outputting a set of sentinel nodes capable of accurately inverting the overall network dynamics through local observations.

[0139] On the other hand, the present invention also provides an electronic device including a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, it implements the method described above.

[0140] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus 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 apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0141] It should be noted that the apparatus and methods disclosed in the embodiments herein can also be implemented in other ways. The apparatus embodiments described above are merely illustrative; for example, the flowcharts and block diagrams in the accompanying drawings show the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments herein. In this regard, each block in a flowchart or block diagram may represent a module, program, or part of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system to perform the specified function or action, or can be implemented using a combination of dedicated hardware and computer instructions.

[0142] In addition, the functional modules in the various embodiments of this article can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0143] 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 identifying sentinel nodes in a power-transportation coupled network, characterized in that, include: Step S1: Based on the optimized gradient value, the collected multi-source heterogeneous data of power and transportation are valued and filtered to output high-value data; Step S2: Construct a unified hypergraph model for the power-transportation coupled network, which defines power nodes, transportation nodes and charging station nodes as a set of heterogeneous nodes. Based on physical connections, traffic similarity and cross-network supply and demand relationships, construct multiple types of hyperedges, and use hypergraph convolutional neural networks to extract high-order topological features of nodes and embed them into the matrix. Step S3: Based on the reinforcement learning algorithm, the sentinel node identification problem is modeled as a Markov decision process, and the set of sentinel nodes that minimizes the normalized mean square error function is dynamically searched and output.

2. The method for identifying sentry nodes in a power-transportation coupled network as described in claim 1, characterized in that, Step S1 includes: Step S1.1: Using a strategy combining Monte Carlo approximation and single-sample gradient descent, calculate the optimized gradient value of each data point in the multi-source heterogeneous data of power and transportation. Step S1.2: Compare the optimized gradient value with the set noise filtering threshold and redundancy filtering threshold respectively to filter the multi-source heterogeneous data of power and transportation and obtain high-value data. Step S1.3: Using the high-value data, construct power grid topology maps respectively. Transportation network map Map showing the distribution of charging facilities ; The multi-source heterogeneous data on power and transportation includes power grid flow data, transportation network traffic data, and charging network operation data in the early stages of the disaster.

3. The method for identifying sentry nodes in a power-transportation coupled network as described in claim 2, characterized in that, Step S1.1 includes: The i-th data point in the multi-source heterogeneous data of power and transportation is calculated using the following formula. Optimization gradient value , In the formula, M The cumulative number of Monte Carlo samplings is used to ensure the statistical significance of the value assessment; For the first m In the random permutation, located at the data point Previous data subset; For the model utility function; Step S1.2 includes: setting a noise filtering threshold. and redundant filtering thresholds : like Determine the data point Data with incorrect labels or maliciously injected data will be removed. like Determine the data point Redundant features that have no significant impact on model performance are removed. Step S1.3 includes: Using the high-value data retained after filtering, power grid flow data, traffic network flow data, and charging network operation data are extracted respectively; the node attributes and line connection relationships in the power grid flow data are analyzed to construct a power grid topology map. ; Analyze the road segment connectivity and geographical location information in the traffic network traffic data to construct a traffic network map. ; Analyze the spatial coordinates and equipment configuration information of charging stations in the charging network operation data to construct a charging facility distribution map. .

4. The method for identifying sentry nodes in a power-transportation coupled network as described in claim 3, characterized in that, Step S2 includes: Step S2.1: Using the aforementioned power grid topology diagram The aforementioned transportation network diagram and the distribution map of the charging facilities Based on this, a unified hypergraph model of the power-transportation coupled network is constructed. In the formula, V Represents a set of heterogeneous nodes. E Denotes the set of superedges. H This represents the correlation matrix of the unified hypergraph model of the power-transportation coupled network. X Represents the node feature matrix; Among them, the set of heterogeneous nodes Includes: power node set Charging station node set and traffic node set ; Power node set This includes: generator nodes, substation nodes, and ordinary load nodes; Charging station node set As a coupling hub; Traffic node set This includes: road intersections and road segment endpoints; Hyperedge set The following three types of hyperedges can represent higher-order one-to-many relationships: Power topology superedge Traffic similarity hyperedge and cross-network coupling hyperedge ; Traffic similarity hyperedge In n This represents the total number of traffic similarity hyperedges constructed. Indicates the first Traffic similarity hyperedge; The correlation matrix of the unified hypergraph model of the power-transportation coupled network This is used to mathematically represent the relationship between heterogeneous nodes in a set of heterogeneous nodes and hyperedges in a set of hyperedges, where... Represents the set of real numbers; Represents a set of heterogeneous nodes The total number of nodes in; Represents the set of superedges The total number of superedges in; Step S2.2: Construct a node feature matrix of physical properties and operating states of heterogeneous nodes. ,in, Represents the set of real numbers. The total number of all heterogeneous nodes. For node feature dimensions; Step S2.3: Use a hypergraph convolutional neural network to simulate the propagation process of disaster impact in the power and transportation coupled network. Through the alternating propagation mechanism of heterogeneous node to hyperedge aggregation and hyperedge to heterogeneous node update, feature learning is performed to extract the high-order topological features of heterogeneous nodes and output the node high-order topological feature embedding matrix.

5. The method for identifying sentry nodes in a power-transportation coupled network as described in claim 4, characterized in that, Step S2.2 includes: constructing normalized feature vectors for the three types of heterogeneous node sets in the heterogeneous node set, and concatenating them according to the node index to form a global node feature matrix, thus forming the node feature matrix. ; Step S2.2.1: Construct the feature vector of the power node: For any power node Its eigenvector is defined as In the formula, The node load factor represents the proportion of the current power load to the rated capacity, reflecting the operating margin of the power node. The upper limit of power supply capacity characterizes the physical power supply or transmission limit of a power node and determines its support potential in disaster recovery. Voltage amplitude represents the voltage stability of power nodes and is a key indicator for determining whether a power system exceeds its limits. Step S2.2.2: Construct the feature vector of the charging station node: For any charging station node As a coupling hub connecting power and transportation, its feature vector is defined as follows. In the formula, The number of charging piles represents the static service capacity base of the charging station node; This provides real-time facility utilization, reflecting the current level of activity and queuing congestion risk at charging station nodes. The real-time total charging load not only represents its power extraction demand from the power network, but also indirectly reflects the intensity of traffic flow convergence at this charging station node. Step S2.2.3: Construct traffic node feature vectors: For any traffic node Its eigenvector is defined as In the formula, The road grade coefficient quantifies the road type and determines the upper limit of traffic capacity and the benchmark for traffic speed at traffic nodes. Cross-sectional traffic flow reflects the number of vehicles passing through the traffic node per unit time and characterizes the dynamic distribution of traffic flow. The Road Congestion Index is calculated based on the ratio of average vehicle speed to free-flow vehicle speed, representing traffic efficiency and vehicle congestion. Step S2.2.4 Stack all the feature vectors of power nodes, charging station nodes, and traffic nodes according to the global index order of the corresponding heterogeneous nodes in the unified hypergraph model of the power-transport coupling network to form the initial node feature matrix. X ; Step S2.3 includes: processing the node feature matrix X and the correlation matrix H Used as input to a hypergraph convolutional neural network for feature updates: Step S2.3.1, Heterogeneous node to hyperedge aggregation: The information of each heterogeneous node is determined based on the correlation matrix. They converge at their superedge, forming a process l The hyperedge features output after processing by a layer hypergraph convolutional neural network To simulate the impact of local heterogeneous node states on hyperedges; In the formula: This is a hyperdiagonal matrix used for normalization; It is an incidence matrix. For the process l After processing by a hypergraph convolutional neural network, the output node feature matrix, if heterogeneous nodes v If it belongs to a hyperedge, then the incidence matrix The corresponding element in the array is 1 if it is 1, otherwise it is 0; For the process l The output of the hypergraph convolutional neural network is a trainable feature transformation weight matrix. Step S2.3.2, Update the superedge to the heterogeneous node: Aggregated hyperedge features The updated node feature matrix is ​​then passed back to all heterogeneous nodes contained in the hyperedge. To simulate the feedback and adjustment of the hyperedge to the state of the internal heterogeneous nodes: In the formula: It is a diagonal matrix of node degrees; This is the hyperedge weight matrix, used to characterize the differences in importance of different types of hyperedges in feature propagation; It is a non-linear activation function; Through multiple layers of convolution stacking, the final output is a high-order topological feature embedding matrix of nodes. In the formula: Indicates the process The final output of the hypergraph convolutional neural network is the node feature matrix.

6. The method for identifying sentry nodes in a power-transportation coupled network as described in claim 5, characterized in that, Step S3 includes: Step S3.1: Construct a Markov decision process for the interaction between the intelligent agent and the unified hypergraph model of the power-transportation coupled network; Step S3.2: The agent uses a trained deep Q-network model to solve the Markov decision process and outputs the set of sentinel nodes that minimizes the normalized mean square error function.

7. The method for identifying sentry nodes in a power-transportation coupled network as described in claim 6, characterized in that, Step S3.1 includes: Step S3.1.1, State Space: Embedded by the higher-order topological features of the nodes and current decision-making steps t The set of sentinel nodes below mask vector Jointly composed of state vectors The set of sentinel nodes mask vector It is a one-dimensional Boolean vector. Let be the total number of heterogeneous nodes in the entire network, where is the th in the mask vector. The nth element being 1 indicates that the nth element is 1. Each heterogeneous node belongs to the current decision step. t The set of sentinel nodes below The first in the mask vector If any element is 0, it indicates that the heterogeneous node is a non-sentinel node, and the decision is made through the current step. t The mask vector of the sentinel node set below The intelligent agent can perceive the deployment location of heterogeneous nodes in the network topology and the current observation scheme in real time; Step S3.1.2, Motion Space: Define the action as node replacement: in the decision step Based on the current strategy, the agent decides to select from the current set of sentinel nodes. Remove the sentinel node with the lowest contribution. And select the heterogeneous node with the highest potential value from the remaining non-sentinel nodes. Join to form a new set of sentinel nodes. ; Step S3.1.3, Reward Function: Define rewards To represent the negative increment of the approximation error, a regularization term based on the Fisher information matrix is ​​introduced: In the formula: the first term This indicates encouragement to reduce error. Indicates the previous decision step Normalized mean square error function for the sentinel node set; Indicates the current decision step t The normalized mean square error function of the sentinel node set; the second term is the regularization constraint. For coefficients, These are the diagonal elements of the Fisher information matrix, used to penalize parameters important to historical tasks. Dramatic changes This represents the fixed optimal first position after the historical task training is completed. Each network weight parameter, This indicates the th in the currently training deep Q-network. Each network weight parameter; Diagonal elements of the Fisher information matrix This reflects the first in the deep Q-network model Weight parameters The importance of historical missions In the formula, For the agent in state The policy probability of the action defined in the action space is selected below. This serves as an experience replay pool for storing the data distribution of historical tasks. This represents the average of the sampling results over the historical data distribution. This represents the probability distribution of actions in a given state. Represents action variables; This represents the set of weight parameters for a deep Q-network model. Step S3.2 includes: Step S3.2.1: The deep Q-network architecture of the deep Q-network model includes: Input layer, used to input state vectors ; Feature fusion layer, which combines state vectors The embedded features and mask features are concatenated or weighted through an attention mechanism to extract the comprehensive feature representation of heterogeneous nodes under the current topology coverage, and the fused state feature representation is output. The hidden layer, connected to the feature fusion layer, is used to perform nonlinear transformation on the fused state feature representation, extract deep topological and state association patterns, and output the hidden feature representation. The output layer is used to perform linear mapping on the latent feature representation, output the predicted Q value of the node replacement action, and select the optimal action based on the predicted Q value to update the current set of sentinel nodes until the set of sentinel nodes with the smallest approximation error is found and output. Step S3.2.2: Construct the loss function for the deep Q-network architecture. : Introduce a target network with the same structure but lagging weight parameter updates into the loss function calculation and weight parameter update process of the deep Q-network architecture. The network weight parameters of the target network are denoted as... ; Define the loss function for a deep Q-network architecture. Predicted Q-values ​​for the current deep Q-network architecture With the target Q value Mean square error between: In the formula, Indicates from the experience replay pool A transition sample quadruple randomly sampled from the middle The calculated average loss value, Represents the state vector at the current moment; Indicates an action; Indicates a reward; This represents the state vector that the user transitions to at the next moment after performing an action. The weight parameters represent the deep Q-network architecture; the target Q-value. Calculations based on Bellman equations: , It is the negative increment of the approximation error. As a discount factor, This refers to finding the optimal action that yields the greatest expected return. This represents the predicted action value based on the action performed in the state vector at the next time step. This represents the candidate actions that may be executed under the state vector at the next moment; Transfer Sample Experience Replay Pool This is used to store the transition sample quadruples generated by the agent's interaction with the current hypergraph environment, for training the deep Q-network architecture; the transition sample quadruples are denoted as... In the formula, Indicates the current decision step The state vector; This represents the specific actions of the intelligent agent; This represents the immediate reward after the agent performs an action; Indicates the next decision step The state vector; Step S3.2.3: Train the constructed deep Q-network architecture: First, from the experience replay pool Randomly select a small batch of transfer sample quadruples Then the current state vector in the transition sample quadruple is... Input a deep Q-network architecture and compute the actions to be performed. The corresponding predicted Q value, and utilizing the instant reward. and the state vector of the next decision step The target Q-value obtained based on the Bellman equation and the target network is calculated, and then the mean squared error loss between the two is obtained; then, the backpropagation algorithm is used to obtain the loss function. Relative to the current network weight parameters gradient Finally, an adaptive moment estimation optimizer, such as Adam, is used to adjust the network weight parameters according to the gradient direction. Iterative updates are performed to gradually optimize the action value function; An adaptive moment estimation optimizer is used to optimize the current network weight parameters. Update: In the formula, For learning rate, The loss function is expressed with respect to the network weight parameters. gradient operator, The loss function represents the deep Q-network architecture; When the loss function After convergence or reaching the preset number of training iterations, fix the network weight parameters. The training of the deep Q-network architecture was completed.

8. A sentinel node identification system for a power-transport coupled network, characterized in that, include: The data value assessment module is used to quantify and filter the value of the collected multi-source heterogeneous data on power and transportation based on the optimization gradient value, and output high-value data. The Unified Hypergraph Modeling Module is used to construct a unified hypergraph model of the power-transportation coupled network and extract high-order topological features of nodes into the embedding matrix. The sentinel node identification module is used to model the sentinel node identification problem as a Markov decision process. It uses a trained deep Q-network model to solve the Markov decision process and outputs the set of sentinel nodes that minimizes the approximation error.

9. The power-transportation coupled network sentinel node identification system as described in claim 8, characterized in that, The data value assessment module specifically employs a strategy combining Monte Carlo approximation and single-sample gradient descent to calculate the optimized gradient value of the data points. The optimized gradient value is then compared with the set noise filtering threshold and redundancy filtering threshold to filter the multi-source heterogeneous data of the power and transportation systems, thereby obtaining high-value data.

10. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, which, when executed by the processor, implements the method of any one of claims 1 to 7.

Images