Method, recovery optimization method and device for identifying key nodes in a power system network

Abstract

The present invention discloses a method for identifying key nodes in a power system network, a restoration optimization method and a device. The identification method includes: constructing a DQN model to learn actions for selecting the best nodes to be removed under different network states, where the actions in the model are to remove nodes in the network and the corresponding layer nodes; training the DQN model using a training data set composed of power system network information under different states to generate a damaged key node identification model, and constructing a reward function using the influence ratio of the network path, the proportion of affected nodes, and the connectivity value of the affected path after removing the nodes and the layer nodes during the training process; obtaining the operation data of the target power system network, constructing a network topology structure and a distance matrix, and inputting them into the damaged key node identification model to obtain the identification result of the damaged key nodes. The present invention has the advantages of simple implementation method, low implementation cost and complexity, high execution efficiency and accuracy, and strong generalization ability and robustness.

Description

Technical Field

[0001] The present invention relates to the technical field of power system security monitoring, and in particular to a method for identifying key nodes in a power system network, a restoration optimization method, and a device therefor. Background Art

[0002] During the operation of a power system network, when sudden events such as natural disasters, equipment failures, and human attacks occur, it may cause node damage and affect the normal operation of the power grid. If the damaged key nodes in the disaster scenario can be accurately identified, it is convenient to perform restoration optimization on the damaged nodes to ensure the stability and reliability of the power grid. For the identification of key nodes, existing technologies usually adopt methods based on complex network theory or methods based on electrical parameters. The method based on complex network theory identifies key nodes in the power grid by calculating the topological attributes of nodes (such as node degree, betweenness centrality, etc.), and the method based on electrical parameters monitors the electrical parameters of each node in the power grid (such as voltage, current, power, etc.), analyzes the parameter changes of the nodes before and after the fault, and thus identifies the damaged nodes. However, the above methods based on complex network theory or methods based on electrical parameters have high computational complexity and low efficiency, and are difficult to apply to the current power grid environment with an expanding scale. Moreover, they can only identify relatively important nodes based on the node attributes in the power grid, and cannot quantify the impact of disasters on the connectivity of the power grid, so it is difficult to determine which are the key nodes under actual disasters.

[0003] Using machine learning and deep learning technologies to automatically identify damaged nodes, that is, by analyzing the images, videos or text data of power grid equipment to automatically identify damaged nodes, can improve the efficiency of identification and the degree of automation. However, the complexity of the traditional machine learning model is actually still high and its generalization ability is insufficient, and it is difficult to adapt to different power grid structures and operating conditions. Moreover, the power grid structure and operating state are dynamically changing, and the traditional identification method based on the machine learning model is difficult to adapt to the changing power grid structure and operating state, resulting in a still low actual identification accuracy and insufficient stability under different operating conditions, and still difficult to meet the high reliability and high stability requirements of power grid operation. Summary of the Invention

[0004] The technical problem to be solved by the present invention is: aiming at the problems existing in the prior art, the present invention provides a method for identifying key nodes in a power system network, a restoration optimization method, and a device therefor, which has a simple implementation method, low implementation cost and complexity, high execution efficiency and accuracy, and strong generalization ability and robustness, and can quickly and accurately identify the key damaged nodes in the power grid under different disaster states, and then can determine an effective node restoration optimization strategy.

[0005] To solve the above technical problems, the technical solution proposed by the present invention is as follows:

[0006] A method for identifying key nodes in a power system network, the steps including:

[0007] Construct a DQN (Deep Q-Network) model to learn the action of selecting the best node to remove under different network states, where the network states include the normal connected state of the network and the network states under different disaster types. The input features of the model are the state representations of the power system network, and the output of the model is a sequence of nodes arranged in the removal order and used as the identified damaged key nodes. The action of the agent in the model is to remove nodes and corresponding layer nodes in the network, and the layer nodes are the nodes within a specified distance centered on the target node;

[0008] Use a training data set containing power system network information under different network states to train the DQN model to generate a damaged key node identification model. During the training process, use the node removal impact index to construct a reward function, and the node removal impact index includes the impact ratio of the network path, the affected node ratio, and the connectivity value of the affected path after removing the node and the layer nodes;

[0009] Obtain the operation data of the target power system network, and construct a network topology structure and a distance matrix according to the operation data of the target power system network. Each element in the distance matrix corresponds to the shortest path distance between nodes in the network;

[0010] Input the operation data, topology structure, and distance matrix of the target power system network into the trained damaged key node identification model to obtain the identification result of the damaged key nodes in the current target power system network.

[0011] Further, the constructing the network topology structure and the distance matrix according to the operation data of the target power system network includes:

[0012] Collect the operation data of the power system network, and the operation data includes voltage level, current capacity of the connection, and condition of the transmission line;

[0013] Identify the node and connection information in the power system network according to the collected operation data of the power system network, obtain the mutual relationship between the nodes and the connections, determine the connection mode and path between each node in the power grid, and construct a topology diagram of the power system network;

[0014] Calculate the adjacency matrix of the graph based on the node and connection information in the power system network where is the set of nodes in the power system network, is the set of edges in the power system network, represents a node to node the weight of the edge;

[0015] According to the adjacency matrix obtain the shortest distance between any pair of nodes to obtain the distance matrix.

[0016] Furthermore, the adjacency matrix is configured according to the following formula:

[0017]

[0018] The distance matrix each element in is determined according to the following formula:

[0019]

[0020]

[0021] wherein, represents using the adjacency matrix , node and node as input to execute the Dijkstra algorithm, represents from node to node the shortest path distance.

[0022] Furthermore, the calculation expression of the constructed reward function is:

[0023]

[0024] wherein, and respectively represent the weight parameters for measuring the influence path ratio and the affected node ratio, represents the influence ratio of the network path after removing the node and its surrounding nodes, represents the ratio of the affected nodes after removing the node and its surrounding nodes, represents the connectivity value affected after removing the node and its surrounding nodes;

[0025] The influence ratio of the network path after removing the node and its surrounding nodes is calculated according to the following formula:

[0026]

[0027] wherein, represents the number of nodes that the th node in the remaining network can reach after removing the node and its surrounding nodes, represents the number of reachable node pairs in the power system network, represents the total number of nodes in the network, represents the node set of the remaining network after removing the node and its layer nodes;

[0028] The proportion of nodes affected after removing the node and its layer nodes The calculation expression is:

[0029]

[0030] where represents the number of layer nodes, represents the shortest distance centered on the removed node and the layer node set formed by the nodes within;

[0031] The connectivity value affected after removing the node and its layer nodes The calculation expression is:

[0032]

[0033] where, represents the removed node and the number of nodes in the largest connected subgraph after removing the corresponding layer nodes , represents the set of removed nodes, represents the number of nodes in the largest connected subgraph of the power system network .

[0034] Furthermore, training the DQN model using a training dataset composed of power system network information in different states includes:

[0035] Taking the training dataset composed of power system network information in different states as the input of the DQN model;

[0036] During the training process, at each step, all possible removal actions are evaluated based on the current network state, that is, the nodes to be removed and the corresponding layer nodes, where first a node is selected as the current removed node, and the reward is predicted by the DQN model according to the reward function, and the node with the maximum reward is found , at this time the removed node set , then the next node is selected as the current removed node, and the node with the maximum reward is determined , at this time the removed node set , and so on, adopting a step-by-step expansion strategy, at the th time, by predicting the node with the maximum reward , a set of node removal sequences is found , represents the number of removed nodes;

[0037] Judge whether the set of removed nodes contains all the nodes of the original network, that is, whether it is equal to . If so, it is determined that the entire network has been completely damaged, the training stops, and a sequence of nodes arranged in the removal order is obtained as the identified damaged critical nodes. Otherwise, the model continues to be trained. The set of removed nodes is the union of the node set and the corresponding layer node set .

[0038] An optimization method for power system network disaster simulation and restoration, the steps include:

[0039] Identify the set of damaged critical nodes in the target power system network according to the above method , L represents the scope of influence of the damaged nodes;

[0040] Randomly select a set of nodes from the identified set of damaged critical nodes and the corresponding layer nodes for interruption to simulate different disaster scenarios, and obtain the current set of damaged nodes as , is the union of the node set and the corresponding layer node set , then the set of critical nodes to be repaired first is ;

[0041] Obtain the network state and the number of connected subgraphs of the target power system network under the current disaster scenario simulation state;

[0042] Select nodes for restoration according to the network state and the number of connected subgraphs under the current disaster scenario simulation state, and enable the backup lines between the selected nodes and the remaining network nodes. First, restore the damaged critical nodes in the network. Among them, if the network is in a disconnected state, identify the nodes that can connect the connected subgraphs fastest and use them as the target repair nodes; preferentially repair the target repair nodes until the network forms a single connected subgraph to make the network reach a connected state; when the network reaches a connected state, sequentially select the target nodes that optimize the network performance for restoration and enable the backup lines of the target nodes to obtain the optimal order of repair nodes.

[0043] Further, by constructing a hierarchical reinforcement learning model, the hierarchical reinforcement learning model is used for node recovery. Among them, in the hierarchical reinforcement learning model, the underlying network preferentially repairs the target repair nodes until the network forms a single connected subgraph and the network reaches a connected state. The state of the underlying network is defined as the number and size distribution of the remaining network connected subgraphs, and the action is defined as selecting a node for repair and enabling the backup line of the node to be repaired; in the hierarchical reinforcement learning model, the high-level network is used to sequentially select the target nodes that optimize the network capacity for recovery. The state of the high-level network is defined as the current remaining network, the set of nodes to be repaired, and the connectivity metric of the entire network, and the action is defined as selecting a node that can improve the network performance for repair and enabling the backup line of the selected node.

[0044] Further, the reward function of the underlying network is defined as , represents the size of the largest connected subgraph of the network after repairing the node, represents the number of connected subgraphs of the network after repairing the node, and respectively represent the weight parameters for measuring the size of the largest connected subgraph of the network and the number of connected subgraphs of the network, represents the first repaired node the average clustering coefficient of the network after is the weight for preferentially repairing key nodes, is the first indicator function, which takes the value of 1 when the first repaired node belongs to the set of key nodes to be preferentially repaired and 0 otherwise;

[0045] The reward function of the high-level network is defined as , represents the average clustering coefficient of the network after repairing the node , represents the connectivity metric of the network determined according to the number of nodes in the largest connected subgraph in the remaining network except the second repaired node , represents the remaining network except the second repaired node , and represent weight parameters, is the second indicator function, which takes the value of 1 when the second repaired node belongs to the set of key nodes to be preferentially repaired and 0 otherwise.

[0046] Further, using the hierarchical reinforcement learning model for node recovery includes:

[0047] Establish a low-level state table and a high-level state table for storing the expected return values of each state-action pair of the low-level network and the high-level network respectively;

[0048] Obtain the set of removed nodes and the remaining network data in the power system network under different disaster types and input them into the hierarchical reinforcement learning model;

[0049] In the low-level network of the hierarchical reinforcement learning model, at each state according to the low-level state table and the exploration rate select an action with probability select the action with the largest value in the low-level state table and with probability randomly select an action, execute the selected action and then obtain the return and the next state according to the reward function, and update the low-level state table according to , where represents the discount factor; repeat the execution until the number of connected subgraphs of the network becomes 1 after repairing the nodes, and then transfer to the high-level network for node recovery;

[0050] When performing node recovery in the high-level network, at each state according to the high-level state table and the exploration rate select an action with probability select the action with the largest value in the low-level state table and with probability randomly select an action, execute the selected action and then obtain the return and the next state according to the reward function, and update the low-level state table according to , repeat the execution until the preset iteration stop condition is reached, and determine the optimal low-level state table and high-level state table , and obtain a hierarchical reinforcement learning model with optimal parameters;

[0051] Obtain the set of removed nodes and the remaining network data in the current power system network and input them into the hierarchical reinforcement learning model with optimal parameters for node recovery.

[0052] A power system network disaster simulation and recovery optimization device includes a processor and a memory. The memory is used to store a computer program, and the processor is used to execute the computer program to perform the above method.

[0053] Compared with the prior art, the advantages of the present invention are as follows:

[0054] 1. By combining the quantification of disaster impacts and the deep reinforcement learning method, the present invention constructs a DQN model to learn the action of selecting the best nodes to be removed under different network states. The actions performed by the agent in the model each time include removing nodes in the network and the corresponding layer nodes, which can comprehensively consider the impact of removed nodes on the nodes within the layer, and use the removed node impact index to construct a reward function during the model training process, which can accurately quantify the impact of the removed nodes on the grid connectivity, accurately evaluate the impact of different disasters on the grid connectivity, use the deep reinforcement learning algorithm to dynamically learn and optimize the nodes in the power grid, can quickly and accurately identify the damaged key nodes affected by different disasters, and the identification accuracy is not easily affected by changes in the grid structure and operating state, which can improve the generalization ability of the model and the robustness of the identification.

[0055] 2. The present invention can further evaluate the damage conditions of key nodes and paths by simulating different disaster scenarios to simulate the impact of natural disasters or other emergencies on the power grid. At the same time, it uses the reinforcement learning strategy method to guide the node repair work of the power grid, determines the repair priority of nodes according to the network state of whether the network is connected, and adopts a phased method to perform node recovery. First, it preferentially repairs the target repair nodes and enables their backup lines until the network forms a single connected subgraph to make the network reach the connected state, which can accurately achieve the rapid repair of the power network after the disaster. Then, when the network reaches the connected state, it sequentially selects the target nodes that optimize the network performance for recovery and enables their backup lines, so that not only can the rapid recovery of the network be achieved, but also the optimization of the network performance after recovery can be effectively realized, ensuring the rapid response and efficient recovery of the power grid in the face of disasters. Description of the Drawings

[0056] Figure 1 It is a schematic flow chart of the implementation of the method for identifying key nodes in the power system network in Embodiment 1 of the present invention.

[0057] Figure 2 It is a schematic diagram of the architecture of the DQN model adopted in Embodiment 1 of the present invention.

[0058] Figure 3 It is a schematic flow chart of the implementation of node recovery using a hierarchical reinforcement learning model in Embodiment 2 of the present invention.

[0059] Figure 4It is a schematic diagram of the principle for the present invention to achieve key node recognition and node recovery in specific application embodiments. Detailed implementation manners

[0060] The present invention will be further described below in conjunction with the accompanying drawings of the specification and specific preferred embodiments, but the protection scope of the present invention is not limited thereby.

[0061] By combining the quantification of the impact of disasters and the deep reinforcement learning method, the present invention introduces the influence ratio of the network path, the ratio of affected nodes, and the connectivity value of the affected path after removing nodes and circle-layer nodes as the influence index of the removed nodes to quantify the influence state after removing nodes under different network states. At the same time, a DQN model is constructed to learn the actions of selecting the best removed nodes under different network states. The actions performed by the agent in the model each time include removing nodes and corresponding circle-layer nodes in the network. The circle-layer nodes are the nodes within a specified distance centered on the target node, which can comprehensively consider the impact of the removed nodes on the nodes within the circle layer, and use the influence index of the removed nodes to construct a reward function during the model training process, so as to reward or punish the actions according to the influence state of the removed nodes, which can accurately quantify the impact of the removed nodes on the power grid connectivity, accurately evaluate the impact of different disasters on the power grid connectivity, and thus use the deep reinforcement learning algorithm to dynamically learn and optimize the nodes in the power grid, quickly and accurately identify the damaged key nodes affected by different disasters, and the recognition accuracy is not easily affected by the changes in the power grid structure and operation state, which can improve the generalization ability of the model and the robustness of the recognition.

[0062] As Figure 1 shown, the steps of the method for identifying key nodes of the power system network in this embodiment include:

[0063] Step S01. Construct a DQN model to learn the actions of selecting the best removed nodes under different network states. The network states include the normal connected state of the network and the network states under different disaster types, where the disaster types include various types of extreme weather disasters such as typhoons, heavy rains, and floods. The input features of the model are the state representations of the power system network, and the state representations include information such as nodes and edges in the network. The output of the model is a sequence of nodes arranged in the removal order and used as the identified damaged key nodes. The actions of the agent in the model are to remove nodes and corresponding circle-layer nodes in the network, and the circle-layer nodes are the nodes within a specified distance centered on the target node.

[0064] In order to quantify the impact of the removed nodes on the surrounding nodes and the performance of the entire network, this embodiment defines an influence circle layer to simulate the impact of the removed nodes on the circle-layer nodes and the performance of the entire network. Specifically, according to the impact of different levels of disasters, the affected range of the damaged nodes caused by the disasters is defined as , and any node , the disaster impact circle layer nodes are defined as , where represents the shortest path length from node to node . For example, when takes the value of 1, the circle layer nodes at this time are the direct neighbor nodes of node . When takes other values, according to the Dijkstra algorithm, the shortest distances from each node in the node set to other nodes can be obtained, that is, the nodes within the shortest distance centered on this node are the circle layer nodes.

[0065] In order to build a DQN model to learn the action of selecting the best removal node under different network states, in a specific application embodiment, as Figure 2 shown, the deep reinforcement learning model can be defined in the following way:

[0066] Step S101. Initialize the current value network and the target value network, and set the number of iterations C, various weights, and initialize the experience replay buffer D (experience pool).

[0067] Step S102. Obtain the power system network information under different network states to form a training data set, and output . The power system network information includes information such as network nodes and connections. For example, an undirected network with 15 nodes and 24 connections can be used as the input of the model.

[0068] Step S103. Select the action a to be executed in the way, that is, select the action with the highest Q value in the current state with a probability, and randomly select an action with a probability. The expression can be represented as:

[0069] , (5)

[0070] where, is the selected action, is the current state, is the exploration probability, represents the current number of training times, M represents the total number of training times, is the action with the highest Q value in the current state, represents random selection.

[0071] As the training progresses, will gradually become smaller. At this time, the maximum Q value is used with a higher probability for action selection, which can accelerate the training process.

[0072] Step S104. Select an action and obtain a reward R, and at the same time obtain the next state .

[0073] Step S105. Store the sample in the experience replay buffer D. If the experience replay buffer D is full, delete the old sample and then store the new sample.

[0074] Step S106. Randomly draw a group of samples from the experience replay buffer D and update the weights of the DQN network using the gradient descent method:

[0075] (6)

[0076] where is the learning rate is the gradient of the current network parameters.

[0077] Step S107. After multiple rounds of iteration, update the weights of the DQN network until the maximum number of iterations is reached or other preset iteration conditions are met, save the optimal model parameters to obtain the trained model. During the iteration process, the following loss function can be specifically minimized:

[0078] (7)

[0079] where is the parameter of the current network.

[0080] Step S02. Use the training dataset composed of power system network information in different network states to train the DQN model to generate a damaged critical node recognition model. During the training process, use the node removal impact index to construct the reward function. The node removal impact index includes the impact ratio of the network path, the affected node ratio, and the connectivity value of the affected path after removing the node and its surrounding nodes. The impact ratio of the network path after removing the node and its surrounding nodes is used to represent the ratio of the network paths affected after removing the node and its surrounding nodes. The affected node ratio is used to represent the ratio of the nodes affected after removing the node and its surrounding nodes. The connectivity value of the affected path is used to represent the magnitude of the impact on the network path connectivity after removing the node and its surrounding nodes.

[0081] In this embodiment, by combining graph theory analysis and a deep reinforcement learning model, the DQN model is trained to learn the action of selecting the best node to be removed in different network states. The operation data of the power grid is analyzed using this model, and the reward function is defined using the node removal impact index. A key node sequence crucial for the operation of the power grid can be identified, and the degree of influence of the lost nodes on the stability and power supply capacity of the entire power grid can be determined. These nodes play a core role in the power grid, and their faults or damages may have a significant impact on the power grid, realizing the identification of damaged key nodes.

[0082] To quantify the indicators of the disaster-affected network, in this embodiment, the specific steps for defining the impact index of each removed node after removing the node and the layer nodes are as follows:

[0083] (1) Influence ratio of network paths

[0084] When several nodes in the network are removed, the set of damaged nodes is defined as , represents the number of removed nodes, and the union of the layer nodes of the disaster-affected node set is defined as: where traverses all the nodes in, and the set of removed nodes is defined as The set of nodes remaining in the network after removing the node and the layer nodes is .

[0085] Assume that the number of reachable node pairs in the original network is , represents the total number of nodes in the network. Then, for each node in the remaining network, the Dijkstra algorithm can be used to calculate the shortest path to other nodes , and record the number of nodes that each can reach (excluding itself). Furthermore, the number of reachable node pairs in the remaining network is . Then, as an alternative implementation, the influence ratio of the network path after removing the node and the layer nodes can be defined as:

[0086] (1)

[0087] where, represents the number of nodes that the -th node in the remaining network after removing the node and the layer nodes can reach, represents the set of nodes in the remaining network.

[0088] (2) Proportion of affected nodes

[0089] As an alternative implementation, the proportion of nodes affected after removing a node and its surrounding layer nodes can be defined as the ratio of the number of nodes in the union of the surrounding layer nodes of the node set to the total number of nodes in the network. For example, the calculation expression can be represented as:

[0090] (2)

[0091] where, represents the number of surrounding layer nodes, represents the shortest distance centered on the removed node within which the set of surrounding layer nodes formed by the nodes is located.

[0092] (3)The affected connectivity value

[0093] As an alternative implementation, the affected connectivity value can be defined as the ratio of the difference between the number of nodes in the largest connected subgraph of the original network and the number of nodes in the largest connected subgraph of the remaining network to the number of nodes in the largest connected subgraph of the original network. For example, the affected connectivity value of the calculation expression can be represented as:

[0094] (3)

[0095] where, represents the removed node and the number of nodes in the largest connected subgraph after the corresponding surrounding layer nodes are removed , represents the set of nodes after the removed node is removed, represents the number of nodes in the largest connected subgraph of the power system network , is the connectivity metric of the network after the node is removed, and can specifically be defined as the number of nodes in the largest connected subgraph in the remaining network.

[0096] In this embodiment, by introducing the surrounding layer nodes and the removed node impact indicators, a comprehensive evaluation index system for the impact on the network after removing a node in the network can be formed. Using this system to analyze the set of each node and its surrounding layer nodes in the network, the impact on the grid connectivity under different disaster events can be accurately quantified, breaking through the limitations of traditional single-node evaluation, effectively evaluating the impact of nodes and edges in the power network in a disaster scenario, and thus being able to more comprehensively identify critical nodes.

[0097] As an alternative implementation, in this embodiment, for the node removal sequence, considering the change in network connectivity after removing a node and its surrounding layer nodes, the reward function can be constructed according to the following formula using the above various removed node impact indicators:

[0098] (4)

[0099] Among them, and respectively represent the weight parameters for measuring the proportion of the influence path and the proportion of affected nodes, represents the influence proportion of the path, represents the proportion of affected nodes, represents the connectivity value of the influence.

[0100] The reward function constructed in the above manner can measure the degree to which a certain node affects and damages the entire network by using the influence proportion of the network path and the proportion of affected nodes. At the same time, considering the connectivity metric, it can determine whether the network is disconnected after each node is affected by the disaster. Therefore, by comprehensively considering the influence proportion of the network path, the proportion of affected nodes, and the connectivity value of the influence path after removing the nodes and the layer nodes, the reward function can preferentially identify the nodes with a high degree of network damage and those that disconnect the network into several connected subgraphs, thereby accurately identifying the key damaged nodes.

[0101] In this embodiment, by training the DQN model, a set of node sequences sorted by key importance found by the reinforcement learning model is finally output. The nodes in the node sequence are arranged in the order of removal, and the arrangement order can represent the magnitude of the impact of each node on the network destructiveness. As an alternative implementation, training the DQN model using the training data set includes:

[0102] Step S201. Use the training data set composed of power system network information in different states as the input of the DQN model;

[0103] Step S202. During the training process, at each step, evaluate all possible removal actions based on the current network state, that is, the nodes to be removed and the corresponding layer nodes. Among them, by first selecting a node as the current removal node, the DQN model predicts the reward according to the reward function, and finds the node with the maximum reward , and at this time the node removal set is . Then select the next node as the current removal node, determine the node with the maximum reward , and at this time the node removal set is , and so on. Adopting a step-by-step expansion strategy, the node removal set at the th time is . At the th time, by predicting the node with the maximum reward, a set of node removal sequences is found;

[0104] Step S203. Determine whether the removed node set contains all the nodes of the original network, that is, , the set of nodes and the set of layer nodes If it is the union, it is determined that the entire network has been completely damaged at this time, the training stops, and a node sequence arranged in the removal order is obtained As the identified damaged critical nodes, the node sorting order is sorted according to the comprehensive impact degree on the network, where the set of removed nodes is the set of nodes and the corresponding set of layer nodes of the union.

[0105] Step S03. Obtain the operation data of the target power system network, construct the network topology structure and the distance matrix according to the operation data of the target power system network, and each element in the distance matrix corresponds to the shortest path distance between nodes in the network.

[0106] As an optional implementation manner, constructing the network topology structure and the distance matrix according to the operation data of the target power system network includes:

[0107] Step S301. Collect the operation data of the power system network, and the operation data includes voltage level, current capacity of the connection, condition of the transmission line, etc.

[0108] Optionally, after collecting the operation data of the power grid, it also includes cleaning and preprocessing the collected data to ensure the accuracy and availability of the data.

[0109] Step S302. Identify the node and connection information in the power system network according to the collected operation data of the power system network, obtain the mutual relationship between the nodes and the connections, determine the connection method and path between each node in the power grid, and construct the topological structure diagram of the power system network.

[0110] By sorting out the node and connection information of the power grid through the operation data of the power system network, the topological structure diagram of the power grid can be constructed. By analyzing the mutual relationship between each node and connection, the connection method and path between each node in the power grid can be determined.

[0111] Step S303. Calculate the adjacency matrix of the graph based on the node and connection information in the power system network , where is the set of nodes in the power system network, is the set of connections in the power system network, represents the weight of the edge from node to node , and the shortest distance between any node pair is obtained according to the adjacency matrix to obtain the distance matrix .

[0112] The adjacency matrix is used to represent the connection relationship between nodes in the power grid, where the elements of the matrix represent the transmission capacity or weight between nodes. The distance matrix can be constructed by using the Dijkstra algorithm based on the adjacency matrix That is, the Dijkstra algorithm can be used to calculate and find the shortest path between any node pairs in the power grid with the help of the adjacency matrix so as to obtain the distance matrix . Specifically, a graph is constructed based on the node and edge information of the power grid , and the total number of nodes in the network is determined and the total number of edges . For each edge , a weight is assigned , and the weight represents the transmission capacity of the edge

[0113] As an alternative implementation, the weights of all edges can be set to 1, indicating that the transmission capacities of the edges in the power grid are approximate, that is, set to represent the weight of the edge from node to node . If there is no edge between node and node , then , indicating unreachable, and the diagonal elements , and the adjacency matrix is configured according to the following formula

[0114] (8)

[0115] Then, based on the adjacency matrix, the Dijkstra algorithm is used to find the distance between any node pairs. Specifically, for each node , the Dijkstra algorithm is run to find the shortest path distance from to other nodes : , and the distance matrix is obtained , that is, the elements in the distance matrix are determined according to the following formula

[0116] (9)

[0117]

[0118] where represents executing the Dijkstra algorithm with the adjacency matrix , node and node as inputs Indicates the slave node to the node of the shortest path distance.

[0119] Step S04. Input the operation data, topology structure, and distance matrix of the target power system network into the trained damaged critical node recognition model to obtain the recognition result of the damaged critical node.

[0120] In this embodiment, the damaged critical node recognition model trained based on the DQN model in step S01 is utilized. When it is necessary to identify the critical nodes of the target power system network, first obtain the operation data of the target power system network. By analyzing the operation data, construct the network topology structure and distance matrix, and then input them into the trained damaged critical node recognition model. The model can quickly and accurately identify the damaged critical nodes of the target power system network.

[0121] Through the above steps, this embodiment can combine the quantification of the impact of disasters on nodes and the deep reinforcement learning method, evaluate the impact on the layer nodes of the power grid nodes according to the disaster impact, use the deep reinforcement learning method to dynamically learn and optimize the nodes in the power grid, quickly and accurately identify the damaged critical nodes of the network under different disasters, and improve the recognition efficiency, reliability, and robustness of the critical nodes.

[0122] Embodiment 2:

[0123] Based on Embodiment 1, this embodiment simulates different disaster scenarios to evaluate the damage conditions of critical nodes and paths to simulate the impact of natural disasters or other emergencies on the power grid. At the same time, use the reinforcement learning strategy method to guide the node repair work of the power grid, determine the repair priority of the nodes according to the network state of whether the network is connected, and adopt a phased manner to restore the nodes. First, preferentially repair the target repair nodes and enable their backup lines until the network forms a single connected subgraph to make the network reach the connected state, which can accurately achieve the rapid repair of the power network after the disaster. Then, when the network reaches the connected state, sequentially select the target nodes that optimize the network performance for restoration and enable their backup lines, so that not only can the rapid restoration of the network be achieved, reducing the number of connected subgraphs in the network, but also the optimization of the network performance after restoration can be effectively realized, and it can ensure the rapid response and efficient restoration of the power grid in the face of disasters.

[0124] The steps of the power system network disaster simulation recovery optimization method in this embodiment include:

[0125] Step 1: Identify the set of damaged critical nodes in the target power system network according to Embodiment 1 , L indicating the affected range of the damaged nodes;

[0126] Step 2: Randomly select a set of node sets from the identified set of damaged critical nodes and the corresponding layer nodes to interrupt and simulate different disaster scenarios, and obtain the current set of damaged nodes as , which is the union of the node set and the corresponding set of layer nodes . Then the set of critical nodes to be repaired first is ;

[0127] Step 3: Obtain the network state of the target power system network in the current disaster scenario simulation state and the number of connected subgraphs;

[0128] Step 4: Select nodes for restoration according to the network state and the number of connected subgraphs in the current disaster scenario simulation state, and enable the backup lines between the selected nodes and the remaining network nodes. First, restore the damaged critical nodes in the network. Among them, if the network is in a disconnected state, identify the nodes that can connect the connected subgraphs fastest and use them as the target repair nodes; preferentially repair the target repair nodes and enable their backup lines until the network forms a single connected subgraph and the network reaches a connected state; when the network reaches a connected state, sequentially select the target nodes that optimize the network performance for restoration and enable the backup lines of the target nodes to obtain the optimal order of repair nodes.

[0129] Specifically, according to the impact of disasters of different degrees, define the affected range of nodes damaged by disasters as , the set of nodes damaged by disasters (damaged nodes) . Then the union of the layer nodes of the node set is also removed, where traverse all the nodes in , and according to the Dijkstra algorithm, the shortest distances from each node in the node set to other nodes can be calculated, and then the nodes that meet the shortest distance centered on this node can be determined, that is, the layer nodes. Furthermore, the set of nodes to be removed can be determined as , and the set of nodes in the remaining network is , and the set of nodes to be repaired is also . .

[0130] In specific application embodiments, to achieve the optimization of power system network disaster simulation and recovery, first analyze the post-disaster state of the power grid to determine whether the network is disconnected and the number of connected subgraphs; if the network is disconnected, identify the nodes that can connect each connected subgraph the fastest, and preferentially repair these nodes that can connect each connected subgraph the fastest and enable their backup lines to reduce the number of connected subgraphs, and then continue the repair operation until the network forms a single connected subgraph; at this time, the network is already connected, select the optimal nodes from the remaining set of removed nodes for recovery, recover by selecting the nodes that can maximize the improvement of network capabilities and enable the backup lines of the nodes to be recovered, and then based on the current network state, continue to select the next optimal node for recovery, repeat the selection and recovery process until the optimal repair node order is obtained, complete the post-disaster recovery of the power grid, and ensure that the connectivity and various capabilities of the power grid are maximally improved. Further, the recovered power grid can be tested and evaluated to ensure stability and reliability.

[0131] As an alternative implementation, a hierarchical reinforcement learning model can be constructed and used for node recovery. Among them, in the hierarchical reinforcement learning model, the underlying network preferentially repairs the target repair nodes until the network forms a single connected subgraph, making the network reach a connected state. The state of the underlying network is defined as the number and size distribution of the remaining network connected subgraphs, and the action is defined as selecting a node for repair and enabling the backup line between the selected node and the remaining network nodes. The high-level network in the hierarchical reinforcement learning model is used to sequentially select the target nodes that optimize the network capabilities for recovery. The state of the high-level network is defined as the current remaining network, the set of nodes to be repaired, and the connectivity metric of the entire network, and the action is defined as selecting a node that improves the network performance for repair and enabling the backup line between the selected node and the remaining network nodes. For example, the selection of this node is based on its increase in the network clustering coefficient and the optimization of the entire network performance. That is, in a hierarchical reinforcement manner, the underlying network connects the entire network, that is, merges the connected subgraphs of the remaining network, reduces the number of connected subgraphs in the network, and preferentially repairs the nodes that can connect multiple connected subgraphs. When the number of connected subgraphs in the network becomes 1, it automatically changes from the strategy of focusing on the merger of connected subgraphs to the strategy of optimizing network performance, realizing dynamic hierarchical adjustment, and then using the high-level network to optimize the connectivity and clustering coefficient of the network. After the number of connected subgraphs in the network becomes 1 during the repair process, considering the improvement of network connectivity and the increase in the clustering coefficient, it can effectively optimize the performance of the recovered network and improve the efficiency and effect of repair.

[0132] Furthermore, the reward function of the underlying network can be defined as , represents the size of the largest connected subgraph of the network after repairing the node, represents the number of connected subgraphs of the network after repairing the node, and respectively represent the weight parameters for measuring the size of the largest connected subgraph of the network and the number of connected subgraphs of the network, represents the first repaired node and the average clustering coefficient of the network after that, is the weight for measuring the priority of repairing critical nodes, is the first indicator function, which takes the value of 1 when the first repaired node belongs to the set of critical nodes to be repaired preferentially and 0 otherwise; the reward function of the high-level network can be defined as , represents the average clustering coefficient of the network after repairing the node and represents the connectivity metric of the network determined according to the number of nodes in the largest connected subgraph in the remaining network except the second repaired node , represents the remaining network except the second repaired node , and represent the weight parameters, is the second indicator function, which takes the value of 1 when the second repaired node belongs to the set of critical nodes to be repaired preferentially and 0 otherwise.

[0133] To calculate the above average clustering coefficient , , the following method can be adopted:

[0134] The average clustering coefficient is the clustering coefficient of the network, which is an index for measuring the local connectivity of the network and can describe the degree of mutual connection between the neighbor nodes of a node. A high clustering coefficient indicates that the neighbor nodes of a node form a tight community or group.

[0135] For a node in the network, its clustering coefficient can be expressed as , where is the number of edges actually existing between the neighbor nodes of node , is the degree of node , then for the remaining network, the average clustering coefficient can be expressed as: , where is the node selected for repair, represents the set of nodes in the remaining network.

[0136] As an alternative implementation, as Figure 3 shown, using a hierarchical reinforcement learning model for node recovery includes:

[0137] Step 401. Establish a low-level state table and a high-level state table for storing the expected return values of each state-action pair of the low-level network and the high-level network, respectively.

[0138] During the initialization setup of the model, it includes initializing the low-level state table and the high-level state table , setting the learning rate and the discount factor as well as the exploration rate , and obtaining the set of removed nodes in the original network and the remaining network data as input data.

[0139] Step 402. Obtain the set of removed nodes in the power system network under different disaster types and the remaining network data and input them into the hierarchical reinforcement learning model.

[0140] Step 403. In each state of the low-level network in the hierarchical reinforcement learning model , according to the low-level state table and the exploration rate , select an action , with probability select the action with the largest value in the low-level state table and with probability randomly select an action, execute the selected action and then obtain the return and the next state according to the reward function, and update the low-level state table according to , where represents the discount factor; repeat the execution until the number of connected subgraphs of the network becomes 1 after repairing the nodes, and then transfer to the high-level network for node recovery.

[0141] Specifically, in each state, according to the table and the exploration rate , select an action, with probability select the action with the largest value in the table, and with probability

[0142] randomly select an action, that is:

[0143] where is the selected action is the current state, is the exploration probability, m represents the current number of training times, and M represents the total number of training times. is the action with the highest Q value in the current state. As the training progresses, the value of will continuously decrease. At this time, the maximum Q value is used with a higher probability to select actions, which can accelerate the training process.

[0144] Then execute the selected action and obtain the reward and the next state and update the table using the Q-learning update formula: Update the table. Repeat the above steps. When the number of connected subgraphs of the network becomes 1 after repairing the node, automatically transfer to the high-level logic to use the table and the new reward function That is, find the repaired node according to the average clustering coefficient of the network and the connectivity metric of the network.

[0145] Step 404. When performing node recovery in the high-level network, at each state select an action according to the high-level state table and the exploration rate with probability select the action with the largest value in the low-level state table and with probability randomly select an action, execute the selected action and obtain the reward according to the reward function after execution and the next state and update the low-level state table according to Repeat the execution until the preset iteration stop condition is reached, and determine the optimal low-level state table and the high-level state table to obtain a hierarchical reinforcement learning model with optimal parameters. and the high-level state table to obtain a hierarchical reinforcement learning model with optimal parameters.

[0146] Specifically, select an action according to the table and the exploration rate with probability select the action with the largest value in the table, with probability randomly select an action, execute the selected action and obtain the reward and the next state ; update the table using the Q-learning update formula: , repeat the above steps until the maximum number of iterations is reached or other stopping conditions are met. Finally, save the optimal and table as the model parameters to obtain a hierarchical reinforcement learning model with optimal parameters.

[0147] Step 405. Obtain the set of removed nodes in the current power system network and the remaining network data, and input them into the hierarchical reinforcement learning model with optimal parameters for node restoration.

[0148] In summary, in this embodiment, by considering the impact of disasters on the entire power network. By constructing a removal strategy including layer nodes, combining deep reinforcement learning to identify the damaged key nodes in the network, and then simulating different disaster situations, and combining the hierarchical reinforcement learning method to generate node restoration strategies for different disaster situations, it can accurately evaluate the impact of disasters on the grid connectivity, identify key nodes and optimize the disaster restoration strategy at the same time. Using the hierarchical reinforcement learning strategy not only focuses on the rapid physical connection restoration of the network, but also can take into account the optimization of the network performance after restoration, ensuring that the power grid can quickly resume normal power supply in the face of disasters, reducing the social and economic losses caused by power outages, thereby improving the recovery ability of the power grid in disaster situations, and enhancing the overall robustness of the power grid, thus significantly improving the response speed and recovery efficiency of the power grid during disasters, enhancing the overall stability and robustness of the power grid, and facilitating power grid management and rapid disaster response, making power grid management more refined, effectively preventing and reducing large-scale power outage events, and ensuring the stable operation of the power grid.

[0149] To verify the effect of the present invention, in a specific application embodiment, the above method of the present invention is used to identify key nodes and restore nodes under the influence of disasters in a 15-node, 24-edge undirected network, as Figure 4 shown. First, analyze the operation data of the network to determine the influence layer nodes when each node is removed, and then learn through the DQN model to identify a set of nodes sorted by criticality as the identified node sequence. Then, through hierarchical reinforcement learning, determine the optimal node restoration order. The numbers marked on the node side in the figure are the restoration numbers. The results can verify that the present invention can quickly and accurately identify the damaged key nodes in the network, effectively determine the node restoration order, and optimize the network performance.

[0150] To implement the methods of Embodiment 1 to Embodiment 2 above, the key node identification device for the power system network in this embodiment includes:

[0151] A model training module is used to build a DQN model to learn the actions of selecting the best nodes to be removed under different network states. The network states include the normal connected state of the network and the network states under different disaster types. The input features of the model are the state representations of the power system network, and the output of the model is a sequence of nodes arranged in the removal order and used as the identified damaged critical nodes. The actions of the agent in the model are to remove the nodes and the corresponding layer nodes in the network. The DQN model is trained using a training data set containing power system network information under different network states to generate a damaged critical node identification model. During the training process, a reward function is constructed using the node removal impact metrics. The node removal impact metrics include the impact ratio of the network path, the affected node ratio, and the connectivity value of the affected path after removing the nodes and the layer nodes. The layer nodes are the nodes within a specified distance centered on the target node;

[0152] An identification module is used to obtain the operation data of the target power system network, construct a network topology structure and a distance matrix according to the operation data of the target power system network. Each element in the distance matrix corresponds to the shortest path distance between nodes in the network. The operation data, topology structure, and distance matrix of the target power system network are input into the trained damaged critical node identification model to obtain the identification result of the damaged critical nodes in the current target power system network.

[0153] Furthermore, a power grid information collection and analysis module can be constructed to collect and analyze the operation data of the power grid, identify key information such as nodes, edges, and transmission capabilities in the power grid to analyze the normal operation state of the power grid; and a power grid topology structure construction module to construct the topology structure of the power grid by analyzing the edges and nodes of the power grid, and construct the adjacency matrix of the power grid using methods such as complex network analysis; and a power grid distance matrix construction module to calculate the shortest path distances between each pair of nodes in the power grid, and then calculate the distance matrix of the shortest paths between nodes.

[0154] The power system network critical node identification device in this embodiment corresponds one-to-one with the above power system network critical node identification method, and will not be elaborated here one by one.

[0155] Furthermore, it also includes a power system network disaster simulation and recovery optimization device, including:

[0156] The power system network critical node identification device as described above;

[0157] A disaster simulation module is used to randomly select several nodes and the corresponding layer nodes from the identified damaged critical nodes for interruption to simulate different disaster scenarios;

[0158] A data acquisition module, configured to acquire the network state of the target power system network in the current disaster scenario simulation state and the number of connected subgraphs.

[0159] A node recovery module, configured to select nodes for recovery according to the network state and the number of connected subgraphs in the current disaster scenario simulation state, and enable backup lines between the selected nodes and the remaining network nodes, and preferentially recover damaged critical nodes in the network. Wherein, if the network is in a disconnected state, identify the nodes that can connect each connected subgraph fastest and use them as target repair nodes; preferentially repair the target repair nodes until the network forms a single connected subgraph so that the network reaches a connected state; when the network reaches a connected state, sequentially select target nodes that optimize the network performance for recovery and enable the backup lines of the target nodes to obtain the optimal order of repair nodes.

[0160] The power system network disaster simulation recovery optimization device in this embodiment corresponds one-to-one with the above-mentioned power system network disaster simulation recovery optimization method, and will not be elaborated here one by one.

[0161] This embodiment further provides a computer device, including a processor and a memory. The memory is used to store a computer program, and the processor is used to execute the computer program to execute the above-mentioned method.

[0162] It can be understood that the above-mentioned method in this embodiment can be executed by a single device, such as a computer or a server, etc., or can also be applied to a distributed scenario where multiple devices cooperate with each other to complete. In the case of a distributed scenario, one of the multiple devices can only execute one or more steps of the above-mentioned method in this embodiment, and the multiple devices interact with each other to complete the above-mentioned method. The processor can be implemented in the form of a general-purpose CPU, a microprocessor, an application-specific integrated circuit, or one or more integrated circuits, etc., and is used to execute relevant programs to implement the above-mentioned method in this embodiment. The memory can be implemented in the form of a read-only memory ROM, a random access memory RAM, a static storage device, and a dynamic storage device, etc. The memory can store an operating system and other application programs. When implementing the above-mentioned method in this embodiment through software or firmware, the relevant program codes are stored in the memory and called by the processor for execution.

[0163] Those skilled in the art should understand that the above embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be realized by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for realizing the functions specified in the process Figure 1 one process or multiple processes and / or blocks Figure 1 These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means that realizes the functions specified in the process Figure 1 one process or multiple processes and / or blocks Figure 1 These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for realizing the functions specified in the process Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0164] The above are only the preferred embodiments of the present invention and do not impose any formal limitations on the present invention. Although the present invention has been disclosed above with the preferred embodiments, it is not intended to limit the present invention. Therefore, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention shall fall within the scope of the protection of the technical solution of the present invention.

Claims

1. A method for identifying key nodes in a power system network, characterized in that the steps Including: Construct a DQN model to learn the action of selecting the best node to be removed in different network states. The network states include the normal connected state of the network and the network states under different disaster types. The input features of the model are the state representations of the power system network, and the output of the model is a sequence of nodes arranged in the removal order and used as the identified damaged critical nodes. The action of the agent in the model is to remove the nodes in the network and the corresponding layer nodes, where the layer nodes are the nodes within a specified distance centered on the target node; Train the DQN model using a training dataset composed of power system network information under different network states to generate a damaged critical node identification model. During the training process, construct a reward function using the removed node impact metrics, where the removed node impact metrics include the impact ratio of the network path, the affected node ratio, and the connectivity value of the affected path after removing the node and its surrounding nodes. Among them, the affected node ratio after removing the node and its surrounding nodes is calculated by the following expression: wherein represents the number of nodes in the layer represents the shortest distance centered on the removed node and represents the set of layer nodes formed by the nodes within this distance Obtain the operation data of the target power system network, and construct a network topology structure and a distance matrix according to the operation data of the target power system network. Each element in the distance matrix corresponds to the shortest path distance between nodes in the network; Input the network topology structure and the distance matrix of the target power system network into the trained damaged critical node identification model to obtain the identification result of the damaged critical nodes in the current target power system network.

2. The method for identifying key nodes in a power system network according to claim 1, wherein The constructing the network topology structure and the distance matrix according to the operation data of the target power system network includes: Collect the operation data of the power system network. The operation data includes voltage levels, current capacities of connected edges, and the conditions of transmission lines; Identify the node and edge information in the power system network according to the collected operation data of the power system network, obtain the mutual relationship between the nodes and edges, determine the connection methods and paths between each node in the power grid, and construct a topology diagram of the power system network; Calculate the graph based on the node and edge information in the power system network adjacency matrix , where is the set of nodes in the power system network, is the set of edges in the power system network, represents the weight of the edge from node to node ; Obtain the shortest distance between any node pair according to the adjacency matrix to obtain the distance matrix.

3. The method for identifying key nodes in a power system network according to claim 2, characterized in that, Adjacency matrix Configure according to the following formula: Distance matrix Each element in it is determined according to the following formula: Among them, represents the execution of Dijkstra's algorithm with the adjacency matrix , node and node as the input, represents the shortest path distance from node to node .

4. The method for identifying key nodes in a power system network according to claim 1, characterized in that The calculation expression of the constructed reward function is: Among them, and respectively represent the weight parameters for measuring the proportion of influence paths and the proportion of affected nodes, represents the influence proportion of the network path after removing the node and the nodes in the circle layer, represents the proportion of affected nodes after removing the node and the nodes in the circle layer, represents the connectivity value of the influence after removing the node and the nodes in the circle layer; Influence ratio of network path after removing nodes and layer nodes The calculation expression is as follows: Among them, represents the number of nodes reached by the th node in the remaining network after removing the node and the layer nodes, represents the number of reachable node pairs in the power system network, represents the total number of nodes in the network, represents the node set of the remaining network after removing the node and the layer nodes; Connectivity value affected after removing nodes and layer nodes The calculation expression is as follows: Among them, represents the maximum connected subgraph after removing the node and the corresponding layer nodes, the number of nodes of represents the set of nodes to be removed, represents the number of nodes of the maximum connected subgraph of the power system network of.

5. The method for identifying key nodes in a power system network according to any one of claims 1 to 4, characterized in that, Training the DQN model using a training data set composed of power system network information in different states includes: Using the training data set composed of power system network information in different states as the input of the DQN model; During the training process, at each step, all possible removal actions are evaluated based on the current network state, that is, the nodes to be removed and the corresponding layer nodes. Among them, by first selecting a node as the current removal node, the reward is predicted by the DQN model according to the reward function, and the node with the maximum reward is found . At this time, the set of removed nodes . Then, select the next node as the current removal node, determine the node with the maximum reward . At this time, the set of removed nodes . And so on, using a step-by-step expansion strategy, at the th time, by predicting the node with the maximum reward , a set of node removal sequences is found. represents the number of removed nodes; Determine the set of removed nodes whether it contains all the nodes of the original network, that is whether it is equal to the set of nodes in the power system network , if so, it is determined that the entire network has been completely damaged, the training stops, and a sequence of nodes arranged in the removal order is obtained as the identified damaged critical nodes, otherwise the model continues to be trained. The set of removed nodes is the node set and the corresponding set of nodes in the circle is the union of them.

6. A method for simulating, restoring and optimizing power system network disasters, characterized in that the steps Including: Identify the set of damaged critical nodes in the target power system network according to the method described in any one of claims 1 to 5 , indicating the scope of influence of the damaged nodes; From the identified set of damaged critical nodes Randomly select a set of nodes And the corresponding layer nodes Interrupt to simulate different disaster scenarios, and the current set of damaged nodes is , Is the set of nodes And the corresponding set of layer nodes The union of, then the set of critical nodes to be repaired first is ; Obtain the network state and the number of connected subgraphs of the target power system network in the current disaster scenario simulation state; Select nodes for restoration according to the network state and the number of connected subgraphs in the current disaster scenario simulation state, and enable the backup lines between the selected nodes and the remaining network nodes. Prioritize the restoration of the damaged critical nodes in the network. Among them, if the network is in a disconnected state, identify the nodes that can connect each connected subgraph fastest and use them as the target repair nodes; prioritize the repair of the target repair nodes until the network forms a single connected subgraph and the network reaches a connected state; when the network reaches a connected state, sequentially select the target nodes that optimize the network performance for restoration and enable the backup lines of the target nodes to obtain the optimal repair node order.

7. The power system network disaster simulation recovery optimization method according to claim 6, characterized in that By constructing a hierarchical reinforcement learning model and using the hierarchical reinforcement learning model for node recovery, wherein in the hierarchical reinforcement learning model, the underlying network preferentially repairs the target repair nodes until the network forms a single connected subgraph to make the network reach a connected state. The state of the underlying network is defined as the number and size distribution of the remaining network connected subgraphs, and the action is defined as selecting a node for repair and enabling the backup line of the node to be repaired; in the hierarchical reinforcement learning model, the high-level network is used to sequentially select the target nodes that optimize the network capabilities for recovery. The state of the high-level network is defined as the current remaining network, the set of nodes to be repaired, and the connectivity metric of the entire network, and the action is defined as selecting a node that improves the network performance for repair and enabling the backup line of the selected node.

8. The power system network disaster simulation recovery optimization method according to claim 7, characterized in that The reward function of the underlying network is defined as , represents the size of the largest connected subgraph of the network after repairing the node, represents the number of connected subgraphs of the network after repairing the node, and represent the weight parameters for measuring the size of the largest connected subgraph of the network and the number of connected subgraphs of the network respectively, represents the first repaired node The average clustering coefficient of the network after that, is the weight for measuring the priority of repairing critical nodes, is the first indicator function, which is used to take the value of 1 when the first repaired node belongs to the set of critical nodes to be repaired preferentially and 0 otherwise; The reward function of the high-level network is defined as , denotes the average clustering coefficient of the network after repairing node . denotes the connectivity metric of the network determined according to the number of nodes in the largest connected subgraph of the remaining network except the second repair node . denotes the remaining network except the second repair node . and denotes the weight parameter is the second indicator function, which takes the value of 1 when the second repair node belongs to the set of priority repair critical nodes , and 0 otherwise.

9. The power system network disaster simulation recovery optimization method according to claim 7 or 8, characterized in that Using the hierarchical reinforcement learning model for node recovery includes: Establish a low-level state table and a high-level state table for correspondingly storing the expected return values of each state-action pair of the low-level network and the high-level network; Obtaining the set of removed nodes and the remaining network data in the power system network under different disaster types and inputting them into the hierarchical reinforcement learning model; In the underlying network of the hierarchical reinforcement learning model, at each state according to the underlying state table and the exploration rate select an action with probability select the action with the largest value in the underlying state table and with probability randomly select an action, execute the selected action and then obtain a reward according to the reward function and the next state and update the underlying state table according to where , represents the discount factor; repeat the execution until the number of connected subgraphs of the network becomes 1 after repairing the node, and then transfer to the high-level network for node recovery;​ When performing node recovery in a high-level network, at each state select an action according to the high-level state table and the exploration rate with probability select the action with the largest value in the low-level state table and with probability randomly select an action, execute the selected action and then obtain a reward and the next state according to the reward function and update the low-level state table Repeat the execution until a preset iteration stop condition is reached, and determine the optimal low-level state table and the high-level state table to obtain a hierarchical reinforcement learning model with optimal parameters; Obtaining the set of removed nodes and the remaining network data in the current power system network and inputting them into the hierarchical reinforcement learning model with optimal parameters for node recovery.

10. A computer device, comprising a processor and a memory, wherein the memory is used for storing a computer program, characterized in that, The processor is used to execute the computer program to execute the method according to any one of claims 1 to 9.

Images