Power grid topology toughness enhancement method based on pre-disaster prediction of graph neural network

Abstract

The invention discloses a power grid topology toughness enhancement method based on pre-disaster prediction of a graph neural network. The method comprises the following steps: constructing a power grid graph structure taking power grid equipment as nodes; predicting a node damage probability through a graph attention network, and defining a high-risk node set; according to the high-risk nodes and the propagation paths thereof, a mixed integer programming model is constructed and solved, and a pre-disaster optimal power grid structure adjustment strategy is generated; in combination with the strategy, a resource scheduling scheme is generated by using a deep reinforcement learning algorithm; calculating a power grid toughness core index after simulation operation, and if the index is lower than a preset threshold value, optimizing a structure adjustment strategy; and uploading the optimized strategy and scheduling scheme to a scheduling platform to complete pre-disaster active defense deployment. According to the method, collaborative linkage of pre-disaster risk prediction, topology reconstruction and resource deployment is realized, a differentiated pre-disaster defense scheme is automatically generated, post-disaster first-aid repair is converted into pre-disaster deployment, and the pre-disaster prevention and control response efficiency is remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power grid disaster prevention and mitigation and intelligent operation and inspection, in particular to a power grid topology resilience enhancement method based on pre-disaster prediction of graph neural networks. BACKGROUND

[0002] Under the background of the prior art, the power grid disaster risk assessment mainly stays in the mode of "risk prediction + post-disaster response", which has many shortcomings: Risk prediction stays in passive response after the event; Existing researches mainly focus on training prediction models using meteorological, geological and historical fault data to output fault probability or risk level. However, these results are only used as a reference for post-disaster emergency response, and cannot form pre-disaster intervention means, making it difficult to break the cascade link of "typhoon → landslide → tower collapse → line trip".

[0003] The power grid topology structure is static and cannot be adjusted; Traditional power grid operation is based on static topology. Even if high-risk nodes and critical paths are identified before a disaster, there is still a lack of structural reconstruction capability to perform "disconnect-reserve-switch" before the disaster, so that the disaster impact is directly transmitted to the main network, causing large-scale power outages.

[0004] Emergency resource scheduling is lagging behind; The existing system mainly deploys emergency teams and resources according to the damage after the disaster, often missing the "golden defense window". This "post-disaster remedy" mode makes it difficult for repair teams to arrive in time, and mobile energy storage devices cannot be deployed in advance, delaying the recovery of power grid resilience.

[0005] The application range of graph neural networks is limited; In recent years, graph neural networks (GNN) have been applied in power grid risk prediction, which can model the topology structure and disaster propagation path between nodes. However, existing methods are mostly used for "disaster probability prediction" or "cascade failure simulation", and lack of linkage design with power grid topology reconstruction and resource deployment, failing to answer the question of "how to actively enhance resilience".

[0006] Lack of pre-disaster defense paradigm; In general, the current technology still stays in the "prediction-response" chain, lacking a new paradigm of "pre-disaster intervention-active deployment". It is urgently needed to use intelligent methods such as graph neural networks to identify high-risk nodes, reconstruct topology structure and optimize resource pre-deployment in coordination before the disaster, so as to improve the resilience of power grid in advance. SUMMARY

[0007] In view of the shortcomings of the prior art, the present application provides a power grid topology resilience enhancement method based on pre-disaster prediction of graph neural networks, which aims to solve the problems in the background art.

[0008] To achieve the above object, the present application provides the following technical scheme: a power grid topology resilience enhancement method based on pre-disaster prediction of graph neural network, comprising the following steps: Step S1: establishing a power grid graph structure taking power grid equipment as nodes; predicting the damage probability of each node based on the graph attention network, and defining a high-risk node set based on the damage probability; Step S2: based on the high-risk node set and its propagation path obtained in step S1, a mixed integer programming model is constructed and solved to generate a pre-disaster optimal power grid structure adjustment strategy; Step S3: based on the pre-disaster optimal power grid structure adjustment strategy of step S2, a resource scheduling scheme is generated using a deep reinforcement learning algorithm; Step S4: based on the pre-disaster optimal power grid structure adjustment strategy output by step S2 and the resource scheduling scheme output by step S3, a simulation run is performed, and the core indicators measuring the resilience of the power grid are calculated within the simulation run period. Compare the core indicators with the preset resilience threshold. If it is less than the resilience threshold, optimize the optimal power grid structure adjustment strategy , otherwise it remains unchanged; Step S5: upload the pre-disaster optimal power grid structure adjustment strategy and the resource scheduling scheme output by step S4 to the dispatching platform to complete the active defense deployment before the disaster occurs.

[0009] Further, the specific process of step S1 is: Step S1.1: fuse multi-source heterogeneous data including meteorological data, geological data, power grid equipment attributes, and historical fault data; take the power grid equipment as the nodes in the graph structure; take the relationship between the power grid equipment as the edges in the graph structure, and represent the graph structure in the form of a three-value adjacency matrix; Step S1.2: construct a graph attention network and initialize the network parameters; Step S1.3: input the graph structure into the graph attention network for node-level disaster risk probability prediction; the feature vector of each node in the first layer is: ; In the formula, denotes the feature vector of node in the first layer; denotes the attention weight between node and node ; denotes a learnable parameter matrix; denotes the neighbor node set of node ; represents the feature vector of the node in the th layer; After the propagation through multiple layers, the impaired probability of each node is output: ; wherein, represents the transpose of the learnable weight vector of the output layer; represents the feature vector of the node in the th layer; According to the impaired probability of each node and a preset threshold , a high-risk node set is defined: .

[0010] Further, the specific process of step S2 is as follows: Step S2.1: defining the decision variables of the mixed integer programming model: represents the power supply state of the node ; represents the power supply of the reserved node ; represents the power supply of the disconnected node ; represents the operation state of the path ; represents the activated path between nodes ; represents the inactivated path between nodes ; Step S2.2: establishing the objective function of the mixed integer programming model: ; wherein, represents the load size borne by the node ; represents the operation time required by the activated path ; represents the set of paths; represents the dynamic weight factor; Step S2.3: setting the constraint conditions, including the power grid connectivity constraint, the radial operation constraint, the power balance constraint, and the equipment capacity constraint; Step S2.4: solving the mixed integer programming model to obtain the optimal decision variable, i.e., the optimal power grid structure adjustment strategy before the disaster , represents the node​​​ The optimal power supply state, Indicates the path Optimal operational status.

[0011] Furthermore, deep reinforcement learning algorithm is used for resource scheduling, which is expressed as: ; Where, Represents the state-action value function, which represents when the power grid is in state When the scheduling center selects and executes a resource scheduling action After , the expected sum of all future rewards obtained; express The disaster graph state at the moment is used to describe the dynamic evolution characteristics of disasters in the power grid topology. Indicates the state of the disaster map, express The solidified disaster map, Represents resource state variables, Indicates the damage status, represented by Calculated; express Resource scheduling decisions at every moment; represents the learning rate; Indicates execution After that, the feedback benefit obtained at the current moment; represents the discount factor; express The state of the hazard map at the moment; Indicates the candidate actions for the next moment; Indicates Disaster map status at the moment Next, candidate actions Corresponding The maximum value in .

[0012] Furthermore, in step S4, the core indicators for measuring grid resilience are calculated : ; Where, Indicates The total amount of load in the power grid that has not yet been restored at the moment; Indicates the total load of the power grid under normal conditions; represents the simulation cycle; Represents integral.

[0013] Furthermore, in step S4, if the value is less than the resilience threshold, the optimal grid structure adjustment strategy is implemented. The specific process of optimization is: <resilience threshold , then calculate the core indicator by causal inference Attention weights of graph attention network Gradient: , denotes the partial derivative symbol; the attention weights are corrected in reverse according to the calculated gradient .

[0014] Further, the graph attention network is trained using a cross-entropy loss function to minimize the prediction error; ; wherein, denotes the true label of the node; denotes the set of all nodes in the graph structure.

[0015] An electronic device comprising a processor, a memory and a bus, the processor and the memory being connected through the bus, wherein the memory is used to store a set of program codes, and the processor is used to invoke the program codes stored in the memory to execute a power grid topology resilience enhancement method based on graph neural network pre-disaster prediction.

[0016] A non-volatile computer storage medium, the computer storage medium storing computer executable instructions, the computer executable instructions executing a power grid topology resilience enhancement method based on graph neural network pre-disaster prediction.

[0017] Compared with the prior art, the present application has the following beneficial effects:

[0018] (1) In the dynamic risk environment of multiple disaster superposition, the present application can model the disaster coupling effect and cascading propagation path in real time through the graph neural network, accurately identify the high-risk nodes and key links, and realize the pre-disaster global risk situation awareness and device-level resilience evaluation.

[0019] (2) The present application can realize the coordinated linkage of pre-disaster risk prediction results and topology reconstruction, resource deployment, and introduce the core indicator of power grid resilience as a unified evaluation index. This index is used not only to quantitatively evaluate the simulation results in step S4 to determine whether the best power grid structure adjustment strategy before the disaster needs further optimization, but also to compare and select among multiple candidate schemes, thereby automatically generating differentiated pre-disaster defense schemes. Through this mechanism, the traditional "post-disaster repair" can be transformed into "pre-disaster defense", significantly improving the response efficiency of pre-disaster prevention and control of the power grid.

[0020] (3) The application considers the region data scarcity and model migration difficulty, continuously verifies and optimizes the reconstruction strategy through digital twin counter simulation mechanism, guarantees the strategy rapid adaptation across regions without large-scale target domain data, ensures early deployment and safe operation of high-risk regions. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 The method flowchart of the application. DETAILED DESCRIPTION

[0022] As Figure 1 shown, the application provides a technical solution: a power grid topology resilience enhancement method based on graph neural network pre-disaster prediction, including the following steps:

[0023] Step S1: establish a power grid graph structure with power grid equipment as nodes; predict the damage probability of each node based on the graph attention network, and define a high-risk node set based on the damage probability.

[0024] Step S1.1: fuse multi-source heterogeneous data (meteorological, geological, power grid equipment attributes, historical fault data, etc.); abstract power grid equipment as nodes in the graph structure (node attributes include device type, exposure, disaster resistance, etc.); abstract the relationship between power grid equipment as edges (such as physical connection, redundancy relationship, geographical proximity), and use a three-value adjacency matrix form to represent the graph structure.

[0025] Step S1.2: construct a graph attention network (GAT) and initialize network parameters (learnable weight matrix and attention mechanism parameters).

[0026] Step S1.3: input the graph structure into the graph attention network (GAT) for node-level disaster risk probability prediction; each node In the first layer, the feature vector of node is: In the formula, represents the feature vector of node in the first layer; represents the attention weight between node and node ; represents the learnable parameter matrix; represents the neighbor node set of node ; represents the feature vector of node in the first layer.

[0027] After multi-layer propagation, the damage probability of each node is output : ; wherein, denotes the transpose of the learnable weight vector of the output layer; denotes the feature vector of the node in the layer (the last layer).

[0028] wherein, the graph attention network (GAT) is trained using a cross-entropy loss function to minimize the prediction error; ; wherein, denotes the real label of the node; denotes the set of all nodes in the graph structure.

[0029] According to the damage probability of each node and a preset threshold , a high-risk node set is defined: .

[0030] This step abstracts the power grid equipment as nodes and the spatial adjacency relationship as edges by fusing multi-source heterogeneous data such as meteorology, geology, power grid equipment, and historical failures, and constructs a unified spatio-temporal graph model, and gives the nodes multi-dimensional attributes such as equipment type, exposure, and disaster resistance. The role is to use the graph attention network (GAT) to model the nonlinear cascading propagation of multi-disaster coupling, accurately output the pre-disaster damage probability distribution of each node, so as to provide quantifiable risk basis for subsequent "who to cut, who to protect, and how to reconstruct", and avoid the misjudgment brought by previous experience or single index evaluation.

[0031] Step S2: Based on the high-risk node set and its propagation path obtained in step S1, a mixed integer programming model (MIP) is constructed and solved to generate a pre-disaster optimal power grid structure adjustment strategy.

[0032] wherein, the propagation path refers to: the link set of the high-risk node in the power grid graph structure, which transmits the failure risk to the adjacent nodes relying on the edge set (power transmission line). The propagation path describes the cascading failure link formed by the disaster through the power grid topology, which not only includes the damage probability of the high-risk node itself, but also reflects the potential influence range of the adjacent nodes and downstream loads.

[0033] By introducing the high-risk node set and its propagation path, the mixed integer programming model can identify both the "risk source node" and the "risk transmission channel", and realize the active interruption of disaster cascading propagation by disconnecting, switching or retaining part of the nodes and lines before the disaster, so as to generate a pre-disaster optimal power grid structure adjustment strategy.

[0034] Step S2.1: Define the decision variables of the mixed integer programming model: denote the power supply state of node ; denote the power supply of reserved node ; denote the power supply of disconnected node .

[0035] denote the operation state of path ; denote the activated path between nodes ; denote the inactivated path between nodes .

[0036] Step S2.2: Establish the objective function of the mixed integer programming model (MIP) aiming to minimize the load loss and recovery time simultaneously; ; wherein, denote the load size of node ; denote the operation time (such as switch switching time) required by the activated path ; denote the set of paths; denote the dynamic weight factor, , denote the power load of node , denote the preset base weight constant.

[0037] Step S2.3: Set the constraint conditions:

[0038] Power grid connectivity constraint: Ensure that the optimized network structure is connected and there is no island.

[0039] Radial operation constraint: The distribution network usually requires radial operation, i.e., no loop network.

[0040] Power balance constraint: Ensure that the power generated by the power source is equal to the power consumed by the load (including line loss).

[0041] Device capacity constraint: Ensure that the lines, transformers, etc. do not overload.

[0042] Step S2.4: Solve the mixed integer programming model (MIP) to obtain the optimal decision variable, i.e., the optimal pre-disaster power grid structure adjustment strategy , denote the node The optimal power supply state, Indicates the path Optimal operational status.

[0043] Step S3: Based on the optimal pre-disaster grid structure adjustment strategy in step S2, a deep reinforcement learning (DRL) algorithm is used to generate a resource scheduling plan (to achieve efficient blocking of cascading fault propagation and transform the traditional "post-disaster repair" into "pre-disaster defense"), which is expressed as: ; Where, Represents the state-action value function, which represents when the power grid is in state When the scheduling center selects and executes a resource scheduling action The expected sum of all future rewards that can be obtained after accumulating (i.e., long-term return); express The disaster graph state at a certain moment (a joint representation of the grid topology + node damage probability + failed grid equipment + resource distribution at a certain moment) is used to describe the dynamic evolution characteristics of disasters in the grid topology. Indicates the state of the disaster map, express The solidified disaster map, Represents resource state variables (such as resource reserves and distribution), Indicates the damage status, represented by Calculated; express Resource scheduling decisions at every moment; represents the learning rate; Indicates execution After that, the feedback benefit obtained at the current moment; represents the discount factor; express The state of the hazard map at the moment; Indicates the candidate actions for the next moment; Indicates Disaster map status at the moment Next, candidate actions Corresponding The maximum value in .

[0044] This step, based on the determined pre-disaster topology, employs deep reinforcement learning (e.g., graph policy networks (GPNs)) based on graph embedding to pre-deploy emergency resources such as repair teams, mobile energy storage vehicles, and temporary ring main units to "risk blocking points." This utilizes the trial-and-error and reward mechanisms of reinforcement learning to find the optimal spatial and quantitative configuration to block the propagation of cascading faults before a disaster strikes. This maximizes resource utilization, shortens post-disaster recovery time, and transforms traditional "post-disaster repair" into "pre-disaster defense."

[0045] Step S4: Simulation running based on the pre-disaster optimal power grid structure adjustment strategy output in step S2 and the resource scheduling scheme output in step S3, calculating the core index measuring the resilience of the power grid in the simulation running period, comparing the core index with the preset resilience threshold, and if it is less than the resilience threshold, optimizing the optimal power grid structure adjustment strategy, otherwise, keeping it unchanged.

[0046] Wherein, the core index measuring the resilience of the power grid is calculated : ; In the formula, represents the total amount of loads that have not been restored in the power grid at the time t; represents the total load of the power grid in the normal state; represents the simulation period; represents the integral.

[0047] Wherein, the specific process of optimizing the optimal power grid structure adjustment strategy if it is less than the resilience threshold is as follows: If < resilience threshold , the gradient of the attention weight of the graph attention network (GAT) is calculated through causal inference: , represents the partial derivative symbol; the attention weight is reversely corrected according to the calculated gradient , forming a rolling optimization closed loop. This step injects the topology reconstruction scheme and resource deployment scheme generated in steps S2-S3 into the power grid digital twin, simulates the evolution process of typhoon, thunderstorm and other multi-disaster species through adversarial simulation, calculates the resilience index R in real time, and identifies the key control variables with the help of the causal inference model. Its role is to verify the effectiveness of the strategy at low cost and high concurrency in the virtual environment, find potential vulnerabilities, reversely optimize the graph neural network parameters and reconstruction strategy, form a continuous self-evolution closed loop of “data-model-strategy”, and ensure the migratability and robustness of the strategy across regions and disaster species.

[0048] Step S5: uploading the pre-disaster optimal power grid structure adjustment strategy and the resource scheduling scheme output in step S4 to the dispatching platform (power grid dispatching system (EMS / SCADA)) to complete the active defense deployment before the disaster occurs.

[0049]

[0050] ​​​​​​​An electronic device comprises a processor, a memory and a bus, the processor and the memory are connected through the bus, wherein the memory is used to store a set of program codes, and the processor is used to call the program codes stored in the memory to execute a power grid topology resilience enhancement method based on graph neural network disaster prediction.

[0051] A non-volatile computer storage medium, the computer storage medium stores computer executable instructions, the computer executable instructions execute a power grid topology resilience enhancement method based on graph neural network disaster prediction.

[0052] Although embodiments of the present application have been shown and described, it is to be understood that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for enhancing power grid topology resilience based on graph neural network pre-disaster prediction, characterized in that: The steps include: Step S1: Establish a power grid graph structure with power grid devices as nodes; predict the damage probability of each node based on the graph attention network, and define a high-risk node set based on the damage probability; Step S2: Based on the high-risk node set and its propagation path obtained in step S1, a mixed integer programming model is constructed and solved to generate the optimal pre-disaster grid structure adjustment strategy; Step S3: Based on the optimal grid structure adjustment strategy before the disaster in step S2, a resource scheduling plan is generated using a deep reinforcement learning algorithm; Step S4: Optimal pre-disaster grid structure adjustment strategy based on the output of step S2 The resource scheduling scheme outputted in step S3 is simulated and run, and the core indicators for measuring grid resilience are calculated during the simulation operation cycle. The core indicators are compared with the preset resilience threshold. If the core indicators are less than the resilience threshold, the optimal grid structure adjustment strategy is implemented. Optimize, otherwise remain unchanged; Step S5: Adjust the optimal grid structure before the disaster outputted in step S4 The resource scheduling plan is uploaded to the scheduling platform to complete active defense deployment before a disaster occurs.

2. The method for enhancing power grid topology resilience based on graph neural network pre-disaster prediction according to claim 1 is characterized by: The specific process of step S1 is: Step S1.1: Integrate multi-source heterogeneous data including meteorological, geological, power grid equipment attributes, and historical fault data; treat power grid equipment as nodes in a graph structure; treat relationships between power grid equipment as edges in the graph structure, and represent the graph structure using a ternary adjacency matrix; Step S1.2: Construct the graph attention network and initialize the network parameters; Step S1.3: Input the graph structure into the graph attention network to predict the node-level disaster risk probability; each node In the The feature vector of the layer is: ； Where, Indicates the Nodes in the layer The eigenvector of Representation node and nodes The attention weight between represents the learnable parameter matrix; Representation node The set of neighbor nodes of Indicates the Nodes in the layer The eigenvector of After multi-layer propagation, the damage probability of each node is output : ； Where, represents the transpose of the learnable weight vector of the output layer; Indicates the Nodes in the layer The eigenvector of According to the damage probability of each node and preset thresholds , define a set of high-risk nodes : 。 3. The method for enhancing power grid topology resilience based on graph neural network pre-disaster prediction according to claim 2 is characterized by: The specific process of step S2 is: Step S2.1: Define the decision variables for the mixed integer programming model: Representation node Power supply status; Retained nodes Power supply; Indicates a disconnected node Power supply; Indicates the path Operational status; Represents the path between activated nodes ; Indicates that the path between nodes is not activated ; Step S2.2: Establish the objective function of the mixed integer programming model: ； Where, Representation node The size of the load carried; Indicates the activation path Required operating time; Represents a collection of paths; represents the dynamic weight factor; Step S2.3: Setting constraints, including grid connectivity constraints, radial operation constraints, power balance constraints, and equipment capacity constraints; Step S2.4: Solve the mixed integer programming model to obtain the optimal decision variables, that is, the optimal grid structure adjustment strategy before the disaster. , Representation node The optimal power supply state, Indicates the path Optimal operational status.

4. The method for enhancing power grid topology resilience based on graph neural network pre-disaster prediction according to claim 3 is characterized by: Using deep reinforcement learning algorithm for resource scheduling, it can be expressed as: ； Where, Represents the state-action value function, which represents when the power grid is in state When the scheduling center selects and executes a resource scheduling action After , the expected sum of all future rewards obtained; express The disaster graph state at the moment is used to describe the dynamic evolution characteristics of disasters in the power grid topology. Indicates the state of the disaster map, express The solidified disaster map, Represents resource state variables, Indicates the damage status, represented by Calculated; express Resource scheduling decisions at every moment; represents the learning rate; Indicates execution After that, the feedback benefit obtained at the current moment; represents the discount factor; express The state of the hazard map at the moment; Indicates the candidate actions for the next moment; Indicates Disaster map status at the moment Next, candidate actions Corresponding The maximum value in .

5. The method for enhancing power grid topology resilience based on graph neural network pre-disaster prediction according to claim 4 is characterized by: In step S4, the core indicators for measuring grid resilience are calculated : ； Where, Indicates The total amount of load in the power grid that has not yet been restored at the moment; Indicates the total load of the power grid under normal conditions; represents the simulation cycle; Represents integral.

6. The method for enhancing power grid topology resilience based on graph neural network pre-disaster prediction according to claim 5 is characterized by: In step S4, if the value is less than the resilience threshold, the optimal grid structure adjustment strategy is implemented. The specific process of optimization is as follows: <Toughness threshold , then calculate the core indicators through causal inference Attention weights for graph attention networks Gradient: , Indicates the symbol for partial derivative; the attention weight is reversed according to the calculated gradient .

7. The method for enhancing power grid topology resilience based on graph neural network pre-disaster prediction according to claim 6, characterized in that: Using the cross entropy loss function Train the graph attention network to minimize prediction error; ； Where, represents the true label of the node; Represents the set of all nodes in the graph structure.

8. An electronic device, characterized in that: It includes a processor, a memory and a bus, wherein the processor and the memory are connected via the bus, wherein the memory is used to store a set of program codes, and the processor is used to call the program codes stored in the memory to execute a method for enhancing the topology resilience of a power grid based on pre-disaster prediction of a graph neural network as described in any one of claims 1 to 7.

9. A non-volatile computer storage medium storing computer-executable instructions, characterized in that: The computer executable instructions execute a method for enhancing power grid topology resilience based on pre-disaster prediction using a graph neural network as described in any one of claims 1-7.

Images