Evaluation methods and systems for the effectiveness of artificial lightning suppression intervention

By training a graph neural network model and optimizing the loss function in conjunction with physical constraints, the problem of scarce tower-level annotation data was solved, and the effectiveness of manual lightning suppression intervention was accurately evaluated.

CN123115885APending Publication Date: 2026-11-17CHINA ELECTRIC POWER RESEARCH INSTITUTE CO LTD +3
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611018689.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-09
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies cannot accurately assess the effectiveness of manual lightning suppression intervention, mainly due to the scarcity of tower-level labeled data, making it impossible to effectively train purely data-driven graph neural networks.

Method used

By training a graph neural network model and utilizing graph-structured data, combined with standard supervised loss function, physical consistency constraints, and physical equation residual constraints, the total loss function is optimized to evaluate the effectiveness of manual lightning suppression intervention.

Benefits of technology

In situations where labeled data is scarce, graph neural network models can accurately predict the reduction rate of lightning strike risk at each tower node, enabling a refined assessment of the effectiveness of artificial lightning suppression intervention at the tower level.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115885A_ABST
    Figure CN123115885A_ABST
Patent Text Reader

Abstract

This invention provides a method and system for evaluating the effectiveness of artificial lightning suppression intervention, comprising: acquiring graph structure data of the target transmission line and associated power grid equipment in the power grid system before and after artificial lightning suppression operations; training a graph neural network model using the graph structure data to predict the reduction rate of lightning fault risk for each tower node in the graph structure data, and optimizing it using a total loss function until the total loss function is minimized to obtain a trained graph neural network model; the total loss function includes a standard supervised loss function, physical consistency constraints, and physical equation residual constraints; inputting the graph structure data into the trained graph neural network model for prediction to obtain the predicted value of the reduction rate of lightning fault risk for each tower node in the graph structure data; and evaluating the effectiveness of artificial lightning suppression intervention based on the predicted value of the reduction rate of lightning fault risk for each tower node; this invention can accurately evaluate the effectiveness of tower-level artificial lightning suppression intervention.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of power grid disaster prevention and mitigation and weather modification technology, specifically to a method and system for evaluating the effectiveness of artificial lightning suppression intervention. Background Technology

[0002] Artificial lightning suppression typically refers to an active disaster prevention technology that suppresses or weakens lightning activity by seeding high-performance suppression materials (such as conductive particles, ice nuclei, etc.) into strong convective thunderstorm clouds, thereby altering the cloud's charge structure and electrification process. Evaluating the effectiveness of artificial lightning suppression operations requires answering a key question: by how much has manual intervention reduced the risk of lightning strikes on each tower? Graph neural networks, due to their inherent graph-based data processing capabilities, are theoretically well-suited for modeling power grid topology as a graph and achieving refined evaluation at the tower level through neighbor aggregation. However, in practical applications, purely data-driven graph neural networks face a fundamental obstacle in the field of artificial lightning suppression: a lack of data for training. Specifically, tower-level labeled data is extremely scarce, making it impossible to effectively train purely data-driven models.

[0003] Therefore, designing a method to accurately evaluate the effectiveness of artificial lightning suppression intervention is an important problem that urgently needs to be solved. Summary of the Invention

[0004] To address the problem that existing technologies cannot accurately evaluate the effectiveness of artificial lightning suppression intervention, this invention proposes a method for evaluating the effectiveness of artificial lightning suppression intervention. The steps include: conducting artificial lightning suppression operations during thunderstorms and acquiring graph structure data of the target transmission lines and associated power grid equipment in the power grid system before and after the artificial lightning suppression operations; based on the graph structure data, training a graph neural network model to predict the reduction rate of lightning fault risk for each tower node in the graph structure data, and optimizing it using a total loss function until the total loss function is minimized to obtain a trained graph neural network model; the total loss function includes a standard supervised loss function, physical consistency constraints, and physical equation residual constraints; based on the graph structure data, using the trained graph neural network model to predict the predicted value of the reduction rate of lightning fault risk for each tower node in the graph structure data; and evaluating the effectiveness of the artificial lightning suppression intervention based on the predicted value of the reduction rate of lightning fault risk for each tower node.

[0005] Optionally, the graph structure data includes multiple tower nodes, the electrical characteristics of each tower node, the physical connection edges and spatial proximity edges between every two tower nodes, and the true value of the lightning strike fault risk reduction rate corresponding to the marked tower nodes among the multiple tower nodes.

[0006] Optionally, the step of training a graph neural network model based on the graph structure data to predict the lightning strike risk reduction rate of each tower node in the graph structure data, and optimizing it using a total loss function until the total loss function is minimized to obtain a trained graph neural network model, specifically includes: using the graph neural network model to predict and obtain the predicted value of the lightning strike risk reduction rate corresponding to each tower node; calculating the standard supervised loss value using a standard supervised loss function based on multiple tower nodes, the predicted value of the lightning strike risk reduction rate corresponding to each tower node, and the actual value of the lightning strike risk reduction rate; and calculating the standard supervised loss value based on every two tower nodes and every two... For each tower node, the physical connection edges and spatial proximity edges are calculated using physical consistency constraints to obtain physical consistency constraint values. Based on multiple tower nodes, the predicted lightning fault risk reduction rate for each tower node, and the electrical characteristics of the towers, physical equation residual constraints are calculated to obtain physical equation residual constraint values. The standard supervised loss value, the physical consistency constraint value, and the physical equation residual constraint value are weighted and summed to obtain the total loss function value. Based on the total loss function value, the graph neural network model is adjusted through backpropagation until the total loss function value is minimized, thus obtaining the trained graph neural network model.

[0007] Optionally, after evaluating the effectiveness of artificial lightning suppression intervention based on the predicted reduction rate of lightning fault risk for each of the aforementioned tower nodes, the method further includes: using the graph structure data and a trained graph neural network model, performing multiple independent forward propagations on several unlabeled tower nodes among the multiple tower nodes using the Monte Carlo dropout method to obtain a predicted value for each unlabeled tower node; determining the prediction variance of each unlabeled tower node in the graph structure data based on the predicted values ​​of the several unlabeled tower nodes; sorting the prediction variances of multiple unlabeled tower nodes in the graph structure data from largest to smallest, selecting the top-ranked unlabeled tower nodes to obtain recommended labeled tower nodes for the graph structure data; performing actual measurements on the recommended labeled tower nodes in the graph structure data to obtain measured values ​​of the lightning fault risk reduction rate of the recommended labeled tower nodes in the graph structure data; updating the graph structure data using the measured values ​​of the lightning fault risk reduction rate of the recommended labeled tower nodes in the graph structure data; and using the updated graph structure data for the evaluation of the effectiveness of the next artificial lightning suppression intervention.

[0008] Optionally, the physical consistency constraint is calculated using the following formula:

[0009] in, This represents the physical consistency constraint value of the total loss function in the graph neural network model. For graph-structured data, it is the set of spatially nearest edges. Tower nodes predicted by a graph neural network model The rate of reduction in the risk of lightning strike failure; Tower nodes predicted by a graph neural network model The rate of reduction in the risk of lightning strike failure; This refers to the set of physical connection edges in graph-structured data. For smoothing coefficients;

[0010] in, For physical similarity weights; Tower nodes in graph structure data Line characteristics; Tower nodes in graph structure data Line characteristics; This is the Gaussian kernel bandwidth parameter.

[0011] Optionally, the residual constraint of the physical equation is calculated using the following formula:

[0012] in, The physical equation residual constraint value for the total loss function in the graph neural network model; Tower nodes predicted by a graph neural network model The rate of reduction in the risk of lightning strike failure; This refers to all tower nodes in the graph structure data;

[0013] in, Tower nodes in graph structure data The theoretical risk reduction rate; This refers to the ground flash density before artificial lightning suppression operations. This refers to the ground flash density after artificial lightning suppression operations. The distribution of lightning current amplitude before manual lightning suppression operations; The distribution of lightning current amplitude after artificial lightning suppression operations; Tower nodes in graph structure data The shielding failure probability function; This refers to the amplitude of the lightning current. This represents the lower limit of the lightning current withstand capability of the target transmission line in the graph structure data.

[0014] Optionally, the total loss function value is calculated using the following formula:

[0015] in, This represents the total loss function value of the graph neural network model. These are the weighting coefficients for physical consistency constraints. The value is inversely proportional to the number of tower nodes already marked in the graph structure data; These are the weighting coefficients for the residual constraints of the physical equations. The value is inversely proportional to the number of tower nodes already marked in the graph structure data; This is the standard supervised loss function for graph neural network models; Physical consistency constraints for the total loss function in a graph neural network model; The physical equation residual constraint value for the total loss function in the graph neural network model.

[0016] Optionally, the step of evaluating the effectiveness of manual lightning suppression intervention based on the predicted reduction rate of lightning fault risk for each of the tower nodes includes: based on each of the tower nodes: if the predicted reduction rate of lightning fault risk for the tower node is within a preset high-efficiency range, then the evaluation result corresponding to the predicted reduction rate of lightning fault risk for that tower node is determined to be that the manual intervention is effective; if the predicted reduction rate of lightning fault risk for the tower node is within a preset medium-efficiency range, then the evaluation result corresponding to the predicted reduction rate of lightning fault risk for that tower node is determined to be that the manual intervention is effective; if the predicted reduction rate of lightning fault risk for the tower node is within a preset low-efficiency range, then the evaluation result corresponding to the predicted reduction rate of lightning fault risk for that tower node is determined to be that the manual intervention is ineffective; if the predicted reduction rate of lightning fault risk for the tower node is within a preset ineffective range, then the evaluation result corresponding to the predicted reduction rate of lightning fault risk for that tower node is determined to be that the manual intervention is insignificant.

[0017] Furthermore, this application also provides an evaluation system for the effectiveness of artificial lightning suppression intervention, comprising: a data acquisition module, used to conduct artificial lightning suppression operations during thunderstorms and acquire graph structure data of the target transmission line and associated power grid equipment in the power grid system before and after the artificial lightning suppression operations; a loss optimization training module, used to train a graph neural network model based on the graph structure data to predict the reduction rate of lightning fault risk of each tower node in the graph structure data, and to optimize using a total loss function until the total loss function is minimized, thereby obtaining a trained graph neural network model; the total loss function includes a standard supervised loss function, physical consistency constraints, and physical equation residual constraints; a model prediction module, used to predict the predicted value of the reduction rate of lightning fault risk of each tower node in the graph structure data using the trained graph neural network model based on the graph structure data; and an intervention effect evaluation module, used to evaluate the effectiveness of artificial lightning suppression intervention based on the predicted value of the reduction rate of lightning fault risk of each tower node.

[0018] Optionally, the graph structure data includes multiple tower nodes, the electrical characteristics of each tower node, the physical connection edges and spatial proximity edges between every two tower nodes, and the true value of the lightning strike fault risk reduction rate corresponding to the marked tower nodes among the multiple tower nodes.

[0019] Optionally, the loss optimization training module is specifically used for: using a graph neural network model to predict and obtain the predicted value of the lightning fault risk reduction rate corresponding to each tower node; calculating the standard supervised loss value using a standard supervised loss function based on multiple tower nodes, the predicted value of the lightning fault risk reduction rate corresponding to each tower node, and the actual value of the lightning fault risk reduction rate; calculating the physical consistency constraint value using physical consistency constraints based on every two tower nodes, the physical connection edge and the spatial proximity edge corresponding to every two tower nodes; calculating the physical equation residual constraint value using physical equation residual constraints based on multiple tower nodes, the predicted value of the lightning fault risk reduction rate corresponding to each tower node, and the electrical related features of the towers; weighted summing the standard supervised loss value, the physical consistency constraint value, and the physical equation residual constraint value to obtain the total loss function value; and adjusting the graph neural network model through backpropagation based on the total loss function value until the total loss function value is minimized, thereby obtaining the trained graph neural network model.

[0020] Optionally, the system further includes: an active learning sample recommendation module; the active learning sample recommendation module is used to perform multiple independent forward propagations on several unlabeled tower nodes among multiple tower nodes using the Monte Carlo dropout method based on the graph structure data and the trained graph neural network model, to obtain the predicted value of each unlabeled tower node; based on the predicted values ​​of the several unlabeled tower nodes, determine the prediction variance of each unlabeled tower node in the graph structure data; sort the prediction variances of multiple unlabeled tower nodes in the graph structure data from largest to smallest, select the top-ranked unlabeled tower nodes to obtain the recommended labeled tower nodes of the graph structure data; perform actual measurements on the recommended labeled tower nodes of the graph structure data to obtain the measured value of the lightning strike fault risk reduction rate of the recommended labeled tower nodes in the graph structure data; update the graph structure data using the measured value of the lightning strike fault risk reduction rate of the recommended labeled tower nodes in the graph structure data; and use the updated graph structure data for the evaluation of the effect of the next manual lightning suppression intervention.

[0021] Optionally, the physical consistency constraint is calculated using the following formula:

[0022] in, This represents the physical consistency constraint value of the total loss function in the graph neural network model. For graph-structured data, it is the set of spatially nearest edges. Tower nodes predicted by a graph neural network model The rate of reduction in the risk of lightning strike failure; Tower nodes predicted by a graph neural network model The rate of reduction in the risk of lightning strike failure; This refers to the set of physical connection edges in graph-structured data. For smoothing coefficients;

[0023] in, For physical similarity weights; Tower nodes in graph structure data Line characteristics; Tower nodes in graph structure data Line characteristics; This is the Gaussian kernel bandwidth parameter.

[0024] Optionally, the residual constraint of the physical equation is calculated using the following formula:

[0025] in, The physical equation residual constraint value for the total loss function in the graph neural network model; Tower nodes predicted by a graph neural network model The rate of reduction in the risk of lightning strike failure; This refers to all tower nodes in the graph structure data;

[0026] in, Tower nodes in graph structure data The theoretical risk reduction rate; This refers to the ground flash density before artificial lightning suppression operations. This refers to the ground flash density after artificial lightning suppression operations. The distribution of lightning current amplitude before manual lightning suppression operations; The distribution of lightning current amplitude after artificial lightning suppression operations; Tower nodes in graph structure data The shielding failure probability function; This refers to the amplitude of the lightning current. This represents the lower limit of the lightning current withstand capability of the target transmission line in the graph structure data.

[0027] Optionally, the total loss function value is calculated using the following formula:

[0028] in, This represents the total loss function value of the graph neural network model. These are the weighting coefficients for physical consistency constraints. The value is inversely proportional to the number of tower nodes already marked in the graph structure data; These are the weighting coefficients for the residual constraints of the physical equations. The value is inversely proportional to the number of tower nodes already marked in the graph structure data; This is the standard supervised loss function for graph neural network models; Physical consistency constraints for the total loss function in a graph neural network model; The physical equation residual constraint value for the total loss function in the graph neural network model.

[0029] Optionally, the intervention effect evaluation module is specifically used for: based on each of the tower nodes: if the predicted value of the lightning strike risk reduction rate of the tower node is in a preset high-efficiency range, then the evaluation result corresponding to the predicted value of the lightning strike risk reduction rate of the tower node is determined to be that the artificial intervention effect is significant; if the predicted value of the lightning strike risk reduction rate of the tower node is in a preset medium-efficiency range, then the evaluation result corresponding to the predicted value of the lightning strike risk reduction rate of the tower node is determined to be that the artificial intervention effect is obvious; if the predicted value of the lightning strike risk reduction rate of the tower node is in a preset low-efficiency range, then the evaluation result corresponding to the predicted value of the lightning strike risk reduction rate of the tower node is determined to be that the artificial intervention effect is limited; if the predicted value of the lightning strike risk reduction rate of the tower node is in a preset ineffective range, then the evaluation result corresponding to the predicted value of the lightning strike risk reduction rate of the tower node is determined to be that the artificial intervention effect is not significant.

[0030] In another aspect, this application also provides a computing device, including: one or more processors; processors for executing one or more programs; when the one or more programs are executed by the one or more processors, implementing an evaluation method for the effect of artificial lightning suppression intervention as described above.

[0031] In another aspect, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed, implements the method for evaluating the effect of artificial lightning suppression intervention as described above.

[0032] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention provides a method and system for evaluating the effectiveness of artificial lightning suppression intervention. It trains a graph neural network model using graph-structured data to predict the reduction rate of lightning strike risk for each tower node in the graph-structured data. The model is then optimized using a total loss function, which includes a standard supervised loss function, physical consistency constraints, and physical equation residual constraints. The dual constraints of physical consistency and physical equation residual constraints prevent overfitting during model training when labeled data is scarce in the graph-structured data. This allows the graph neural network model to accurately predict the reduction rate of lightning strike risk for each tower node even with a small number of labeled samples, thus enabling accurate evaluation of the tower-level artificial lightning suppression intervention effectiveness based on the predicted reduction rate of lightning strike risk for each tower node. Attached Figure Description

[0033] Figure 1 This is a flowchart of the method for evaluating the effect of artificial lightning suppression intervention according to the present invention; Figure 2 This is a schematic diagram of the overall steps of the present invention; Figure 3 This is a schematic diagram of the power grid topology model structure for the evaluation method of the artificial lightning suppression intervention effect of the present invention; Figure 4 This is a schematic diagram of the graph neural network message passing and node embedding process for evaluating the effect of artificial lightning suppression intervention according to the present invention. Figure 5 This is a schematic diagram of the loss function structure with enhanced physical constraints in the evaluation method for the effect of artificial lightning suppression intervention of the present invention; Figure 6 This is a schematic diagram of the active learning sample recommendation process for the evaluation method of the artificial lightning suppression intervention effect of the present invention; Figure 7 This is a structural block diagram of the device for evaluating the effect of artificial lightning suppression intervention according to the present invention; Figure 8 This is a schematic diagram of the structure of the evaluation system for the effect of artificial lightning suppression intervention of the present invention; Figure 9 This is a schematic diagram of the structure of an electronic device provided by the present invention. Detailed Implementation

[0034] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0035] Example 1: A method for evaluating the effectiveness of artificial lightning suppression intervention, such as Figure 1 As shown, it includes the following steps: Step 1: Conduct artificial lightning suppression operations during thunderstorms and obtain graphical structure data of the target transmission lines and associated power grid equipment in the power grid system before and after the artificial lightning suppression operations.

[0036] The graph structure data includes multiple tower nodes, the electrical characteristics of each tower node, the physical connection edges and spatial proximity edges between every two tower nodes, and the true value of the lightning strike fault risk reduction rate for the labeled tower nodes among the multiple tower nodes.

[0037] The detailed explanation of step 1 is as follows: S101: Construct a power grid topology model.

[0038] Reference Figure 3 The target transmission line and associated power grid equipment are modeled as a graph. , where the node set The edge set includes all tower nodes of the target line and adjacent tower nodes of the line with a spatial distance less than a preset threshold (e.g., 500m). It includes physical connection edges (connection of conductors and lightning protection wires between adjacent towers on the same line) and spatial proximity edges (spatial association between tower pairs that cross the line and whose spatial distance is less than a preset threshold).

[0039] S102: Construct node feature vectors.

[0040] For each tower node in the diagram Constructing the initial feature vector It includes features in the following dimensions, as shown in Table 1: Table 1 Feature vector of each tower node

[0041] All features are normalized.

[0042] Among them, historical failure vulnerability middle, Indicates the number of historical power outages. This represents the historical fault weighting coefficient, with a common value range of 0.3 to 0.8. Here, we take... =0.5, which is a moderate empirical value and has good applicability in most scenarios.

[0043] Step 2: Based on the graph structure data, train a graph neural network model to predict the reduction rate of lightning strike failure risk for each tower node in the graph structure data, and optimize it using the total loss function until the total loss function is minimized to obtain the trained graph neural network model; the total loss function includes the standard supervised loss function, physical consistency constraints, and physical equation residual constraints.

[0044] Specifically, a graph neural network model is used to predict the lightning strike risk reduction rate for each tower node. Based on multiple tower nodes, the predicted lightning strike risk reduction rate for each tower node, and the actual lightning strike risk reduction rate, a standard supervised loss function is used to calculate the standard supervised loss value. Based on every two tower nodes, the physical connection edges and spatial proximity edges corresponding to each two tower nodes, physical consistency constraints are used to calculate the physical consistency constraint value. Based on multiple tower nodes, the predicted lightning strike risk reduction rate for each tower node, and the electrical characteristics of the towers, physical equation residual constraints are used to calculate the physical equation residual constraint value. The standard supervised loss value, the physical consistency constraint value, and the physical equation residual constraint value are weighted and summed to obtain the total loss function value. Based on the total loss function value, the graph neural network model is adjusted through backpropagation until the total loss function value is minimized, thus obtaining the trained graph neural network model.

[0045] Step 2 includes the following details: S103: Message passing and node embedding based on graph neural networks.

[0046] Reference Figure 4 ,use layer( =2~4) Graph neural networks are used for message passing and node embedding learning. In the first... Layer, for each node : (a) Aggregate the embedding vectors of the neighboring nodes in the previous layer:

[0047] in, For nodes The set of neighboring nodes, For nodes A neighboring node in the graph In the middle, if there exists an edge connecting the nodes and nodes ,but belong . Includes all nodes Directly connected nodes. The size of the set of neighboring nodes, i.e., the number of nodes. The number of directly connected edges. In the mean aggregation of GNNs, Used as a normalization denominator, For nodes The Layer embedding vector.

[0048] (b) Update its own embedding vector:

[0049] in For the first The learnable weight matrix of the layer, This is the ReLU activation function. After... After layer propagation, the final embedding vector of each node is output. .

[0050] It should be noted that during the message passing and node embedding phases, the node... and nodes This refers to all nodes in the diagram—including target line tower nodes, adjacent line tower nodes, and substation and converter station nodes; however, it is not relevant to subsequent discussions on lightning strike risk reduction rates. In the calculation and evaluation, nodes Specifically refers to tower nodes.

[0051] Step S104: Training the loss function based on physical constraint enhancement.

[0052] Reference Figure 5 This is the core innovative step of the present invention. To address the overfitting problem caused by the extreme scarcity of labeled data, a dual physical constraint loss term based on the Electrogeometric Model (EGM) is introduced on top of the standard supervised loss function of graph neural networks—physical consistency constraint. and physical equation residual constraints .

[0053] (a) Standard monitoring loss For the labeled tower node set Calculate the predicted value Compared with measured values Mean square error between:

[0054] in, The acquisition method is as follows: for towers in the work area where manual lightning suppression operations have been completed, extract the towers from the lightning location system. Ground flash density before and after artificial lightning suppression operation _ and _ and towers Lightning current amplitude probability density function and The line characteristic parameters (protection angle) of the tower. Grounding resistance Insulation level Topographic coefficient (etc.) and lightning parameters before and after intervention are input into the electrical geometric model, and the lightning trip rate before intervention is calculated respectively. and lightning tripping rate after intervention ; and then calculate =( - ) / ×100%. For towers in non-operational areas, due to the lack of measured lightning parameters after intervention, it is impossible to calculate... These poles are the unlabeled nodes in the GNN training.

[0055] (b) Physical consistency constraints Based on the electrical geometry model, a definite functional relationship exists between the line lightning trip rate and the ground flash density, lightning current amplitude distribution, and line characteristic parameters. Physical consistency constraints employ differentiated constraint strategies for spatially adjacent edges and physically connected edges.

[0056] in, This represents the physical consistency constraint value of the total loss function in the graph neural network model. For graph-structured data, it is the set of spatially nearest edges. Tower nodes predicted by a graph neural network model The rate of reduction in the risk of lightning strike failure; Tower nodes predicted by a graph neural network model The rate of reduction in the risk of lightning strike failure; This refers to the set of physical connection edges in graph-structured data. The smoothing coefficient is used. For spatially adjacent edges, it is weighted by physical similarity. Perform differential smoothing — With the two-node line characteristic parameter sub-vector The Euclidean distance is inversely proportional:

[0057] in, For physical similarity weights; Tower nodes in graph structure data Line characteristics; Tower nodes in graph structure data Line characteristics; For nodes The line characteristic parameter sub-vectors contain only three physical dimensions: insulation vulnerability, grounding vulnerability, and shielding vulnerability. This is the Gaussian kernel bandwidth parameter. Line characteristic parameter subvectors of labeled towers Empirical specification of statistical variance: That is, take the median Euclidean distance of the feature differences between the two endpoints of all edges in the graph, so that the weights are... The graph exhibits a reasonable dynamic range. Cross-validation can also be used to select the value from the empirical candidate set {0.1, 0.5, 1.0, 2.0} that minimizes the MSE of the unlabeled tower validation set. For physically connected edges, an independent... Equal weight smoothing constraint, Smoothing coefficient of physical connection edge ( >1, such as =2.0~5.0), applying stronger equal-weight smoothing constraints to the physical connection edges.

[0058] The physical basis for distinguishing between spatially adjacent edges and physically connected edges and applying differentiated constraints is as follows: Two towers connected by a physically connected edge belong to the same transmission line and share the same line design standards (voltage level, insulation coordination, protection angle design value). Therefore, they have consistent shielding failure characteristics in the electrical geometry model and should exhibit a physically appropriate smooth transition trend after manual intervention. Two towers connected by a spatially adjacent edge may belong to different lines, and their line characteristic parameters may differ significantly, requiring further differentiation through... After performing physical similarity assessment, differentiated constraints are applied. If the same constraint is applied to physically connected edges... Weighting, on the contrary, may lead to unnecessary constraint attenuation due to engineering noise; using a larger equal weighting coefficient γ can ensure the same line tower The predictions are consistent with the engineering design.

[0059] The physical meaning of the physical consistency constraint is: the more similar the insulation, grounding, and shielding parameters of two towers, the more similar their risk reduction effects should be after artificial intervention; towers on the same line should naturally have a continuous and smooth risk reduction trend. This loss term causes the graph neural network prediction results to obey known physical laws.

[0060] (c) Physical equation residual constraints To further enhance the substantive nature of physical constraints, residual terms from the physical equations of the electrical geometric model are introduced into the loss function. Based on the electrical geometric model, the line lightning trip rate is... The theoretical calculation formula is:

[0061] in For ground flash density, This is the lightning interception coefficient. Let be the probability density function of the lightning current amplitude. To mask the failure probability function, This represents the lower limit of the line's withstand lightning current. The physical meaning of this formula is that the line's lightning trip rate is jointly determined by the ground flash density, the lightning current amplitude distribution, and the line's shielding performance.

[0062] After artificial intervention, the ground flash density changed from Down to (Assume the rate of change is) The distribution of lightning current amplitude is as follows: Become According to the electrical geometry model, the tower Theoretical risk reduction rate This can be deduced as:

[0063] in, Tower nodes in graph structure data The theoretical risk reduction rate; This refers to the ground flash density before artificial lightning suppression operations. This refers to the ground flash density after artificial lightning suppression operations. The distribution of lightning current amplitude before manual lightning suppression operations; The distribution of lightning current amplitude after artificial lightning suppression operations; Tower nodes in graph structure data The shielding failure probability function; This represents the amplitude of the lightning current, in kA. This represents the lower limit of the lightning current withstand capability of the target transmission line in the graph structure data. Protected by the tower's angle Grounding resistance and topography coefficient A joint decision.

[0064] In one specific embodiment of the present invention, the probability density function of lightning current amplitude... The cumulative distribution of lightning current amplitude recommended in the guidelines for improving lightning protection performance of overhead power distribution lines (IEEE Std 1410-2010) is adopted, and its complementary cumulative distribution function is: The corresponding probability density function for:

[0065] Before and after intervention and The shielding failure probability function is obtained by substituting the statistical median values ​​of the lightning current amplitude before and after the intervention (32kA and 25kA respectively in Application 1) into the above formula. Using empirical formulas from electrical geometry models:

[0066] in For towers The shielding failure characteristic current is determined by the protection angle of the tower. Tower height and topography coefficient Decide, The scale parameter, which is used to mask the failure probability distribution, is usually taken as 5~8kA.

[0067] The residual constraint of the physical equation is defined as the predicted value of the graph neural network. Compared with the theoretical value of the electrical geometric model Mean square error between:

[0068] in, The physical equation residual constraint value for the total loss function in the graph neural network model; Tower nodes predicted by a graph neural network model The rate of reduction in the risk of lightning strike failure; For all tower nodes in the graph structure data, this constraint directly penalizes the deviation between the graph neural network predictions and the theoretical values ​​of the electrical geometry model, ensuring that the predictions are not only smooth between adjacent towers but also consistent with the engineering theory of lightning tripping in terms of physical mechanism. (Compared to standard supervised loss) Only for the labeled tower node set The monitoring signals provided are different. For all tower nodes in the diagram All provide physical monitoring signals—including unlabeled tower nodes—so that physical constraints can also prevent overfitting in areas not covered by labeled data.

[0069] (d) Total loss function:

[0070] in, This represents the total loss function value of the graph neural network model. The weighting coefficients for physical consistency constraints (recommended) =0.1~0.5) The value is inversely proportional to the number of tower nodes already marked in the graph structure data; These are the weighting coefficients for the residual constraints of the physical equations (suggested). =0.05~0.3) The value is inversely proportional to the number of tower nodes already marked in the graph structure data; This is the standard supervised loss function for graph neural network models; Physical consistency constraints for the total loss function in a graph neural network model; This represents the residual constraint value of the physical equation for the total loss function in the graph neural network model. When labeled data is extremely limited (| |<500), increase and To enhance the guiding role of physical constraints; as the labeled data gradually becomes richer (| (>2000), both can be appropriately reduced to allow data-driven approaches to gradually take the lead.

[0071] (e) Model training: Minimizing via backpropagation and gradient descent Update the weight matrices of each layer in the graph neural network. And prediction head parameters.

[0072] In practical applications, traditional technologies, particularly purely data-driven graph neural networks, face a fundamental obstacle in the field of artificial mine suppression: a lack of data for training. Specifically, this manifests as follows: Firstly, labeled data at the pole level is extremely scarce, making it impossible to effectively train a purely data-driven model. Training a graph neural network requires a large number of labeled samples—that is, data with known "input features → output..." The data on towers is limited. However, artificial lightning suppression is currently in the research and exploration stage, and the actual number of field operations is extremely limited. Assuming that one operation covers the observation range of about 200 towers, even after 10 operations, only about 2,000 labeled samples can be accumulated. For transmission line maps that often contain thousands of towers (such as ±800kV lines which can have more than 3,200 tower nodes), the labeling coverage is less than 60% of the entire map, and the labeled nodes are highly concentrated in the operation area. Under such sparse and uneven labeling conditions, purely data-driven graph neural networks will suffer from severe overfitting—the model's predictions are acceptable in the operation area, but its generalization ability to towers in non-operation areas is extremely poor, and the prediction results lack physical credibility. It should be noted that the scarcity of labeled data in the field of artificial lightning suppression is not a common data acquisition problem that can be solved by "increasing the number of experiments," but rather a unique technical obstacle in this field determined by the research and exploration stage of artificial lightning suppression technology itself. Unlike conventional machine learning applications where the scarcity of annotations can be alleviated through crowdsourcing, data augmentation, or transfer learning, the pole-level annotation data for artificial mine suppression must be obtained through actual field operations. Each operation costs hundreds of thousands of yuan, involves meteorological condition assessment, airspace approval, and multi-party collaboration, and can only be carried out a few times a year. Furthermore, the annotable poles are limited to the coverage area of ​​the work zone. This results in the accumulation of annotation data being measured in years, creating an order-of-magnitude gap compared to the training requirements of deep learning models, which often require tens of thousands of labeled samples. Traditional methods for solving small-sample problems, such as semi-supervised learning, transfer learning, and data augmentation, are all based on the assumption of at least several hundred to several thousand labeled samples, while the labeled samples in the early stages of artificial mine suppression technology promotion may be less than a hundred. Therefore, the technical obstacle in this field is not simply a lack of data volume, but a significant difficulty in meeting the requirements of model training with the rate of annotation data accumulation, necessitating fundamental innovation at the technical solution level.

[0073] Secondly, traditional graph neural network methods lack utilization of known physical laws. The physical mechanism of lightning tripping of transmission lines has been quantitatively described well through methods such as electrical geometric models—there are clear functional relationships between the line lightning tripping rate and ground flash density, lightning current amplitude distribution, line insulation level, grounding resistance, protection angle, and terrain coefficient. When a graph neural network predicts vastly different results for two towers with similar basic characteristics... At this point, the prediction is physically contradictory. Purely data-driven graph neural networks lack a mechanism to penalize such physically inconsistent predictions.

[0074] It should be noted that the solution to the above problems faces non-trivial obstacles in practice due to cross-domain technology combinations: (1) Smoothing regularization in graph semi-supervised learning (such as Laplacian regularization) applies equal-weight smoothing constraints to all graph edges, without distinguishing between "pairs of towers with similar physical features but existing edges due to spatial proximity" and "pairs of towers with different physical features but also existing edges". In transmission line diagrams, spatially adjacent edges may connect towers with different voltage levels and different design standards - applying equal-weight smoothing constraints to such tower pairs will introduce noise and reduce generalization performance. Therefore, directly applying graph smoothing regularization to this field cannot solve the problem of scarce annotations, and may instead exacerbate overfitting. (2) Physics-Informed Neural Networks (PINN) usually rely on a large number of collocations to calculate the residuals of physical equations - in traditional PINN applications, collocations can be sampled arbitrarily in the solution domain, and the number is not limited. However, in the assessment of artificial lightning suppression, the calculation of the residual equations of the electrical geometric model requires specific physical parameters (protection angle, grounding resistance, terrain coefficient) for each tower. These parameters are only defined at the tower nodes—the number of tower nodes is fixed and cannot be arbitrarily increased as in traditional PINN. Therefore, simply applying the PINN framework to this field results in the signal density of physical constraints being limited by the number of tower nodes, and the effect of traditional PINN cannot be achieved. (3) The superposition of the above two obstacles means that graph smoothing regularization (which requires distinguishing edge types) and PINN physical constraints (which are limited by the number of nodes) cannot solve the problem in this field on their own. Only by deeply integrating the two under a specific graph structure framework—using GNN message passing to propagate the physical constraint signal along the graph edges to the unlabeled nodes, and using physical features to weight and distinguish the constraint strength of different types of edges—can their respective limitations be overcome.

[0075] In step 2, the dual constraints of physical consistency constraints and physical equation residual constraints can prevent overfitting during model training when labeled data is scarce in graph structure data. Furthermore, physical consistency constraints employ differentiated constraint strategies for spatially adjacent edges and physically connected edges. To further enhance the substantive nature of physical constraints, physical equation residual constraints are added. Ultimately, the total loss function of the graph neural network model includes the standard supervised loss function, physical consistency constraints, and physical equation residual constraints. By using the total loss function to optimize the training of the graph neural network model, a graph neural network model that meets the requirements can be obtained.

[0076] Step 3: Based on the graph structure data, use the trained graph neural network model to predict the lightning strike failure risk reduction rate of each tower node in the graph structure data.

[0077] Step 4: Evaluate the effectiveness of manual lightning suppression intervention based on the predicted reduction rate of lightning strike risk at each tower node.

[0078] Specifically, based on each tower node: if the predicted reduction rate of lightning strike risk for a tower node is within a preset high-benefit range, then the assessment result corresponding to the predicted reduction rate of lightning strike risk for that tower node is determined to be that the effect of human intervention is significant; if the predicted reduction rate of lightning strike risk for a tower node is within a preset medium-benefit range, then the assessment result corresponding to the predicted reduction rate of lightning strike risk for that tower node is determined to be that the effect of human intervention is obvious; if the predicted reduction rate of lightning strike risk for a tower node is within a preset low-benefit range, then the assessment result corresponding to the predicted reduction rate of lightning strike risk for that tower node is determined to be that the effect of human intervention is limited; if the predicted reduction rate of lightning strike risk for a tower node is within a preset ineffective range, then the assessment result corresponding to the predicted reduction rate of lightning strike risk for that tower node is determined to be that the effect of human intervention is not significant.

[0079] The details of steps 3 and 4 are as follows: Step S105: Refined evaluation of the intervention effect at the tower / line section level.

[0080] Based on the trained graph neural network model, the predicted reduction rate of lightning strike risk for each tower is output. The line segments are divided into four levels according to the intervals shown in Table 2: Table 2 Evaluation Interval Division

[0081] Step 5: Based on the graph structure data and the trained graph neural network model, use Monte Carlo Dropout to perform multiple independent forward propagations on several unlabeled tower nodes among multiple tower nodes to obtain the predicted value of each unlabeled tower node; based on the predicted values ​​of several unlabeled tower nodes, determine the prediction variance of each unlabeled tower node in the graph structure data; sort the prediction variances of multiple unlabeled tower nodes in the graph structure data from largest to smallest, select the top-ranked unlabeled tower nodes to obtain the recommended labeled tower nodes for the graph structure data; conduct field measurements on the recommended labeled tower nodes in the graph structure data to obtain the measured values ​​of the lightning strike risk reduction rate of the recommended labeled tower nodes in the graph structure data; update the graph structure data using the measured values ​​of the lightning strike risk reduction rate of the recommended labeled tower nodes in the graph structure data; use the updated graph structure data for the evaluation of the effect of the next manual lightning suppression intervention.

[0082] The details of step 5 are as follows: Step S106: Active learning of sample recommendation based on prediction uncertainty.

[0083] Reference Figure 6 To address the problem of optimizing the allocation of limited annotation resources, this paper proactively recommends the most worthwhile towers to be annotated in the next batch based on the prediction uncertainty of graph neural networks.

[0084] (a) Estimation of prediction uncertainty: Monte Carlo dropout method is used, with Dropout enabled during the inference phase for each unlabeled tower. Perform T independent forward propagations (e.g., T=50) to obtain T predicted values. , The average of the T predicted values. For the first The next independent forward propagation takes values ​​from 1 to T. The prediction variance is calculated as a measure of uncertainty.

[0085] (b) Active learning sampling strategy: Sort by prediction variance from largest to smallest, and select the top B unlabeled towers with the largest variance as the next batch of recommended labeling targets. Towers with large prediction variance are given priority because labeling these towers is expected to minimize model uncertainty and improve generalization performance.

[0086] (c) Iterative Closed-Loop Design of Resources: The recommended list of towers is provided to the manual lightning suppression operation team. In the next operation, monitoring equipment is deployed on these towers to obtain actual measurement data. Label the data; add the newly labeled data to the training set, retrain the model, and update. Prediction forms a closed-loop iteration from model prediction to uncertainty assessment to recommended labeling to new data acquisition to model updates.

[0087] Traditional techniques lack an active learning mechanism to guide the optimal allocation of limited annotation resources. Monitoring and evaluating the effectiveness of each manual lightning suppression operation is costly, making it impossible to thoroughly annotate all towers along the entire line. Traditional methods do not provide an effective solution for selecting the most noteworthy towers from a large number of unannotated towers—those that provide the greatest improvement in model accuracy after annotation—to guide the allocation of limited field observation resources. Step 5, however, recommends noteworthy towers from among the unannotated towers and updates the graph structure data using the measured reduction rate of lightning strike risk at the recommended tower nodes. This updated graph structure data is then used to evaluate the effectiveness of the next manual lightning suppression intervention. This closed-loop design achieves an active learning mechanism, where each updated graph structure data improves model accuracy and guides the optimal allocation of limited annotation resources.

[0088] The overall flowchart of the method of this invention is as follows: Figure 2As shown, the summary is as follows: The power grid topology is modeled as graph data; a graph neural network is used for node embedding learning; a total loss function integrating standard supervised loss and dual physical constraint loss based on the electrical geometry model is constructed for training—physical consistency constraints constrain the predictive smoothness of spatially adjacent edges through feature similarity weights and impose equal-weight smoothness constraints on physically connected edges; physical equation residual constraints constrain the physical rationality of predicted values ​​through theoretical values ​​of the electrical geometry model and also provide supervised signals for unlabeled nodes. The dual constraints work together to prevent overfitting caused by scarce labeled data; active learning sample recommendation is performed based on prediction uncertainty. This invention solves the unique technical obstacle in the field of artificial lightning suppression, where the extreme scarcity of labeled data prevents the effective training of purely data-driven models, and for the first time achieves deep integration of electrical geometry models and graph neural networks in the evaluation of artificial lightning suppression effects.

[0089] The core technology of this invention is named the Edge Type Aware Dual Physical Constraint Physics-Informed Graph Neural Network (PI-GNN). Edge type awareness refers to the differentiated processing strategy for spatially adjacent edges and physically connected edges in the physical consistency constraint—it can automatically identify different types of edges in the graph and apply different constraint strengths; dual physical constraints refer to the loss function simultaneously including two physical constraint terms—physical consistency constraint and physical equation residual constraint.

[0090] The method for evaluating the effectiveness of artificial lightning suppression intervention of the present invention is applied to an evaluation device for the effectiveness of artificial lightning suppression intervention, such as... Figure 7 As shown, the system includes: a power grid topology graph construction module: modeling the target transmission line and associated equipment as graph structure data, including physical connection edges and spatial proximity edges; a node feature engineering module: constructing a ten-dimensional initial node feature vector including insulation vulnerability, grounding vulnerability, shielding vulnerability, terrain vulnerability, historical fault vulnerability, ground flash density and rate of change, node type, and node degree, and performing normalization processing; a graph neural network inference module: performing K-layer message passing and node embedding learning, outputting node embedding vectors; and a physical constraint enhancement training module: constructing a system that integrates standard supervised loss and physical consistency constraints. and physical equation residual constraints Total loss function Execute graph neural network model training; incorporates an adaptive dynamic adjustment strategy for physical constraint weight coefficients—when the number of labeled towers is less than 500. and The larger value is selected, and both gradually decrease as the number of labeled towers increases. The physical consistency constraint calculation unit has a built-in edge type discrimination subunit to distinguish between spatially adjacent edges and physically connected edges, and applies a physical similarity weighted smoothing strategy to spatially adjacent edges and an equal-weight smoothing strategy to physically connected edges. The refined evaluation module outputs the values ​​of each tower based on the node embedding vector. The system predicts values ​​and performs a four-level segmented evaluation. The active learning sample recommendation module estimates prediction uncertainty based on the Monte Carlo dropout method and recommends a list of priority towers for the next batch of labeling based on variance.

[0091] Compared with traditional technologies, the present invention has the following advantages: (1) This invention pioneers a dual-physical-constraint-enhanced graph neural network with edge type awareness, overcoming the unique technical obstacle in this field where the extreme scarcity of manually labeled lightning suppression data prevents the training of purely data-driven models. For the first time, this invention embeds the physical laws of the electrical geometry model into the loss function of the graph neural network in a dual form—physical consistency constraints. Spatial neighbor edges are weighted by physical similarity Differential smoothing is performed, and equal-weight constraints are applied to physical connection edges to reflect engineering design consistency; physical equation residual constraints are also applied. The predicted values ​​are directly aligned with the physical theoretical values ​​using the theoretical formulas of the electrical geometry model. This dual constraint mechanism effectively prevents overfitting under conditions of extremely sparse labeled data, enabling the graph neural network to achieve reliable tower-level fine-grained evaluation even with only a few hundred labeled samples. (2) The dual physical constraints produce a non-obvious synergistic and complementary effect, and their combined effect exceeds the sum of the individual parts. Physical consistency constraint Using edge structure in graph topology to constrain the predictive smoothness of adjacent towers—this is a local relative constraint (adjacent towers with similar features) (should be similar), physical equation residual constraints Using theoretical formulas from electrical geometry models to constrain the absolute deviation between predicted values ​​and engineering theory—this is a global absolute constraint (per tower). (It should approximate the theoretical value of EGM). Local relative constraints and global absolute constraints constrain the model from two dimensions: the relationship between towers and the mechanism of a single tower, respectively, forming a positive interactive complement. Ablation experimental data directly prove the non-additive nature of the synergistic effect: using alone... The MSE of the unmarked tower is 0.018, used alone. The unlabeled tower's MSE was 0.022. If the two methods only had an additive effect, the expected improvement from their combined use should be a linear sum of their individual improvements—however, the actual MSE from the combined use was 0.015, meaning the actual improvement far exceeded the expected value corresponding to the theoretical sum of improvements. This difference indicates... The provided local smoothing is Global alignment provides a more stable and optimized landscape. The physical anchoring, in turn, reinforced The ability to distinguish anomalous tower pairs—there exists a cooperative mechanism between the two beyond addition. Furthermore, compared to the standard PINN method (only... A comparison with a graph without GNN topology (MSE = 0.026) shows that this invention naturally utilizes the implicit physical priors in the power grid topology through the message passing mechanism of GNN, enabling physical constraint signals to propagate along graph edges to unlabeled areas, achieving a generalization error reduction of approximately 42% compared to the standard PINN. Further γ ablation experiments demonstrate that the smoothing coefficient of physical connection edges has an optimal range of 2.0–5.0—when γ = 1.0, the edge type awareness degenerates to an effect similar to the unified constraint version (MSE = 0.015). =2.0~3.0 is optimal (MSE=0.012~0.013). =5.0, due to excessive smoothing, the generalization performance slightly degraded (MSE=0.013), proving that the selection of this parameter is not arbitrary but has a non-obvious optimal range. (3) Establish an active learning closed loop based on prediction uncertainty to realize the intelligent allocation of limited labeling resources. Estimate the prediction variance by Monte Carlo dropout method, automatically recommend the most worthwhile towers to be labeled, so that the limited field monitoring resources can maximize the improvement of model accuracy. (4) Adaptive adjustment strategy of physical constraint weight coefficient to realize the dynamic balance between physical model and data-driven. When the labeled data is very small, and The physical constraints automatically increase, guiding model training primarily; as labeled data gradually accumulates, both automatically decrease, allowing data-driven approaches to take the lead, achieving a smooth transition from physical to data-driven approaches.

[0092] The specific applications of this invention are as follows: Application 1: PI-GNN Refined Evaluation for ±800kV UHVDC Lines. Scenario: A ±800kV UHVDC transmission line is approximately 1200km long with 3200 towers. Currently, only 3 manual lightning suppression operations have been completed, accumulating data for approximately 480 towers. Actual measurement and labeling data showed a labeling coverage rate of only 15%, concentrated in three work areas along the middle section of the line. The method of this invention is now being used to cover the entire line at the tower level. Detailed assessment.

[0093] Steps S101~S102: Graph modeling and feature extraction.

[0094] Construction Graph , | |≈4052 (including 3200 target towers, 850 adjacent 500kV line towers, and 2 converter station nodes),| |≈5600, where the physical connection edges| ≈4400 spatially adjacent edges| ≈1200 entries. Extract ten-dimensional node feature vectors and normalize them. Taking tower number 156 as an example: =1950kV, =18Ω, =8°, terrain is hillside =2 times, =8.6, =5.2, node degree is 3. The normalized feature vector is: [0.51, 0.72, 0.36, 0.68, 0.52, 0.86, 0.52, 0.40, 0, 0.30].

[0095] Steps S103~S104: Graph Neural Network Training and Physical Constraint Enhancement.

[0096] A 3-layer graph neural network (GraphSAGE) architecture is used, with a hidden layer dimension of 64. Only 480 labeled towers provide supervision signals, while the remaining approximately 2772 towers are unlabeled. Taking the median Euclidean distance of the feature differences between the two endpoints of all edges in the graph as 0.42, the smoothing coefficient of the physically connected edges is... =3.0.

[0097] First, only standard supervised loss is used. train( =0, =0), after 50 training rounds, the model's MSE on labeled towers decreased to 0.008, but manual evaluation was performed on 30 randomly selected unlabeled towers—predicting With reality The MSE is as high as 0.042, indicating significant overfitting. Then, physical constraint loss is added. and ( =0.25, =0.15), retrained for 50 rounds: labeled tower MSE = 0.012 (slightly increased due to the trade-off between physical constraints and data fitting), but unlabeled tower MSE decreased to 0.012, resulting in an approximately 71% improvement in generalization performance. As shown in Table 3: Table 3 Comparison of Physical Constraint Effects

[0098] To further verify the smoothness coefficient of physical connection edges The selection of γ was not arbitrary; ablation experiments were conducted to test the model performance under different values ​​of γ. All experiments were performed using 480 labeled samples, the same GNN architecture, and... =0.25、 The test was conducted under the condition of 0.15, as shown in Table 4: Table 4 ablation experiment

[0099] The ablation experiments showed a non-monotonic trend of "first decreasing and then increasing"— When the value is 1.0, the MSE is 0.015 (close to uniformity). Version), When the value is 3.0, it decreases to the optimal value of 0.012. When γ = 5.0, it rises back to 0.013. This trend indicates that the smoothness coefficient of the physical connection edge has an optimal range of 2.0 to 5.0, and is not necessarily better the larger or the smaller: if γ is too small, it cannot fully utilize the prior knowledge of the engineering design consistency of the same line tower. If the value is too large, the local differences between towers along the same line will be excessively smoothed, leading to prediction distortion. This non-monotonic trend further proves that the selection of the γ parameter in the edge type-aware strategy cannot be easily determined by those skilled in the art without creative effort. To further verify the unique technical effect of the physical constraint scheme of this invention, it is supplemented with the pure graph smoothing regularization method (without physical weights). Laplacian regularization, i.e., the weight of all edges. The comparative experiments with a uniform value of 1) are shown in Table 5: Table 5 Comparative Test Table

[0100] The above comparison results show that: pure graph smoothing regularization ( =1) It improves generalization performance to some extent (MSE decreases from 0.042 to 0.031), but the improvement (approximately 26%) is far less than that of the physical constraint scheme of this invention. The edge type-aware version further reduces MSE from 0.015 to 0.012, demonstrating the effectiveness of the strategy of distinguishing between spatially adjacent edges and physically connected edges and applying differentiated constraints—the equal-weighted strong smoothing on physically connected edges fully utilizes the prior engineering design consistency of the same line tower, avoiding... Unnecessary attenuation under weighted engineering design noise.

[0101] Furthermore, to verify that the technical effect of this invention truly stems from the deep fusion of "physical constraints + GNN + edge type awareness," rather than any single regularization method achieving the same effect, a comparison with the standard PINN method is provided, as shown in Table 6: Table 6 Comparison of PI-GNN and Standard PINN Methods

[0102] While the standard PINN method utilizes physical equations for constraints, it neglects the physical information inherent in the graph topology—based on the electrical geometry model, adjacent towers on the same line share similar lightning exposure environments and line design standards, thus their... There should be a smooth physical transition relationship. This invention naturally utilizes this graph topological prior through the message passing mechanism of GNN, and combines it with an edge type-aware strategy and EGM residual constraints to achieve a generalization error reduction of approximately 54% compared to the standard PINN.

[0103] Step S105: Detailed evaluation of the entire line.

[0104] PI-GNN Prediction Based on Edge Type Aware Dual Physical Constraints: High-Efficiency Segments ≥50%) Approximately 38% of the towers, medium-efficiency section (30%≤ <50%) approximately 32%, low-efficiency segment (10%≤) <30%) approximately 18%, invalid segments ( <10%) Approximately 12%. Invalid segments are mainly concentrated in the cross-canyon section ( =1.75) and sections with high grounding resistance ( >25Ω). High-efficiency sections are concentrated in open plain areas with low grounding resistance. =0.8~1.0, <10Ω).

[0105] Step S106: Active learning recommendation.

[0106] A Monte Carlo dropout method was used to perform 50 forward propagations on all unlabeled towers along the entire route, and the prediction variance was calculated. Of the 30 towers with the highest prediction variance, 22 (73%) were located in the transition zone between high-efficiency and low-efficiency sections—these boundary towers are precisely the regions where graph neural networks struggle to classify. It is recommended to prioritize labeling these 30 towers in the next assignment. Simulation results showed that after adding 30 labels, the MSE of unlabeled towers using the type-aware dual-physical-constraint PI-GNN further decreased to 0.009.

[0107] Application 2: Dynamic adaptive adjustment of physical constraint weight coefficients.

[0108] In the above application 1, a fixed method is used. =0.25、 =0.15. This application demonstrates the adaptive adjustment strategy for both. Initially, the labeled data consisted of only 100 towers: =0.4、 =0.25 (physical constraint dominant). As the labeled data gradually increases to 480 bases, both decrease linearly to... =0.25、 =0.15. When the number of labeled data exceeds 2000, both decrease to below 0.1, allowing data-driven approaches to gradually take the lead, as shown in Table 7: Table 7. Annotated Data Results

[0109] The adaptive strategy significantly improved model performance in the early stages when labeled data was scarce.

[0110] Example 2: Based on the same inventive concept, this invention also provides an evaluation system for the effectiveness of artificial lightning suppression intervention, such as... Figure 8 As shown, it includes: The data acquisition module is used to acquire graph structure data of the target transmission lines and related power grid equipment in the power grid system before and after artificial lightning suppression operations during thunderstorms. The loss optimization training module is used to train a graph neural network model based on the graph structure data to predict the reduction rate of lightning fault risk for each tower node in the graph structure data, and to optimize it using a total loss function until the total loss function is minimized, obtaining the trained graph neural network model. The total loss function includes a standard supervised loss function, physical consistency constraints, and physical equation residual constraints. The model prediction module is used to predict the lightning fault risk reduction rate for each tower node in the graph structure data using the trained graph neural network model. The intervention effect evaluation module is used to evaluate the effect of artificial lightning suppression intervention based on the predicted lightning fault risk reduction rate for each tower node.

[0111] Furthermore, the graph structure data includes multiple tower nodes, the electrical characteristics of each tower node, the physical connection edges and spatial proximity edges between every two tower nodes, and the true value of the lightning fault risk reduction rate corresponding to the labeled tower nodes among the multiple tower nodes.

[0112] Furthermore, the loss optimization training module is specifically used for: using a graph neural network model to predict the lightning fault risk reduction rate for each tower node; calculating the standard supervised loss value using a standard supervised loss function based on multiple tower nodes, the predicted lightning fault risk reduction rate for each tower node, and the actual lightning fault risk reduction rate; calculating the physical consistency constraint value using physical consistency constraints based on every two tower nodes, the physical connection edges and spatial proximity edges corresponding to every two tower nodes; calculating the physical equation residual constraint value using physical equation residual constraints based on multiple tower nodes, the predicted lightning fault risk reduction rate for each tower node, and the electrical characteristics of the towers; weighted summing the standard supervised loss value, the physical consistency constraint value, and the physical equation residual constraint value to obtain the total loss function value; and adjusting the graph neural network model through backpropagation based on the total loss function value until the total loss function value is minimized, thus obtaining the trained graph neural network model.

[0113] Furthermore, it also includes: an active learning sample recommendation module; specifically, the active learning sample recommendation module is used to: based on graph structure data and a trained graph neural network model, perform multiple independent forward propagations on several unlabeled tower nodes among multiple tower nodes using the Monte Carlo dropout method to obtain the predicted value of each unlabeled tower node; based on the predicted values ​​of several unlabeled tower nodes, determine the prediction variance of each unlabeled tower node in the graph structure data; sort the prediction variances of multiple unlabeled tower nodes in the graph structure data from largest to smallest, select the top-ranked unlabeled tower nodes to obtain the recommended labeled tower nodes in the graph structure data; conduct actual measurements on the recommended labeled tower nodes in the graph structure data to obtain the measured value of the lightning strike fault risk reduction rate of the recommended labeled tower nodes in the graph structure data; update the graph structure data using the measured value of the lightning strike fault risk reduction rate of the recommended labeled tower nodes in the graph structure data; and use the updated graph structure data for the evaluation of the effect of the next manual lightning suppression intervention.

[0114] Furthermore, the physical consistency constraint is calculated using the following formula:

[0115] in, This represents the physical consistency constraint value of the total loss function in the graph neural network model. For graph-structured data, it is the set of spatially nearest edges. Tower nodes predicted by a graph neural network model The rate of reduction in the risk of lightning strike failure; Tower nodes predicted by a graph neural network model The rate of reduction in the risk of lightning strike failure; This refers to the set of physical connection edges in graph-structured data. For smoothing coefficients;

[0116] in, For physical similarity weights; Tower nodes in graph structure data Line characteristics; Tower nodes in graph structure data Line characteristics; This is the Gaussian kernel bandwidth parameter.

[0117] Furthermore, the residual constraints of the physical equations are calculated using the following formula:

[0118] in, The physical equation residual constraint value for the total loss function in the graph neural network model; Tower nodes predicted by a graph neural network model The rate of reduction in the risk of lightning strike failure; This refers to all tower nodes in the graph structure data;

[0119] in, Tower nodes in graph structure data The theoretical risk reduction rate; This refers to the ground flash density before artificial lightning suppression operations. This refers to the ground flash density after artificial lightning suppression operations. The distribution of lightning current amplitude before manual lightning suppression operations; The distribution of lightning current amplitude after artificial lightning suppression operations; Tower nodes in graph structure data The shielding failure probability function; This refers to the amplitude of the lightning current. This represents the lower limit of the lightning current withstand capability of the target transmission line in the graph structure data.

[0120] Furthermore, the total loss function value is calculated using the following formula:

[0121] in, This represents the total loss function value of the graph neural network model. These are the weighting coefficients for physical consistency constraints. The value is inversely proportional to the number of tower nodes already marked in the graph structure data; These are the weighting coefficients for the residual constraints of the physical equations. The value is inversely proportional to the number of tower nodes already marked in the graph structure data; This is the standard supervised loss function for graph neural network models; Physical consistency constraints for the total loss function in a graph neural network model; The physical equation residual constraint value for the total loss function in the graph neural network model.

[0122] Furthermore, the intervention effect evaluation module is specifically used for: based on each tower node: if the predicted reduction rate of lightning strike risk for a tower node is within a preset high-efficiency range, then the evaluation result corresponding to the predicted reduction rate of lightning strike risk for that tower node is determined to be that the human intervention effect is significant; if the predicted reduction rate of lightning strike risk for a tower node is within a preset medium-efficiency range, then the evaluation result corresponding to the predicted reduction rate of lightning strike risk for that tower node is determined to be that the human intervention effect is obvious; if the predicted reduction rate of lightning strike risk for a tower node is within a preset low-efficiency range, then the evaluation result corresponding to the predicted reduction rate of lightning strike risk for that tower node is determined to be that the human intervention effect is limited; if the predicted reduction rate of lightning strike risk for a tower node is within a preset ineffective range, then the evaluation result corresponding to the predicted reduction rate of lightning strike risk for that tower node is determined to be that the human intervention effect is not significant.

[0123] Example 3: Based on the same inventive concept, the present invention also provides a computer device, such as... Figure 9 As shown, the computer device includes a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, and is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the computer storage medium to implement the corresponding method flow or corresponding function, so as to realize the steps of the method for evaluating the effect of artificial lightning suppression intervention in the above embodiment.

[0124] Example 4: Based on the same inventive concept, this invention also provides a storage medium, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the steps of the method for evaluating the effect of artificial lightning suppression intervention in the above embodiments.

[0125] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0126] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0127] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0128] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0129] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of the claims of the present invention pending approval.

Claims

1. A method for evaluating the effectiveness of artificial lightning suppression intervention, characterized by the following steps: include: During thunderstorms, artificial lightning suppression operations are carried out to obtain graphical structure data of the target transmission lines and associated power grid equipment in the power grid system before and after the artificial lightning suppression operations. Based on the graph structure data, a graph neural network model is trained to predict the reduction rate of lightning strike failure risk for each tower node in the graph structure data, and optimized using a total loss function until the total loss function is minimized to obtain the trained graph neural network model; the total loss function includes a standard supervised loss function, physical consistency constraints, and physical equation residual constraints; Based on the graph structure data, the trained graph neural network model is used to predict the lightning strike failure risk reduction rate of each tower node in the graph structure data. The effectiveness of artificial lightning suppression intervention is evaluated based on the predicted reduction rate of lightning strike risk at each of the aforementioned tower nodes.

2. The method as described in claim 1, characterized in that, The graph structure data includes multiple tower nodes, the electrical characteristics of each tower node, the physical connection edges and spatial proximity edges between every two tower nodes, and the true value of the lightning strike fault risk reduction rate corresponding to the marked tower nodes among the multiple tower nodes.

3. The method as described in claim 2, characterized in that, The step of training a graph neural network model based on the graph structure data to predict the reduction rate of lightning strike failure risk for each tower node in the graph structure data, and optimizing it using a total loss function until the total loss function is minimized, thereby obtaining the trained graph neural network model, includes the following steps: The graph neural network model is used to predict the reduction rate of lightning strike risk for each tower node. Based on multiple tower nodes, the predicted value of the lightning fault risk reduction rate corresponding to each tower node and the actual value of the lightning fault risk reduction rate, the standard supervised loss value is calculated using the standard supervised loss function. Based on each pair of tower nodes, the physical connection edge and spatial proximity edge corresponding to each pair of tower nodes, the physical consistency constraint value is obtained by calculating using the physical consistency constraint. Based on multiple tower nodes, the predicted reduction rate of lightning fault risk for each tower node, and the electrical characteristics of the tower, the residual constraint value of the physical equation is obtained by calculation using the physical equation residual constraint. The total loss function value is obtained by weighted summing of the standard supervision loss value, the physical consistency constraint value, and the physical equation residual constraint value. Based on the total loss function value, the graph neural network model is adjusted through backpropagation until the total loss function value is minimized, thus obtaining the trained graph neural network model.

4. The method as described in claim 1 or 2, characterized in that, After evaluating the effectiveness of artificial lightning suppression intervention based on the predicted reduction rate of lightning fault risk for each of the aforementioned tower nodes, the method further includes: Based on the graph structure data and the trained graph neural network model, the Monte Carlo dropout method is used to perform multiple independent forward propagations on several unlabeled tower nodes among multiple tower nodes to obtain the predicted value of each unlabeled tower node. Based on the predicted values ​​of the aforementioned unlabeled tower nodes, the prediction variance of each unlabeled tower node in the graph structure data is determined; Sort the prediction variances of multiple unlabeled tower nodes in the graph structure data from largest to smallest, and select the top-ranked unlabeled tower nodes to obtain the recommended labeled tower nodes for the graph structure data. The recommended labeled tower nodes in the graph structure data were measured to obtain the measured value of the lightning strike failure risk reduction rate of the recommended labeled tower nodes in the graph structure data. The graph structure data is updated using the measured values ​​of the lightning strike failure risk reduction rate of the recommended labeled tower nodes in the graph structure data; The updated graph structure data will be used to evaluate the effectiveness of the next artificial lightning suppression intervention.

5. The method as described in claim 3, characterized in that, The physical consistency constraint is calculated using the following formula: in, This represents the physical consistency constraint value of the total loss function in the graph neural network model. For graph-structured data, it is the set of spatially nearest edges. Tower nodes predicted by a graph neural network model The rate of reduction in the risk of lightning strike failure; Tower nodes predicted by a graph neural network model The rate of reduction in the risk of lightning strike failure; This refers to the set of physical connection edges in graph-structured data. For smoothing coefficients; in, For physical similarity weights; Tower nodes in graph structure data Line characteristics; Tower nodes in graph structure data Line characteristics; This is the Gaussian kernel bandwidth parameter.

6. The method as described in claim 3, characterized in that, The residual constraint of the physical equation is calculated using the following formula: in, The physical equation residual constraint value for the total loss function in the graph neural network model; Tower nodes predicted by a graph neural network model The rate of reduction in the risk of lightning strike failure; This refers to all tower nodes in the graph structure data; in, Tower nodes in graph structure data The theoretical risk reduction rate; This refers to the ground flash density before artificial lightning suppression operations. This refers to the ground flash density after artificial lightning suppression operations. The distribution of lightning current amplitude before manual lightning suppression operations; The distribution of lightning current amplitude after artificial lightning suppression operations; Tower nodes in graph structure data The shielding failure probability function; This refers to the amplitude of the lightning current. This represents the lower limit of the lightning current withstand capability of the target transmission line in the graph structure data.

7. The method as described in claim 3, characterized in that, The total loss function value is calculated using the following formula: in, This represents the total loss function value of the graph neural network model. These are the weighting coefficients for physical consistency constraints. The value is inversely proportional to the number of tower nodes already marked in the graph structure data; These are the weighting coefficients for the residual constraints of the physical equations. The value is inversely proportional to the number of tower nodes already marked in the graph structure data; This is the standard supervised loss function for graph neural network models; Physical consistency constraints for the total loss function in a graph neural network model; The physical equation residual constraint value for the total loss function in the graph neural network model.

8. The method as described in claim 1, characterized in that, The evaluation of the effectiveness of artificial lightning suppression intervention based on the predicted reduction rate of lightning fault risk for each of the aforementioned tower nodes includes the following steps: Based on each of the aforementioned tower nodes: If the predicted reduction rate of lightning strike risk of the tower node is within the preset high-efficiency range, then the evaluation result corresponding to the predicted reduction rate of lightning strike risk of the tower node is determined to be that the artificial intervention effect is significant. If the predicted reduction rate of lightning strike fault risk of the tower node is within the preset medium benefit range, then the evaluation result corresponding to the predicted reduction rate of lightning strike fault risk of the tower node is determined to be that the effect of artificial intervention is obvious. If the predicted reduction rate of lightning strike fault risk of the tower node is in the preset low-efficiency range, then the evaluation result corresponding to the predicted reduction rate of lightning strike fault risk of the tower node is determined to be that the effect of human intervention is limited. If the predicted reduction rate of lightning strike risk for the tower node is within a preset invalid interval, then the evaluation result corresponding to the predicted reduction rate of lightning strike risk for the tower node is determined to be that the effect of manual intervention is not significant.

9. A system for evaluating the effectiveness of artificial lightning suppression intervention, characterized in that, include: The data acquisition module is used to carry out artificial lightning suppression operations during thunderstorms and acquire graphical structure data of the target transmission lines and related power grid equipment in the power grid system before and after the artificial lightning suppression operations. The loss optimization training module is used to train a graph neural network model based on the graph structure data to predict the reduction rate of lightning strike failure risk for each tower node in the graph structure data, and to optimize it using the total loss function until the total loss function is minimized, thereby obtaining the trained graph neural network model; the total loss function includes the standard supervised loss function, physical consistency constraints, and physical equation residual constraints; The model prediction module is used to predict the lightning strike risk reduction rate of each tower node in the graph structure data based on the graph structure data and using a trained graph neural network model. The intervention effect evaluation module is used to evaluate the effect of artificial lightning suppression intervention based on the predicted value of the lightning fault risk reduction rate of each of the tower nodes.

10. The system as described in claim 9, characterized in that, The graph structure data includes multiple tower nodes, the electrical characteristics of each tower node, the physical connection edges and spatial proximity edges between every two tower nodes, and the true value of the lightning strike fault risk reduction rate corresponding to the marked tower nodes among the multiple tower nodes.

11. The system as described in claim 9, characterized in that, The loss optimization training module is specifically used for: The graph neural network model is used to predict the lightning strike risk reduction rate for each tower node. Based on multiple tower nodes, the predicted lightning strike risk reduction rate for each tower node, and the actual lightning strike risk reduction rate, the standard supervised loss value is calculated using the standard supervised loss function. Based on every two tower nodes, the physical connection edges and spatial proximity edges corresponding to every two tower nodes, the physical consistency constraint value is calculated using physical consistency constraints. Based on multiple tower nodes, the predicted lightning fault risk reduction rate for each tower node, and the electrical characteristics of the towers, physical equation residual constraints are used to calculate and obtain physical equation residual constraint values. The standard supervised loss value, the physical consistency constraint value, and the physical equation residual constraint value are weighted and summed to obtain the total loss function value. Based on the total loss function value, the graph neural network model is adjusted through backpropagation until the total loss function value is minimized, thus obtaining the trained graph neural network model.

12. The system as described in claim 9, characterized in that, Also includes: Active learning sample recommendation module; The active learning sample recommendation module is used to perform multiple independent forward propagations on several unlabeled pole nodes among multiple pole nodes using the Monte Carlo dropout method based on the graph structure data and the trained graph neural network model, to obtain the predicted value of each unlabeled pole node; based on the predicted values ​​of the several unlabeled pole nodes, determine the prediction variance of each unlabeled pole node in the graph structure data; sort the prediction variances of multiple unlabeled pole nodes in the graph structure data from largest to smallest, and select the several unlabeled pole nodes at the top of the sort to obtain the recommended labeled pole nodes for the graph structure data; Actual measurements were taken of the recommended labeled tower nodes in the graph structure data to obtain the measured values ​​of the lightning strike risk reduction rate of the recommended labeled tower nodes in the graph structure data; the graph structure data was updated using the measured values ​​of the lightning strike risk reduction rate of the recommended labeled tower nodes in the graph structure data; the updated graph structure data was used for the evaluation of the effect of the next manual lightning suppression intervention.

13. The system as described in claim 9, characterized in that, The physical consistency constraint is calculated using the following formula: in, This represents the physical consistency constraint value of the total loss function in the graph neural network model. For graph-structured data, it is the set of spatially nearest edges. Tower nodes predicted by a graph neural network model The rate of reduction in the risk of lightning strike failure; Tower nodes predicted by a graph neural network model The rate of reduction in the risk of lightning strike failure; This refers to the set of physical connection edges in graph-structured data. For smoothing coefficients; in, For physical similarity weights; Tower nodes in graph structure data Line characteristics; Tower nodes in graph structure data Line characteristics; This is the Gaussian kernel bandwidth parameter.

14. The system as described in claim 9, characterized in that, The residual constraint of the physical equation is calculated using the following formula: in, The physical equation residual constraint value for the total loss function in the graph neural network model; Tower nodes predicted by a graph neural network model The rate of reduction in the risk of lightning strike failure; This refers to all tower nodes in the graph structure data; in, Tower nodes in graph structure data The theoretical risk reduction rate; This refers to the ground flash density before artificial lightning suppression operations. This refers to the ground flash density after artificial lightning suppression operations. The distribution of lightning current amplitude before manual lightning suppression operations; The distribution of lightning current amplitude after artificial lightning suppression operations; Tower nodes in graph structure data The shielding failure probability function; This refers to the amplitude of the lightning current. This represents the lower limit of the lightning current withstand capability of the target transmission line in the graph structure data.

15. The system as described in claim 9, characterized in that, The total loss function value is calculated using the following formula: in, This represents the total loss function value of the graph neural network model. These are the weighting coefficients for physical consistency constraints. The value is inversely proportional to the number of tower nodes already marked in the graph structure data; These are the weighting coefficients for the residual constraints of the physical equations. The value is inversely proportional to the number of tower nodes already marked in the graph structure data; This is the standard supervised loss function for graph neural network models; Physical consistency constraints for the total loss function in a graph neural network model; The physical equation residual constraint value for the total loss function in the graph neural network model.

16. The system as described in claim 9, characterized in that, The intervention effect evaluation module is specifically used for: Based on each of the aforementioned tower nodes: if the predicted value of the lightning strike failure risk reduction rate of the tower node is within the preset high-efficiency range, then the evaluation result corresponding to the predicted value of the lightning strike failure risk reduction rate of the tower node is determined to be that the artificial intervention effect is significant. If the predicted reduction rate of lightning strike fault risk of the tower node is within the preset medium benefit range, then the evaluation result corresponding to the predicted reduction rate of lightning strike fault risk of the tower node is determined to be that the effect of artificial intervention is obvious. If the predicted reduction rate of lightning strike fault risk of the tower node is in the preset low-efficiency range, then the evaluation result corresponding to the predicted reduction rate of lightning strike fault risk of the tower node is determined to be that the effect of human intervention is limited. If the predicted reduction rate of lightning strike risk for the tower node is within a preset invalid interval, then the evaluation result corresponding to the predicted reduction rate of lightning strike risk for the tower node is determined to be that the effect of manual intervention is not significant.

17. A computer device, characterized in that, include: One or more processors; The processor is used to store one or more programs; When the one or more programs are executed by the one or more processors, an evaluation method for the effect of artificial lightning suppression intervention as described in any one of claims 1 to 8 is implemented.

18. A computer-readable storage medium, characterized in that, It contains a computer program, which, when executed, implements an evaluation method for the effect of artificial lightning suppression intervention as described in any one of claims 1 to 8.