Power grid hidden danger detection method, system and equipment based on graph fusion

Abstract

The invention discloses a power grid hidden danger detection method, system and device based on graph fusion, and relates to the technical field of power grid hidden danger detection, and the method comprises the following steps: constructing a power grid technology semantic graph and a power grid technology common graph; a comprehensive association graph is obtained by combining a power grid technical feature graph and a power grid equipment feature graph, the power grid technical feature graph is obtained according to a power grid technical semantic graph and a power grid technical common graph, and the power grid equipment feature graph is obtained according to power grid topology; performing primary clustering on the comprehensive association graph to obtain a hidden danger feature cluster, and mapping according to the hidden danger feature cluster to obtain a risk judgment condition cluster; and carrying out secondary clustering on the risk judgment condition cluster, and carrying out hidden danger detection on the target section in combination with a secondary clustering result and power grid operation data. The problem that an existing power grid hidden danger detection technology is poor in adaptability is solved.

Description

Technical Field

[0001] This invention relates to the field of power grid hazard detection technology, and more specifically, to a power grid hazard detection method, system, and equipment based on graph fusion. Background Technology

[0002] As a critical national infrastructure, the safe and stable operation of the power grid is crucial to the national economy, people's livelihood, and social stability. With the deepening of smart grid and digital construction, the scale of the power grid is constantly expanding, the types of equipment are becoming increasingly diverse, and the operational data is experiencing explosive growth. Traditional power grid hazard detection mainly relies on regular inspections, fixed threshold alarms, and expert experience judgment. These methods are gradually revealing their limitations, such as low efficiency and poor adaptability, when faced with massive amounts of heterogeneous data and complex relationships.

[0003] In recent years, the development of artificial intelligence technology, especially natural language processing and graph neural networks, has provided a new technical path for intelligent operation and maintenance of power grids. However, the ability to extract key information from unstructured text data such as equipment ledgers, inspection records, and fault reports, and to effectively integrate multi-source data to achieve accurate and adaptive hazard identification remains a significant challenge in the field.

[0004] Traditional methods for detecting potential power grid hazards largely rely on periodic manual inspections and alarm mechanisms based on fixed thresholds. Manual inspections suffer from low efficiency, high risk, and are heavily influenced by environmental factors and subjective experience, making comprehensive and timely coverage difficult. While simple threshold alarms can alert to significant voltage and current exceedances, they cannot effectively identify early, latent faults that have deviated from normal conditions but have not yet exceeded limits. These methods become increasingly limited when facing new power grids with increasingly complex structures and dynamically changing operating modes, failing to achieve comprehensive and timely monitoring coverage and lacking the ability to accurately predict and provide early warnings of potential risks. To address these challenges, power grid hazard detection technology is rapidly developing towards intelligence and digitalization. By introducing technologies such as big data and artificial intelligence, it is possible to deeply mine and analyze massive amounts of real-time operational data and historical fault records.

[0005] For example, the invention patent announcement CN113283704B discloses a knowledge graph-based intelligent power grid fault handling system. The system's fault handling plan parsing module acquires fault handling entity and relationship knowledge; the graph generation module stores the identified entities and entity relationships in triples to obtain equipment entity knowledge graphs, accident plan knowledge graphs, and handling process knowledge graphs; the fault perception module intelligently senses and collects information related to power grid faults within the power grid system; the fault risk assessment module identifies risks; and the intelligent fault handling module, after a line fault occurs in the power grid, infers suitable handling measures based on the accident plan knowledge graph and the handling process knowledge graph, combined with topology changes, power flow, and AC power supply frequency before and after the power grid fault.

[0006] For example, the invention patent announcement CN120355407B discloses a method and system for distribution network fault location based on intelligent decision-making. This method acquires a set of target state data collected by the intelligent sensing network covering the entire distribution network, performs spatiotemporal semantic fusion processing to generate a multi-dimensional feature map, uses a dynamic decision reasoning engine to perform context-aware discrimination on the multi-dimensional feature map, generates a fault risk discrimination result, performs causal link tracing on the target state data set based on the fault risk discrimination result, generates a fault section location result, generates a fault location command based on the fault section location result and sends it to the distribution network intelligent operation and maintenance platform to trigger a precise maintenance process.

[0007] The above-disclosed technical solutions have at least the following technical problems: Existing methods rely heavily on pre-set rules, resulting in a relatively rigid decision-making process that is difficult to cope with complex failure scenarios. Furthermore, they lack the ability to predict the evolution trend of potential hazards and cannot identify early risk signals.

[0008] To address the above problems, this invention proposes a solution. Summary of the Invention

[0009] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a power grid hazard detection method, system, and device based on graph fusion. By constructing a power grid technology feature graph and combining it with a power grid equipment feature graph to generate a comprehensive correlation graph, a hazard feature cluster is obtained through primary clustering and a fault condition cluster is obtained through secondary clustering. This achieves comprehensive capture of the deep correlation between power grid equipment and technology, thereby solving the problem of poor adaptability of existing power grid hazard detection technologies.

[0010] To achieve the above objectives, the present invention provides the following technical solution: The power grid hazard detection method based on graph fusion includes the following steps: constructing a power grid technology semantic graph and a power grid technology shared graph; combining the power grid technology feature graph and the power grid equipment feature graph to obtain a comprehensive correlation graph, wherein the power grid technology feature graph is obtained based on the power grid technology semantic graph and the power grid technology shared graph, and the power grid equipment feature graph is obtained based on the power grid topology; performing a first clustering on the comprehensive correlation graph to obtain hazard feature clusters, and mapping the hazard feature clusters to obtain risk judgment condition clusters; performing a second clustering on the risk judgment condition clusters, and combining the second clustering results with power grid operation data to detect hazards in the target section.

[0011] In a preferred embodiment, the construction of the power grid technology semantic graph and the power grid technology shared graph specifically involves: obtaining technical step fields based on the acquired power grid fault diagnosis technology method, and converting the technical step fields into power grid technology word vectors based on NLP technology; constructing a power grid technology semantic graph using the technical step fields as technical feature nodes and the semantic similarity between power grid technology word vectors as edges; and constructing a power grid technology shared graph by combining the co-occurrence frequency of any acquired technical step fields as edges and the technical feature nodes.

[0012] In a preferred embodiment, the method for constructing the power grid technology feature map and the power grid equipment feature map specifically includes: fusing the power grid technology semantic map and the power grid technology shared map to obtain the power grid technology feature map, which includes technology feature nodes and technology association edges; the technology association edges are obtained by fusing the edges of the semantic map and the shared map; the physical connection edges of the power grid equipment are obtained based on the equivalent impedance distance between the power grid equipment; based on a pre-built operation and maintenance function weight association library, the operation and maintenance function dependencies between the power grid equipment are mapped to function association edges; the physical connection edges and function association edges are weighted and fused to obtain equipment association edges; the power grid equipment is used as equipment feature nodes; and the power grid equipment feature map is obtained based on the equipment association edges and equipment feature nodes.

[0013] In a preferred embodiment, the step of combining the power grid technology feature map and the power grid equipment feature map to obtain the comprehensive association map specifically involves: based on the power grid fault diagnosis technology method, obtaining equipment technology edges according to the co-occurrence frequency of power grid equipment and technology step fields, and normalizing the technology association edges, equipment association edges, and equipment technology edges; taking the union of technology feature nodes and equipment feature nodes as equipment technology nodes; constructing a comprehensive association map based on equipment technology nodes, technology association edges, equipment association edges, and equipment technology edges; optimizing the comprehensive association map based on the similarity of equipment technology nodes; and updating the comprehensive association map according to the real-time state of the power grid using a pre-built DGNN model.

[0014] In a preferred embodiment, the optimization of the comprehensive association graph based on the device technology nodes specifically involves: converting the device technology nodes of the comprehensive association graph into node word vectors based on NLP technology; calculating the semantic similarity of the node word vectors; merging device technology nodes with semantic similarity higher than a preset similarity threshold; fusing the shared edges of the merged node pairs based on an attention mechanism; and inheriting the non-shared edges to the merged nodes.

[0015] In a preferred embodiment, the step of performing a single clustering of the comprehensive correlation graph to obtain a hazard feature cluster, and mapping the hazard feature cluster to obtain a risk judgment condition cluster, specifically involves: obtaining a graph Laplacian matrix from the comprehensive correlation graph and extracting matrix feature vectors using a stochastic SVD algorithm; clustering the matrix feature vectors using spectral clustering and fuzzy clustering methods, and introducing a meta-learning method to obtain a hazard feature cluster; obtaining conditional semantic vectors based on a pre-constructed risk judgment condition library; calculating the correlation between the conditional semantic vectors and node word vectors, and filtering risk judgment conditions that exceed a preset correlation threshold to obtain a risk judgment condition cluster.

[0016] In a preferred embodiment, the secondary clustering of risk judgment condition clusters, combined with the secondary clustering results and power grid operation data, is used to detect potential hazards in the target section. Specifically, this involves: obtaining the early warning utility vector of each judgment condition for each type of fault based on the accuracy, recall, and F1 score of each judgment condition in the risk judgment condition cluster; performing secondary clustering on the risk judgment condition clusters based on the early warning utility vectors, and verifying and filtering the secondary clustering results using a causal inference method to obtain fault condition clusters corresponding to each type of fault, wherein each fault condition cluster includes several risk judgment conditions; calculating several risk judgment values ​​in each fault condition cluster of the target section, performing normalization processing, and taking the average value to obtain the fault condition cluster value; obtaining the weight of the corresponding fault condition cluster based on the frequency proportion of each historical fault type in the target section in the historical power grid data; and weighting and fusing the fault condition cluster values ​​according to the weights of the fault condition clusters to obtain a comprehensive hazard value, which is then combined with power grid operation data for hazard detection.

[0017] In a preferred embodiment, the verification and screening of the secondary clustering results using the causal inference method specifically involves: obtaining a causal dataset using the PSM method based on historical power grid data and the results of the secondary clustering; obtaining the causal effect of the secondary clustering results based on machine learning according to the causal dataset; and verifying and screening the effectiveness of the secondary clustering results based on the causal effect.

[0018] The power grid hazard detection system based on graph fusion includes: a basic graph and vector construction module for constructing a power grid technology semantic graph and a power grid technology common graph; a graph fusion module for combining the power grid technology feature graph and the power grid equipment feature graph to obtain a comprehensive correlation graph; a primary clustering module for performing primary clustering on the comprehensive correlation graph to obtain hazard feature clusters, and mapping the hazard feature clusters to obtain risk judgment condition clusters; and a secondary clustering module for performing secondary clustering on the risk judgment condition clusters, and combining the secondary clustering results with power grid operation data to detect hazards in the target section.

[0019] An electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to execute the aforementioned graph fusion-based power grid hazard detection method.

[0020] The technical effects and advantages of the graph fusion-based power grid hazard detection method, system, and equipment of this invention are as follows: This invention constructs a power grid technology feature map and a power grid equipment feature map, and then merges them to generate a comprehensive correlation map, thereby achieving a comprehensive capture of the multi-level relationships between power grid equipment, technical methods, and fault phenomena. By performing a first clustering of the comprehensive correlation map to extract hidden danger feature clusters and mapping them to risk judgment condition clusters, it achieves the capture of global features of the graph structure. By performing a second clustering of the judgment condition clusters and combining them with real-time power grid operation data, it achieves hidden danger detection in target sections, improving the accuracy and coverage of hidden danger identification and effectively solving the problem of poor adaptability of existing power grid hidden danger detection technologies. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the power grid hidden danger detection method based on graph fusion provided in an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of the power grid hidden danger detection system based on graph fusion provided in an embodiment of the present invention.

[0023] Figure 3 This is a schematic diagram of the power grid hazard detection device based on graph fusion provided in an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0025] Example 1, Figure 1 The present invention provides a power grid hazard detection method based on graph fusion, comprising the following steps: S1, construct the power grid technology semantic graph and the power grid technology shared graph; S2, A comprehensive correlation diagram is obtained by combining the power grid technology feature diagram and the power grid equipment feature diagram. The power grid technology feature diagram is obtained based on the power grid technology semantic diagram and the power grid technology common diagram, and the power grid equipment feature diagram is obtained based on the power grid topology. S3, perform a clustering operation on the comprehensive correlation graph to obtain the hidden danger feature cluster, and map the risk judgment condition cluster based on the hidden danger feature cluster; S4 performs secondary clustering on the risk assessment condition clusters, and combines the secondary clustering results with power grid operation data to detect potential hazards in the target section.

[0026] This embodiment constructs a power grid technology feature map and a power grid equipment feature map, and then merges them to generate a comprehensive correlation map, thereby achieving a comprehensive capture of the multi-level relationships between power grid equipment, technical methods, and fault phenomena. By performing a first clustering on the comprehensive correlation map to extract hidden danger feature clusters and mapping them to risk judgment condition clusters, it achieves the capture of global features of the graph structure. By performing a second clustering on the judgment condition clusters and combining them with real-time power grid operation data, it achieves hidden danger detection in target sections, improving the accuracy and coverage of hidden danger identification and effectively solving the problem of poor adaptability of existing power grid hidden danger detection technologies.

[0027] S1, construct the power grid technology semantic graph and the power grid technology shared graph.

[0028] In this embodiment, S1 specifically refers to: The technical step field is obtained based on the acquired power grid fault diagnosis technology method, and the technical step field is converted into power grid technology word vectors based on NLP technology. Using the technical step field as the technical feature node and the semantic similarity between power grid technology word vectors as the edge, a power grid technology semantic graph is constructed. Using the co-occurrence frequency of any obtained technical step field as an edge, and combining it with technical feature nodes, a shared graph of power grid technologies is constructed.

[0029] In this embodiment, the conversion of technical step fields into power grid technical word vectors based on NLP technology specifically involves using a pre-trained BERT model for the conversion.

[0030] It should be noted that BERT is a pre-trained language model based on the Transformer architecture. It significantly improves the ability of natural language processing through bidirectional contextual understanding. It learns rich language representations from large-scale unlabeled text by using two pre-training tasks: masked language model and next sentence prediction. Through this design, the word vectors generated by BERT can dynamically adapt to different contexts and effectively solve the problem of polysemy.

[0031] It should be noted that this embodiment transforms the technical step field into word vectors to deeply capture the subtle semantic differences of the technical step field in the context of power grid professional terminology. At the same time, it uses a shared graph to reflect the collaborative mode of technology in practical applications by using co-occurrence frequency, transforming unstructured text data into a standardized graph representation, providing a high-quality semantic foundation for subsequent analysis, and improving the computability and scalability of the data.

[0032] S2, A comprehensive correlation diagram is obtained by combining the power grid technology feature diagram and the power grid equipment feature diagram. The power grid technology feature diagram is obtained based on the power grid technology semantic diagram and the power grid technology common diagram. The power grid equipment feature diagram is obtained based on the power grid topology. The power grid technology feature diagram is obtained based on the power grid technology semantic diagram and the power grid technology common diagram.

[0033] In this embodiment, the method for constructing the power grid technology feature map and the power grid equipment feature map is specifically as follows: A power grid technology feature graph is obtained by integrating the power grid technology semantic graph and the power grid technology shared graph. The power grid technology feature graph includes technology feature nodes and technology association edges. The technical association edges are obtained by fusing the edges of the semantic graph and the shared graph; The physical connection edges of the power grid equipment are obtained based on the equivalent impedance distance between the power grid equipment. Based on the pre-built operation and maintenance function weight association library, the operation and maintenance function dependencies between the power grid equipment are mapped to function association edges. The physical connection edges and functional association edges are weighted and fused to obtain the device association edges. The power grid equipment is used as the device feature nodes, and the power grid equipment feature graph is obtained based on the device association edges and device feature nodes.

[0034] In this embodiment, the specific formula for obtaining the physical connection edge of the power grid equipment based on the equivalent impedance distance between the power grid equipment is as follows:

[0035] In the formula, For equipment and equipment impedance distance, The preset smoothing parameters, This represents the weight of the physically connected edges.

[0036] In this embodiment, the operation and maintenance function weight association library specifically divides the operation and maintenance function dependencies between devices into three categories and sets basic weights. The three categories of operation and maintenance function dependencies include dynamic protection relationships, data monitoring relationships, and logical control relationships. The basic weights are modified according to the accuracy of protection actions, the integrity of monitoring data, and the success rate of control signal execution.

[0037] In this embodiment, the process of obtaining a comprehensive correlation diagram by combining the power grid technology feature map and the power grid equipment feature map specifically involves: Based on the power grid fault diagnosis technology method, the equipment technology edge is obtained according to the co-occurrence frequency of power grid equipment and technology step fields, and the technology association edge, equipment association edge and equipment technology edge are normalized. The union of technical feature nodes and device feature nodes is taken as the device technical node; Construct a comprehensive association graph based on equipment technology nodes, technology association edges, equipment association edges, and equipment technology edges; The comprehensive correlation graph is optimized based on the similarity of equipment technical nodes; The integrated correlation diagram is updated based on the real-time status of the power grid using a pre-built DGNN model.

[0038] In this embodiment, the optimization of the integrated correlation diagram based on the device technology nodes specifically involves: Based on NLP technology, the device technology nodes of the comprehensive association graph are converted into node word vectors; Semantic similarity is calculated on the word vectors of nodes, and device technology nodes with semantic similarity higher than a preset similarity threshold are merged. The shared edges of the merged node pairs are merged using an attention mechanism, and the non-shared edges are inherited to the merged node.

[0039] In this embodiment, the specific formula for obtaining the equipment technology edge based on the co-occurrence frequency of the power grid equipment and technology step fields is as follows:

[0040] In the formula, For equipment, For diagnostic techniques and methods, The initial value for the equipment technology edge. The total number of devices in the power grid. For those who have used diagnostic techniques The number of devices, Diagnostic techniques for equipment Number of times, equipment The maximum frequency of diagnostic techniques used.

[0041] In this embodiment, the shared edge of the merged node pair refers to the two edges corresponding to the shared node that is connected to both merged nodes, and the non-shared edge refers to the edge of the non-shared node that is connected to only one merged node.

[0042] In this embodiment, the fusion of the shared edges of the merged node pairs based on the attention mechanism is specifically implemented through a shallow neural network.

[0043] In this embodiment, the step of updating the integrated correlation graph based on the real-time state of the power grid using a pre-built DGNN model specifically involves: Periodically acquire the real-time status of the power grid, convert the real-time status of the power grid into a comprehensive correlation diagram, and convert the comprehensive correlation diagrams of different periods into a sequence of graph snapshots; The DGNN model is trained based on the graph snapshot sequence to learn the evolutionary features of the integrated association graph; The edge weights and device technology nodes of the integrated association graph are updated using the trained DGNN model.

[0044] In this embodiment, the sequence learning structure of the DGNN model adopts the GRU structure.

[0045] It should be noted that the DGNN model is a dynamic graph neural network model, which is a neural network model specifically designed to process graph structure data that changes over time. By combining graph neural networks with sequence learning techniques, it can not only capture the spatial topological relationships of nodes and edges in the graph at each time step, but also model the dynamic patterns of graph structures that change over time, enabling more accurate modeling of the dynamic nature of networks in the real world.

[0046] It should be noted that this embodiment obtains a comprehensive association graph by quantifying physical connection edges and functional association edges, and by weightedly fusing technical semantic association, shared relationships, and equipment physical topology and functional dependencies. Then, by introducing node semantic similarity optimization and attention mechanism edge fusion, the unification of multi-dimensional information at the technical and equipment levels is achieved, overcoming the limitations of a single data source. While maintaining key information, it improves computational efficiency and enables the comprehensive association graph to dynamically adapt to real-time changes in the power grid status.

[0047] S3. Perform a clustering operation on the comprehensive correlation graph to obtain the hidden danger feature cluster, and map the risk judgment condition cluster based on the hidden danger feature cluster.

[0048] In this embodiment, S3 specifically refers to: The graph Laplacian matrix is ​​obtained from the comprehensive correlation graph, and the matrix eigenvectors are extracted using the stochastic SVD algorithm. By combining spectral clustering and fuzzy clustering methods to cluster matrix eigenvectors, and introducing meta-learning methods, potential hazard feature clusters are obtained. The conditional semantic vector is obtained based on the pre-built risk assessment condition library; The correlation between the conditional semantic vector and the node word vector is calculated, and risk judgment conditions that exceed the preset correlation threshold are filtered to obtain a risk judgment condition cluster.

[0049] In this embodiment, the specific formula for the graph Laplacian matrix is ​​as follows:

[0050] In the formula, For degree matrix, For Laplace matrix, It is an identity matrix.

[0051] In this embodiment, the degree matrix is ​​specifically formulated as follows:

[0052] In the formula, For equipment technology nodes and Edge weights between them.

[0053] In this embodiment, the risk assessment condition library includes the name, description, and calculation method of the risk assessment condition.

[0054] In this embodiment, the correlation between the conditional semantic vector and the node word vector is specifically obtained by the Mahalanobis distance between the conditional semantic vector and the node word vector.

[0055] It should be noted that stochastic SVD is an efficient algorithm for providing low-rank approximations of large-scale matrices. It reduces computational complexity by introducing randomness. The core idea is to use a random matrix to project the original high-dimensional data into a low-dimensional subspace, then perform accurate SVD decomposition in this smaller space, and finally reconstruct the result back into the original space to obtain an approximate decomposition of the original matrix. This allows it to significantly reduce computation time and memory consumption while maintaining approximate accuracy when processing large-scale matrices, effectively avoiding the memory overflow problem of traditional methods.

[0056] In this embodiment, the method of combining spectral clustering and fuzzy clustering to cluster matrix feature vectors and introducing a meta-learning method is as follows: Based on the features of the comprehensive association graph, the clustering parameters are optimized using a pre-trained meta-learner; The clustering parameters include the number of clusters and the ambiguity parameter; Based on the clustering parameters, the feature vectors are clustered using a combination of spectral clustering and fuzzy clustering methods.

[0057] In this embodiment, the meta-learner uses MAML.

[0058] It should be noted that meta-learning is a subfield of machine learning. Its core goal is to enable machines to learn how to learn, that is, to enable models to quickly adapt to new tasks by acquiring experience from multiple related tasks. Unlike traditional machine learning models that are trained for a single task, meta-learning generalizes at the task level. Its training process involves many small tasks, each of which simulates a few-shot learning scenario, so that the model can master a general learning paradigm. By optimizing the initial parameters of the model, it can quickly adapt to new tasks through a small number of gradient updates.

[0059] It should be noted that MAML stands for Model Independent Meta-Learning. Its core idea is to find task-sensitive initial parameters for the model so that when faced with a new task, the model can quickly achieve excellent performance with only a small number of samples from the new task and a few gradient update steps.

[0060] It should be noted that this embodiment extracts the feature vectors of the graph Laplacian matrix using the random SVD algorithm, combines the advantages of spectral clustering and fuzzy clustering, and introduces meta-learning to optimize the clustering parameters to cluster the matrix feature vectors to obtain the hidden danger feature clusters. Based on the correlation between the conditional semantic vectors of the pre-constructed risk judgment condition library and the node word vectors, the risk judgment condition clusters corresponding to the hidden danger feature clusters are obtained. This achieves the capture of global features of the graph structure, and can also handle the uncertainty of node affiliation, reduce manual intervention, and improve the accuracy and generalization of risk judgment condition cluster division.

[0061] S4 performs secondary clustering on the risk assessment condition clusters, and combines the secondary clustering results with power grid operation data to detect potential hazards in the target section.

[0062] In this embodiment, S4 specifically refers to: Based on the accuracy, recall, and F1 score of each judgment condition in the risk judgment condition cluster for each type of fault, the warning utility vector of each judgment condition for each type of fault is obtained. The risk judgment condition clusters are clustered twice based on the early warning utility vector. The results of the second clustering are verified and filtered using the causal inference method to obtain the fault condition clusters corresponding to each type of fault. The fault condition clusters include several risk judgment conditions. Calculate several risk judgment values ​​in each fault condition cluster of the target section, normalize them, and take the average value to obtain the fault condition cluster value; The weights of the corresponding fault condition clusters are obtained based on the frequency proportion of each historical fault type in the target section in the historical power grid data. The values ​​of fault condition clusters are weighted and fused according to their respective weights to obtain a comprehensive hidden danger value, which is then combined with power grid operation data for hidden danger detection.

[0063] It's important to note that the F1 score combines information from precision and recall; it's the harmonic mean of the two. Precision measures the proportion of truly positive samples out of those that are correctly identified, focusing on the accuracy of the prediction. Recall measures the proportion of correctly identified true positive samples out of all true positive samples. F1 provides a more comprehensive reflection of the overall performance of the judgment condition in imbalanced datasets or scenarios with varying tolerances for different types of errors than precision.

[0064] In this embodiment, the weight of the corresponding fault condition cluster is obtained based on the frequency proportion of each historical fault type in the target section in the historical power grid data. The specific formula is as follows:

[0065] In the formula, For the first The frequency percentage of each type of failure For the first The number of times each type of failure occurs. This represents the total number of occurrences for all fault types.

[0066] In this embodiment, the verification and screening of the secondary clustering results using the causal inference method specifically involves: Based on historical power grid data and the results of secondary clustering, the causal dataset was obtained by using the PSM method. Based on the causal dataset, the causal effect of the secondary clustering results is obtained using machine learning; The effectiveness of the secondary clustering results is verified and screened based on causal effects.

[0067] In this embodiment, the causal dataset is obtained by partitioning using the PSM method, specifically as follows: The causal dataset includes a treatment group, a control group, treatment variables, outcome variables, and confounding variables; Set a specified time period, and determine whether to trigger an alert based on whether the conditions within the specified time period exceed a preset threshold; The control group consists of a sample of all devices that trigger warnings for a specific cluster of fault conditions within a pre-defined time period; The control group consists of equipment samples that did not trigger the fault condition cluster warning during a preset time period, but are similar to the treatment group samples in other characteristics; The processing variable is whether it belongs to a specific fault condition cluster and triggers an early warning; The result variable is whether the corresponding type of fault has occurred; Confounding variables include data from when the device is running.

[0068] In this embodiment, the causal effect obtained from the secondary clustering results based on the causal dataset using machine learning specifically includes: Train two machine learning models, with confounding variables as inputs and processing and outcome variables as outputs, respectively. Obtain the residuals of the two machine learning models, and then obtain the causal effect estimate through residual regression; The confidence interval is obtained from the causal effect estimate and then a significance test is performed.

[0069] In this embodiment, the residual is specifically formulated as follows:

[0070] In the formula, and These are the residuals of the two models, To truly handle variables, For the actual outcome variable, and These are the predicted values ​​from the two models, respectively.

[0071] In this embodiment, the causal effect is specifically formulated as follows:

[0072] In the formula, For causal effect, This is the error term.

[0073] In this embodiment, the estimation of causal effects uses ordinary least squares, and the specific formula is as follows:

[0074] In the formula, and No. One residual.

[0075] Through mathematical derivation, we obtain:

[0076] In the formula, This is an estimate of the causal effect. This is the matrix transpose.

[0077] It should be noted that obtaining confidence intervals based on causal effect estimates and performing significance tests is existing technology and will not be elaborated here.

[0078] It should be noted that PSM stands for Propensity Score Matching, a statistical method used to process observational research data. Its core objective is to reduce selection bias caused by confounding variables in the data by finding one or more individuals with similar background characteristics in the control group for each individual in the experimental group, thereby making the estimation of treatment effects closer to the reliability of randomized controlled trials.

[0079] It should be noted that in this embodiment, risk judgment condition clusters are clustered according to the constructed early warning utility vector, and a machine learning model is introduced to verify causal inference to obtain fault condition clusters. The frequency ratio of historical fault types is converted into weights, and hidden danger detection is carried out by combining real-time power grid operation data and fault condition clusters. This enables the detection system to automatically adapt to the specific risk characteristics of different target sections. When detecting hidden dangers, it can reflect the global risk pattern while taking into account the characteristics of local areas.

[0080] Example 2, Figure 2 The present invention provides a power grid hazard detection system based on graph fusion, comprising: The basic graph and vector construction module is used to construct the power grid technology semantic graph and the power grid technology common graph. The graph fusion module is used to combine power grid technology feature maps and power grid equipment feature maps to obtain a comprehensive correlation map. The primary clustering module is used to perform primary clustering on the comprehensive correlation graph to obtain hazard feature clusters, and to map risk judgment condition clusters based on the hazard feature clusters. The secondary clustering module is used to perform secondary clustering on the risk assessment condition clusters, and combine the secondary clustering results with power grid operation data to detect hidden dangers in the target section.

[0081] Example 3: This example provides a computer electronic device, such as... Figure 3 As shown, the electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to execute the aforementioned graph fusion-based power grid hazard detection method.

[0082] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0083] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0084] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0085] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0086] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0087] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A power grid hidden danger detection method based on graph fusion, characterized in that, The method comprises the following steps: constructing a power grid technology semantic graph and a power grid technology common graph; obtaining a comprehensive correlation graph by combining a power grid technology feature graph and a power grid equipment feature graph, wherein the power grid technology feature graph is obtained according to the power grid technology semantic graph and the power grid technology common graph, and the power grid equipment feature graph is obtained according to a power grid topology; performing primary clustering on the comprehensive correlation graph to obtain a hidden danger feature cluster, and mapping the hidden danger feature cluster to obtain a risk judgment condition cluster; performing secondary clustering on the risk judgment condition cluster, and combining the secondary clustering result and power grid operation data to detect hidden dangers in a target section.

2. The graph fusion based power grid hazard detection method of claim 1, wherein, The construction of the power grid technology semantic graph and the power grid technology common graph is specifically as follows: obtaining a technology step field according to an obtained power grid fault diagnosis technology method, and converting the technology step field into a power grid technology word vector based on NLP technology; constructing a power grid technology semantic graph by taking the technology step field as a technology feature node and taking the semantic similarity between the power grid technology word vectors as an edge; constructing a power grid technology common graph by taking the occurrence frequency of any technology step field as an edge and combining the technology feature node.

3. The graph fusion based power grid hazard detection method of claim 2, wherein, The power grid technology feature graph and the power grid equipment feature graph construction method is specifically as follows: obtaining a power grid technology feature graph by fusing the power grid technology semantic graph and the power grid technology common graph, wherein the power grid technology feature graph comprises a technology feature node and a technology correlation edge; the technology correlation edge is obtained by fusing the edges of the semantic graph and the common graph; obtaining a physical connection edge of the power grid equipment according to the equivalent impedance distance between the power grid equipments, and mapping the operation and maintenance function dependency relationship between the power grid equipments into a function correlation edge based on a pre-constructed operation and maintenance function weight association library; obtaining a device correlation edge by fusing the physical connection edge and the function correlation edge, taking the power grid equipment as a device feature node, and obtaining a power grid equipment feature graph according to the device correlation edge and the device feature node.

4. The graph fusion based power grid hazard detection method of claim 3, wherein, The comprehensive correlation graph is obtained by combining the power grid technology feature graph and the power grid equipment feature graph, specifically as follows: obtaining a device technology edge according to the occurrence frequency of the power grid equipment and the technology step field based on the power grid fault diagnosis technology method, and normalizing the technology correlation edge, the device correlation edge and the device technology edge; taking the union of the technology feature node and the device feature node as a device technology node; constructing a comprehensive correlation graph according to the device technology node, the technology correlation edge, the device correlation edge and the device technology edge; optimizing the comprehensive correlation graph according to the similarity of the device technology node; updating the comprehensive correlation graph according to the real-time state of the power grid through a pre-constructed DGNN model.

5. The graph fusion based power grid hazard detection method of claim 4, wherein, The optimization of the comprehensive correlation graph according to the device technology node is specifically as follows: converting the device technology node of the comprehensive correlation graph into a node word vector based on NLP technology; performing semantic similarity calculation on the node word vector, and merging the device technology nodes with a semantic similarity higher than a preset similarity threshold; fusing the common edges of the merged nodes based on an attention mechanism, and inheriting the non-common edges to the merged nodes.

6. The graph fusion based power grid hazard detection method of claim 5, wherein, The hidden danger feature cluster is obtained by performing primary clustering on the comprehensive correlation graph, and the risk judgment condition cluster is obtained by mapping the hidden danger feature cluster, specifically as follows: obtaining a graph Laplacian matrix from the comprehensive correlation graph, and extracting a matrix feature vector by using a random SVD algorithm; The matrix characteristic vectors are clustered by combining a spectral clustering method and a fuzzy clustering method, and a meta-learning method is introduced to obtain hazard feature clusters; A condition semantic vector is obtained according to a pre-constructed risk judgment condition library; The relevance of the condition semantic vector and the node word vector is calculated, risk judgment conditions higher than a preset relevance threshold are screened, and a risk judgment condition cluster is obtained.

7. The graph fusion based power grid hazard detection method of claim 6, wherein, The risk judgment condition cluster is subjected to secondary clustering, and the target section is subjected to hazard detection by combining the secondary clustering result and power grid operation data, specifically as follows: According to the accuracy, recall rate and F1 score of each judgment condition in the risk judgment condition cluster of each type of fault, an early warning utility vector of each judgment condition for each type of fault is obtained; The risk judgment condition cluster is subjected to secondary clustering according to the early warning utility vector, and the secondary clustering result is verified and screened by a causal inference method to obtain a fault condition cluster corresponding to each type of fault, wherein the fault condition cluster includes a plurality of risk judgment conditions; A plurality of risk judgment values in each fault condition cluster of the target section are calculated, normalized and averaged to obtain a fault condition cluster value; According to the frequency proportion of each historical fault type of the target section in the power grid historical data, a weight of the corresponding fault condition cluster is obtained; The fault condition cluster value is weighted and fused according to the weight of the fault condition cluster to obtain a comprehensive hazard value, and hazard detection is performed in combination with the power grid operation data.

8. The graph fusion based power grid hazard detection method of claim 7, wherein, The secondary clustering result is verified and screened by a causal inference method, specifically as follows: Based on the power grid historical data and the secondary clustering result, a causal data set is divided by a PSM method; According to the causal data set, a causal effect of the secondary clustering result is obtained based on machine learning; The effectiveness of the secondary clustering result is verified and screened according to the causal effect.

9. A system using the power grid hazard detection method based on graph fusion according to any one of claims 1-8, comprising: a basic graph and vector construction module for constructing a power grid technical semantic graph and a power grid technical common graph; a graph fusion module for obtaining a comprehensive correlation graph by combining a power grid technical feature graph and a power grid equipment feature graph; a primary clustering module for obtaining hazard feature clusters by primary clustering the comprehensive correlation graph, and obtaining risk judgment condition clusters according to the hazard feature clusters; a secondary clustering module for secondary clustering the risk judgment condition clusters, and performing hazard detection on a target section by combining the secondary clustering result and power grid operation data.

10. An electronic device, comprising: The electronic device comprises: at least one processor; and a memory connected in communication with the at least one processor; wherein the memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor to enable the at least one processor to execute the power grid hazard detection method based on graph fusion according to any one of claims 1-8.

Images