Power distribution network topology correction method fusing topology anomaly recognition and structure credible reasoning

Abstract

The invention relates to a power distribution network topology correction method fusing topology anomaly recognition and structure credible reasoning. The method comprises the following steps: recognizing an abnormal region of a topological structure, and recognizing and generating candidate edges and unobserved nodes; constructing a structure behavior joint feature tensor considering behavior collaboration between nodes, adjacent structure similarity and sudden change synchronism, and inputting the structure behavior joint feature tensor into a topological connection credibility scoring model to obtain an edge credibility score; and carrying out credibility correction on the initial topological structure by combining the edge credibility score obtained by the topological connection credibility scoring model, the candidate edge and the unobserved node, and generating a final topological structure. Compared with the prior art, the method has the advantages that omission edges, error edges and unmodeled nodes in the original topological structure are comprehensively repaired, and the integrity and credibility of the topological structure of the power distribution network are remarkably improved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of power distribution network topology identification, in particular to a power distribution network topology correction method fusing topology abnormality identification and structure reliable inference. BACKGROUND

[0002] With the large number of access of distributed energy and the continuous improvement of power distribution automation level, the structure of the power distribution network is becoming more and more complex, showing the characteristics of many points, long lines and flexible structure. The traditional topology identification method relying on drawings and manual checking has been difficult to meet the real-time and accuracy requirements. In recent years, the application of graph neural network (GNN) in topology identification has made certain progress, but still faces two key challenges: one is that part of the edge-node connection relationship is limited by imperfect collection device deployment, and there are data abnormalities or missing problems; the other is that there are a large number of transient disturbances and data fluctuations under complex working conditions, and the traditional identification method based on single time or static graph structure is easy to misjudge.

[0003] Meanwhile, in the process of power distribution network topology identification, the attributes and connection relationships of nodes (such as ring network cabinets, switch stations, etc.) are partially observable and partially unobservable. How to infer the unknown node relationship under the condition that part of the node information is known is the key to improve the accuracy and robustness of topology identification. Therefore, a new auxiliary verification method fusing time sequence information mining and graph structure propagation mechanism is urgently needed to effectively supplement and correct the results of the main identification model such as graph neural network. SUMMARY

[0004] The purpose of the application is to provide a power distribution network topology correction method fusing topology abnormality identification and structure reliable inference.

[0005] The purpose of the application can be achieved by the following technical solutions: A power distribution network topology correction method fusing topology abnormality identification and structure reliable inference, the method steps comprising: Identifying the abnormal area of the topology structure, and identifying the candidate edges and unobserved nodes; Constructing a structure behavior joint feature tensor considering the behavior cooperation, adjacent structure similarity and mutation synchronism between nodes, and inputting the topology connection reliability scoring model to obtain the edge reliability score; Combining the edge reliability score obtained by the topology connection reliability scoring model and the candidate edges and unobserved nodes, performing reliable correction on the initial topology structure, and generating the final topology structure.

[0006] As a preferred technical solution, the identification of the abnormal area of the topology structure and the identification of the candidate edges and unobserved nodes are as follows: The power residual and voltage residual of each node in the calculation topology are calculated, the mean and standard deviation of the residual time sequence are calculated in a sliding window, and dynamic stability is judged; Nodes with residual mean and standard deviation greater than the set threshold in the sliding window interval are determined as abnormal nodes and added to the abnormal node set; The similarity matrix is constructed using dynamic time warping distance for abnormal nodes, and the potential connection edge between two nodes is obtained through the spectral clustering algorithm; The graph structure completion is modeled as a low-rank matrix completion problem, and the singular value thresholding algorithm is used for iterative solution to obtain the completed adjacency matrix; Determine that the non-zero elements higher than the threshold in the completed adjacency matrix are potential connections, i.e. candidate edges; if the nodes corresponding to a row or column of the completed adjacency matrix are all zero or approximately zero, the node is determined as a potential unobserved node.

[0007] As a preferred technical solution, the spectral clustering algorithm is used to obtain the potential connection edge between two nodes, which is as follows: The node behavior characteristics are constructed based on the node current, voltage and power factor, and the dynamic time warping distance between nodes is calculated based on the node behavior characteristics; The similarity matrix is calculated based on the dynamic time warping distance between nodes, and a Laplacian matrix is further constructed; The eigenvectors corresponding to the first k The minimum non-zero eigenvalue are solved to form a spectral embedding matrix; Each row of the spectral embedding matrix is clustered, and nodes with similar characteristics are divided into the same cluster; For any two nodes, if they belong to the same cluster but are not connected in the existing topology, it is determined that there may be a potential connection edge between the two nodes, and the candidate edge set is added.

[0008] As a preferred technical solution, the graph structure completion is modeled as a low-rank matrix completion problem, which is represented as:

[0009] In the formula, is the original observed adjacency matrix; is the adjacency matrix to be completed; is the observation position projection operator; is the observation set; is the Frobenius norm; is the kernel norm; is the regularization coefficient.

[0010] As a preferred technical solution, the topological connection credibility scoring model models the connection credibility of the candidate edge by using a graph attention network: the structural and behavioral joint feature tensor is projected into a high-dimensional space by using a linear transformation to obtain an embedded feature representation of the edge; the embedded vector is input into an attention mechanism for scoring, and then the attention score is input into an activation function to obtain the final edge credibility.

[0011] As a preferred technical solution, the structural and behavioral joint feature tensor specifically includes: current correlation, voltage correlation, mutation synchronization degree, mirror change rate, and adjacent coincidence rate between each pair of candidate nodes.

[0012] As a preferred technical solution, the mutation synchronization degree is calculated as follows: For each node, a set of mutation events is constructed If the current or voltage change at any time t is greater than a set threshold, the time is added to the set of mutation events. For two nodes, the sets of mutation events , Matching is performed, and if the first time There is a second time So that the absolute value of the difference between the first time And the second time Is less than or equal to the tolerance synchronization error, it is considered to be a pair of synchronous mutations, and the mutation synchronization degree is defined as:

[0013] In the formula, Indicates the set of mutation events of node i,j ; Is a constant used to prevent division by zero.

[0014] As a preferred technical solution, the mirror change rate is used to detect whether the current mutation of node i Corresponds to the opposite change of the voltage direction of node j , and is calculated as follows: Get the set of mutation times of the detection node, and for each mutation point, determine whether the current and voltage change directions of the node are opposite, and count the number of times that meet the condition and divide by the total number of mutations:

[0015] In the formula, Is the set of mutation times of node i ; And Are the current and voltage change amounts of node i At time t ; Is a sign function, i.e., the sign of the return value.

[0016] As a preferred technical solution, the edge credibility score obtained by combining the topology connection credibility scoring model is used to modify the credibility of the initial topology structure, specifically as follows: Based on the scoring results of each candidate edge, the candidate edges with a scoring result lower than the set scoring threshold are removed; The nodes that are still completely isolated after edge modification are screened for structure matching prediction, and each unobserved node is combined with all nodes in the original topology structure in the current topology j For each candidate edge The topology connection credibility scoring model is used to calculate the scoring result, and the connection relationship is judged; Based on the connection relationship judgment result, the edge structure is adjusted and completed to obtain the modified topology graph.

[0017] As a preferred technical solution, the connection relationship judgment is specifically as follows: For each unobserved node, if the maximum candidate edge scoring result is less than the first scoring threshold, the unobserved node is not connected; If there is a node such that the candidate edge scoring result is greater than or equal to the second scoring threshold, and the number of nodes satisfying the condition is less than the set number, then connect the nodes satisfying the condition; If there is a node such that the candidate edge scoring result is greater than or equal to the second scoring threshold, and the number of nodes satisfying the condition is greater than or equal to the set number, then connect the top set number of nodes with the highest score; If there is no node with a candidate edge scoring result greater than or equal to the second scoring threshold, but there is a node such that the candidate edge scoring result is between the first scoring threshold and the second scoring threshold, then connect the node with the highest score.

[0018] Compared with the prior art, the present application has the following beneficial effects: 1) The present application proposes a power distribution network topology modification method that combines topology anomaly recognition and structure credibility reasoning. By introducing a scoring-driven edge credibility judgment mechanism and combining a structured access strategy for unobserved nodes, the present application realizes accurate modification and closed-loop completion of the power distribution network topology structure. While ensuring low computational resource consumption, the present application achieves comprehensive repair of missing edges, incorrect edges, and unmodeled nodes in the original topology structure, significantly improving the integrity and credibility of the power distribution network topology structure.

[0019] 2) The application identifies suspected unobserved or unmodeled nodes in the main model by methods such as graph structure completion and spectral clustering. Abnormal nodes are screened based on the residual indicators of power and voltage, a similarity matrix is constructed, and spectral clustering algorithm is used to find nodes with similar behavior but unconnected structure to mine the implicit connections that may exist in the abnormal area. Finally, the topology of the reconstructed graph is modeled as a low-rank matrix completion problem, and the singular value thresholding algorithm is used for iterative solution, which can locate the abnormality from the power flow and voltage behavior and generate the edge set to be completed and the possible missed nodes. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 A flowchart of the topology auxiliary verification and correction method based on operation data driving of the application; Figure 2 A flowchart of topology structure abnormality identification and candidate information extraction of the application; Figure 3 A flowchart of topology connection credibility scoring model construction of the application; Figure 4 A flowchart of structure correction mechanism and topology output of the application. DETAILED DESCRIPTION

[0021] The application will be described in detail below in combination with the drawings and specific embodiments. The embodiments are implemented on the premise of the technical solution of the application, and detailed implementation modes and specific operation processes are given, but the protection scope of the application is not limited to the following embodiments.

[0022] Embodiment 1 The application proposes a topology auxiliary verification and correction method based on operation data driving, which aims to find the structural omissions, connection errors and unobserved devices in the main identification topology result of the distribution network, and proposes a credible correction suggestion. As shown in the figure, the method is divided into three steps: Figure 1 Step 1: Topology structure abnormality identification and candidate information extraction: use power flow residual, voltage imbalance and inconsistent behavior mode, combined with spectral clustering and graph completion algorithm, to locate the suspected structural defect area and obtain the candidate edge and unobserved node.

[0023] Step 2: Topology connection credibility scoring model construction: construct a behavior+structure joint feature tensor, use graph attention network (GAT) to output edge credibility score, and quantify the connection reliability.

[0024] Step 3: Topology structure correction mechanism and topology output: based on the scoring and structure reasoning results, optimize the node connection, and output the corrected topology structure and credible graph result.

[0025] Among them, the specific implementation process of each step is as follows: ​Step 1: Identify topological anomalies and extract candidate information. The goal of this step is to detect areas in the operating data that are inconsistent with the topological structure, locate anomalies from the flow and voltage behavior, and generate a set of edges to be completed and nodes that may be missed. The specific implementation process is as follows: Figure 2 As shown: 1.1) Multi-source residual consistency analysis A residual index based on power and voltage is constructed, and the sliding statistical method is used to screen abnormal nodes.

[0026] The power residual is calculated as follows:

[0027] Where: For nodes i At the moment t The active power residual of For all i Neighboring nodes that transmit power, For nodes k To the node i The power transmitted, For nodes i Output to adjacent nodes j Power, For nodes i The power of the local load.

[0028] The voltage residual is calculated as follows:

[0029] Where: For nodes i At the moment t The voltage residual; For nodes i Voltage value; For nodes i The set of adjacent nodes of is the number of adjacent nodes.

[0030] Then, the mean and standard deviation of the residual time series are calculated using a fixed-length sliding window to determine dynamic stability.

[0031] For the sliding window, the window length is set to , the current time is t , then the window interval is:

[0032] The average residual is calculated within the window interval to measure the overall power fluctuation level:

[0033] The standard deviation is calculated, and the larger the standard deviation, the worse the stability:

[0034] If the current time satisfies the following formula:

[0035] In the formula, is the average residual decision threshold, is the fluctuation amplitude threshold.

[0036] The node i is determined as an abnormal node, and is added to the abnormal node set , that is:

[0037] The obtained abnormal node set will focus on completing their connections, finding and their close but missed connections in the subsequent steps, and also pay attention to whether these abnormal nodes become isolated nodes due to measurement problems.

[0038] 1.2) Behavior pattern clustering and spectral clustering In order to further mine the implicit connections that may exist in the abnormal area, the behavior patterns between nodes need to be clustered. A similarity matrix based on dynamic time warping (DTW) is constructed, and a spectral clustering algorithm is used to find nodes with similar behavior but unconnected structure.

[0039] First, the node behavior characteristics are constructed as follows:

[0040] In the formula: , , The distribution is the node current, voltage and power factor, T is the time length.

[0041] Then calculate the dynamic time warping (DTW) distance:

[0042] In the formula: is the time alignment path, is the vector length.

[0043] Calculate the similarity matrix :

[0044] In the formula: is the Gaussian kernel width coefficient.

[0045] Then construct the Laplacian matrixL

[0046] where S is the similarity matrix, i.e. ; D is the degree matrix, and ; Spectral decomposition and eigenvector extraction are performed, and the following equation is solved:

[0047] where L is the Laplacian matrix; x is the eigenvector (corresponding to a dimension in the graph structure); is the eigenvalue, i.e., the strength of the tension in that direction.

[0048] Solve the first k nonzero eigenvalue corresponding to the eigenvector , which constitutes the spectral embedding matrix:

[0049] where each row represents the coordinates of the i th node in the spectral space.

[0050] Finally, the k-means algorithm is used to cluster each row X of the spectral embedding matrix , dividing nodes with similar features into the same cluster.

[0051] For any two nodes i and j , if they belong to the same cluster but are not connected in the existing topology, it is determined that there may be a potential connection between the two nodes, and the candidate edge set is added.

[0052] The candidate edges output by spectral clustering are important inputs or prior information for low-rank matrix completion. Low-rank completion can optimize and improve the network structure more comprehensively and strictly based on these preliminary candidates.

[0053] 1.3) Low-rank matrix completion based on singular value thresholding The goal of graph structure completion is to fill in the missing connection edges in the adjacency matrix based on the partially known node connection relationships and similarities, and to reconstruct the topology of the graph.

[0054] This problem can be modeled as a low-rank matrix completion problem, and the singular value thresholding (SVT) algorithm is used for iterative solution.

[0055] Let the original observed adjacency matrix be​ but some of its elements are not observable.

[0056] Define the observation set as Then construct the following optimization problem:

[0057] where is the adjacency matrix to be completed; is the observation position projection operator; is the Frobenius norm; is the nuclear norm, i.e., the sum of all singular values, reflecting the rank of the matrix; is the regularization coefficient, .

[0058] The singular value thresholding (SVT) algorithm proceeds as follows: 1.3.1) Algorithm initialization, set the initial matrix , set the threshold parameter, step size parameter, convergence tolerance, maximum number of iterations K .

[0059] 1.3.2) For each iteration k =1,2,3,... K ; perform singular value decomposition on the matrix :

[0060] where , is the orthogonal matrix; is the singular value diagonal matrix, r is the effective rank, each represents the strong principal component amplitude of the current graph structure.

[0061] Continue soft thresholding operation on all singular values:

[0062] where, if , the singular value is set to 0, suppressing noise and redundant structure; is the sparse singular value matrix.

[0063] Reconstruct the adjacency matrix using the sparse singular values:

[0064] where is the completed adjacency matrix estimate value of the k +1th iteration.

[0065] Introduce the reconstruction error into the next iteration:

[0066] Where, To preserve the error of the observed position; is the step size parameter.

[0067] The iteration is terminated if one of the following conditions is met:

[0068] Where, is the convergence tolerance.

[0069] Finally, the completed adjacency matrix is ​​output , and Above threshold The non-zero elements of It is determined that there is a potential connection, that is, a candidate edge. If the corresponding nodes of a row or column are all zero or approximately zero, then the node is considered to be a potential unobserved node.

[0070] The set of candidate edges obtained by low-rank matrix completion includes both possible connections discovered by the spectral clustering method and new connections inferred by the completion algorithm based on global consistency.

[0071] In this step, for topological anomaly identification and candidate information extraction, first, multi-source residual consistency analysis is used to find abnormal nodes. Then, behavioral pattern clustering and spectral clustering are used to cluster these abnormal nodes and surrounding nodes. It can be found that although some nodes are not connected together in the original network, their power consumption patterns are very similar (in this case, there may actually be a real but missed connection between them). The connections of these nodes are added to the candidate edge set. Finally, the low-rank matrix completion based on singular value thresholding is based on known connections, candidate connections found through behavioral pattern clustering and spectral clustering, and similarities between nodes. Through the SVT algorithm, during the algorithm iteration, the entire network information can be integrated to fill in those missing connection relationships, and even some isolated island nodes that have not been noticed before may be discovered.

[0072] Step 2: Topological connection credibility scoring model construction. This step aims to evaluate each edge in the candidate edge set ( i,j ) Establish a credibility scoring function to quantify the possibility of it being a real connection edge. The process of this step is as follows Figure 3 As shown: By constructing a structural-behavioral joint feature tensor and introducing the graph attention mechanism (GAT), this model can comprehensively judge the behavioral synergy, adjacency structure similarity, and mutation synchronization between nodes, and output a score. The higher the score, the greater the possibility that the connection edge exists in the real topological structure.

[0073] 2.1) For each pair of candidate nodes , construct five-dimensional features:

[0074] Where each index is calculated as follows: is the current correlation, i.e. the correlation coefficient of the current between nodes i,j .

[0075]

[0076] Where, and are the current time series between nodes i,j . Cov is the covariance; is the standard deviation.

[0077] is the voltage correlation, i.e. the correlation coefficient of the voltage between nodes i,j .

[0078]

[0079] Where, and are the voltage time series between nodes i,j .

[0080] is the mutation synchronization degree, calculated as follows: For each node construct a set of mutation events , if at any time t , satisfies:

[0081] then add this time to the set of mutation events , i.e.

[0082] For the set of mutation events i,j of node , perform matching: if , there exists such that ( is the tolerance synchronization error) then consider it a pair of synchronous mutations, define the mutation synchronization degree as:

[0083] Where, is used to prevent division by zero.

[0084] Mirror change rate, used to detect nodes i Current abrupt change corresponds to the node j The voltage direction is opposite, and the calculation method is as follows: Get the detection node i The set of abrupt change moments ; At each abrupt point, determine whether the following formula is satisfied:

[0085] In the formula, and are the current and voltage change of node i at time t , is a sign function, that is, the positive and negative of the return value.

[0086] Finally, the number of times that satisfy the formula is divided by the total number of abrupt changes:

[0087] In the formula, is an indicator function, which takes 1 when the condition is met, and 0 otherwise, is the number of parameters in the set of abrupt change moments.

[0088] The adjacent coincidence rate is obtained by calculating the intersection and union of the adjacent sets of two nodes.

[0089]

[0090] In the formula, and are the adjacent node sets of nodes i and j, respectively.

[0091] 2.2) Connection credibility score model construction This scoring mechanism uses a lightweight graph attention network (GAT) to model the connection credibility of candidate edges ( i,j ). The core process includes feature mapping, attention calculation, weight normalization, and final score output.

[0092] (2.2.1) Feature mapping Use linear transformation to project low-dimensional input features into high-dimensional space to enhance the nonlinear interaction ability between different scoring factors:

[0093] In the formula, is a learnable parameter matrix; is an edgeEmbedding feature representation of. .

[0094] (2.2.2) Attention score calculation Input embedding vectors into attention mechanism for scoring:

[0095] where, is a learnable attention weight vector; is a nonlinear activation function, denotes the original attention value of edge .

[0096] (2.2.3) Final score output Send the attention score into a sigmoid activation function to compress it into the interval [0, 1] to get the final edge confidence:

[0097] where, is an optional edge score output layer; is a sigmoid function, is the probability of the existence of the candidate edge.

[0098] Step 3: Structure correction mechanism and topology output The purpose of this step is to modify the initial topology structure based on the candidate structure, combine the output results of the scoring model with the potential edges and unknown node information identified in the graph structure completion process, and generate a complete and reliable output topology structure. The process is shown in Figure 4 : 3.1) Input information Candidate edge set: = { (from step 1 structure completion process)}; Unobserved node (candidate node) set: = { (from step 1 structure completion process)}; = Score result of each edge: (from step 2 scoring result); Original topology graph structure: .

[0099] 3.2) Edge correction mechanism According to the scoring result of each candidate edge obtained in step 2 , the following strategy is used.

[0100] When , it is determined as a real edge, and the completion is performed, if ​If so, the candidate edge is determined, and manual review or further verification is performed. If If so, the edge is determined to be non-existent, and the elimination process is performed.

[0101] 3.3) Unobserved node connection mechanism In step 1, a batch of suspected unobserved or unmodeled nodes are identified through methods such as graph structure completion and spectral clustering. These nodes have obvious power and voltage activity in their behavior time series data, but are not connected to any known nodes. Therefore, in the structure modification stage, their connection edges need to be supplemented to achieve the structural closure of the graph.

[0102] (3.3.1) First, filter the nodes that are completely isolated after the edge modification. These nodes are not connected in the candidate edge completion and must be structure matching predicted.

[0103] (3.3.2) For each unobserved node Combine all in the current topology, establish a candidate edge set, and for each candidate edge , use the scoring method in step 2 to calculate its score result .

[0104] Based on the score results of each candidate edge, construct the score table as follows:

[0105] Determine the connection relationship based on the mechanism shown in the following table.

[0106] Table 1 Connection relationship judgment table

[0107] 3.4) Structure graph update and result output According to the above process, adjust the edge structure and complete the final output of the modified topology graph:

[0108]

[0109]

[0110] In the formula, is the final output of the power grid topology graph, including all nodes and the modified connection relationship; is the node set of the final topology graph, including the original nodes and the identified unobserved nodes; is the edge set of the final topology graph, including the original retained edges, the completed edges, and the predicted connected edges; is the set of nodes that have been modeled in the original topology structure; The set of unobserved nodes identified in the structure completion process; The set of all edges (connection relationships) existing in the original topology graph; The set of edges that are determined to be low in credibility and should be removed in the scoring process; The set of edges that are scored higher than the credibility threshold in the candidate edge scoring and are determined to be completed; The set of access edges generated by unobserved node prediction to ensure structural connectivity.

[0111] The method of the present application realizes accurate correction and closed-loop completion of the power distribution network topology by introducing a scoring-driven edge credibility judgment mechanism combined with unobserved node structured access strategy. Compared with existing topology identification models, it has obvious improvement in multiple key indicators, as shown in the following table.

[0112] Table 2 Comparison of the present application scheme and existing topology identification model identification structure

[0113] As can be seen from the results, the present application ensures low computing resource consumption while achieving comprehensive repair of missing edges, incorrect edges and unmodeled nodes in the original topology structure, significantly improving the integrity and credibility of the power distribution network topology. The method has high deployability and can be integrated into existing power distribution network topology identification systems as an auxiliary verification and correction module, and is especially suitable for substation structure modeling optimization tasks for automated operation and maintenance scenarios.

[0114] Embodiment 2 As a second aspect of the present application, the present application also provides an electronic device comprising: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the power distribution network topology correction method as described above. In addition to the above-mentioned processors, memories and interfaces, any device with data processing capability in the embodiment usually includes other hardware according to the actual function of the device with data processing capability, which will not be described here.

[0115] Embodiment 3 As a third aspect of the present application, the present application also provides a computer readable storage medium having stored thereon computer instructions which, when executed by a processor, implement the power grid topology correction method of fusing topology anomaly identification and structure trusted reasoning as described above. The computer readable storage medium can be an internal storage unit of any data processing capable device, such as a hard disk or a memory, of any of the aforementioned embodiments. The computer readable storage medium can also be an external storage device, such as a plug-in hard disk, a Smart Media Card (SMC), an SD card, a Flash Card, etc. equipped on the device. Further, the computer readable storage medium can include both the internal storage unit of any data processing capable device and the external storage device. The computer readable storage medium is used to store the computer program and other programs and data required by the data processing capable device, and can also be used to temporarily store data that has been output or will be output.

[0116] The above functions, if implemented in the form of software function units and sold or used as independent products, can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application, in essence or the part that contributes to the prior art, or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0117] The preferred embodiments of the present application are described in detail above. It should be understood that those skilled in the art can make many modifications and changes without creative labor based on the concept of the present application. Therefore, any technical solution that can be obtained by logical analysis, reasoning or limited experiments by those skilled in the art based on the concept of the present application and the prior art shall be within the protection scope determined by the claims.

Claims

1. A power distribution network topology correction method that fuses topology anomaly identification and structure credible inference, characterized in that, The method steps include: Identify the abnormal area of the topology structure, identify the candidate edge and the unobserved node; Construct a structure and behavior joint feature tensor considering the behavior synergy, adjacency structure similarity and mutation synchronization between nodes, and input the topology connection credibility scoring model to obtain the edge credibility score; Combine the edge credibility score obtained by the topology connection credibility scoring model and the candidate edge and the unobserved node to perform credible correction on the initial topology structure to generate the final topology structure.

2. The power distribution network topology correction method of claim 1, wherein, The identification of the abnormal area of the topology structure, the identification of the candidate edge and the unobserved node are as follows: Calculate the power residual and voltage residual of each node in the topology structure, and perform mean and standard deviation statistics on the residual time series in a sliding window, and perform dynamic stability discrimination; Determine the nodes whose residual mean and standard deviation in the sliding window interval are greater than the set threshold as abnormal nodes, and add them to the abnormal node set; Use dynamic time warping distance to construct a similarity matrix for the abnormal nodes, and obtain the potential connection edge between two nodes through a spectral clustering algorithm; Model the graph structure completion as a low-rank matrix completion problem, and solve it iteratively using a singular value thresholding algorithm to obtain a completed adjacency matrix; Determine that the non-zero elements in the completed adjacency matrix that are higher than the threshold are potential connections, i.e., candidate edges; if the nodes corresponding to a row or column of the completed adjacency matrix are all zero or approximately zero, then the node is determined to be a potential unobserved node.

3. The power distribution network topology correction method of claim 2, wherein, The spectral clustering algorithm for obtaining the potential connection edge between two nodes is as follows: Based on the node current, voltage and power factor, construct the node behavior feature, and calculate the dynamic time warping distance between nodes based on the node behavior feature; Calculate the similarity matrix based on the dynamic time warping distance between nodes, and further construct a Laplacian matrix; before solving the laplacian matrix k the eigenvector corresponding to the smallest non-zero eigenvalue of the laplacian matrix, to form a spectral embedding matrix Cluster each row of the spectral embedding matrix, and divide nodes with similar features into the same cluster; For any two nodes, if they belong to the same cluster but are not connected in the existing topology, it is determined that there may be a potential connection edge between the two nodes, and the candidate edge set is added.

4. The power distribution network topology correction method of claim 2, wherein, The graph structure completion is modeled as a low-rank matrix completion problem as follows: wherein is the original observed adjacency matrix; is the adjacency matrix to be completed; is the observed position projection operator; is the observed set; is the Frobenius norm; is the kernel norm; is the regularization coefficient.

5. The power distribution network topology correction method of claim 1, wherein, The topology connection credibility scoring model uses a graph attention network to model the connection credibility of the candidate edge: the structure and behavior joint feature tensor is projected into a high-dimensional space using linear transformation to obtain the embedded feature representation of the edge; Input the embedded vector into the attention mechanism for scoring, and then input the attention score into the activation function to obtain the final edge credibility.

6. The power distribution network topology correction method of claim 1, wherein, The structure and behavior joint feature tensor specifically includes: current correlation, voltage correlation, mutation synchronization degree, mirror change rate and adjacency overlap rate between each pair of candidate nodes.

7. The power distribution network topology correction method of claim 6, wherein, The mutation synchronization degree is calculated as follows: For each node, a set of mutation events is constructed If at any time t the current or voltage change is greater than a set threshold, then that time is added to the set of mutation events. A set of mutation events for two nodes , Perform matching, if the first time , there is a second time Make the absolute value of the difference between the first time And the second time Less than or equal to the tolerance synchronization error, it is considered as a pair of synchronous mutations, and the mutation synchronization degree is defined as: wherein represents a set of mutation events of a node i,j ; is a constant for preventing division by zero.

8. The power distribution network topology correction method of claim 6, wherein, The mirror change rate is used to detect the node i Whether the current mutation corresponds to the node j The voltage direction is opposite, and the calculation method is as follows: Get the set of mutation times of the detected nodes, and for each mutation point, determine whether the current and voltage change directions of the nodes are opposite, and count the number of times and divide by the total number of mutations: In the formula, is a set of mutation moments of the node i . and are respectively current and voltage variation amounts of the node i at the moment t . is a sign function, that is, the sign of the return value.

9. The power distribution network topology correction method of claim 1, wherein, Based on the scoring results of each candidate edge, the edge is corrected, and the candidate edges with a scoring result lower than a set scoring threshold are removed. ​ Screening the nodes which are still completely isolated after the edge correction to perform structure matching prediction, and predicting each unobserved node with all original topologies in the current topology j In combination, for each candidate edge Calculate the score result by using the topology connection credibility scoring model, and judge the connection relationship; Based on the connection relationship judgment result, the edge structure is adjusted and completed to obtain a revised topological graph.

10. The power distribution network topology correction method of claim 9, wherein, The connection relationship judgment is specifically as follows: For each unobserved node, if the maximum candidate edge score result is less than the first score threshold, the unobserved node is not connected; If there is a node such that the candidate edge score result is greater than or equal to the second score threshold, and the number of nodes satisfying the condition is less than the set number, then the nodes satisfying the condition are connected; If there is a node such that the candidate edge score result is greater than or equal to the second score threshold, and the number of nodes satisfying the condition is greater than or equal to the set number, then the set number of nodes with the highest score are connected; If there is no node with a candidate edge score result greater than or equal to the second score threshold, but there is a node such that the candidate edge score result is between the first score threshold and the second score threshold, then the node with the highest score is connected.

Images