A distributed photovoltaic substation metering data anomaly tracing method

Abstract

The present application relates to a kind of distributed photovoltaic substation metering data abnormality tracing method, belong to the field of intelligent distribution network, solve the problem of low positioning efficiency of abnormal user caused by the loss of topology information in low-voltage photovoltaic substation.The electrical data of distribution transformer and each user in the distributed photovoltaic substation at night without photovoltaic output is collected, and the approximate topology of substation main line is constructed.Based on the voltage data of each user in the collected substation during the high output period of photovoltaic, the final adjacency matrix between nodes is constructed.The users are divided into multiple user clusters based on the final adjacency matrix between nodes clustering.The cluster power curve is constructed for each user cluster, and the fluctuation consistency between the voltage of each user in the cluster and the cluster power curve of the user's cluster is calculated based on the approximate topology of substation main line to identify suspected abnormal clusters.Node-level identification is carried out on suspected abnormal clusters to obtain abnormal metering users.The automatic tracing and accurate positioning of metering data abnormality are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent power distribution network, and particularly relates to a distributed photovoltaic substation metering data abnormality tracing method. BACKGROUND

[0002] In the intelligent power distribution network, the large-scale access of distributed photovoltaic makes the substation operation structure more complex, and the traditional metering data abnormality troubleshooting method faces challenges. On the one hand, the low-voltage substation generally lacks complete topological structure information, and cannot accurately obtain the connection relationship between users; on the other hand, the number of users in a single substation is large, and abnormality troubleshooting often needs to check each household, resulting in low positioning efficiency. In the existing method, rule threshold judgment, statistical analysis (such as clustering, principal component analysis) or shallow machine learning method is often used for abnormality detection. These methods can identify overall data fluctuations, but most of them rely on manual experience to set parameters, lack self-adaptive ability, and especially when the substation lacks topological support or faces high-dimensional time series data, the recognition accuracy is limited, and it is difficult to locate specific abnormal metering users in real time.

[0003] In addition, the existing method can only determine that the substation as a whole is abnormal, lacks in-depth modeling of the voltage response correlation between users and the physical connection structure, and is difficult to accurately identify the abnormal source, and there is a risk of missed detection and misjudgment. SUMMARY

[0004] In view of the above analysis, the embodiments of the present application aim to provide a distributed photovoltaic substation metering data abnormality tracing method to solve the technical problem of low positioning efficiency of abnormal metering users in low-voltage photovoltaic substations caused by the lack of topological information in the existing method.

[0005] The purpose of the present application is mainly realized through the following technical solutions:

[0006] The present application provides a distributed photovoltaic substation metering data abnormality tracing method, comprising the following steps:

[0007] Collecting electrical data of the distribution transformer and each user in the distributed photovoltaic substation at night without photovoltaic output, and constructing a substation main line approximate topology;

[0008] Based on the collected voltage data of each user in the substation during the high photovoltaic output period, an inter-node final adjacency matrix between nodes is constructed; and based on the inter-node final adjacency matrix, the users are divided into multiple user clusters through clustering;

[0009] Constructing a cluster power curve for each user cluster, and calculating the fluctuation consistency between the voltage of each user in the cluster and the cluster power curve of the cluster to which the user belongs based on the substation main line approximate topology, so as to identify a suspected abnormal cluster;

[0010] The abnormal cluster is identified at a node level to obtain an abnormal metering user.

[0011] Wherein, the distribution transformer and the user are regarded as nodes.

[0012] Further, the electrical data includes voltage amplitude of the distribution transformer, voltage, active power and reactive power of each user node; and the constructing the approximate topology of the main line of the transformer area includes:

[0013] The user node with the highest voltage is selected from the nodes closest to the distribution transformer as a current node, and the cost function value between the current node and each remaining node is calculated, and the node with the minimum cost function value is selected as a new current node; the remaining nodes are polled until all nodes are polled.

[0014] Based on the voltage of the selected user node, a Lasso regression model is used for data fitting, and the coupling strength between nodes is used as a matrix element to obtain an initial adjacency matrix between nodes.

[0015] The main chain processing is performed based on the initial adjacency matrix between nodes to obtain the approximate topology of the main line of the transformer area.

[0016] Further, the cost functions of the current node and the next node are represented as follows:

[0017] ;

[0018] Wherein, are the voltages of the current node and the next node, respectively; are the active powers of the current node and the next node, respectively; are weight coefficients of the voltage difference term and the active power term, respectively.

[0019] Further, the Lasso regression model is represented as follows:

[0020] ;

[0021] Wherein, is a connection vector of the node , and λ is an L1 regularization coefficient;

[0022] The Lasso regression model is used for data fitting to obtain the coupling strength value between nodes.

[0023] The coupling strength value between nodes is used as a matrix element to obtain the initial adjacency matrix between nodes.

[0024] Further, the main chain processing based on the initial adjacency matrix between nodes includes:

[0025] symmetrizing the initial adjacency matrix between nodes to obtain a symmetrized adjacency strength;

[0026] comparing the symmetrized adjacency strength with a preset coupling strength threshold, retaining the adjacency edges greater than the preset coupling strength threshold to obtain an undirected sparse graph;

[0027] taking each edge of the undirected sparse graph as a coupling strength between nodes as a graph distance to obtain a weighted graph, and using Kruskal algorithm to obtain a corresponding minimum spanning tree; wherein the edges in the minimum spanning tree satisfying the main chain optimal path criterion constitute a main chain path;

[0028] if the node belongs to the main chain path, it is marked as a main node, otherwise it is a branch node; all main nodes and branch nodes, and the coupling strength between nodes, constitute the approximate topology of the transformer main line.

[0029] Further, the main chain optimal path criterion is as follows:

[0030] ;

[0031] wherein, is the main chain path; is the minimum spanning tree, represents a set of all traversable paths in the minimum spanning tree, represents each pair of adjacent nodes in the current path P, , are weight coefficients of the influence of the control voltage stability term and the phase angle difference term on the path score, respectively; are the voltage of the node , and the phase angle of the node , , are the voltage amplitude of the node

[0032] Further, the construction of the final inter-node adjacency matrix between nodes includes:

[0033] normalizing the voltage data sequence of each user in the midday photovoltaic high-power period;

[0034] based on the normalized user voltage sequence data, calculating the standard deviation of the user node voltage sequence, and calculating the Pearson correlation coefficient between the user nodes;

[0035] based on the Pearson correlation coefficient between the user nodes and the coupling strength in the undirected sparse graph, constructing the voltage fluctuation correlation matrix of the user nodes and the node .

[0036] computing user nodes with voltage fluctuation covariance of nodes ;

[0037] based on voltage fluctuation correlation matrix of user nodes with nodes and voltage fluctuation covariance of nodes ; computing inter-node final adjacency matrix of nodes with nodes

[0038] ; further, constructing voltage fluctuation correlation matrix of user nodes with nodes as follows:

[0039] ;

[0040] wherein, is voltage fluctuation similarity between user nodes and ; is Pearson correlation coefficient between user nodes and , is structure prior enhancement factor, is coupling strength in undirected sparse graph, is power adjustment term for enhancing structure influence.

[0041] further, computing inter-node final adjacency matrix of nodes with nodes as follows:

[0042] ;

[0043] wherein, is voltage fluctuation covariance of user nodes and nodes , are respectively voltage sequence standard deviations of user nodes and nodes , are respectively measurement weights of fusing two similarity degrees.

[0044] further, based on the inter-node final adjacency matrix, constructing symmetric normalized Laplacian matrix and performing spectral clustering to obtain multiple user clusters with similar voltage fluctuation patterns.

[0045] further, based on the Laplacian matrix, performing spectral clustering to obtain multiple user clusters with similar voltage fluctuation patterns, comprising:

[0046] Solving symmetric normalized Laplacian matrix The first eigenvalue corresponding to the eigenvector, the eigenvector is composed of spectral feature space ;

[0047] According K-means clustering by row, get multiple user clusters, each cluster includes at least 1 user node.

[0048] Further, the identification of suspected abnormal cluster, comprising:

[0049] The voltage and power of each user node in each cluster are standardized to obtain standardized voltage and power;

[0050] Calculate the Mahalanobis distance between the standardized voltage and the standardized power of each user node in each cluster;

[0051] Calculate the dynamic time warping distance between the standardized voltage of each user node and the cluster standardized power curve value, and the Pearson consistency; wherein the cluster standardized power curve value is the average of all user power in the cluster;

[0052] Based on the Mahalanobis distance, dynamic time warping distance and Pearson consistency, the multi-dimensional consistency anomaly score value of each user node in each cluster is calculated by weighted calculation;

[0053] Define the power response distribution entropy of each user voltage fluctuation in each cluster, and calculate the entropy variance in the cluster based on the power response distribution entropy;

[0054] Construct the corresponding probability distribution of each cluster and the reference normal cluster, and calculate the KL divergence of the probability distribution of each cluster and the probability distribution of the reference normal cluster;

[0055] For each cluster, if the mathematical expectation of the multi-dimensional consistency anomaly score value is greater than the consistency threshold, the entropy variance in the cluster is greater than the entropy threshold in the cluster, and the KL divergence is greater than the divergence threshold, any one, then the cluster is judged as a suspected abnormal cluster.

[0056] Further, the multi-dimensional consistency anomaly score value of the user node is as follows:

[0057] ;

[0058] Wherein, is the Mahalanobis distance between the standardized voltage and the standardized power of the user node ; is the dynamic time warping distance between the standardized voltage of the user node and the cluster standardized power curve value; is the Pearson consistency; , and are respectively , and weight.

[0059] Further, the suspected abnormal cluster is subjected to node-level identification to obtain an abnormal metering user, comprising:

[0060] For the suspected abnormal cluster, a graph structure subgraph of the suspected abnormal cluster in the final adjacency matrix between the nodes is extracted;

[0061] Based on the graph structure subgraph, a symmetric normalized adjacency submatrix is constructed;

[0062] An input feature vector of each user node in the suspected abnormal cluster is extracted;

[0063] Based on the symmetric normalized adjacency submatrix and the input feature vector, a two-layer graph convolution network is used to output a user node abnormal probability;

[0064] If the user node abnormal probability is greater than the upper limit range of the probability threshold, the user node is determined to be an abnormal user.

[0065] Further, the two-layer graph convolution network GCN is used to output a user node abnormal probability, comprising:

[0066] Based on the symmetric normalized adjacency submatrix and the input feature vector, a first layer propagation is performed to obtain a hidden layer node feature matrix;

[0067] Based on the hidden layer node feature matrix and the symmetric normalized adjacency submatrix, a second layer propagation is performed to obtain an output abnormal probability;

[0068] The output abnormal probability is subjected to binary classification determination to obtain a user node abnormal probability.

[0069] Further, the input feature vector comprises a voltage mean value, a voltage standard deviation, a power mean value, a power standard deviation, a voltage kurtosis, a voltage skewness, a voltage fluctuation upper limit, a voltage fluctuation lower limit, a similarity correlation feature, and a frequency domain feature vector.

[0070] Compared with the prior art, the present application can at least achieve one of the following beneficial effects:

[0071] 1、The present application can mine the internal law between voltage and power data without topology drawings by fusing intelligent power distribution network physical modeling and data-driven algorithm, and realize accurate identification and real-time positioning of abnormal users in the transformer area. This method combining physical model and data-driven has higher adaptive ability and identification accuracy than the traditional method relying only on artificial experience setting parameters, especially in the absence of topology support or facing high-dimensional time series data, which can more accurately identify abnormal sources and reduce the risk of missed detection and misjudgment;

[0072] 2、The present application can realize rapid and accurate identification of abnormal users by constructing a model based on a two-layer graph convolutional network (GCN, Graph Convolutional Network) and combining topology information and user power consumption characteristics, avoiding the low efficiency problem of checking each household, thereby significantly improving the efficiency of abnormal positioning. This method not only can quickly identify abnormal users and improve the efficiency of abnormal positioning, but also can reduce the waste of time and resources caused by checking each household;

[0073] 3、The present application can support intelligent operation and maintenance and rapid abnormal disposal of power distribution network, and provide real-time decision support for power distribution network operation and maintenance personnel through automatic abnormal detection and positioning, support intelligent operation and maintenance, and improve the operation efficiency and reliability of intelligent power distribution network. This intelligent operation and maintenance method helps to reduce human errors, improve response speed, and ensure the stable operation of the power grid.

[0074] In the present application, the above-mentioned technical solutions can be combined with each other to realize more preferred combination schemes. Other features and advantages of the present application will be described in the subsequent specification, and some advantages will become apparent from the specification or by implementing the present application. The purpose and other advantages of the present application can be realized and obtained from the contents specifically pointed out in the specification and the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0075] The accompanying drawings are included to provide a better understanding of the embodiments, and should not be considered as limiting the application, in which the same reference numerals refer to the same components throughout the drawings;

[0076] Figure 1 A distributed photovoltaic transformer area metering data abnormality tracing method flowchart in the embodiments of the present application;

[0077] Figure 2 A typical low-voltage line load distribution diagram of a photovoltaic power generation access in the embodiments of the present application;

[0078] Figure 3 A user voltage curve over time diagram when photovoltaic power generation in the embodiments of the present application;

[0079] Figure 4For the embodiment of the present application, the method in the present application is used to identify that there is an abnormal user schematic diagram in the second suspected abnormal cluster. DETAILED DESCRIPTION

[0080] The preferred embodiments of the present application will be described in detail below with reference to the drawings, wherein the drawings constitute a part of this application and are used to explain the principles of the embodiments of the present application, but are not used to limit the scope of the present application.

[0081] To solve the above problems, an online abnormality tracing method is needed, which fuses power distribution network physical modeling and data-driven algorithm, mines the internal law between voltage and power data under the premise of no transformer substation topology paper, realizes accurate identification and real-time positioning of abnormal users in the transformer substation, and supports intelligent operation and maintenance and abnormal rapid disposal of the power distribution network.

[0082] One specific embodiment of the present application discloses a distributed photovoltaic transformer substation metering data abnormality tracing method, as shown in the figure, comprising the following steps: Figure 1

[0083] Step S1, collecting electrical data of the distribution transformer and each user in the distributed photovoltaic transformer substation at night without photovoltaic output, and constructing a transformer substation main line approximate topology;

[0084] Step S2, based on the collected voltage data of each user in the transformer substation at noon photovoltaic high output period, constructing a final inter-node adjacency matrix between nodes; based on the final inter-node adjacency matrix, clustering and dividing the users into multiple user clusters;

[0085] Step S3, constructing a cluster power curve for each user cluster, calculating the fluctuation consistency between the voltage of each user in the cluster and the cluster power curve of the cluster based on the transformer substation main line approximate topology, and identifying a suspected abnormal cluster;

[0086] Step S4, performing node-level identification on the suspected abnormal cluster to obtain an abnormal metering user;

[0087] Wherein, the transformer and the user are regarded as nodes.

[0088] The distributed photovoltaic transformer substation refers to a low-voltage power supply area widely accessing roof photovoltaic, electric vehicles and other distributed power sources, and has the characteristics of strong bidirectional power flow, frequent voltage fluctuation, complex user structure, etc.

[0089] Step S1 comprises steps S11-S12.

[0090] Step S11, collecting electrical data of the distribution transformer and each user in the distributed photovoltaic transformer substation at night without photovoltaic output.

[0091] ​The electrical data includes voltage amplitude of distribution transformer, voltage, active power and reactive power of each user node.

[0092] Night data selection, select the stable period of night without photovoltaic output, obtain the voltage and power data of distribution transformer and each user.

[0093] Exemplarily, the obtained electrical data includes: voltage data, active power data and reactive power data of each user; voltage amplitude of distribution transformer (the bus of the transformer area, unified as 1.0 p.u.), p.u. is per unit, unit value. The obtained data is shown in Table 1:

[0094] Based on the voltage of each user, the user voltage amplitude vector ;

[0095] Based on the active power data of each user, the user power amplitude vector ;

[0096] Based on the reactive power data of each user, the user reactive power amplitude vector The number of users in the distributed photovoltaic transformer area is .

[0097] Table 1: Obtained electrical data

[0098]

[0099] Select the most stable data in the night 1-3 hour window for analysis. Among them, is the voltage value of the user; is the active power of the user; is the reactive power of the user.

[0100] The voltage of the user in the distributed photovoltaic transformer area is exemplarily shown in Figure 3 , an example of 24-hour voltage curve, which shows the voltage change of 8 users in a day in a certain distributed photovoltaic transformer area, which can be used for intuitive understanding of abnormal tracing scene.

[0101] Horizontal axis: t / h: time (hour), range 0.0→24.0h, covering the whole day.

[0102] Vertical axis: voltage (unit value p.u.), reference 1.0 p.u. corresponds to the bus voltage of the distribution transformer.

[0103] The scale in the figure is 0.96→1.14 p.u., which is converted to actual voltage about 220V→250V (according to 220V bus).

[0104] The step S11 is to collect the electrical data of the distribution transformer and each user in the stable period without photovoltaic power at night, including voltage and power data, to provide basic data support for constructing the approximate topology of the main line of the transformer area and subsequent abnormal tracing analysis.

[0105] The step S12 is to construct the approximate topology of the main line of the transformer area.

[0106] The voltage of the transformer area bus is 1.0 p.u., and the voltage of the transformer area bus is taken as the reference voltage, and other electrical data values are represented relative to the reference voltage, as shown in Table 1.

[0107] According to the electrical physical law, in the low-voltage transformer area network, the voltages of adjacent nodes satisfy the approximate Ohm's law, and the coupling relationship of node voltage-current is obtained, as follows:

[0108] Formula (1)

[0109] wherein, is the complex voltage of the user node is the complex impedance of the branch is the resistance of the branch is the branch current from the user node to .

[0110] is the complex impedance of the branch is the resistance of the branch is the branch current from the node to ; is the reactance of the branch . As shown in . Figure 2 The formula is essentially Ohm's law used in an alternating current system to describe that in a low-voltage distribution network, the voltage difference between two adjacent nodes is equal to the product of the branch impedance and the current flowing through the branch.

[0111] In the present application, the node refers to the distribution transformer and all terminal users supplied by it, each node corresponds to a sampling point at the metering level, has independent voltage, current and power measurement records, and can be used to construct the electrical relationship and topology structure in the transformer area.

[0112] For simplicity, the voltage phase angle and nonlinear effects are ignored, and the voltage drop is estimated by the voltage amplitude, as follows:

[0113]

[0114] ​​​​​Formula (2)

[0115] wherein, is the voltage amplitude voltage drop between node and ; is the resistance between node and ; is the reactance between node and ; is the active load power of node ; is the reactive load power of node ; , are the voltage amplitudes of nodes , respectively.

[0116] Compared with the traditional voltage difference approximation equal to power multiplied by resistance theory, the formula can more accurately consider the joint action of active and reactive power and is applicable to physical connection inference between any nodes.

[0117] The constructed main line route of the transformer district approximates topology, comprising:

[0118] The user node with the highest voltage is selected from the node closest to the distribution transformer as the current node, the cost function value between the current node and each of the remaining nodes is calculated, the node with the minimum cost function is selected as the new current node, and the remaining nodes that have not been selected are polled until all nodes are polled.

[0119] Based on the voltage of the selected user node, data fitting is performed using a Lasso regression model, the coupling strength between nodes is taken as a matrix element, and an initial adjacency matrix between nodes is obtained.

[0120] The main chain processing is performed based on the initial adjacency matrix between nodes to obtain the main line route of the transformer district approximates topology.

[0121] The cost function reflects the degree of physical consistency between node and node in voltage response and power characteristics, and is used to describe whether there is a real electrical coupling relationship between the two. The smaller the value, the more likely the connection.

[0122] The cost functions of the current node and the next node are as follows:

[0123] Formula (3)

[0124] wherein, V and V+1 are the voltage of the current node and the next node respectively; P and P+1 are the active power of the current node and the next node respectively; α and β are the weight coefficients of the voltage difference term and the active power term respectively.

[0125] α and β are the weight coefficients of the voltage difference term and the active power term respectively, used to adjust the relative influence of the voltage difference and the power difference;

[0126] The meaning of adding the second term power difference term in formula (3) is that, from the physical mechanism of the power distribution system, users with similar load power usually have the following characteristics:

[0127] ① The electrical appliances are close to each other (for example, connected to the same branch or the same power distribution transformer);

[0128] ② The probability of simultaneous change is higher, which is reflected in the voltage fluctuation and is more consistent;

[0129] ③ Even if the voltage itself cannot distinguish "proximity", the power characteristics can make up for the judgment (especially when the phase interference or voltage noise is large); for example, two users have very close voltages |V i -V j ≈0, but one of them has a load of 10kW and the other has a load of 0.5kW, the probability of the two nodes being physically connected is much lower than that of two users with loads of about 10kW.

[0130] Users with similar load power, reflected in the cost function, are that the voltage difference is approximately 0 and the power difference is also approximately 0; the smaller the cost function, the closer the two nodes.

[0131] L1 regularization is introduced to control sparsity, and a Lasso optimization problem is constructed. The purpose of constructing the Lasso optimization model is to deduce the node adjacency matrix A and obtain the approximate physical topology structure of the main line of the transformer area without relying on the known topology, by fitting and sparsity compression of the voltage data between users.

[0132] The Lasso regression model is represented as follows:

[0133] Formula (4)

[0134] wherein, is the connection vector of node , and λ is the L1 regularization coefficient;

[0135] The Lasso regression model is used for data fitting to obtain the coupling strength value between nodes.

[0136] The coupling strength value between nodes is used as the matrix element to obtain the initial adjacency matrix between nodes.

[0137] is a vector, the connection vector of node , represents the coupling relationship between user node and all other user nodes; its component is the coupling strength value between nodes , represents node , the coupling strength of node to node , the voltage of node , the voltage of node , the influence weight of node .

[0138] The physical meaning of formula (4) is to solve by Lasso optimization problem , so that the voltage of node can be best fitted by the linear combination of other user node voltages .

[0139] The coupling strength value between nodes represents the connection possibility between node and node . is a vector, the coupling degree of user node to other user nodes.

[0140] λ is the L1 regularization coefficient, which controls the sparsity of the graph. Exemplarily, λ takes the following values:

[0141] (1) Generally, when the number of users is small (10-50), take ;

[0142] (2) When the number of users is medium (50-200), take ;

[0143] (3) When the number of users is large (200+), take 0.1-1.0.

[0144] Lasso is a one-way fitting, which may occur In power systems, the "electrical connection" between distribution lines connecting users is bidirectional conduction, unless directional elements such as SOP (Soft open point, flexible soft switch) are added, but in ordinary low-voltage distribution networks, it is not considered. Therefore, whether the user nodes are electrically connected should be a symmetric relationship, that is, the physical structure diagram is a directed graph.

[0145] The backbone chain processing based on the initial adjacency matrix between nodes includes:

[0146] Symmetrize the initial adjacency matrix between nodes to obtain the symmetrized adjacency strength;

[0147] Compare the symmetrized adjacency strength with a preset coupling strength threshold, retain the adjacency edges greater than the preset coupling strength threshold, and obtain an undirected sparse graph;

[0148] Take each edge of the undirected sparse graph as the coupling strength between nodes as the graph distance, obtain a weighted graph, and use Kruskal algorithm to obtain the corresponding minimum spanning tree; wherein the edges in the minimum spanning tree that satisfy the main chain optimal path criterion constitute the main chain path;

[0149] If the node belongs to the main chain path, it is marked as a main node, otherwise it is a branch node; all main nodes and branch nodes, and the coupling strength between nodes, constitute the approximate topology of the transformer area main line.

[0150] To meet the requirements of the undirected graph, the adjacency matrix between nodes is processed in a symmetrical manner as follows:

[0151] Formula (5)

[0152] Wherein, is the symmetrized adjacency strength used to construct the undirected graph structure; is the coupling strength between nodes and solved by Lasso regression; is the reverse coupling strength between nodes and . and are both solved by Lasso fitting.

[0153] The adjacency matrix element represents the coupling strength between user nodes and in electrical behavior, the larger the value, the more similar the voltage power characteristics, and thus reflecting the possible physical connection relationship between the two, which can be used to construct the sparse topology graph inside the transformer area.

[0154] The preset coupling strength threshold retains the high coupling edge, and the link less than the preset coupling strength threshold is set to 0 to achieve graph sparsity, which is represented as follows:

[0155] Formula (6)

[0156] Wherein, is the adjacency matrix element of the final undirected sparse graph constructed.

[0157] The undirected sparse graph As a weighted undirected graph, each edge in the undirected sparse graph represents the coupling strength between nodes. The coupling strength between nodes is transformed into graph distance, and a weighted graph is constructed as follows:

[0158] Formula (7)

[0159] in, Let be the edge weight matrix. It is an exponential factor.

[0160] For example, A value of 2 is used to enhance the distinguishability of edge weights; A value of 2 amplifies the non-linearity and increases the difference: it further amplifies the importance of strong connections, making the minimum spanning tree more biased towards trunk connections.

[0161] The Kruskal algorithm outputs a tree with the following number of edges. An undirected graph with connected nodes.

[0162] The core idea of ​​Kruskal's algorithm is to sort edges by weight in ascending order, select the connecting edges with the smallest weights from the undirected sparse graph, and construct a backbone electrical connection structure that covers all nodes and has the shortest overall graph distance. The specific steps are as follows:

[0163] Undirected sparse graph , Let be the set of nodes, representing all user nodes in the transformer area; Let be the set of edges in an undirected sparse graph, and let the weight matrix of the edges be... .

[0164] (1) Edge sorting: sort all edges of the undirected sparse graph ( , )∈E according to edge weight Sort by size from smallest to largest;

[0165] (2) Initialize the set: Each node forms a separate set;

[0166] (3) Selecting edges in sequence: Start enumerating from the smallest edge, if the two endpoints of the edge , If they belong to different sets, select the edge and add it to the spanning tree; merge the two sets.

[0167] (4) Termination condition: Repeat step 3 until the number of selected edges equals The minimum spanning tree construction is now complete.

[0168] Obtaining the minimum spanning tree based on the weighted graph ,in, Let be the set of nodes, representing all user nodes in the transformer area; is the edge set of the minimum spanning tree, representing the backbone electrical connection path between user nodes, satisfying:

[0169] Equation (8)

[0170] wherein, is the undirected sparse graph, indicates that the minimum spanning tree belongs to the undirected sparse graph G and is connected.

[0171] The minimization objective indicates that the sum of the graph distances of all selected edges is minimized, ensuring that the total "electrical cost" of the backbone path is minimized.

[0172] The backbone chain refers to a main path selected from the set of all passable paths in the minimum spanning tree, which has the best path stability, the strongest coupling, and the strongest voltage support, and is used as a representative path of the distributed photovoltaic district backbone topology. The backbone chain is determined by the backbone chain optimal path criterion.

[0173] The backbone chain optimal path criterion is as follows:

[0174] Equation (9)

[0175] wherein, is the backbone chain path; is the minimum spanning tree, indicates the set of all traversable paths in the minimum spanning tree, indicates each pair of adjacent nodes in the current path P, , are weight coefficients of the influence of the control voltage stability term and the phase angle difference term on the path score, respectively; are the voltage amplitudes of the nodes , and the phase angles of the nodes , , .

[0176] This objective can be regarded as a multi-factor main path maximization problem. The meaning of this formula is to score each path by enumerating all paths in the minimum spanning tree, and to select a path with the most stable structure, the strongest coupling, and the most consistent parameter fluctuation as the "backbone chain".

[0177] is the coupling strength of the nodes , in the sparse adjacency graph, and the sparse adjacency graph constructs the final edge weight, indicating the connection strength between two users; is the score factor, representing the coupling score of each edge in the path, directly reflecting the physical connection strength. Using the logarithmic form can suppress abnormal edge weights and enhance the discrimination of weak connection edges, making them more physically interpretable in path scoring. The greater the coupling strength, the higher the path score; taking the logarithm moderates the extreme difference and prevents the weight span from being too large;

[0178] 、 are the weight coefficients of the control voltage stability term and the phase angle difference term, respectively, on the influence of the path score; 、 are the phase angles of the node 、 voltage.

[0179] For example, the voltage stability term weight , the phase angle difference term weight take the following values:

[0180] The value range is 0.6-0.8; The value range is 0.2-0.4.

[0181] If the area is dominated by resistive load and the voltage amplitude changes significantly, the =0.7, =0.3;

[0182] If the load is dominated by distributed photovoltaic inverters and the phase angle fluctuates significantly, it can be adjusted to =0.5, =0.5.

[0183] In the following way, the "trunk-branch" structure of the electrical connection in the distributed photovoltaic area is represented.

[0184] Equation (10)

[0185] where, is marked as the trunk structure if the user node ∈ main chain path ; otherwise, it is marked as a branch node .

[0186] The role of step S1 is to form a set of low-voltage distributed photovoltaic area trunk line approximate topology without prior topological information, providing a basis for subsequent clustering analysis and abnormal tracing.

[0187] Step S2, in particular.

[0188] User voltage data is selected during the midday peak photovoltaic output period, as detailed below. Data is obtained from smart meters in the distributed photovoltaic area, and nighttime electrical data for the distributed photovoltaic area is obtained in the same step S11.

[0189] The final adjacency matrix between the constructed nodes includes:

[0190] Normalize the voltage data sequences of each user during the midday peak photovoltaic power output period;

[0191] Based on the normalized user voltage sequence data, the standard deviation of the user node voltage sequence is calculated, and the Pearson correlation coefficient between user nodes is calculated.

[0192] Based on the Pearson correlation coefficient between user nodes and the coupling strength in the undirected sparse graph, user nodes are constructed. With nodes Voltage fluctuation correlation matrix;

[0193] Compute user nodes With nodes Voltage fluctuation covariance;

[0194] Based on user nodes With nodes The voltage fluctuation correlation matrix and voltage fluctuation covariance are used to calculate the node. With nodes The final adjacency matrix between nodes.

[0195] To eliminate the differences in absolute voltage values ​​among different users and highlight the relative patterns of fluctuation behavior, voltage time-series data were normalized, and a structure-guided voltage fluctuation correlation matrix was constructed.

[0196] The voltage time-series data of each user during the midday peak photovoltaic output period are normalized as follows:

[0197] Formula (11)

[0198] in, For users In time The voltage value; for The average voltage over a given time period; The time series length is the total sampling duration of voltage data for each user; the user node is calculated. Mean of voltage series.

[0199] Based on the normalized user voltage sequence data, the variance of the user node voltage sequence is calculated as follows:

[0200] Equation (12)

[0201] wherein, is the user node The variance of the voltage time series; is the user The voltage value standardized sequence at time t, with a mean of 0 and a standard deviation of 1;

[0202] Equation (12) calculates the user node The standard deviation of the voltage sequence minus the mean, which can emphasize the fluctuation behavior, and the output variance reflects the fluctuation intensity, The larger the value, the more volatile the node voltage, indicating that the photovoltaic fluctuation is more intense.

[0203] Based on the variance of the user node voltage time series, the standard deviation of the user node voltage sequence is calculated as follows:

[0204] Equation (13)

[0205] Equation (13) normalizes the voltage sequence, which is convenient for subsequent Pearson correlation coefficient calculation. The reason for normalization is that the voltage amplitude of different users is very different, and it must be normalized for comparison.

[0206] Equation (14)

[0207] wherein, is the user node The Pearson correlation coefficient between measures the similarity of the fluctuation shape; is the standard deviation of the voltage sequence of the user node and

[0208] Calculate the Pearson correlation coefficient, the input is the normalized voltage sequence of two nodes, and the output is a similarity index with a value range of [-1, 1], the larger the value, the stronger the fluctuation synchronization.

[0209] Build a voltage fluctuation correlation matrix between user node and node , as follows:

[0210] Equation (15)

[0211] wherein, is the voltage fluctuation similarity between user node and is the user node and ​​Pearson correlation coefficient between them, is the structure prior enhancement factor, is the coupling strength in the undirected sparse graph, is the power adjustment term to enhance the structure influence.

[0212] is the link strength between users in the above undirected sparse graph; , is the structure prior enhancement factor, exemplarily, generally taken as 0.1-2; is the power adjustment term to enhance the structure influence; if =0, the term does not enhance, if both user nodes are on the backbone, the term will significantly improve their .

[0213] Formula (15) is to calculate the structure-enhanced correlation index (introducing the coupling graph prior ), the function of the formula is to fuse the voltage fluctuation correlation between nodes and the electrical structure prior to construct a more physically credible enhanced similarity matrix for subsequent spectral clustering or graph clustering. The Pearson correlation coefficient is a pure statistical index, which only reflects whether the voltage fluctuation curves of two users are consistent, but ignores their electrical connection relationship. In other words, if the voltage curves of two user nodes are very similar, they may have no electrical connection; while two user nodes are closely connected electrically, if the fluctuation is a little different, a lower value will be given, which will mislead clustering or clustering. Therefore, the formula introduces a structure-guided enhancement factor , for example, =0.55, the correlation is medium and low, =0.95, the structure connection is very strong, μ=1.0, γ=2, then =0.55(1+0.95) 2 =0.55*3.8025=2.09, the original value 0.55 is amplified to 2.09, which obviously reflects the structure compensation improvement.

[0214] The voltage fluctuation covariance of user-user node and node is calculated as follows:

[0215] Formula (16)

[0216] wherein, is the user and a voltage fluctuation covariance value between voltage sequences, for supplementing the similarity evaluation; for the acquired user node at a time of a voltage original value; for the user node a voltage mean value of the entire time sequence.

[0217] Formula (16) is used to solve the covariance term, reflecting the common trend of voltage fluctuation.

[0218] calculate the user node and the final adjacency matrix between nodes , as follows:

[0219] Formula (17)

[0220] wherein, is the voltage fluctuation covariance between the user node and the node , are the standard deviations of the voltage sequences of the user node and the node , respectively, are the measurement weights of the fusion and two similarity degrees.

[0221] is the final adjacency matrix between nodes after fusion; , for fusing two similarity degrees. Formula (17) fuses the final similarity index, and outputs .

[0222] This part of logic is to establish the original fluctuation trend similarity → graph structure enhancement → covariance term supplement → final fusion .

[0223] Based on the final adjacency matrix between nodes, a symmetric normalized Laplacian matrix is constructed and spectral clustering is performed to obtain multiple user clusters with similar voltage fluctuation patterns.

[0224] The purpose is to transform the similarity information between high-dimensional users into a low-dimensional feature representation suitable for clustering, starting from the final voltage fluctuation correlation matrix between nodes , a symmetric normalized Laplacian matrix is constructed, and the feature vector is reduced to a low-dimensional space and clustered.

[0225] First, the computation degree matrix is established and the Laplacian matrix is normalized (Stabilizing the eigenvector in the spectral space avoids clustering bias caused by the degree of different nodes, and is more suitable for stable representation in the spectral space.) as follows:

[0226] Equation (18)

[0227] wherein, is the degree value of the user node , representing the sum of similarity (or connection strength) between the user node and all other nodes, i.e., the "total connection strength" of the user node ;

[0228] is the degree matrix, which is a diagonal matrix with the degrees of each node on the main diagonal , and the rest of the elements are zero; is the matrix dimension, representing is an N×N real matrix, N is the number of user nodes; is a symmetric normalized Laplacian matrix used for spectral clustering, which can stabilize the feature space and make the difference in connection strength of different nodes not cause bias; is the unit matrix; is the weighted similarity matrix guided by voltage fluctuation.

[0229] In equation (18), the diagonal element of the node degree matrix D represents the overall coupling strength between the user node and all other nodes, i.e., the connection ability or total relevance of the node in the weighted similarity graph structure, which is used to reflect the weight status of the node in the global network structure.

[0230] The Laplacian matrix introduces the normalization operation of node degree, effectively weakening the bias effect caused by the difference in connection strength of different nodes, thereby improving the stability and aggregability of the feature representation in the spectral space. This matrix can ensure that the eigenvector obtained in the subsequent feature decomposition process has a clearer geometric structure in the spectral space, and is suitable for further performing K-means clustering.

[0231] Based on the Laplacian matrix, spectral clustering is performed to obtain multiple user clusters with similar voltage fluctuation patterns, including:

[0232] Solving the eigenvector corresponding to the first minimum eigenvalue of the symmetric normalized Laplacian matrix , and composing the spectral feature space with the eigenvector;

[0233] ​ K-means clustering is used by row to obtain multiple user clusters, each of which includes at least one user node.

[0234] The eigenvalue decomposition (spectral dimension reduction) is continued as follows The first eigenvalues corresponding to the eigenvectors of the smallest eigenvalues form a new spectral feature space, in which the structure is easier to be captured by the clustering algorithm. And assemble the eigenvectors Spliced into a feature matrix by column. As follows:

[0235] Equation (19)

[0236] where, is the feature matrix; is the eigenvector corresponding to the first minimum eigenvalue of Lsym; is the dimension of the matrix , indicating that there are user nodes in total, each node is represented by a dimensional vector.

[0237] where, is the diagonal matrix of the graph, and each diagonal element represents the total connection strength of node ; is the low-dimensional feature representation of the th user in the spectral space; is the target cluster number (which can be estimated by Gap Statistic).

[0238] K-means clustering is used by row to achieve automatic clustering and dimension reduction representation of transformer users. (1) Input the feature matrix

[0239] , , which represents the low-dimensional feature vector of the th user node in the spectral space. (2) Initialize the cluster center vector

[0240] .

[0241] (3) The iterative update process is as follows: ① Clustering assignment stage

[0242] For each user node , according to its embedded feature vector , calculate its Euclidean distance with all cluster centers : ​

[0243] Equation (20)

[0244] wherein is a user node a low-dimensional feature in the spectral space; is the center vector of the cluster;

[0245] assigning a user to the cluster closest to it , obtaining if and only if ; is the set of users within the cluster; this step outputs the cluster to which each user belongs .

[0246] ② Center updating stage

[0247] For each cluster , update the center point as the mean of all user feature vectors within the cluster, as follows:

[0248] Equation (21)

[0249] Repeat steps ①-② until the clustering center no longer changes or the maximum number of iterations is reached, completing the clustering.

[0250] Exemplarily, the maximum number of iterations is 50 times.

[0251] Output result: obtain the clustering label corresponding to each user , i.e., the cluster class to which the user belongs.

[0252] The role of step S2 is to divide users with highly similar voltage fluctuation patterns into the same cluster, realize dimension reduction and preliminary grouping of users in the transformer area, and mine structural electrical behavior characteristics.

[0253] Step S3, identify suspected abnormal clusters.

[0254] The identification of suspected abnormal clusters comprises:

[0255] standardizing the voltage and power of each user node within each cluster to obtain standardized voltage and power;

[0256] calculating the Mahalanobis distance between the standardized voltage and the standardized power of each user node within each cluster;

[0257] calculating the dynamic time warping distance between the standardized voltage of each user node and the cluster standardized power curve value, and the Pearson consistency; wherein the cluster standardized power curve value is the average of all user powers within the cluster;

[0258] Based on the Mahalanobis distance, dynamic time warping distance and Pearson consistency, a weighted calculation is performed to obtain the multidimensional consistency anomaly score of user nodes within each cluster.

[0259] Define the power response distribution entropy for each user voltage fluctuation within each cluster, and calculate the intra-cluster entropy variance based on the power response distribution entropy;

[0260] Construct the corresponding probability distribution of each cluster and the reference normal cluster, and calculate the KL divergence between the probability distribution of each cluster and the probability distribution of the reference normal cluster;

[0261] For each cluster, if any one of the following conditions is met: the expected value of the multidimensional consistency anomaly score is greater than the consistency threshold, the intra-cluster entropy variance is greater than the intra-cluster entropy threshold, and the KL divergence is greater than the divergence threshold, then the cluster is judged to be a suspected anomaly cluster.

[0262] User nodes The voltage timing sequence is , A joint normalization feature is constructed to standardize the voltage and power curves of each user while preserving their shape characteristics and eliminating the influence of absolute values, in order to prepare for the consistency of the calculated response, as follows:

[0263] Formula (22)

[0264] in, For users Voltage time series standardization; For users The average value of the voltage time series within this time period; 1 is a column vector of length T consisting entirely of 1s, used to construct a vector similar to the original vector. Mean vectors of the same dimension; For users voltage time series Standard deviation; For users The power time series vector; For users The mean of the power sequence over a time period; For users The standard deviation of the power sequence; For users The standardized power sequence.

[0265] Construct the covariance matrix And define the Mahalanobis uniformity distance (the difference between the reaction and the cluster as a whole). The covariance matrix is ​​used to measure the cluster size. The uniformity variance between voltage and power fluctuations among all users in a cluster reflects whether the cluster exhibits a consistent overall trend. Large differences indicate inconsistent voltage responses and low cluster stability. The Mahalanobis uniformity distance... The larger the value, the more significant the difference between the user's voltage response and the overall power trend of the cluster. Furthermore, when considering load abrupt changes or response delays, the Dynamic Time Warping (DTW) distance is used. A larger DTW distance indicates a greater deviation in the trend pattern between the user's response and the overall power trajectory, which can be used to identify delay anomalies or structural anomalies. Users whose fluctuations remain large after nonlinear alignment between sequences should be given special attention.

[0266] Formula (23)

[0267] Formula (24)

[0268] Formula (25)

[0269] Formula (26)

[0270] in, The Mahalanobis uniformity distance (the difference between the response and the cluster as a whole); No. The normalized power sequence of each cluster; For users With the The dynamic time warping distance of the cluster average power curve; Π is the set of legal matching paths used to align moments in two time series; , users respectively , At any moment Standardized voltage value; No. The normalized power curve value of each cluster at time s For users With the Cluster voltage fluctuation correlation coefficient.

[0271] The multidimensional consistency anomaly scores of user nodes are as follows:

[0272] Formula (27)

[0273] in, For user nodes The Mahalanobis distance between the standardized voltage and the standardized power; The dynamic time-warped distance between the standardized voltage of the user node and the cluster standardized power curve value; is the Pearson consistency; , and are the weights of , and respectively.

[0274] is the multi-dimensional consistency anomaly score value of the user in the th cluster; the higher the multi-criteria consistency anomaly score value of the user node , the more inconsistent the user's response is with the overall power change of the cluster, and the more likely it is to be an abnormal user; is the weighting coefficient, .

[0275] The role of step S3 is to perform multi-criteria consistency scoring on each user cluster by constructing a joint feature space and a consistency measure function, to identify and determine abnormal clusters, thereby providing a basis for further abnormal user identification.

[0276] The power response distribution entropy of each user voltage fluctuation is defined, and the entropy variance within the cluster is calculated as a stability indicator:

[0277] Equation (28)

[0278] wherein, , is a normalization constant (to prevent log explosion); is the total power curve (normalized) of the th cluster at time ; represents the noise of the user response; is the entropy variance within the th cluster, and the corresponding probability distribution of the current normal response cluster and the reference normal cluster is constructed , , and the KL divergence is defined as follows:

[0279] Equation (29)

[0280] wherein, is the th clustering cluster, and is the current generation of judgment cluster; is the reference normal cluster, such as the trunk cluster. If the fluctuation behavior distribution of a cluster deviates greatly from the trunk cluster (or normal cluster), it is an abnormal cluster. is the normalized power fluctuation distribution of the cluster , Probability distribution of normal cluster power response; KL divergence value, the larger the value, the greater the difference between the two clusters.

[0281] Normalization constant, normalize the power and voltage to a value between [0, 1], so that Limit between 0-1.

[0282] Final abnormal cluster judgment:

[0283] Formula (30)

[0284] Where, The weight coefficients of the consistency score mean, KL divergence value and distribution deviation threshold stability index, respectively; Abnormal cluster.

[0285] Where, Indicates as follows:

[0286] Formula (31)

[0287] Indicates the mean of the user abnormal score in the cluster, reflecting the overall deviation degree.

[0288] Exemplarily, The value is 0.8; The value is 0.5; The value is 1.0; in specific application, it can be changed according to specific needs.

[0289] The role of step S3 is to construct a cluster overall power curve for each user cluster, and to determine whether the cluster is an abnormal cluster by calculating the fluctuation consistency index between the voltage sequence of each user in the cluster and the cluster power curve.

[0290] Step S4, specifically.

[0291] Node-level identification is performed on the suspected abnormal cluster to obtain an abnormal metering user, including:

[0292] For the suspected abnormal cluster, extract the graph structure subgraph of the suspected abnormal cluster in the final adjacency matrix between nodes;

[0293] Based on the graph structure subgraph, construct a symmetric normalized adjacency submatrix;

[0294] Extract the input feature vector of each user node in the suspected abnormal cluster;

[0295] ​​​Based on the symmetric normalized adjacency sub-matrix and the input feature vector, a two-layer graph convolutional network is used to output a user node anomaly probability;

[0296] If the user node anomaly probability is greater than the upper limit of the probability threshold, the user node is determined to be an abnormal user.

[0297] Abnormal user identification and trace positioning based on graph neural network. Assuming that the abnormal cluster is , extract its subgraph in the original adjacency matrix:

[0298] Formula (32)

[0299] wherein, is the sparse graph adjacency matrix mentioned above; is the graph structure subgraph of the i-th cluster; is the number of nodes in the cluster.

[0300] Construct a symmetric normalized adjacency matrix, which is a standard GCN (Graph Neural Network) preprocessing to prevent neighbor node scale imbalance in graph convolution. The normalized adjacency matrix is as follows:

[0301] Formula (33)

[0302] wherein, The symmetric normalized adjacency matrix represents the standardized structure graph used for graph neural network.

[0303] The input feature vector includes voltage mean, voltage standard deviation, power mean, power standard deviation, voltage kurtosis, voltage skewness, voltage fluctuation upper limit, voltage fluctuation lower limit, similarity correlation feature, and frequency domain feature vector.

[0304] For user , convert its voltage and power sequence into fixed-dimension embedding features, basic statistical feature vector (extract statistical description features of user voltage and power time series to reflect their fluctuation pattern, distribution form and stability), similarity correlation feature (reflecting voltage response hysteresis) and high frequency domain feature (catch implicit periodic disturbance patterns such as photovoltaic output and other high frequency features) are as follows, and the final node input feature is the splicing vector . As follows:

[0305] Formula (34)

[0306] ​Equation (35)

[0307] Equation (36)

[0308] Equation (37)

[0309] wherein , are the voltage, power mean values, respectively; , are the voltage, power standard deviations, respectively; is the kurtosis, indicating the sharpness, whether an abnormal jump occurs; is the skewness, indicating the degree of deviation from the center, whether voltage imbalance occurs; , are the upper and lower limits of voltage fluctuation, respectively; is the basic statistical feature vector; is the similarity correlation feature; , are the normalized voltage, power, respectively; τ is the time lag quantity; FFT(.) is the fast Fourier transform operation; is the frequency component amplitude; is the frequency domain reserved dimension (e.g. F=5); is the frequency domain feature vector; is the total dimension of node input.

[0310] The upper limit of voltage fluctuation refers to the maximum value of the user voltage time series in a specified time period; the lower limit of voltage fluctuation refers to the minimum value of the user voltage time series in a specified time period.

[0311] The two-layer graph convolution network GCN is used to output the user node anomaly probability, including:

[0312] Based on the symmetric normalized adjacent submatrix and the input feature vector, the first layer propagation is performed to obtain the hidden layer node feature matrix;

[0313] Based on the hidden layer node feature matrix and the symmetric normalized adjacent submatrix, the second layer propagation is performed to obtain the output anomaly probability;

[0314] The output anomaly probability is subjected to binary classification determination to obtain the user node anomaly probability.

[0315] Subsequently, a two-layer graph convolution network (GCN) is used to construct node representation, as follows:

[0316] First layer propagation:

[0317] Equation (38)

[0318] The second layer propagation (output layer) is:

[0319] Formula (39)

[0320] Where X is the node input feature matrix, from the concatenation of the above ; is a symmetric normalized adjacency matrix; is the first layer weight; is the second layer weight, is a nonlinear activation function ReLU; is the hidden layer node feature matrix; is the output anomaly probability. Define the output probability and perform a binary classification decision rule:

[0321] Formula (40)

[0322] Formula (41)

[0323] Apply the Sigmoid activation to the to obtain the anomaly probability , The closer the numerical value is to 1, the more likely it is an abnormal user.

[0324] A preset probability threshold θ∈[0.5,0.9], after completing the node-level anomaly scoring calculation of the graph neural network, the system determines the abnormality of each user node according to the set threshold, The user is determined as the final abnormal user.

[0325] As shown in Figure 4 , an abnormal user in the second suspected abnormal cluster is identified using the method in the application.

[0326] All abnormal users are established as an abnormal user set u, which includes key users whose voltage response and power fluctuation deviate significantly from the behavior mode of the same cluster, while retaining the cluster label, structure position and abnormal score index thereof, for forming a final metering data anomaly report, and realizing automatic and high-precision tracing of abnormal users in a distributed area.

[0327] The function of step S4 is to extract the voltage and power time series features of each user in the user cluster determined as abnormal, and use the graph neural network model and graph convolution to aggregate local neighbor information for node-level anomaly recognition, thereby realizing the process of automatic identification and fault positioning of metering data anomalies in a distributed photovoltaic area.

[0328] In summary, the distributed photovoltaic area metering data anomaly tracing method according to the embodiment of the application has the following beneficial effects:

[0329] 1、The present application can mine the internal laws between voltage and power data without topology drawings by fusing intelligent power distribution network physical modeling and data-driven algorithms, realizing accurate identification and real-time positioning of abnormal users in the transformer area. This method combining physical model and data-driven has higher adaptive ability and identification accuracy than the traditional method relying only on artificial experience setting parameters, especially in the absence of topology support or facing high-dimensional time series data, it can more accurately identify abnormal sources and reduce the risk of missed detection and misjudgment;

[0330] 2、The present application can realize rapid and accurate identification of abnormal users by constructing a model based on graph convolution network (GCN), combining topology information and user power consumption characteristics, avoiding the low efficiency problem of checking house by house, thereby significantly improving the efficiency of abnormal positioning. This method not only can quickly identify abnormal users and improve the efficiency of abnormal positioning, but also can reduce the waste of time and resources caused by checking house by house;

[0331] 3、The present application can support intelligent operation and maintenance and rapid disposal of abnormality of power distribution network, through automatic abnormal detection and positioning, providing real-time decision support for power distribution network operation and maintenance personnel, supporting intelligent operation and maintenance, thereby improving the operation efficiency and reliability of intelligent power distribution network. This intelligent operation and maintenance method helps to reduce human errors, improve response speed and ensure stable operation of the power grid.

[0332] Those skilled in the art can understand that all or part of the processes of the above-mentioned embodiments can be completed by a computer program instructing related hardware, and the program can be stored in a computer readable storage medium. The computer readable storage medium includes a magnetic disk, an optical disk, a read-only memory or a random access memory.

[0333] The above is only the preferred embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for tracing the source of anomalies in metering data of distributed photovoltaic power stations, characterized in that, The method comprises the following steps: Collecting electrical data of distributed photovoltaic substation and users at night without photovoltaic output, and constructing an approximate topology of main lines in the substation; Based on the collected voltage data of users at the high photovoltaic output period in the afternoon, an inter-node final adjacency matrix between nodes is constructed; Based on the inter-node final adjacency matrix, clustering is performed to divide the users into multiple user clusters; For each user cluster, a cluster power curve is constructed to identify suspected abnormal clusters; For the suspected abnormal clusters, node-level identification is performed to obtain abnormal metering users; Both the distribution transformer and the user are regarded as nodes; The electrical data includes the voltage amplitude of the distribution transformer, the voltage, active power and reactive power of each user node; the construction of the approximate topology of the main lines in the substation comprises: Selecting the user node with the highest voltage from the nodes closest to the distribution transformer as the current node, traversing the remaining nodes, calculating the cost function value between the current node and each of the remaining nodes, and selecting the node with the minimum cost function value as the new current node; continue polling the remaining nodes that have not been selected until all nodes are polled; Based on the voltage of the user nodes selected in turn, data fitting is performed using a Lasso regression model to obtain an initial adjacency matrix between nodes with the coupling strength between nodes as the matrix elements; Based on the initial adjacency matrix between nodes, main chain processing is performed to obtain the approximate topology of the main lines in the substation; The main chain processing based on the initial adjacency matrix between nodes comprises: Symmetric adjacency processing is performed on the initial adjacency matrix between nodes to obtain the adjacency strength after symmetry; The adjacency strength after symmetry is compared with a preset coupling strength threshold, and the adjacency edges greater than the preset coupling strength threshold are retained to obtain an undirected sparse graph; Each edge of the undirected sparse graph represents the coupling strength between nodes as the graph distance to obtain a weighted graph, and the Kruskal algorithm is used to obtain the corresponding minimum spanning tree; wherein the edges in the minimum spanning tree that meet the main chain optimal path criterion constitute the main chain path; If the node belongs to the main chain path, it is marked as a main node, otherwise it is a branch node; all main nodes and branch nodes, as well as the coupling strength between nodes, constitute the approximate topology of the main lines in the substation; The construction of the final adjacency matrix between nodes comprises: Normalizing the voltage data sequence of each user at the high photovoltaic output period in the afternoon; Based on the normalized user voltage sequence data, the standard deviation of the user node voltage sequence is calculated, and the Pearson correlation coefficient between the user nodes is calculated; constructing user nodes based on Pearson correlation coefficients between the user nodes, and coupling strengths in the undirected sparse graph correlation matrix of voltage fluctuations with nodes ​ Computing user nodes with nodes voltage fluctuation covariance; based on the user nodes a matrix of voltage fluctuations and a covariance of voltage fluctuations associated with the nodes a matrix of final adjacencies between the nodes a matrix of final adjacencies between the nodes ​ The method for identifying suspected abnormal clusters comprises: Standardizing the voltage and power of the user nodes in each cluster to obtain standardized voltage and power; Calculating the Mahalanobis distance between the standardized voltage and the standardized power of each user node in each cluster; Calculating the dynamic time warping distance between the standardized voltage of each user node and the cluster standardized power curve value, and the Pearson consistency; wherein the cluster standardized power curve value is the average of all user powers in the cluster; Based on the Mahalanobis distance, dynamic time warping distance and Pearson consistency, a multi-dimensional consistency anomaly score value of the user nodes in each cluster is calculated by weighted calculation. An entropy of power response distribution of each user voltage fluctuation in each cluster is defined, and an intra-cluster entropy variance is calculated based on the entropy of power response distribution; A corresponding probability distribution of each cluster and a reference normal cluster is constructed, and a KL divergence of the probability distribution of each cluster and the probability distribution of the reference normal cluster is calculated; For each cluster, if the mathematical expectation of the multi-dimensional consistency anomaly score value is greater than a consistency threshold, the intra-cluster entropy variance is greater than an intra-cluster entropy threshold, and the KL divergence is greater than a divergence threshold, any one of which is true, the cluster is determined as a suspected abnormal cluster.

2. The distributed photovoltaic regional metering data anomaly tracing method according to claim 1, characterized in that, The cost function of the current node and the next node is represented as follows: ； wherein, Vcur and Vnext are the voltages of the current node and the next node, respectively; Pcur and Pnext are the active powers of the current node and the next node, respectively; a and b are the weight coefficients of the voltage difference term and the active power term, respectively.

3. The distributed photovoltaic metropolitan area metering data anomaly root cause method of claim 1, wherein, The Lasso regression model is represented as follows: ； wherein, is the connection vector of the node , and λ is the L1 regularization coefficient; The coupling strength values between nodes are obtained by using a Lasso regression model to fit the data ; with the coupling strength value between the nodes As a matrix element, the initial adjacency matrix between the nodes is obtained.

4. The distributed photovoltaic district metering data anomaly tracing method according to claim 1, which is adjusted in that the backbone chain optimal path criterion is as follows: ； wherein, is the backbone chain path; is the minimum spanning tree, represents the set of all traversable paths in the minimum spanning tree, represents each pair of adjacent nodes in the current path P, , are the weight coefficients of the influence of the control voltage stability term and the phase angle difference term on the path score, respectively; are the voltage amplitudes of the nodes , the phase angle of the voltage; are the voltage amplitudes of the nodes , , respectively, is the coupling strength of the nodes , in the sparse adjacency graph.

5. The distributed photovoltaic metropolitan area metering data anomaly root cause method of claim 1, wherein, Constructing user nodes Correlation matrix with voltage fluctuations of nodes as follows: ； wherein, is the user node has a voltage fluctuation similarity with ; is the user node has a Pearson correlation coefficient with , is a structure prior enhancement factor, is a coupling strength in the undirected sparse graph, is a power adjustment term to enhance the structure impact.

6. The distributed photovoltaic metropolitan metering data anomaly root cause tracing method according to claim 5, characterized in that, Computing user nodes With nodes The inter-node final adjacency matrix of nodes, as follows: ； where, is the voltage of the user node is the voltage of the node is the voltage fluctuation covariance of the user node and the node is the voltage sequence standard deviation of the user node and the node is the measurement weight of the fusion and the two similarity measures.

7. The distributed photovoltaic metropolitan metering data anomaly root cause tracing method according to claim 6, characterized in that, Based on the final inter-node adjacency matrix, a symmetric normalized Laplacian matrix is constructed and spectral clustering is performed to obtain a plurality of user clusters with similar voltage fluctuation patterns.

8. The distributed photovoltaic metropolitan metering data anomaly root cause tracing method according to claim 7, characterized in that, Spectral clustering is performed based on the Laplacian matrix to obtain a plurality of user clusters with similar voltage fluctuation patterns, including: Solving symmetric normalized laplacian matrix The eigenvectors corresponding to the first minimum eigenvalues of the normalized laplacian matrix are obtained ; To K-means clustering is used by row to obtain a plurality of user clusters, each cluster including at least one user node.

9. The distributed photovoltaic metropolitan area metering data anomaly root cause method of claim 1, wherein, A multi-dimensional consistency anomaly score value of a user node is as follows: ； wherein, is the Mahalanobis distance between the normalized voltage of the user node and the normalized power of the cluster; is the dynamic time warping distance between the normalized voltage of the user node and the cluster normalized power curve value; is the Pearson consistency; , and are the weights of , and respectively.

10. The distributed photovoltaic metropolitan area metering data anomaly root cause method of claim 1, wherein, Node-level identification is performed on the suspected abnormal cluster to obtain an abnormal metering user, including: For the suspected abnormal cluster, a graph structure subgraph of the suspected abnormal cluster in the final inter-node adjacency matrix is extracted; Based on the graph structure subgraph, a symmetric normalized adjacency submatrix is constructed; An input feature vector of each user node in the suspected abnormal cluster is extracted; Based on the symmetric normalized adjacency submatrix and the input feature vector, a user node anomaly probability is output by a two-layer graph convolutional network; If the user node anomaly probability is greater than an upper limit range of a probability threshold, the user node is determined as an abnormal user.

11. The distributed photovoltaic metropolitan metering data anomaly root cause tracing method of claim 10, wherein, A two-layer graph convolutional network GCN is used to output a user node anomaly probability, including: Based on the symmetric normalized adjacency submatrix and the input feature vector, a first layer propagation is performed to obtain a hidden layer node feature matrix; Based on the hidden layer node feature matrix and the symmetric normalized adjacency submatrix, a second layer propagation is performed to obtain an output anomaly probability; The output anomaly probability is subjected to binary classification determination to obtain a user node anomaly probability.

12. The distributed photovoltaic metropolitan metering data anomaly root cause method according to any one of claims 10-11, characterized in that, The input feature vector includes a voltage mean value, a voltage standard deviation, a power mean value, a power standard deviation, a voltage kurtosis, a voltage skewness, a voltage fluctuation upper limit, a voltage fluctuation lower limit, a similarity correlation feature, and a frequency domain feature vector.

Images