Distributed power supply access unit remote positioning method and system

Abstract

The invention discloses a distributed power supply access unit remote positioning method and system, and the method comprises the steps: solving an initial value of a connectivity matrix and an initial value of an inverse covariance matrix of each bus, and carrying out the building of a power distribution network topology model based on the initial values; according to the method, inverse covariance matrix real-time estimation is carried out through data collected in real time, the node connectivity condition is judged, a real-time connectivity matrix is constructed, topological changes of the power distribution network are tracked, and topological updating is achieved; based on the connectivity matrix estimated in real time, obtaining the connection state of a line in the system, and obtaining a key monitoring area; searching the key monitoring area through a breadth-first search algorithm to complete coarse positioning; a suspected access point is marked for the roughly positioned area, a voltage phase angle sequence of the suspected access point is extracted, feature vectors extracted by a SeriesTransform model are input into a classifier, and a prediction result is obtained; the method has the advantage that the positioning of a large number of distributed power supply access units in the power distribution network is realized.

Description

Technical Field

[0001] This invention relates to the field of distributed power sources in power systems, and specifically to a method and system for remotely locating distributed power source access units. Background Technology

[0002] Distributed energy resources (DERs) distribute electricity production across numerous small power generation units, typically located near loads. This reduces transmission distance, lowers transmission losses, and improves power supply reliability. However, with the rapid increase in the number of DERs, accurately, efficiently, and scalably locating the vast number of DER units connected to the distribution network remotely remains a challenge.

[0003] Traditional equipment struggles to support the large-scale grid connection demands of DERs: First, frequent changes in network topology due to switching operations and fault isolation make it difficult for centralized control systems, such as SCADA systems, to achieve real-time network status updates due to communication delays or data gaps. Second, measurement equipment in medium and low-voltage distribution networks has significant blind spots, making it impossible to directly locate massive numbers of DER nodes using limited sensors. Third, the number of DERs is growing exponentially, and centralized systems cannot meet the high-concurrency positioning requirements, making it difficult to complete the task within a second-level response time.

[0004] Chinese patent publication CN117200234A discloses a method for fast tracking of intermittent distributed energy power flow. It constructs a topology analysis engine based on a graph database. The topology model of the distributed power supply system is input into the topology analysis engine, which stores the model in the graph database and performs a topology search based on a graph search algorithm to obtain the power supply range of each distributed power source in the system. However, this patent application cannot accurately determine the location of the numerous distributed power source access units in the distribution network. Summary of the Invention

[0005] The technical problem to be solved by this invention is that the existing technology cannot determine the location of the large number of distributed power generation access units in the distribution network.

[0006] This invention solves the above-mentioned technical problems through the following technical means: a remote positioning method for a distributed power supply access unit, comprising the following steps:

[0007] S1. In the offline phase, acquire historical data of SCADA switch status, μPMU measurement data and smart meter data, solve for the initial values ​​of the connectivity matrix and inverse covariance matrix of each bus, and build the distribution network topology model.

[0008] S2. Real-time estimation of the inverse covariance matrix is ​​performed using real-time collected data to determine node connectivity, construct a real-time connectivity matrix, track changes in the distribution network topology, and achieve topology updates.

[0009] S3. Based on the real-time estimated connectivity matrix, the connectivity status of the lines in the system is obtained. Sensitivity detection is performed on the buses in the connected state to obtain key monitoring areas. The key monitoring areas are searched using a breadth-first search algorithm to exclude disconnected feeder sections and complete coarse localization. Suspected access points are marked in the coarsely located areas. For nodes marked as suspected access points, their voltage phase angle sequence is extracted and input into the feature vector extracted by the pre-trained SeriesTransformer model. The prediction result is obtained through the classifier, and the output is whether the node is a distributed power supply access unit.

[0010] This invention utilizes historical data to obtain initial values ​​for the connectivity matrix and the inverse covariance matrix. It then combines these initial values ​​with real-time acquired data to estimate the inverse covariance matrix in real time, completing topology updates. Based on the updated topology, coarse localization is performed. Suspected access points are marked in the coarsely located areas. For nodes marked as suspected access points, their voltage phase angle sequences are extracted and input into a pre-trained SeriesTransformer model. These extracted feature vectors are then fed into a classifier to determine whether the node is a distributed power supply access unit, thus completing the localization of the vast number of distributed power supply access units in the distribution network.

[0011] Furthermore, S1 includes:

[0012] Historical data acquisition of S11, SCADA switch status, μPMU measurement data and smart meter data. μPMU measurement data is the phase angle and amplitude of bus voltage, and smart meter data is the power injection data.

[0013] S12. Model the distribution network as a graph structure G(v,e), where node v represents a bus and edge e represents a line switch. The sample covariance matrix S0 is obtained by summing and averaging the bus voltage phase angle change sequence and the transpose of the sequence within a time window.

[0014] S13. Assume the inverse covariance matrix Θ has the same form as the sample covariance matrix. Initially, the inverse covariance matrix Θ is a zero matrix. The inverse covariance matrix Θ is solved using either the first solution method, the second solution method, or a weighted fusion of both methods. The first solution method involves continuously and randomly adjusting the value of the inverse covariance matrix Θ during the solution process and calculating the value of the first objective function. The inverse covariance matrix Θ corresponding to the minimum value of the first objective function is the estimated value of the solved inverse covariance matrix Θ. The first objective function is: Where Tr() represents the trace of the matrix, i.e., the sum of the elements on the main diagonal of the matrix, ln det() represents the natural logarithm of the matrix determinant, and Θ prev Let be the inverse covariance matrix of the previous time step, β be the regularization parameter, and ||||1 denote the L1 norm;

[0015] The second solution method involves dividing the distribution network into several sub-regions. Within each sub-region, the value of the inverse covariance matrix is ​​randomly adjusted, and the value of the second objective function is calculated. The inverse covariance matrix corresponding to the minimum value of the second objective function is the estimated value of the inverse covariance matrix for the current round obtained from solving for that sub-region. The second objective function is... i represents the i-th sub-region, Θ i Let ||Θ| represent the inverse covariance matrix of the i-th sub-region; when the ||Θ| of adjacent iterations (k) -Θ (k-1) || F / ||Θ (k) || F <10 -4 The iteration stops when the estimated value of the inverse covariance matrix of the sub-region is obtained, Θ. (k) Let |||| represent the inverse covariance matrix obtained in the k-th iteration. F The Frobenius norm of a matrix is ​​a measure of its magnitude. It is calculated by summing the squares of the absolute values ​​of all elements in the matrix and then taking the square root.

[0016] S14. For the inverse covariance matrix Θ, if Θ ij If the value is less than or equal to -η, then nodes ij are considered connected; otherwise, they are considered disconnected. An estimated connectivity matrix A is then formed based on the connectivity assessment results between each node. est ; where Θ ij Let η be the element in the i-th row and j-th column of the inverse covariance matrix Θ, and let η be the threshold and η = |μ|. non-conn |+3σ non-conn μ non-conn σ is the mean of the elements in the inverse covariance matrix Θ. non-conn Let Θ be the standard deviation of the elements in the inverse covariance matrix; compare the estimated connectivity matrix A. est By comparing the estimated connectivity between nodes with the actual SCADA switch states, it is determined whether the estimated connectivity between nodes is consistent with the actual switch states. The accuracy and recall are then calculated. If the accuracy or recall does not meet the requirements, the process returns to step S13 above to re-estimate the inverse covariance matrix until the recall and accuracy meet the requirements, thus obtaining the final inverse covariance matrix.

[0017] S15. Based on the final inverse covariance matrix, the final estimated connectivity matrix is ​​obtained, the connectivity of nodes in the distribution network is obtained, and the distribution network topology model is built by combining the data collected in the offline stage.

[0018] Furthermore, S2 includes:

[0019] Real-time acquisition of S21, SCADA switch status, μPMU measurement data and smart meter data;

[0020] S22. Real-time acquired data is timestamped and the time delay is compensated. The μPMU measurement data is input into the trained isolated forest model for detection to obtain anomaly scores. When the anomaly score exceeds the preset score threshold, the voltage phase angle data corresponding to the anomaly score is removed. For SCADA switch states, the difference between the real-time value and the historical mean of each switch state is calculated. If it exceeds the historical mean ±3σ and there is no corresponding switch event, the data point is marked as bad data and repaired by interpolation of data before and after the time. σ is the standard deviation of the switch state data. After the above processing, the standardized dataset D(t) = [Δδ(t), V(t), P(t), Q(t), switch state at time t] is obtained, where Δδ(t) is the phase angle change sequence at time t, P(t) is the active power at time t, Q(t) is the reactive power at time t, and V(t) is the voltage amplitude at time t.

[0021] S23. Construct a state vector x = [δ1, V1, δ2, V2, ..., δ] using a standardized dataset. n V n ] T , where δ i Let V be the voltage phase angle at node i. i Let z be the voltage magnitude at node i, and let z = [z1, z2, ..., z2]. n ] T The expression z = h(x) + e satisfies z = h(x) + e, where z n Let represent the injected power and branch power flow of the nth node, h(x) be a nonlinear measurement function describing the relationship between the measured value and the system state, and e be the measurement error vector.

[0022] The distribution network is divided into several sub-regions, and each edge gateway is responsible for the local state estimation of one sub-region. The local optimization objective function for each sub-region k is: Among them, z k h represents the measurement value of the edge gateway corresponding to sub-region k. k (x) is the nonlinear measurement function of subregion k, w k h is continuously adjusted to assign measurement weights to the edge gateways corresponding to sub-region k. kThe matrix parameters in (x) are used to calculate the value of the local optimization objective function. When the objective function is minimized, the update is stopped, and the corresponding optimized h is obtained. k (x), will optimize h k Substituting the node injected power and branch power flow into the formula z=h(x)+e, we obtain the state vector x, which is also the estimated result of the node's voltage phase angle and voltage amplitude. Based on this, we complete the voltage phase angle and voltage amplitude data that were removed from the standardized dataset. In practical applications, the nonlinear measurement function is the power flow equation, or it can be other set relationships between injected power and voltage phase angle and voltage amplitude. The purpose of this step is to fit the relationship, so that the voltage phase angle and voltage amplitude can be measured based on the injected power and branch power flow.

[0023] S24. Real-time estimation of the inverse covariance matrix. This real-time estimation process uses either a third objective function, the ADMM algorithm, or a weighted fusion of the results from both methods. The third objective function is: Where, Σ S Let Θ(t) be the sample covariance matrix of the current time window, Θ(t) be the inverse covariance matrix at time t, and λ be another regularization parameter. During the solution process using the third objective function, the value of the inverse covariance matrix is ​​continuously and randomly adjusted, and the value of the third objective function is calculated. The inverse covariance matrix Θ corresponding to the minimum value of the third objective function is the real-time estimate of the solved inverse covariance matrix Θ. It should be noted that the inverse covariance matrix represents the connectivity of the network; therefore, although its value is randomly selected during the training process, it should be constrained to be within the range of 0 to 1. The random selection of the inverse covariance matrix in other places follows the same principle.

[0024] S25. Compare the matrix elements in the currently obtained real-time estimate of the inverse covariance matrix with the threshold η to obtain the real-time estimated connectivity matrix. Based on the connectivity matrix, obtain the connectivity status of nodes in the distribution network. Combine the estimated results of the node voltage phase angle and voltage amplitude, as well as the phase angle change sequence and power data, to update the distribution network topology. The principle and process of comparing the matrix elements in the real-time estimate of the inverse covariance matrix with the threshold η to obtain the real-time estimated connectivity matrix are the same as those in S15 above, and will not be repeated here.

[0025] Furthermore, the ADMM algorithm decomposes the problem into two sub-problems: covariance fitting and sparsity constraints. An auxiliary variable Z' = Θ(t) is introduced. Through this constraint, the original problem can be split into two independent sub-problems for alternating optimization. One sub-problem is updating Θ, which handles the data fitting part; the other sub-problem is updating Z', which handles the sparsity constraint part. Therefore, the ADMM algorithm's solution process is as follows:

[0026] S241. Initialization: Let Θ(0) = Z′(0) = (0) = 0, the Lagrange multiplier U and the penalty parameter ρ = 10, Z(0) means that this auxiliary variable is initialized to a zero matrix at the 0th iteration of the algorithm, and U(0) means that the initial penalty for the constraint Z′ = Θ is zero at the beginning of the algorithm, providing an unbiased starting point;

[0027] S242, Θ Update: Fix Z′( k) and U (k) Solve

[0028]

[0029] Where k represents the k-th iteration; Θ in this formula is randomly adjusted, and each round of Θ update process is based on the previous round of Θ and fine-tuned until the formula reaches the minimum value to obtain the current round of Θ. Therefore, compared with general random adjustment, this adjustment process is based on the result of the previous round, and the convergence speed is faster.

[0030] S243, Z′ Update: Fixed Θ (k+1) and U (k) Solve

[0031]

[0032] The above calculation principle is based on a fixed Θ. (k+1) and U (k) In this formula, Z′ is randomly adjusted. Each round of Z′ update is based on the previous round of Z′ and fine-tuned until the formula reaches the minimum value to obtain the current round of Z′. Therefore, compared with the general random adjustment, this adjustment process is based on the result of the previous round and converges faster.

[0033] S244, Lagrange multiplier update: U (k+1) =U (k) +Z'( k+1) -Θ (k+1) ;

[0034] S245. Determine whether the convergence condition is met, and determine ||Θ (k+1) -Z′ (k+1) If || < ∈, and this condition is not met, return to step S242 to continue iterating; if it is met, stop iterating, and Θ is then... (k+1) This is the real-time estimate of the inverse covariance matrix, where ∈ is a pre-defined positive number.

[0035] Furthermore, S3 includes:

[0036] S31. Based on the real-time estimated connectivity matrix, the connectivity status of the lines in the system is determined. Sensitivity detection is performed on the buses in the connectivity state. The sensitivity detection involves calculating the sensitivity matrix of each bus to the μPMU measurement. δ PMU This indicates the voltage phase angle measured by the μPMU. This represents finding the partial derivative, P. DER This represents the active power of the bus in a connected state; when S sen When the number of elements exceeds a set threshold, the bus area corresponding to that element is designated as a key monitoring area. The key monitoring area is searched using a breadth-first search algorithm to exclude disconnected feeder segments and complete the coarse positioning.

[0037] S32. Solve the power flow equations using the Newton-Raphson method to inversely deduce the node injection power; if node P i (t)>1.2×rated capacity and A est (y) indicates that the node is connected to the distributed power supply feeder, then it is marked as a suspected access point, P i (t) represents the injected power of the i-th node at time t, A est (t) represents the connectivity matrix at time t, and the distributed power supply access feeder refers to the feeder segment determined by S31;

[0038] S33. For nodes marked as suspected access points, extract their voltage phase angle sequence, convert it into an input format suitable for the SeriesTransformer model, input it into the pre-trained SeriesTransformer model, input the feature vector extracted by the pre-trained SeriesTransformer model into the classifier, obtain the prediction result through the classifier, and output whether the node is a distributed power access unit. The input format suitable for the SeriesTransformer model is a three-dimensional tensor, which consists of the number of samples in a single input, the voltage phase angle sequence, and the dimension of the input features.

[0039] Furthermore, the processing procedure of the SeriesTransformer model is as follows:

[0040] The input data is processed by the inner layer, and the result is used to calculate attention. The result of the attention calculation and the inner layer processing is added together and then fed into the normalization layer. The output of the normalization layer is fed into the feedforward layer. The output of the feedforward layer and the output of the normalization layer are added together, normalized again, and then linearly projected to output the final result.

[0041] The present invention also provides a remote positioning system for distributed power supply access units, comprising:

[0042] The initialization configuration module is used to acquire historical data of SCADA switch status, μPMU measurement data and smart meter data during the offline phase, solve for the initial values ​​of the connectivity matrix and inverse covariance matrix of each bus, and build the distribution network topology model.

[0043] The real-time update module is used to estimate the inverse covariance matrix in real time using real-time collected data, determine the connectivity of nodes, construct a real-time connectivity matrix, track changes in the distribution network topology, and realize topology updates.

[0044] The collaborative localization module is used to determine the connectivity status of lines in the system based on the real-time estimated connectivity matrix, perform sensitivity detection on buses in the connected state to identify key monitoring areas, and search the key monitoring areas using a breadth-first search algorithm to exclude disconnected feeder sections, thus completing coarse localization. Suspected access points are marked in the coarsely located areas. For nodes marked as suspected access points, their voltage phase angle sequences are extracted and input into the feature vectors extracted by the pre-trained SeriesTransformer model, which are then fed into a classifier. The classifier obtains the prediction result and outputs whether the node is a distributed power supply access unit.

[0045] Furthermore, the initialization configuration module is also used for:

[0046] Historical data acquisition of S11, SCADA switch status, μPMU measurement data and smart meter data. μPMU measurement data is the phase angle and amplitude of bus voltage, and smart meter data is the power injection data.

[0047] S12. Model the distribution network as a graph structure G(v,e), where node v represents a bus and edge e represents a line switch. The sample covariance matrix S0 is obtained by summing and averaging the bus voltage phase angle change sequence and the transpose of the sequence within a time window.

[0048] S13. Assume the inverse covariance matrix Θ has the same form as the sample covariance matrix. Initially, the inverse covariance matrix Θ is a zero matrix. The inverse covariance matrix Θ is solved using either the first solution method, the second solution method, or a weighted fusion of both methods. The first solution method involves continuously and randomly adjusting the value of the inverse covariance matrix Θ during the solution process and calculating the value of the first objective function. The inverse covariance matrix Θ corresponding to the minimum value of the first objective function is the estimated value of the solved inverse covariance matrix Θ. The first objective function is: Where Tr() represents the sum of the elements on the main diagonal of the matrix, ln det() represents the natural logarithm of the matrix determinant, and Θ prev Let be the inverse covariance matrix of the previous time step, β be the regularization parameter, and ||||1 denote the L1 norm;

[0049] The second solution method involves dividing the distribution network into several sub-regions. Within each sub-region, the value of the inverse covariance matrix is ​​randomly adjusted, and the value of the second objective function is calculated. The inverse covariance matrix corresponding to the minimum value of the second objective function is the estimated value of the inverse covariance matrix for the current round obtained from solving for that sub-region. The second objective function is... i represents the i-th sub-region, Θ i Let ||Θ| represent the inverse covariance matrix of the i-th sub-region; when the ||Θ| of adjacent iterations (k) -Θ (k-1) || F / ||Θ (k) || F <10 -4 The iteration stops when the estimated value of the inverse covariance matrix of the sub-region is obtained, Θ. (k) Let |||| represent the inverse covariance matrix obtained in the k-th iteration. F Denotes the Frobenius norm of a matrix;

[0050] S14. For the inverse covariance matrix Θ, if Θ ij If the value is less than or equal to -η, then nodes ij are considered connected; otherwise, they are considered disconnected. An estimated connectivity matrix A is then formed based on the connectivity assessment results between each node. est ; where Θ ij Let η be the element in the i-th row and j-th column of the inverse covariance matrix Θ, and let η be the threshold and η = |μ|. non-conn |+3σ non-conn μ non-conn σ is the mean of the elements in the inverse covariance matrix Θ. non-conn Let Θ be the standard deviation of the elements in the inverse covariance matrix; compare the estimated connectivity matrix A. est By comparing the estimated connectivity between nodes with the actual SCADA switch states, it is determined whether the estimated connectivity between nodes is consistent with the actual switch states. The accuracy and recall are then calculated. If the accuracy or recall does not meet the requirements, the process returns to step S13 above to re-estimate the inverse covariance matrix until the recall and accuracy meet the requirements, thus obtaining the final inverse covariance matrix.

[0051] S15. Based on the final inverse covariance matrix, the final estimated connectivity matrix is ​​obtained, the connectivity of nodes in the distribution network is obtained, and the distribution network topology model is built by combining the data collected in the offline stage.

[0052] Furthermore, the real-time update module is also used for:

[0053] Real-time acquisition of S21, SCADA switch status, μPMU measurement data and smart meter data;

[0054] S22. Real-time acquired data is timestamped and the time delay is compensated. The μPMU measurement data is input into the trained isolated forest model for detection to obtain anomaly scores. When the anomaly score exceeds the preset score threshold, the voltage phase angle data corresponding to the anomaly score is removed. For SCADA switch states, the difference between the real-time value and the historical mean of each switch state is calculated. If it exceeds the historical mean ±3σ and there is no corresponding switch event, the data point is marked as bad data and repaired by interpolation of data before and after the time. σ is the standard deviation of the switch state data. After the above processing, the standardized dataset D(t) = [Δδ(t), V(t), P(t), Q(t), switch state at time t] is obtained, where Δδ(t) is the phase angle change sequence at time t, P(t) is the active power at time t, Q(t) is the reactive power at time t, and V(t) is the voltage amplitude at time t.

[0055] S23. Construct a state vector x = [δ1, V1, δ2, V2, ..., δ] using a standardized dataset. n V n ] T , where δ i Let V be the voltage phase angle at node i. i Let z be the voltage magnitude at node i, and let z = [z1, z2, ..., z2]. n ] T The expression z = h(x) + e satisfies z = h(x) + e, where z n Let represent the injected power and branch power flow of the nth node, h(x) be a nonlinear measurement function describing the relationship between the measured value and the system state, and e be the measurement error vector.

[0056] The distribution network is divided into several sub-regions, and each edge gateway is responsible for the local state estimation of one sub-region. The local optimization objective function for each sub-region k is: Among them, z k h represents the measurement value of the edge gateway corresponding to sub-region k. k (x) is the nonlinear measurement function of subregion k, w k h is continuously adjusted to assign measurement weights to the edge gateways corresponding to sub-region k. k The matrix parameters in (x) are used to calculate the value of the local optimization objective function. When the objective function is minimized, the update is stopped, and the corresponding optimized h is obtained. k (x), will optimize h kSubstituting the node injected power and branch power flow into the formula z=h(x)+e, we obtain the state vector x, which is also the estimated result of the node's voltage phase angle and voltage amplitude. This step describes the method of completing the voltage amplitude and voltage phase angle in the standardized dataset. In actual operation, some data may be removed due to interference in the μPMU measurement data. Therefore, the removed voltage amplitude and voltage phase angle data are completed through step S23.

[0057] S24. Real-time estimation of the inverse covariance matrix. This real-time estimation process uses either a third objective function, the ADMM algorithm, or a weighted fusion of the results from both methods. The third objective function is: Where, Σ S Θ(t) is the sample covariance matrix of the current time window, Θ(t) is the inverse covariance matrix at time t, and λ is another regularization parameter. During the solution process using the third objective function, the value of the inverse covariance matrix is ​​continuously and randomly adjusted, and the value of the third objective function is calculated. When the value of the third objective function is minimized, the corresponding inverse covariance matrix Θ is the real-time estimate of the inverse covariance matrix Θ obtained by the solution.

[0058] S25. Compare the matrix elements in the real-time estimated value of the inverse covariance matrix obtained at the current time with the threshold η to obtain the real-time estimated connectivity matrix. Based on the connectivity matrix, obtain the connectivity status of the nodes in the distribution network. Combine the estimated results of the voltage phase angle and voltage amplitude of the nodes with the phase angle change sequence and power data to update the distribution network topology.

[0059] Furthermore, the process of solving S24 using the ADMM algorithm is as follows:

[0060] S241. Initialization: Let Θ(0)=Z′(0)=(0)=0, U is the Lagrange multiplier, initialize the penalty parameter ρ=10, Z′(0) means that the auxiliary variable is initialized to a zero matrix in the 0th iteration, and U(0) means that the Lagrange multiplier is 0 in the 0th iteration;

[0061] S242, Θ Update: Fixed Z′ (k) and U (k) Solve

[0062]

[0063] Where k represents the k-th iteration;

[0064] S243, Z′ Update: Fixed Θ (k+1) and U (k) Solve

[0065]

[0066] S244, Lagrange multiplier update: U (k+1) =U (k) +Z'( k+1) -Θ (k+1) ;

[0067] S245. Determine whether the convergence condition is met, and determine ||Θ (k+1) -Z′( k+1) If || < ∈, and this condition is not met, return to step S242 to continue iterating; if it is met, stop iterating, and Θ is then... (k+1) This is the real-time estimate of the inverse covariance matrix, where ∈ is a pre-defined positive number.

[0068] Furthermore, the collaborative positioning module is also used for:

[0069] S31. Based on the real-time estimated connectivity matrix, the connectivity status of the lines in the system is determined. Sensitivity detection is performed on the buses in the connectivity state. The sensitivity detection involves calculating the sensitivity matrix of each bus to the μPMU measurement. δ PMU This indicates the voltage phase angle measured by the μPMU. This represents finding the partial derivative, P. DER This represents the active power of the bus in a connected state; when S sen When the number of elements exceeds a set threshold, the bus area corresponding to that element is designated as a key monitoring area. The key monitoring area is searched using a breadth-first search algorithm to exclude disconnected feeder segments and complete the coarse positioning.

[0070] S32. Solve the power flow equations using the Newton-Raphson method to inversely deduce the node injection power; if node P i (t)>1.2×rated capacity and A est (t) indicates that the node is connected to the distributed power supply feeder, then it is marked as a suspected access point, P i (t) represents the injected power of the i-th node at time t, A est (t) represents the connectivity matrix at time t, and the distributed power supply access feeder refers to the feeder segment determined by S31;

[0071] S33. For nodes marked as suspected access points, extract their voltage phase angle sequence, convert it into an input format suitable for the SeriesTransformer model, input it into the pre-trained SeriesTransformer model, input the feature vector extracted by the pre-trained SeriesTransformer model into the classifier, obtain the prediction result through the classifier, and output whether the node is a distributed power access unit. The input format suitable for the SeriesTransformer model is a three-dimensional tensor, which consists of the number of samples in a single input, the voltage phase angle sequence, and the dimension of the input features.

[0072] Furthermore, the processing procedure of the SeriesTransformer model is as follows:

[0073] The input data is processed by the inner layer, and the result is used to calculate attention. The result of the attention calculation and the inner layer processing is added together and then fed into the normalization layer. The output of the normalization layer is fed into the feedforward layer. The output of the feedforward layer and the output of the normalization layer are added together, normalized again, and then linearly projected to output the final result.

[0074] The advantages of this invention are:

[0075] (1) This invention uses historical data to obtain the initial values ​​of the connectivity matrix and the inverse covariance matrix. It uses the initial values ​​of the connectivity matrix and the inverse covariance matrix combined with real-time collected data to perform real-time estimation of the inverse covariance matrix, completes the topology update, performs coarse positioning based on the updated topology, marks the coarsely positioned areas as suspected access points, extracts the voltage phase angle sequence of the nodes marked as suspected access points, inputs the feature vector extracted by the pre-trained SeriesTransformer model into the classifier to obtain whether the node is a distributed power access unit, thereby completing the positioning of a large number of distributed power access units in the distribution network.

[0076] (2) This invention employs edge-layer data acquisition and local state estimation technology. By deploying edge devices at the power grid site, sensor data is aggregated in real time, and data cleaning, time synchronization, and distributed weighted least squares state estimation are performed to achieve rapid calculation of node voltage phase angle and amplitude. This technology relies on the localized processing advantages of edge computing to significantly reduce the communication latency of data uploading to the cloud, shortening the state estimation time from seconds in traditional centralized solutions to milliseconds, providing high-frequency, low-latency basic data support for real-time positioning.

[0077] (3) The dynamic topology update strategy of this invention combines multi-source data acquisition and topology change modeling to achieve high-frequency dynamic updates of the network topology, preventing positioning errors caused by topology changes. By estimating the inverse covariance matrix and using the weighted least squares method, the spatiotemporal correlation between electrical connections between distribution network nodes and measurement data is captured in real time, enabling the positioning accuracy of distributed power supply access units to reach the bus level, significantly improving the adaptability and reliability of the positioning method, and ensuring the long-term accuracy of remote positioning of distributed power supply access units.

[0078] (4) The present invention uses the feature mining capability of the SeriesTransformer model, which can fully capture the complex correlations and features in the network topology and electrical quantity data, improve the accuracy of the positioning results, maintain the efficiency of the positioning process, and ensure the real-time and accuracy of the remote positioning of the distributed power access unit.

[0079] (5) The edge-cloud collaborative computing architecture of the present invention deploys real-time tasks such as state estimation and timestamp calibration on the edge gateway, and combines the multivariate time series feature analysis capability of the cloud SeriesTransformer model to achieve hierarchical processing of "coarse-grained region screening and fine-grained precise positioning", which can handle high-concurrency positioning in large-scale scenarios.

[0080] (6) This invention precisely locates distributed power source access units, transforming passive, decentralized power sources into active, controllable grid assets. This enables refined monitoring and management of the grid status, ensuring the safe and stable operation of the system under a high proportion of renewable energy access. Based on this, by optimizing aggregation strategies and improving scheduling efficiency, efficient operating models such as virtual power plants can be constructed, enhancing resource utilization and economic benefits. Simultaneously, in the event of a grid fault, quickly identifying the source of the anomaly and employing isolation strategies can effectively shorten power outage recovery time and improve overall power supply reliability. Attached Figure Description

[0081] Figure 1 This is a schematic diagram of the architecture of a remote positioning method for a distributed power access unit disclosed in an embodiment of the present invention;

[0082] Figure 2 This is a flowchart of a remote positioning method for a distributed power access unit disclosed in an embodiment of the present invention;

[0083] Figure 3 This is a SeriesTransformer model structure diagram of a remote positioning method for a distributed power access unit disclosed in an embodiment of the present invention;

[0084] Figure 4This is a schematic diagram of the input data encoding format of the SeriesTransformer model for a remote positioning method for a distributed power access unit disclosed in an embodiment of the present invention. Detailed Implementation

[0085] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0086] Example 1

[0087] Embodiment 1 of this invention provides a remote positioning method for distributed power supply access units. This method integrates dynamic topology tracking, hierarchical distributed algorithms, and interdisciplinary technologies. It achieves real-time data acquisition and local state estimation through an edge layer, continuously tracks changes in electrical connections between network nodes using a dynamic inverse covariance matrix, and designs a step-by-step positioning strategy. In the coarse positioning stage, it narrows down the candidate region, and in the fine positioning stage, it combines electrical features to achieve accurate identification, effectively balancing computational efficiency and positioning accuracy. Compared to traditional solutions, its advantages lie in relying on edge computing for rapid localized data processing, improving the response speed to topology changes through dynamic matrix modeling, and enhancing noise resistance through multi-source data fusion, effectively addressing dynamic changes in the distribution network topology. This provides a practical technical path for the refined monitoring and management of smart distribution networks. Figure 1 The diagram illustrates the architecture of the proposed method. The perception layer consists of a μPMU and a smart meter, with a built-in IEEE 1588PTP clock synchronization module. The edge layer is deployed on the industrial gateway of the distribution transformer, handling real-time data processing and lightweight computation, reducing cloud load and improving response speed. The edge layer's data preprocessing steps include time synchronization, anomaly detection, and state estimation, as well as localized computational tasks such as preliminary processing of the inverse covariance matrix, integrating a dynamic topology model and an ADMM solver. The cloud layer handles global data processing, model training, and large-scale computational tasks, such as running the SeriesTransformer model and implementing a step-by-step positioning strategy, and needs to update the inverse covariance model and SeriesTransformer parameters based on historical data. The edge layer outputs preprocessed data to the cloud, and the cloud layer outputs dynamic model parameters and positioning commands to the edge layer. Figure 2 As shown, the method proposed in this invention specifically includes the following steps:

[0088] S1. Initialization Configuration: During the offline phase, historical SCADA switch status and μPMU measurement data are used to optimize sparsity with L1 regularization, train a dynamic topology model, and determine the initial values ​​of the estimated connectivity matrix and inverse covariance matrix for each bus. The specific process is as follows:

[0089] S11. Multi-source data acquisition and annotation

[0090] Basic data is collected, specifically including SCADA switch status, μPMU measurement data, smart meter data, and line parameters. SCADA switch status includes historical status data (closed / open) of all circuit breakers (CBs) and disconnectors in the distribution network, with a time resolution ≤1 minute, covering at least 3 months of normal operation scenarios (including planned maintenance, fault isolation, and other typical topology change events). μPMU measurement data includes bus voltage phase angle and amplitude. Smart meters collect active / reactive power injection and load curves at the nodes to construct a Gaussian distribution model of power fluctuations. Line parameters include resistance, reactance, ground capacitance, transformer turns ratio, and inverter parameters at the DER connection point.

[0091] For each switch operation event (such as CB closing / opening), a timestamp and the set of affected lines are labeled to form a "topology change-measurement response" sample pair, labeled in the format {t k ,ΔA(t k ),{Δδ ij (t k )}}, where ΔA(t) k ) represents the change in the connectivity matrix, Δδ ij (t k ) represents the phase angle difference change of node ij.

[0092] S12. Construct the initial connectivity matrix

[0093] The distribution network is modeled as a graph structure G(v,e), where node v represents a bus and edge e represents a line switch. First, a static matrix needs to be constructed based on the switch states, while also considering the changes in line parameters caused by DER integration. Based on the historical stable operating states of SCADA, i.e., periods without switch operation, an initial connectivity matrix is ​​constructed. Where A0(i,j) = 1 indicates that node i is connected to node j, and 0 indicates otherwise.

[0094] Then, the covariance matrix of the voltage phase angle change is calculated. For the phase angle change sequence Δδ(t) during historical stable operating periods (no topological changes, load fluctuations ≤10%), the sample covariance matrix is ​​calculated. Where T is the total number of samples.

[0095] S13, Model Training

[0096] An optimization problem using L1 regularization is used to estimate the inverse covariance matrix Θ, with the objective function being:

[0097]

[0098] Where Tr() represents the sum of the elements on the main diagonal of the matrix, ln det() represents the natural logarithm of the matrix determinant, ln det(Θ) ensures the positive definiteness of the matrix, and β||Θ-Θ prev ||1 is the regularization term, Θ prev Let Θ be the inverse covariance at the previous time step (initially set to a 0 matrix), and β be the regularization parameter. The inverse covariance matrix Θ has the same form as the sample covariance matrix and is initially a 0 matrix. During the solution process, the value of the inverse covariance matrix Θ is continuously and randomly adjusted, and the value of the objective function is calculated. The inverse covariance matrix Θ corresponding to the minimum objective function value is the estimated value of the solved inverse covariance matrix Θ.

[0099] This step uses the L1 norm to promote sparsity, which aligns with the assumption that topological changes typically involve a small number of switches.

[0100] The above-described solution process for the inverse covariance matrix Θ is a relatively simplified method. In practical applications, the distributed ADMM algorithm can also be used to solve the problem, decomposing the large-scale network into sub-regions (such as partitioning by feeder), and solving the local optimization problem in each sub-region.

[0101]

[0102] Where i represents the i-th sub-region, Θ i Let ||Θ| represent the inverse covariance matrix of the i-th sub-region. The principle for solving the inverse covariance matrix of a sub-region is similar to that for the entire inverse covariance matrix Θ: values ​​are randomly assigned to the inverse covariance matrix, and then the objective function is calculated. The inverse covariance matrix corresponding to the minimum objective function value is the estimated value of the solved inverse covariance matrix Θ. For each sub-region's solved inverse covariance matrix, these matrices are concatenated according to the region to form the inverse covariance matrix Θ of the entire network. Each iteration yields a global inverse covariance matrix Θ, representing the entire network. However, this Θ is not necessarily optimal. Therefore, the global Θ is iteratively updated until the ||Θ| of the adjacent iteration is obtained. (k) -Θ (k-1) || F / ||Θ (k) || F <10 -4 The iteration stops when the optimal inverse covariance matrix Θ is obtained. (k) Let represent the inverse covariance matrix obtained in the k-th iteration. F Let Frobenius norm denote the matrix.

[0103] It should be noted that either of the two methods for solving the inverse covariance matrix Θ can be used in practical applications, or the results of the two methods can be combined. For example, the results of the first and second methods for solving the inverse covariance matrix Θ can be weighted and combined to obtain a new inverse covariance matrix.

[0104] S14, Model Validation

[0105] Generate the corresponding inverse covariance matrix for multiple (1000+) topological scenarios. Calculate the Θ values ​​for pairs of disconnected nodes (actually unconnected nodes ij). ij Distribution, calculate its mean μ non-conn and standard deviation σ non-conn Set the threshold η = |μ non-conn |+3σ non-conn Θ ij This represents the value of the element in the i-th row and j-th column of the inverse covariance matrix.

[0106] If Θ ij If the value is less than or equal to -η, then node ij is considered connected (switch closed); otherwise, it is considered disconnected (switch open). An estimated connectivity matrix A is then formed based on the connectivity determination results between each node. est .

[0107] Comparative estimation of connectivity matrix A est Compared with the actual SCADA status, determine whether the estimated connectivity between nodes is consistent with the actual switching status, and then calculate the precision and recall. The precision is required to be ≥95% and the recall is required to be ≥90%. If the precision or recall does not meet the requirements, return to step S13 above to re-estimate the inverse covariance matrix until the recall and precision meet the requirements, and obtain the final inverse covariance matrix.

[0108] S15, Model Deployment

[0109] Save the estimated connectivity matrix, inverse covariance matrix Θ, regularization parameter β, threshold η, and line parameter table. Distribute the trained model parameters (A0, Θ0, η) to the edge gateway and preload them into local storage. Since this step involves calculations based on historical data, the calculation results are uniformly used as initial values. The data calculated in step S2 below for real-time topology updates depends on these initial values. Therefore, A0 represents the initial value of the estimated connectivity matrix, and Θ0 represents the initial value of the inverse covariance matrix Θ.

[0110] The edge layer initialization loads offline training data A0, Θ0, β, and η, providing the edge gateway with the initial topology and historical statistical characteristics of the distribution network, serving as benchmark data for real-time tracking. A0 clarifies the initial connection state between nodes, serving as the basis for subsequent real-time topology change detection; Θ0 reflects the initial correlation between node voltage phase angles, helping to capture correlation changes caused by topology shifts during real-time tracking. Through these steps, the dynamic topology model constructed in the offline phase accurately maps topology change patterns in historical operating scenarios, providing reliable prior knowledge and initial states for the real-time tracking phase, ensuring positioning accuracy and robustness in dynamic network environments.

[0111] S2. Real-time data acquisition via edge computing, combined with the Dynamic Inverse Covariance Matrix (ADMM) algorithm to track topology changes, and threshold values ​​to determine node connectivity. When a switch operation or device malfunction is detected, topology updates are achieved within seconds. The specific process is as follows:

[0112] S21, Multi-source data acquisition

[0113] Distributed power sources (such as photovoltaic inverters and energy storage systems) incorporate smart meters and μPMUs (micro phasor measurement units) to collect data at a fixed frequency. Specifically, the μPMU collects the voltage phase angle δ in real time at a sampling rate of 25Hz. i (t) and amplitude V i (t), SCADA obtains the switch status every minute, and the smart meter uploads power injection data at 15-minute intervals.

[0114] S22, Edge Preprocessing

[0115] Various sensors have built-in local clocks, which are initialized upon device startup. When a sensor acquires each data point, it immediately appends the current local clock timestamp to that data point. For example, when a μPMU acquires voltage phase angle data, it uses the millisecond-accurate time displayed on its local clock as the timestamp for that data. Smart meters perform the same operation when acquiring power data. The timestamp format should adopt a unified standard, such as ISO 8601 format (YYYY-MM-DDTHH:MM:SS.sssZ), to facilitate subsequent data processing and analysis.

[0116] There is a certain communication delay during data transmission from the sensor to the edge gateway. The edge gateway records the timestamp of each data point being sent from the sensor and the timestamp of its arrival at the edge gateway, calculating the communication delay by the difference between the two. For example, if the smart meter sends data at time t1 and the edge gateway receives the data at time t2, then the communication delay Δt = t2 - t1. To compensate for this delay, the edge gateway adjusts the data's timestamp to t1 + Δt. Because there may be a discrepancy between the sensor's local clock and the edge gateway's clock, the edge gateway periodically (e.g., weekly) compares and calibrates the two.

[0117] Next, anomaly detection is performed. Isolation forest detection is used for μPMU data. During a period of normal system operation (e.g., one week), voltage phase angle data collected by the μPMU is continuously collected. This data is stored chronologically to form a historical dataset, which is then used to train the Isolation Forest algorithm. The core idea of ​​the Isolation Forest algorithm is to randomly partition the data space, construct multiple decision trees, and calculate the average path length of each data point in these decision trees. The shorter the path length, the more isolated the data point, and the more likely it is to be an anomaly. During system operation, the voltage phase angle data of the μPMU is collected in real time and input into the trained Isolation Forest model for detection. The model calculates an anomaly score for each data point; the higher the score, the more likely the data point is to be an anomaly. When the anomaly score exceeds a preset score threshold, the voltage phase angle data corresponding to that anomaly score is removed.

[0118] The residual test method is used for SCADA status. Status data of the SCADA system over a period of time is collected, including parameters such as switch status and voltage. Statistical analysis is performed on the preprocessed data to calculate the historical mean and standard deviation of each parameter, which will serve as the benchmark for subsequent anomaly detection. During system operation, SCADA status data is collected in real time, and the difference between the real-time value and the historical mean of each parameter is calculated, i.e., the residual. If the residual exceeds the historical mean ± 3σ and there is no corresponding switch event, the data point is marked as bad data. Data interpolation between consecutive time points is used for data repair. For example, for voltage data, if the data at time t is marked as bad data, while the voltage value at time t-1 is V1 and the voltage value at time t+1 is V2, the linear interpolation formula Vt = (V1 + V2) / 2 can be used to repair the voltage data at time t.

[0119] The standardized dataset D(t) = [Δδ(t), V(t), P(t), Q(t), switch state at time t] is input for the next stage. Here, Δδ(t) is the phase angle change sequence at time t, P(t) is the active power at time t, and Q(t) is the reactive power at time t.

[0120] S23, State Estimation

[0121] Input measurement data, including power measurements at each node (active power P). i and reactive power Q i The voltage amplitude measurements of some nodes, etc., are obtained from various measuring devices installed in the power system. The goal of state estimation is to combine these measurement data with the system model to estimate the true voltage phase angle and amplitude of each node in the system.

[0122] Suppose the system has n nodes and the state vector is x = [δ1, V1, δ2, V2, ..., δ n V n ] T , where δ i Let V be the voltage phase angle (rad) at node i. i The voltage amplitude (pu). Measurement vector z = [z1, z2, ..., z...]. n ] T Includes node injected power and branch power flow, satisfying z = h(x) + e, where z n Let h(x) represent the injected power and branch power flow of the nth node, and h(x) be a nonlinear measurement function (such as the power flow equation) that describes the relationship between the measured value and the system state. The measurement error vector e ~ N(0,R) is zero-mean Gaussian noise.

[0123] Distributed weighted least squares (WLS) is used to estimate the system state, obtaining the voltage phase angle and magnitude of the nodes. Specifically, the state estimation process is constrained by the regularized inverse covariance matrix in the model loaded in step S1, and the topological difference β||Θ(t)-Θ(t-1)||1 between adjacent time steps is penalized by the l1 norm, introducing model priors into the weighted least squares (WLS) method. To address the computational bottleneck of large-scale distribution networks, the distribution network is divided into M sub-regions, with each edge gateway responsible for the local state estimation of one sub-region. Global convergence is achieved through information exchange between boundary buses (such as the common node phase angles of adjacent sub-regions).

[0124] The local optimization objective function for each sub-region k is:

[0125]

[0126] Among them, z k h represents the measurement value of the edge gateway corresponding to sub-region k. k (x) is the nonlinear measurement function of subregion k, w k Let h be the measurement weights for the edge gateway corresponding to sub-region k (μPMU weight = 100, smart meter weight = 10, SCADA switch status weight = 50). During the above solution process, h is continuously adjusted. kThe matrix parameters in (x) are used to calculate the value of the local optimization objective function. When the objective function is minimized, the update is stopped, and the corresponding optimized h is obtained. k (x), will optimize h k Substituting the node injected power and branch power flow into the formula z=h(x)+e, we obtain the state vector x, which is also the estimated result of the node's voltage phase angle and voltage amplitude. Based on this, we can complete the voltage phase angle and voltage amplitude data that were removed from the standardized dataset.

[0127] S24. Real-time estimation of the inverse covariance matrix

[0128] Next, the inverse covariance matrix is ​​estimated in real time. This step is to capture the correlation of voltage phase angle changes in real time and capture topological dynamics. The voltage phase angle and magnitude data of each node at different times t obtained from the previous state estimation are denoted as x(t). The goal of inverse covariance matrix estimation is to estimate a symmetric positive definite matrix Θ(t) that reflects the correlation between nodes. Off-diagonal elements Θ ij The physical meaning of corresponds to the mutual admittance between nodes. The larger the negative value, the higher the probability of connectivity. Its sparse structure corresponds to the network topology.

[0129] Specifically, it calculates the sample covariance matrix within the current window. Optimization issues:

[0130]

[0131] The first term is the sample covariance matrix, calculated from the state estimation result x. Its calculation principle is the same as that of the sample covariance matrix S0 within the current window, calculated using a similar formula to S0 based on the time window corresponding to the current time t, only the time window differs. λ is a regularization parameter used to control the sparsity of matrix Θ, determined through cross-validation (here λ = 0.5). The third term is the dynamic constraint: minimizing the topological difference between adjacent time steps. During the solution process, the value of the inverse covariance matrix is ​​continuously and randomly adjusted, and the objective function is calculated. The inverse covariance matrix Θ corresponding to the minimum objective function value is the real-time estimate of the solved inverse covariance matrix Θ.

[0132] The real-time estimation of the inverse covariance matrix can also be achieved using the ADMM algorithm, a highly efficient distributed optimization algorithm capable of solving large network optimization problems. The local covariance matrix is ​​calculated at the edge layer, and the solution is obtained iteratively in the cloud using ADMM. The problem is decomposed into two sub-problems: covariance fitting and sparse constraints, introducing an auxiliary variable Z′=Θ(t).

[0133] ADMM iteration steps:

[0134] (a) Initialization: Let Θ(0)=Z′(0)=(0)=0, U is the Lagrange multiplier, initialize the penalty parameter ρ=10, Z′(0) means that the auxiliary variable is initialized to a zero matrix in the 0th iteration, and U(0) means that the Lagrange multiplier is 0 in the 0th iteration.

[0135] (b) Θ Update: Fixed Z′ (k) and U (k) Solve

[0136]

[0137] Where k represents the k-th iteration. The solution is obtained iteratively using the matrix inversion gradient descent method, ensuring that Θ... (k+1) Approximating the inverse covariance of the sample.

[0138] (c)Z Update: Fixed Θ (k+1) and U (k) Solve

[0139]

[0140] For the L1 norm minimization problem, the soft-thresholding operator is used to minimize Z′. (k+1) Sparsify the off-diagonal elements, force tiny values ​​to zero, and preserve significant connections.

[0141] (d) Lagrange multiplier update: U (k+1) =U (k) +Z' (k+1) -Θ (k+1) .

[0142] (e) Determine if the convergence condition is met, for example, ||Θ (k+1) -Z′ (k+1) ||<∈, (∈ is a pre-defined small positive number, here it is 10) -4 If the condition is not met, return to step (b) and continue iterating; if the condition is met, stop iterating, and Θ is then... (k+1) This is the estimated inverse covariance matrix.

[0143] S25, Dynamic Update of Connectivity Matrix

[0144] Based on the current inverse covariance matrix Θ(t), the node connectivity is determined element by element using a threshold η to generate an estimated connectivity matrix A. est (t), if Θ ij (t)≤-η, then A est (t) ij =1, otherwise 0. Detection matrix A est (t) and the previous time A est(t-1), the difference, the set of switches for identifying changes ΔA={(i,j)|A est (t) ij ≠A est (t-1) ij Extract the switch set {(i,j)} corresponding to the non-zero elements. For detected changes, perform power flow residual verification to calculate whether it conforms to the following formula; if it does not conform, mark it as a suspected false positive.

[0145] |P i 计算 -P i 量测 |<5%×rated power.

[0146] S3. The edge layer collects data from the μPMU and smart meters every second, performs anomaly detection, timestamp calibration, and distributed state estimation, and generates a real-time state vector x(t) and a local connectivity matrix A. est (t), reducing cloud computing load. The cloud aggregates edge layer data every minute, uses breadth-first search to narrow down the location range, detects topology changes using the SeriesTransformer model, and updates the global connectivity matrix A(t). For suspected areas, it combines power flow inversion and feature matching to output the precise location of the DER access unit. The specific process is as follows:

[0147] S31, Coarse Positioning

[0148] In power systems, sensitivity represents the degree to which a change in one variable affects another. This is used to quantify the active power P of a bus in a connected state (i.e., potentially connected to distributed generation units). DER The change affects the voltage phase angle δ measured by the μPMU (micro-synchronous phasor measurement unit). PMU To assess the impact of the μPMU measurement, the sensitivity matrix of each busbar to the μPMU measurement is calculated.

[0149]

[0150] In simple terms, it reflects how much the voltage phase angle measured by the μPMU changes when the active power of a suspected access unit changes slightly. Based on this model, the P... DER Make a small perturbation and then observe δ. PMU The changes in power are observed. Through repeated experiments and data recording, the sensitivity values ​​at different busbars can be calculated. After calculating the sensitivity values ​​for each busbar, a threshold needs to be set, here it is 0.5 / kW. If the sensitivity value of a certain busbar is greater than this threshold, it indicates that the busbar is more sensitive to changes in the active power of the suspected connected unit, and it is classified as a key monitoring area.

[0151] Using the dynamic connectivity matrix A est The traceable area for μPMU measurements is determined through breadth-first search (BFS), excluding disconnected feeder sections, thus narrowing down the location range. This process relies on the offline model's learning and representation of the distribution network topology, enabling rapid and effective filtering of potential areas during real-time location. The specific steps are as follows:

[0152] (a) First, starting from the node where μPMU is located, mark it as visited and add it to a queue.

[0153] (b) Then, remove a node from the queue and check all unvisited nodes connected to that node. Mark these unvisited nodes as visited and add them to the queue.

[0154] (c) Repeat step (b) until the queue is empty. This results in a set of nodes connected to the μPMU node, which is the traceable area of ​​μPMU measurements.

[0155] Based on the connectivity information obtained from the BFS algorithm, it can be determined which feeder sections are not connected to the μPMU node. Since the power generated by the distributed generation source propagates through the connected electrical network after its connection, disconnected feeder sections cannot be areas where the distributed generation source is connected. Therefore, these disconnected feeder sections can be excluded, thus narrowing down the location range.

[0156] S32, Precise Positioning

[0157] (1) Solving power flow equations based on the Newton-Raphson method to back-calculate node injection power

[0158] In power flow calculations, a set of nonlinear equations describing the power balance at nodes needs to be solved. For a node i, its active power P i and reactive power Q i It can be represented as:

[0159]

[0160] Among them, V i and V j These are the voltage magnitudes at nodes i and j, respectively, G ij and G ij These are the conductance and susceptance between node i and node j, respectively, δ ij =δ i -δ j It is the voltage phase angle difference between node i and node j.

[0161] The basic idea of ​​the Newton-Raphson method is to iteratively approximate the solution to a system of equations. First, an initial guess of the voltage magnitude and phase angle is given. Then, the error in the system of equations is calculated based on this guess. Next, a linear system of equations is solved to update the voltage magnitude and phase angle, gradually reducing the error. This process is repeated until the error meets a certain accuracy requirement. Taking active power as an example, the power flow equations at the current iteration point (V...) are... (k) ,δ (k) Taylor expansion at () and ignoring higher-order terms yields the linear correction equation:

[0162]

[0163] The elements of the Jacobian matrix J, the partial derivatives of active power with respect to phase angle. This reflects the impact of the phase angle change at node j on the active power at node i. The partial derivative of active power with respect to voltage amplitude. The active power at node i is reflected by the effect of changes in the voltage amplitude at node j on line parameters. The partial derivative of reactive power with respect to phase angle is also considered. The partial derivative of reactive power with respect to voltage amplitude has the opposite sign to that of active power, reflecting the coupling characteristics between reactive power and phase angle.

[0164] Calculate power error:

[0165]

[0166] The Jacobian matrix of a power distribution network is sparse (each node connects to an average of 3-5 lines), and the proportion of non-zero elements is usually <5%. Sparse LU decomposition can be used to accelerate the solution:

[0167] Decompose J = L·U, where L is a lower triangular matrix and U is an upper triangular matrix. The decomposition process only processes non-zero elements, and the time complexity is O(N). 3 The value decreases to O(M·N) (where M is the number of non-zero elements).

[0168] Forward substitution: Solve L·y=ΔP / ΔQ to obtain the intermediate variable y;

[0169] Backward substitution: Solve for U·Δx=y to obtain the voltage correction Δx=[Δδ,ΔV / V] T .

[0170] Extract ΔV from the correction vector Δx i / V i Calculate the actual correction amount ΔV i =V i ·(ΔV i / V i ).

[0171] The voltage amplitude correction V is obtained through sparse LU decomposition. i (k) and phase angle correction Update state variables:

[0172]

[0173] V i (k+1) =V i (k) +ΔV i (k)

[0174] When the power error of all nodes is less than the threshold (e.g., 10) -6 (pu), or voltage / phase angle correction less than 10 -4 pu / , stop iteration, and obtain the final P i and Q i .

[0175] This step is based on the real-time state estimation results x(t) and A. est (t) The power flow equations are solved using the Newton-Raphson method, thereby deriving the injected power at each node. The Newton-Raphson method strictly follows Kirchhoff's laws and Ohm's law, ensuring that the derived power conforms to the physical laws of the power grid and avoiding the "uninformed" errors of purely data-driven methods.

[0176] It should be noted that the fine-localization process here involves the inverse calculation of injected power. The injected power mentioned in the measurements above is a known quantity, primarily because the measurements serve as one of many input pieces of evidence, their main purpose being to help estimate a more reliable global voltage state. In the fine-localization stage, the measured injected power is not directly used for the final judgment because the original data has limitations in coverage, accuracy, and timeliness. The fine-localization method of this invention first uses all data, including high-precision μPMU data, to estimate the most reliable global voltage state. Then, based on this, a power flow inversion method that strictly follows physical laws is used to calculate the theoretically consistent injected power covering all nodes, thereby enabling discovery and localization.

[0177] (2) Marking of suspected access points

[0178] If a node P i (t)>1.2×rated capacity and A est (t) indicates that the node is connected to the DER access feeder, and is therefore marked as a potential access point. This is because the integration of distributed generation sources injects additional power into the grid, increasing the active power of the node. Furthermore, only nodes connected to the DER access feeder are likely to be access points.

[0179] (3) SeriesTransformer model

[0180] For nodes marked as potential access points, their voltage phase angle sequence (length = 100 cycles, i.e., 4 seconds) is extracted and input into a pre-trained SeriesTransformer model. The architecture of the SeriesTransformer model is as follows: Figure 3 As shown, the model has been trained offline using a large amount of historical data, learning the temporal correlation and spatial distribution characteristics of multivariate variables (voltage, current, power, etc.). The preprocessed phase angle sequence is converted into a format suitable for the SeriesTransformer model input, including a three-dimensional tensor, the number of samples in a single input, the phase angle sequence, and the dimension of the input features. The self-attention mechanism in the SeriesTransformer model can capture the temporal correlation and spatial distribution characteristics in the phase angle sequence, converting the input phase angle sequence into a representative feature vector. Then, the extracted feature vector is input into the classifier, and the prediction result is obtained through the forward propagation process of the classifier, outputting whether the node is a DER access unit.

[0181] Figure 3 This invention demonstrates the SeriesTransformer model architecture proposed in this invention, with the specific encoding format provided by [the relevant source]. Figure 4 This model, as a novel pure encoder structure, is specifically designed for representation learning and adaptive association of multivariate sequences. Specifically, in the SeriesTransformer model, each time series driven by a complex underlying process is labeled to describe the attributes of the variables. These labeled time series interact with each other through a self-attention mechanism and are processed separately using a feedforward network to complete the learning of sequence representations. The task of generating predicted sequences is primarily achieved by linear layers.

[0182] The principle of SeriesTransformer is as follows:

[0183] Based on the input sequence X :,n Predict each variable The process can be simply described as follows:

[0184]

[0185] M l+1 =TrmBlock(M l ), l=0,…,L-1

[0186]

[0187] Where M = {m1,…,m} N}∈R N×D It contains N embedded tokens of dimension D. Both the embedding layer and the linear projection layer are implemented using a multilayer perceptron. In this process, the embedding layer transforms the raw time-series data into a feature representation suitable for model processing; the linear projection layer predicts and projects future time series based on the features learned by the model. After processing by the embedding layer, the resulting variable tokens interact with each other through a self-attention mechanism. This interaction effectively captures the correlation between different variable tokens, thereby uncovering the potential relationships between variables in multivariate time series. Simultaneously, the shared feedforward network in each TrmBlock processes the variable tokens independently. The feedforward network further extracts complex features from the variable tokens through a series of linear and nonlinear transformations, providing a more representative sequence representation for subsequent prediction tasks. Furthermore, traditional Transformer models require dedicated positional embeddings to encode sequence order information to help the model identify element order. However, in this model, since sequence order information is implicitly stored in the neuron arrangement of the feedforward network, positional embedding is unnecessary as in traditional Transformers. This design not only simplifies the model structure but also avoids the information loss problem that may be caused by position embedding, thereby improving model performance and efficiency.

[0188] Layer normalization, to ensure variable independence and avoid the influence of sampling misalignment or time delays on the cross-correlation of modeling variables, standardizes each variable individually. In typical Transformer-based prediction models, to improve the convergence speed and training stability of deep networks, this module normalizes the multivariate representations of the same timestamp, which gradually merges the variables. However, if the collected time points do not represent the same event, this operation can introduce interactive noise between non-causal or delayed processes. Furthermore, since all sequences of tokens are normalized to a Gaussian distribution, discrepancies caused by inconsistent measurements are reduced. In contrast, in previous architectures, tokens at different time steps are normalized separately, which often leads to over-smoothing of the time series and loss of some key information. In this SeriesTransformer model, layer normalization is applied to the sequence representation of a single variable according to the following formula, which is effective in handling non-stationary problems. Moreover, since all sequences of tokens are normalized to a Gaussian distribution, discrepancies caused by measurement inconsistencies are reduced. In contrast, in previous architectures, tokens at different time steps were normalized separately, which often led to over-smoothing of the time series and loss of some key information.

[0189]

[0190] In the SeriesTransformer architecture, a feedforward network (FFN) is used to process the sequence representation of each variable token. Based on the general approximation theorem, FFNs are able to extract complex features to describe time series. By stacking backward modules, FFNs encode the observed time series and decode the representation of future sequences using dense nonlinear connections.

[0191] Regarding the self-attention mechanism, seriesTransformer treats the entire time series of a single variable as an independent process. For the comprehensive extracted representation of each time series, the self-attention module generates queries, keys, and values ​​through linear projection. d k This is the projection dimension. Self-attention is obtained using the following formula, where each score element can reveal variable-level correlations to some extent:

[0192]

[0193] Based on the above technical solutions, this invention selects a distribution network in a new energy demonstration park as the implementation scenario. This area has a complex distribution network environment with typical multi-DER access. First, data from the park over the past year is collected, and topology change events (including planned maintenance, photovoltaic grid connection switching, etc.) are labeled to train a dynamic topology model. During system operation, statistical analysis of real-time measurement data and dynamic matrix tracking are used to achieve automated detection and preliminary location of abnormal events. When the edge layer detects that the voltage phase angle fluctuation of a node exceeds twice the standard deviation of historical voltage phase angle fluctuations, and simultaneously, the elements of the inverse covariance matrix representing the electrical connection strength between nodes show significant changes with absolute values ​​exceeding the threshold η, it is marked as a suspected topology change or new DER access event. This mechanism, combining statistical regularity and electrical correlation analysis, reduces the false alarm rate compared to a single threshold judgment, avoiding false alarms caused by load fluctuations. The edge layer uses breadth-first search to traverse the dynamic connectivity matrix to determine the feeder to which the abnormal node belongs. Combining this with historical SCADA switch operation records, no recent switch operations on this feeder are found, ruling out topology changes caused by switch status changes, and the event is preliminarily determined to be a new DER access event. This step leverages the radial structure of the distribution network to rapidly converge the positioning range from the entire network level to the feeder level, avoiding redundant calculations for all network nodes. The cloud receives abnormal node data uploaded from the edge layer, inverts the node's active power using the Newton-Raphson method, and calculates the real-time injected power of the node. If the injected power exceeds the rated load, it indicates additional power injection. The voltage phase angle sequence of the node is further extracted and input into a pre-trained SeriesTransformer model. The model captures the periodic characteristics of phase angle fluctuations through a self-attention mechanism and outputs classification results, such as correlation with light intensity, identifying it as a photovoltaic access feature. This application enables real-time and accurate positioning of distributed energy units, optimizes aggregation strategies, and improves scheduling efficiency. In fault scenarios, it quickly locates the DER access point corresponding to abnormal voltage / power fluctuations, assisting in isolation strategies and shortening power outage recovery time. The entire process employs edge layer data acquisition and local state estimation technology. Edge devices deployed on the power grid site aggregate sensor data in real time, performing data cleaning, time synchronization, and distributed weighted least squares state estimation to achieve rapid calculation of node voltage phase angles and amplitudes. This technology leverages the localized processing advantages of edge computing to significantly reduce the communication latency of data uploading to the cloud, shortening the state estimation time from seconds in traditional centralized solutions to milliseconds, and providing high-frequency, low-latency basic data support for real-time positioning.

[0194] Example 2

[0195] Based on Embodiment 1, Embodiment 2 of the present invention also provides a remote positioning system for a distributed power supply access unit, comprising:

[0196] The initialization configuration module is used to acquire historical data of SCADA switch status, μPMU measurement data and smart meter data during the offline phase, solve for the initial values ​​of the connectivity matrix and inverse covariance matrix of each bus, and build the distribution network topology model.

[0197] The real-time update module is used to estimate the inverse covariance matrix in real time using real-time collected data, determine the connectivity of nodes, construct a real-time connectivity matrix, track changes in the distribution network topology, and realize topology updates.

[0198] The collaborative localization module is used to determine the connectivity status of lines in the system based on the real-time estimated connectivity matrix, perform sensitivity detection on buses in the connected state to identify key monitoring areas, and search the key monitoring areas using a breadth-first search algorithm to exclude disconnected feeder sections, thus completing coarse localization. Suspected access points are marked in the coarsely located areas. For nodes marked as suspected access points, their voltage phase angle sequences are extracted and input into the feature vectors extracted by the pre-trained SeriesTransformer model, which are then fed into a classifier. The classifier obtains the prediction result and outputs whether the node is a distributed power supply access unit.

[0199] Specifically, the initialization configuration module is also used for:

[0200] Historical data acquisition of S11, SCADA switch status, μPMU measurement data and smart meter data. μPMU measurement data is the phase angle and amplitude of bus voltage, and smart meter data is the power injection data.

[0201] S12. Model the distribution network as a graph structure G(v,e), where node v represents a bus and edge e represents a line switch. The sample covariance matrix S0 is obtained by summing and averaging the bus voltage phase angle change sequence and the transpose of the sequence within a time window.

[0202] S13. Assume the inverse covariance matrix Θ has the same form as the sample covariance matrix. Initially, the inverse covariance matrix Θ is a zero matrix. The inverse covariance matrix Θ is solved using either the first solution method, the second solution method, or a weighted fusion of both methods. The first solution method involves continuously and randomly adjusting the value of the inverse covariance matrix Θ during the solution process and calculating the value of the first objective function. The inverse covariance matrix Θ corresponding to the minimum value of the first objective function is the estimated value of the solved inverse covariance matrix Θ. The first objective function is: Where Tr() represents the sum of the elements on the main diagonal of the matrix, ln det() represents the natural logarithm of the matrix determinant, and Θ prev Let be the inverse covariance matrix of the previous time step, β be the regularization parameter, and ||||1 denote the L1 norm;

[0203] The second solution method involves dividing the distribution network into several sub-regions. Within each sub-region, the value of the inverse covariance matrix is ​​randomly adjusted, and the value of the second objective function is calculated. The inverse covariance matrix corresponding to the minimum value of the second objective function is the estimated value of the inverse covariance matrix for the current round obtained from solving for that sub-region. The second objective function is... i represents the i-th sub-region, Θ i Let ||Θ| represent the inverse covariance matrix of the i-th sub-region; when the ||Θ| of adjacent iterations (k) -Θ (k-1) || - / ||Θ (k) || F <10 -4 The iteration stops when the estimated value of the inverse covariance matrix of the sub-region is obtained, Θ. (k) Let |||| represent the inverse covariance matrix obtained in the k-th iteration. F Denotes the Frobenius norm of a matrix;

[0204] S14. For the inverse covariance matrix Θ, if Θ ij If the value is less than or equal to -η, then nodes ij are considered connected; otherwise, they are considered disconnected. An estimated connectivity matrix A is then formed based on the connectivity assessment results between each node. est ; where Θ ij Let η be the element in the i-th row and j-th column of the inverse covariance matrix Θ, and let η be the threshold and η = |μ|. non-conn |+3σ non-conn μ non-conn σ is the mean of the elements in the inverse covariance matrix Θ. non-conn Let Θ be the standard deviation of the elements in the inverse covariance matrix; compare the estimated connectivity matrix A. est By comparing the estimated connectivity between nodes with the actual SCADA switch states, it is determined whether the estimated connectivity between nodes is consistent with the actual switch states. The accuracy and recall are then calculated. If the accuracy or recall does not meet the requirements, the process returns to step S13 above to re-estimate the inverse covariance matrix until the recall and accuracy meet the requirements, thus obtaining the final inverse covariance matrix.

[0205] S15. Based on the final inverse covariance matrix, the final estimated connectivity matrix is ​​obtained, the connectivity of nodes in the distribution network is obtained, and the distribution network topology model is built by combining the data collected in the offline stage.

[0206] More specifically, the real-time update module is also used for:

[0207] Real-time acquisition of S21, SCADA switch status, μPMU measurement data and smart meter data;

[0208] S22. Real-time acquired data is timestamped and the time delay is compensated. The μPMU measurement data is input into the trained isolated forest model for detection to obtain anomaly scores. When the anomaly score exceeds the preset score threshold, the voltage phase angle data corresponding to the anomaly score is removed. For SCADA switch states, the difference between the real-time value and the historical mean of each switch state is calculated. If it exceeds the historical mean ±3σ and there is no corresponding switch event, the data point is marked as bad data and repaired by interpolation of data before and after the time. σ is the standard deviation of the switch state data. After the above processing, the standardized dataset D(t) = [Δδ(t), V(t), P(t), Q(t), switch state at time t] is obtained, where Δδ(t) is the phase angle change sequence at time t, P(t) is the active power at time t, Q(t) is the reactive power at time t, and V(t) is the voltage amplitude at time t.

[0209] S23. Construct a state vector x = [δ1, V1, δ2, V2, ..., δ] using a standardized dataset. n V n ] T , where δ i Let V be the voltage phase angle at node i. i Let z be the voltage magnitude at node i, and let z = [z1, z2, ..., z2]. n ] T The expression z = h(x) + e satisfies z = h(x) + e, where z n Let represent the injected power and branch power flow of the nth node, h(x) be a nonlinear measurement function describing the relationship between the measured value and the system state, and e be the measurement error vector.

[0210] The distribution network is divided into several sub-regions, and each edge gateway is responsible for the local state estimation of one sub-region. The local optimization objective function for each sub-region k is: Among them, z k h represents the measurement value of the edge gateway corresponding to sub-region k. k (x) is the nonlinear measurement function of subregion k, w k h is continuously adjusted to assign measurement weights to the edge gateways corresponding to sub-region k. k The matrix parameters in (x) are used to calculate the value of the local optimization objective function. When the objective function is minimized, the update is stopped, and the corresponding optimized h is obtained. k (x), will optimize h k Substituting (x), node injected power and branch power flow into the formula z=h(x)+e, we obtain the state vector x, which is also the estimated results of the node's voltage phase angle and voltage amplitude. Based on this, we can complete the voltage phase angle and voltage amplitude data that were removed from the standardized dataset.

[0211] S24. Real-time estimation of the inverse covariance matrix. This real-time estimation process uses either a third objective function, the ADMM algorithm, or a weighted fusion of the results from both methods. The third objective function is: Where, Σ S Θ(t) is the sample covariance matrix of the current time window, Θ(t) is the inverse covariance matrix at time t, and λ is another regularization parameter. During the solution process using the third objective function, the value of the inverse covariance matrix is ​​continuously and randomly adjusted, and the value of the third objective function is calculated. When the value of the third objective function is minimized, the corresponding inverse covariance matrix Θ is the real-time estimate of the inverse covariance matrix Θ obtained by the solution.

[0212] S25. Compare the matrix elements in the real-time estimated value of the inverse covariance matrix obtained so far with the threshold η to obtain the real-time estimated connectivity matrix.

[0213] More specifically, the process of solving S24 using the ADMM algorithm is as follows:

[0214] S241. Initialization: Let Θ(0)=Z′(0)=(0)=0, U is the Lagrange multiplier, initialize the penalty parameter ρ=10, Z′(0) means that the auxiliary variable is initialized to a zero matrix in the 0th iteration, and U(0) means that the Lagrange multiplier is 0 in the 0th iteration;

[0215] S242, Θ Update: Fixed Z′ (k) and U (k) Solve

[0216]

[0217] Where k represents the k-th iteration;

[0218] S243, Z Update: Fixed Θ (k+1) and U (k) Solve

[0219]

[0220] S244, Lagrange multiplier update: U (k+1) =U (k) +Z'( k+1) -Θ (k+1) ;

[0221] S245. Determine whether the convergence condition is met, and determine ||Θ (k+1) -Z′( k+1) If || < ∈, and this condition is not met, return to step S242 to continue iterating; if it is met, stop iterating, and Θ is then... (k+1) This is the real-time estimate of the inverse covariance matrix, where ∈ is a pre-defined positive number.

[0222] More specifically, the collaborative positioning module is also used for:

[0223] S31. Based on the real-time estimated connectivity matrix, the connectivity status of the lines in the system is determined. Sensitivity detection is performed on the buses in the connectivity state. The sensitivity detection involves calculating the sensitivity matrix of each bus to the μPMU measurement. δ PMU This indicates the voltage phase angle measured by the μPMU. This represents finding the partial derivative, P. DER This represents the active power of the bus in a connected state; when S sen When the number of elements exceeds a set threshold, the bus area corresponding to that element is designated as a key monitoring area. The key monitoring area is searched using a breadth-first search algorithm to exclude disconnected feeder segments and complete the coarse positioning.

[0224] S32. Solve the power flow equations using the Newton-Raphson method to inversely deduce the node injection power; if node P i (t)>1.2×rated capacity and A est (t) indicates that the node is connected to the distributed power supply feeder, then it is marked as a suspected access point, P i (t) represents the injected power of the i-th node at time t, A est (t) represents the connectivity matrix at time t, and the distributed power supply access feeder refers to the feeder segment determined by S31;

[0225] S33. For nodes marked as suspected access points, extract their voltage phase angle sequence, convert it into an input format suitable for the SeriesTransformer model, input it into the pre-trained SeriesTransformer model, input the feature vector extracted by the pre-trained SeriesTransformer model into the classifier, obtain the prediction result through the classifier, and output whether the node is a distributed power access unit. The input format suitable for the SeriesTransformer model is a three-dimensional tensor, which consists of the number of samples in a single input, the voltage phase angle sequence, and the dimension of the input features.

[0226] More specifically, the processing procedure of the SeriesTransformer model is as follows:

[0227] The input data is processed by the inner layer, and the result is used to calculate attention. The result of the attention calculation and the inner layer processing is added together and then fed into the normalization layer. The output of the normalization layer is fed into the feedforward layer. The output of the feedforward layer and the output of the normalization layer are added together, normalized again, and then linearly projected to output the final result.

[0228] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for remotely locating a distributed power supply access unit, characterized in that, Includes the following steps: S1. In the offline phase, acquire historical data of SCADA switch status, μPMU measurement data and smart meter data, solve for the initial values ​​of the connectivity matrix and inverse covariance matrix of each bus, and build the distribution network topology model. S2. Real-time estimation of the inverse covariance matrix is ​​performed using real-time collected data to determine node connectivity, construct a real-time connectivity matrix, track changes in the distribution network topology, and achieve topology updates. S3. Based on the real-time estimated connectivity matrix, the connectivity status of the lines in the system is obtained. Sensitivity detection is performed on the buses in the connected state to obtain key monitoring areas. The key monitoring areas are searched using a breadth-first search algorithm to exclude disconnected feeder sections and complete coarse localization. Suspected access points are marked in the coarsely located areas. For nodes marked as suspected access points, their voltage phase angle sequence is extracted and input into the feature vector extracted by the pre-trained SeriesTransformer model. The prediction result is obtained through the classifier, and the output is whether the node is a distributed power supply access unit.

2. The remote positioning method for a distributed power supply access unit according to claim 1, characterized in that, S1 includes: Historical data acquisition of S11, SCADA switch status, μPMU measurement data and smart meter data. μPMU measurement data is the phase angle and amplitude of bus voltage, and smart meter data is the power injection data. S12. Model the distribution network as a graph structure G(v,e), where node v represents a bus and edge e represents a line switch. The sample covariance matrix S0 is obtained by summing and averaging the bus voltage phase angle change sequence and the transpose of the sequence within a time window. S13. Assume the inverse covariance matrix Θ has the same form as the sample covariance matrix. Initially, the inverse covariance matrix Θ is a zero matrix. The inverse covariance matrix Θ is solved using either the first solution method, the second solution method, or a weighted fusion of both methods. The first solution method involves continuously and randomly adjusting the value of the inverse covariance matrix Θ during the solution process and calculating the value of the first objective function. The inverse covariance matrix Θ corresponding to the minimum value of the first objective function is the estimated value of the solved inverse covariance matrix Θ. The first objective function is: Where Tr() represents the sum of the elements on the main diagonal of the matrix, lndet() represents the natural logarithm of the matrix determinant, and Θ prev Let be the inverse covariance matrix of the previous time step, β be the regularization parameter, and || ||1 denote the L1 norm; The second solution method involves dividing the distribution network into several sub-regions. Within each sub-region, the value of the inverse covariance matrix is ​​randomly adjusted, and the value of the second objective function is calculated. The inverse covariance matrix corresponding to the minimum value of the second objective function is the estimated value of the inverse covariance matrix for the current round obtained from solving for that sub-region. The second objective function is... i represents the i-th sub-region, Θ i Let ||Θ| represent the inverse covariance matrix of the i-th sub-region; when the ||Θ| of adjacent iterations (k) -Θ (k-1) || F / ||Θ (k) || F <10 -4 The iteration stops when the estimated value of the inverse covariance matrix of the sub-region is obtained, Θ. (k) Let || denote the inverse covariance matrix obtained in the k-th iteration, || || F Denotes the Frobenius norm of a matrix; S14. For the inverse covariance matrix Θ, if Θ ij If the value is less than or equal to -η, then nodes ij are considered connected; otherwise, they are considered disconnected. An estimated connectivity matrix A is then formed based on the connectivity assessment results between each node. est ; where Θ ij Let η be the element in the i-th row and j-th column of the inverse covariance matrix Θ, and let η be the threshold and η = |μ|. non-conn |+3σ non-conn μ non-conn σ is the mean of the elements in the inverse covariance matrix Θ. non-conn Let Θ be the standard deviation of the elements in the inverse covariance matrix; compare the estimated connectivity matrix A. est By comparing the estimated connectivity between nodes with the actual SCADA switch states, it is determined whether the estimated connectivity between nodes is consistent with the actual switch states. The accuracy and recall are then calculated. If the accuracy or recall does not meet the requirements, the process returns to step S13 above to re-estimate the inverse covariance matrix until the recall and accuracy meet the requirements, thus obtaining the final inverse covariance matrix. S15. Based on the final inverse covariance matrix, the final estimated connectivity matrix is ​​obtained, the connectivity of nodes in the distribution network is obtained, and the distribution network topology model is built by combining the data collected in the offline stage.

3. The remote positioning method for a distributed power supply access unit according to claim 2, characterized in that, S2 include: Real-time acquisition of S21, SCADA switch status, μPMU measurement data and smart meter data; S22. Real-time acquired data is timestamped and the time delay is compensated. The μPMU measurement data is input into the trained isolated forest model for detection to obtain anomaly scores. When the anomaly score exceeds the preset score threshold, the voltage phase angle data corresponding to the anomaly score is removed. For SCADA switch states, the difference between the real-time value and the historical mean of each switch state is calculated. If it exceeds the historical mean ±3σ and there is no corresponding switch event, the data point is marked as bad data and repaired by interpolation of data before and after the time. σ is the standard deviation of the switch state data. After the above processing, the standardized dataset D(t) = [Δδ(t), V(t), P(t), Q(t), switch state at time t] is obtained, where Δδ(t) is the phase angle change sequence at time t, P(t) is the active power at time t, Q(t) is the reactive power at time t, and V(t) is the voltage amplitude at time t. S23. Construct a state vector x = [δ1, V1, δ2, V2, ..., δ] using a standardized dataset. n V n ] T , where δ i Let V be the voltage phase angle at node i. i Let z be the voltage magnitude at node i, and let z = [z1, z2, ..., z2]. n ] T The expression z = h(x) + e satisfies z = h(x) + e, where z n Let represent the injected power and branch power flow of the nth node, h(x) be a nonlinear measurement function describing the relationship between the measured value and the system state, and e be the measurement error vector. The distribution network is divided into several sub-regions, and each edge gateway is responsible for the local state estimation of one sub-region. The local optimization objective function for each sub-region k is: Among them, z k h represents the measurement value of the edge gateway corresponding to sub-region k. k (x) is the nonlinear measurement function of subregion k, w k h is continuously adjusted to assign measurement weights to the edge gateways corresponding to sub-region k. k The matrix parameters in (x) are used to calculate the value of the local optimization objective function. When the objective function is minimized, the update is stopped, and the corresponding optimized h is obtained. k (x), will optimize h k Substituting (x), node injected power and branch power flow into the formula z=h(x)+e, we obtain the state vector x, which is also the estimated results of the node's voltage phase angle and voltage amplitude. Based on this, we can complete the voltage phase angle and voltage amplitude data that were removed from the standardized dataset. S24. Real-time estimation of the inverse covariance matrix. This real-time estimation process uses either a third objective function, the ADMM algorithm, or a weighted fusion of the results from both methods. The third objective function is: Where, Σ S Θ(t) is the sample covariance matrix of the current time window, Θ(t) is the inverse covariance matrix at time t, and λ is another regularization parameter. During the solution process using the third objective function, the value of the inverse covariance matrix is ​​continuously and randomly adjusted, and the value of the third objective function is calculated. When the value of the third objective function is minimized, the corresponding inverse covariance matrix Θ is the real-time estimate of the inverse covariance matrix Θ obtained by the solution. S25. Compare the matrix elements in the real-time estimated value of the inverse covariance matrix obtained at the current time with the threshold η to obtain the real-time estimated connectivity matrix. Based on the connectivity matrix, obtain the connectivity status of the nodes in the distribution network. Combine the estimated results of the voltage phase angle and voltage amplitude of the nodes with the phase angle change sequence and power data to update the distribution network topology.

4. The remote positioning method for a distributed power supply access unit according to claim 3, characterized in that, The process of solving S24 using the ADMM algorithm is as follows: S241. Initialization: Let Θ(0)=Z′(0)=(0)=0, U is the Lagrange multiplier, initialize the penalty parameter ρ=10, Z′(0) means that the auxiliary variable is initialized to a zero matrix in the 0th iteration, and U(0) means that the Lagrange multiplier is 0 in the 0th iteration; S242, Θ Update: Fixed Z′ (k) and U (k) Solve Where k represents the k-th iteration; S243, Z′ Update: Fixed Θ (k+1) and U (k) Solve S244, Lagrange multiplier update: U (k+1) =U (k) +Z'( k+1) -Θ (k+1) ; S245. Determine whether the convergence condition is met, and determine ||Θ (k+1) -Z′( k+1) If || < ∈, and this condition is not met, return to step S242 to continue iterating; if it is met, stop iterating, and Θ is then... (k+1) This is the real-time estimate of the inverse covariance matrix, where ∈ is a pre-defined positive number.

5. The remote positioning method for a distributed power supply access unit according to claim 3, characterized in that, S3 include: S31. Based on the real-time estimated connectivity matrix, the connectivity status of the lines in the system is determined. Sensitivity detection is performed on the buses in the connectivity state. The sensitivity detection involves calculating the sensitivity matrix of each bus to the μPMU measurement. δ PMU This indicates the voltage phase angle measured by the μPMU. This represents finding the partial derivative, P. DeR This represents the active power of the bus in a connected state; when S sen When the number of elements exceeds a set threshold, the bus area corresponding to that element is designated as a key monitoring area. The key monitoring area is searched using a breadth-first search algorithm to exclude disconnected feeder segments and complete the coarse positioning. S32. Solve the power flow equations using the Newton-Raphson method to inversely deduce the node injection power; if node P i (t)>1.2×rated capacity and A est (t) indicates that the node is connected to the distributed power supply feeder, then it is marked as a suspected access point, P i (t) represents the injected power of the i-th node at time t, A est (t) represents the connectivity matrix at time t, and the distributed power supply access feeder refers to the feeder segment determined by S31; S33. For nodes marked as suspected access points, extract their voltage phase angle sequence, convert it into an input format suitable for the SeriesTransformer model, input it into the pre-trained SeriesTransformer model, input the feature vector extracted by the pre-trained SeriesTransformer model into the classifier, obtain the prediction result through the classifier, and output whether the node is a distributed power supply access unit. The input format suitable for the SeriesTransformer model is a three-dimensional tensor, which consists of the number of samples in a single input, the voltage phase angle sequence, and the dimension of the input features.

6. The remote positioning method for a distributed power supply access unit according to claim 5, characterized in that, The processing procedure for the SeriesTransformer model is as follows: The input data is processed by the inner layer, and the result is used to calculate attention. The result of the attention calculation and the inner layer processing is added together and then fed into the normalization layer. The output of the normalization layer is fed into the feedforward layer. The output of the feedforward layer and the output of the normalization layer are added together, normalized again, and then linearly projected to output the final result.

7. A remote positioning system for a distributed power supply access unit, characterized in that, include: The initialization configuration module is used to acquire historical data of SCADA switch status, μPMU measurement data and smart meter data during the offline phase, solve for the initial values ​​of the connectivity matrix and inverse covariance matrix of each bus, and build the distribution network topology model. The real-time update module is used to estimate the inverse covariance matrix in real time using real-time collected data, determine the connectivity of nodes, construct a real-time connectivity matrix, track changes in the distribution network topology, and realize topology updates. The collaborative localization module is used to determine the connectivity status of lines in the system based on the real-time estimated connectivity matrix, perform sensitivity detection on buses in the connected state to identify key monitoring areas, and search the key monitoring areas using a breadth-first search algorithm to exclude disconnected feeder sections, thus completing coarse localization. Suspected access points are marked in the coarsely located areas. For nodes marked as suspected access points, their voltage phase angle sequences are extracted and input into the feature vectors extracted by the pre-trained SeriesTransformer model, which are then fed into a classifier. The classifier obtains the prediction result and outputs whether the node is a distributed power supply access unit.

8. A remote positioning system for a distributed power supply access unit according to claim 7, characterized in that, The initialization configuration module is also used for: Historical data acquisition of S11, SCADA switch status, μPMU measurement data and smart meter data. μPMU measurement data is the phase angle and amplitude of bus voltage, and smart meter data is the power injection data. S12. Model the distribution network as a graph structure G(v,e), where node v represents a bus and edge e represents a line switch. The sample covariance matrix S0 is obtained by summing and averaging the bus voltage phase angle change sequence and the transpose of the sequence within a time window. S13. Assume the inverse covariance matrix Θ has the same form as the sample covariance matrix. Initially, the inverse covariance matrix Θ is a zero matrix. The inverse covariance matrix Θ is solved using either the first solution method, the second solution method, or a weighted fusion of both methods. The first solution method involves continuously and randomly adjusting the value of the inverse covariance matrix Θ during the solution process and calculating the value of the first objective function. The inverse covariance matrix Θ corresponding to the minimum value of the first objective function is the estimated value of the solved inverse covariance matrix Θ. The first objective function is: Where Tr() represents the sum of the elements on the main diagonal of the matrix, ln det() represents the natural logarithm of the matrix determinant, and Θ prev Let be the inverse covariance matrix of the previous time step, β be the regularization parameter, and || ||1 denote the L1 norm; The second solution method involves dividing the distribution network into several sub-regions. Within each sub-region, the value of the inverse covariance matrix is ​​randomly adjusted, and the value of the second objective function is calculated. The inverse covariance matrix corresponding to the minimum value of the second objective function is the estimated value of the inverse covariance matrix for the current round obtained from solving for that sub-region. The second objective function is... i represents the i-th sub-region, Θ i Let ||Θ| represent the inverse covariance matrix of the i-th sub-region; when the ||Θ| of adjacent iterations (k) -Θ (k-1) || F / ||Θ (k) || F <10 -4 The iteration stops when the estimated value of the inverse covariance matrix of the sub-region is obtained, Θ. (k) Let || denote the inverse covariance matrix obtained in the k-th iteration, || || F Denotes the Frobenius norm of a matrix; S14. For the inverse covariance matrix Θ, if Θ ij If the value is less than or equal to -η, then nodes ij are considered connected; otherwise, they are considered disconnected. An estimated connectivity matrix A is then formed based on the connectivity assessment results between each node. est ; where Θ ij Let η be the element in the i-th row and j-th column of the inverse covariance matrix Θ, and let η be the threshold and η = |μ non-conn +3σ non-conn μ non-conn σ is the mean of the elements in the inverse covariance matrix Θ. non-conn Let Θ be the standard deviation of the elements in the inverse covariance matrix; compare the estimated connectivity matrix A. est By comparing the estimated connectivity between nodes with the actual SCADA switch states, it is determined whether the estimated connectivity between nodes is consistent with the actual switch states. The accuracy and recall are then calculated. If the accuracy or recall does not meet the requirements, the process returns to step S13 above to re-estimate the inverse covariance matrix until the recall and accuracy meet the requirements, thus obtaining the final inverse covariance matrix. S15. Based on the final inverse covariance matrix, the final estimated connectivity matrix is ​​obtained, the connectivity of nodes in the distribution network is obtained, and the distribution network topology model is built by combining the data collected in the offline stage.

9. A remote positioning system for a distributed power supply access unit according to claim 8, characterized in that, The real-time update module is also used for: Real-time acquisition of S21, SCADA switch status, μPMU measurement data and smart meter data; S22. Real-time acquired data is timestamped and the time delay is compensated. The μPMU measurement data is input into the trained isolated forest model for detection to obtain anomaly scores. When the anomaly score exceeds the preset score threshold, the voltage phase angle data corresponding to the anomaly score is removed. For SCADA switch states, the difference between the real-time value and the historical mean of each switch state is calculated. If it exceeds the historical mean ±3σ and there is no corresponding switch event, the data point is marked as bad data and repaired by interpolation of data before and after the time. σ is the standard deviation of the switch state data. After the above processing, the standardized dataset D(t) = [Δδ(t), V(t), P(t), Q(t), switch state at time t] is obtained, where Δδ(t) is the phase angle change sequence at time t, P(t) is the active power at time t, Q(t) is the reactive power at time t, and V(t) is the voltage amplitude at time t. S23. Construct a state vector x = [δ1, V1, δ2, V2, ..., δ] using a standardized dataset. n V n ] T , where δ i Let V be the voltage phase angle at node i. i Let z be the voltage magnitude at node i, and let z = [z1, z2, ..., z2]. n ] T The expression z = h(x) + e satisfies z = h(x) + e, where z n Let represent the injected power and branch power flow of the nth node, h(x) be a nonlinear measurement function describing the relationship between the measured value and the system state, and e be the measurement error vector. The distribution network is divided into several sub-regions, and each edge gateway is responsible for the local state estimation of one sub-region. The local optimization objective function for each sub-region k is: Among them, z k h represents the measurement value of the edge gateway corresponding to sub-region k. k (x) is the nonlinear measurement function of subregion k, w k h is continuously adjusted to assign measurement weights to the edge gateways corresponding to sub-region k. k The matrix parameters in (x) are used to calculate the value of the local optimization objective function. When the objective function is minimized, the update is stopped, and the corresponding optimized h is obtained. k (x), will optimize h k Substituting (x), node injected power and branch power flow into the formula z=h(x)+e, we obtain the state vector x, which is also the estimated results of the node's voltage phase angle and voltage amplitude. Based on this, we can complete the voltage phase angle and voltage amplitude data that were removed from the standardized dataset. S24. Real-time estimation of the inverse covariance matrix. This real-time estimation process uses either a third objective function, the ADMM algorithm, or a weighted fusion of the results from both methods. The third objective function is: Where, Σ S Θ(t) is the sample covariance matrix of the current time window, Θ(t) is the inverse covariance matrix at time t, and λ is another regularization parameter. During the solution process using the third objective function, the value of the inverse covariance matrix is ​​continuously and randomly adjusted, and the value of the third objective function is calculated. When the value of the third objective function is minimized, the corresponding inverse covariance matrix Θ is the real-time estimate of the inverse covariance matrix Θ obtained by the solution. S25. Compare the matrix elements in the real-time estimated value of the inverse covariance matrix obtained at the current time with the threshold η to obtain the real-time estimated connectivity matrix. Based on the connectivity matrix, obtain the connectivity status of the nodes in the distribution network. Combine the estimated results of the voltage phase angle and voltage amplitude of the nodes with the phase angle change sequence and power data to update the distribution network topology.

10. A remote positioning system for a distributed power supply access unit according to claim 9, characterized in that, The process of solving S24 using the ADMM algorithm is as follows: S241. Initialization: Let Θ(0)=Z′(0)=(0)=0, U is the Lagrange multiplier, initialize the penalty parameter ρ=10, Z′(0) means that the auxiliary variable is initialized to a zero matrix in the 0th iteration, and U(0) means that the Lagrange multiplier is 0 in the 0th iteration; S242, Θ Update: Fixed Z′ (k) and U (k) Solve Where k represents the k-th iteration; S243, Z′ Update: Fixed Θ (k+1) and U (k) Solve S244, Lagrange multiplier update: U (k+1) =U (k) +Z'( k+1) -Θ (k+1) ; S245. Determine whether the convergence condition is met, and determine ||Θ (k+1) -Z′ (k+1) If || < ∈, and this condition is not met, return to step S242 to continue iterating; if it is met, stop iterating, and Θ is then... (k+1) This is the real-time estimate of the inverse covariance matrix, where ∈ is a pre-defined positive number.

Images