Power distribution network topology intelligent identification method considering structural similarity

Abstract

The invention discloses a power distribution network topology intelligent identification method considering structural similarity, relates to the technical field of power distribution network modeling, and aims to construct a novel power distribution network topology intelligent identification method oriented to a high-permeability distributed energy access scene. The method comprises the following steps: firstly, extracting potential correlation characteristics of a light storage and charging equipment group in dimensions such as power fluctuation and voltage response, and overcoming the problem of characteristic space fragmentation caused by a traditional independent hypothesis; secondly, designing a transfer learning framework based on the maximum mean value difference, realizing identification knowledge transfer from a known topological structure scene to a novel station access scene, and reducing a model generalization error in a small sample scene; and finally, proposing a hybrid optimization algorithm fusing a wolf pack cooperation mechanism and gradient direction guidance. According to the method, the problem of topology identification misalignment caused by new energy random access is solved, an explainable and extensible technical base is provided for situation awareness of a novel power system, and the method has remarkable subject cross innovation value and engineering application prospects.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network modeling, and specifically, to an intelligent identification method for distribution network topology considering structural similarity. Background Art

[0002] Driven by the "dual carbon" goal, the active distribution network containing a high proportion of distributed photovoltaics, energy storage, and electric vehicle charging facilities has become a typical form of the new power system. However, the spatio-temporal coupling characteristics of a large number of heterogeneous energy units result in strong time-varying and high-dimensional non-linear characteristics of the distribution network topology. Traditional topology identification methods based on steady-state measurements face severe challenges such as a sharp drop in accuracy and an exponential increase in computational complexity. According to statistics, the misoperation rate of protection caused by topology identification errors in the main power grids of the United States and Europe has risen to 2.3 times per thousand nodes per year, seriously threatening power supply reliability. Developing a new topology identification system that takes into account physical mechanisms and data-driven approaches has become the core technical bottleneck for ensuring the safe and economic operation of the new distribution system, and has great engineering application value and interdisciplinary innovation significance.

[0003] There are currently three major technical gaps in the research: First, at the data level, existing methods mostly use independent assumptions to process heterogeneous data of photovoltaic, energy storage, and charging, ignoring the operational correlation characteristics generated by the physical structure similarity among equipment groups; second, at the model level, mainstream machine learning methods are limited by insufficient spatio-temporal coverage of training data, and the generalization performance drops by more than 40% when new types of stations are connected; third, at the algorithm level, traditional optimization methods face the curse of dimensionality - the system identification of the distribution network takes too long and it is difficult to meet the real-time requirements. Summary of the Invention

[0004] The purpose of the present invention is to provide an intelligent identification method for distribution network topology considering structural similarity. First, by exploring the structural similarity of photovoltaic, energy storage, and charging stations, a feature space embedded with physical information can be constructed to provide prior knowledge constraints for overcoming "data fragmentation"; second, introducing transfer learning can achieve cross-regional and cross-scenario topology feature knowledge distillation to solve the problem of scarce training samples caused by the random access of new energy; finally, the wolf hill-climbing algorithm that combines swarm intelligence and gradient optimization can search for the global optimal solution within polynomial time complexity to solve the problems raised in the background art.

[0005] To achieve the above purpose, the present invention provides the following technical solution: An intelligent identification method for distribution network topology considering structural similarity, including the following steps:

[0006] S1. Construct a multi-dimensional topology identification information set considering the operational similarity of photovoltaic, energy storage, and charging stations;

[0007] S2. Conduct cross-scenario knowledge traction based on transfer learning and expand the information set constructed in S1;

[0008] S3. Based on the wolf hill climbing algorithm, the information set is optimized from global to local topology identification.

[0009] According to the above technical solution, in step S1, the following steps are included:

[0010] S101, collecting and preprocessing multi-source data of the distribution network;

[0011] S102. Construct a structural similarity quantification model for the data preprocessed in S101.

[0012] According to the above technical solution, in step S101, it is assumed that the distribution network includes N photovoltaic storage and charging station nodes, and each node collects m-dimensional operating data within the time window T, including but not limited to active power P t , reactive power Q t , voltage amplitude V t , harmonic distortion rate THD t , forming the original data matrix:

[0013]

[0014] in:

[0015] T: the length of the data collection time window, in time steps, for example, minutes or hours;

[0016] m: The physical quantity dimension collected by each node, such as active power, reactive power, and other different types of data;

[0017] X i ∈R m×T : The original data matrix of the i-th node, each row corresponds to a physical quantity, and each column corresponds to a time step;

[0018] The original value of the k-th dimension physical quantity of the i-th node at the 1st, 2nd, ...Tth time step. For example, k = 1 represents active power, k = 2 represents reactive power, k = 3 represents voltage amplitude, and k = 4 represents harmonic distortion rate.

[0019] According to the above technical solution, in step S102, the following steps are included:

[0020] S102-1. Calculate the DTW distance for the k-dimensional time series of any two station nodes i and j.

[0021] S102-2. Calculate nonlinear correlations between device response characteristics;

[0022] S102-3. Perform similarity feature fusion on the nonlinear correlation between the DTW distances between the station nodes and the device response characteristics.

[0023] According to the above technical solution, in step S102-1, for the calculation of the DTW distance, it is specifically as follows:

[0024]

[0025] Among them, represents the dynamic time warping distance between the station nodes i and j in the k-th dimensional physical quantity. The smaller the value, the more similar the time series fluctuation patterns are; π represents the warping path; tp, tq represent the time step alignment indexes of the station nodes i and j in the warping path;

[0026] In step S102-2, for the calculation of the non-linear correlation, it is specifically as follows:

[0027]

[0028] Among them, represents the mutual information between the k-th dimension of the station node i and the l-th dimension of the station node j in the physical quantity, which measures the non-linear correlation; p(x), p(y) represent the marginal probability distribution functions of the physical quantities x and y; p(x, y) is the joint probability distribution;

[0029] In step S102-3, for the similarity feature fusion, it is specifically as follows:

[0030]

[0031] Among them, S i,j is the comprehensive similarity matrix. The closer it is to 1, the stronger the topological correlation between the station nodes i and j; w k is the feature weight of the k-th dimensional physical quantity; H(x i ) is the information entropy of the multi-dimensional data of the node i;

[0032] The comprehensive similarity matrix constitutes an information set.

[0033] According to the above technical solution, in step S2, it includes the following steps:

[0034] S201. Structurally similar weighted domain adaptation;

[0035] S202. Minimize the joint loss to achieve knowledge transfer.

[0036] According to the above technical solution, in step S201, use the generated similarity matrix S i,j ,

[0037] to improve the feature alignment process of transfer learning; define the weighted distribution difference between the source domain (known topology) and the target domain (new scenario):

[0038]

[0039] Among them, is the feature vector of the source domain node j; is the feature vector of the target domain node i; N s , N t are the numbers of nodes in the source domain and the target domain respectively; φ is the reproducing kernel Hilbert space mapping function;

[0040] In step S202, the knowledge transfer is realized by minimizing the joint loss as follows:

[0041]

[0042] Among them, h is the topological classifier; is the true topological label of the source domain node j; λ is the trade-off parameter; L is the knowledge transfer loss function.

[0043] According to the above technical solution, in step S3, node pairs that meet the threshold conditions are extracted from the information set in step S1, and at the same time, combined with the new scenario obtained by migrating in step S2, the initial wolf pack individuals are constructed together. Each individual represents a possible topological connection hypothesis, ensuring that the initial solution set has both relevance and data-driven characteristics;

[0044] Select the top a% individuals with the highest fitness as the leading wolves, that is, α wolves, and the remaining wolf pack is β wolves. The β wolves perform refined searches around the positions of the α wolves;

[0045] Utilize the gradient change trend between the node feature vectors in step S202 to locate the region with strong topological relevance in the feature space and quickly narrow the search range;

[0046] Based on the similarity matrix in step S1, divide the β wolves into several subgroups. Individuals within the same subgroup only optimize the topological connections between nodes with high similarity to avoid ineffective searches;

[0047] Call the migration scenario in step S2 to verify the consistency of the current topological hypothesis in real time; if a certain hypothesis makes the migration scenario meaningless, immediately eliminate this individual;

[0048] When the fitness change of the optimal solution in b consecutive iterations is less than c%; and the passing rate of the migration scenario verification in step S2 reaches more than d%, the termination determination is made.

[0049] Compared with the prior art, the beneficial effects of the present invention are:

[0050] The present invention aims to construct a new method for intelligent identification of distribution network topology for scenarios with high-penetration distributed energy access. First, potential correlation features of the photovoltaic-storage-charging device group in dimensions such as power fluctuation and voltage response are extracted to overcome the problem of fragmented feature space caused by traditional independent assumptions. Second, a transfer learning framework based on maximum mean discrepancy is designed to achieve the transfer of identification knowledge from known topology scenarios to new substation access scenarios, reducing the model generalization error in small-sample scenarios. Finally, a hybrid optimization algorithm that combines the wolf pack cooperation mechanism and gradient direction guidance is proposed. The invention solves the problem of inaccurate topology identification caused by random access of new energy, provides an interpretable and extensible technical foundation for the situation awareness of new power systems, and has significant interdisciplinary innovation value and engineering application prospects. Description of the Drawings

[0051] Figure 1 Schematic diagram of the computational node similarity matrix for an embodiment of the present invention;

[0052] Figure 2 Schematic diagram of generating a new substation after transfer learning for an embodiment of the present invention;

[0053] Figure 3 Schematic diagram of the convergence of the wolf hill-climbing algorithm after 15 iterations for an embodiment of the present invention. Detailed Embodiments

[0054] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0055] The present invention provides a technical solution for an intelligent identification method of distribution network topology considering structural similarity, including the following steps:

[0056] S1. Construct a multi-dimensional topology identification information set considering the operation similarity of photovoltaic-storage-charging substations;

[0057] Including the following steps:

[0058] S101. Perform multi-source data collection and preprocessing of the distribution network;

[0059] Suppose the distribution network includes N photovoltaic-storage-charging substation nodes, and each node collects m-dimensional operation data within the time window T, including but not limited to active power P t , reactive power Q t , voltage amplitude V t , harmonic distortion rate THD t , to form the original data matrix:

[0060]

[0061] Wherein:

[0062] T: The length of the data acquisition time window, in time steps, such as minutes, hours;

[0063] m: The dimension of the physical quantity collected by each node, such as different types of data like active power, reactive power, etc.;

[0064] X i ∈R m×T : The original data matrix of the i-th node, where each row corresponds to a physical quantity and each column corresponds to a time step;

[0065] The original value of the k-th dimension physical quantity of the i-th node at the 1st, 2nd, …, T-th time steps, for example: k = 1 represents active power, k = 2 represents reactive power, k = 3 represents voltage amplitude, k = 4 represents harmonic distortion rate.

[0066] S102. Construct a structural similarity quantization model for the data preprocessed in S101;

[0067] Including the following steps:

[0068] S102-1. Calculate the DTW distance for the k-dimensional time series of any two substation nodes i, j;

[0069] For the calculation of the DTW distance, specifically as follows:

[0070]

[0071] Wherein, represents the dynamic time warping distance between substation nodes i and j on the k-th dimension physical quantity, and the smaller the value, the more similar the time series fluctuation patterns; π represents the warping path; tp, tq represent the time step alignment indices of substation nodes i and j in the warping path;

[0072] S102-2. Calculate the non-linear correlation between equipment response characteristics;

[0073] For the calculation of the non-linear correlation, specifically as follows:

[0074]

[0075] Wherein, represents the mutual information between the k-th dimension of substation node i and the l-th dimension of substation node j of the physical quantity, measuring the non-linear correlation; p(x), p(y) represent the marginal probability distribution functions of physical quantities x, y; p(x, y) is the joint probability distribution;

[0076] S102-3. Perform similarity feature fusion on the non-linear correlation between the DTW distance between station nodes and the equipment response characteristics;

[0077] For the similarity feature fusion, it is as follows:

[0078]

[0079] Among them, S i,j is the comprehensive similarity matrix. The closer it is to 1, the stronger the topological correlation between station nodes i and j; w k is the feature weight of the k-th dimensional physical quantity; H(x i ) is the information entropy of the multi-dimensional data of node i;

[0080] The comprehensive similarity matrix constitutes the information set.

[0081] S2. Perform cross-scenario knowledge traction based on transfer learning and expand the information set constructed in S1;

[0082] It includes the following steps:

[0083] S201. Structure similarity weighted domain adaptation;

[0084] Use the generated similarity matrix S i,j to improve the feature alignment process of transfer learning; Define the weighted distribution difference between the source domain (known topology) and the target domain (new scenario):

[0085]

[0086] Among them, is the feature vector of source domain node j; is the feature vector of target domain node i; N s , N t are the number of nodes in the source domain and the target domain respectively; φ is the reproducing kernel Hilbert space mapping function;

[0087] S202. Minimize the joint loss to achieve knowledge transfer;

[0088] Specifically as follows:

[0089]

[0090] Among them, h is the topology classifier; is the true topology label of source domain node j; λ is the trade-off parameter; L is the knowledge transfer loss function.

[0091] S3. Perform global-to-local optimized topology identification on the information set based on the wolf hill climbing algorithm.

[0092] Specifically, node pairs that meet the threshold conditions are extracted from the information set in step S1, and together with the new scenarios migrated in step S2, initial wolf pack individuals are constructed. Each individual represents a possible topological connection hypothesis, ensuring that the initial solution set has both relevance and data-driven characteristics;

[0093] Select the top a% of individuals with the highest fitness as the leading wolves, i.e., α wolves, and the remaining wolf pack as β wolves. The β wolves perform refined searches around the positions of the α wolves;

[0094] Utilize the gradient change trend between the node feature vectors in step S202 to locate the regions with strong topological relevance in the feature space and quickly narrow the search scope;

[0095] Based on the similarity matrix in step S1, divide the β wolves into several subgroups. Individuals within the same subgroup only optimize the topological connections between nodes with high similarity to avoid ineffective searches;

[0096] Call the migration scenario in step S2 to verify the consistency of the current topological hypothesis in real time; if a certain hypothesis makes the migration scenario meaningless, immediately eliminate this individual;

[0097] When the fitness change of the optimal solution in 100 consecutive iterations is less than 1%; and the passing rate of the migration scenario verification in step S2 reaches more than 95%, terminate the determination.

[0098] Example:

[0099] To verify the effectiveness of the present invention, we select a distribution network connected to a photovoltaic-storage-charging station as the research object. It includes 12 photovoltaic stations (total capacity 8.2 MW), 6 energy storage stations (total capacity 4.5 MWh), and 18 charging stations (daily charging volume 35 MWh). The node topological structure changes 3 - 5 times a day due to the fluctuation of new energy output.

[0100] Training set: Historical data for 20 days (including 15 typical topological structures).

[0101] Test set: Operation data for 5 days after adding 3 new photovoltaic-storage-charging stations.

[0102] Implementation steps

[0103] Step 1 execution: Extract features such as the active power and voltage harmonics of each node.

[0104] Calculate the node similarity matrix, as Figure 1 shown, indicating that there is strong coupling in geographical proximity.

[0105] Step 2 execution: Generate new stations after transfer learning, as Figure 2As shown, the confidence threshold τ = 0.85. The information set expands from the original 12,000 groups to 18,000 groups, and 94.3% of the newly added data passes the consistency verification.

[0106] Step 3 is executed: The wolf climbing algorithm converges after 15 iterations, as Figure 3 shown, taking 4 minutes and 12 seconds. It has a lower convergence threshold compared to the traditional PSO algorithm.

[0107] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention. Any reference signs in the claims should not be construed as limiting the claimed rights.

Claims

1. An intelligent identification method for the topology of a distribution network considering structural similarity, characterized in that It includes the following steps: S1. Construct a multi-dimensional topological identification information set considering the operational similarity of the optical storage and charging stations; S2. Conduct cross-scenario knowledge traction based on transfer learning and expand the information set constructed in S1; S3. Conduct global-to-local optimal topological identification on the information set based on the wolf hill climbing algorithm.

2. The intelligent identification method for the distribution network topology considering structural similarity according to claim 1, characterized in that In step S1, it includes the following steps: S101. Collect and preprocess multi-source data of the distribution network; S102. Construct a structural similarity quantification model for the data preprocessed in S101.

3. The intelligent identification method for the topology of a distribution network considering structural similarity according to claim 2, characterized in that, In step S101, assume that the distribution network contains N optical storage and charging station nodes, and each node collects m-dimensional operation data within the time window T to form an original data matrix: Where: T: The length of the data collection time window, in time steps; m: The dimension of the physical quantity collected by each node; X i ∈R m×T : The original data matrix of the i-th node, where each row corresponds to a physical quantity and each column corresponds to a time step; The original value of the physical quantity of the k-th dimension of the i-th node at the 1st, 2nd, …, T-th time steps.

4. The intelligent identification method for the topology of a distribution network considering structural similarity according to claim 3, characterized in that In step S102, it includes the following steps: S102-1. Calculate the DTW distance for the k-dimensional time series of any two station nodes i and j; S102-2. Calculate the non-linear correlation between the device response characteristics; S102-3. Conduct similarity feature fusion on the DTW distance between the station nodes and the non-linear correlation between the device response characteristics.

5. The intelligent identification method for the topology of a distribution network considering structural similarity according to claim 4, characterized in that In step S102-1, for the calculation of the DTW distance, it is as follows: Among them, represents the dynamic time warping distance between the station nodes i and j in the k-th physical quantity. The smaller the value, the more similar the time series fluctuation patterns are; π represents the warping path; tp and tq represent the time step alignment indices of the station nodes i and j in the warping path. In step S102-2, for the calculation of the non-linear correlation, it is as follows: Among them, represents the mutual information between the k-th dimension of the station node i and the l-th dimension of the physical quantity of the station node j, measuring the non-linear correlation; p(x) and p(y) represent the marginal probability distribution functions of the physical quantities x and y; p(x,y) is the joint probability distribution; In step S102-3, for the similarity feature fusion, it is as follows: Among them, S i,j is the comprehensive similarity matrix. The closer it is to 1, the stronger the topological correlation between the station nodes i and j; w k is the characteristic weight of the k-th dimensional physical quantity; H(x i ) is the information entropy of the multi-dimensional data of node i. The comprehensive similarity matrix constitutes the information set.

6. The intelligent identification method for the topology of a distribution network considering structural similarity according to claim 5, characterized in that, In step S2, it includes the following steps: S201. Structural similarity weighted domain adaptation; S202. Minimize the joint loss to achieve knowledge transfer.

7. The intelligent identification method for the topology of a distribution network considering structural similarity according to claim 6, characterized in that In step S201, the generated similarity matrix S is utilized i,j , Improve the feature alignment process of transfer learning; define the weighted distribution difference between the source domain and the target domain: Among them, is the feature vector of the source domain node j; is the feature vector of the target domain node i; N s , N t are the number of nodes in the source domain and the target domain respectively; φ is the reproducing kernel Hilbert space mapping function; In step S202, minimize the joint loss to achieve knowledge transfer as follows: Among them, h is the topological classifier; is the true topological label of the source domain node j; λ is the trade-off parameter; L is the knowledge transfer loss function.

8. The intelligent identification method for the distribution network topology considering structural similarity according to claim 1, characterized in that In step S3, extract the node pairs that meet the threshold conditions from the information set in step S1, and jointly construct the initial wolf pack individuals in combination with the new scenarios migrated in step S2. Each individual represents a possible topological connection hypothesis, ensuring that the initial solution set has both relevance and data-driven characteristics; Select the top a% individuals with the highest fitness as the leading wolves, that is, the α wolves, and the remaining wolf pack are the β wolves. The β wolves conduct refined searches around the positions of the α wolves; Use the gradient change trend between the node feature vectors in step S202 to locate the region with strong topological relevance in the feature space; Based on the similarity matrix in step S1, divide the β wolves into several subgroups, and the individuals within the same subgroup only optimize the topological connections between nodes with high similarity; Call the migrated scenarios in step S2 to verify the consistency of the current topological hypothesis in real time; If a certain hypothesis makes the migrated scenario meaningless, immediately eliminate this individual; When the fitness change of the optimal solution in b consecutive iterations is less than c%; and the verification pass rate of the migrated scenarios in step S2 reaches more than d%, terminate the determination.

Images