CN-122020363-A - District topology identification method and system based on graphic neural network

CN122020363ACN 122020363 ACN122020363 ACN 122020363ACN-122020363-A

Abstract

The invention provides a district topology identification method and a district topology identification system based on a graphic neural network, and relates to the technical field of power distribution networks of power systems, wherein the method comprises the steps that an HPLC concentrator is used as an edge module, a district is modeled into a graphic structure, a transformer and a smart meter are used as nodes of the graphic structure, and potential electrical connection between the nodes is used as edges of the graphic structure; the method comprises the steps of collecting multi-source data of each intelligent ammeter in a platform area by an edge module, obtaining node characteristic vectors and edge characteristic matrixes according to the preprocessed multi-source data after preprocessing the multi-source data, learning node characteristics and edge characteristics by the edge module based on the obtained node characteristic vectors and the obtained edge characteristic matrixes by using a graph neural network, calculating the probability of electrical connection between any two nodes, and generating a platform area topological structure which accords with the set physical constraint by the edge module according to the obtained result by adopting a minimum spanning tree algorithm. The invention can realize the high-precision automatic identification of the topology of the area.

Inventors

Wang Zita
HUANG CHUZHI
HUANG ZONGWEI
ZHUANG SHAOYANG
CAI SHANSHAN
CHEN WEIPENG

Assignees

泉州亿兴电力工程建设有限公司鲤城自动化分公司

Dates

Publication Date: 20260512
Application Date: 20251231

Claims (10)

1. The method for identifying the region topology based on the graph neural network is characterized by comprising the following steps: Step S1, an HPLC concentrator is used as an edge module, a platform area is modeled into a graph structure, a transformer and a smart meter are used as nodes of the graph structure, and potential electrical connection between the nodes is used as edges of the graph structure; S2, acquiring multi-source data of each intelligent ammeter in the platform area by an edge module, preprocessing the multi-source data, and acquiring node feature vectors and edge feature matrixes according to the preprocessed multi-source data; Step S3, the edge module learns node characteristics and edge characteristics by using a graph neural network based on the node characteristic vector and the edge characteristic matrix obtained in the step S2, and calculates the probability of electrical connection between any two nodes; And S4, generating a platform region topological structure conforming to the set physical constraint by adopting a minimum spanning tree algorithm according to the result obtained in the step S3 by the edge module.
2. The method for identifying a region topology based on a neural network according to claim 1, wherein in the step S2, the multi-source data includes voltage data, current data, HPLC time delay, signal intensity, power factor and harmonic content, and the preprocessing of the multi-source data includes missing value filling, outlier detection and rejection, time alignment and normalization.
3. The method for identifying a region topology based on a graphic neural network as recited in claim 2, wherein in the step S2, the node feature vector includes a voltage average, a voltage variance, a current average, a current variance, a power factor average, a third harmonic, a fifth harmonic, and a seventh harmonic of the node, and the edge feature matrix includes a voltage correlation between the node i and the node j Distance to electrical Poor signal strength Difference in power factor Harmonic content difference , wherein, , , 、 Representing the voltage data sequences of nodes i and j respectively, In order to solve for the covariance, In order to obtain the standard deviation of the standard, As the propagation speed of the electromagnetic wave, As a characteristic coefficient of the cable, Is the HPLC delay between node i and node j.
4. The method for identifying the topology of the area based on the graphic neural network of claim 3, wherein in the step S3, the graphic neural network comprises at least two layers of graphic attention layers and an edge prediction module, the graphic attention layers learn node characteristics and edge characteristics by adopting a multi-head attention mechanism, and the edge prediction module combines the learned node characteristics and the edge characteristics through an MLP network to predict the probability of electrical connection between each pair of nodes one by one.
5. The method for identifying the topology of the area based on the graphic neural network according to claim 1,2,3 or 4, wherein in the step S4, a probability adjacency matrix is constructed according to the probability predicted in the step S3, the physical constraint comprises taking the transformer as a root node, taking the tree topology as a loop-free topology, and setting the upper limit of the number of sub-nodes of each node, and each side in the obtained topology of the area has the original probability reserved as the confidence coefficient.
6. The method for identifying the topology of the area based on the graph neural network according to claim 1, 2, 3 or 4, wherein in the step S4, each node in the topology of the area is judged, if the node is an explicit node, the node is kept in the topology, if the node is a fuzzy node, the node, a neighborhood node of the node and an edge feature matrix of the node are packaged and reported to a cloud server, the cloud server uses a complete graph neural network, performs secondary reasoning by combining a plurality of area history data and more set physical constraints, and issues a reasoning result to an edge module, and the edge module updates the topology of the area, wherein the graph neural network used by the edge module is obtained by adopting a model compression technology for the complete graph neural network.
7. The method of claim 6, wherein the node is a distinct node when the confidence levels of all reserved edges of the node are greater than a set high confidence threshold, and the node is a fuzzy node when the confidence levels of all reserved edges of the node are less than a set low confidence threshold, or the difference between the maximum confidence level and the next-largest confidence level is less than 0.1.
8. The method for identifying the region topology based on the graphic neural network according to claim 3 or 4, further comprising the steps of S5, continuously collecting multi-source data by an edge module, obtaining a current fusion feature vector according to an edge feature matrix obtained by the multi-source data, and calculating L2 norms of the current fusion feature vector and a historical fusion feature vector in the last topological stabilization at intervals of set time When (when) When the difference threshold value is larger than the set difference threshold value, the topology change area is positioned and extracted, incremental learning is adopted to conduct reasoning so as to update the confidence coefficient of the edges of the nodes of the topology change area, and when And when the difference value is smaller than the set difference threshold value, the topology is kept unchanged.
9. The method for identifying a cell topology based on a neural network of claim 8, wherein in step S5, the comparison is performed And when the voltage correlation change is greater than 0.3 or the HPLC delay change is greater than 20%, or the new node is on-line or the node is off-line, positioning and extracting a topology change area, and updating the topology.
10. The platform region topology identification system based on the graph neural network is used for realizing the identification method according to any one of claims 1 to 9, and is characterized by comprising an edge module and a cloud server, wherein an HPLC concentrator is used as the edge module, the edge module models a platform region as a graph structure, transformers and intelligent electric meters are used as nodes of the graph structure, potential electric connection among the nodes is used as edges of the graph structure, the edge module collects multi-source data of all intelligent electric meters in the platform region, after preprocessing the multi-source data, node feature vectors and edge feature matrixes are obtained according to the preprocessed multi-source data, the node feature vectors and the edge feature matrixes are learned by utilizing a lightweight graph neural network, the probability of electric connection between any two nodes is calculated, a minimum spanning tree algorithm is used for generating a platform region topology structure conforming to the set physical constraint according to the calculated probability, nodes in the generated platform region topology structure are judged to be clear nodes or cloud nodes according to confidence, the fuzzy nodes, the neighborhood nodes of the fuzzy nodes and the edge feature matrixes of the fuzzy nodes are packed to the server, the fuzzy nodes are used for the graph structure, the graph neural network is used for achieving the complete graph structure, the graph structure is updated by utilizing the fuzzy neural network, and the fuzzy neural network is used for the graph structure is more completely updating the edge constraint graph network, and the result is obtained by adopting the fuzzy neural network.

Description

District topology identification method and system based on graphic neural network Technical Field The invention relates to the technical field of power distribution networks of power systems, in particular to a district topology identification method and system based on a graphic neural network. Background The transformer area (distribution transformer power supply area) is a basic management unit of the power distribution network, and the accurate topological relation of the transformer area is a basis for realizing line loss calculation, load prediction and fault positioning. The traditional platform topology identification mainly depends on a manual line inspection mode, but the mode has a plurality of problems, such as low efficiency due to manual field inspection, untimely updating of file data, high user-to-user relationship error rate of 15% -30%, and incapability of coping with frequent line transformation and user migration. In order to solve the problems existing in manual line inspection, some new solutions are provided, but all have limitations. If the voltage correlation is used for analysis, the influence of load fluctuation is large, the identification accuracy is less than 85%, extra hardware investment is needed by using a platform area identification instrument, the maintenance cost is high, the influence of the line length and the branches is caused by using the carrier signal intensity of the HPLC, the misjudgment rate is high, and the coverage is insufficient by using the matching of outage events and depending on the historical outage data. Besides, the access of new energy sources also brings updating challenges, such as changing the power flow direction of a platform area by a large amount of distributed photovoltaic access, affecting voltage characteristics by random loads of electric automobile charging piles, and dynamically changing topological relations caused by switching of energy storage equipment, which affect the identification of the topology of the platform area. Disclosure of Invention The invention mainly aims to provide a district topology identification method and system based on a graphic neural network, which realize high-precision automatic identification of district topology. The invention is realized by the following technical scheme: the district topology identification method based on the graph neural network comprises the following steps: Step S1, an HPLC concentrator is used as an edge module, a platform area is modeled into a graph structure, a transformer and a smart meter are used as nodes of the graph structure, and potential electrical connection between the nodes is used as edges of the graph structure; S2, acquiring multi-source data of each intelligent ammeter in the platform area by an edge module, preprocessing the multi-source data, and acquiring node feature vectors and edge feature matrixes according to the preprocessed multi-source data; Step S3, the edge module learns node characteristics and edge characteristics by using a graph neural network based on the node characteristic vector and the edge characteristic matrix obtained in the step S2, and calculates the probability of electrical connection between any two nodes; And S4, generating a platform region topological structure conforming to the set physical constraint by adopting a minimum spanning tree algorithm according to the result obtained in the step S3 by the edge module. Further, in the step S2, the multi-source data includes voltage data, current data, HPLC delay, signal strength, power factor and harmonic content, and the preprocessing of the multi-source data includes missing value filling, outlier detection and rejection, time alignment and normalization. Further, in the step S2, the node feature vector includes a voltage average, a voltage variance, a current average, a current variance, a power factor average, a third harmonic, a fifth harmonic, and a seventh harmonic of the node, and the edge feature matrix includes a voltage correlation between the node i and the node jDistance to electricalPoor signal strengthDifference in power factorHarmonic content difference, wherein,,,、Representing the voltage data sequences of nodes i and j respectively,In order to solve for the covariance,In order to obtain the standard deviation of the standard,As the propagation speed of the electromagnetic wave,As a characteristic coefficient of the cable,Is the HPLC delay between node i and node j. Further, in step S3, the graph neural network includes at least two layers of graph attention layers and an edge prediction module, the graph attention layers learn node features and edge features by using a multi-head attention mechanism, and the edge prediction module combines the learned node features and edge features through the MLP network to predict the probability of electrical connection between each pair of nodes one by one. Further, in the step S4, a probability adjacency matrix is constructe