Low-voltage distribution network node topology identification and user node dynamic adjustment method

Abstract

The invention provides a low-voltage distribution network node topology identification and user node dynamic adjustment method, and aims to solve the problems of poor real-time performance, high calculation complexity and insufficient adaptability in dynamic topology identification of a traditional method. High-precision topology identification is realized by collecting voltage, current and power time sequence data of devices such as an intelligent electric meter and a PMU, extracting time sequence dependence of node features by using LSTM, modeling spatial association of a power grid topology in combination with a graph convolutional network, generating a reconstructed adjacency matrix and calculating binary cross entropy loss. Furthermore, the connection state of the user nodes is dynamically adjusted based on the reconstruction topology and the power flow data change, and the operation efficiency and stability of the power grid are improved. According to the method, time sequence analysis and graph structure modeling are fused, dependence on feature engineering is remarkably reduced, the model generalization ability and robustness are enhanced, the method is suitable for real-time topology inference, fault location and energy efficiency optimization scenes of the low-voltage power distribution network, and an efficient and reliable technical scheme is provided for dynamic management of an intelligent power grid.

Description

Technical Field

[0001] The present invention belongs to the technical field of smart grids and low-voltage distribution networks, and relates to a method for identifying low-voltage distribution network node topology and dynamically adjusting user nodes. Background Art

[0002] With the rapid development of smart grid technology, topology identification of low-voltage distribution networks (LVDNs) has become a crucial research topic for ensuring stable distribution network operation. Accurately understanding the topological relationships of distribution networks not only facilitates fault location and isolation, but also optimizes energy efficiency management and enhances demand-side responsiveness. However, LVDNs have long faced challenges such as a wide variety of equipment, inconsistent implementation standards, and varying levels of intelligence. Traditional topology identification methods, which rely on manual inspections and static models, suffer from high cost, low efficiency, and poor adaptability to dynamic changes. Although physical modeling approaches such as convex optimization and wavelet transforms have achieved some success in early research, they are sensitive to measurement errors and data omissions, and their high computational complexity makes them difficult to meet the real-time requirements of large-scale LVDNs. In recent years, data-driven statistical learning methods, combined with new monitoring devices such as smart meters and synchronized phasor measurement units (PMUs), have provided more comprehensive data support for topology identification. However, data incompleteness, noise interference, and dynamic topology changes still hinder identification accuracy.

[0003] Although topology recognition methods based on machine learning (such as XGBoost and random forests) reduce computing requirements, they are limited by the quality of feature engineering, have poor generalization capabilities, and are difficult to adapt to complex and dynamic topological changes. Deep learning methods (such as graph neural networks (GNNs) and long short-term memory networks (LSTMs)) have improved recognition accuracy, but require a large amount of labeled data for training, and the decision-making process is difficult to explain, which limits its practical application in the power industry. In addition, although reinforcement learning (such as deep Q networks (DQNs)) has shown potential in topology optimization, its training cycle is long, it relies on a high-quality simulation environment, and it is difficult to meet real-time requirements. Existing technologies still have shortcomings in the real-time, accuracy, and robustness of dynamic topology recognition. There is an urgent need for an efficient method that combines time series modeling and graph structure analysis.

[0004] Low-voltage distribution network topology undergoes frequent dynamic changes. Factors such as user load randomness, line faults (such as short circuits and disconnections), and communication delays can all lead to topology reconstruction, making traditional methods difficult to infer in real time. Furthermore, device measurement errors, missing data, and noise further exacerbate the challenge of model robustness. To address these issues, existing research has attempted to integrate various techniques (such as the CNN-LSTM joint model and the Graph Attention Network (GAT)), but these models have limited generalization capabilities and are highly dependent on specific grid data. This paper proposes a dynamic topology recognition method based on LSTM-GCN. By combining time series feature extraction with graph structure modeling, this method not only captures the dynamic changes of nodes in the temporal dimension but also explores spatial dependencies, significantly improving recognition accuracy and stability. Experiments demonstrate that this method demonstrates excellent performance in both an IEEE test system and a low-voltage distribution network simulation model. It particularly demonstrates significant advantages in dynamic topology recognition for user node addition and removal, providing new insights for intelligent management and fault diagnosis of low-voltage distribution networks. Summary of the Invention

[0005] The purpose of the present invention is to provide a low-voltage distribution network node topology identification and user node dynamic adjustment method, which aims to solve the problem.

[0006] To solve the above technical problems, the present invention provides a method for identifying node topology in a low-voltage distribution network and dynamically adjusting user nodes, comprising the following steps:

[0007] S1. Collect time series power flow data of nodes in the low-voltage distribution network. The data includes node voltage, current and power, expressed as:

[0008]

[0009] Where T is the time step, N is the number of nodes, and F is the feature dimension;

[0010] S2. Use the long short-term memory network to extract the time dependency of node features. The hidden state h of the long short-term memory network t , the hidden state matrix of the output time step is:

[0011]

[0012] S3: Input the node features output by the long short-term memory network into the graph convolutional network to aggregate the node features and generate the node feature matrix H. LSTM , and reconstruct the adjacency matrix A * , long short-term memory network part:

[0013] H GCN =σ(A·H LSTM W GCN )

[0014] in, is the learning parameter of the long short-term memory network layer, F′ is the output feature dimension, and the predicted adjacency matrix is:

[0015] A * =σ·H GCN

[0016] Here σ is the sigmoid function, which determines the probability of the existence of an edge in the adjacency matrix;

[0017] S4. Calculate the binary cross entropy loss Loss between the reconstructed adjacency matrix and the original adjacency matrix. The formula is:

[0018]

[0019] Final judgment accuracy Acc:

[0020]

[0021] S5. According to the reconstructed adjacency matrix A * Dynamically adjust the connection status of user nodes based on changes in node flow data, including the addition or removal of user nodes.

[0022] Further preferably, in step S1, the normalization processing of the node flow data adopts the formula:

[0023]

[0024] When the length is long or short, the voltage U, current I, and power P are normalized to eliminate the dimension difference.

[0025] Further preferably, in step S2, the long short-term memory network module is connected to the memory module by a forget gate. Input Gate Candidate memory state and output gate The joint operation generates the hidden state of each time step The specific formula is:

[0026]

[0027] Further preferably, in step S3, the node feature aggregation formula of the graph convolutional network is:

[0028]

[0029] is the feature representation of node v in the lth layer, N(v) represents the set of neighbor nodes of node v, d v is the degree of node v, W (l)is the weight matrix of the lth layer, and σ is the activation function of the layer.

[0030] Further preferably, in step S5, the public node N c.n The voltage of the remaining user nodes n user Voltage change:

[0031]

[0032] The power of the public node becomes:

[0033]

[0034] Calculate the power change of public nodes:

[0035]

[0036] According to the power change Determine whether the user node is cut off or put into use.

[0037] Further preferably, in step S1, the node power flow data is collected by a smart meter, a synchronized phasor measurement unit, or a distribution automation system, with a sampling interval of 15 minutes to 1 hour.

[0038] Further preferably, in step S2, the time step T of the long short-term memory network is processed by a sliding window method to enhance the local dependency of the time series data.

[0039] The present invention is further configured to have the following beneficial effects compared with the prior art:

[0040] First, the LSTM-GCN model demonstrates high accuracy in identifying distribution network topology and demonstrates superior robustness compared to other deep learning methods. By inputting time-series flow data from distribution network nodes, the LSTM captures the temporal changes in the data and, combined with the GCN to learn the connections between nodes, the model accurately reconstructs the adjacency matrix and identifies the network's topology.

[0041] Secondly, the method proposed in the present invention utilizes the time series characteristics of the distribution network node flow combined with the graph structure data of its node composition, which provides a new idea for the low-voltage distribution network topology identification.

[0042] Third, the proposed method demonstrates application potential in identifying complex distribution network topologies of different voltage levels and scales, and also has good effects on the topological structure changes of low-voltage distribution networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is the LSTM structure diagram;

[0044] Figure 2 This is the GCN structure diagram;

[0045] Figure 3 It is a structural diagram combining LSTM module and GCN module;

[0046] Figure 4 This is a schematic diagram of the LVDN tree structure;

[0047] Figure 5 is the adjacency matrix of the undirected graph;

[0048] Figure 6 It is a simplified equivalent circuit and voltage drop phasor diagram;

[0049] Figure 7 This is the LSTM-GCN topology recognition flow chart;

[0050] Figure 8 This is a simplified model diagram of the IEEE 33-bus system;

[0051] Figure 9 is the accuracy of the IEEE-bus33 system under different machine learning methods;

[0052] Figure 10 This is the comparison of the accuracy of deep learning models under different noise levels;

[0053] Figure 11 This is the comparison of the F1 value of the deep learning model under different noise levels;

[0054] Figure 12 It is a network diagram of the simulation system;

[0055] Figure 13 It is a schematic diagram of the topological structure changes of user node input and removal;

[0056] Figure 14 It is the power variation diagram of the public node. DETAILED DESCRIPTION

[0057] The following is a further detailed description of a low-voltage power distribution network node topology identification and user node dynamic adjustment method proposed by the present invention in conjunction with the accompanying drawings and specific embodiments. According to the following description, the advantages and features of the present invention will become clearer. It should be noted that the drawings are all in a very simplified form and are not in precise proportions, and are only used to conveniently and clearly assist in explaining the purpose of the embodiments of the present invention. The same or similar reference numerals in the drawings represent the same or similar components.

[0058] Example

[0059] 1 Research Methods and Process

[0060] 1.1 LSTM Model and GCN Structure

[0061] Long Short-Term Memory (LSTM) is a special type of Recurrent Neural Network (RNN) that is designed to solve the gradient vanishing and gradient exploding problems encountered by traditional RNN when processing long sequence data. Figure 1 LSTM has memory cells (cell states) and gating mechanisms (gates), which enable it to capture long-term dependencies and maintain memory for a long time. LSTM cells include the following core components: forget gate, input gate, candidate cell state, cell state, output gate, and hidden state.

[0062] Voltage and current data in distribution networks fluctuate due to factors such as load changes and equipment failures. LSTM can predict voltage and current fluctuations by learning from historical operating data and identifying patterns within it. When processing time-dependent data such as load forecasts, voltage fluctuations, and current changes in distribution networks, data from distribution networks is typically sequential and contains complex patterns and long-term dependencies. LSTM can effectively capture these patterns, enabling more accurate predictions and optimization.

[0063] GCN (Graph Convolutional Network) is a deep learning model for graph data, especially suitable for processing data with graph structure. Its structure is as follows Figure 2 As shown in the figure, unlike traditional convolutional neural networks (CNNs), GCNs mainly perform convolution operations on graph-structured data to learn the relationships between nodes and the structural information of the graph. The core idea of ​​GCN is to perform convolution operations on the adjacency matrix of the graph, rather than directly applying it to grid-shaped data (such as images) like CNN. The goal of GCN is to learn node representations and the overall structure of the graph by aggregating node neighbor information.

[0064] Graph theory G = {V, E, A}, where V represents the node set and E represents the edge set. V represents the node set consisting of the input nodes, and E represents the matrix consisting of the edges between each node, where the elements can be expressed as e ij , represents the edge between node i and node j. i is an element in the node set, v j is an element of another node. If there is a connection between i and j, then the element a in the corresponding adjacency matrix A (where A has a dimension of n×n) is ijThe value of is 1 or 0. Each node v ij Has a characteristic vector h v , the representation (or feature) of each node is updated through graph convolution operation. Its operation can be expressed as:

[0065]

[0066] is the feature representation of node v in the lth layer, N(v) represents the set of neighbor nodes of node v, d v is the degree of node v, W (l) is the weight matrix of the lth layer, and σ is the activation function of the layer. GCN can aggregate the node representations of the entire graph into a graph-level representation, which is then used for graph classification tasks. In node classification tasks, GCN can be used for node label prediction. The GCN used in this article is for link prediction. The goal of the link prediction task is to predict the possible edges in the graph. Through GCN, the model can learn the potential associations in the graph structure and predict missing links. Mapping to the distribution network corresponds to the physical connection relationship between the corresponding nodes and the mutual influence between the nodes.

[0067] 1.2 Fusion Model Construction and Methodology

[0068] Figure 3 In the study of the topological structure of the distribution network for the LSTM-GCN model framework, the graph theory method is used to study G = {V, E, A}, where V represents the node set consisting of input nodes, E represents the matrix composed of edges between each node, and the elements can be expressed as e ij , represents the edge between node i and node j. If v i is an element in the node set, v j is an element of another node. If there is a connection between i and j, then the element a in the corresponding adjacency matrix A (where A has a dimension of n×n) is ij The value is 1 or 0. In the complete distribution network system, the operation of the node is in dynamic stability, and the flow operation data of the node is time-varying. The time series flow operation data of each node in the distribution network is extracted as input, which can be expressed as X(X∈R T×N×F ). Where T represents the time step, N represents the number of nodes, and F represents the feature dimension. This paper chooses to extract the voltage, current, and power of the node as the data sample input.

[0069]

[0070] In the LSTM module, the time series data of the input node is divided into the node time series data X corresponding to all time scales according to the input time step T.i (1,2,3,…,n∈N), that is, the time series flow data of all nodes corresponding to time i, V i,j The voltage data of node j at time i is represented by LSTM, and the rest of the data is similar. The long-term dependencies of the time series data are extracted through LSTM, and the feature representation is generated, and then the hidden state of the last time step is output.

[0071]

[0072] For time series data processing, input data X and perform basic operations in the LSTM module: forget gate, input gate, and output gate. The simplified calculation can be expressed as:

[0073]

[0074] in is the output of the forget gate, The output of the input gate, is a candidate memory state, is the memory cell state, The output of the output gate, The hidden state at each time step, W and b correspond to the weight and bias terms in each step respectively, The hidden state of the previous time step. Finally, the output of LSTM is:

[0075]

[0076] In the GCN module, the node features and the original adjacency matrix output by LSTM are input to learn the relationship between nodes and reconstruct the adjacency matrix. GCN part:

[0077]

[0078] in, is the learning parameter of the GCN layer, F′ is the output feature dimension, and the predicted adjacency matrix is:

[0079] A * =σ·H GCN (12)

[0080] Here σ is the sigmoid function, which determines the probability of the existence of an edge in the adjacency matrix. The loss function BCE is:

[0081]

[0082] The final judgment accuracy,

[0083]

[0084] 2. Research on the identification of low-voltage distribution network topology

[0085] 2.1 Overview and framework of low voltage distribution station area

[0086] LVDN is the part of the power system that directly connects to the user and is also the last link in the distribution network. Usually, LVDN is a radial structure, from the transformer down to the user. The topological structure analysis adopts a tree structure consisting of three parts: root, stem, and leaf. Figure 4 shown.

[0087] 2.2 Research on the changes of node topology in low-voltage distribution network

[0088] In the distribution network, due to switch actions (active actions of the distribution network), line faults, node load shedding, etc., the distribution network topology structure will change, or node data will be lost. From this moment on, the topology structure between the nodes of the distribution network changes. At this time, the flow data measured in the data acquisition device of the node is inconsistent with the flow operation rules when the topology structure has not changed. By comparing the node flow data before and after the two moments, the node adjacency matrix after the change is inferred, that is, the predicted adjacency matrix. Assume that each time scale corresponds to an adjacency matrix A, in which the element a ij There are only 0 and 1, 0 means there is no connection between node i and node j, and 1 means there is a connection. Figure 5 Represents the graph structure composed of nodes and its adjacency matrix.

[0089] When analyzing the circuit relationship in the distribution network branch, the most intuitive change is the voltage drop at both ends of the branch. Figure 6 To simplify the equivalent circuit and voltage drop phasor diagram. In a series branch, the current at both ends is the same, and the voltage drop expression is as follows:

[0090]

[0091] Assume that the voltage between the two ends is known (i.e. corresponding ), the current in the branch It can be expressed by the starting voltage and the starting power (or the ending voltage and the ending power):

[0092]

[0093] ΔU1 is the longitudinal component of the voltage drop, and δU1 is the transverse component of the voltage drop. These two indicators are used to describe the voltage changes at the beginning and end of a series branch. When there are multiple branches, the node admittance matrix Y obtained from the original topology is b , represents the admittance between two nodes. According to the corresponding mutual admittance values ​​between the nodes, the original adjacency matrix between the nodes can be obtained:

[0094]

[0095] y ij is the node's self-admittance, y ij (i≠j) represents the admittance between node i and node j. When y ij =0, it means that the mutual admittance of the two nodes is 0, that is, there is no direct connection between the two nodes.

[0096] Large-scale distribution networks typically have a wide distribution area, a large number of nodes, a relatively uniform load distribution, and minimal overall load fluctuations, resulting in dynamically stable system operation. Changes in topology are typically due to proactive changes such as system switching, distribution scheduling, and demand response, or to changes caused by faults such as large-scale equipment failures and line disconnections. These topology changes are easy to detect and prevent in medium- and high-voltage distribution networks. Low-voltage distribution substations encompass a relatively smaller area than medium- and high-voltage distribution networks and typically employ a radial or tree-like structure. Topology changes within low-voltage distribution substations are mostly caused by the addition and removal of user nodes. The electricity load of ordinary residential users fluctuates over time, so topology changes are relatively frequent and more dependent on changes in user demand.

[0097] In the low-voltage distribution area, the branch boxes, meter boxes and users are all equivalent to nodes, and the connection relationship can be expressed by the original admittance matrix of the node, as follows:

[0098]

[0099] Due to the special structure of the low-voltage matching area, the original admittance matrix and the original adjacency matrix composed of nodes are usually sparse matrices, so the elements y ij (i≠j) is mostly 0. When a user is removed at a certain moment, the node admittance value associated with it becomes 0.

[0100]

[0101] The topological connection relationship of the low-voltage distribution station area can be equivalent to many public nodes connected to multiple user nodes. The original admittance matrix of the station area is split according to the connection relationship of the user nodes. The public nodes are N c.n , the voltage is The current is The public node connects to m user nodes n user1 , n user2 ,…,n usermSince there are very few direct connections between users in the low-voltage distribution area, such connection relationships are ignored. After splitting according to each public node, the admittance matrix composed of one public node and its connected users can be expressed as Y broken :

[0102]

[0103] When the user node connected to the public node is cut off, the current Y broken All the values ​​related to the removed nodes in the , become 0, and the other user flows in this part will also change. Before the user node is removed, the public node N c.n The power change is determined by the complex power To calculate. Assume that at a certain time t, user node k (1≤k≤m) is removed from the public node, then the admittance matrix changes:

[0104] ΔY=Y broken -Y k (twenty four)

[0105] The shape is:

[0106]

[0107] In the low-voltage distribution area, the removal of a single-phase user will not change the current of other users on the public node, so ΔI i =0(i≠k). Public node N c.n The voltage of the remaining user nodes n user Voltage change:

[0108]

[0109] The power of the public node becomes:

[0110]

[0111] Active power change:

[0112]

[0113] because so That is, the active power change value of the public node can be equivalent to the contribution of the removed user.

[0114] In the low-voltage area, the impedance Z(R+jX) of the transmission line connected to the user side is small, and the admittance Y(G+jB) is large. Therefore, after the user is cut off, The change is minimal, and the impact on other users connected to the public node is negligible. Ultimately, the most noticeable change is in the power flow data at the public node corresponding to the removed user. In a low-voltage distribution network, users with different phase sequences are considered relatively independent, and the switching on and off of users with different phase sequences does not affect each other. Research on the topology of low-voltage distribution networks, and the issues of user switching on and off, primarily focuses on low-voltage users connected to a public node and with the same phase sequence.

[0115] 2.3LSTM-GCN Identification of Low-Voltage Distribution Network Topology Structure Process ( Figure 7 )

[0116] 3. Data Acquisition and Experimental Design

[0117] 3.1 Acquisition of distribution network node flow data

[0118] In any distribution network, obtaining the flow data (such as voltage, power, etc.) of the distribution network nodes requires calculation based on the power flow equation of the power system. If two sets of specific data are determined, including the node injected active power, node injected reactive power, node voltage amplitude, and voltage phase angle difference between adjacent nodes, the current state of the distribution network can be determined based on the power flow equation. The active power of each node is the relationship between the electric energy injected by the node and the current flowing through the node. For a distribution network node i, its active power P i and reactive power Q i It can be expressed by the following formula:

[0119]

[0120] Where V i and V j are the voltage amplitudes of nodes i and j, θ i and θ j is the phase angle difference between node i and node j, G ij and B ij are the real and imaginary parts of the admittance between node i and node j, respectively.

[0121] 3.2 Data collection and preprocessing

[0122] The experimental data in this paper is based on the IEEE 33-node standard model ( Figure 8 The node operating flow data in the model is recorded, with a sampling time of 24 hours and a sampling interval of 15 minutes. The voltage, current, and power time series data for each node, along with the original adjacency matrix of the distribution network, are obtained as input data. When obtaining node flow data in the distribution network, the sampling time and sampling interval can be adjusted accordingly, as the scale and complexity of different distribution networks vary.

[0123] In actual distribution networks, the methods for obtaining node data usually involve the following. Smart meters (SmartMeters), installed on the user side or at key nodes of the distribution network, can measure and record voltage, current, power and other data in real time. High-frequency data acquisition and remote transmission are the basis for analyzing the real-time topology of the distribution network. Distribution automation systems (DAS), mostly used in medium and high voltage distribution systems, obtain current, voltage, power and other data of each node in real time through online monitoring equipment (such as sensors, data acquisition terminals, etc.). The system can collect, process, analyze and feed back data to the central control system in real time. Synchronous phasor measurement units (PMUs) can measure voltage, phase, frequency and other data of each node in the power grid in real time. The measurement data obtained by these types of devices are highly accurate and timely, which can greatly reduce the experimental errors caused by the data.

[0124] The network structure in the distribution network is complex, with a large number of nodes, and the node flow data obtained is huge. In addition, there is also the problem of mutual influence between adjacent nodes in the distribution network. The voltage of adjacent nodes will vary due to factors such as load changes, line length and voltage regulation. Generally, the voltage change of adjacent nodes is continuous, but under different loads or operating conditions, the voltage difference may become larger. The voltage per unit value difference of such nodes is small. If it is not processed and used directly in the model, it will affect the model performance and thus affect the accuracy of topology identification. Therefore, it is necessary to standardize and normalize the measured data:

[0125]

[0126] When the number of nodes in the distribution network is large, the node flow data is more abundant and the dimension of the original adjacency matrix is ​​larger. A sliding window method can be considered to improve the time dependence of the data. To reduce the computational complexity, a sparse matrix can be used to store the original adjacency matrix.

[0127] 3.3 Experimental content

[0128] The recognition accuracy of the model for different distribution networks was verified using multiple IEEE test systems. The recognition results of each model are shown in Table 1:

[0129] Table 1 IEEE-bus system topology recognition accuracy

[0130]

[0131] From the above table data, we can see that the LSTM-GCN model proposed in this paper has a good recognition effect on the topological structure of nodes in the distribution network, and has a high accuracy rate in topological structure recognition for distribution networks of various types and complex structures. In order to verify the advantages of this model in identifying the topological structure of distribution networks, other deep learning models are selected for comparison. The recognition results of each model are as follows Figure 9 :

[0132] To verify the robustness of the proposed model, we added noise of different proportions to five models: LSTM-GCN, GRU-GCN, Transformer, TGNet, and GAT, and analyzed the model robustness and F1 value change trends. Figure 10 is the comparison result of the accuracy of deep learning models under different noise levels. Figure 11 The following are comparisons of the F1 scores of deep learning models under different noise levels. These two figures show that the LSTM-GCN method proposed in this paper can maintain an accuracy rate of over 98% in topological structure recognition even in the presence of noise, while being less affected by noise and exhibiting superior robustness compared to other deep learning methods.

[0133] The simulation system proposed in the book Intelligent Identification of Low Voltage Distribution Network Topology

[31] , as a research model of low-voltage distribution substation, there are 9 distribution nodes and 13 user nodes in the system. The system network structure is as follows Figure 12 As shown in Table 2, the user distribution is as follows:

[0134] Table 2 is the system user distribution and simplified topology diagram (corresponding to Figure 12 ) in the corresponding label.

[0135] Table 2 System user distribution

[0136]

[0137] The simulation runs for one day with a sampling interval of 15 minutes, during which the power of the user nodes varies according to a typical residential load pattern. The model takes the user node traffic data as input and generates three sets of traffic data by configuring the node insertion and removal operations. Figure 13 As shown, topology identification confirms the insertion of nodes M73 (Phase C) and M82 (Phase A) and the removal of nodes M41 (Phase C) and M71 (Phase A), demonstrating the effectiveness of the 16-model in detecting topology changes caused by single-phase user operations. Figure 14 The power changes at the common node are shown. Disconnecting single-phase user nodes reduces the load and lowers the total power, while connecting them increases the load and correspondingly increases the power. The magnitude of these changes is closely related to the load power of the user nodes added or removed.

[0138] The above description is only a description of the preferred embodiments of the present invention and does not limit the scope of the present invention. Any changes and modifications made by ordinary technicians in the field of the present invention based on the above disclosure shall fall within the scope of protection of the claims.

Claims

1. A method for identifying node topology in a low-voltage power distribution network and dynamically adjusting user nodes, characterized in that: The following steps are involved: S1. Collect time series power flow data of nodes in the low-voltage distribution network. The data includes node voltage, current and power, expressed as: Where T is the time step, N is the number of nodes, and F is the feature dimension; S2. Use the long short-term memory network to extract the time dependency of node features. The hidden state h of the long short-term memory network t , the hidden state matrix of the output time step is: S3: Input the node features output by the long short-term memory network into the graph convolutional network to aggregate the node features and generate the node feature matrix H. LSTM , and reconstruct the adjacency matrix A * , long short-term memory network part: H GCN =σ(A·H LSTM ·W GCN ) in, is the learning parameter of the long short-term memory network layer, F′ is the output feature dimension, and the predicted adjacency matrix is: A * =σ·H GCN Here σ is the sigmoid function, which determines the probability of the existence of an edge in the adjacency matrix; S4. Calculate the binary cross entropy loss Loss between the reconstructed adjacency matrix and the original adjacency matrix. The formula is: Final judgment accuracy Acc: S5. According to the reconstructed adjacency matrix A * Dynamically adjust the connection status of user nodes based on changes in node flow data, including the addition or removal of user nodes.

2. A method for identifying node topology of a low-voltage power distribution network and dynamically adjusting user nodes according to claim 1, characterized in that: In step S1, the node power flow data is standardized using the formula: When the length is long or short, the voltage U, current I, and power P are normalized to eliminate the dimension difference.

3. A method for identifying node topology of a low-voltage power distribution network and dynamically adjusting user nodes according to claim 1, characterized in that: In step S2, the long short-term memory network module passes the forget gate f t i , input gate Candidate memory state and output gate The joint operation generates the hidden state of each time step The specific formula is:

4. A method for identifying low-voltage power distribution network node topology and dynamically adjusting user nodes according to claim 1, characterized in that: In step S3, the node feature aggregation formula of the graph convolutional network is: is the feature representation of node v in the lth layer, N(v) represents the set of neighbor nodes of node v, d v is the degree of node v, W (l) is the weight matrix of the lth layer, and σ is the activation function of the layer.

5. A method for identifying node topology of a low-voltage power distribution network and dynamically adjusting user nodes according to claim 1, characterized in that: In step S5, the public node N c.n The voltage of the remaining user nodes n user Voltage change: The power of the public node becomes: Calculate the power change of public nodes: According to the power change Determine whether the user node is cut off or put into use.

6. A method for identifying node topology of a low-voltage power distribution network and dynamically adjusting user nodes according to claim 1, characterized in that: In step S1, node power flow data is collected through smart meters, synchronized phasor measurement units, or distribution automation systems, with a sampling interval of 15 minutes to 1 hour.

7. A method for identifying low-voltage power distribution network node topology and dynamically adjusting user nodes according to claim 1, characterized in that: In step S2, the time step T of the long short-term memory network is processed by a sliding window method to enhance the local dependency of the time series data.

Images