A Power Grid Area Division Method and System Based on Dynamic Graph Neural Network
By constructing a dynamic graph of the power grid using a dynamic graph neural network, extracting hidden state data of nodes, analyzing node importance and local density, and dividing the power grid into regions, the problem that the power grid topology segmentation method cannot respond to dynamic changes in real time is solved, and real-time fault isolation and load balancing of the power grid are realized.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- STATE GRID ZHEJIANG ELECTRIC POWER CO LTD HANGZHOU POWER SUPPLY CO
- Filing Date
- 2026-05-07
- Publication Date
- 2026-06-02
AI Technical Summary
In existing technologies, power grid topology splitting methods are based on static data and cannot respond to dynamic changes in the power grid in real time, resulting in the failure of the splitting results and the inability to achieve fault isolation, load balancing and optimized control.
A power grid region partitioning method based on dynamic graph neural network is adopted. By acquiring static topology data and dynamic measurement data, a dynamic graph of the power grid is constructed. The hidden state data of nodes is extracted by dynamic graph neural network, the importance and local density of nodes are analyzed, and clustering and region partitioning are performed to realize real-time dynamic management of power grid topology.
It significantly improves the real-time adaptability of the power grid topology splitting strategy and the accuracy of the splitting results, enhances fault isolation efficiency and load balancing, and solves the problems of insufficient splitting accuracy and low topology matching degree in traditional methods.
Smart Images
Figure CN122132861A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power grid management technology, and in particular to a method and system for dividing power grid areas based on dynamic graph neural networks. Background Technology
[0002] With the rapid development of smart grids, the scale of power grids is constantly expanding and their topology is becoming increasingly complex, resulting in highly dynamic power grid operation. Power grid topology partitioning, as a key technology to ensure the safe and stable operation of power grids, aims to divide complex power grids into several relatively independent sub-regions to achieve functions such as fault isolation, load balancing, and optimized control.
[0003] In existing technologies, power grid topology partitioning is usually based on offline static topology data and traditional clustering algorithms to partition the grid. For example, the partitioning boundary is determined by calculating the betweenness centrality of nodes or the line load rate. However, these methods ignore the real-time changes in the power grid's operating status, causing the partitioning results to fail when the power grid is dynamically adjusted, and thus failing to achieve the intended functions such as fault isolation, load balancing, and optimized control. Summary of the Invention
[0004] This invention provides a method and system for dividing power grid regions based on dynamic graph neural networks, in order to solve the technical problem of how to optimize the division of power grid topology and realize real-time dynamic management of power grid topology.
[0005] To address the aforementioned technical problems, embodiments of the present invention provide a power grid area division method based on a dynamic graph neural network, comprising: In response to a power grid partitioning trigger signal, static topology data and dynamic measurement data of the target power grid are acquired, and a dynamic power grid diagram is constructed based on the static topology data and the dynamic measurement data. The power grid dynamic graph is input into a pre-constructed dynamic graph neural network for feature extraction to obtain the node hidden state data of the target power grid; The feature extraction is designed to extract the state sequence data of each node in the power grid dynamic diagram, perform feature association processing on the obtained static features and dynamic time-series features, and obtain node hidden state data that integrates time-series dependencies and topological connection relationships. The importance of each node in the target power grid is analyzed based on the node hidden state data, and each node in the target power grid is screened based on the importance to generate a set of key nodes; Calculate the local density of each node in the set of key nodes, and perform a secondary screening of the set of key nodes based on the local density of nodes to construct a set of candidate nodes; Based on the candidate node set, clustering processing is performed on each node in the target power grid to obtain the clustering partitioning result of the target power grid; The target power grid is divided into regions based on the clustering results to obtain the region division results of the target power grid.
[0006] As one preferred embodiment, the static topology data includes the rated voltage, rated power, and rated line capacity of each device in the target power grid; the dynamic measurement data includes switch status and line load rate. The construction of the power grid dynamic diagram based on the static topology data and dynamic measurement data includes: The static topology data is parsed to construct an initial power grid topology; wherein the parsing process includes establishing an initial association between nodes and edges, using equipment in the target power grid as nodes and transmission lines as edges; The connection relationships between nodes in the initial power grid topology are updated in real time based on the switch status, and the weight values of edges in the initial power grid topology are updated in real time based on the line load rate, so as to obtain a dynamic graph of the power grid.
[0007] As one preferred embodiment, the step of inputting the dynamic graph of the power grid into a pre-constructed dynamic graph neural network for feature extraction to obtain the node hidden state data of the target power grid includes: Extract historical state sequence data of nodes in the power grid dynamic diagram, wherein the historical state sequence data includes static sequence data and dynamic sequence data; Feature extraction is performed on the static sequence data and the dynamic sequence data respectively to obtain a time-series feature vector containing static features and dynamic time-series features; The static features and the dynamic time-series features are correlated to obtain a feature correlation matrix; The feature correlation matrix is subjected to graph structure propagation processing to obtain the node hidden state data.
[0008] As one preferred embodiment, the step of analyzing the importance of each node in the target power grid based on the node hidden state data includes: The influence relationship between nodes is analyzed based on the hidden state data of the nodes to obtain the influence weight matrix between nodes; The hidden state data of the nodes are integrated with the influence weight matrix between the nodes, and the result of the integration is mapped to a preset interval to obtain an importance score that reflects the importance of each node in the target power grid.
[0009] As one preferred embodiment, calculating the local node density of each node in the set of key nodes includes: The neighborhood radius is dynamically adjusted based on the obtained load rate of the target power grid to construct an adaptive density calculation space; The similarity between nodes in the key node set is measured by a multidimensional feature distance metric, and a weighted density matrix is generated based on the similarity. Density values are calculated and optimized based on the adaptive density calculation space and the weighted density matrix to obtain the local density values of the nodes.
[0010] As one preferred embodiment, the step of clustering each node in the target power grid based on the candidate node set to obtain the clustering result of the target power grid includes: The candidate node set is processed using the farthest point iteration method to obtain a number of initial cluster centers; Clustering is performed on the set of key nodes based on the initial cluster centers to obtain the clustering results of the target power grid.
[0011] As one preferred embodiment, the process of processing the candidate node set based on the farthest point iteration method yields a number of initial cluster centers, including: The node with the highest local density in the candidate node set is determined as the first cluster center; Iteratively calculate the minimum distance between the candidate node set and the first cluster center, and take the node with the largest minimum distance as the next cluster center, until the preset number of clusters is met; The obtained cluster centers are verified based on the topological constraints of the target power grid, and a number of initial cluster centers are generated based on the verification results.
[0012] As one preferred embodiment, the step of clustering the key node set based on the initial cluster centers to obtain the clustering results includes: During the clustering process, the Euclidean distance between each node in the key node set and each initial cluster center is calculated. Based on the Euclidean distance, each node is assigned to the nearest initial cluster center to generate a first cluster. The cluster center is then updated to the feature mean of each node in the first cluster. The clustering process is iterated up to a preset number of times to obtain the initial clustering result; Calculate the deviation rate between the total load and the average load of each cluster in the initial clustering partitioning result. If the deviation rate exceeds a preset deviation threshold, adjust the assignment of the boundary nodes in the initial clustering partitioning result to obtain the clustering partitioning result.
[0013] As one preferred embodiment, the step of dividing the target power grid into regions based on the clustering results to obtain the region division results of the target power grid includes: Extract the connecting edges from the clustering results, and filter the connecting edges whose edge weights exceed a preset threshold as candidate splitting edges; The dynamic graph of the power grid is divided based on the candidate cutting edges to obtain the regional division result of the target power grid.
[0014] Another embodiment of the present invention provides a power grid area division system based on a dynamic graph neural network, comprising: The acquisition module is used to acquire static topology data and dynamic measurement data of the target power grid in response to the power grid partitioning trigger signal, and to construct a dynamic power grid map based on the static topology data and dynamic measurement data. The extraction module is used to input the dynamic graph of the power grid into a pre-constructed dynamic graph neural network for feature extraction to obtain the node hidden state data of the target power grid; The feature extraction is designed to extract the state sequence data of each node in the power grid dynamic diagram, perform feature association processing on the obtained static features and dynamic time-series features, and obtain node hidden state data that integrates time-series dependencies and topological connection relationships. The first filtering module is used to analyze the importance of each node in the target power grid based on the node hidden state data, and to filter each node in the target power grid based on the importance to generate a set of key nodes; The second filtering module is used to calculate the local density of each node in the key node set, and to perform a second filtering on the key node set based on the local density of the nodes to construct a candidate node set. The clustering module is used to perform clustering processing on each node in the target power grid based on the candidate node set, so as to obtain the clustering partitioning result of the target power grid; The partitioning module is used to partition the target power grid into regions based on the clustering partitioning results, thereby obtaining the region partitioning results of the target power grid.
[0015] Compared with the prior art, the beneficial effects of the embodiments of the present invention are at least one of the following: 1) This invention integrates static power grid topology and dynamic measurement data through dynamic graph neural network to construct a graph model that evolves dynamically over time. Combined with the ability of gated cyclic unit to capture the temporal dependence of node state, the power grid topology partitioning strategy can respond in real time to dynamic operating scenarios such as load fluctuations and switch state changes, significantly improving the adaptability to the real-time state of the power grid and overcoming the shortcomings of traditional static partitioning methods in dealing with dynamic changes in the power grid. 2) Based on the clustering optimization mechanism of node importance score and adaptive density calculation, this invention can accurately identify key nodes and topological community structure of the power grid. With the improved clustering initialization and load balancing constraints, the segmentation results are more in line with the actual operation requirements of the power grid, effectively improving the load balancing and fault isolation efficiency after segmentation, and solving the problems of insufficient segmentation accuracy and low topology matching degree in traditional methods. Attached Figure Description
[0016] Figure 1 This is a flowchart illustrating a power grid area division method based on a dynamic graph neural network in one embodiment of the present invention. Figure 2 This is a schematic diagram of a power grid area division system based on a dynamic graph neural network in one embodiment of the present invention; Figure label: Among them, 11 is the acquisition module; 12 is the extraction module; 13 is the first filtering module; 14 is the second filtering module; 15 is the clustering module; and 16 is the partitioning module. Detailed Implementation
[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0018] In the description of this invention, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.
[0019] In the description of this invention, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.
[0020] One embodiment of the present invention provides a method for dividing power grid areas based on a dynamic graph neural network. For details, please refer to [link to relevant documentation]. Figure 1 , Figure 1 The diagram shown is a flowchart of a power grid area division method based on a dynamic graph neural network according to one embodiment of the present invention, which includes steps S1-S6: S1: In response to the grid partitioning trigger signal, acquire the static topology data and dynamic measurement data of the target grid, and construct a grid dynamic diagram based on the static topology data and dynamic measurement data; Preferably, in one embodiment of the present invention, the static topology data includes the rated voltage, rated power, and rated line capacity of each device in the target power grid; the dynamic measurement data includes switch status and line load rate. A dynamic power grid diagram is constructed based on static topology data and dynamic measurement data, including: The static topology data is parsed and processed to construct the initial power grid topology; the parsing process includes establishing the initial association between nodes and edges, with equipment in the target power grid as nodes and transmission lines as edges; The connection relationships between nodes in the initial power grid topology are updated in real time based on the switch status, and the weight values of edges in the initial power grid topology are updated in real time based on the line load rate to obtain the dynamic graph of the power grid.
[0021] Specifically, this embodiment uses IEC61970 CIM / E format files as static topology data input, and extracts power grid equipment parameters and connection relationships using analytical tools. Specifically, physical equipment such as substations, generators, and loads are abstracted as nodes V in a graph model, and transmission lines are abstracted as edges E; the node feature matrix... This includes static parameters such as rated voltage and rated power, and dynamic measurement parameters such as real-time voltage. Edge sets represent the connection relationships between devices, represented by an adjacency matrix. This indicates that if the node and Direct connection via line =1, otherwise 0.
[0022] When establishing the initial association between nodes and edges, the principle of device line mapping is followed: each node corresponds to a unique device, and each edge corresponds to a power transmission line; node attributes are synchronously mapped to rated parameters; when edge attributes are initialized, only the physical connection state is retained, and dynamic operating parameters are not included.
[0023] Real-time acquisition of switch status data, and dynamic adjustment of the adjacency matrix according to the following rules. : in, This refers to the current moment.
[0024] The real-time load rate of the line is used as the edge weight, and the calculation formula is as follows: in, For line load power, The weight is the line's rated capacity, and its value ranges from [0, 1]. near When the load factor approaches 1, it indicates that the line load is close to the capacity limit, and it is treated as a critical edge in the dynamic graph.
[0025] A data synchronization cycle is set, and the data is updated based on the latest switch status and load rate after each sampling. Edge weights are coupled with the adjacency matrix to form a weighted adjacency matrix, which controls the information transmission strength between nodes. Static data is imported in batches via the CIM / E file interface, while dynamic data is synchronized in real time via the IEC61850 protocol. A data verification mechanism is established, which triggers a redundancy verification process when the switch status is inconsistent with the adjacency matrix.
[0026] When the line load rate exceeds the threshold, the edge weight update frequency is automatically increased to strengthen the monitoring of heavily loaded lines; when a sudden change in switch status is detected, the adjacency matrix is immediately refreshed and the corresponding edge is marked as faulty, the weight is temporarily adjusted to 0 and the splitting strategy is recalculated.
[0027] Understandably, traditional static graph models rely on fixed adjacency matrices, which cannot reflect dynamic scenarios such as switch state transitions or line overload. This invention, by updating the adjacency matrix and edge weights in real time, controls the error between the graph model and the physical power grid state within one sampling period, improving dynamic adaptability by more than 40%.
[0028] Existing methods mostly consider only topological connectivity and do not incorporate operational parameters such as load rate into edge attributes. This invention quantifies line operational pressure through edge weights, enabling the dynamic graph to simultaneously include topological structure and operational status, providing a more comprehensive decision-making basis for subsequent segmentation strategies.
[0029] Furthermore, this invention employs an incremental update mechanism, only locally updating the changing node connections and edge weights, rather than reconstructing the entire graph model. Testing shows that in a 1000-node power grid, a single dynamic update takes less than 5ms, far shorter than the full reconstruction time of traditional methods, thus meeting the real-time control requirements of the power grid.
[0030] S2: Input the dynamic graph of the power grid into a pre-constructed dynamic graph neural network for feature extraction to obtain the hidden state data of the nodes of the target power grid; Among them, feature extraction is designed to extract the state sequence data of each node in the dynamic diagram of the power grid, perform feature association processing on the obtained static features and dynamic time-series features, and obtain the hidden state data of the nodes that integrates the time-series dependency relationship and the topological connection relationship. Preferably, in one embodiment of the present invention, the dynamic graph of the power grid is input into a pre-constructed dynamic graph neural network for feature extraction to obtain the node hidden state data of the target power grid, including: Extract historical state sequence data of nodes in the power grid dynamic diagram, where the historical state sequence data includes static sequence data and dynamic sequence data; Feature extraction is performed on static sequence data and dynamic sequence data respectively to obtain a time series feature vector containing static features and dynamic time series features; The static features and dynamic time-series features are correlated to obtain the feature correlation matrix; The feature correlation matrix is processed by graph structure propagation to obtain the hidden state data of the nodes.
[0031] Specifically, in this embodiment, the static attribute time series of nodes are extracted from the power grid dynamic diagram, including: rated parameters: equipment design parameters such as rated voltage and rated power, which remain constant over time; equipment type labels: the equipment type to which the node belongs (such as generator or bus), forming a type encoding vector. Real-time acquisition of dynamic parameter sequences reflecting the power grid operating status is also performed, including: electrical quantities: real-time measurements such as voltage and power; topology status: real-time operating parameters such as switch on / off status and line load rate; and rate of change: the time change rate of parameters such as power and voltage, capturing transient dynamic characteristics.
[0032] In this embodiment, before performing feature extraction, the architecture of the dynamic graph neural network based on GRU needs to be pre-constructed. Specifically, the network architecture adopts a series structure of GRU temporal feature extraction layer + graph convolutional network (GCN) topology feature propagation layer. The GRU layer is responsible for capturing the temporal dependency relationship of node states, and the GCN layer is responsible for fusing the topological connection relationship between nodes. The network input is the "node historical state sequence + dynamic adjacency matrix" of the power grid dynamic graph, and the output is the node hidden state data.
[0033] Based on the characteristics of the two types of sequence data, a differential extraction method is adopted to obtain static features that can characterize the inherent attributes of nodes and dynamic temporal features that can characterize the temporal change patterns.
[0034] Specifically, since the static sequence data has constant values over T time steps, we first take the static data from any time step and compress the static sequence matrix from N×T×D_s to N×D_s; then we perform Z-score normalization (to eliminate the dimensional differences of parameters such as rated voltage and rated power) to finally obtain the static feature matrix, where each row vector corresponds to the static feature of a single node.
[0035] A GRU layer is used to capture the temporal dependencies of dynamic sequences. First, the dynamic sequence matrix is split into N time steps T × dynamic feature dimension D_d samples by node, and input into a pre-trained GRU layer. The GRU learns the temporal patterns of the dynamic sequence (such as the upward trend of load rate and voltage fluctuation cycle) through a reset gate (controlling the degree of forgetting historical information) and an update gate (controlling the degree of integration of current information). Finally, the hidden state of the last time step of the GRU layer is taken as the dynamic temporal feature of the node to form a dynamic temporal feature matrix.
[0036] Furthermore, a sliding time window is used to store historical data, forming nodes. The state matrix; the historical state sequence is processed by a gated recurrent unit, and the temporal dependencies of the node states are captured by the output temporal hidden states, which are identified as: The static nominal parameters are concatenated with the temporal hidden state output by GRU to form a fused feature vector. The feature dimension difference is eliminated by the weight matrix to generate a feature association matrix of uniform scale.
[0037] Based on the power grid topology adjacency matrix and edge weights, feature propagation is achieved through GCN, and a weighted adjacency matrix is generated by combining switch states and load rates. After propagation through GCN, the feature matrices of each node are integrated into hidden state data, with each row corresponding to the spatiotemporal fusion features of the node.
[0038] Specifically, the normalized adjacency matrix and the optimized feature association matrix are input into the GCN layer. The GCN achieves graph structure propagation through neighbor feature aggregation, where the features of each node are integrated with the feature information of its neighboring nodes (e.g., a node will aggregate the features of connected high-load line nodes). Then, the learnable weight matrix and activation function of the GCN are used for nonlinear transformation to finally obtain the node hidden state data. Each row vector of the node hidden state data integrates the node's own static inherent attributes, dynamic temporal change patterns, and topological cooperative relationships with neighboring nodes.
[0039] S3: Analyze the importance of each node in the target power grid based on the node hidden state data, and filter each node in the target power grid based on the importance to generate a set of key nodes; Preferably, in one embodiment of the present invention, analyzing the importance of each node in the target power grid based on node hidden state data includes: An analysis of the influence relationships between nodes is performed on the hidden state data of the nodes to obtain the influence weight matrix between the nodes; The hidden state data of nodes are integrated with the influence weight matrix between nodes, and the result of the integration is mapped to a preset interval to obtain an importance score that reflects the importance of each node in the target power grid.
[0040] In this embodiment, a linear transformation is performed on the node hidden state matrix to generate a feature vector for influence relationship analysis. The influence correlation between nodes is analyzed, and effective correlations are screened based on the power grid topology. The correlation between nodes is normalized to generate an influence weight matrix, ensuring that the sum of the weights in each row is 1. The node hidden states and the influence weight matrix are weighted and aggregated, and a nonlinear transformation is used to map the scores to a preset interval, ultimately obtaining the node importance score vector, represented as: in, The model can learn parameters, and the higher the score, the more critical the node's impact on power grid stability.
[0041] The distribution of node importance scores in historical fault scenarios is statistically analyzed to determine an initial threshold; this threshold is then adaptively adjusted based on the current power grid operating status. (Filtering) Nodes exceeding a preset threshold are designated as nodes, forming a set of key nodes. Subsequent processing only applies to this set of key nodes to generate a standardized feature matrix. ,in, To preserve the number of nodes, For feature dimensions.
[0042] S4: Calculate the local density of each node in the key node set, and perform a secondary screening of the key node set based on the local density of the nodes to construct a candidate node set; Preferably, in one embodiment of the present invention, calculating the local node density of each node in the set of key nodes includes: The neighborhood radius is dynamically adjusted based on the obtained load rate of the target power grid to construct an adaptive density computation space; The similarity between nodes in the key node set is measured by multidimensional feature distance, and a weighted density matrix is generated based on the similarity. Density values are calculated and optimized based on the adaptive density computation space and the weighted density matrix to obtain the local density values of the nodes.
[0043] Specifically, the power grid load rate is divided into three levels: low (≤40%), medium (40%-70%), and high (>70%), with corresponding neighborhood radius scaling factors of 1.2, 1.0, and 0.8, respectively. The base radius is determined by the average node spacing of the power grid.
[0044] In this embodiment, dynamic features such as real-time node voltage, load rate, and power change rate are fused with static features such as rated voltage and equipment type to form a multi-dimensional feature vector. Euclidean distance is used to calculate the feature distance between nodes. A Gaussian kernel function is used to convert the distance into similarity weights to construct a weighted density matrix.
[0045] The local density is calculated using the radius method, specifically: in, For nodes and Euclidean distance, According to the formula Sort by density in descending order and select those with density values higher than a threshold. The nodes constitute the candidate set threshold The model uses a dynamic threshold and a gradient boosting regression tree to automatically adjust to fluctuations in grid load.
[0046] The specific training process is as follows: Using historical fault data thresholds as initial prediction values, calculate the negative gradient (i.e., pseudo-residual). Fit the residuals using a training regression tree to reduce the error of the current model. Then, multiply the new tree predictions by the learning rate and add them to the existing model. Use the above steps as iterative training, stopping once the preset number of trees (n_estimators) is reached.
[0047] S5: Based on the candidate node set, perform clustering processing on each node in the target power grid to obtain the clustering partitioning result of the target power grid; Preferably, in one embodiment of the present invention, clustering is performed on each node in the target power grid based on the candidate node set to obtain the clustering result of the target power grid, including: The candidate node set is processed using the farthest point iteration method to obtain a number of initial cluster centers; The key node set is clustered based on the initial cluster centers to obtain the clustering results of the target power grid.
[0048] Preferably, in one embodiment of the present invention, the candidate node set is processed based on the farthest point iteration method to obtain a number of initial cluster centers, including: The node with the highest local density in the candidate node set is determined as the first cluster center; Iteratively calculate the minimum distance between other nodes in the candidate node set that are not the first cluster center and the first cluster center, and take the node with the largest minimum distance as the next cluster center, until the preset number of clusters is met; The obtained cluster centers are verified based on the topological constraints of the target power grid, and a number of initial cluster centers are generated based on the verification results.
[0049] In this embodiment, the node with the highest density is selected from the candidate set as the first initial center. For each candidate node, the minimum distance from it to an existing center is calculated. Each time you choose The largest node is selected as the new center. The process stops when the number of selected centers reaches a preset value K or the minimum distance between all remaining nodes is less than a threshold, thus forming the initial center set. The specific iterative formula is as follows: in, For the candidate set of nodes, Candidate nodes With the selected center Euclidean distance, the first individual centers For candidate centers and selected centers The node with the largest minimum distance.
[0050] Preferably, in one embodiment of the present invention, clustering of the key node set based on the initial cluster centers to obtain clustering results includes: During the clustering process, the Euclidean distance between each node in the key node set and each initial cluster center is calculated. Based on the Euclidean distance, the nodes are assigned to the nearest initial cluster center to generate the first cluster. The cluster center is then updated to the feature mean of each node in the first cluster. The clustering process is iterated up to a preset number of times to obtain the initial clustering result; Calculate the deviation rate between the total load and the average load of each cluster in the initial clustering partitioning result. If the deviation rate exceeds the preset deviation threshold, adjust the assignment of the boundary nodes in the initial clustering partitioning result to obtain the clustering partitioning result.
[0051] Z-score standardization is applied to the multidimensional features of the key node set to eliminate the influence of dimensions. The standardized Euclidean distance from node i to the initial center is calculated, and the node is assigned to the nearest center to generate the initial cluster. After each iteration, the cluster center is updated to the feature mean of the nodes within the cluster. The iteration stops when the number of iterations reaches a preset value or the change in the center is less than a threshold, and the initial clustering result is obtained.
[0052] Specifically, for each cluster Node extraction Obtain the cluster The set of all nodes, where, , For clusters The number of nodes. This can be determined using the formula: Calculate the feature dimension of nodes within a cluster The average value below, where Indicates the first The node at the th The value on each feature.
[0053] Combine the means of each dimension to form a new centroid vector. Repeat the above steps until the centroid change is less than the threshold or the maximum number of iterations is reached. Since only the set of critical nodes is processed, the computational complexity is reduced from... Down to The number of iterations was reduced by 50%. The overall computation time was shortened from 30 minutes to 1 minute.
[0054] S6: Divide the target power grid into regions based on the clustering results to obtain the region division results of the target power grid.
[0055] Preferably, in one embodiment of the present invention, the target power grid is divided into regions based on the clustering results to obtain the region division results of the target power grid, including: Extract the connecting edges from the clustering results, and filter the connecting edges whose edge weights exceed a preset threshold as candidate splitting edges; The dynamic graph of the power grid is divided based on candidate cutting edges to obtain the regional division results of the target power grid.
[0056] In this embodiment, edges connecting different clusters are extracted, edge weights are calculated, and edges with weights higher than a threshold are selected as the splitting boundaries to form a set of splitting edges. Circuit breaker tripping commands are generated based on the set of splitting edges, prioritizing the disconnection of lines with high loads and low importance. After splitting, the power balance of each islanded area is verified. If a power deficit exceeds 10%, the splitting is reversed and new splitting edges are selected.
[0057] After collecting parameters such as voltage and frequency of each cluster after segmentation, if the deviation exceeds 5%, the incremental learning of the dynamic graph neural network is triggered. The DGNN parameters are updated through the elastic weight solidification algorithm to make the model adapt to the new topology.
[0058] Specifically, based on Elastic Weight Consolidation (EWC), important parameters are preserved through regularization, thus protecting old knowledge.
[0059] in, For Fisher's information matrix, To balance parameters, when the power grid topology changes (such as adding nodes), the input dimensions of the GCN are dynamically expanded, and the model parameters are updated batch by batch to adapt to the real-time data stream.
[0060] Another embodiment of the present invention provides a power grid area division system based on a dynamic graph neural network. For details, please refer to [link to relevant documentation]. Figure 2 , Figure 2The diagram shown illustrates a power grid area division system based on a dynamic graph neural network in one embodiment of the present invention, comprising: The acquisition module 11 is used to acquire static topology data and dynamic measurement data of the target power grid in response to the power grid partitioning trigger signal, and to construct a dynamic diagram of the power grid based on the static topology data and dynamic measurement data. Extraction module 12 is used to input the dynamic graph of the power grid into a pre-constructed dynamic graph neural network for feature extraction to obtain the node hidden state data of the target power grid; Among them, feature extraction is designed to extract the state sequence data of each node in the dynamic diagram of the power grid, perform feature association processing on the obtained static features and dynamic time-series features, and obtain the hidden state data of the nodes that integrates the time-series dependency relationship and the topological connection relationship. The first screening module 13 is used to analyze the importance of each node in the target power grid based on the node hidden state data, and to screen each node in the target power grid based on the importance to generate a set of key nodes. The second filtering module 14 is used to calculate the local density of each node in the key node set, and to perform secondary filtering on the key node set based on the local density of the nodes to construct a candidate node set. Clustering module 15 is used to perform clustering processing on each node in the target power grid based on the candidate node set to obtain the clustering partitioning result of the target power grid; The partitioning module 16 is used to partition the target power grid into regions based on the clustering results, and obtain the region partitioning results of the target power grid.
[0061] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.
Claims
1. A method for dividing power grid areas based on dynamic graph neural networks, characterized in that, include: In response to a power grid partitioning trigger signal, static topology data and dynamic measurement data of the target power grid are acquired, and a dynamic power grid diagram is constructed based on the static topology data and the dynamic measurement data. The power grid dynamic graph is input into a pre-constructed dynamic graph neural network for feature extraction to obtain the node hidden state data of the target power grid; The feature extraction is designed to extract the state sequence data of each node in the power grid dynamic diagram, perform feature association processing on the obtained static features and dynamic time-series features, and obtain node hidden state data that integrates time-series dependencies and topological connection relationships. The importance of each node in the target power grid is analyzed based on the node hidden state data, and each node in the target power grid is screened based on the importance to generate a set of key nodes; Calculate the local density of each node in the set of key nodes, and perform a secondary screening of the set of key nodes based on the local density of nodes to construct a set of candidate nodes; Based on the candidate node set, clustering processing is performed on each node in the target power grid to obtain the clustering partitioning result of the target power grid; The target power grid is divided into regions based on the clustering results to obtain the region division results of the target power grid.
2. The power grid area division method based on dynamic graph neural network as described in claim 1, characterized in that, The static topology data includes the rated voltage, rated power, and rated line capacity of each device in the target power grid. The dynamic measurement data includes switch status and line load rate; The construction of the power grid dynamic diagram based on the static topology data and the dynamic measurement data includes: The static topology data is parsed to construct an initial power grid topology; wherein the parsing process includes establishing an initial association between nodes and edges, using equipment in the target power grid as nodes and transmission lines as edges; The connection relationships between nodes in the initial power grid topology are updated in real time based on the switch status, and the weight values of edges in the initial power grid topology are updated in real time based on the line load rate, so as to obtain a dynamic graph of the power grid.
3. The power grid area division method based on dynamic graph neural network as described in claim 1, characterized in that, The step of inputting the dynamic graph of the power grid into a pre-constructed dynamic graph neural network for feature extraction to obtain the node hidden state data of the target power grid includes: Extract historical state sequence data of nodes in the power grid dynamic diagram, wherein the historical state sequence data includes static sequence data and dynamic sequence data; Feature extraction processing is performed on the static sequence data and the dynamic sequence data respectively to obtain a time-series feature vector containing static features and dynamic time-series features; The static features and the dynamic time-series features are correlated to obtain a feature correlation matrix; The feature correlation matrix is subjected to graph structure propagation processing to obtain the node hidden state data.
4. The power grid area division method based on dynamic graph neural network as described in claim 1, characterized in that, The step of analyzing the importance of each node in the target power grid based on the node hidden state data includes: The influence relationship between nodes is analyzed based on the hidden state data of the nodes to obtain the influence weight matrix between nodes; The hidden state data of the nodes are integrated with the influence weight matrix between the nodes, and the result of the integration is mapped to a preset interval to obtain an importance score that reflects the importance of each node in the target power grid.
5. The power grid area division method based on dynamic graph neural network as described in claim 1, characterized in that, The calculation of the local node density of each node in the set of key nodes includes: The neighborhood radius is dynamically adjusted based on the obtained load rate of the target power grid to construct an adaptive density calculation space; The similarity between nodes in the key node set is measured by a multidimensional feature distance metric, and a weighted density matrix is generated based on the similarity. Density values are calculated and optimized based on the adaptive density calculation space and the weighted density matrix to obtain the local density values of the nodes.
6. The power grid area division method based on dynamic graph neural network as described in claim 1, characterized in that, The process of clustering nodes in the target power grid based on the candidate node set to obtain the clustering result of the target power grid includes: The candidate node set is processed using the farthest point iteration method to obtain a number of initial cluster centers; Clustering is performed on the set of key nodes based on the initial cluster centers to obtain the clustering results of the target power grid.
7. The power grid area division method based on dynamic graph neural network as described in claim 6, characterized in that, The process of the candidate node set based on the farthest point iteration method yields a number of initial cluster centers, including: The node with the highest local density in the candidate node set is determined as the first cluster center; Iteratively calculate the minimum distance between the candidate node set and the first cluster center, and take the node with the largest minimum distance as the next cluster center, until the preset number of clusters is met; The obtained cluster centers are verified based on the topological constraints of the target power grid, and a number of initial cluster centers are generated based on the verification results.
8. The power grid area division method based on dynamic graph neural network as described in claim 6, characterized in that, The step of clustering the key node set based on the initial cluster centers to obtain the clustering results includes: During the clustering process, the Euclidean distance between each node in the key node set and each initial cluster center is calculated. Based on the Euclidean distance, each node is assigned to the nearest initial cluster center to generate a first cluster. The cluster center is then updated to the feature mean of each node in the first cluster. The clustering process is iterated up to a preset number of times to obtain the initial clustering result; Calculate the deviation rate between the total load and the average load of each cluster in the initial clustering partitioning result. If the deviation rate exceeds a preset deviation threshold, adjust the assignment of the boundary nodes in the initial clustering partitioning result to obtain the clustering partitioning result.
9. The power grid area division method based on dynamic graph neural network as described in claim 1, characterized in that, The step of dividing the target power grid into regions based on the clustering results to obtain the region division results of the target power grid includes: Extract the connecting edges from the clustering results, and filter the connecting edges whose edge weights exceed a preset threshold as candidate splitting edges; The dynamic graph of the power grid is divided based on the candidate cutting edges to obtain the regional division result of the target power grid.
10. A power grid area division system based on a dynamic graph neural network, characterized in that, include: The acquisition module is used to acquire static topology data and dynamic measurement data of the target power grid in response to the power grid partitioning trigger signal, and to construct a dynamic power grid map based on the static topology data and dynamic measurement data. The extraction module is used to input the dynamic graph of the power grid into a pre-constructed dynamic graph neural network for feature extraction to obtain the node hidden state data of the target power grid; The feature extraction is designed to extract the state sequence data of each node in the power grid dynamic diagram, perform feature association processing on the obtained static features and dynamic time-series features, and obtain node hidden state data that integrates time-series dependencies and topological connection relationships. The first filtering module is used to analyze the importance of each node in the target power grid based on the node hidden state data, and to filter each node in the target power grid based on the importance to generate a set of key nodes; The second filtering module is used to calculate the local density of each node in the key node set, and to perform a second filtering on the key node set based on the local density of the nodes to construct a candidate node set. The clustering module is used to perform clustering processing on each node in the target power grid based on the candidate node set, so as to obtain the clustering partitioning result of the target power grid; The partitioning module is used to partition the target power grid into regions based on the clustering partitioning results, thereby obtaining the region partitioning results of the target power grid.