Power grid correlation section identification method based on time sequence diagram convolutional network

Abstract

The invention discloses a power grid correlation section identification method based on a time sequence diagram convolutional network, and the method comprises the following steps: collecting the historical power flow data or disturbance response data of the operation of a power grid, and extracting the power change sequence of each key power transmission line; constructing an initial graph structure taking a power plant and a transformer substation as nodes and taking a power transmission line as an edge; introducing a time sequence diagram convolutional network to perform joint modeling on time sequence features and diagram structure features of line states, and learning correlation coefficients between lines; and judging and identifying the associated section through a threshold value. According to the method, the section affiliation of the line in the power grid can be fully mined, more accurate and dynamic section association identification is realized, and support is provided for scheduling decision, section division and cooperative control.

Description

Technical Field

[0001] This invention relates to the field of power system operation and dispatch, and in particular to a method for identifying power grid correlation sections based on time-series graph convolutional networks. Background Technology

[0002] Transmission sections are critical analytical objects in power system operation and dispatch, widely used in scenarios such as power flow distribution control, security constraint verification, and system stability assessment. In actual operation, complex power coupling and dynamic interactions may exist between different transmission sections. When a section experiences power limit exceedance or structural disturbance, other related sections may be affected simultaneously, even leading to risk contagion. Therefore, accurately identifying strongly correlated sections in the power grid is of great significance for improving the global awareness and risk prevention and control capabilities of power grid dispatch.

[0003] Currently, the identification of inter-section correlations mainly relies on traditional methods such as physical topology analysis, sensitivity calculation, and linear power flow models. These methods are often based on static model assumptions and are difficult to fully characterize the dynamic behavior changes of sections caused by factors such as load fluctuations, power generation fluctuations, and power source disturbances during actual operation. They also cannot effectively capture the implicit coupling relationships between sections that evolve over time. In addition, existing methods have insufficient generalization ability in high-dimensional operating data and complex disturbance environments, and lack the ability to jointly model the temporal and graph structure features of sections.

[0004] With the continuous improvement of the digitalization and automation level of the power grid, the spatiotemporal resolution of operation monitoring data has been significantly enhanced, providing a foundation for data-driven identification of complex relationships. Graph Neural Networks (GNNs), as a class of deep learning methods that have emerged in recent years, have been widely applied to tasks such as power grid topology analysis, state estimation, and risk identification. However, traditional GNN models are mainly designed for static graph structures and are difficult to directly model the dynamic evolution relationships between power grid sections.

[0005] Therefore, there is an urgent need for a new method that can integrate power grid topology information with the time-series characteristics of operational data to achieve dynamic and accurate identification of the correlation between transmission sections, so as to better support the safe operation and intelligent dispatch of the power grid. Summary of the Invention

[0006] The purpose of this invention is to provide a method for identifying power grid cross-section correlations based on temporal graph convolutional networks (TGCNs), primarily used for identifying and analyzing the correlations between cross-sections in a power grid dispatching system. This method innovatively proposes a power grid cross-section correlation identification technique based on temporal graph convolutional networks by collecting historical operating data and disturbance response data of the power grid and combining the advantages of TGCNs and temporal convolutional networks (TCNs). This technique can effectively reveal the dynamic correlations between different cross-sections in the power grid, thereby providing support for power grid safety dispatching, risk warning, and power flow optimization.

[0007] Technical Solution: To address the aforementioned technical problems, this invention proposes a method for identifying power grid associated sections based on temporal graph convolutional networks. This method includes the following steps:

[0008] S1 Power Grid Operation Data Acquisition and Processing: Collects historical power flow data of the power grid under different operating scenarios. The data includes the power flow of transmission lines and the voltage amplitude, frequency, and node injected power index of power grid nodes, and performs cleaning, standardization, and time-series reconstruction.

[0009] This method first collects historical power flow data or disturbance response data of the power grid under different operating scenarios. This data includes indicators such as power flow of transmission lines, voltage amplitude at grid nodes, frequency, and injected power. After cleaning, standardization, and temporal reconstruction, this raw data is formatted to suit a time-series graph convolutional network model.

[0010] S2 cross-sectional diagram structure construction

[0011] During the power grid graph structure construction phase, the system constructs a directed graph structure with transmission lines as the edges and power plants and substations as the nodes.

[0012] S3 model input data construction

[0013] For each graph node and edge in S2, the corresponding data in S1 is obtained. The graph node data includes voltage amplitude, frequency, and node injected power, while the edge data is power flow. The data of the graph nodes and edges are used to establish a data sequence within a continuous time window T, forming a time-series feature tensor that reflects dynamic behavior, which is used as the model input data.

[0014] S4 model output correlation coefficient matrix construction

[0015] Assume that the set of transmission lines represented by the edges in the power grid diagram structure is V = {v1, v2, ..., v...} D Line correlation weight w ij Describes line v i and v jThe coupling strength between them. The coupling strength is modeled using historical power flow data and power grid physical topology information, and the calculation formula is:

[0016] w ij =α·corr(P i ,P j )+β·dist(v i ,v j )

[0017] Among them, P i ,P j Represents the power time series of lines i and j, corr(P) i ,P j ) represents the linear correlation coefficient between line i and line j in the power flow time series, v i ,v j Dist(v) represents the edge between line i and line j in the power grid diagram structure. i ,v j ) represents the electrical distance between line i and line j in the power grid topology, and α and β are hyperparameters for adjusting the weights. w is calculated for each pair of all transmission lines. ij This yields the line correlation coefficient matrix.

[0018] S5 Model Construction and Training

[0019] In the model building phase, a recognition model based on Temporal Graph Convolutional Network (T-GCN) is used to fuse graph structure features and temporal state features. This model mainly includes the following modules:

[0020] Graph Structure Encoding Layer: This layer is responsible for sparsely encoding the adjacency matrix of the power grid cross-section graph, effectively inputting the graph structure into the model;

[0021] Temporal convolutional layer: This layer extracts local dynamic patterns by performing one-dimensional convolution on the time series of cross-sectional states. It captures the temporal dependencies of the cross-sectional states through convolution operations, enhancing the expressive power of temporal features.

[0022] Graph convolutional layer: Uses Chebyshev multinomials or GCN operators to perform feature aggregation, propagation and aggregation of graph features, and calculates the embedding representation of each node in the graph;

[0023] Output layer: Calculates the line correlation coefficient and outputs the line correlation coefficient matrix.

[0024] S6 Association Recognition and Output

[0025] The trained T-GCN model can generate a correlation coefficient matrix A between different lines in the power grid, where each element A ij Indicates line vi With v j The correlation coefficient between them. This matrix can be analyzed using threshold judgment to identify clusters of cross-sections with strong correlations. The method is to set a threshold of ε, if w ij If the correlation coefficient is greater than ε, then lines i and j are considered to belong to the same cross section. In actual operation data, this correlation coefficient matrix can be used to identify potential risk propagation paths or coupling relationships between cross sections in a timely manner, assisting dispatchers in optimizing the operation of the power grid.

[0026] To optimize the training process, the Adam optimizer was used to update the model parameters. The loss function includes cross-section state prediction error and correlation coefficient constraint terms. The loss function is defined as follows:

[0027] L = L state +λ·L relation

[0028] Among them, L state It is the cross-sectional condition prediction error, L relation This represents the constraint error of the cross-sectional correlation coefficient, and λ is the regularization parameter used to balance the influence of the two parts. Training employs a mini-batch mechanism, utilizing actual runtime data and simulation-generated disturbance response data for model training and optimization.

[0029] This method supports online deployment. Upon receiving a new round of operational data, it can update the data sequence of nodes and edges within a continuous time window T as model input in real time and dynamically output new cross-section association identification results. This feature enables the method to adapt to real-time changes in the power grid and provide timely decision support for dispatchers. The system of this invention can output a visualized cross-section association graph, using graph visualization tools to show the strength of coupling relationships between cross-sections, helping dispatchers determine potential risk propagation paths. This function can be applied to practical dispatching scenarios such as cross-section constraint construction, power flow adjustment scheme formulation, over-limit cross-section risk identification, and tie-line safety control. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention. Detailed Implementation

[0031] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0032] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0033] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0034] A regional power grid (let's call it the "X City Power Grid") consists of multiple transmission sections, substations, and transmission lines, involving several major power transmission lines, such as "Line A-1," "Line B-3," and "Line C-4." Within this power grid, dispatchers need to analyze the interrelationships between these sections to ensure that potential risk propagation paths and dynamic coupling relationships between sections can be identified promptly in the event of a fault. This embodiment employs the Temporal Graph Convolutional Network (T-GCN) method of this invention to identify and analyze the interrelationships of this power grid.

[0035] 1. Power Grid Operation Data Acquisition and Processing

[0036] In this embodiment, the power grid operation data comes from the dispatch center of the X City power grid. The power grid includes six main transmission lines (line "A-1" to "A-3"), involving multiple substations and power plants. The specific line names and information are as follows:

[0037] Line A-1: ​​Connecting substations "BS-1" and "BS-2", it is one of the main transmission lines of the power grid;

[0038] Line A-2: Connects substation "BS-2" and substation "BS-3", mainly responsible for the transmission of regional load;

[0039] Line A-3: Connects substations "BS-4" and "BS-5", serving as an emergency backup and load regulator in this power grid.

[0040] The data collected includes multiple parameters for each line (e.g., "A-1"), such as power flow, voltage, current, and frequency. The sampling frequency is once every 5 minutes, and the collected data includes:

[0041] Power flow data: power input and output for each line;

[0042] Voltage data: voltage amplitude and phase at each node;

[0043] Current data: Current variations in each line;

[0044] Frequency data: Overall frequency changes in the power grid.

[0045] These data, after being cleaned, standardized, and temporally reconstructed, are formatted for processing by time-series graph convolutional networks. To avoid excessive noise interference, incomplete or anomalous samples were removed during data cleaning, and the standardization process ensured that all data were dimensionally consistent. Temporal reconstruction reorganized the data into time windows (30 minutes in length), with each window containing metrics such as power flow, voltage, current, and frequency for that time period.

[0046] 2. Construction of power grid diagram structure

[0047] Construct a directed graph structure for the power grid, treating each transmission line as an edge in the graph.

[0048] 3. Model Input Data Construction

[0049] The state sequence of each grid node is constructed by collecting its power change sequence over a continuous time window. For example, for section D1, its power flow data over a continuous time window is represented as follows:

[0050]

[0051] Where F represents the state characteristic dimension of each cross section, such as power flow, voltage, current, etc.

[0052] The state data of all cross sections within each time window are formed into a temporal feature tensor, which is then fed into a temporal graph convolutional network for processing.

[0053] 4. Model Output Data Construction

[0054] For example, in this power grid, the coupling strength w between lines A-1 and A-2 12 It is determined by the following two factors: power flow correlation corr(P1,P2): measured by calculating the linear correlation coefficient of the power flow time series of the two lines in historical data;

[0055] Electrical distance dist(v1,v2): represents the sum of the impedances of the two lines along their paths in the power grid topology (for example, if the two lines are adjacent, the distance between them is short and the coupling strength is high).

[0056] Taking into account the effects of these two parts, the edge weight w 12 The calculation formula is: w ij =α·corr(P i ,P j )+β·dist(v i ,v j );

[0057] Here, α and β are adjustment coefficients used to balance the effects of the two parts.

[0058] 5. Model Building and Training

[0059] In this embodiment, a model based on a Temporal Graph Convolutional Network (T-GCN) is constructed. This model includes the following modules:

[0060] Graph structure coding layer: Sparsely encodes the adjacency matrix of the power grid to generate the structural features of the graph;

[0061] Temporal convolutional layer: Performs one-dimensional convolution on time-series data of cross-sectional state to extract local temporal dynamic patterns;

[0062] Graph convolutional layer: uses Chebyshev multinomials to propagate and aggregate graph features, and calculates the embedded representation of nodes;

[0063] Output layer: Calculates the similarity between node embedding vectors and generates a cross-sectional correlation coefficient matrix.

[0064] During training, historical data of the regional power grid were used to train the model. We employed the Adam optimizer to optimize the network parameters, and the loss function included constraints on cross-sectional state prediction error and correlation coefficient.

[0065] 6. Association Identification and Output

[0066] After the model training is completed, we obtain the power grid line correlation coefficient matrix A, where each element A... ij Indicates cross section v i With v j The correlation coefficient between them.

[0067] For example, the correlation coefficient matrix obtained by the calculation model identified a strong correlation between "Line A-1" and "Line A-2", indicating that they have strong dynamic coupling in the system, that is, they are on the same cross section.

[0068] The method of this invention supports online deployment. During the real-time operation of the power grid, when new operational data (such as load changes or equipment failures) enters the system, the model can automatically update the cross-sectional state sequence and dynamically output cross-sectional association identification results based on the new data. The method involves setting a threshold of 0.5; if w ij If the correlation coefficient is greater than 0.5, then lines i and j are considered to belong to the same cross section. For example, after a fault event of "Line A-3" occurs, the system will update the correlation coefficient matrix in real time and mark the cross section clusters related to the fault for dispatchers to refer to.

[0069] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for identifying power grid associated sections based on temporal graph convolutional networks, characterized in that, The method includes the following steps: S1 collects historical power flow data of the power grid under different operating scenarios. The data includes the power flow of transmission lines and the voltage amplitude, frequency, and node injected power index of the power grid nodes, and performs cleaning, standardization, and time-series reconstruction. S2 uses power plants and substations as nodes and transmission lines as edges to construct a power grid graph structure. For each graph node and edge in S2, S3 obtains the corresponding data in S1. The graph node data includes voltage amplitude, frequency, and node injected power, while the edge data is power flow. The data of the graph nodes and edges are used to establish a data sequence within a continuous time window T, forming a time-series feature tensor that reflects dynamic behavior, which is used as the model input data. Based on the historical data obtained by S1, S4 calculates the correlation index of the historical data of each transmission line and generates a line correlation coefficient matrix based on the correlation index. S5 constructs a recognition model based on Temporal Graph Convolutional Network (T-GCN), which integrates graph structure features and temporal state features, and uses the input and output data samples constructed in S3 and S4 to train the model; S6 applies the trained model to the actual operating system, obtains the data sequence of nodes and edges in the graph within a continuous time window T as the model input, generates the line correlation coefficient matrix, and obtains the section to which the line belongs by judging through the threshold.

2. The method for identifying power grid correlation sections based on temporal graph convolutional networks according to claim 1, characterized in that, In step S3, the time dimension of the time-series feature tensor has a length of T, the device dimension is the set of all topological nodes and edges, the graph node data includes voltage amplitude, frequency, node injected power, and the edge data is power flow.

3. The method for identifying power grid correlation sections based on temporal graph convolutional networks according to claim 1, characterized in that, In step S4, the correlation index is calculated using the following formula: w ij =α·corr(P i ,P j )+β·dist(v i ,v j ) Among them, P i ,P j Represents the power time series of lines i and j, corr(P) i ,P j ) represents the linear correlation coefficient between line i and line j in the power flow time series, v i ,v j Dist(v) represents the edge between line i and line j in the power grid diagram structure. i ,v j ) represents the electrical distance between line i and line j in the power grid topology, and α and β are hyperparameters for adjusting the weights. w is calculated for each pair of all transmission lines. ij Thus, the correlation coefficient matrix of the line is obtained.

4. The method for identifying power grid correlation sections based on temporal graph convolutional networks according to claim 1, characterized in that, In step S5, the graph convolutional network includes the following modules: The graph structure coding layer performs sparse coding on the adjacency matrix of the power grid topology graph; Temporal convolutional layers perform one-dimensional convolution processing on the input state time series to extract local dynamic patterns; Graph convolutional layers employ Chebyshev multinomials or GCN operators for feature aggregation. The output layer calculates the line correlation coefficients and outputs the line correlation coefficient matrix.

5. The method for identifying power grid correlation sections based on temporal graph convolutional networks according to claim 1, characterized in that, In step S6, the method for determining whether they belong to the same associated section based on a set threshold is as follows: The threshold is set to ε; if w ij If the value is greater than ε, then lines i and j are considered to belong to the same cross section.

6. The method for identifying power grid correlation sections based on temporal graph convolutional networks according to claim 1, characterized in that, The model training process uses the Adam optimizer, and the loss function includes constraints on cross-sectional state prediction error and correlation coefficient. The training adopts a Mini-batch mechanism.

7. The method for identifying power grid correlation sections based on temporal graph convolutional networks according to claim 1, characterized in that, The method also includes: outputting a visual cross-sectional correlation diagram, which uses a graph visualization tool to show the coupling relationship between each cross-section, in order to assist dispatchers in judging potential risk propagation paths.

Images