Substation voltage anomaly monitoring method based on Fourier transform and graph neural network

Abstract

The present invention discloses a method for monitoring substation voltage anomaly based on Fourier transform and graph neural network, which belongs to the field of voltage anomaly monitoring. This method is based on a voltage data prediction model, the input of which includes the voltage time series matrix of the substation in a certain time period and the graph network structure of the substation, and the output is the voltage data prediction vector of the substation at the next moment after the time period. The model includes a time-frequency feature extraction module, a mask matrix construction module, an adjacency matrix construction module and a graph neural network. The present invention explicitly describes the topological relationship of the substation voltage level by constructing a graph network structure, and calculates the correlation fusion matrix based on prior experience, and then filters the correlation fusion matrix based on the graph network structure and corresponding rules, so that the model can more accurately capture the voltage characteristics of the metering device itself and multiple devices, reduce noise interference, and improve the accuracy of anomaly monitoring.

Description

Technical Field

[0001] The present invention belongs to the field of voltage anomaly monitoring, and in particular relates to a substation voltage anomaly monitoring method based on Fourier transform and graph neural network. Background Art

[0002] During substation operation, transformers typically connect high, medium, and low voltage levels. Several metering devices (such as energy meters and voltage transformers) are deployed on each voltage level. The metering circuit can collect voltage time series data in real time, generating a multi-channel time series. When problems such as metering inaccuracies and transformer failures occur, voltage data will be the first to show abnormalities. Prompt detection and early warning can effectively ensure the stable operation of the power system. Therefore, rapid and accurate monitoring of voltage anomalies in metering devices is essential.

[0003] At present, there are mainly the following methods for monitoring substation voltage anomalies:

[0004] (1) Identification method based on the correlation of voltage time series data: This method uses the correlation characteristics of the voltage data of metering devices in the substation (the correlation of the same voltage level is higher than that of different voltage levels) to identify outlier metering devices through clustering, or uses expert rules to identify metering devices with abnormal correlation. The advantage of this method is that it is simple to implement, but it can only perform qualitative analysis and cannot quantify metering deviations.

[0005] (2) Discrimination method based on time series regression model: Using models such as ARIMA, LSTM, and Transformer, the prediction model is trained using historical normal data, and anomalies are identified by comparing the error distribution between the predicted value and the actual value. Although this method supports deviation quantification, it only relies on the historical data of a single device, ignoring the correlation between multiple devices, is easily interfered with in a high-noise environment, and has limited accuracy.

[0006] (3) Multi-device joint modeling method: The voltage data of all metering devices are input into the model, and the historical patterns of individual devices and the correlations between multiple devices are mined at the same time. Although this method uses historical data and considers the correlations between metering devices, due to the lack of prior knowledge, the model cannot capture effective features in the massive data, and the prediction effect is uncontrollable. Summary of the Invention

[0007] The present invention proposes a substation voltage anomaly monitoring method based on Fourier transform and graph neural network, the purpose of which is to solve the problem that the existing method has uncontrollable prediction effect due to the lack of intervention of prior experience.

[0008] The technical solutions of the present invention are as follows:

[0009] A method for monitoring substation voltage anomaly based on Fourier transform and graph neural network, comprising the following steps:

[0010] Step S1: constructing a voltage data prediction model;

[0011] The input of the voltage data prediction model includes the voltage time series matrix of the substation in a certain time period and the graph network structure of the substation. The voltage time series matrix includes the voltage data measurement values ​​of all metering devices in the substation in the time period, and the graph network structure defines the topological relationship of all metering devices.

[0012] The output of the voltage data prediction model is the voltage data prediction vector of the substation at the next moment after the time period, and the voltage data prediction vector includes the voltage data prediction value of each metering device;

[0013] The voltage data prediction model includes a time-frequency feature extraction module, a mask matrix construction module, an adjacency matrix construction module and a graph neural network;

[0014] Step S2: training a voltage data prediction model;

[0015] Step S3: Perform rolling prediction using the trained voltage data prediction model to obtain multiple continuous voltage data prediction vectors. Simultaneously, obtain voltage data measurement vectors corresponding to the voltage data prediction vectors obtained by the rolling prediction. The voltage data measurement vectors have the same shape as the voltage data prediction vectors and include the voltage data measurement values ​​of each metering device at the corresponding moment. Then, calculate the residual value between each set of voltage data prediction vectors and the voltage data measurement vectors.

[0016] Step S4: Count the residual values ​​and determine whether voltage anomaly occurs based on the statistical characteristics.

[0017] As a further improvement of the substation voltage anomaly monitoring method based on Fourier transform and graph neural network: the time-frequency feature extraction module is based on the input voltage time series matrix Get the voltage timing window matrix , then based on the voltage timing window matrix Get the time-frequency feature matrix ;

[0018] Voltage Timing Matrix , is the number of substation metering devices, voltage timing matrix The columns are arranged in chronological order, and each column corresponds to a sampling moment; the voltage timing matrix Middle Rank Elements of a column For substation The metering device is The voltage data measurement value at the time corresponding to the column;

[0019] The mask matrix construction module constructs a mask matrix based on the input graph network structure ;

[0020] The adjacency matrix construction module is based on the voltage timing matrix Use correlation weighted fusion operation to obtain the correlation fusion matrix , and then use the mask matrix Correlation Fusion Matrix Filter and then calculate the normalized adjacency matrix through the degree matrix ;

[0021] The graph neural network is based on the time-frequency feature matrix and the normalized adjacency matrix Get the voltage data prediction vector.

[0022] As a further improvement of the substation voltage anomaly monitoring method based on Fourier transform and graph neural network, the working process of the time-frequency feature extraction module is as follows:

[0023] Step A1: According to the preset window size, the voltage timing matrix Extract the last The voltage timing window matrix is ​​obtained by column ;

[0024] Step A2: Voltage timing window matrix Perform Fourier transform on each row of The amplitude of the frequency component;

[0025] For the The Fourier transform of the first The complex value of the frequency component is:

[0026] ;

[0027] The amplitude of the corresponding frequency component ;

[0028] Step A3: Calculate the voltage timing window matrix The mean and standard deviation of each row;

[0029] No. The row means are: ;

[0030] No. The standard deviation of the rows is: ;

[0031] Step A4: Based on the voltage timing window matrix Each row selected The amplitude of the frequency components, the mean and standard deviation of each row, and the final element values, and concatenate them to obtain the time-frequency feature matrix ; Time-frequency feature matrix No. OK .

[0032] As a further improvement of the substation voltage anomaly monitoring method based on Fourier transform and graph neural network, the working process of the mask matrix construction module is as follows:

[0033] Step B1: The graph network structure includes the physical connection relationships between buses of different voltage levels, transformers, and metering devices in the substation. The metering devices are regarded as nodes in the graph network structure, and edges are constructed between the nodes.

[0034] Step B2: Construct edge filtering rules and assign thresholds to each edge.

[0035] Filtering rules include:

[0036] Rule a: For a certain edge, if the two nodes connected by the edge are the busbar side metering device and the transformer side metering device of the same voltage level, then the threshold value is assigned to the edge. , threshold Greater than 0 and less than 1;

[0037] Rule b: For a certain edge, if the two nodes connected by the edge are bus-side metering devices of two different voltage levels, then the threshold value is assigned to the edge. ;

[0038] Rule c: For edges that do not meet rules a and b, assign a threshold to them , threshold greater than 1;

[0039] Step B3: Construct a mask matrix based on the threshold of each edge ; Mask matrix Middle Rank Elements of a column The value of nodes and The threshold for edges between nodes.

[0040] As a further improvement of the substation voltage anomaly monitoring method based on Fourier transform and graph neural network, the working process of the adjacency matrix construction module is as follows:

[0041] Step C1: Voltage timing matrix With a fixed number of columns Split in reverse chronological order, and we get submatrices, where the The sub-matrix is ​​recorded as ;

[0042] Step C2: For each submatrix, calculate the correlation coefficient between each row; In the sub-matrix Row and The correlation coefficient between rows is recorded as , then:

[0043] ;

[0044] in, Indicates the sub-matrix Rank The element value of the column, Indicates the In the sub-matrix The mean of the rows, Indicates the sub-matrix Rank The element value of the column, Indicates the In the sub-matrix The mean of the rows;

[0045] Step C3, assign weights to each sub-matrix; The weight of each sub-matrix is ​​recorded as ;

[0046] Step C4: Based on the correlation coefficient calculation results and weights of each sub-matrix, a correlation weighted fusion operation is used to obtain a correlation fusion matrix ;

[0047] Correlation Fusion Matrix Middle Rank The column element values ​​are:

[0048] ;

[0049] Step C5: Use the mask matrix Correlation Fusion Matrix Filter and obtain the filtered fusion matrix ;

[0050] The filtering process is as follows: For the correlation fusion matrix Middle Rank Elements of a column , if its value is greater than or equal to the corresponding mask matrix The element value in , then the corresponding filtered fusion matrix Middle Rank Elements of a column ,otherwise ;

[0051] Step C6: Filtered fusion matrix Add self-loops to get the initial adjacency matrix ; , is the identity matrix;

[0052] Step C7: Based on the initial adjacency matrix Calculate degree matrix ; Degree matrix is a diagonal matrix, all elements outside the diagonal are 0, and the elements on the diagonal are: ,in Degree matrix No. Rank Elements of the column, Represents the initial adjacency matrix No. Rank Elements of the column;

[0053] Step C8: Based on the initial adjacency matrix Sum degree matrix Get the normalized adjacency matrix : .

[0054] As a further improvement of the substation voltage anomaly monitoring method based on Fourier transform and graph neural network, the specific process of step C3 is as follows:

[0055] Step C3-1: Select the attenuation factor ;

[0056] Step C3-2, assign original weights to each sub-matrix; The original weights of the sub-matrices are ;

[0057] Step C3-3: Normalize the original weights of each submatrix to obtain the weights of each submatrix: .

[0058] As a further improvement of the substation voltage anomaly monitoring method based on Fourier transform and graph neural network: the graph neural network includes two graph convolution layers and one fully connected layer;

[0059] The processing process of the first graph convolution layer is:

[0060] ;

[0061] in, is the weight matrix of the first graph convolutional layer; , and To seek Parameters set when is a hidden dimension; is the activation function; The data matrix output by the first graph convolution layer;

[0062] The processing process of the second graph convolution layer is:

[0063] ;

[0064] in, is the weight matrix of the second graph convolutional layer, is a hidden dimension; is the activation function; The data matrix output by the second graph convolution layer;

[0065] The processing process of the fully connected layer is:

[0066] ;

[0067] in, is the weight matrix of the fully connected layer; , indicating Add the same value to each element of ; is the voltage data prediction vector, No. Elements Indicates the The predicted value of voltage data of each metering device.

[0068] As a further improvement of the substation voltage anomaly monitoring method based on Fourier transform and graph neural network, the specific steps of step S2 are:

[0069] Step S2-1: Obtain one or more original voltage time series matrices of the target substation ;in, , is the voltage timing window matrix The number of columns;

[0070] Step S2-2: From the original voltage timing matrix Extract multiple voltage timing matrices from As a training sample, each voltage timing matrix In the original voltage timing matrix The next column in is used as the label value of the training sample ; Tag value No. Elements Indicates the Voltage data measurement value of each metering device;

[0071] Step S2-3: Input the extracted training samples and the graph network structure of the target substation into the voltage data prediction model to obtain the corresponding voltage data prediction vector , and then predict the vector based on the voltage data and the corresponding label value Calculate the loss function, and then adjust the weight matrix of the graph neural network in the voltage data prediction model based on the loss function.

[0072] As a further improvement of the substation voltage anomaly monitoring method based on Fourier transform and graph neural network: in step S3, during rolling prediction, firstly based on the voltage time series matrix initially set in the target monitoring range Obtain the corresponding voltage data prediction vector, and then add the voltage data measurement values ​​of all metering devices at the next moment as a new column to the voltage time series matrix At the end, the new voltage timing matrix is ​​obtained , and then the next voltage data prediction vector is obtained, and so on, multiple consecutive voltage data prediction vectors are obtained.

[0073] As a further improvement of the substation voltage anomaly monitoring method based on Fourier transform and graph neural network: in step S4, it is assumed that The voltage data prediction vector and voltage data measurement vector are combined, then The voltage data prediction vector Predicted value of voltage data of each metering device and the corresponding voltage data measurement vector Voltage data measurement value of each metering device The error between ;

[0074] The mean of all error values ​​is:

[0075] ;

[0076] The standard deviation of all error values ​​is:

[0077] ;

[0078] If a satisfy , then the corresponding The metering device is There is a residual anomaly at the moment corresponding to the voltage data prediction vector;

[0079] If a metering device If the proportion of residual anomalies in the residual values ​​exceeds the preset proportion value, it is determined that the voltage of the metering device is abnormal; if the metering device is A metering device, which The abnormal deviation value at the time corresponding to the voltage data prediction vector is .

[0080] Compared with the prior art, the present invention has the following beneficial effects:

[0081] 1. The present invention explicitly describes the topological relationship of substation voltage levels by constructing a graph network structure, calculates edge weights (correlation fusion matrix) based on prior experience, and then filters the correlation fusion matrix based on the graph network structure and corresponding rules, so that the model can more accurately capture the voltage characteristics of the metering device itself and multiple devices, reduce noise interference, and improve the accuracy of abnormal monitoring.

[0082] 2. Based on the time series regression model architecture, the present invention introduces prior experience (substation topology and the corresponding voltage fluctuation correlation rules), which not only retains the quantitative analysis ability of the regression model, but also makes up for the defect of traditional regression methods that ignore the correlation between multiple devices. At the same time, it also makes the model more adaptable and can be applied to various business scenarios such as fault analysis and power replenishment. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 Schematic diagram of the architecture of the voltage data prediction model in the present invention;

[0084] Figure 2 Schematic diagram of the network structure in a specific implementation manner. DETAILED DESCRIPTION

[0085] The technical solution of the present invention will be described below with reference to the accompanying drawings. Obviously, the embodiments described are only a part of the embodiments of the present invention, rather than all the embodiments.

[0086] A method for monitoring substation voltage anomaly based on Fourier transform and graph neural network, comprising the following steps:

[0087] Step S1: construct a voltage data prediction model.

[0088] The input of the voltage data prediction model includes the voltage time series matrix of the substation in a certain time period and the graph network structure of the substation. The voltage time series matrix includes the voltage data measurement values ​​of all metering devices in the substation during the time period, and the graph network structure defines the topological relationship of all metering devices.

[0089] The output of the voltage data prediction model is the voltage data prediction vector of the substation at the next moment after the time period, and the voltage data prediction vector includes the voltage data prediction value of each metering device.

[0090] like Figure 1 The voltage data prediction model includes a time-frequency feature extraction module, a mask matrix construction module, an adjacency matrix construction module and a graph neural network. The time-frequency feature extraction module is based on the input voltage time series matrix Get the voltage timing window matrix , then based on the voltage timing window matrix Get the time-frequency feature matrix The mask matrix construction module constructs a mask matrix based on the input graph network structure. The adjacency matrix construction module is based on the voltage timing matrix Use correlation weighted fusion operation to obtain the correlation fusion matrix , and then use the mask matrix Correlation Fusion Matrix Filter and then calculate the normalized adjacency matrix through the degree matrix The graph neural network is based on the time-frequency feature matrix and the normalized adjacency matrix Get the voltage data prediction vector.

[0091] Among them, the voltage timing matrix , is the number of substation metering devices, voltage timing matrix The columns are arranged in chronological order, and each column corresponds to a sampling moment. Voltage timing matrix Middle Rank Elements of a column For substation The metering device is The voltage data measurement value at the time corresponding to the column.

[0092] The specific working principles of each module are as follows:

[0093] (1) Time-frequency feature extraction module.

[0094] Step A1: According to the preset window size, the voltage timing matrix Extract the last The voltage timing window matrix is ​​obtained by column .

[0095] In this embodiment, Take 192, that is, extract two days of voltage data (4 times per hour, 4*24*2=192) with a collection density of 15 minutes (the general cycle of energy collection in current substations).

[0096] Step A2: Voltage timing window matrix Perform Fourier transform on each row of The amplitude of a frequency component.

[0097] For the The Fourier transform of the first The complex value of the frequency component is:

[0098] ;

[0099] The amplitude of the corresponding frequency component .

[0100] In this embodiment, .

[0101] Step A3: Calculate the voltage timing window matrix The mean and standard deviation of each row.

[0102] No. The row means are: ;

[0103] No. The standard deviation of the rows is: .

[0104] Step A4: Based on the voltage timing window matrix Each row selected The amplitude of the frequency components, the mean and standard deviation of each row, and the final element values, and concatenate them to obtain the time-frequency feature matrix . Time-frequency feature matrix No. OK .

[0105] In this embodiment, .

[0106] (2) Mask matrix construction module.

[0107] Step B1: The graph network structure includes the physical connection relationship between the substation's different voltage level buses, transformers, and metering devices. The metering devices are regarded as nodes of the graph network structure, and then edges are constructed between the nodes.

[0108] Figure 2 This is a schematic diagram of the graph network structure. The purpose of introducing the graph network structure is to enable the model to learn the spatial relationship between nodes in graph convolution.

[0109] Step B2: Construct edge filtering rules and assign thresholds to each edge.

[0110] Rule a: For a certain edge, if the two nodes connected by the edge are the busbar side metering device and the transformer side metering device of the same voltage level, then the threshold value is assigned to the edge. Threshold Greater than 0 and less than 1.

[0111] like Figure 2 , the edges connecting Table 1 and Table 3, the edges connecting Table 2 and Table 3, the edges connecting Table 4 and Table 5, the edges connecting Table 4 and Table 6, and the edges connecting Table 4 and Table 7 are all assigned thresholds In this embodiment, the threshold .

[0112] Rule b: For a certain edge, if the two nodes connected by the edge are bus-side metering devices of two different voltage levels, then the threshold value is assigned to the edge. .

[0113] like Figure 2 , the edges connecting Table 1 and Table 5, the edges connecting Table 1 and Table 6, the edges connecting Table 1 and Table 7, the edges connecting Table 2 and Table 5, the edges connecting Table 2 and Table 6, and the edges connecting Table 2 and Table 7 are all assigned thresholds .

[0114] Rule c: For edges that do not meet rules a and b, assign a threshold to them Threshold Greater than 1.

[0115] Edges that meet rules a and b are likely to be retained (but not necessarily, as further judgment is required based on size comparison). However, edges that meet rule c will not be retained (filtering is achieved by setting a threshold greater than 1).

[0116] Step B3: Construct a mask matrix based on the threshold of each edge . Mask matrix Middle Rank Elements of a column The value of nodes and The threshold for edges between nodes.

[0117] (3) Adjacency matrix construction module.

[0118] Step C1: Voltage timing matrix With a fixed number of columns Split in reverse chronological order, and we get submatrices, where the The sub-matrix is ​​recorded as .

[0119] Note that since the division is performed in reverse time order, the first sub-matrix corresponds to the voltage timing matrix At the end of The sub-matrix corresponds to the voltage timing matrix The beginning of.

[0120] Preferably, If the number of columns of the last sub-matrix is ​​insufficient during the split , then extract the columns connected to the last submatrix from the penultimate submatrix to complete the matrix.

[0121] Step C2: For each submatrix, calculate the correlation coefficient between each row. In the sub-matrix Row and The correlation coefficient between rows is recorded as , then:

[0122] ;

[0123] in, Indicates the sub-matrix Rank The element value of the column, Indicates the In the sub-matrix The mean of the rows, Indicates the sub-matrix Rank The element value of the column, Indicates the In the sub-matrix The mean of the rows.

[0124] It should be noted that the calculation result of the correlation coefficient is generally -1 to +1, but in theory, negative correlation will not occur in this scenario.

[0125] Step C3: Assign weights to each sub-matrix. The weight of each sub-matrix is ​​recorded as .

[0126] The specific steps are:

[0127] Step C3-1: Select the attenuation factor .

[0128] In this embodiment, .

[0129] Step C3-2: Assign original weights to each sub-matrix. The original weights of the sub-matrices are .

[0130] Obviously, the later the submatrix is ​​in time, the larger the original weight is, and the original weight of the first submatrix is ​​the largest, which is 1.

[0131] Step C3-3: Normalize the original weights of each submatrix to obtain the weights of each submatrix: .

[0132] Step C4: Based on the correlation coefficient calculation results and weights of each sub-matrix, a correlation weighted fusion operation is used to obtain a correlation fusion matrix .

[0133] Correlation Fusion Matrix Middle Rank The column element values ​​are:

[0134] .

[0135] Step C5: Use the mask matrix Correlation Fusion Matrix Filter and obtain the filtered fusion matrix .

[0136] The filtering process is as follows: For the correlation fusion matrix Middle Rank Elements of a column , if its value is greater than or equal to the corresponding mask matrix The element value in , then the corresponding filtered fusion matrix Middle Rank Elements of a column ,otherwise .

[0137] because , is greater than 1, and It must be greater than 0 and less than 1. Therefore, the above filtering process is essentially to retain the correlation fusion matrix The element values ​​that meet rules a and b and are greater than or equal to 0.7 are set, and the other elements are set to 0.

[0138] Step C6: Filtered fusion matrix Add self-loops to get the initial adjacency matrix , so that each node can retain its own characteristics during convolution. , is the identity matrix.

[0139] Since there is no edge between the same nodes, the mask matrix All elements on the diagonal are , resulting in the filtered fusion matrix All diagonal elements are 0. By adding self-loops, the diagonal elements can be corrected to the correct value of 1.

[0140] Step C7: Based on the initial adjacency matrix Calculate degree matrix . Degree matrix is a diagonal matrix, all elements outside the diagonal are 0, and the elements on the diagonal are: ,in Degree matrix No. Rank Elements of the column, Represents the initial adjacency matrix No. Rank Elements of a column.

[0141] Step C8: Based on the initial adjacency matrix Sum degree matrix Get the normalized adjacency matrix : .

[0142] (4) Graph neural networks.

[0143] The graph neural network consists of two graph convolutional layers and one fully connected layer.

[0144] The processing process of the first graph convolution layer is:

[0145] ;

[0146] in, is the weight matrix of the first graph convolutional layer, To hide the dimension, in this embodiment . is the activation function, such as ReLU, Sigmoid, etc. The data matrix output by the first graph convolutional layer.

[0147] The processing process of the second graph convolution layer is:

[0148] ;

[0149] in, is the weight matrix of the second graph convolutional layer, To hide the dimension, in this embodiment . is the activation function. The data matrix output by the second graph convolutional layer.

[0150] The processing process of the fully connected layer is:

[0151] ;

[0152] in, is the weight matrix of the fully connected layer. , indicating Add the same value to each element of . is the voltage data prediction vector, No. Elements Indicates the The predicted value of voltage data of each metering device.

[0153] Step S2: training a voltage data prediction model.

[0154] The specific steps are:

[0155] Step S2-1: Obtain one or more original voltage time series matrices of the target substation . , is the voltage timing window matrix The number of columns.

[0156] Step S2-2: From the original voltage timing matrix Extract multiple voltage timing matrices from As a training sample, each voltage timing matrix In the original voltage timing matrix The next column in is used as the label value of the training sample Tag value No. Elements Indicates the The voltage data measurement value of each metering device.

[0157] Preferably, a sliding window is used to extract the original voltage timing matrix Extract multiple voltage timing matrices from , the starting point of the sliding window is the original voltage timing matrix The first column of the original voltage timing matrix is ​​the initial position of the end point. No. When sliding, the starting point of the sliding window remains stationary and the end point slides to the right. Each time a column is slid, the part covered by the current sliding window is used as a voltage time series matrix. Extract it.

[0158] Step S2-3: Input the extracted training samples and the graph network structure of the target substation into the voltage data prediction model to obtain the corresponding voltage data prediction vector , and then predict the vector based on the voltage data and the corresponding label value Calculate the loss function, and then adjust the weight matrix of the graph neural network in the voltage data prediction model based on the loss function.

[0159] In this embodiment, the loss function is: .

[0160] Through supervised training methods, using optimizers such as SGD, Adam, and RMSProp, a graph network regression model based on time-frequency data features can be trained.

[0161] Step S3: Perform rolling prediction using the trained voltage data prediction model to obtain multiple continuous voltage data prediction vectors. Simultaneously, obtain voltage data measurement vectors corresponding to the voltage data prediction vectors obtained from the rolling prediction. The voltage data measurement vectors have the same shape as the voltage data prediction vectors and contain the voltage data measurement values ​​of each metering device at the corresponding moment. Then, calculate the residual value between each set of voltage data prediction vectors and the voltage data measurement vectors.

[0162] During rolling forecasting, the voltage timing matrix initially set in the target monitoring range is first Obtain the corresponding voltage data prediction vector, and then add the voltage data measurement values ​​of all metering devices at the next moment as a new column to the voltage time series matrix At the end, the new voltage timing matrix is ​​obtained , and then the next voltage data prediction vector is obtained, and so on, multiple consecutive voltage data prediction vectors are obtained.

[0163] Step S4: Count the residual values ​​and determine whether voltage anomaly occurs based on the statistical characteristics.

[0164] Specifically, suppose we have The voltage data prediction vector and voltage data measurement vector are combined, then The voltage data prediction vector Predicted value of voltage data of each metering device and the corresponding voltage data measurement vector Voltage data measurement value of each metering device The error between .

[0165] The mean of all error values ​​is:

[0166] ;

[0167] The standard deviation of all error values ​​is:

[0168] .

[0169] If a satisfy , then the corresponding The metering device is There is a residual anomaly at the moment corresponding to the voltage data prediction vector. If the proportion of residual anomalies in the residual values ​​exceeds a preset proportion value (80% in this embodiment), it is determined that the voltage of the metering device is abnormal. A metering device, which The abnormal deviation value at the time corresponding to the voltage data prediction vector is .

[0170] This method not only identifies voltage anomalies but also quantitatively assesses abnormal deviations, enabling quantitative analysis. This method can be used for both metering device failure analysis and retroactive power calculation, demonstrating its high practical value in engineering applications.

[0171] It should be noted that it is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. The scope of the present invention is defined by the claims rather than the foregoing description.

Claims

1. A method for monitoring substation voltage anomaly based on Fourier transform and graph neural network, characterized in that the steps include: Step S1: constructing a voltage data prediction model; The input of the voltage data prediction model includes the voltage time series matrix of the substation in a certain time period and the graph network structure of the substation. The voltage time series matrix includes the voltage data measurement values ​​of all metering devices in the substation in the time period, and the graph network structure defines the topological relationship of all metering devices. The output of the voltage data prediction model is the voltage data prediction vector of the substation at the next moment after the time period, and the voltage data prediction vector includes the voltage data prediction value of each metering device; The voltage data prediction model includes a time-frequency feature extraction module, a mask matrix construction module, an adjacency matrix construction module and a graph neural network; The time-frequency feature extraction module is based on the input voltage time series matrix Get the voltage timing window matrix , then based on the voltage timing window matrix Get the time-frequency feature matrix ; Voltage Timing Matrix , is the number of substation metering devices, voltage timing matrix The columns are arranged in chronological order, and each column corresponds to a sampling moment; the voltage timing matrix Middle Rank Elements of a column For substation The metering device is The voltage data measurement value at the time corresponding to the column; The mask matrix construction module constructs a mask matrix based on the input graph network structure ; The adjacency matrix construction module is based on the voltage timing matrix Use correlation weighted fusion operation to obtain the correlation fusion matrix , and then use the mask matrix Correlation Fusion Matrix Filter and then calculate the normalized adjacency matrix through the degree matrix ; The filtering process is as follows: For the correlation fusion matrix Middle Rank Elements of a column , if its value is greater than or equal to the corresponding mask matrix The element value in , then the corresponding filtered fusion matrix Middle Rank Elements of a column ,otherwise ; The graph neural network is based on the time-frequency feature matrix and the normalized adjacency matrix Obtain voltage data prediction vector; The working process of the time-frequency feature extraction module is as follows: Step A1: According to the preset window size, the voltage timing matrix Extract the last The voltage timing window matrix is ​​obtained by column ; Step A2: Voltage timing window matrix Perform Fourier transform on each row of The amplitude of the frequency component; For the The Fourier transform of the first The complex value of the frequency component is: ； The amplitude of the corresponding frequency component ; Step A3: Calculate the voltage timing window matrix separately The mean and standard deviation of each row; No. The row means are: ; No. The standard deviation of the rows is: ; Step A4: Based on the voltage timing window matrix Each row selected The amplitude of the frequency components, the mean and standard deviation of each row, and the final element values, and concatenate them to obtain the time-frequency feature matrix ; Time-frequency feature matrix No. OK ; The working process of the mask matrix construction module is: Step B1: The graph network structure includes the physical connection relationships between buses of different voltage levels, transformers, and metering devices in the substation. The metering devices are regarded as nodes in the graph network structure, and edges are constructed between the nodes. Step B2: Construct edge filtering rules and assign thresholds to each edge. Filtering rules include: Rule a: For a certain edge, if the two nodes connected by the edge are the busbar side metering device and the transformer side metering device of the same voltage level, then the threshold value is assigned to the edge. , threshold Greater than 0 and less than 1; Rule b: For a certain edge, if the two nodes connected by the edge are bus-side metering devices of two different voltage levels, then the threshold value is assigned to the edge. ; Rule c: For edges that do not meet rules a and b, assign a threshold to them , threshold greater than 1; Step B3: Construct a mask matrix based on the threshold of each edge ; Mask matrix Middle Rank Elements of a column The value of nodes and The threshold for edges between nodes; Step S2: training a voltage data prediction model; Step S3: Perform rolling prediction using the trained voltage data prediction model to obtain multiple continuous voltage data prediction vectors. Simultaneously, obtain voltage data measurement vectors corresponding to the voltage data prediction vectors obtained by the rolling prediction. The voltage data measurement vectors have the same shape as the voltage data prediction vectors and include the voltage data measurement values ​​of each metering device at the corresponding moment. Then, calculate the residual value between each set of voltage data prediction vectors and the voltage data measurement vectors. Step S4: Count the residual values ​​and determine whether voltage anomaly occurs based on the statistical characteristics.

2. The method for monitoring substation voltage anomaly based on Fourier transform and graph neural network according to claim 1, characterized in that: The working process of the adjacency matrix construction module is: Step C1: Voltage timing matrix With a fixed number of columns Split in reverse chronological order, and we get submatrices, where the The sub-matrix is ​​recorded as ; Step C2: For each submatrix, calculate the correlation coefficient between each row; In the sub-matrix Row and The correlation coefficient between rows is recorded as , then: ； in, Indicates the sub-matrix Rank The element value of the column, Indicates the In the sub-matrix The mean of the rows, Indicates the sub-matrix Rank The element value of the column, Indicates the In the sub-matrix The mean of the rows; Step C3, assign weights to each sub-matrix; The weight of each sub-matrix is ​​recorded as ; Step C4: Based on the correlation coefficient calculation results and weights of each sub-matrix, a correlation weighted fusion operation is used to obtain a correlation fusion matrix ; Correlation Fusion Matrix Middle Rank The column element values ​​are: ； Step C5: Use the mask matrix Correlation Fusion Matrix Filter and obtain the filtered fusion matrix ; Step C6: Filtered fusion matrix Add self-loops to get the initial adjacency matrix ; , is the identity matrix; Step C7: Based on the initial adjacency matrix Calculate degree matrix ; Degree matrix is a diagonal matrix, all elements outside the diagonal are 0, and the elements on the diagonal are: ,in Degree matrix No. Rank Elements of the column, Represents the initial adjacency matrix No. Rank Elements of the column; Step C8: Based on the initial adjacency matrix Sum degree matrix Get the normalized adjacency matrix : .

3. The method for monitoring substation voltage anomaly based on Fourier transform and graph neural network according to claim 2, characterized in that: The specific process of step C3 is: Step C3-1: Select the attenuation factor ; Step C3-2, assign original weights to each sub-matrix; The original weights of the sub-matrices are ; Step C3-3: Normalize the original weights of each submatrix to obtain the weights of each submatrix: .

4. The method for monitoring substation voltage anomaly based on Fourier transform and graph neural network according to claim 1, characterized in that: The graph neural network includes two graph convolutional layers and one fully connected layer; The processing process of the first graph convolution layer is: ； in, is the weight matrix of the first graph convolutional layer; , and To seek Parameters set when is a hidden dimension; is the activation function; The data matrix output by the first graph convolution layer; The processing process of the second graph convolution layer is: ； in, is the weight matrix of the second graph convolutional layer, is a hidden dimension; is the activation function; The data matrix output by the second graph convolution layer; The processing process of the fully connected layer is: ； in, is the weight matrix of the fully connected layer; , indicating Add the same value to each element of ; is the voltage data prediction vector, No. Elements Indicates the The predicted value of voltage data of each metering device.

5. The method for monitoring substation voltage anomaly based on Fourier transform and graph neural network according to claim 1, characterized in that: The specific steps of step S2 are: Step S2-1: Obtain one or more original voltage time series matrices of the target substation ;in, , is the voltage timing window matrix The number of columns; Step S2-2: From the original voltage timing matrix Extract multiple voltage timing matrices from As a training sample, each voltage timing matrix In the original voltage timing matrix The next column in is used as the label value of the training sample ; Tag value No. Elements Indicates the Voltage data measurement value of each metering device; Step S2-3: Input the extracted training samples and the graph network structure of the target substation into the voltage data prediction model to obtain the corresponding voltage data prediction vector , and then predict the vector based on the voltage data and the corresponding label value Calculate the loss function, and then adjust the weight matrix of the graph neural network in the voltage data prediction model based on the loss function.

6. The method for monitoring substation voltage anomaly based on Fourier transform and graph neural network according to claim 1, characterized in that: In step S3, during rolling prediction, the voltage time series matrix initially set in the target monitoring range is first Obtain the corresponding voltage data prediction vector, and then add the voltage data measurement values ​​of all metering devices at the next moment as a new column to the voltage time series matrix At the end, the new voltage timing matrix is ​​obtained , and then the next voltage data prediction vector is obtained, and so on, multiple consecutive voltage data prediction vectors are obtained.

7. The method for monitoring substation voltage anomaly based on Fourier transform and graph neural network according to claim 6, characterized in that: In step S4, it is assumed that The voltage data prediction vector and voltage data measurement vector are combined, then The voltage data prediction vector Voltage data prediction value of each metering device and the corresponding voltage data measurement vector Voltage data measurement value of each metering device The error between ; The mean of all error values ​​is: ； The standard deviation of all error values ​​is: ； If a satisfy , then the corresponding The metering device is There is a residual anomaly at the moment corresponding to the voltage data prediction vector; If a metering device If the proportion of residual anomalies in the residual values ​​exceeds the preset proportion value, it is determined that the voltage of the metering device is abnormal; if the metering device is A metering device, which The abnormal deviation value at the time corresponding to the voltage data prediction vector is .

Images