A power grid safety and stability control system state monitoring and fault diagnosis method

Abstract

The application discloses a power grid safety and stability control system state monitoring and fault diagnosis method, and belongs to the technical field of power system automation. The method acquires voltage phasor data, current phasor data and frequency data collected by multiple synchronous phasor measurement units in a wide-area measurement system of a power grid, and constructs a time-space synchronous data matrix; based on an electromechanical transient time scale constraint of a power system, state topology reconstruction processing is performed on the time-space synchronous data matrix to generate a device-level operation situation feature map; a pre-trained time-space graph convolution network model is called to perform multi-dimensional feature extraction processing on the device-level operation situation feature map to obtain a power grid time-space operation state feature; and power grid operation state evaluation and fault diagnosis operations are performed according to the power grid time-space operation state feature. The method can more comprehensively retain device-level correlation characteristics in power grid transient behaviors, and improves the accuracy of fault type discrimination and fault positioning.

Description

Technical Field

[0001] This invention relates to the field of power system automation technology, specifically to a method for monitoring the status and diagnosing faults in a power grid safety and stability control system. Background Technology

[0002] The power grid safety and stability control system is a crucial line of defense for ensuring the reliable operation of the power system. Its condition monitoring and fault diagnosis capabilities directly affect the timeliness and accuracy of fault handling. The deployment of wide-area measurement systems enables the power grid to simultaneously acquire voltage phasor, current phasor, and frequency data from multiple nodes, providing a data foundation for real-time monitoring of system dynamics. Utilizing high-frequency sampling data from synchronous phasor measurement units to assess power grid operating conditions and diagnose faults has become a focus of attention in this field.

[0003] Existing power grid condition monitoring and fault diagnosis methods are mostly based on a model-driven approach. This involves constructing a power system transient simulation model and inputting real-time measurement data into the model to calculate transient stability margins or match preset fault modes. This method is highly dependent on the accuracy of the system model and has a heavy online computational burden. When faced with complex power grid topologies and multiple uncertainties, model errors are easily amplified, leading to deviations from actual conditions in the condition assessment. Another type of data-driven method directly extracts statistical or time-frequency features from synchronous phasor measurement data and then uses a classifier to achieve fault diagnosis. This type of method typically treats multi-node measurement information as isolated time-series signals, failing to fully utilize the spatial correlations inherent in the power grid topology and struggling to characterize the equipment-level dynamic interaction features during electromechanical transient processes. When the power grid experiences transient instability or complex faults, the spatiotemporal correlation patterns implicit in the wide-area measurement information cannot be fully explored, limiting the accuracy of fault component location and the refinement of condition assessment.

[0004] Therefore, it is necessary to solve the problems of how to reconstruct feature maps that can characterize the equipment-level operating status from synchronous phasor measurement data under electromechanical transient time scale, and how to effectively extract power grid operating status information from such feature maps containing complex spatiotemporal dependencies. Summary of the Invention

[0005] This invention aims to provide a method for monitoring the state and diagnosing faults in a power grid safety and stability control system. It reconstructs equipment-level operating status feature maps from synchronous phasor measurement data and extracts the spatiotemporal operating status features of the power grid using a spatiotemporal graph convolutional network. This addresses the problems of insufficient mining of the spatiotemporal correlation characteristics of the power grid and limited fault diagnosis accuracy in existing methods.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a method for state monitoring and fault diagnosis of a power grid safety and stability control system. This method includes the following steps: acquiring voltage phasor data, current phasor data, and frequency data collected by multiple synchronous phasor measurement units in a power grid wide-area measurement system; constructing a spatiotemporal synchronization data matrix based on the voltage phasor data, current phasor data, and frequency data; performing state topology reconstruction processing on the spatiotemporal synchronization data matrix based on the electromechanical transient time scale constraints of the power system to generate a device-level operating status feature map; calling a pre-trained spatiotemporal graph convolutional network model to perform multi-dimensional feature extraction processing on the device-level operating status feature map to obtain the spatiotemporal operating status features of the power grid; and performing power grid operating status assessment and fault diagnosis operations based on the spatiotemporal operating status features of the power grid. This method, by fusing wide-area measurement information with a deep learning model, can effectively capture the temporal and spatial coupling characteristics of the power grid dynamic process, improving the accuracy of state perception and fault identification.

[0007] As a technical solution of the present invention, the process of acquiring voltage phasor data, current phasor data, and frequency data collected by multiple synchronous phasor measurement units in a power grid wide-area measurement system includes: determining the device identifier and communication address of the synchronous phasor measurement units deployed at each node in the power grid; sending high-frequency sampling commands to each synchronous phasor measurement unit through a wide-area communication network; receiving voltage amplitude, voltage phase angle, current amplitude, current phase angle, and frequency measurement values ​​returned by each synchronous phasor measurement unit, which include time tags; and performing clock synchronization calibration on the voltage amplitude, voltage phase angle, current amplitude, current phase angle, and frequency measurement values ​​to generate voltage phasor data, current phasor data, and frequency data with a unified time reference. This solution utilizes high-frequency synchronous measurement data with a unified time reference across the entire network to provide a high-precision data foundation for subsequent spatiotemporal feature analysis.

[0008] Preferably, when constructing the spatiotemporal synchronized data matrix, the voltage phasor data, current phasor data, and frequency data are aligned along the time dimension according to time labels to generate continuous time series data; a sliding time window is set to perform frame processing on the continuous time series data, resulting in multiple time window data frames; the voltage phasor data, current phasor data, and frequency data in each time window data frame are spatially sorted according to the electrical distance between nodes in the power grid topology; the sorted data frames are stacked along the time dimension to generate a three-dimensional spatiotemporal synchronized data matrix. This processing method preserves the inherent temporal and spatial correlation of the measurement data, facilitating the extraction of spatiotemporal evolution patterns from subsequent models.

[0009] Furthermore, when generating the equipment-level operational status feature map, voltage phase angle data is extracted from the spatiotemporal synchronization data matrix, and the phase angle change rate of each node's voltage phase angle within a preset time window is calculated. Based on the phase angle change rate, the dominant oscillation mode in the electromechanical transient process of the power system is identified using time-domain signal processing technology. The power angle swing trajectory of each generator node is determined according to the oscillation frequency and damping ratio of the dominant oscillation mode. The power angle swing trajectory is subjected to phase space reconstruction processing to generate state trajectory mapping points for each generator node. The state trajectory mapping points of each generator node are spatially topologically associated with the voltage amplitude trajectory mapping points of the load nodes to construct an equipment-level operational status feature map containing node states and edge connection relationships. This feature map can intuitively represent the dynamic behavior and interaction relationships of key equipment in the electromechanical transient process, providing structured input for subsequent graph model processing.

[0010] In a preferred embodiment of the present invention, when using a pre-trained spatiotemporal graph convolutional network model for multi-dimensional feature extraction, the spatiotemporal graph convolutional network model includes an input layer, multiple spatiotemporal graph convolutional modules, and an output layer. The device-level operational status feature map is input into the input layer, and node feature normalization and edge weight initialization are performed on the feature map. Each spatiotemporal graph convolutional module performs spatial domain graph convolution on the feature map output from the previous layer to extract node neighborhood spatial features. Temporal domain causal convolution is then performed on the node neighborhood spatial features to extract time-dependent features. A gating mechanism dynamically fuses the spatial and temporal features output from each spatiotemporal graph convolutional module to generate an intermediate feature map. The intermediate feature map output from the last spatiotemporal graph convolutional module is input into the output layer and subjected to fully connected mapping to obtain the spatiotemporal operational status features of the power grid. This model can adaptively capture the spatial propagation characteristics and temporal evolution patterns under power grid topology constraints, and the fused features have stronger discriminative capabilities.

[0011] In the above scheme, when performing spatial domain graph convolution, the feature vector of each node and the adjacency matrix between nodes are obtained from the feature map output by the previous layer; the Laplacian matrix of the adjacency matrix is ​​calculated, and the Laplacian matrix is ​​eigenvalued and eigenvectord to obtain the eigenvalues ​​and eigenvectors of the Laplacian matrix; a graph Fourier transform basis is constructed based on the eigenvalues ​​and eigenvectors, and the feature vector of each node is mapped to the spectral domain through the graph Fourier transform basis; learnable filter kernel parameters are introduced into the spectral domain, and convolution is performed on the mapped spectral domain features; the convolved spectral domain features are mapped back to the spatial domain through inverse graph Fourier transform to obtain the spatial features of the node neighborhood. Preferably, the learnable filter kernel parameters are parameterized using Chebyshev polynomial approximation to reduce the computational complexity of spectral domain convolution and improve the practicality of the model in large-scale power grids.

[0012] As another technical solution of the present invention, when performing power grid operation status assessment and fault diagnosis, the spatiotemporal operation status characteristics of the power grid are input into a pre-trained multi-level state assessment model to calculate the transient stability margin index of each region of the power grid. Based on the comparison results of the transient stability margin index and the preset stability threshold, the operation status of each region of the power grid is classified into levels, generating an operation status label set. Simultaneously, the spatiotemporal operation status characteristics of the power grid are input into a pre-trained fault diagnosis classifier. Through the fully connected layer and Softmax activation function in the fault diagnosis classifier, the power grid fault type discrimination result, fault component location result, and fault severity assessment result are output. This solution achieves quantitative assessment of the power grid operation status and refined diagnosis of fault events.

[0013] Specifically, the processing flow of the fault diagnosis classifier is as follows: The spatiotemporal operating state features of the power grid are input into the first fully connected layer of the fault diagnosis classifier, and a first-dimensional compression operation is performed to obtain a compressed feature vector. The compressed feature vector is then input into the fault type discrimination branch of the fault diagnosis classifier. Through the second fully connected layer and the first Softmax activation function, the probability distribution of power grid fault types is output, and the fault type corresponding to the maximum probability is selected as the power grid fault type discrimination result. The compressed feature vector is then input into the fault component location branch of the fault diagnosis classifier. Through the third fully connected layer and the second Softmax activation function, the probability distribution of faults in each power grid component is output, and the component corresponding to the maximum probability is selected as the fault component location result. Finally, the compressed feature vector is input into the fault severity assessment branch of the fault diagnosis classifier. Through the fourth fully connected layer and the third Softmax activation function, the probability distribution of power grid fault severity levels is output, and the level corresponding to the maximum probability is selected as the fault severity assessment result. This multi-branch parallel processing method can simultaneously complete fault type identification, component location, and severity assessment, meeting the real-time requirements of the power grid stability control system for multiple diagnostic information.

[0014] In terms of model training, the spatiotemporal graph convolutional network model adopts a supervised learning strategy based on historical fault data in the pre-training stage, and uses a weighted combination of cross-entropy loss and mean squared error loss as the loss function for model training, so that the model can take into account both classification accuracy and feature representation ability.

[0015] Using the above method, the wide-area measurement data of the power grid is integrated into a spatiotemporal synchronization matrix and reconstructed into an equipment-level operating status feature map. By using a spatiotemporal graph convolutional network to mine deep spatiotemporal features, the online assessment of operating status and intelligent fault diagnosis are finally realized, providing a reliable decision-making basis for the safe and stable control of the power grid.

[0016] The technical effects and advantages provided by the present invention in the above technical solution are as follows: After constructing a spatiotemporal synchronized data matrix based on voltage phasor data, current phasor data, and frequency data, state topology reconstruction is performed based on the time-scale constraints of electromechanical transients in the power system. Specifically, voltage phase angle data is extracted from the spatiotemporal synchronized data matrix, and the phase angle change rate within a preset time window is calculated. The dominant oscillation mode in the electromechanical transient process is identified using time-domain signal processing techniques, and the power angle swing trajectory of each generator node is determined based on the oscillation frequency and damping ratio. Phase space reconstruction is performed on the power angle swing trajectory to generate state trajectory mapping points for each generator node. Simultaneously, spatial topological association is established between the state trajectory mapping points and the voltage amplitude trajectory mapping points of the load nodes to construct a device-level operational status feature map containing node states and edge connections. This processing no longer treats the measurement data of each node as an independent time series, but rather embeds the dynamic characteristics of generator power angle swing and the electrical connections between nodes into the graph structure. This allows the feature map to directly express the behavioral differences and interrelationships of each device during the transient process, effectively preserving the device-level transient evolution details in the wide-area measurement information.

[0017] A pre-trained spatiotemporal graph convolutional network model is used to perform multi-dimensional feature extraction on the device-level operational status feature map. Each spatiotemporal graph convolutional module in the model performs spatial domain graph convolution on the feature map output from the previous layer to extract spatial features of the node's neighborhood; then, temporal domain causal convolution is performed on these neighborhood spatial features to extract temporal dependent features. A gating mechanism is used to dynamically fuse spatial and temporal features. In the spatial domain graph convolution, learnable filter kernel parameters enable adaptive feature extraction of the graph structure, while Chebyshev polynomial approximation parameterization reduces the computational cost of spectral domain convolution. The sequential stacking of spatial domain graph convolution and temporal causal convolution allows the model to abstract the temporal and spatial correlation patterns in the power grid transient process layer by layer. The gating fusion mechanism adaptively adjusts the information retention ratio of spatial and temporal features in each layer, thereby obtaining more discriminative spatiotemporal operational status features of the power grid. When classifying faults based on these characteristics, the fault type discrimination branch, fault component location branch, and fault severity assessment branch can output more accurate discrimination and location results, respectively. They are less likely to confuse the responses of adjacent components during electromechanical transients and are more adaptable to complex operating conditions such as oscillation center migration. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0019] Figure 1This is a flowchart of a method for state monitoring and fault diagnosis of a power grid safety and stability control system. Figure 2 This is a flowchart illustrating the method for constructing equipment-level operational status characteristic diagrams in power systems. Figure 3 It is a probability distribution curve of the severity of the fault. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] See Figure 1 This invention provides a method for monitoring the status and diagnosing faults in a power grid safety and stability control system, comprising the following steps: Acquire voltage phasor data, current phasor data, and frequency data collected by multiple synchronous phasor measurement units in the power grid wide-area measurement system; A spatiotemporal synchronization data matrix is ​​constructed based on the voltage phasor data, current phasor data, and frequency data. Based on the time scale constraints of electromechanical transients in the power system, the spatiotemporal synchronization data matrix is ​​subjected to state topology reconstruction processing to generate a device-level operating status feature map. The pre-trained spatiotemporal graph convolutional network model is invoked to perform multi-dimensional feature extraction processing on the device-level operation status feature map to obtain the spatiotemporal operation status features of the power grid; Based on the spatiotemporal operating characteristics of the power grid, power grid operating status assessment and fault diagnosis operations are performed.

[0022] Example 1

[0023] The device identifiers and communication addresses of the synchronous phasor measurement units deployed at each node in the power grid are determined. In specific implementation, the unique device identifier and corresponding network communication address of each synchronous phasor measurement unit are read from the configuration management database of the power grid wide-area measurement system. The network communication address of the synchronous phasor measurement unit adopts a combination of IP address and port number. A high-frequency sampling command is sent to each synchronous phasor measurement unit through the wide-area communication network. The high-frequency sampling command carries a sampling rate parameter, which is set to 100Hz, indicating that data is collected once every 10 milliseconds. Measurement data packets returned by each synchronous phasor measurement unit are received. The measurement data packets include voltage amplitude, voltage phase angle, current amplitude, current phase angle, and frequency measurements with GPS time stamps. The received voltage amplitude, voltage phase angle, current amplitude, current phase angle, and frequency measurements are subjected to clock synchronization calibration. The clock synchronization calibration includes: using GPS time as a unified time reference, obtaining the GPS time tag from the measurement data packet reported by each synchronized phasor measurement unit, comparing the GPS time tag with the standard GPS reference time, calculating the time offset of each synchronized phasor measurement unit, correcting the voltage amplitude sequence, voltage phase angle sequence, current amplitude sequence, current phase angle sequence, and frequency sequence respectively using the Lagrange interpolation method, aligning the measurement data of different synchronized phasor measurement units to a unified time coordinate, and generating voltage phasor data, current phasor data, and frequency data with a unified time reference.

[0024] The generated voltage phasor data, current phasor data, and frequency data are aligned along the time dimension using a unified time base time label to form continuous time series data. In specific implementation, voltage phasor data, current phasor data, and frequency data from all synchronized phasor measurement units at the same time are extracted and merged to form a multidimensional time series dataset indexed by timestamps. A sliding time window is set with a length of 500 milliseconds and a sliding step size of 100 milliseconds. The continuous time series data is framed using the sliding time window, resulting in a series of time window data frames. Each time window data frame contains voltage phasor data, current phasor data, and frequency data from all synchronized phasor measurement units within that window. The voltage phasor data, current phasor data, and frequency data in each time window data frame are spatially sorted based on the electrical distance between nodes in the power grid topology. The electrical distance is defined using a quantitative index based on the node impedance matrix. and nodes electrical distance between Calculate using the following formula:

[0025] in, This represents the electrical distance between node m and node n. Represents a node Self-impedance, Represents a node Self-impedance, Represents a node With nodes The mutual impedance, self-impedance, and mutual impedance between nodes are all obtained from the elements of the power grid node impedance matrix. Electrical distance The smaller the value, the stronger the node. and nodes The closer the electrical connections, the better. The specific process of spatial dimension sorting is as follows: For each time window data frame, a node in the power grid topology is selected as the reference node for spatial sorting. The reference node is the generator node with the largest capacity in the power grid. The electrical distance between all other nodes and the reference node is calculated. The voltage phasor data, current phasor data, and frequency data corresponding to each node are arranged in ascending order of electrical distance, forming a spatially ordered two-dimensional data matrix. All time window data frames obtained after spatial dimension sorting are stacked according to the time dimension to generate a three-dimensional spatiotemporal synchronization data matrix. The dimensions of the three-dimensional spatiotemporal synchronization data matrix are the time frame number, spatial order position, and measurement feature channel, where the measurement feature channel includes voltage phasor features, current phasor features, and frequency features.

[0026] Example 2

[0027] See Figure 2 The voltage phase angle data is extracted from the spatiotemporal synchronization data matrix, and the phase angle change rate of each node within a preset time window is calculated. In specific implementation, data slices corresponding to the voltage phase angle channel are extracted from the measurement feature channel of the three-dimensional spatiotemporal synchronization data matrix to obtain the voltage phase angle value of each node at each time frame number. The length of the preset time window is set to 200 milliseconds, corresponding to 20 consecutive time frame numbers. For each node, the voltage phase angle value at the start and end of the preset time window is taken, and the phase angle change rate is calculated according to the following formula:

[0028] in, Represents a node Voltage phase angle change rate, node number The value ranges from 1 to integers, This represents the total number of nodes in the power grid that have synchronous phasor measurement units deployed. This represents the rate of change of the voltage phase angle at node p. Indicates the start time of the preset time window. Represents a node At the start of the preset time window The voltage phase angle value, Represents a node At the end of the preset time window The voltage phase angle value, This indicates the duration of the preset time window. Set to 200 milliseconds. The value is determined by the time frame number of the sliding time window. For all The voltage phase angle change rate is calculated for each node to obtain the voltage phase angle change rate vector.

[0029] Based on the voltage phase angle change rate vector, the dominant oscillation modes in the electromechanical transient process of the power system are identified using time-domain signal processing techniques. In specific implementations, the Prony analysis method is used for time-domain signal processing. The input signal of the Prony analysis method is the voltage phase angle change rate vector, and the output is the set of oscillation mode parameters contained in the electromechanical transient process of the power system. The specific implementation process of the Prony analysis method is as follows: the voltage phase angle change rate vector is used as a discrete-time signal sequence, and the length of the discrete-time signal sequence is... Equal to the total number of nodes A Hankel matrix is ​​constructed from the discrete-time signal sequence. Singular value decomposition is performed on the Hankel matrix, and the order of the signal model is determined based on the relative magnitudes of the singular values. The order of the signal model The cumulative singularity contribution rate exceeds 95% of the previous The number of singular values ​​is determined. The linear prediction coefficients of the signal model are solved using the least squares method. A characteristic polynomial is constructed based on these coefficients, and the eigenvalues ​​of the characteristic polynomial are solved. The frequency and damping ratio of each oscillation mode are extracted from the eigenvalues. The amplitude and initial phase of each oscillation mode are calculated. All oscillation modes are sorted from largest to smallest amplitude, and the oscillation mode corresponding to the maximum amplitude is selected as the dominant oscillation mode. The oscillation frequency of the dominant oscillation mode is then obtained. Damping ratio of the dominant oscillation mode .

[0030] Based on the oscillation frequency of the dominant oscillation mode Damping ratio of the dominant oscillation mode The power angle swing trajectory of each generator node is determined. In specific implementation, a complete list of generator nodes is obtained from the configuration database of the power grid wide-area measurement system; the total number of generator nodes is [number missing]. The voltage phase angle time series for each generator node is extracted from the voltage phase angle data. The length of the voltage phase angle time series is determined by the number of sliding time windows. The calculation method for the power angle swing trajectory is as follows: For the th... One generator node, power angle swing trajectory It is obtained by subtracting the reference phase angle of the grid inertia center under the dominant oscillation mode from the voltage phase angle time series, where Number the generator nodes. The value ranges from 1 to The reference phase angle of the power grid's center of inertia is calculated by weighted averaging of the voltage phase angle time series of all generator nodes. The weight of the average is the inertial time constant of each generator node, which is obtained from the power grid equipment parameter database. The power angle swing trajectory reflects the relative motion state of the generator rotor with respect to the system's center of inertia during electromechanical transients.

[0031] The power angle swing trajectory is reconstructed in phase space to generate state trajectory mapping points for each generator node. In specific implementation, a time-delay embedding method is used for phase space reconstruction. For the first... A sequence of power angle swing trajectories of generator nodes Embedding dimension of phase space reconstruction Set to 3, time delay The time delay is calculated using the mutual information method. The calculation process of the mutual information method is as follows: calculate the self-mutual information function for the power angle swing trajectory sequence, and take the time delay corresponding to the first local minimum point of the self-mutual information function as the time delay. The value of . In the embedding dimension For 3 and time delay Under the conditions, the first Each generator node at time The state trajectory mapping points are represented as three-dimensional vectors. ,in Take all time points in the sequence of oscillation trajectories with angular momentum that can form a complete three-dimensional vector. For all... Each generator node undergoes phase space reconstruction processing to generate a set of state trajectory mapping points for each generator node.

[0032] Spatial topological association is performed between the state trajectory mapping points of each generator node and the voltage amplitude trajectory mapping points of the load nodes to construct a device-level operational status feature map containing node states and edge connections. In specific implementation, a power grid topology graph structure containing all generator nodes and all load nodes is obtained from a power grid topology database. The power grid topology graph structure is represented as an undirected graph, and the set of nodes in the undirected graph... Include generator nodes and There are load nodes, among which Represents the total number of load nodes, and the set of edges in an undirected graph. This represents the transmission line connection relationships between nodes. For each load node, the voltage amplitude data is extracted and processed using the same time-delay embedding method as the phase space reconstruction of the generator node's power angle swing trajectory. The embedding dimension is set to 3, and the time delay is calculated using the mutual information method of the load node's voltage amplitude sequence. This generates voltage amplitude trajectory mapping points for the load nodes, which are represented as three-dimensional vectors. The construction method of the equipment-level operating status feature map is as follows: the node state characteristics of each node in the undirected graph consist of the three-dimensional coordinate values ​​of the corresponding generator node's state trajectory mapping point or the corresponding load node's voltage amplitude trajectory mapping point. The connection relationship of each edge in the undirected graph is related to the edge set in the power grid topology graph structure. Consistent with each other, the weight of an edge is determined by the reciprocal of the impedance modulus of the corresponding transmission line, which is obtained from the power grid equipment parameter database. The node state features and edge connectivity are combined to form a complete equipment-level operational status feature graph data structure. The node features of the equipment-level operational status feature graph are 3-dimensional, and the edge weights are scalar values.

[0033] Example 3

[0034] A spatiotemporal graph convolutional network model is constructed, consisting of an input layer, three spatiotemporal graph convolutional modules, and a cascaded output layer. In the specific implementation, the input layer receives a device-level operational status feature map as input data. This map is represented by a graph data structure, which includes a node feature matrix and an adjacency matrix. The node feature matrix has a dimension of [missing information]. ,in This represents the total number of nodes in the device-level operational status characteristic graph. Equal to the total number of generator nodes Total number of load nodes the sum of Indicates the number of feature channels of the input node. The value is set to 3, corresponding to the 3D features of the node state. The equipment-level operational status feature map is input into the input layer, which performs node feature normalization and edge weight initialization on the map. Node feature normalization uses the Z-score standardization method. The mean and standard deviation of each column of the node feature matrix are calculated. The node feature value of each column is subtracted from the mean of that column and then divided by the standard deviation to obtain the normalized node feature matrix. The edge weight initialization process involves extracting the edge connections and weights from the equipment-level operational status feature map. The edge weights are determined by the reciprocal of the impedance modulus of the corresponding transmission line. These edge weights are used as the element values ​​at the corresponding positions in the initial adjacency matrix of the spatiotemporal graph convolutional network model. Self-loop connections are added to the initial adjacency matrix, and the weight value of each self-loop connection is set to 1.0.

[0035] The first spatiotemporal graph convolution module processes the normalized node feature matrix and initial adjacency matrix output from the input layer, performing spatial domain graph convolution and temporal causal convolution operations. The specific process of the spatial domain graph convolution operation is as follows: Obtain the feature vector of each node in the feature map output from the previous layer and the adjacency matrix between nodes. When the previous layer is the input layer, the feature vector is the row vector of the normalized node feature matrix, and the adjacency matrix is ​​the initial adjacency matrix. Calculate the Laplacian matrix of the adjacency matrix. Defined as ,in Represents the identity matrix. Representing the adjacency matrix degree matrix, degree matrix Diagonal matrix, degree matrix The diagonal elements represent the adjacency matrix. The sum of the elements in the corresponding row. For the Laplace matrix. Perform eigenvalue decomposition to obtain the set of eigenvalues ​​of the Laplacian matrix. and eigenvector matrix The eigenvalue decomposition satisfies ,in It is a diagonal matrix composed of eigenvalues. Each column is an eigenvector corresponding to the eigenvalue.

[0036] Graph Fourier transform bases are constructed based on the eigenvalues ​​and eigenvectors of the Laplace matrix. The graph Fourier transform bases are composed of the eigenvector matrix. The column vectors are used to map the feature vector of each node to the spectral domain through a graph Fourier transform basis. The mapping method is to use the spectral domain feature vectors. Equal to the eigenvector matrix The transpose of the vector is multiplied by the node feature vector. Learnable filter kernel parameters are introduced into the spectral domain, and these parameters are parameterized using Chebyshev polynomial approximation. The formula for the spectral domain convolution operation, approximated by Chebyshev polynomials, is as follows:

[0037] in, This represents the feature vector output after the spectral domain convolution operation. Denotes the order of the Chebyshev polynomial. Set to 5. For summation index variables, Indicates the first Learnable filter kernel parameters corresponding to the Chebyshev polynomial of order 1 At the start of model training, a normal distribution with a mean of 0 and a variance of 0.01 is used for random initialization. During model training, the distribution is updated using the backpropagation algorithm. Indicates the first Chebyshev polynomials This represents a scaled eigenvalue diagonal matrix. The relationship with the eigenvalues ​​of the Laplace matrix is ​​as follows: , This represents the maximum value among all eigenvalues. This represents the nodal eigenvectors mapped to the spectral domain. Chebyshev polynomials. Calculated using recursion: , , ,in .

[0038] The feature vector output after spectral domain convolution operation Mapping back to the spatial domain via the inverse graphical Fourier transform, the inverse graphical Fourier transform is performed such that the eigenvectors in the spatial domain are equal to the eigenvector matrix. Multiply This yields the spatial features of the node's neighborhood. The above spatial domain graph convolution operation is then performed on each node in the device-level operational status feature map to obtain the spatial feature matrix of the node's neighborhood for all nodes.

[0039] A temporal causal convolution operation is performed on the node neighborhood spatial feature matrix to extract temporal dependent features. In the specific implementation, a one-dimensional causal convolution kernel is used. This kernel performs convolution operations only on features from the current time frame and previous time frames in the temporal dimension. The length of the one-dimensional causal convolution kernel is set to three time frames, and the number of channels in the kernel is the same as the number of channels in the node neighborhood spatial features. The weight parameters of the one-dimensional causal convolution kernel are initialized using the Xavier initialization method at the beginning of model training. The node neighborhood spatial feature matrix is ​​unfolded along the time frame dimension, and a one-dimensional causal convolution operation is applied to the temporal series features of each node to obtain the temporal dependent feature matrix.

[0040] A gating mechanism is used to dynamically fuse the spatial and temporal features output by each spatiotemporal graph convolutional module. In practice, the gating mechanism employs a sigmoid activation function to weight and fuse the spatial and temporal features; the calculation formula is: [Intermediate Feature Map] equals the spatial characteristic matrix Multiply by sigmoid (time feature matrix) Add the time feature matrix Multiply by (1-sigmoid(time feature matrix)) )),in Represents the feature matrix of the node's neighborhood space. represents the time-dependent feature matrix, and sigmoid represents the sigmoid activation function. The output of the gating mechanism serves as the intermediate feature map generated by the current spatiotemporal graph convolutional module, and the dimension of the intermediate feature map is consistent with the dimension of the input feature map.

[0041] The intermediate feature map generated by the first spatiotemporal graph convolutional module is passed to the second spatiotemporal graph convolutional module. The structure and operations of the second spatiotemporal graph convolutional module are the same as those of the first. The input of the second spatiotemporal graph convolutional module is the intermediate feature map output by the first spatiotemporal graph convolutional module and the initial adjacency matrix. The output is the intermediate feature map generated by the second spatiotemporal graph convolutional module. The intermediate feature map generated by the second spatiotemporal graph convolutional module is then passed to the third spatiotemporal graph convolutional module. The structure and operations of the third spatiotemporal graph convolutional module are the same as those of the first spatiotemporal graph convolutional module. The output is the intermediate feature map generated by the third spatiotemporal graph convolutional module.

[0042] The intermediate feature map output from the third spatiotemporal graph convolutional module is input into the output layer. The output layer undergoes fully connected mapping to obtain the spatiotemporal operating state features of the power grid. In specific implementation, the output layer contains a fully connected layer. The input feature dimension of the fully connected layer is the feature dimension of the intermediate feature map output from the third spatiotemporal graph convolutional module, and the output feature dimension of the fully connected layer is set to 128 dimensions. The weight matrix of the fully connected layer is initialized using the Xavier initialization method at the beginning of model training, and the bias vector of the fully connected layer is initialized to a zero vector. The intermediate feature map output from the third spatiotemporal graph convolutional module is flattened into a one-dimensional vector and input into the fully connected layer. After linear transformation, a 128-dimensional spatiotemporal operating state feature vector of the power grid is obtained.

[0043] The spatiotemporal graph convolutional network model employs a supervised learning strategy based on historical fault data during the pre-training phase. This historical fault data includes measurement data from the synchronization phasor measurement unit recorded under different fault conditions in the power grid, along with corresponding fault type labels, fault component labels, and fault severity labels. During pre-training, equipment-level operational status feature maps are extracted from the historical fault data as training samples. The labels for these training samples include fault type labels, fault component labels, and fault severity labels. The output of the spatiotemporal graph convolutional network model connects to three branches, which are used for fault type identification, fault component localization, and fault severity assessment, respectively. Each branch contains a fully connected layer and a softmax activation function. The loss function used in model training is a weighted combination of cross-entropy loss and mean squared error loss. Defined as:

[0044] in, This represents the total loss function value during the pre-training phase of the spatiotemporal graph convolutional network model. The cross-entropy loss represents the branch used to determine the fault type. This represents the cross-entropy loss of the faulty component location branch. The cross-entropy loss represents the branch used to assess the severity of the fault. This represents the mean square error loss between the feature vector of the spatiotemporal operation state of the power grid and the preset reference feature vector. , , and These are the loss weighting coefficients. Set to 1.0, Set to 1.0, Set to 0.5. The loss weight coefficient is set to 0.1, based on the balance of importance of each branch task in the overall fault diagnosis objective. During the pre-training phase, the Adam optimization algorithm is used to update all learnable parameters of the spatiotemporal graph convolutional network model. The initial learning rate is set to 0.001, the batch size to 32, and the number of training iterations to 200. During training, when the validation set loss no longer decreases after 10 consecutive iterations, the learning rate is reduced to 0.1 times its original value. After training, the parameters of the spatiotemporal graph convolutional network model are saved.

[0045] Example 4

[0046] The spatiotemporal operating characteristics of the power grid are input into a pre-trained multi-level state assessment model to calculate the transient stability margin index for each region of the power grid. In specific implementation, the multi-level state assessment model consists of a region partitioning module and a regional transient stability assessment network. The region partitioning module reads the power grid's partitioning configuration information from the power grid topology database. This configuration information divides all generator nodes and load nodes in the power grid into... The system is divided into several regions, each containing a number of nodes. The regions are defined based on the control zone boundaries in the power grid dispatching and operation procedures. The value is the total number of actual control zones of the power grid. The input of the regional transient stability assessment network is the regional feature vector aggregated by region from the spatiotemporal operating state features of the power grid. The spatiotemporal operating state features of the power grid are 128-dimensional vectors. The output is the transient stability margin index for each region. The aggregation method of the regional feature vector is as follows: for the , From the spatiotemporal operational status characteristics of the power grid in the region, the feature belonging to the first region is extracted. The characteristics of all nodes in the region are used to calculate the average of the characteristics of all nodes, thus obtaining the th region. Regional feature vectors of each region ,in For area code, The value ranges from 1 to The regional transient stability evaluation network consists of three cascaded fully connected layers. The first fully connected layer has an input dimension of 128 and an output dimension of 64, with the ReLU activation function. The second fully connected layer has an input dimension of 64 and an output dimension of 32, with the ReLU activation function. The third fully connected layer has an input dimension of 32 and an output dimension of 1, with the Sigmoid activation function. The output value is used as the integer. Transient stability margin indices for each region. The calculation formula is:

[0047] in, Indicates the first Transient stability margin index for each region The numerical range of is real numbers between 0 and 1. This represents the Sigmoid activation function. To modify the activation function of the linear unit, This represents the weight matrix of the first fully connected layer, with dimension 1. , This represents the bias vector of the first fully connected layer, with a dimension of 64. This represents the weight matrix of the second fully connected layer, with dimension 1. , This represents the bias vector of the second fully connected layer, which has a dimension of 32. This represents the weight matrix of the third fully connected layer, with dimension 1. , This represents the bias vector of the third fully connected layer, with a dimension of 1. Indicates the first The regional feature vectors for each region are 128-dimensional. The multi-level state assessment model is trained using historical stable data samples during the pre-training phase. These samples include 128-dimensional spatiotemporal operating state features of the power grid under different operating conditions, along with corresponding regional stability labels. The regional stability labels are either 0 or 1, with 0 indicating an unstable state and 1 indicating a stable state. The mean squared error loss function is used during pre-training, and the Adam optimization algorithm is selected. The initial learning rate is set to 0.0005, the batch size to 64, and the number of training iterations to 150.

[0048] Based on the transient stability margin index The comparison results with preset stability thresholds are used to classify the operating status of each region of the power grid and generate an operating status label set. In specific implementation, the preset stability thresholds include a first stability threshold. Second stability threshold First stability threshold Set to 0.8, the second stability threshold Set to 0.2. First stability threshold. The rationale for setting it to 0.8 is as follows: On the Sigmoid output characteristic curve, when the input value exceeds the linear region and enters the saturation region, the output value tends to 1. The input value corresponding to 0.8 is in the strong positive response region, which can reliably characterize the stable state. Second stability threshold. The rationale for setting it to 0.2 is as follows: when the output value is less than 0.2, the input value is in a strong negative response region, which can reliably characterize the unstable state. The rule for classifying the operating state level is: when... At that time, the first The operational status level of each region is classified as "stable"; when At that time, the first The operational status level of each region is classified as "critically stable"; when At that time, the first The operational status level of each region is classified as "unstable". For all... Each region performs an operation to classify its operational status level, generating an operational status label set. The operational status label set contains the region number and operational status level label for each region.

[0049] The spatiotemporal operating state features of the power grid are input into a pre-trained fault diagnosis classifier. Through a fully connected layer and a Softmax activation function in the classifier, the classifier outputs power grid fault type discrimination results, fault component location results, and fault severity assessment results. In specific implementation, the pre-trained fault diagnosis classifier consists of a feature compression fully connected layer, a fault type discrimination sub-network, a fault component location sub-network, and a fault severity assessment sub-network. The input to the feature compression fully connected layer is 128-dimensional power grid spatiotemporal operating state features, and the output is a 64-dimensional compressed feature vector. The ReLU activation function is used. The fault type discrimination sub-network contains one fully connected layer and one Softmax activation function. The fully connected layer of the fault type discrimination sub-network has an input dimension of 64 dimensions and an output dimension of... dimension, This represents the total number of predefined fault types. The Softmax activation function outputs the probability distribution of power grid fault types, and the fault type corresponding to the maximum probability is selected as the power grid fault type discrimination result. The fault component localization subnetwork contains a fully connected layer and a Softmax activation function. The fully connected layer of the fault component localization subnetwork has an input dimension of 64 dimensions and an output dimension of... dimension, This indicates the total number of components in the power grid that need to be located. The value is equal to the total number of generator nodes. Total number of load nodes and the total number of transmission lines The sum of these factors, along with the Softmax activation function, outputs the fault probability distribution of each power grid component. The component with the highest probability is selected as the fault location result. The fault severity assessment subnetwork contains a fully connected layer and a Softmax activation function. The fully connected layer of the fault severity assessment subnetwork has an input dimension of 64 dimensions and an output dimension of [missing information]. dimension, This represents the total number of predefined fault severity levels. The value is set to 3, corresponding to three levels: minor, moderate, and severe faults. The Softmax activation function outputs the probability distribution of the power grid fault severity levels, and the level corresponding to the highest probability is selected as the fault severity assessment result. During the pre-training phase, the fault diagnosis classifier adopts a strategy of joint training with the spatiotemporal graph convolutional network model. The training data and loss function are consistent with those of the spatiotemporal graph convolutional network model during pre-training. During training, the feature compression fully connected layer of the fault diagnosis classifier receives the spatiotemporal operating state features of the power grid output by the spatiotemporal graph convolutional network model as input. The three sub-networks of the fault diagnosis classifier independently calculate the cross-entropy loss. The cross-entropy losses of the three sub-networks are added with equal weights and then weighted and combined with other loss functions of the spatiotemporal graph convolutional network model to jointly participate in backpropagation and update parameters.

[0050] Example 5

[0051] The spatiotemporal operating state features of the power grid are input into the first fully connected layer of the fault diagnosis classifier, where a first-dimensional compression operation is performed to obtain a compressed feature vector. In specific implementation, the spatiotemporal operating state features of the power grid are 128-dimensional vectors generated by the output layer of the spatiotemporal graph convolutional network model. The input dimension of the first fully connected layer is 128-dimensional, and the output dimension is 64-dimensional. The ReLU function is used as the activation function for the first fully connected layer. The weight matrix of the first fully connected layer is shown below. The dimension is The bias vector of the first fully connected layer The dimension is 64. The formula for calculating the first dimension compression operation is:

[0052] in, This represents a compressed feature vector with a dimension of 64. This represents the weight matrix of the first fully connected layer. The Xavier initialization method is used for initialization at the beginning of model training. This represents a feature vector representing the spatiotemporal operating state of the power grid, with a dimension of 128. This indicates a modified linear unit activation function. , This represents the bias vector of the first fully connected layer. Initialize to a zero vector. and The model is updated using the backpropagation algorithm during training.

[0053] Compress feature vectors The input to the fault diagnosis classifier is the fault type discrimination branch. This branch consists of a second fully connected layer and a first Softmax activation function. In a specific implementation, the input dimension of the second fully connected layer is 64-dimensional, and the output dimension is... dimension, This indicates the total number of predefined fault types. The values ​​are determined by the power grid fault classification criteria. The weight matrix of the second fully connected layer. The dimension is The bias vector of the second fully connected layer The dimension is Dimension. The second fully connected layer compresses the feature vector. Perform a linear transformation to obtain the original score vector for fault type discrimination. , Original score vector for fault type discrimination The dimension is Dimension, original score vector for fault type discrimination The Middle element Indicates that the power grid belongs to the first The original scores for each fault type, among which The value ranges from 1 to Integers. The original score vector for fault type discrimination. The input is the first Softmax activation function, which determines the original score vector for fault type identification. After performing an exponential operation on each element and then normalizing it, the probability distribution of power grid fault types is output. Probability distribution of power grid fault types The Middle element The calculation method is as follows ,in To accumulate index variables, Take all from 1 to All integer values. Select the probability distribution of power grid fault types. The fault type corresponding to the highest probability in the middle is used as the result of power grid fault type discrimination.

[0054] Compress feature vectors The input to the fault diagnosis classifier is the fault component localization branch. This branch consists of a third fully connected layer and a second Softmax activation function. In a specific implementation, the input dimension of the third fully connected layer is 64-dimensional, and the output dimension is... dimension, This indicates the total number of components in the power grid that need to be located. The value is equal to the total number of generator nodes. Total number of load nodes and the total number of transmission lines The sum of the total number of generator nodes and the total number of load nodes The total number of transmission lines was obtained from the power grid topology database. Represents the set of edges in the power grid topology graph structure. The number of elements. The weight matrix of the third fully connected layer. The dimension is The bias vector of the third fully connected layer The dimension is Dimension. The third fully connected layer compresses the feature vector. Perform a linear transformation to obtain the original score vector for fault component location. , Original score vector for faulty component location The dimension is Dimension, original score vector for fault component location The Middle element Indicates the first in the power grid The original score for each component failure, of which The value ranges from 1 to Integers. Locate the faulty component using the original score vector. The input is a second Softmax activation function, which is used to locate the original score vector of the faulty component. Each element is exponentially calculated and then normalized to output the fault probability distribution of each component in the power grid. Fault probability distribution of various power grid components The Middle element The calculation method is as follows ,in To accumulate index variables, Take all from 1 to All integer values. Select the fault probability distribution for each component of the power grid. The component with the highest probability in the middle is used as the location result of the faulty component.

[0055] Compress feature vectors The input to the fault diagnosis classifier is the fault severity evaluation branch. This branch consists of a fourth fully connected layer and a third Softmax activation function. In this implementation, the input dimension of the fourth fully connected layer is 64-dimensional, and the output dimension is... dimension, This represents the total number of predefined fault severity levels. The value is set to 3, with three fault severity levels: minor, moderate, and severe. (Weight matrix of the fourth fully connected layer.) The dimension is The bias vector of the fourth fully connected layer The dimension is Dimension. The fourth fully connected layer compresses the feature vector. Perform a linear transformation to obtain the original score vector for assessing the severity of the fault. , Original score vector for assessing the severity of the fault The dimension is Dimension, original score vector for assessing fault severity The Middle element This indicates that the power grid fault belongs to the first The original scores for each severity level, of which The value ranges from 1 to Integers. The original score vector is used to assess the severity of the fault. The input is a third Softmax activation function, which evaluates the original score vector of the fault severity. After performing an exponential operation on each element and then normalizing it, the probability distribution of the severity level of the power grid fault is output. Probability distribution of power grid fault severity levels The Middle element The calculation method is as follows ,in To accumulate index variables, Take all from 1 to All integer values. Select the probability distribution of power grid fault severity levels. The level corresponding to the highest probability is used as the assessment result of the fault severity.

[0056] See Figure 3In the graph, the horizontal axis represents the test sample number, ranging from 0 to 500, and the vertical axis represents the probability value, ranging from 0 to 1. The three curves in the legend correspond to the probability of minor faults (solid green line), moderate faults (dashed orange line), and severe faults (dotted red line), respectively. The trend of the curves is analyzed based on the output results of the fault diagnosis classifier in Example 5.

[0057] Within the sample number range of 0 to 200, the green minor fault probability curve fluctuates relatively stably and remains between 0.7 and 0.9, indicating that the probability of samples being diagnosed as minor faults is relatively high during this period; the orange moderate fault probability curve and the red severe fault probability curve both remain below 0.1, indicating that the possibility of moderate and severe faults is low.

[0058] Within the sample sequence number range of 200 to 380, the probability of minor green faults decreased significantly to below 0.2, the probability of moderate orange faults increased significantly, with fluctuations concentrated between 0.6 and 0.9, and the probability of severe red faults remained at a low level. This indicates that the samples in this stage were mainly judged as moderate faults and had a high diagnostic confidence.

[0059] Within the sample sequence number range of 380 to 500, the probability curve for severe red faults rises rapidly and stabilizes within the range of 0.7 to 0.9, while the probability curves for minor green faults and moderate orange faults both drop below 0.2. This indicates that the samples in this stage are judged to have the highest probability of severe faults, and the fault diagnosis classifier accurately distinguishes the change process of fault severity.

[0060] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for state monitoring and fault diagnosis of a power grid safety and stability control system, characterized in that, Includes the following steps: Acquire voltage phasor data, current phasor data, and frequency data collected by multiple synchronous phasor measurement units in the power grid wide-area measurement system; A spatiotemporal synchronization data matrix is ​​constructed based on the voltage phasor data, current phasor data, and frequency data. Based on the time scale constraints of electromechanical transients in the power system, the spatiotemporal synchronization data matrix is ​​subjected to state topology reconstruction processing to generate a device-level operating status feature map. The pre-trained spatiotemporal graph convolutional network model is invoked to perform multi-dimensional feature extraction processing on the device-level operation status feature map to obtain the spatiotemporal operation status features of the power grid; Based on the spatiotemporal operating characteristics of the power grid, power grid operating status assessment and fault diagnosis operations are performed.

2. The method for state monitoring and fault diagnosis of a power grid safety and stability control system according to claim 1, characterized in that, The acquisition of voltage phasor data, current phasor data, and frequency data collected by multiple synchronous phasor measurement units in the power grid wide-area measurement system specifically involves: Determine the device identifier and communication address of the synchronous phasor measurement unit deployed at each node in the power grid; High-frequency sampling commands are sent to each synchronous phasor measurement unit via a wide-area communication network; Receive the voltage amplitude, voltage phase angle, current amplitude, current phase angle, and frequency measurements returned by each synchronous phasor measurement unit, which include time tags; The voltage amplitude, voltage phase angle, current amplitude, current phase angle, and frequency measurements are clock-synchronized and calibrated to generate voltage phasor data, current phasor data, and frequency data with a unified time reference.

3. The method for state monitoring and fault diagnosis of a power grid safety and stability control system according to claim 1, characterized in that, The construction of the spatiotemporal synchronization data matrix based on the voltage phasor data, current phasor data, and frequency data specifically involves: The voltage phasor data, current phasor data, and frequency data are aligned according to time labels to generate continuous time series data. A sliding time window is set to divide the continuous time series data into frames, resulting in multiple time window data frames. The voltage phasor data, current phasor data, and frequency data in each time window data frame are sorted spatially according to the electrical distance between nodes in the power grid topology. The arranged data frames are stacked according to the time dimension to generate a three-dimensional spatiotemporal synchronization data matrix.

4. The method for state monitoring and fault diagnosis of a power grid safety and stability control system according to claim 1, characterized in that, Based on the electromechanical transient time scale constraints of the power system, the spatiotemporal synchronization data matrix is ​​subjected to state topology reconstruction processing to generate a device-level operating status feature map, specifically as follows: Extract the voltage phase angle data from the spatiotemporal synchronization data matrix and calculate the rate of change of the voltage phase angle of each node within a preset time window; Based on the phase angle change rate, the dominant oscillation mode in the electromechanical transient process of the power system is identified by time-domain signal processing technology; Based on the oscillation frequency and damping ratio of the dominant oscillation mode, the power angle swing trajectory of each generator node is determined; The power angle swing trajectory is reconstructed in phase space to generate state trajectory mapping points for each generator node; Spatial topological association is performed between the state trajectory mapping points of each generator node and the voltage amplitude trajectory mapping points of the load node to construct a device-level operational status feature map that includes node status and edge connection relationships.

5. The method for state monitoring and fault diagnosis of a power grid safety and stability control system according to claim 1, characterized in that, The pre-trained spatiotemporal graph convolutional network model is invoked to perform multi-dimensional feature extraction processing on the device-level operational status feature map to obtain the spatiotemporal operational status features of the power grid, specifically: A spatiotemporal graph convolutional network model is constructed, which includes an input layer, multiple spatiotemporal graph convolutional modules, and an output layer. The device-level operational status feature map is input into the input layer, and node feature normalization and edge weight initialization are performed on the device-level operational status feature map. Each spatiotemporal graph convolution module performs spatial domain graph convolution on the feature map output from the previous layer to extract the spatial features of the node's neighborhood. Perform temporal causal convolution on the spatial features of the node's neighborhood to extract temporal dependency features; The spatial and temporal features output by each spatiotemporal graph convolutional module are dynamically fused using a gating mechanism to generate an intermediate feature map. The intermediate feature map output by the last spatiotemporal graph convolution module is input into the output layer and fully connected mapping is performed to obtain the spatiotemporal operating state features of the power grid.

6. The method for state monitoring and fault diagnosis of a power grid safety and stability control system according to claim 5, characterized in that, The step of extracting node neighborhood spatial features by performing spatial domain graph convolution operation on the feature map output by the previous layer through each spatiotemporal graph convolution module is as follows: Obtain the feature vector of each node and the adjacency matrix between nodes in the feature map output by the previous layer; Calculate the Laplacian matrix of the adjacency matrix, and perform eigenvalue decomposition on the Laplacian matrix to obtain the eigenvalues ​​and eigenvectors of the Laplacian matrix; Based on the eigenvalues ​​and eigenvectors, a graph Fourier transform basis is constructed, and the eigenvectors of each node are mapped to the spectral domain through the graph Fourier transform basis; Learnable filter kernel parameters are introduced into the spectral domain, and convolution operations are performed on the mapped spectral domain features. The spectral domain features after convolution are mapped back to the spatial domain through inverse graph Fourier transform to obtain the spatial features of the node neighborhood.

7. The method for state monitoring and fault diagnosis of a power grid safety and stability control system according to claim 6, characterized in that, The learnable filter kernel parameters are parameterized using Chebyshev polynomial approximation to reduce the computational complexity of spectral domain convolution.

8. The method for state monitoring and fault diagnosis of a power grid safety and stability control system according to claim 1, characterized in that, The operation of power grid operation status assessment and fault diagnosis based on the spatiotemporal operation status characteristics of the power grid specifically includes: The spatiotemporal operating state characteristics of the power grid are input into a pre-trained multi-level state assessment model to calculate the transient stability margin index of each region of the power grid. Based on the comparison results between the transient stability margin index and the preset stability threshold, the operating status of each region of the power grid is classified into different levels, and an operating status label set is generated. The spatiotemporal operating state characteristics of the power grid are input into a pre-trained fault diagnosis classifier. Through the fully connected layer and Softmax activation function in the fault diagnosis classifier, the power grid fault type discrimination result, fault component location result, and fault severity assessment result are output.

9. The method for state monitoring and fault diagnosis of a power grid safety and stability control system according to claim 8, characterized in that, The process involves inputting the spatiotemporal operating state characteristics of the power grid into a pre-trained fault diagnosis classifier. Through the fully connected layer and Softmax activation function in the fault diagnosis classifier, the classifier outputs power grid fault type discrimination results, fault component location results, and fault severity assessment results. Specifically: The spatiotemporal operating state features of the power grid are input into the first fully connected layer of the fault diagnosis classifier, and the first dimension compression operation is performed to obtain the compressed feature vector. The compressed feature vector is input into the fault type discrimination branch of the fault diagnosis classifier. Through the second fully connected layer and the first Softmax activation function in the fault type discrimination branch, the probability distribution of power grid fault types is output. The fault type corresponding to the maximum probability is selected as the power grid fault type discrimination result. The compressed feature vector is input into the fault component localization branch of the fault diagnosis classifier. Through the third fully connected layer and the second Softmax activation function in the fault component localization branch, the fault probability distribution of each power grid component is output, and the component corresponding to the maximum probability is selected as the fault component localization result. The compressed feature vector is input into the fault severity assessment branch of the fault diagnosis classifier. Through the fourth fully connected layer and the third Softmax activation function in the fault severity assessment branch, the probability distribution of the power grid fault severity level is output. The level corresponding to the maximum probability is selected as the fault severity assessment result.

10. The method for state monitoring and fault diagnosis of a power grid safety and stability control system according to claim 1, characterized in that, The spatiotemporal graph convolutional network model adopts a supervised learning strategy based on historical fault data during the pre-training stage, using a weighted combination of cross-entropy loss and mean squared error loss as the loss function for model training.

Images