Power grid transient voltage stability calculation method based on neural network

Abstract

This application discloses a neural network-based method for calculating transient voltage stability of power grids. By analyzing circuit breaker state signals in real time and constructing a dynamic adjacency matrix sequence, the model can accurately capture real-time changes in the power grid structure caused by events such as fault clearing. This ensures that the feature aggregation process strictly follows the dynamic evolution of physical connections, achieving accurate modeling of energy flow paths during transient processes. This approach enhances the model's ability to capture changes in energy flow paths during transient processes and improves its physical consistency. Consequently, it significantly improves the accuracy, robustness, and interpretability of voltage stability assessments under complex fault, noise interference, and dynamic topology changes, providing more reliable intelligent support for power grid security defense and control decisions.

Description

Technical Field

[0001] This application relates to the field of intelligent assessment, and more specifically, to a method for calculating the transient voltage stability of power grids based on neural networks. Background Technology

[0002] With the continuous expansion of modern power systems and the increasing sophistication of power electronics, the operating characteristics of power grids are becoming more complex, and transient voltage stability has become a core indicator for ensuring the safe operation of large power grids. To rapidly assess system risks and implement preventative control measures after large disturbances, constructing an efficient and reliable transient voltage stability calculation scheme for power grids is of significant engineering importance in avoiding voltage collapse and large-scale power outages. Currently, academia and industry are actively exploring artificial intelligence-based assessment methods, particularly utilizing graph neural networks to process massive measurement data under power grid topology, aiming to balance the need for computational efficiency with the extraction of spatial correlations.

[0003] However, existing power grid transient voltage stability assessment schemes, especially those based on graph convolutional networks (GCNs), suffer from a significant lack of dynamic topology adaptability when dealing with real-world power grid fault scenarios. These schemes typically rely on the idealized assumption that the power grid topology remains static, or they simply use the adjacency matrix from before the fault occurred during modeling. However, in real-world operating environments, fault clearing caused by relay protection actions inevitably leads to the instantaneous interruption of transmission lines, resulting in drastic real-time physical abrupt changes in the network topology. Because existing algorithms cannot dynamically perceive and identify these topological changes, their models cannot accurately represent the energy flow path after the fault during feature aggregation. Considering that changes in physical connections directly determine the system's voltage support capability for nodes, static graph models exhibit severe physical distortion when facing topological abrupt changes, leading to fatal biases in the assessment results and significantly limiting the model's generalization performance and practical value under complex power system conditions.

[0004] Therefore, an optimized neural network-based method for calculating the transient voltage stability of power grids is needed. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides a method for calculating the transient voltage stability of a power grid based on neural networks.

[0006] According to one aspect of this application, a method for calculating the transient voltage stability of a power grid based on a neural network is provided, comprising: S1: Perform synchronous feature mapping on the acquired raw measurement data stream and raw state data stream to generate analog characteristic sequences and digital topology sequences; S2: Construct a dynamic adjacency matrix sequence based on digital topology sequences; S3: Based on the dynamic adjacency matrix sequence, perform graph feature aggregation on the analog characteristic sequence under topological constraints to obtain a spatial feature map containing physical topological constraints; S4: Perform transient evolution trend capture and global pooling on the spatial feature map to obtain the spatiotemporal feature vector; S5: The spatiotemporal feature vectors are classified and mapped for probability and stability are determined by a multilayer perceptron to obtain stability evaluation results.

[0007] Compared with existing technologies, this application provides a neural network-based method for calculating transient voltage stability of power grids. By analyzing circuit breaker state signals in real time and constructing a dynamic adjacency matrix sequence, the model can accurately capture real-time changes in the power grid structure caused by events such as fault clearing. This ensures that the feature aggregation process strictly follows the dynamic evolution of physical connections, achieving accurate modeling of energy flow paths during transient processes. This approach enhances the model's ability to capture changes in energy flow paths during transient processes and improves its physical consistency. Consequently, it significantly improves the accuracy, robustness, and interpretability of voltage stability assessment in scenarios involving complex faults, noise interference, and dynamic topology changes, providing more reliable intelligent support for power grid security defense and control decisions. Attached Figure Description

[0008] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1 This is a flowchart of a neural network-based method for calculating transient voltage stability of a power grid according to an embodiment of this application; Figure 2 This is a schematic diagram of the data flow of the neural network-based power grid transient voltage stability calculation method according to an embodiment of this application; Figure 3 This is a flowchart of step S2 in the neural network-based power grid transient voltage stability calculation method according to an embodiment of this application; Figure 4 This is a flowchart of step S3 in the neural network-based power grid transient voltage stability calculation method according to an embodiment of this application. Detailed Implementation

[0010] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0011] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0012] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0013] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0014] In the technical solution of this application, a method for calculating the transient voltage stability of power grid based on neural networks is proposed. Figure 1 This is a flowchart of a neural network-based method for calculating the transient voltage stability of a power grid according to an embodiment of this application. Figure 2 This is a system architecture diagram of a neural network-based method for calculating power grid transient voltage stability according to an embodiment of this application. Figure 1 and Figure 2 As shown, the neural network-based method for calculating the transient voltage stability of a power grid according to an embodiment of this application includes the following steps: S1, performing synchronous feature mapping on the acquired raw measurement data stream and raw state data stream to generate an analog feature sequence and a digital topology sequence; S2, constructing a dynamic adjacency matrix sequence based on the digital topology sequence; S3, performing graph feature aggregation under topological constraints on the analog feature sequence based on the dynamic adjacency matrix sequence to obtain a spatial feature map containing physical topological constraints; S4, performing transient evolution trend capture and global pooling processing on the spatial feature map to obtain a spatiotemporal feature vector; S5, performing classification probability mapping and stability discrimination on the spatiotemporal feature vector using a multilayer perceptron to obtain a stability evaluation result.

[0015] Specifically, in step S1, synchronous feature mapping is performed on the acquired raw measurement data stream and raw state data stream to generate analog characteristic sequences and digital topology sequences. It should be understood that the power grid transient process involves massive amounts of heterogeneous data. The raw measurement data reflects the continuous physical evolution of the system, while the raw state data reflects discrete topological changes; the two differ in time scale and data dimension. To enable subsequent deep learning models to accurately capture spatiotemporal correlations, the technical solution of this application performs synchronous feature mapping on the acquired raw measurement data stream and raw state data stream to transform these multi-source heterogeneous data into standardized input sequences aligned on the time axis and logically correlated, thereby providing a unified data benchmark for dynamic topology sensing.

[0016] The raw measurement data stream includes time series of voltage amplitude, voltage phase angle, active power, and reactive power at key nodes across the entire network. These data characterize the real-time operating status of the power system. The raw status data stream includes digital signals of the opening / closing of circuit breakers at both ends of key transmission lines across the entire network, used to depict the physical connectivity of the power grid. The final generated analog characteristic sequence is a set of continuous physical quantities after normalization, while the digital topology sequence is a discrete coded sequence used to map the real-time connectivity of the power grid.

[0017] In practical implementation, firstly, the original measurement data stream and the original state data stream are time-synchronized and missing value repaired to obtain a synchronized and aligned dataset. In actual engineering, due to differences in the sampling frequency or transmission delay of the measurement devices, the technical solution of this application performs time synchronization and missing value repair on the original measurement data stream and the original state data stream to ensure that analog and digital quantities match at the same time plane. In this process, time-stamp remapping of all data points can be performed using a high-precision unified clock signal to eliminate time jitter caused by differences in communication links. For data points with missing values, a linear interpolation algorithm based on time series is used for filling. That is, the values ​​of the two nearest valid data points before and after the missing time point are linearly calculated according to the proportion of time distance to estimate the value of the missing point, ensuring the continuity and integrity of the data in the time dimension.

[0018] Next, based on the fault trigger time, a sliding window is used to truncate the synchronized alignment dataset to obtain a time-series slice. During this process, the status data stream is continuously monitored, and when a specific circuit breaker tripping signal characterizing the fault is identified, that moment is marked as the fault trigger time. Subsequently, a sliding time window of preset length is used to truncate the synchronized alignment dataset before and after that moment, resulting in a three-dimensional data block focusing on the transient process of the fault, i.e., the time-series slice. This slice retains complete dynamic evolution information before and after the fault.

[0019] Furthermore, the time-series slices undergo analog-to-digital separation and feature normalization reconstruction to obtain analog feature sequences and digital topology sequences. In this process, firstly, based on the physical properties of the data, continuously changing quantities are classified as analog quantities, and discrete state quantities are classified as digital quantities. Next, the analog quantities undergo feature normalization, that is, all values ​​of the feature within the current sliding window are linearly scaled to a standard range of zero to one, specifically by subtracting the minimum value within the window and dividing by the range. For digital quantities, direct binarization mapping is performed. Finally, the normalized analog quantities are used to construct the analog feature sequence, while the binarized digital quantities are used to construct the digital topology sequence.

[0020] Specifically, S2 involves constructing a dynamic adjacency matrix sequence based on a digital topology sequence. It should be understood that traditional graph convolutional networks typically assume a static power grid topology or utilize only pre-fault connectivity. However, in actual operation, fault clearing caused by relay protection actions can trigger instantaneous interruptions in transmission lines, resulting in drastic real-time physical abrupt changes. To address the problem that simple topology connections cannot characterize electrical strength, the technical solution in this application constructs a dynamic adjacency matrix sequence to perceive topology abrupt changes caused by faults in real time. This ensures that the model can correctly characterize the energy flow path after the fault during feature aggregation, improving the model's evaluation accuracy and generalization ability under complex fault conditions.

[0021] Figure 2 This is a schematic diagram of the data flow in the neural network-based power grid transient voltage stability calculation method according to an embodiment of this application. Figure 2 As shown, S2 includes: S21, performing physical connection benchmark initialization on the digital topology sequence based on a preset power grid physical wiring diagram to obtain a static reference matrix; S22, performing connectivity blocking on the digital topology sequence and the static reference matrix based on circuit breaker switch state mapping logic to obtain a real-time mask matrix; S23, performing physical topology reconstruction on the static reference matrix and the real-time mask matrix to obtain a dynamic adjacency matrix sequence.

[0022] Specifically, in step S21, based on a preset power grid physical wiring diagram, the digital topology sequence is initialized with a physical connection reference to obtain a static reference matrix. The power grid physical wiring diagram defines nodes with physical line connections when all circuit breakers are closed. Based on this wiring diagram, a static reference matrix can be constructed. This matrix is ​​a binary symmetric matrix, and its element assignment rules are as follows: if a node... and nodes If a direct physical line connection exists in the preset physical wiring diagram, the value of the corresponding matrix element is set to 1 to indicate the existence of a potential connection path; otherwise, the value is set to 0, indicating that there is no direct physical line between the two nodes. This matrix represents the maximum possible set of connections in the power grid topology and serves as the baseline starting point for all dynamic changes.

[0023] Specifically, in step S22, based on the circuit breaker switch state mapping logic, connectivity blocking is applied to the digital topology sequence and the static reference matrix to obtain a real-time mask matrix. During this process, for each time step in the digital topology sequence... Read the open / close status of all relevant circuit breakers at that moment. Generate a value corresponding to the time step based on a defined circuit breaker switch status mapping logic. The real-time mask matrix is ​​generated. Here, the circuit breaker switch state mapping logic is: a line is electrically connected only when both circuit breakers at both ends are in the closed state. The final generated real-time mask matrix is ​​also a binary matrix, and its element assignment rule is: if the connected node... and nodes The circuit breakers at both ends of the physical line at any time If all circuit breakers are in the closed state, the corresponding element in the mask matrix is ​​set to 1, indicating that the connection is valid at the current moment; otherwise, the value is set to 0, indicating that the connection is blocked at the current moment due to the circuit breaker tripping. This matrix accurately reflects the actual on / off status of the power grid at each moment.

[0024] Specifically, in step S23, physical topology reconstruction is performed on the static reference matrix and the real-time mask matrix to obtain a dynamic adjacency matrix sequence. During this process, the obtained static reference matrix is ​​compared with that of each time step... The real-time mask matrix is ​​multiplied element-wise (i.e., Hadamard product) to filter all potential connections described by the static reference matrix. Only connections with a value of 1 in the static reference matrix (indicating the existence of a physical line) and a corresponding value of 1 in the real-time mask matrix (indicating that the line is currently conducting) are retained as 1 in the result matrix; otherwise, the result is 0. This result is the dynamic adjacency matrix for time step t. This filtering process is repeated for all time steps to obtain a sequence of dynamic adjacency matrices that accurately reflects the dynamic changes of the power grid topology over time.

[0025] Specifically, S3, based on the dynamic adjacency matrix sequence, performs graph feature aggregation under topological constraints on the analog characteristic sequence to obtain a spatial feature map containing physical topological constraints. It should be understood that in actual power system operation, the strength of mutual influence between nodes is not solely determined by the connection state, but is strictly limited by the admittance or impedance parameters of the lines. A short-distance, low-impedance, heavily loaded line provides far greater support for voltage stability than a long-distance, high-impedance, weakly connected line. Therefore, in the technical solution of this application, by introducing prior physical knowledge to quantify the substantial physical closeness between nodes, the problem that simple topological connections cannot characterize electrical strength is solved. This ensures that the feature propagation path strictly follows the energy flow trajectory defined by Kirchhoff's laws, effectively suppressing attention divergence caused by measurement noise, and improving the accuracy and interpretability of critical voltage stability assessment.

[0026] Figure 4 This is a flowchart of step S3 in the neural network-based power grid transient voltage stability calculation method according to an embodiment of this application. Figure 4 As shown, S3 includes: S31, performing symmetric Laplace normalization on the dynamic adjacency matrix sequence to obtain a normalized dynamic adjacency sequence; S32, performing physical topology-based feature transformation and aggregation on the normalized dynamic adjacency sequence and the analog feature sequence to obtain an aggregated feature tensor; S33, performing redundant feature filtering on the aggregated feature tensor to obtain a spatial feature map.

[0027] Specifically, in step S31, the dynamic adjacency matrix sequence is subjected to symmetric Laplace normalization to obtain a normalized dynamic adjacency sequence. It should be understood that the dynamic adjacency matrix sequence can accurately describe the physical trajectory of the power grid topology evolution over time, reflecting the dynamic characteristics of the system structure. However, due to the significant differences in degree (number of connected lines) among different nodes in the power system, directly using the original adjacency matrix for feature aggregation can lead to numerical explosion or gradient vanishing problems during feature propagation. Therefore, in the technical solution of this application, symmetric Laplace normalization is performed to provide a stable numerical propagation basis for the graph neural network, ensuring that the propagation ratio of features remains consistent under different topological structures, thereby improving the convergence speed and robustness of the model when dealing with real-time physical abrupt changes caused by fault clearing.

[0028] In this process, firstly, a self-loop augmentation process is performed on each time-step matrix in the dynamic adjacency matrix sequence. This involves introducing an identity matrix so that each node incorporates its own features during aggregation, resulting in an augmented adjacency matrix. Next, the degree matrix corresponding to the augmented adjacency matrix is ​​calculated. Then, the augmented adjacency matrix is ​​symmetrically scaled using the degree matrix to scale each element to the inverse space of the geometric mean of the degrees of its connected nodes, thereby eliminating the influence of differences in the number of node connections. The resulting normalized dynamic adjacency sequence is the core input for subsequent physical topology-based feature transformation and aggregation steps, determining the flow path and intensity basis of spatial features in the network.

[0029] Specifically, in step S32, a physical topology-based feature transformation and aggregation are performed on the normalized dynamic adjacency sequence and analog characteristic sequence to obtain an aggregated feature tensor. It should be understood that in existing attempts, the aggregation process of Graph Convolutional Networks (GCNs) relies solely on the binary state (0 or 1) of the adjacency matrix to characterize the connection relationships between nodes, completely ignoring the crucial differences in physical parameters in the power system, such as line admittance or impedance. Specifically, in a physical power grid, the intensity of mutual influence between nodes is not solely determined by the connection state, but is strictly limited by the line admittance or impedance parameters. A short-distance, low-impedance, heavily loaded line provides far greater support for voltage stability than a long-distance, high-impedance, weakly connected line. Existing aggregation methods treat both as equivalent, resulting in a lack of weighting of physical parameters in the feature aggregation process, making it impossible to distinguish between strongly electrically coupled neighbors and weakly electrically coupled neighbors. Furthermore, existing aggregation methods rely on purely data-driven attention mechanisms that only calculate the statistical similarity of data features. In transient processes, they are highly susceptible to interference from measurement noise or spurious oscillations, causing the model to learn false correlations that are not physically causal. This blind aggregation method, which lacks physical constraints, violates the energy flow laws determined by Kirchhoff's laws and severely limits the model's evaluation accuracy and generalization ability under complex fault conditions.

[0030] Therefore, in the technical solution of this application, a physical-data dual-stream attention aggregation mechanism based on dynamic admittance gating is constructed to inject the physical prior knowledge of the power grid into the graph feature aggregation process of the neural network, thereby ensuring that the feature propagation path strictly follows the real energy flow law of the power grid, thereby improving the evaluation accuracy, physical consistency and generalization ability of the model.

[0031] In this process, firstly, based on the normalized dynamic adjacency sequence and preset line physical parameters, a dynamic topology mask is used to extract physical parameters and calculate nonlinear strength mapping to obtain an electrical coupling strength matrix that can characterize the strength of physical connections between nodes. This step aims to introduce prior physical knowledge to quantify the substantial physical tightness between nodes, thus solving the problem that simple topology connections cannot characterize electrical strength. Specifically, the normalized dynamic adjacency sequence and the set of line physical parameters (including the magnitude of the admittance of the entire network lines) are received. The dynamic adjacency sequence is used as a mask to extract the line admittance under the current effective topology, and a learnable scaling factor is introduced to perform nonlinear mapping on the admittance values, thereby generating an electrical coupling strength matrix that can dynamically characterize the strength of physical connections between nodes. This process not only provides the model with connection relationship information but also clarifies the tightness of electrical connections between nodes, enabling the neural network to adaptively perceive the decisive influence of physical parameters on voltage stability.

[0032] Specifically, the calculation process for dynamic electrical coupling strength can be expressed by the formula: in, The electrical coupling strength matrix represents the time step. Next node With nodes The coupling strength element between them; Represents a node With nodes The admittance magnitude of the physical circuits between them; This is an indicator function. It takes a value of 1 when the condition in parentheses is met (i.e., the line is connected in the dynamic topology), and 0 otherwise, to ensure that the coupling strength of the disconnected line is zero. These are learnable physical scaling parameters used to adjust the magnitude of the influence of admittance on strength. For activation function, In time step In a dynamic adjacency matrix, nodes With nodes Whether they are connected.

[0033] Secondly, based on the electrical coupling strength matrix and the analog signal feature sequence, the hybrid attention coefficient tensor is determined. This step aims to implement a dual-stream driven strategy, injecting the physical electrical coupling strength as a bias term into the attention mechanism to correct potential biases from pure data-driven approaches and ensure that the model still follows physical laws under data noise interference. Specifically, a dual-stream parallel processing method is adopted: on one hand, feature similarity scores of the data stream are calculated through linear transformation and shared attention vectors; on the other hand, the physical flow bias term is directly constructed using the previously generated electrical coupling strength matrix and logarithmically superimposed onto the unnormalized attention score. This scheme is significant in the scenario, because when there is a very strong electrical connection (high admittance) between two nodes, even if the data features at the moment of failure exhibit temporary dissimilarity, the model will force a higher attention weight due to the physical bias, thus preventing physical laws from being overwhelmed by data noise.

[0034] Specifically, in calculating the mixed attention coefficient, firstly, the unnormalized mixed score is calculated, and then the final coefficient is generated. This process can be expressed by the formula: in, Unnormalized mixed attention score; It is a non-linear activation function; For analog quantity characteristic sequence Time Node eigenvectors; The feature transformation weight matrix; This represents the parameter vector for the attention mechanism. This represents a vector concatenation operation; This is a physical trust gating parameter used to adjust the model's dependence on physical connections; Indicates Logarithmic operation with base; To prevent the logarithm from being a tiny constant with negative infinity; Indicates exponentiation; The final generated hybrid attention coefficients.

[0035] Furthermore, using the hybrid attention coefficient tensor, anisotropic weighted summation and residual fusion are performed on the analog feature sequence to obtain the aggregated feature tensor. This step aims to use weights incorporating physical biases to perform targeted feature aggregation and address the oversmoothing problem in deep networks. Specifically, the calculated hybrid attention coefficients are used to perform weighted summation on the transformation features of neighboring nodes, achieving preferential energy flow along electrically connected paths (i.e., anisotropic aggregation), and the original features are superimposed to preserve the unique state information of each node. This step completes the final fusion of physical constraints and data features. Specifically, this process can be expressed by the following formula: in, This is the final output aggregated feature tensor; It is a non-linear activation function; Represents the set of neighboring nodes of node i; This is a learnable weight matrix.

[0036] In particular, the dynamic admittance gating and physical bias attention mechanism proposed in this application solves the core problem of the lack of physical topology awareness in traditional graph neural networks for power grid assessment, achieving a leap from topology-equivalent aggregation to electrical strength-weighted aggregation. Ultimately, it significantly improves the physical consistency and robustness of the model in complex transient processes, ensuring that the feature propagation path still strictly follows the energy flow trajectory defined by Kirchhoff's laws under topological abrupt changes caused by faults such as N-1 or Nk. It effectively suppresses attention divergence caused by measurement noise, making the assessment results not only dependent on data statistical regularities but also controlled by the physical parameters of the power grid, thereby greatly improving the accuracy and interpretability of critical voltage stability assessment.

[0037] Specifically, in step S33, redundant feature filtering is performed on the aggregated feature tensor to obtain a spatial feature map. It should be understood that although the original feature aggregation process initially integrates topological and electrical information through a physical-data dual-stream mechanism, in complex power grid transient processes, the generated aggregated feature tensor often contains redundant interference information caused by measurement noise, spurious oscillations, or non-physical correlations. Furthermore, purely data-driven aggregation can easily lead to the model learning spurious correlations that violate Kirchhoff's laws. Therefore, in the technical solution of this application, redundant feature filtering is used to effectively remove feature components in the aggregated feature tensor that have low contribution to voltage stability discrimination or negative interference. This ensures that the final generated spatial feature map can accurately depict the core evolution trend strictly constrained by physical topology and electrical parameters, solves the oversmoothing problem of deep networks, and enables the model to maintain physical consistency and robustness even in topological abrupt change scenarios.

[0038] In this process, firstly, the hybrid attention coefficients calculated in the previous steps are received and aggregated into a feature tensor. These coefficients are then used to perform a directional weighted summation of the features of neighboring nodes in the aggregated feature tensor, achieving feature convergence guided by physical bias. Next, the aggregated information is passed through a mapping layer containing a nonlinear activation function and residually superimposed with the original features of the current node to achieve redundancy filtering in the high-dimensional feature space. Through the synergistic effect of this nonlinear transformation and residual fusion, the system effectively filters spatial redundancy while preserving core electrical characteristics. The final spatial feature map is a representation matrix containing high-dimensional spatiotemporal features and physical structural information after physical constraint aggregation and redundancy filtering.

[0039] Specifically, in step S4, transient evolution trend capture and global pooling are performed on the spatial feature map to obtain a spatiotemporal feature vector. It should be understood that the transient voltage stability of the power grid is a dynamic, time-series evolution process. Although the spatial feature map generated in the preceding steps includes physical topological constraints, it is essentially a spatial feature aggregation, lacking a continuous characterization of voltage fluctuation trends over time. By capturing transient evolution trends, dynamic trajectory features of voltage stability affected by disturbances can be extracted. Global pooling aims to aggregate the local spatiotemporal features of all nodes in the network into a global representation, thereby providing the final classifier with a fixed-length feature vector that can characterize the stable state of the entire system, ensuring that the model still possesses strong generalization ability under complex fault scenarios.

[0040] In practice, firstly, a shared-weight gated recurrent unit (GRU) network is used to perform node-level parallel temporal encoding and latent state extraction on the spatial feature map to obtain the node temporal state matrix. In this process, for each node in the spatial feature map, its feature values ​​at all time steps are extracted to form a feature sequence for that node. Next, a shared-weight GRU network is used to independently and in parallel encode the temporal signals of all nodes. Here, shared weight means that the GRU network processing the temporal signals of all nodes is the same network instance, and its internal parameters are shared among all nodes. This greatly reduces the number of model parameters, enhances generalization ability, and implicitly assumes that "the transient dynamics of all nodes follow similar physical laws." The GRU network processes the input sequence of each node stepwise through its internal gating mechanisms (mainly update gates and reset gates). These gating mechanisms can selectively decide which historical information to forget and which new information to update to the current state, thereby effectively capturing long-term dependencies in the sequence, such as whether the voltage is oscillating recovery or a continuous downward trend. For each node, after processing its entire time series, the GRU network outputs a final hidden state vector, which encapsulates the dynamic evolution trend of the node within the entire time window. Finally, the final hidden state vectors of all nodes are stacked according to the node dimension to obtain the node time series state matrix.

[0041] Next, the node temporal state matrix is ​​subjected to dimensionality compression and network-wide feature encoding to obtain a spatiotemporal feature vector. During this process, the node temporal state matrix output from the previous sub-step is received. This matrix contains the condensed temporal information of each node, but its dimensionality is still related to the number of nodes, which may vary with the grid size and is not suitable for direct input into a classifier. Therefore, a global average pooling operation is performed along the node dimensions. That is, the arithmetic mean of the hidden state vectors of all nodes is calculated on each feature dimension. This operation is repeated for all feature dimensions, thereby compressing a two-dimensional node temporal state matrix into a one-dimensional vector, i.e., the spatiotemporal feature vector. In this way, information from all nodes can be aggregated, and it is robust to changes in the number of nodes, while preserving the distribution characteristics along the feature dimensions in a smooth manner.

[0042] Specifically, in step S5, a multilayer perceptron is used to perform classification probability mapping and stability discrimination on the spatiotemporal feature vector to obtain a stability assessment result. It should be understood that although the spatiotemporal feature vector fully encodes the key dynamic information of the power grid in both spatial (topological and physical coupling) and temporal (transient evolution trend) dimensions after a fault occurs, it is still a high-dimensional, continuous numerical vector and does not possess direct engineering decision-making significance. Therefore, in the technical solution of this application, a multilayer perceptron is introduced to learn the nonlinear decision boundary from high-dimensional spatiotemporal features to the classification objective of voltage stability / instability, and outputs the assessment result in probabilistic form, thereby providing dispatchers with a quantitative and reliable basis for judging the current transient voltage stability risk of the power grid.

[0043] In practice, firstly, based on a fully connected neural network with hidden layers, a high-dimensional feature projection is performed on the spatiotemporal feature vector to obtain the original logistic regression vector. That is, through multiple nonlinear transformations, the input spatiotemporal feature vector is mapped to an unnormalized score vector corresponding to the number of target categories, i.e., the original logistic regression vector. During this process, the spatiotemporal feature vector is input into a multilayer perceptron model containing one or more hidden layers. This model projects the input spatiotemporal feature vector into a low-dimensional space corresponding to the number of target categories through a series of nonlinear transformations. The output of the last hidden layer is then transformed by a linear output layer to produce a vector with the same dimension as the number of categories, i.e., the original logistic regression vector. Each element of this vector represents the unnormalized raw score of the sample belonging to the corresponding category.

[0044] Then, probability normalization and category decision are performed on the original logistic regression vector to obtain the stability assessment result. In this process, firstly, the original logistic regression vector is probability normalized to transform the original scores, which could be arbitrary real numbers, into a probability distribution such that the output value for each category is between 0 and 1, and the sum of the probabilities of all categories is 1. The most commonly used normalization method is the Softmax function. Next, the probability values ​​corresponding to all categories are compared, and the category label with the highest probability value is output as the final discrimination result. The final output stability assessment result is a discrete category label (such as "stable" / "unstable"), providing crucial basis for power grid safety analysis.

[0045] In summary, the neural network-based method for calculating transient voltage stability of power grids according to the embodiments of this application is explained. It constructs a dynamic adjacency matrix sequence by real-time analysis of circuit breaker state signals, enabling the model to accurately capture real-time changes in the power grid structure caused by events such as fault clearing. This strictly adheres to the dynamic evolution of physical connections during feature aggregation, thus achieving accurate modeling of energy flow paths during transient processes. In this way, the model's ability to capture changes in energy flow paths during transient processes and its physical consistency are enhanced, significantly improving the accuracy, robustness, and interpretability of voltage stability assessment in scenarios involving complex faults, noise interference, and dynamic topology changes. This provides more reliable intelligent support for power grid security defense and control decisions.

[0046] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for calculating the transient voltage stability of a power grid based on a neural network, characterized in that, include: S1: Perform synchronous feature mapping on the acquired raw measurement data stream and raw state data stream to generate analog characteristic sequences and digital topology sequences; S2: Construct a dynamic adjacency matrix sequence based on digital topology sequences; S3: Based on the dynamic adjacency matrix sequence, perform graph feature aggregation on the analog characteristic sequence under topological constraints to obtain a spatial feature map containing physical topological constraints; S4: Perform transient evolution trend capture and global pooling on the spatial feature map to obtain the spatiotemporal feature vector; S5: The spatiotemporal feature vectors are classified and mapped for probability and stability are determined by a multilayer perceptron to obtain stability evaluation results.

2. The method for calculating the transient voltage stability of a power grid based on a neural network according to claim 1, characterized in that, The raw measurement data stream includes the time series of voltage amplitude, voltage phase angle, active power, and reactive power at key nodes across the entire network; the raw status data stream includes the opening / closing digital signals of circuit breakers at both ends of key transmission lines across the entire network.

3. The method for calculating the transient voltage stability of a power grid based on a neural network according to claim 1, characterized in that, Step S1 includes: Time synchronization and missing value repair are performed on the original measurement data stream and the original state data stream to obtain a synchronized and aligned dataset. Based on the fault triggering time, a sliding window is used to truncate the synchronous alignment dataset to obtain time-series slices. The time-series slices are subjected to analog-digital separation and feature normalization reconstruction to obtain analog feature sequences and digital topology sequences.

4. The method for calculating the transient voltage stability of a power grid based on a neural network according to claim 1, characterized in that, Step S2 includes: Based on a pre-set power grid physical wiring diagram, the digital topology sequence is initialized with physical connection references to obtain a static reference matrix; Based on the circuit breaker switch state mapping logic, connectivity blocking is performed on the digital topology sequence and static reference matrix to obtain the real-time mask matrix; Physical topology reconstruction is performed on the static reference matrix and the real-time mask matrix to obtain a dynamic adjacency matrix sequence.

5. The method for calculating the transient voltage stability of a power grid based on a neural network according to claim 1, characterized in that, Step S3 includes: The dynamic adjacency matrix sequence is subjected to symmetric Laplace normalization to obtain a normalized dynamic adjacency sequence. A physical topology-based feature transformation and aggregation are performed on normalized dynamic adjacency sequences and analog feature sequences to obtain aggregated feature tensors. Redundant feature filtering is performed on the aggregated feature tensor to obtain a spatial feature map.

6. The method for calculating the transient voltage stability of a power grid based on a neural network according to claim 1, characterized in that, Step S4 includes: Based on a gated recurrent unit network with shared weights, node-level parallel temporal encoding and hidden state extraction are performed on the spatial feature map to obtain the node temporal state matrix. The temporal state matrix of the nodes is compressed in dimension and the whole network features are encoded to obtain the spatiotemporal feature vector.

7. The method for calculating the transient voltage stability of a power grid based on a neural network according to claim 1, characterized in that, Step S5 includes: Based on a fully connected neural network with hidden layers, high-dimensional feature projection is performed on the spatiotemporal feature vector to obtain the original logistic regression vector; The original logistic regression vector is subjected to probability normalization and category decision to obtain stability assessment results.

8. The method for calculating the transient voltage stability of a power grid based on a neural network according to claim 5, characterized in that, Physical topology-based feature transformation and aggregation are performed on normalized dynamic adjacency sequences and analog feature sequences to obtain aggregated feature tensors, including: Based on the normalized dynamic adjacency sequence and preset line physical parameters, physical parameters are extracted and nonlinear strength mapping is calculated using dynamic topology mask to obtain an electrical coupling strength matrix that can characterize the strength of physical connection between nodes. The hybrid attention coefficient tensor is determined based on the electrical coupling strength matrix and the analog characteristic sequence. By using the hybrid attention coefficient tensor, anisotropic weighted summation and residual fusion are performed on the analog feature sequence to obtain the aggregated feature tensor.

Images