Power grid transient voltage stability evaluation method and system based on interpretable deep learning

Abstract

The present application relates to the technical field of power system safety evaluation, and provides a power grid transient voltage stability evaluation method and system based on interpretable deep learning, which comprises: power grid partitioning and construction of a sample set containing node features and adjacency matrices of each region, construction of a two-stage transient voltage stability evaluation model: the first stage extracts local features of each region, and after simplifying the local features of the secondary region through average pooling, the local features of the primary region are spliced to form a feature matrix; the second stage extracts inter-regional interaction features to obtain a transient voltage stability evaluation result; the two-stage model is subjected to interpretable analysis using a graph neural network interpretability algorithm to identify key node features and edges; and the rationality of the identified key node features and edges is quantitatively verified using an interpretability index. While realizing transient voltage stability evaluation, the causes of instability are determined to provide reliable decision support for the preventive control of new power systems.

Description

Technical Field

[0001] This invention belongs to the field of power system safety assessment technology, and in particular relates to a method and system for assessing power grid transient voltage stability based on interpretable deep learning. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] The increasing proportion of power generation from new energy sources is exacerbating the uncertainty of power grid operation. The interaction between AC and DC hybrid systems weakens the grid's resilience and makes its dynamic behavior more complex. Fluctuating grid operating conditions and fault disturbances can easily trigger transient voltage stability problems, leading to new energy source disconnection, cascading faults, and even large-scale power outages. Transient voltage stability assessment is a core component in preventing such safety incidents and ensuring stable grid operation.

[0004] Traditional transient voltage stability assessment mainly relies on time-domain simulation methods, which determine stability by solving high-dimensional nonlinear differential-algebraic equations of the power system. However, this method is computationally intensive and time-consuming, failing to meet the application requirements of online real-time power grid assessment. In recent years, with the widespread deployment of synchronous phasor measurement devices, the scale of operational data available to the power system has increased significantly, laying the data foundation for deep learning-based transient voltage stability assessment methods. Existing research has applied deep learning models such as convolutional neural networks, recurrent neural networks, and graph neural networks to transient voltage stability assessment. By learning the nonlinear mapping relationship between massive power grid data and transient voltage stability, fast and high-precision stability assessment has been achieved.

[0005] However, the inherent black-box nature of deep learning models makes it impossible to explain their decision-making logic and judgment criteria. Power grid dispatchers find it difficult to trust the model's predictions, severely limiting the application of deep learning methods in practical power grid dispatching. To address the interpretability issue, existing research has introduced interpretable methods such as decision trees, attention mechanisms, and gradient-weighted activation mapping. However, these methods have limitations when processing power system graph structure data; they can only identify the impact of electrical quantities on model predictions and cannot simultaneously consider the dynamic interactions between power grid topology and nodes. Regarding interpretable methods for Graph Neural Networks (GNNs), research in the field of transient voltage stability assessment remains insufficient, and there is a lack of interpretability metrics suitable for large power grid transient voltage stability assessment, making it impossible to quantitatively verify the rationality of the interpretation results. Summary of the Invention

[0006] To address the technical problems mentioned above, this invention provides a method and system for assessing power grid transient voltage stability based on interpretable deep learning. Considering the significant local characteristics of transient voltage and its influence by certain global interactions, the proposed model employs a strategy of local feature extraction and global feature integration: the first stage extracts local features by region using Cluster-GCN, and the second stage mines inter-regional interaction features using CNN and SE-Net. The proposed feature extraction mechanism aligns with the physical characteristics of transient voltage, helping to improve model interpretability. Based on this, an interpretability algorithm using graph neural networks is employed to perform interpretability analysis of the transient voltage stability assessment model, identifying the key electrical quantities and interactions leading to instability. Interpretability indices adapted to the transient voltage stability assessment scenario are applied to quantitatively verify the rationality of the interpretation results. While achieving rapid transient voltage stability assessment, the core causes of instability can be located, and the interpretation results conform to physical mechanisms, providing reliable decision support for the prevention and control of new power systems.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of the present invention provides a method for evaluating transient voltage stability of a power grid based on interpretable deep learning, comprising: The power grid is divided into zones, and a sample set containing node features and adjacency matrices of each zone is constructed based on power flow information, fault information, and topological connectivity of each zone. This sample set is then used to train a two-stage transient voltage stability assessment model. The two-stage transient voltage stability assessment model includes: a first stage using a clustering graph convolutional network to extract local features of each zone, and simplifying the local features of secondary zones through average pooling, then concatenating them with the local features of the primary zones to form a feature matrix; based on the feature matrix, a second stage using a convolutional neural network and a compression-excitation network to extract interaction features between zones, and based on these interaction features, the transient voltage stability assessment result is obtained. Based on the transient voltage stability assessment results, the interpretability algorithm of graph neural network is used to perform interpretability analysis on the trained two-stage transient voltage stability assessment model, identify key node features and key edges, and use interpretability index to quantitatively verify the rationality of the identified key node features and key edges. After successful verification, the trained two-stage transient voltage stability evaluation model was used to evaluate transient voltage stability.

[0008] Furthermore, the step of partitioning the power grid includes: partitioning the buses in the power grid using agglomerative hierarchical clustering algorithm based on the transient voltage stability margin of each bus under different scenarios; defining the area containing the target bus for transient voltage stability assessment as the primary area and the remaining areas as secondary areas, and further dividing the secondary areas according to the hierarchical representation of the clustering results.

[0009] Furthermore, the method for calculating the transient voltage stability margin is as follows: ; Where V0 is the value of the bus voltage before the fault occurred; V min This represents the lowest bus voltage after the fault is cleared; V th The acceptable threshold for bus voltage; V end To observe the average value of the bus voltage in the last second; T max The maximum duration for which the bus voltage remains below the sustainable threshold after a fault is cleared; T end T represents the simulated observation duration. cut T is the fault clearing time. th This indicates the acceptable duration when the bus voltage is below the threshold.

[0010] Furthermore, the node features are formed by concatenating vectors constructed from static power flow information of each region with fault codes; The adjacency matrix is ​​constructed based on the topological connection relationships of each region.

[0011] Furthermore, the parameters of the two-stage transient voltage stability assessment model and the interpretability algorithm of the graph neural network are adjusted based on the interpretability index.

[0012] Furthermore, when splicing the local features of the secondary region with the local features of the primary region, regions that are geographically adjacent are selected, and the local features corresponding to the adjacent regions are spliced ​​together first.

[0013] Furthermore, the convolutional neural network moves along the feature matrix with a fixed stride, performs convolution operations position by position, extracts the interaction features between regions, and the compression-excitation network performs adaptive calibration on the channel feature response after each layer of the convolutional neural network.

[0014] Furthermore, the interpretability algorithm of the graph neural network includes: parameterizing the adjacency matrix and node features by introducing learnable edge masks and node feature masks respectively to obtain a parameterized graph; inputting the parameterized graph into a trained two-stage transient voltage stability evaluation model to obtain the modified transient voltage stability evaluation result; and updating the edge mask and node feature mask by minimizing the difference between the transient voltage stability evaluation results before and after the modification.

[0015] Furthermore, the interpretability metrics include fidelity, accuracy, sparsity, and stability.

[0016] A second aspect of the present invention provides a power grid transient voltage stability assessment system based on interpretable deep learning, comprising: The training module is configured to: partition the power grid and, based on the power flow information, fault information, and topological connectivity of each region, construct a sample set containing the node features and adjacency matrix of each region, and train a two-stage transient voltage stability assessment model; the two-stage transient voltage stability assessment model includes: in the first stage, a clustering graph convolutional network is used to extract local features of each region, and after simplifying the local features of secondary regions by average pooling, the features are concatenated with the local features of the main regions to form a feature matrix; based on the feature matrix, in the second stage, a convolutional neural network and a compression-excitation network are used to extract the interaction features between each region, and based on the interaction features between each region, the transient voltage stability assessment result is obtained; The analysis module is configured to: based on the transient voltage stability assessment results, use the interpretability algorithm of graph neural network to perform interpretability analysis on the trained two-stage transient voltage stability assessment model, identify key node features and key edges, and use interpretability indicators to quantitatively verify the rationality of the identified key node features and key edges; The application module is configured to perform transient voltage stability evaluation using a trained two-stage transient voltage stability evaluation model after successful verification.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention considers the significant local characteristics of transient voltages and their influence by certain global interactions. The proposed model employs a strategy of local feature extraction and global feature integration: in the first stage, local features are extracted by region using Cluster-GCN; in the second stage, interaction features between sub-regions are mined using CNN and SE-Net. The proposed feature extraction mechanism is consistent with the physical characteristics of transient voltages, which helps improve the interpretability of the model. Based on this, GNNExplainer is used to perform interpretability analysis of the transient voltage stability assessment model, identify the key electrical quantities and interactions leading to instability, and apply interpretability indices adapted to the transient voltage stability assessment scenario to quantitatively verify the rationality of the interpretation results. While achieving rapid transient voltage stability assessment, the core causes of instability can be located, and the interpretation results are consistent with the physical mechanism, providing reliable decision support for the prevention and control of new power systems. Attached Figure Description

[0018] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0019] Figure 1 This is a schematic diagram of the two-stage transient voltage stability evaluation model architecture of Embodiment 1 of the present invention; Figure 2This is a schematic diagram of the GNNExplainer algorithm principle in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the practical application framework of the transient voltage stability evaluation method of Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the Shandong power grid topology and analysis results according to Embodiment 1 of the present invention; Figure 5 This is a schematic diagram of the power flow distribution near busbar 64 of the Shandong power grid in Embodiment 1 of the present invention; Figure 6 This is a framework diagram of the power grid transient voltage stability assessment system based on interpretable deep learning, according to Embodiment 2 of the present invention. 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.

[0021] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0022] Example 1 This embodiment provides a method for evaluating transient voltage stability of power grids based on interpretable deep learning.

[0023] This embodiment presents a power grid transient voltage stability assessment method based on interpretable deep learning. Considering the local characteristics of transient voltage and its global interactive influence, a two-stage deep learning model adapted to the physical characteristics of transient voltage is proposed to improve interpretability. The first stage uses Cluster-GCN (a graph convolutional network) to extract local features of each region, and the second stage uses CNN (convolutional neural network) and SE-Net (a compression-excitation network) to extract the interactions between regions. GNNExplainer (an interpretability algorithm for graph neural networks) is used to perform interpretability analysis of the transient voltage stability assessment model, identifying key electrical quantities and interactions. Interpretability indices adapted to the transient voltage stability assessment scenario are applied to quantitatively verify the rationality of the interpretation, including fidelity, accuracy, sparsity, and stability. While achieving rapid transient voltage stability assessment, this method can also pinpoint the core causes of instability, providing reliable decision support for the prevention and control of new power systems.

[0024] The power grid transient voltage stability assessment method based on interpretable deep learning provided in this embodiment, such as... Figure 3 As shown, it includes the following steps: Step 1: Divide the power grid into zones, grouping buses with similar transient voltage characteristics into the same zone; define the zone that has a greater impact on the target bus for transient voltage stability assessment as the primary zone, and the remaining zones as secondary zones; construct a sample set containing node characteristics and adjacency matrices for each zone based on power flow information, fault information and topology connections of each zone.

[0025] Step 101: The power grid is divided into zones using agglomerative hierarchical clustering algorithm to group buses with similar transient voltage characteristics into the same sub-region.

[0026] First, the transient voltage stability margin of each bus is calculated under different scenarios, and the transient voltage stability margin matrix H is constructed as follows: ; Where, η ji This represents the transient voltage margin of node i in scenario j; m represents the total number of scenarios, and n represents the total number of nodes.

[0027] Then, each column vector in the H matrix can reflect the transient characteristics of the corresponding bus in m scenarios; these column vectors are used as clustering samples, totaling n; the agglomerative hierarchical clustering algorithm is used to cluster buses with similar transient characteristics into the same region.

[0028] The method for calculating the transient voltage stability margin is as follows: ; Where V0 is the value of the bus voltage before the fault occurred; V min This represents the lowest bus voltage after the fault is cleared; V th The acceptable threshold for bus voltage; V end To observe the average value of the bus voltage in the last second; T max The maximum duration for which the bus voltage is below the sustainable threshold after a fault is cleared; T end T represents the simulated observation duration. cut T is the fault clearing time. th This indicates the acceptable duration when the bus voltage is below the threshold.

[0029] Step 102: Define the region containing the target bus for transient voltage stability assessment as the primary region and the remaining regions as secondary regions; and further divide the secondary regions in detail according to the hierarchical representation of the clustering results to simplify their topology and reduce the computational load for subsequent model training.

[0030] Step 103: The adjacency matrix is ​​constructed based on the power grid topology of each region. The node features are formed by concatenating the vector constructed from the static power flow information of each region with the fault code.

[0031] For the adjacency matrix Ak , within region k, each busbar is regarded as a node of subgraph k, and the connections between busbars (including AC lines, DC lines and transformers) constitute the edges of subgraph k. If node i is connected to node j, then A k the corresponding elements at coordinates (i,j) and (j,i) in are set to 1, otherwise they are set to 0.

[0032] The power information of node i is constructed by sequentially concatenating the following features of the corresponding busbar i: the active power and reactive power of connected generators, the active power and reactive power of connected loads, and the active power and reactive power on the i side of connected lines. The size of the feature is 1×F, where F is a constant equal to the maximum number of features for all nodes. If the number of features is less than F, the remaining elements are set to 0. The fault code is a 3-dimensional vector. The first element indicates whether a fault occurs: the nodes corresponding to the busbars at both ends of the faulty line are both set to 1. The second element indicates the fault location: for the busbars at both ends of the faulty line, if the distance from the fault point to busbar i is l% (0<l<100), this element is set to 1 l / 100. The third element indicates the fault duration: if the fault lasts for t seconds, this element of the busbars at both ends of the line is set to t×10.

[0033] Step 2: Construct a two-stage transient voltage stability assessment model, as Figure 1 shown. In the first stage, Cluster-GCN is used to extract local features of each region, the local features of secondary regions are simplified through average pooling, and then concatenated with the local features of the main region to form a feature matrix. In the second stage, CNN and SE-Net are used to extract key features of the main region and the dynamic interaction effects between various regions. With the objective of minimizing the transient voltage stability assessment error, the two-stage assessment model is trained offline to obtain optimal model parameters.

[0034] Step 201: The Cluster-GCN in the first stage independently performs graph convolution operations on each sub-region (that is, the subgraph corresponding to each power grid region), and extracts local features related to transient voltage of each region, as shown below: ; where, z represents the finally obtained feature vector; f represents the graph convolution operation; L is the number of performed graph convolution operations; A k is the adjacency matrix of subgraph k; n is the number of subgraphs; X k is the node feature of subgraph k.

[0035] Main diagonal enhancement is added to the graph convolution operation to strengthen the contribution of local features, and residual connection is introduced to stabilize model training; normalization processing is performed on the adjacency matrix of each sub-region to ensure the numerical stability of feature extraction, as shown below: ; ; in, λ represents the features of subgraph k after graph convolution; W1 and W2 represent the learnable parameters of the graph convolution layer; λ is the main diagonal enhancement coefficient. The normalized adjacency matrix; diag( ) indicates extracting diagonal elements to form a diagonal matrix; E is the identity matrix; D k A represents the degree matrix of subgraph k; k Let be the original adjacency matrix of subgraph k.

[0036] Step 202: Secondary region features are simplified using average pooling, and the calculation formula is as follows: ; Among them, X k Let S be the feature matrix of subregion k, with dimension S. k ×F,S k X is the number of nodes in the secondary subgraph k, and F is the number of features; i,j k For X k The element with coordinates (i,j) in the middle; U k Secondary subgraph The F-dimensional vector obtained by average pooling the features; U j k For U k The element with coordinate (1,j) in the middle.

[0037] Step 203: When stitching the features of each region into a feature matrix, if the geographical locations of two sub-regions are close, their features are stitched together first to preserve the spatial information of the power grid.

[0038] Step 204: The second stage of information extraction is performed using CNN and SE-Net. The CNN convolutional kernel moves along the input feature matrix with a fixed stride, performing convolution operations position by position to mine the local correlation information between regions hidden in the feature matrix and extract the dynamic interaction features of each region of the power grid. SE-Net enhances the model's ability to extract interactions between regions of the power grid by adaptively calibrating the channel feature responses of the convolutional neural network. Specifically, SE-Net performs adaptive calibration of the channel feature responses after each layer of the CNN: first, it generates initial channel weights through global average pooling; then, it adaptively learns the weights through fully connected layers; finally, it multiplies these weights with the original channels to enhance high-value feature channels, strengthening the model's ability to identify interactions between key sub-regions.

[0039] First, for the m-th channel C of the convolutional layer mPerform global average pooling to generate initial weights: ; Where, α m The initial weights of the obtained channel m; q m C represents the number of features in the m-th channel; m i,j This represents the feature with coordinates (i,j) on the m-th channel.

[0040] Then, the weights of all channels are adaptively adjusted using two fully connected layers, as shown below: ; in, These are the adjusted channel weights; δ is the activation function; W a With W b These represent the learnable weight parameters of the two fully connected layers.

[0041] Finally, the learned weights are multiplied by the original feature maps of the corresponding channels to complete the channel attention weighting, as shown below: ; Among them, C 1 m This represents the m-th channel after adaptive calibration.

[0042] Ultimately, based on C 1 m By using logistic regression (softmax), the stability of the transient voltage can be determined.

[0043] Step 3: The interpretability analysis of the trained two-stage transient voltage stability assessment model is performed using the GNNExplainer graph neural network interpretability algorithm to identify key node features and key edges. Key node features reflect important electrical quantities affecting transient voltage stability, while key edges reflect important interactions between buses that affect transient voltage stability.

[0044] GNNExplainer introduces learnable edge masks and node feature masks to parameterize the adjacency matrix and node features respectively, resulting in a parameterized graph. The mask is updated by maximizing the mutual information between the original prediction and the prediction based on the parameterized graph. The closer the mask value is to 1, the more important the corresponding feature is to the model's prediction.

[0045] The goal of GNNExplainer is to identify the subgraph G that is most critical to the model's predictions. s and node feature set X s When the model is based solely on G s and X sWhen making predictions, the results are similar to the original predictions: ; Where H(Y) is the entropy of the original predicted Y; H(Y|G=G s X=X s ) is given (G) s ,X s The conditional entropy of Y under condition ).

[0046] (1) Introduce learnable masks to parameterize the graph, including edge masks M. adj and node feature mask M feat M adj and M feat These are used to parameterize the adjacency matrix A and the node eigenvector X, respectively, as follows: ; ; in, This represents the Hadamard product; δ represents the sigmoid function mapping the mask to [0,1]. This represents the parameterized adjacency matrix. This represents the parameterized node features.

[0047] (2) Subsequently, as Figure 2 As shown, the parameterized adjacency matrix and node features are input into the trained GNN model to obtain the modified prediction results. Then, the difference between the modified prediction and the original prediction is minimized by optimizing the mask. To solve this optimization problem, this difference is quantified as the loss between the original prediction and the modified prediction. The optimization process is achieved by calculating the gradient of the loss function with respect to the mask, and an iterative update is performed using a gradient-based optimization algorithm.

[0048] A feature mask close to 1 indicates that the feature is more important to the model's predictions. Node feature masks reflect the impact of power flow information and faults on transient voltage stability, including the power of generators, loads, and lines, as well as the location and duration of faults. Edge masks reflect the impact of dynamic interactions between power system buses.

[0049] Step 4: The reasonableness of the interpretation results generated by GNNExplainer is quantitatively verified using four interpretability metrics: fidelity, accuracy, sparsity, and stability. The parameters of the two-stage model and GNNExplainer are adjusted based on these interpretability metrics to obtain high-quality interpretation results.

[0050] The four interpretability indicators are calculated as follows: (1) Fidelity: This measures the importance of key features identified by the interpretability algorithm to the model's predictions, reflecting the degree to which the interpretability algorithm reproduces the model's decision-making logic. The calculation formula is as follows: ; Where N is the sample size; P i X represents the original predicted probability of sample i; i For a new sample after setting the identified key features to 0; P(X) i ) represents the predicted probability of a new sample.

[0051] A higher fidelity indicates that the key features identified by the interpretability algorithm are more important to the model's prediction, and that the interpretability algorithm can more accurately reproduce the model's decision-making logic.

[0052] (2) Accuracy: This measures the degree of agreement between the key features identified by the interpretability algorithm and the actual factors that cause transient voltage instability. The calculation steps are as follows: First, for the samples predicted to be unstable, the generator power and load power, which the model determined to be important, were selected as variables to be adjusted. Then, with the goal of minimizing the predicted probability of transient voltage instability, the variables to be adjusted are optimized, as shown below. If the identified key features are consistent with the actual factors that cause transient voltage instability, adjusting these variables will improve the transient voltage stability level and decrease the predicted probability of instability: ; Among them; X Gen and X Load P represents the power of the generator and the load to be adjusted, respectively. unstable This represents the predicted probability of transient voltage instability; grid operation constraints include: power flow constraints, generator output constraints, load power constraints, static voltage constraints, etc. Finally, the change in transient voltage stability margin before and after the adjustment is calculated to obtain the accuracy index of the interpretation: ; Where N is the number of samples; η i,1 With η i,2 These represent the transient voltage stability margins of sample i before and after adjustment.

[0053] A smaller transient voltage stability margin indicates a more stable transient voltage after adjustment. This invention uses the particle swarm optimization algorithm to solve this optimization problem.

[0054] (3) Sparsity: Used to measure the proportion of key features identified by the interpretability algorithm to the total number of features, and to measure the clarity and conciseness of the interpretation. The calculation formula is: ; Where N is the number of samples; m i M represents the number of features in sample i that are identified as key features by the interpretability algorithm; i Let i be the total number of features of sample i.

[0055] High sparsity indicates concise interpretation results, helping dispatchers quickly focus on the core causes of instability. However, higher sparsity is not always better; high sparsity may indicate that important features are being overlooked. Therefore, while ensuring high fidelity and high accuracy, higher sparsity is preferable.

[0056] (4) Stability: Used to measure the numerical stability of the interpretation results, the calculation formula is: ; Where N is the number of samples; P is the number of small perturbations applied to a single sample; E i This is the original interpretation result for sample i; E i,p The sample after the Pth perturbation The interpretation result; D is the L1 norm distance.

[0057] Step 5: Online Application. After successful verification, the trained two-stage transient voltage stability assessment model is used to evaluate transient voltage stability.

[0058] (1) Data input.

[0059] Power grid operation data and zoning results: Real-time collection of current power grid operation data, combined with zoning results from the offline phase, serves as input for online analysis.

[0060] (2) Feature construction.

[0061] Extract features from real-time running data that are consistent with those from offline training to ensure consistency of model input.

[0062] (3) Judgment of transient voltage stability.

[0063] The offline-trained two-stage transient voltage stability assessment model is invoked to make a real-time judgment on the current power grid state. If the transient voltage is determined to be stable: the process ends, but monitoring continues. If the voltage is determined to be transiently unstable, proceed to the subsequent analysis and control phase.

[0064] (4) The GNNExplainer algorithm is used to perform interpretability analysis on the instability judgment and locate the root cause of the problem.

[0065] (5) Key electrical quantities and interactive effects that lead to instability: Identify which nodes / lines have abnormal power and other indicators, and the coupling effects between them, to accurately locate the cause of instability.

[0066] (6) Formulating prevention and control measures.

[0067] Based on key influencing factors, develop targeted prevention and control strategies (such as load shedding, reactive power compensation, and unit adjustment) to intervene in advance and avoid transient voltage instability.

[0068] In this embodiment, a simplified Shandong power grid example is used to verify the effectiveness of the proposed method. This example includes 200 AC lines, 4 DC lines, and 161 buses. First, a power grid partition sample is constructed. Then, the transient voltage stability of the target nodes in the Shandong power grid is evaluated using the proposed two-stage model. Next, the interpretability of the two-stage model is analyzed using the GNNExplainer algorithm. Finally, the interpretability index is used to quantify and verify the rationality of the interpretation results. The results are shown below: (1) Power grid partitioning and sample construction.

[0069] Figure 4 The topology and zoning results of the Shandong power grid are presented. Region 1 was selected as the target region for transient voltage stability assessment because it contains 3 DC lines and has relatively complex transient voltage characteristics. Bus 64 was selected as the target bus for transient voltage stability because it has a heavy load and is close to DC lines, and its transient voltage instability is closely related to DC commutation failure.

[0070] Training samples were generated using time-domain simulation. The proportion of renewable energy generation was 41.32%. Power generation and load fluctuated randomly within the original 40% range. Four candidate fault lines were identified through a series of experiments: lines 35-65, 59-66, 60-76, and 13-26. In each simulation, three lines were randomly selected to experience a three-phase short circuit. The fault occurred at 1.0 second, and the duration was randomly set to 0.1 or 0.2 seconds. The three fault lines were then immediately disconnected. The fault location was set at 25%, 50%, or 75% of the selected lines. If the transient voltage of bus 64 remained below 0.9 pu for more than 1 second after the fault was cleared, the bus was considered transiently unstable. The total dataset size was 14,400 samples, of which 1,064 samples were marked as stable. All samples were randomly divided into training, validation, and test sets in an 8:1:1 ratio.

[0071] (2) Results of transient voltage stability assessment.

[0072] The parameters of the two-stage transient voltage stability evaluation model are as follows. Cluster-GCN consists of four layers: the number of output channels in each layer is set to 40, 20, 15, and 10, respectively; the λ parameter for each layer is set to 1. In the readout layer, the output size after feature concatenation is 46×10. CNN contains three layers: the number of convolutional kernels in each convolutional layer is set to 8, 4, and 2, respectively; the corresponding kernel sizes are 6×2, 6×2, and 5×4, respectively; the padding parameter and stride for all convolutional layers are set to 1 and 2, respectively. For the SE-Net module, the number of neurons in each fully connected layer is set to 16, 8, and 4, respectively.

[0073] The performance of the proposed two-stage transient voltage stability assessment model was tested on a simplified example of the Shandong power grid and compared with the traditional model.

[0074] Table 1. Comparison of performance of different transient voltage stability evaluation models

[0075] As shown in Table 1, the two-stage feature extraction method proposed in this embodiment outperforms other traditional models. Global information aggregation by GCN cannot effectively extract the local characteristics of transient voltage. CNN, limited by its ability to process non-Euclidean data, cannot capture the topological information of the power system, resulting in low performance; the network structure of ANN (Artificial Neural Network) is too simple, making it difficult to accurately extract key features of transient voltage. The proposed method adopts a two-stage feature extraction architecture, integrating Cluster-GCN, CNN, and SE-Net, which can effectively capture key local features of the main region and extract the interaction relationships between different partitions, thus improving model performance.

[0076] (3) Results of model interpretability analysis.

[0077] The GNNExplainer algorithm is used to locate important node features and edges. Taking an unstable sample as an example, the top six important node features and edges are selected and displayed. As shown in Table 2, the faulty lines in this sample include: lines 35-65, lines 59-66, and lines 60-76.

[0078] Table 2. Key power grid features and masks for GNNExplainer positioning

[0079] As shown in Table 2, synchronous generators 313 and 320 are the two most important generators for model prediction. Figure 5 As shown, the two generators provide a total of 11.82 Mvar of reactive power, accounting for 77% of the total reactive power output of generators in the region. Their power support capability is crucial for maintaining the stability of the system's transient voltage.

[0080] The load power connected to buses 62, 63, and 64 is also important for model prediction. The significant reactive power demand of buses 62 and 63 prevents generator 320 from providing sufficient power support to bus 64. The reactive power of lines 60-64 and 64-65 is also a key feature, as these two lines are the only reactive power transmission path for bus 64. Figure 5 As shown, they transmit 3.10 Mvar and 5.32 Mvar of reactive power, respectively. The reactive power of bus 64 cannot be locally balanced, becoming one of the key factors causing transient voltage instability.

[0081] When faulty lines 35-65 and 60-76 disconnect, the system topology changes, and the power sources on the east and west sides of the power system cannot provide sufficient support to bus 64 through lines 60-64 and 64-65. Simultaneously, the model identifies the active power of lines 201-61 and 61-64 as a key characteristic. From a physical perspective, most of the active power in this area is transmitted via UHVDC; the fault triggers commutation failure, leading to a sharp decrease in active power support. This further exacerbates the transient voltage instability of bus 64. The side masks output by the model reveal the bidirectional interaction between buses in the power system: the lines connected to bus 64 have large side masks in both directions, indicating that the transient voltage stability of bus 64 strongly depends on bidirectional power exchange and electrical coupling with adjacent buses; the lines from bus 201 to bus 61 have large edge masks, reflecting the significant impact of DC commutation failure on the AC system.

[0082] The above analysis shows that the key features identified by GNNExplainer are basically consistent with the real factors that cause transient voltage instability.

[0083] (4) Calculate the interpretability index, quantitatively verify the rationality of the interpretation results, and compare them with other traditional methods, as shown in Table 3.

[0084] Table 3. Comparison of interpretability indicators for different methods

[0085] As shown in Table 3, comparing the first two methods, the accuracy of the proposed method's explanation is only about 3% lower than that using GCN (Graph Convolutional Network) and GNNExplainer. However, in terms of sparsity, the proposed method locates 11% fewer key features. This indicates that the features additionally focused on by GCN contribute very little to transient voltage stability. The two-stage feature extraction method proposed in this embodiment can effectively extract local features of transient voltage and focus on the core causal relationships in the main regions, thus providing both accuracy and simplicity in the explanation.

[0086] The interpretability metrics of the method proposed in this embodiment are superior to the other two methods. CNNs, by converting electrical quantities into grid data, disrupt the inherent spatial dependencies between buses. Grad-CAM (Gradient Weighted Class Activation Heatmap) typically uses the last convolutional layer for interpretation, scaling it to the original image size through interpolation, resulting in coarse interpretation granularity. ANNs, with their simple network architecture, cannot accurately capture the high-dimensional nonlinear relationships of power systems. Furthermore, the feature independence assumption of SHAP (Shapley Additive Interpretation) contradicts the strong correlation of power system features, easily generating misleading interpretations. In contrast, the two-stage feature extraction method in this embodiment effectively extracts local features and dynamic interaction effects of transient voltages, and GNNExplainer effectively integrates node features with the power grid topology for interpretation, thus improving interpretability.

[0087] The power grid transient voltage stability assessment method based on interpretable deep learning provided in this embodiment takes into account that transient voltage has significant local characteristics and is affected by certain global interactions. The proposed model adopts a strategy of local feature extraction and global feature integration: in the first stage, local features are extracted by dividing the region through Cluster-GCN, and in the second stage, interactive features between sub-regions are mined by CNN and SE-Net. The proposed feature extraction mechanism is consistent with the physical characteristics of transient voltage and helps to improve the interpretability of the model.

[0088] The power grid transient voltage stability assessment method based on interpretable deep learning provided in this embodiment uses the GNNExplainer algorithm to realize the interpretability analysis of the two-stage model. It can simultaneously identify the key node features and key edges that affect the model decision, which correspond to the key electrical quantities in the power grid and the dynamic interactions between nodes, respectively, providing an explanation for the causes of transient voltage instability.

[0089] The power grid transient voltage stability assessment method based on interpretable deep learning provided in this embodiment uses interpretability indicators to quantitatively verify the rationality of the interpretation results, including four indicators: fidelity, accuracy, sparsity, and stability, which further enhances dispatchers' trust in the transient voltage stability assessment model.

[0090] The power grid transient voltage stability assessment method based on interpretable deep learning provided in this embodiment can realize rapid assessment of transient voltage stability in large power grids and location of instability causes, providing a decision-making basis for dispatchers to implement preventive and control measures, and has engineering application value.

[0091] Example 2 The power grid transient voltage stability assessment system based on interpretable deep learning provided in this embodiment, such as... Figure 6 As shown, it includes: The training module is configured to: partition the power grid and construct a sample set containing node features and adjacency matrices of each region based on power flow information, fault information, and topological connectivity of each region; and train a two-stage transient voltage stability assessment model. The two-stage transient voltage stability assessment model includes: in the first stage, a clustering graph convolutional network is used to extract local features of each region, and after simplifying the local features of secondary regions by average pooling, the features are concatenated with the local features of the main regions to form a feature matrix; in the second stage, a convolutional neural network and a compression-excitation network are used to extract the interaction features between regions; with the goal of minimizing the transient voltage stability assessment error, the two-stage assessment model is trained offline to obtain the optimal model parameters.

[0092] The training module includes: The data acquisition module is configured to: collect real-time operating data from the power grid data acquisition and monitoring system, combine it with power grid dispatching plans, renewable energy power generation forecast data and load forecast data, and obtain accurate current and future operating scenarios of the power grid through state estimation; The sample construction module is configured to: construct the node characteristics and adjacency matrix of each sub-region based on the power system partitioning results and the current and future operation scenarios of the power grid, and generate the sample set required to evaluate transient voltage stability; The transient voltage stability assessment module is configured to: call the trained two-stage transient voltage stability assessment model, perform a fast transient voltage stability assessment on the constructed samples, and output the assessment results of stability / instability.

[0093] The interpretability analysis module (hereinafter referred to as the analysis module) is configured to: when the evaluation result determines transient voltage instability, based on the transient voltage stability evaluation result, use a graph neural network interpretability algorithm to perform interpretability analysis on the trained two-stage transient voltage stability evaluation model, identify key node features and key edges (i.e., locate the key features affecting instability, including key electrical quantities and key interactions that trigger transient voltage instability), and use interpretability indicators to quantitatively verify the rationality of the identified key node features and key edges. Based on the interpretability indicators, the parameters of the two-stage model and the GNNExplainer algorithm are adjusted to obtain high-quality interpretation results.

[0094] The application module is configured to: after successful verification, use the trained two-stage transient voltage stability assessment model to evaluate transient voltage stability and obtain the result indicating whether the transient voltage is stable. The GNNExplainer algorithm is used to identify the key electrical quantities and interactions leading to transient voltage instability.

[0095] It should be noted that each module in this embodiment corresponds one-to-one with each step in Embodiment 1, and their specific implementation processes are the same, so they will not be repeated here.

[0096] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for evaluating transient voltage stability of power grids based on interpretable deep learning, characterized in that, include: The power grid is divided into zones, and a sample set containing node features and adjacency matrices of each zone is constructed based on power flow information, fault information, and topological connectivity of each zone. This sample set is then used to train a two-stage transient voltage stability assessment model. The two-stage transient voltage stability assessment model includes: a first stage using a clustering graph convolutional network to extract local features of each zone, and simplifying the local features of secondary zones through average pooling, then concatenating them with the local features of the primary zones to form a feature matrix; based on the feature matrix, a second stage using a convolutional neural network and a compression-excitation network to extract interaction features between zones, and based on these interaction features, the transient voltage stability assessment result is obtained. Based on the transient voltage stability assessment results, the interpretability algorithm of graph neural network is used to perform interpretability analysis on the trained two-stage transient voltage stability assessment model, identify key node features and key edges, and use interpretability index to quantitatively verify the rationality of the identified key node features and key edges. After successful verification, the trained two-stage transient voltage stability evaluation model was used to evaluate transient voltage stability.

2. The power grid transient voltage stability assessment method based on interpretable deep learning as described in claim 1, characterized in that, The steps for partitioning the power grid include: partitioning the buses in the power grid using agglomerative hierarchical clustering algorithm based on the transient voltage stability margin of each bus under different scenarios; defining the area containing the target bus for transient voltage stability assessment as the primary area, defining the remaining areas as secondary areas, and further dividing the secondary areas according to the hierarchical representation of the clustering results.

3. The power grid transient voltage stability assessment method based on interpretable deep learning as described in claim 2, characterized in that, The method for calculating the transient voltage stability margin is as follows: ； Where V0 is the value of the bus voltage before the fault occurred; V min This represents the lowest bus voltage after the fault is cleared; V th The acceptable threshold for bus voltage; V end To observe the average value of the bus voltage in the last second; T max The maximum duration for which the bus voltage is below the sustainable threshold after a fault is cleared; T end T represents the simulated observation duration. cut T is the fault clearing time. th This indicates the acceptable duration when the bus voltage is below the threshold.

4. The power grid transient voltage stability assessment method based on interpretable deep learning as described in claim 1, characterized in that, The node features are constructed by concatenating vectors built from static power flow information of each region with fault codes; The adjacency matrix is ​​constructed based on the topological connection relationships of each region.

5. The power grid transient voltage stability assessment method based on interpretable deep learning as described in claim 1, characterized in that, The parameters of the two-stage transient voltage stability assessment model and the interpretability algorithm of the graph neural network were adjusted based on the interpretability index.

6. The power grid transient voltage stability assessment method based on interpretable deep learning as described in claim 1, characterized in that, When splicing local features of secondary regions with local features of primary regions, regions that are geographically close to each other are selected, and the local features corresponding to the neighboring regions are spliced ​​together first.

7. The power grid transient voltage stability assessment method based on interpretable deep learning as described in claim 1, characterized in that, The convolutional neural network moves along the feature matrix with a fixed stride, performs convolution operations position by position, extracts the interaction features between regions, and the compression-excitation network performs adaptive calibration on the channel feature response after each layer of the convolutional neural network.

8. The power grid transient voltage stability assessment method based on interpretable deep learning as described in claim 1, characterized in that, The interpretability metrics include fidelity, accuracy, sparsity, and stability.

9. The power grid transient voltage stability assessment method based on interpretable deep learning as described in claim 1, characterized in that, The interpretability algorithm of the graph neural network includes: parameterizing the adjacency matrix and node features by introducing learnable edge masks and node feature masks to obtain a parameterized graph; inputting the parameterized graph into a trained two-stage transient voltage stability evaluation model to obtain the modified transient voltage stability evaluation result; and updating the edge mask and node feature mask by minimizing the difference between the transient voltage stability evaluation results before and after the modification.

10. A power grid transient voltage stability assessment system based on interpretable deep learning, characterized in that, include: The training module is configured to: partition the power grid and, based on the power flow information, fault information, and topological connectivity of each region, construct a sample set containing the node features and adjacency matrix of each region, and train a two-stage transient voltage stability assessment model; the two-stage transient voltage stability assessment model includes: in the first stage, a clustering graph convolutional network is used to extract local features of each region, and after simplifying the local features of secondary regions by average pooling, the features are concatenated with the local features of the main regions to form a feature matrix; based on the feature matrix, in the second stage, a convolutional neural network and a compression-excitation network are used to extract the interaction features between each region, and based on the interaction features between each region, the transient voltage stability assessment result is obtained; The analysis module is configured to: based on the transient voltage stability assessment results, use the interpretability algorithm of graph neural network to perform interpretability analysis on the trained two-stage transient voltage stability assessment model, identify key node features and key edges, and use interpretability indicators to quantitatively verify the rationality of the identified key node features and key edges; The application module is configured to perform transient voltage stability evaluation using a trained two-stage transient voltage stability evaluation model after successful verification.

Images