Adaptive multi-scale transient voltage stability assessment method based on improved SK-BiLSTM

Abstract

An adaptive multi-scale transient voltage stability assessment method based on an improved SK-BiLSTM (Simultaneous Linear Time-Series Least Squares Memory) architecture first considers the impact of indirect connections between nodes in the power grid topology on transient voltage stability. A graph attention network is then used to combine time-series electrical measurement data with graph-structured topological data to form a spatiotemporal fusion dataset. Based on this, an adaptive multi-scale feature extraction model based on the improved SK-BiLSTM architecture is constructed. Through progressive grouped convolution operations, this model comprehensively extracts transient voltage feature information and improves the model's response speed. Validation using modified IEEE 39-node and IEEE 118-node test systems demonstrates that this method can quickly and accurately assess voltage stability. The model's performance in identifying voltage-instability nodes or regions is verified using a defined full coverage metric, further demonstrating the accuracy, effectiveness, and speed of the method.

Description

Technical Field

[0001] The present invention relates to the technical field of transient voltage stability assessment of power systems, and in particular to an adaptive multi-scale transient voltage stability assessment method based on SK-BiLSTM. Background Art

[0002] With the expansion of power system scale and the complexity of operating conditions, the issue of transient voltage stability has become increasingly important. However, existing assessment methods have shortcomings in dealing with the complexity and dynamics of power grids, mainly reflected in the difficulty in ensuring the accuracy and real-time performance of the assessment. Traditional methods such as direct methods and transient energy function methods are powerless when faced with the dynamic changes of large-scale power grids and find it difficult to effectively judge voltage instability. Although deep learning methods have shown potential in recent research, they have difficulty in fully capturing the global and local information of complex transient signals, and the model response speed is insufficient, which affects the assessment effect. Voltage instability may lead to system collapse and large-scale power outages, causing serious economic losses and social impacts. Therefore, there is an urgent need to develop a fast and accurate assessment method that can adaptively capture multi-scale information to improve the ability to identify voltage instability nodes and areas.

[0003] In terms of feature extraction methods, current research mostly uses single-scale convolutional operations (SSCO), which can only extract features at a fixed scale and cannot fully capture information at different scales. This limitation makes it difficult for the model to effectively capture the comprehensive characteristics of global and local information when processing complex transient signals, thus affecting the evaluation effect. Although a small number of studies have used multi-scale convolutional operations (MSCO), they have failed to fully consider the impact of indirect connections between nodes in the power grid topology on transient voltage stability, and there are problems with computational efficiency. In addition, there is a lack of an effective indicator to evaluate the performance of the voltage instability node or area division. Summary of the Invention

[0004] To overcome the shortcomings of existing feature extraction methods, such as the inability to simultaneously extract both global and local information, the failure to fully consider the impact of indirect connections between nodes in the grid topology on transient voltage stability, and insufficient computational efficiency, this paper proposes an adaptive multi-scale transient voltage stability assessment method based on an improved SK-BiLSTM. This method achieves precise delineation of transient voltage stability assessment and voltage instability nodes or regions, improving the accuracy and reliability of the assessment while also ensuring rapid model response.

[0005] The technical solution adopted by the present invention is:

[0006] The adaptive multi-scale transient voltage stability assessment method based on the improved SK-BiLSTM includes the following steps:

[0007] Step 1: Considering the impact of indirect connections between nodes in the power grid topology on transient voltage stability, a graph attention network is used to combine the time-series electrical measurement data features with the graph structure topology data features to form spatiotemporal fusion data features;

[0008] Step 2: Build an adaptive multi-scale assessment model based on the improved SK-BiLSTM. This model introduces a progressive group convolution operation mechanism into the Selective Kernel Convolutional Network (SKNet) and combines it with a Bidirectional Long Short-Term Memory Network (BiLSTM). It also performs a weighted fusion of Focal Loss and Dice Loss to obtain a weighted fusion loss function. This allows the model to leverage the advantages of both methods, thereby improving its performance in transient voltage stability assessment.

[0009] Step 3: Evaluate the transient voltage stability of the power system using the adaptive multi-scale evaluation model based on the improved SK-BiLSTM constructed in Step 2. Verify the model's performance in identifying voltage instability nodes or areas using the defined full coverage index.

[0010] In step 1, a comprehensive study of the transient voltage stability problem is conducted. Considering the sampling characteristics of the time-series electrical measurement data and the sampling characteristics of the graph structure topology data in actual engineering, a comprehensive and integrated time-series electrical measurement data feature and graph structure topology data feature are constructed, and the two are fused through the graph attention network to form a spatial fusion data feature.

[0011] In step 1, among the time series electrical measurement data features, the features related to the generator include G ω , G δ , G Δf , G pm , G P , G Q , which represent the angular velocity, power angle, frequency deviation, mechanical power, active power, and reactive power of different generators respectively; the bus-related characteristics include B u 、B θ , respectively represent the voltage amplitude and phase angle of all buses; the line-related characteristics include C P 、C Q , representing the active power and reactive power of all lines respectively; in the graph structure topology data characteristics, the node degree and the electrical distance between nodes in the electrical system graph structure topology are selected.

[0012] In step 1, transient voltage stability is judged using a practical engineering criterion: the bus node voltage is less than 0.75 pu for no more than 1 second after a fault. Transient voltage stability assessment is defined as a main classification task. In the main classification task, partitioning is not considered. If one of the monitored buses meets the voltage instability criterion, the system is judged to be unstable and the corresponding category label is 0; otherwise, it is judged to be stable and the corresponding category label is 1.

[0013] The classification of voltage instability nodes or regions is defined as a sub-classification task. In this sub-classification task, the transient voltage stability state of each node is determined by considering the partitions. Nodes that meet the criteria are labeled 1, while nodes that do not meet the criteria are labeled 0. If the number of nodes in the system is m, an m-dimensional label is obtained.

[0014] In step 1, since the power system physical quantities are time-series electrical quantities and graph-structured topological quantities, a Graph Attention Network (GAT) is used to fuse the two. As a deep learning model that effectively processes graph-structured topological data, the Graph Attention Network (GAT) uses an attention mechanism to assign different weights to each node when aggregating neighboring node information, dynamically adjusting the influence of different neighboring nodes on the target node. This effectively accounts for the impact of indirect connections between nodes in the power grid topology on transient voltage stability.

[0015] In step 1, a Graph Attention Network (GAT) is used to fuse the graph structure topology data features and time series electrical measurement data features of the power system. The details are as follows:

[0016] The initial feature vector h of node i can be i It is represented by the combination of time series electrical measurement data and graph topology data. Assume that the time series electrical measurement data is x i , the graph topology data is y i , then the initial eigenvector h of node i i As shown in formula (1).

[0017] h i =[x i ||y i ] (1)

[0018] In the above formula, x i Represents time series electrical measurement data; y i Represents graph topology data; || represents vector connection operation; h i Represents the initial feature vector of node i.

[0019] Next, the initial feature vector h of each node i Perform linear transformation and obtain new features as shown in formula (2).

[0020] h′ i =Wh i (2)

[0021] In the above formula, W represents a trainable weight matrix used to map different types of data into a unified feature space; h i represents the initial feature vector of node i; h′ i Represents the vector obtained by linear transformation of the initial eigenvector of node i.

[0022] Afterwards, the attention coefficient between node i and its neighbor node j is calculated as shown in formula (3).

[0023] e ij =LeakyReLU(a T [h′ i ||h′ j ]) (3)

[0024] In the above formula, e ij represents the attention coefficient between node i and its neighbor node j; a represents a trainable weight matrix vector; LeakyReLU represents an activation function used to introduce nonlinearity; [h′ i ||h′ j ] represents the concatenation of the feature vectors of node i and node j.

[0025] The attention coefficients of all neighbor nodes are normalized using the SoftMax function, as shown in Equation (4).

[0026]

[0027] In the above formula, α ij represents the normalized attention coefficient; e ij represents the attention coefficient between node i and its neighbor node j; exp(e ij ) represents the exponential value of the attention coefficient; N i Represents the set of neighbor nodes of node i.

[0028] Finally, the normalized attention coefficient is used to perform weighted summation on the features of neighboring nodes to obtain the new features of node i, as shown in formula (5).

[0029]

[0030] In the above formula, h″ i represents the new feature vector of node i; σ represents the nonlinear activation function; αij represents the normalized attention coefficient; h′ j Represents the feature vector of neighbor node j.

[0031] In step 2, an adaptive multi-scale assessment model based on the improved SK-BiLSTM is constructed, transient voltage stability assessment is defined as the primary classification task, and voltage instability node or area division is defined as the secondary classification task.

[0032] Firstly, the selective kernel convolutional network (SKNet) with progressive group convolution operation is used to preliminarily process the spatiotemporal fusion data features.

[0033] Then, the extracted data features are input into the Bidirectional Long Short-Term Memory Network (Bi-LSTM), which uses its bidirectional time series data processing capabilities to improve the efficiency and accuracy of feature extraction;

[0034] Finally, the data features transmitted by the bidirectional long short-term memory network (Bi-LSTM) are input into the classifiers of the main classification task and the secondary classification task to obtain the evaluation results.

[0035] Through this process, information sharing between the main classification task and the secondary classification task is achieved, as well as the discrimination of the voltage instability node or area division performance of the unstable samples.

[0036] In step 2, feature extraction is achieved using a progressive grouped convolution operation mechanism:

[0037] Traditional convolution operations are insufficient in extracting global and local information and are also slow in response speed. By introducing a progressive grouped convolution mechanism, we can simultaneously extract global and local information and improve the response speed of the model. The progressive grouped convolution operation is shown in formula (7):

[0038]

[0039] In the above formula, G p Indicates the number of groups in the current group volume; G base Indicates the number of basic group convolution groups; epoch indicates the current training round; max_epoch indicates the maximum training round; n is the group adjustment frequency.

[0040] In step 2, the loss function uses a weighted fusion of Focal Loss and Dice Loss. Focal Loss effectively addresses class imbalance, while Dice Loss optimizes classification decision accuracy. By leveraging the strengths of both, the model not only addresses class imbalance but also improves the accuracy and reliability of transient voltage stability assessments, further enhancing the ability to identify and delineate nodes or regions of voltage instability.

[0041] In step 2, the weighted fusion loss function of Focal Loss and Dice Loss is calculated as shown in the following formula (14).

[0042] FL-DL(p)=λFL(p)+(1-λ)DL(p) (14);

[0043] In the above formula, p represents the model's predicted probability for the stable / instability category; λ represents the weight parameter; FL(p) represents Focal Loss; DL(p) represents Dice Loss.

[0044] Focal Loss is used to deal with the problem of class imbalance, and its mathematical expression is shown in formula (15).

[0045] FL(p)=-α(1-p) γ log(p) (15);

[0046] In the above formula, p represents the probability of sample prediction; α represents the balance factor; γ represents the focusing factor; FL(p) represents the Focal Loss.

[0047] Dice Loss is used to optimize the accuracy of classification decisions, and its mathematical expression is shown in formula (16).

[0048]

[0049] In the above formula, p i represents the predicted value; g i represents the real value; ε is a small constant to prevent the denominator from being zero.

[0050] In step 3, the adaptive multi-scale assessment model based on the improved SK-BiLSTM was applied to evaluate the transient voltage stability of the power system. The model's performance in identifying voltage instability nodes or areas was verified using the defined full coverage metric, further demonstrating the model's accuracy, effectiveness, and speed in practical applications.

[0051] In step 3, the performance of the model in the voltage instability node or area partitioning task is measured by defining the Full Coverage Rate (FCR). The mathematical expression of the Full Coverage Rate is shown in the following formula (20).

[0052]

[0053] In the above formula, N1 is the total number of unstable samples evaluated in the main classification task, y ij is the jth true label of the i-th sample, is the jth predicted label of the i-th sample, is an indicator function, which is 1 when all unstable nodes of sample i are correctly predicted, and 0 otherwise.

[0054] In step 3, the transient voltage stability assessment process includes the following steps:

[0055] 1) Construction of sample set of time series data: The sampling step length T of the present invention is set to 0.01s to ensure high accuracy of data. A time window length of 0.4s is used, and the time interval t is set according to different fault removal times. c , extract the time segment of the corresponding length of the original data, the specific sampling interval is [t c -28T,t c +12T];

[0056] 2) Sample set construction of spatial data: The present invention selects node degree and electrical distance between nodes as the topological structure data of the power system diagram;

[0057] 3) Data preprocessing: To facilitate model training and evaluation, Z-score normalization is used to normalize the data input into the model so that all data are in the same dimension and the negative impact of dimensional differences is eliminated. The mathematical expression of Z-score normalization is shown in formula (21).

[0058]

[0059] In the above formula: X * Represents the value after data normalization; μ represents the mean of the data in the sample; σ is the standard deviation of the sample data; ε represents a minimum value; X is the data in the data feature set.

[0060] 4) Model Parameter Selection and Training: Model parameters were adjusted based on accuracy requirements, and the FL-DL loss function was adopted. During model training, the Adam optimizer and Sigmoid classifier were used, and batch processing was employed to accelerate training. The initial learning rate was set to 0.001. To prevent overfitting, the learning rate was halved after the loss value remained stable for five consecutive rounds. Training continued until comprehensive evaluation metrics met requirements and the loss function fluctuations were less than the set value. The model was then retained; otherwise, the model parameters were adjusted and retrained. This entire process ensured the final model parameters.

[0061] 5) Evaluation results: After training, the test set data is passed into the model, and the model performance is comprehensively evaluated based on the comprehensive evaluation indicators output by the model.

[0062] The present invention provides an adaptive multi-scale transient voltage stability assessment method based on an improved SK-BiLSTM, and the technical effects are as follows:

[0063] 1) The method of the present invention can effectively capture the spatiotemporal characteristics of different scales in the power system by introducing the multi-scale convolution operation (MSCO) of the selective kernel convolution network (SKNet) with a progressive grouped convolution mechanism, improve the ability to identify complex voltage transient changes, and achieve accurate division of voltage instability nodes or areas.

[0064] 2) This invention not only intuitively reflects the performance of power system transient voltage stability assessment through the confusion matrix-derived indicators of the main classification task, but also reflects the performance of the model's voltage instability node / region classification through the FCR indicator. Furthermore, the model's rapid response speed can better assist operators in analyzing and assessing power system operational status, providing decision support.

[0065] 3) The adaptive multi-scale transient voltage stability assessment method based on the improved SK-BiLSTM proposed in this paper comprehensively considers the impact of indirect connections between nodes in the power grid topology on transient voltage stability. Using a graph attention network, it combines the characteristics of time-series electrical measurement data with the characteristics of graph structure topology data. This method comprehensively considers the combined effects of temporal and spatial characteristics on voltage stability, achieving transient voltage stability assessment and precise delineation of voltage-instability nodes or regions. By comparing different sub-convolution operation group strategies, the method improves the accuracy and reliability of the assessment while also taking into account the rapidity of the model's response speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] The present invention will be further described below with reference to the accompanying drawings and examples:

[0067] Figure 1 This is the voltage amplitude curve when the system is stable.

[0068] Figure 2 This is the voltage amplitude curve when the system is unstable.

[0069] Figure 3 Schematic diagram of the structure of the adaptive multi-scale transient voltage stability assessment model based on the improved SK-BiLSTM.

[0070] Figure 4 This is the structure diagram of the Selective Kernel Convolutional Network (SKNet) that introduces the progressive group convolution mechanism.

[0071] Figure 5 Schematic diagram of the BiLSTM structure.

[0072] Figure 6 This is the overall flow chart of adaptive multi-scale transient voltage stability assessment based on improved SK-BiLSTM.

[0073] Figure 7 Schematic diagram of the IEEE 39-bus power system topology with wind turbines.

[0074] Figure 8 It is a comprehensive performance evaluation index for models with different loss functions.

[0075] Figure 9 This is the waveform of the loss value of the training set.

[0076] Figure 10 This is the waveform of the test set loss value.

[0077] Figure 11 This is the t-SEN visualization effect diagram of spatiotemporal fusion data.

[0078] Figure 12 This is the t-SEN visualization effect diagram of F1 data.

[0079] Figure 13 This is the t-SEN visualization effect diagram of Y1 data.

[0080] Figure 14 This is the t-SEN visualization effect diagram of F2 data.

[0081] Figure 15 This is the t-SEN visualization effect diagram of Y2 data.

[0082] Figure 16 This is the t-SEN visualization effect diagram of BiLSTM layer data.

[0083] Figure 17 Contribution of time-series electrical measurement data features to transient voltage stability assessment tasks.

[0084] Figure 18 It is the accuracy under the characteristics of datasets of different sizes.

[0085] Figure 19 It is the recall rate under the characteristics of datasets of different sizes.

[0086] Figure 20 It is the accuracy rate under the characteristics of datasets of different sizes.

[0087] Figure 21 It is the complete coverage rate under the characteristics of datasets of different sizes.

[0088] Figure 22 Schematic diagram of the IEEE 118-node system topology with wind turbines introduced. DETAILED DESCRIPTION

[0089] An adaptive multi-scale transient voltage stability assessment method based on an improved SK-BiLSTM (Simultaneous Short-Term Memory) architecture first considers the impact of indirect connections between nodes in the power grid topology on transient voltage stability. A graph attention network is used to combine time-series electrical measurement data features with graph-structured topological data features to form spatiotemporal fusion data features. Next, an improved SK-BiLSTM model is constructed for multi-scale feature extraction. By introducing progressive grouped convolution, the improved SKNet is able to extract both global and local information. Leveraging the BiLSTM's strong bidirectional time-series data processing capabilities, a weighted fusion of Focal Loss and Dice Loss is implemented, leveraging the strengths of both. This enables the model to more comprehensively capture and assess the transient voltage stability of the power system, thereby improving the accuracy and reliability of transient voltage stability assessments. Furthermore, a complete coverage metric is defined to evaluate the model's ability to accurately identify and delineate nodes or regions experiencing voltage instability. Finally, an analysis was conducted in the IEEE39 and IEEE 118 node test systems with wind turbines, verifying that the proposed method can extract global and local information from the dataset while taking into account the response speed of transient voltage stability assessment, and improve the effectiveness of power system voltage stability assessment and voltage instability node or area division.

[0090] Below in conjunction with accompanying drawing, the selected example is described in detail:

[0091] (1) Analysis of temporal and spatial characteristics of transient voltage:

[0092] Transient voltage stability assessments are characterized by high dimensionality, time-varying, and nonlinear characteristics, making it difficult to accurately predict the transient voltage stability state of a power system after a fault occurs. Furthermore, the changing trends of generator power angles and node voltages are closely related to the topological location of the nodes in the power system. Nodes in different locations may experience different reactive power support, resulting in both similarities and differences in the time-series electrical measurement data characteristics of each node. Fusion of time-series electrical measurement data features with graph-structured topological data can fully utilize the dynamic information of time-series data and the static information of the graph structure, significantly improving the accuracy of transient voltage stability assessments and the performance of delineating nodes or regions with voltage instability.

[0093] The transient voltage response characteristics are primarily manifested in the post-fault voltage time series and spatial distribution. Taking a modified IEEE 39-bus system as an example, we study its transient voltage dynamic response. Assume that a three-phase short circuit occurs between nodes 16 and 17 (near node 16), and after a period of time, the fault and the faulty line are removed. Figure 1 It shows the voltage amplitude curve when the system is stable after the fault and fault line are removed; Figure 2 It shows the voltage amplitude curve when the system becomes unstable after the fault and fault line are removed. Figure 2The magnitude of each curve in the graph varies, primarily due to the influence of power system topology, load, generator distribution, electrical distance, impedance, and protection actions. Therefore, it is crucial to fully utilize the dynamic information of time series data and the static information of the graph structure to significantly improve the accuracy of transient voltage stability assessment.

[0094] (2) Construction of adaptive multi-scale transient voltage stability assessment model based on improved SK-BiLSTM:

[0095] This paper proposes an adaptive multi-scale transient voltage stability assessment method based on improved SK-BiLSTM. The overall structure of the model is as follows: Figure 3 The model consists of five parts: raw data, graph attention network (GAT), selective kernel convolution network (SKNet) that introduces the selective kernel convolution mechanism, bidirectional long short-term memory network (Bi-LSTM), and classifier.

[0096] 1) Raw data: This provides basic training data for the model, ensuring that the model can learn and identify the transient stability characteristics of the power system under different operating conditions. The various electrical measurement data features and graph structure topology data features in the raw data are the basis for feature extraction. Through layer-by-layer processing of the model, high-level features useful for evaluating transient stability are gradually extracted. In addition, the raw data is also used for model verification and testing to ensure the model's generalization ability and prediction accuracy. By preprocessing the raw data (such as normalization), the model's training effect and evaluation performance can be improved. Ultimately, the raw data can also be used to establish a baseline model, compare its performance with the new model, and verify the improvement effect and superiority of the new model.

[0097] The present invention considers the influence of the fusion of time series electrical measurement data features and graph structure topology data features on the transient voltage stability assessment performance, and selects a relatively complete set of comprehensive features. Among the time series electrical measurement data features, the features related to the generator include G ω , G δ , G △f , G Pm , G P , G Q , which represent the angular velocity, power angle, frequency deviation, mechanical power, active power, and reactive power of different generators respectively; the bus-related characteristics include B u 、B θ , respectively represent the voltage amplitude and phase angle of all buses; the line-related characteristics include C P 、C Q , respectively representing the active power and reactive power of all lines. In the graph topology data features, the node degree and the electrical distance between nodes in the graph topology are selected.

[0098] Transient voltage stability assessment is defined as the primary classification task, and the classification of voltage instability nodes or areas is defined as the secondary classification task. Transient voltage stability is determined using a practical engineering criterion: the busbar node voltage must remain below 0.75 pu for no more than 1 second after a fault. In the primary classification task, regardless of partitioning, if at least one monitored busbar meets the voltage instability criterion, the system is considered unstable and assigned a class label of 0; otherwise, the system is considered stable and assigned a class label of 1. In the secondary classification task, partitioning is considered and the transient voltage stability of each node is determined. Nodes that meet the criterion are labeled 1, while those that do not are labeled 0. For a system with m nodes, an m-dimensional label is obtained.

[0099] 2) Graph Attention Network (GAT): It can effectively combine the various electrical measurement data features in the original data with the information of the graph structure topology data features, thereby fully utilizing the temporal and spatial characteristics of the power system and significantly improving the overall evaluation performance of the model, and the system's operating status and potential problems can be discovered and resolved earlier.

[0100] Specifically, the initial feature vector h of node i can be i It is represented by the combination of time series electrical measurement data and graph topology data. Assume that the time series electrical measurement data is x i , the graph topology data is y i , then the initial eigenvector h of node i i As shown in formula (1).

[0101] h i =[x i ||y i ] (1)

[0102] In the above formula, x i Represents time series electrical measurement data; y i Represents graph topology data; || represents vector connection operation; h i Represents the initial feature vector of node i.

[0103] Next, the initial feature vector h of each node i Perform linear transformation and obtain new features as shown in formula (2).

[0104] h′ i =Wh i (2)

[0105] In the above formula, W represents a trainable weight matrix used to map different types of data into a unified feature space; h i Represents the initial feature vector of node i.

[0106] Afterwards, the attention coefficient between node i and its neighbor node j is calculated as shown in formula (3).

[0107] e ij =LeakyReLU(a T [h′ i ||h′ j ]) (3)

[0108] In the above formula, e ij represents the attention coefficient between node i and its neighbor node j; a represents a trainable weight matrix vector; LeakyReLU represents an activation function used to introduce nonlinearity; [h′ i ||h′ j ] represents the concatenation of the feature vectors of node i and node j.

[0109] The attention coefficients of all neighbor nodes are normalized using the SoftMax function, as shown in Equation (4).

[0110]

[0111] In the above formula, α ij represents the normalized attention coefficient; e ij represents the attention coefficient between node i and its neighbor node j; exp(e ij ) represents the exponential value of the attention coefficient; N i Represents the set of neighbor nodes of node i.

[0112] Finally, the normalized attention coefficient is used to perform weighted summation on the features of neighboring nodes to obtain the new features of node i, as shown in formula (5).

[0113]

[0114] In the above formula, h″ i represents the new feature vector of node i; σ represents the nonlinear activation function; α ij represents the normalized attention coefficient; h′ j Represents the feature vector of neighbor node j;

[0115] 3). Selective kernel convolution network (SKNet) that introduces selective kernel convolution mechanism: In terms of feature extraction methods, current studies mostly use single-scale convolution operation (SSCO), which can only extract features at a fixed scale and cannot fully capture information at different scales. This limitation makes it difficult for the model to effectively capture the comprehensive characteristics of global and local information when processing complex transient signals, thereby affecting the evaluation effect. Although a small number of studies have adopted multi-scale convolution operation (MSCO), they have failed to fully consider the impact of non-direct connection of nodes in the power grid topology on transient voltage stability, and there are deficiencies in computational efficiency. The present invention introduces the selective kernel convolution network SKNet (SKConv1, SKConv2) with selective kernel convolution mechanism twice, which can adaptively and dynamically extract the spatiotemporal fusion data features input by GAT at multiple scales, so as to comprehensively extract the global and local information in the spatiotemporal fusion data features, improve the accuracy and reliability of transient voltage stability assessment, and judge the performance of dividing voltage instability nodes or areas.

[0116] Specifically, the structure of the selective kernel convolutional network (SKNet) that introduces the progressive group convolution mechanism is shown in the figure below: Figure 4 As shown in the figure, it is divided into three stages: split, fusion and selection.

[0117] First, in the split stage (Split), the dataset feature h" passed to GAT is reshaped into the input feature Figure X , then input the features Figure X Convolution operations are performed in parallel using multiple convolution kernels of different scales. Feature information at different scales is captured, thereby generating multiple feature maps. Specifically, this model uses 3×3, 5×5, and 7×7 convolution kernels to generate feature maps F1, F2, and F3, respectively, as shown in Equation (6).

[0118]

[0119] In the above formula, δ represents the Swish_B activation function; GP represents the progressive group convolution operation; for example, GP(3×3(X)) represents the input feature Figure X Perform progressive grouped convolution with a kernel size of 3×3; BN stands for batch normalization; F1, F2, and F3 represent the input data features respectively. Figure X Data feature maps obtained by performing progressive grouped convolution operations with kernel sizes of 3×3, 5×5, and 7×7.

[0120] By introducing the progressive group convolution mechanism, we can extract global and local information at the same time, improve the response speed of the model, and become more flexible when processing data of different scales. The progressive group convolution operation is shown in formula (7):

[0121]

[0122] In the above formula, G p Indicates the number of groups in the current group volume; G base Indicates the number of basic group convolution groups; epoch indicates the current training round; max_epoch indicates the maximum training round; n is the group adjustment frequency.

[0123] Then, in the fusion stage (Fuse), the feature maps obtained by convolution operations of different sizes are summed element-wise to obtain the fused feature map F, as shown in formula (8).

[0124] F=F1+F2+F3 (8)

[0125] In the above formula, F represents the fused data feature map; F1, F2 and F3 represent the input data features respectively. Figure X Data feature maps obtained by performing progressive grouped convolution operations with kernel sizes of 3×3, 5×5, and 7×7.

[0126] Then the fused feature map F is subjected to global average pooling (GAP) to obtain the global feature vector g, as shown in formula (9).

[0127] g=GAP(F) (9)

[0128] The global feature vector g obtained from Equation (9) is then passed through a two-layer fully connected network to calculate the selective weight. First, the global feature vector is compressed to 1 / 16 of its original dimension through a fully connected layer and then nonlinearly transformed using an activation function (such as ReLU) to generate the feature vector z, as shown in Equation (10).

[0129] z=ReLU(W1g) (10)

[0130] In the above formula, W1 represents the weight matrix of the first fully connected layer, and g represents the global eigenvector. ReLU (Rectified Linear Unit) is a commonly used activation function, expressed as ReLU(x) = max(0,x), and is primarily used to introduce nonlinearity.

[0131] Then, the feature dimension is restored through another fully connected layer and the selective weight vector s is generated, as expressed in Equation (11).

[0132] s=W2z (11)

[0133] In the above formula, W2 represents the weight matrix of the second fully connected layer, and z represents the feature vector obtained by the first fully connected layer and the activation function.

[0134] Finally, in the selection phase (Select), in order to ensure the rationality of the weight distribution between different feature maps, the selective weights need to be normalized using the SoftMax function. The SoftMax function is a commonly used normalization function that can convert the input vector into a probability distribution. Its function is to map a set of real numbers to between (0, 1) and the sum is 1. The selective weight coefficient α i The calculation is shown in formula (12).

[0135]

[0136] In the above formula, si represents the i-th element in the selective weight vector s; i = 1, 2, 3 means that there are three parallel convolution branches in the selective kernel convolution network (SKNet) that introduces the progressive group convolution mechanism, and each branch extracts feature information at a different scale. The output feature maps of these branches are denoted as F1, F2, and F3 respectively, and their weight coefficients α are calculated by the SoftMax function. i , and apply the normalized weights to the corresponding feature maps.

[0137] The normalized selective weight coefficient is used to weight the sum of different feature maps to obtain the output feature map Y, as shown in formula (13).

[0138]

[0139] In the above formula, Y i represents the i-th feature map after selective weighting; α i represents the selectivity weight coefficient of the i-th feature map, which is calculated according to formula (12); F i Represents the i-th feature map, which are F1, F2 and F3, for feature information of different scales; Y represents the output feature map obtained after the final fusion, and the weighted feature map Y i The sum is obtained.

[0140] In order to more comprehensively extract global and local information and become more flexible when processing data of different scales, further optimize the feature representation, improve the robustness and generalization ability of the model, and thus improve the accuracy and reliability of transient voltage stability assessment. The output feature map Y obtained by the selective kernel convolution network (SKNet) with the progressive group convolution mechanism is passed to the selective kernel convolution network again. The first selective kernel convolution operation is called SKConv1 operation, and the second selective kernel convolution operation is called SKConv2 operation. At the same time, in order to distinguish the output feature maps obtained by the final fusion of the two selective kernel convolution operations, the fused feature map obtained by element-wise summation of formula (8) in the SKConv1 operation is denoted as F1, and the fused feature map obtained by weighted summation of formula (13) is denoted as Y1; the fused feature map obtained by element-wise summation of formula (8) in the SKConv2 operation is denoted as F2, and the fused feature map obtained by weighted summation of formula (13) is denoted as Y2.

[0141] 4) Bi-LSTM: The fused feature map Y2 obtained by weighted summation of SKConv2 operations is reshaped and passed into the Bi-LSTM. The forward and backward LSTM layers of BiLSTM are used to simultaneously capture the dependencies between the time series data, thereby fully representing the temporal characteristics of the input data. The BiLSTM structure diagram is shown below. Figure 5 By combining the improved multi-scale spatial feature extraction capability of SKNet with the strong bidirectional time series feature extraction capability of BiLSTM, the spatiotemporal dynamic changes in power system data can be captured more accurately, improving the performance of prediction and classification tasks.

[0142] 5) Classifier: A Sigmoid classifier is used, which converts the model's output into a probability value between 0 and 1, indicating the probability of system stability. By setting an appropriate threshold, these probability values ​​can be converted into clear classification decisions. In addition, the Sigmoid classifier is stable during gradient calculation and model training, further improving the performance and reliability of the model. The mathematical expression of the Sigmoid classifier is shown in Equation (22).

[0143]

[0144] In the above formula, σ1(x) represents the output of the Sigmoid function, and x is the input value, which is usually the output score of the model. Here, it refers to the output score of the original data after being processed by the Graph Attention Network (GAT), two improved selective kernel convolution layers (SKConv1, SKConv2), and the Bidirectional Long Short-Term Memory Network (BiLSTM). The output value of the Sigmoid function can be interpreted as the probability that x belongs to the transient voltage stability of the power system. For example: when σ1(x)>0.5, it usually means that the sample is classified as a stable sample; when σ1(x)≤0.5, it usually means that the sample is classified as an unstable sample;

[0145] (2) Adaptive multi-scale transient voltage stability assessment process based on improved SK-BiLSTM:

[0146] 1) Construction of sample set of time series data: The sampling step length T of the present invention is set to 0.01s to ensure high accuracy of data. A time window length of 0.4s is used, and the time interval t is set according to different fault removal times. c , extract the time segment of the corresponding length of the original data, the specific sampling interval is [t c -28T,t c +12T];

[0147] 2) Sample set construction of spatial data: The present invention selects node degree and electrical distance between nodes as the topological structure data of the power system diagram;

[0148] 3) Data preprocessing: To facilitate model training and evaluation, Z-score normalization is used to normalize the data input into the model so that all data are in the same dimension and the negative impact of dimensional differences is eliminated. The mathematical expression of Z-score normalization is shown in formula (21).

[0149]

[0150] In the above formula: X* represents the value after data normalization; μ represents the mean of the data in the sample; σ is the standard deviation of the sample data; ε represents a minimum value; X is the data in the data feature set.

[0151] 4) Model Parameter Selection and Training: Model parameters were adjusted based on accuracy requirements, and the FL-DL loss function was adopted. During model training, the Adam optimizer and Sigmoid classifier were used, and batch processing was employed to accelerate training. The initial learning rate was set to 0.001. To prevent overfitting, the learning rate was halved after the loss value remained stable for five consecutive rounds. Training continued until comprehensive evaluation metrics met requirements and the loss function fluctuations were less than the set value. The model was then retained; otherwise, the model parameters were adjusted and retrained. This entire process ensured the final model parameters.

[0152] 5) Evaluation results: After training, the test set data is passed into the model, and the model performance is comprehensively evaluated based on the comprehensive evaluation indicators output by the model.

[0153] The present invention considers that in the study of power system suspended voltage stability assessment, the tolerance for evaluating actual transient voltage instability samples as transient voltage stability is much lower than the tolerance for evaluating actual transient voltage stability samples as transient voltage instability. Therefore, based on the binary evaluation confusion matrix shown in Table 1, the accuracy (Ac), precision (Pr), and recall (Re) are defined as shown in Equations (17) to (19).

[0154]

[0155] In equations (17)-(19), Ac provides the overall correct prediction ratio; Re measures the model's ability to identify unstable samples, focusing on reducing missed predictions; and Pr measures the proportion of samples predicted to be unstable that are actually unstable, focusing on reducing false positives. Combining accuracy, recall, and precision allows for a comprehensive assessment of the model's performance in transient voltage stability analysis, ensuring that the model not only performs well overall but also effectively captures and accurately predicts unstable samples.

[0156] Table 1 Confusion matrix

[0157]

[0158] In Table 1, TP and FN represent the number of transient voltage stability samples that are correctly and incorrectly identified, respectively; FP and TN represent the number of transient voltage instability samples that are incorrectly and correctly identified, respectively.

[0159] Furthermore, previous studies on voltage instability nodes or regions typically lack a unified evaluation metric. Therefore, this paper uses the defined complete coverage metric to assess the model's performance in identifying voltage instability nodes or regions, further demonstrating the model's accuracy, effectiveness, and speed in practical applications.

[0160] The Full Coverage Rate (FCR) is a measure of the model's performance in the voltage instability node or area partitioning task. The mathematical expression of the Full Coverage Rate is shown in the following formula (20).

[0161]

[0162] In the above formula, N1 is the total number of unstable samples evaluated in the main classification task, y ij is the jth true label of the i-th sample, is the jth predicted label of the i-th sample, is an indicator function, which is 1 when all unstable nodes of sample i are correctly predicted, and 0 otherwise.

[0163] In summary, the overall flow chart of the adaptive multi-scale transient voltage stability assessment method based on improved SK-BiLSTM proposed in this invention can be expressed as follows: Figure 6 shown.

[0164] The technical effects of the present invention will be further illustrated by way of examples below:

[0165] like Figure 7 As shown, the IEEE 39-bus power system with wind turbines includes seven conventional turbines and three wind farms. PSASP software was used to generate time-series electrical measurement data. To enhance the model's adaptability to actual operating conditions, three wind power penetration rates of 5%, 10%, and 15% were set. The system load ranged from 90% to 110%, with simulations performed at five different load levels in 5% increments. At each load level, a three-phase short-circuit grounding fault occurred on each AC line. The fault clearing time is typically greater than or equal to three cycles (0.05 seconds). This simulation considered a fault duration of 3 to 11 cycles, and cleared the faulty line during fault clearing. For graph-structured topological data, the node degree and internode electrical distance in the power system topology graph were used. The node degree is the number of outgoing lines from the busbar, and the internode electrical distance is represented by the internode impedance. The experiment generated 13,365 samples, including 11,291 stable samples and 2,074 unstable samples in the transient voltage stability assessment task (main classification task);

[0166] Firstly, the time series electrical measurement data features and graph structure topology data features are fused into spatiotemporal fusion data features through the Graph Attention Network (GAT). Then, the spatiotemporal fusion data features are reshaped into X so as to be passed into the Selective Kernel Convolutional Network (SKNet) which introduces the progressive group convolution mechanism.

[0167] The Adam optimizer was used in the model training process. During the initial training, the number of training rounds was set to 100, the batch size N was set to 64, and the learning rate was set to 0.001. To prevent overfitting, the learning rate was halved after the 70th training round.

[0168] In order to facilitate the adjustment of the model structure, it is assumed that the feature Figure X The number of channels is M, and the changes in the feature dimensions of each main hidden layer in the selective kernel convolutional network (SKNet) with the introduction of progressive group convolution mechanism are shown in Table 2.

[0169] Table 2 Changes in feature dimensions of each hidden layer

[0170]

[0171]

[0172] To verify the comprehensive performance of the selective kernel convolutional network (SKNet) with a progressive grouped convolution mechanism combined with a bidirectional long short-term memory network (BiLSTM) in transient voltage stability assessment and voltage instability node or area classification, we compared its comprehensive evaluation index performance with other common deep learning network methods, including single-scale convolutional models (BiLSTM, CNN-BiLSTM) and multi-scale convolutional models (FL-CNN, SE-CGRU, and SKNet). The experimental results are shown in Table 3.

[0173] Table 3 Comparison of comprehensive evaluation index performance of different evaluation models

[0174]

[0175] Accuracy, recall, and precision comprehensively reflect the model's performance in transient voltage stability assessment; complete coverage reflects the model's performance in delineating voltage instability nodes or regions. Table 3 shows that the proposed method performs well in transient voltage stability assessment, achieving high accuracy, recall, and precision. Furthermore, the proposed method significantly outperforms other models in delineating voltage instability nodes or regions.

[0176] Furthermore, to verify that the progressive grouped convolution mechanism employed in the present invention not only comprehensively extracts global and local information from the spatiotemporal fusion feature X, but also offers certain advantages in model response speed, the selective kernel convolutional network (SKNet) employing the progressive grouped convolution mechanism employed in the present invention, combined with a bidirectional long short-term memory network (BiLSTM), was compared with multi-scale convolutional models (FL-CNN, SE-CGRU, and SKNet) in terms of response speed for transient voltage stability assessment and voltage instability node or region partitioning. The experimental results are shown in Table 4.

[0177] Table 4 Comparison of response speed of different multi-scale convolution models

[0178]

[0179] As shown in Table 4, the adaptive multi-scale transient voltage stability assessment method based on the improved SK-BiLSTM proposed in the present invention has a much better response speed than the FL-CNN and SE-CGRU models on both the training set and the test set.

[0180] At the same time, in order to verify the advantages and disadvantages of the progressive group convolution mechanism adopted by the present invention compared with not using group convolution and fixed group convolution, different convolution operation grouping strategies are adopted. The experimental results are shown in Table 5.

[0181] Table 5 Performance comparison under different grouping strategies

[0182]

[0183] As shown in Table 5, the progressive group convolution mechanism significantly improves the model's accuracy and computational efficiency. Although more complex to implement and debug than fixed group convolution, progressive group convolution can better adapt to feature extraction requirements at different levels and optimize resource utilization. Therefore, despite its certain implementation complexity, progressive group convolution still has significant potential and practical value in practical applications.

[0184] Next, the FL-DL loss function (using Focal Loss and Dice Loss weighted fusion) adopted in the present invention is compared with the focal loss function (FL), Dice loss function (DL) and binary cross entropy loss function (BCE), and the comprehensive performance evaluation indicators of transient voltage stability assessment of different loss function models are statistically analyzed. The experimental results are as follows: Figure 8 As shown. The training set loss value waveform is as follows Figure 9 As shown, the test set loss value waveform is as follows Figure 10 shown.

[0185] Depend on Figure 8 As can be seen, the FL-DL method performs well across all metrics, with a particularly strong lead in precision and recall, demonstrating its superiority in identifying stable samples and ensuring accurate classification. Furthermore, the FL-DL method exhibits a more balanced performance across all metrics, demonstrating its overall superiority in classification evaluation tasks. Therefore, the FL-DL method is more reliable and stable in practical applications.

[0186] Depend on Figure 9 and Figure 10 As can be seen, the FL-DL loss function has multiple advantages over other loss functions. First, it exhibits a rapid decline in the early stages of training, enabling the model to quickly learn effective features and thus speeding up training. Second, FL-DL maintains a low final loss value on both the training and validation sets, demonstrating good convergence and stability.

[0187] Since the loss function adopted in the present invention is a FL-DL loss function formed by the weighted fusion of Focal Loss and Dice Loss. Therefore, in order to evaluate the impact of the weight parameter λ in the FL-DL loss function on the model performance, multiple groups of experiments were carried out to train and test the models under different λ values. When λ is 0, the loss function is completely determined by Dice Loss; when λ is 1, it is completely determined by Focal Loss; when λ is between 0 and 1, the loss function is a weighted average of the two. Five different λ values ​​were selected for the experiment, and the comprehensive performance results of transient voltage stability evaluation under different λ values ​​are shown in Table 6.

[0188] Table 6 Comprehensive performance of transient voltage stability assessment under different λ values

[0189]

[0190]

[0191] Table 6 shows that the performance of the model on various evaluation metrics varies significantly with the change in the λ value. In particular, when the λ value is 0.5, the model achieves optimal performance in accuracy (Ac), recall (Re), precision (Pr), and full coverage rate (FCR). This indicates that the combination of Focal Loss and Dice Loss is most effective at this λ value, better balancing different metrics and improving the overall performance of the model.

[0192] In order to further demonstrate the ability of the adaptive multi-scale transient voltage stability assessment method based on the improved SK-BiLSTM proposed in this paper to distinguish stable and unstable samples, the feature distributions of the original dataset, SKConv1 layer, SKConv2 layer and BiLSTM layer were visualized by t-SNE. The experimental results are shown in Figure 2. Figures 11 to 16 As shown. Figure 11 、 Figure 12 、 Figure 13 、 Figure 14 、 Figure 15 and Figure 16 They represent the feature distribution of spatiotemporal fusion data, the feature distribution of F1 data obtained by the SKConv1 layer, the feature distribution of Y1 data obtained by the SKConv1 layer, the feature distribution of F2 data obtained by the SKConv2 layer, the feature distribution of Y2 data obtained by the SKConv2 layer, and the feature distribution of BiLSTM layer data respectively.

[0193] Depend on Figure 11 It can be seen that in the original data distribution, stable samples and unstable samples are mixed together and difficult to distinguish; Figure 12 and Figure 13It can be seen that after being processed by the SKConv1 layer, the features begin to show a certain separation effect, and the boundary between stable samples and unstable samples gradually becomes clear; Figure 14 and Figure 15 It can be seen that after SKConv2 layer processing, the sample boundary is more obvious; Figure 16 As can be seen, after BiLSTM processing, a clear boundary is formed between stable and unstable samples. As feature learning progresses, a clear boundary gradually forms between stable and unstable samples. This phenomenon demonstrates the effectiveness of the adaptive multi-scale transient voltage stability assessment method proposed in this paper based on the improved SK-BiLSTM in feature extraction and distinguishing between unstable and stable samples. It can effectively distinguish between stable and unstable samples in power systems.

[0194] In order to verify that the adaptive multi-scale transient voltage stability assessment method based on the improved SK-BiLSTM proposed in the present invention has good flexibility in dealing with the characteristics of data sets of different sizes. First, 100 trees are integrated through XGBoost to analyze the characteristics of time-series electrical measurement data, and the contribution of generator-related features, generator-related features and generator-related features to transient voltage stability assessment is analyzed. The number of splits is used as a measure of feature contribution and weighted sum is taken to obtain the average. The sample set is divided into training set and test set according to the ratio of 7:3. The model adopts L1-L2 regularization and iterates learning 100 times to obtain the contribution of comprehensive features to voltage problems. Figure 17 shown.

[0195] Depend on Figure 17 It can be seen that in the characteristics of time-series electrical measurement data, the contribution of each electrical quantity to the transient voltage stability assessment is as follows: bus voltage amplitude, branch reactive power, branch active power, bus voltage phase angle, generator reactive power, generator angular velocity, generator power angle, generator active power, generator frequency deviation, and mechanical power.

[0196] Subsequently, the features with lower contribution are eliminated in order from low to high, and different time window lengths are selected to verify the flexibility of the model while ensuring that the time window length can cover the steady state, after fault occurrence and after fault removal data. The experimental results are as follows: Figure 18 、 Figure 19 、 Figure 20 and Figure 21 As shown in the figure, the accuracy, recall, precision and complete coverage of the datasets with different sizes are represented in turn.

[0197] Depend on Figure 18 、 Figure 19 、 Figure 20 and Figure 21The adaptive multi-scale transient voltage stability assessment method proposed in this paper, based on the improved SK-BiLSTM, achieves good results across datasets of varying sizes, demonstrating its high flexibility and ability to adapt to changes in data dimensionality and time series length. Even after removing less-contributing electrical measurement data, the model maintains excellent performance, demonstrating its adaptability in handling data changes and its ability to adaptively extract and enhance features for data with greater impact on the assessment.

[0198] Finally, to test the generalization ability of the adaptive multi-scale transient voltage stability assessment method based on the improved SK-BiLSTM proposed in this paper in large power systems, an N-1 dataset was constructed in an IEEE 118-node system that introduced wind turbines for verification. The system consists of 118 nodes, 47 traditional turbines, 7 wind turbines, 177 AC lines, and 91 loads. Its topology is as follows: Figure 22 As shown in Figure 2, 3248 stable samples and 1447 unstable samples were randomly generated, and the data were divided into training and test sets in a ratio of 3:1, keeping other model parameters unchanged.

[0199] Next, the proposed adaptive multi-scale transient voltage stability assessment method based on the improved SK-BiLSTM was compared with single-scale convolutional models (BiLSTM, CNN-BiLSTM) and multi-scale convolutional models (FL-CNN, SE-CGRU, and SKNet) for comprehensive transient voltage stability assessment performance. The experimental results are shown in Table 7.

[0200] Table 7 Comparison of comprehensive evaluation index performance of different evaluation models

[0201]

[0202] Table 7 shows that the adaptive multi-scale transient voltage stability assessment method proposed in this paper, based on the improved SK-BiLSTM, performs well across various metrics, including accuracy, recall, precision, and complete coverage. In particular, the complete coverage reaches 72.89%. This demonstrates that the model not only operates stably and efficiently in a 39-node system, but also maintains excellent performance in a more complex 118-node system, demonstrating its good generalization capabilities across large-scale power systems.

Claims

1. An adaptive multi-scale transient voltage stability assessment method based on improved SK-BiLSTM is characterized by The following steps are involved: Step 1: Use the graph attention network to combine the time series electrical measurement data features with the graph structure topology data features to form spatiotemporal fusion data features; Step 2: Build an adaptive multi-scale evaluation model based on the improved SK-BiLSTM. In this model, a progressive group convolution operation mechanism is introduced into the selective kernel convolutional network SKNet, combined with a bidirectional long short-term memory network Bi-LSTM, and the FocalLoss and Dice Loss are weightedly fused to obtain a weighted fusion loss function. The feature extraction is achieved using the progressive grouped convolution operation mechanism. The progressive grouped convolution operation is shown in formula (7): (7)； In the above formula, G p Indicates the number of groups in the current group volume; G base Indicates the number of basic group convolution groups; epoch Indicates the current training round; max_epoch Indicates the maximum number of training rounds; n Adjust frequency for grouping; Step 3: Evaluate the transient voltage stability of the power system using the adaptive multi-scale evaluation model based on the improved SK-BiLSTM constructed in Step 2. Verify the model's performance in identifying voltage instability nodes or areas using the defined full coverage index.

2. The adaptive multi-scale transient voltage stability assessment method based on the improved SK-BiLSTM according to claim 1 is characterized by: In step 1, among the time series electrical measurement data features, the features related to the generator include: G ω 、 G δ 、 G Δf 、 G pm 、 G P 、 G Q , respectively representing the angular velocity, power angle, frequency deviation, mechanical power, active power, and reactive power of different generators; bus-related characteristics include B u 、 B θ , respectively represent the voltage amplitude and phase angle of all buses; the line-related characteristics include C P 、 C Q , respectively represent the active power and reactive power of all lines; Among the graph structure topology data features, the node degree and the electrical distance between nodes in the electrical system graph structure topology are selected.

3. The adaptive multi-scale transient voltage stability assessment method based on the improved SK-BiLSTM according to claim 2 is characterized by: In step 1, transient voltage stability is judged using a criterion: transient voltage stability assessment is defined as a main classification task. In the main classification task, without considering partitions, if one of the monitored buses meets the voltage instability criterion, the system is judged to be unstable, and the corresponding category label is 0; otherwise, it is judged to be stable, and the corresponding category label is 1; The division of voltage instability nodes or areas is defined as a sub-classification task. In the sub-classification task, the partition is considered and the transient voltage stability state of each node is judged; the node label that meets the judgment criteria is recorded as 1, and the node label that does not meet the judgment criteria is recorded as 0. If the number of nodes in the system is m, an m-dimensional label will be obtained.

4. The adaptive multi-scale transient voltage stability assessment method based on the improved SK-BiLSTM according to claim 3 is characterized by: In step 1, a graph attention network (GAT) is used to fuse the graph structure topology data features and the time series electrical measurement data features of the power system; The details are as follows: The node i The initial eigenvector of h i It is represented as the combination of time series electrical measurement data and graph topology data; assuming that the time series electrical measurement data is x i , the graph topology data is y i , then the node i The initial eigenvector of h i As shown in formula (1); (1)； In the above formula, x i Represents time-series electrical measurement data; y i Represents graph topology data; || represents the connection operation of vectors; h i Representation node i The initial eigenvector of Next, the initial feature vector of each node is h i Perform linear transformation to obtain new features as shown in formula (2); (2) ； In the above formula, W Represents a trainable weight matrix used to map different types of data into a unified feature space; h i Representation node i The initial eigenvector of Representation node i The vector obtained by linear transformation of the initial eigenvector of ; Afterwards, the computing nodes i Its neighboring nodes j The attention coefficient between them is shown in formula (3); (3) ； In the above formula, e ij Representation node i Its neighboring nodes j The attention coefficient between a Represents a trainable weight matrix vector; LeakyReLU Represents an activation function used to introduce nonlinearity; Representation node i With node j 's feature vector splicing; Normalize the attention coefficients of all neighbor nodes using the SoftMax function, as shown in formula (4); (4) ； In the above formula, represents the normalized attention coefficient; e ij Representation node i Its neighboring nodes j The attention coefficient between exp( e ij ) represents the exponential value of the attention coefficient; N i Representation node i The set of neighbor nodes of Finally, the normalized attention coefficient is used to perform weighted summation on the features of neighbor nodes to obtain the node i The new features are shown in formula (5); (5) ； In the above formula, Representation node i The new eigenvector of ; represents a nonlinear activation function; represents the normalized attention coefficient; Represents neighbor nodes j The eigenvector of .

5. The adaptive multi-scale transient voltage stability assessment method based on the improved SK-BiLSTM according to claim 1 is characterized in that: In step 2, an adaptive multi-scale evaluation model based on an improved SK-BiLSTM is constructed, which includes: a graph attention network GAT, a selective kernel convolution network SKNet that introduces a selective kernel convolution mechanism, a bidirectional long short-term memory network Bi-LSTM, and a classifier; Graph Attention Network (GAT): combines the various electrical measurement data features in the original data with the information of the graph structure topology data features; The selective kernel convolution network SKNet introduces the selective kernel convolution mechanism, which includes three stages: segmentation, fusion, and selection: First, in the segmentation stage, the dataset features passed into the graph attention network GAT are Reshape into input feature map X , then input the feature map X Simultaneously, multiple convolution kernels of different scales are used to perform parallel convolution operations; feature information of different scales is captured to generate multiple feature maps; Then, in the fusion stage, the feature maps obtained by convolution operations of different sizes are summed element-wise to obtain the fused feature map F ; Afterwards, the fusion feature map F The global feature vector is obtained by global average pooling GAP g , and then the global eigenvector is obtained g The selective weights are calculated through two layers of fully connected networks; then, the feature dimensions are restored through another fully connected layer to generate the selective weight vector s ; Finally, in the selection phase, the selective weights are normalized using the SoftMax function; The fused feature map obtained by weighted summation of the second selective kernel convolution operation SKConv2 Y2 The reshaped data is then fed into a bidirectional long short-term memory (Bi-LSTM) network. Using the forward and backward LSTM layers of the Bi-LSTM network, the forward and backward dependencies in the time series data are captured simultaneously, thereby fully characterizing the temporal characteristics of the input data. The classifier uses the Sigmoid classifier, which can convert the output of the model into a probability value between 0 and 1, thereby indicating the probability of whether the system is stable or not.

6. The adaptive multi-scale transient voltage stability assessment method based on the improved SK-BiLSTM according to claim 5 is characterized by: The output feature map obtained by finally fusing the selective kernel convolution network SKNet which introduces the progressive group convolution mechanism Y , and then pass it into the selective kernel convolution network again. The first selective kernel convolution operation is called SKConv1 operation, and the second selective kernel convolution operation is called SKConv2 operation. At the same time, in order to distinguish the output feature map obtained after the final fusion of the two selective kernel convolution operations, the fused feature map obtained by element-level summation in the SKConv1 operation is recorded as F1 , the fusion feature map obtained by weighted summation is recorded as Y1 ; The fused feature map obtained by element-level summation in the SKConv2 operation is recorded as F2 , the fusion feature map obtained by weighted summation is recorded as Y2 .

7. The adaptive multi-scale transient voltage stability assessment method based on the improved SK-BiLSTM according to claim 6 is characterized by: Step 2 includes the following steps: First, the selective kernel convolutional network SKNet, which introduces progressive group convolution operations, is used to preliminarily process the spatiotemporal fusion data features. Then, the extracted data features are input into the bidirectional long short-term memory network Bi-LSTM, which uses its bidirectional time series data processing capabilities to improve the efficiency and accuracy of feature extraction; Finally, the data features transmitted by the bidirectional long short-term memory network Bi-LSTM are input into the classifiers of the main classification task and the secondary classification task to obtain the evaluation results.

8. The adaptive multi-scale transient voltage stability assessment method based on the improved SK-BiLSTM according to claim 4 is characterized by: In step 2, the weighted fusion loss function of Focal Loss and Dice Loss is calculated as shown in the following formula (14); (14) ； In the above formula, p represents the model's predicted probability for the stable / instability category; 𝜆 represents the weight parameter; FL ( p ) represents FocalLoss; DL ( p ) represents Dice Loss; Focal Loss is used to deal with the problem of class imbalance, and its mathematical expression is shown in formula (15); (15) ； In the above formula, represents the balance factor; represents the focusing factor; FL ( p ) represents Focal Loss; Dice Loss is used to optimize the accuracy of classification decisions, and its mathematical expression is shown in formula (16); (16) ； In the above formula, p i represents the predicted value; g i represents the true value; Expressed as a small constant to prevent the denominator from being zero.

9. The adaptive multi-scale transient voltage stability assessment method based on the improved SK-BiLSTM according to claim 1 is characterized in that: In step 3, by defining the full coverage index FCR The performance of the model in the voltage instability node or area division task is measured; the mathematical expression of the full coverage index is shown in the following formula (20); (20) ； In the above formula, N1 is the total number of unstable samples evaluated in the main classification task, y ij It is i The first sample j True labels, ŷ ij It is i The first sample j predicted labels, is the indicator function, when the sample i It is 1 if all unstable nodes are correctly predicted, and 0 otherwise.

Images