Transient voltage stability assessment method fusing critical sample enhancement and spatiotemporal feature mining

Abstract

This invention discloses a transient voltage stability assessment method that integrates critical sample enhancement and spatiotemporal feature mining. Belonging to the field of power system stability analysis, it first obtains the bus voltage disturbance trajectory using time-domain simulation / PMU measured data, calculates the stability margin to classify sample labels, and transforms it into a voltage heatmap characterizing spatiotemporal evolution. Second, it uses a least-squares generative adversarial network to directionally amplify critical stable samples, and combines this with the SSIM index to remove distorted data, constructing a balanced, high-quality training set. Subsequently, it constructs an adaptive hybrid spatiotemporal feature fusion network AHSTF-Net, introducing Transformer and deformable residual convolution modules to extract spatiotemporal features at multiple scales and output voltage stability assessment results. This invention significantly improves noise robustness and the accuracy of transient voltage stability assessment.

Description

Technical Field

[0001] This invention belongs to the field of power system stability analysis, and in particular relates to a transient voltage stability assessment method that integrates critical sample enhancement and spatiotemporal feature mining. Background Technology

[0002] With the interconnection of large-scale power grids and the widespread integration of high-proportion renewable energy sources, the randomness of power system operation has increased, the grid topology has become more complex, and the system operating state often approaches its stability boundary. If faults are not cleared in a timely manner or control measures are inappropriate, transient voltage instability can easily occur, leading to large-scale grid disconnection or even major blackouts. Therefore, conducting rapid and accurate transient voltage stability margin assessment based on the voltage disturbance trajectory at the initial stage of a fault can ensure the early implementation of control measures and guarantee the safe and stable operation of high-proportion renewable energy power systems.

[0003] Existing transient voltage stability analyses largely rely on power system differential-algebraic equations, determining stability margins through time-domain simulation, sensitivity analysis, or isometry curves. While these methods have clear physical meanings, they are highly dependent on system topology, parameter accuracy, and dynamic models, resulting in a large modeling workload and difficulty in handling scenarios with frequent changes in operating conditions after a high proportion of renewable energy integration. Furthermore, time-domain simulation-based methods often fail to meet the requirements of online analysis.

[0004] In recent years, with the widespread application of wide-area measurement systems and phasor measurement units (PMUs), and the accumulation of massive amounts of high-dimensional data, data-driven transient stability assessment methods have received widespread attention. However, two key problems still exist in practical applications: First, due to the limited number of transient voltage instability or critical instability samples, the training sample distribution is uneven, and deep learning models are prone to misjudgment near the stability boundary. Second, traditional deep learning networks have limited ability to simultaneously characterize the spatial coupling and cross-timescale evolution of multiple bus voltages, making it difficult to fully utilize the high-dimensional spatiotemporal information in the voltage disturbance trajectory, and lacking robustness to noise and operating condition changes.

[0005] Therefore, it is necessary to propose a new data-driven model that balances critical sample enhancement and efficient spatiotemporal feature extraction. By constructing a deep learning model that adapts to complex topologies and strong uncertainties, a rapid and high-precision assessment of the transient voltage stability margin of the power system can be achieved. Summary of the Invention

[0006] Objective: To overcome the limitations of existing deep learning models, such as insufficient critical stable and unstable samples for transient voltage and restricted spatiotemporal feature extraction capabilities, this invention proposes a method and system for evaluating transient voltage stability in power systems based on critical sample enhancement and an adaptive hybrid spatiotemporal feature fusion network. This method rapidly searches for the critical cut-off time of the system using the golden section method, utilizes a least-squares generative adversarial network to enhance critical stable samples, and combines structural similarity indices to filter samples, constructing a high-quality training set with balanced distribution. An adaptive hybrid spatiotemporal feature fusion network, AHSTF-Net, integrating Transformer modules and deformable residual convolution modules, is constructed to achieve multi-scale spatiotemporal feature mining of voltage disturbance trajectories and rapid evaluation of transient voltage stability.

[0007] Technical solution: The present invention provides a transient voltage stability evaluation method that integrates critical sample enhancement and spatiotemporal characteristics, comprising the following steps:

[0008] S1. Construct a transient simulation model of the power system. Based on the operating ranges of generators and loads in the system, set various operating modes and fault scenarios. Obtain the disturbance trajectory of each bus voltage based on time-domain simulation. Use the golden section method to search for the critical cutoff time t for each operating condition. CCT The samples are divided into stable samples, critically stable samples, and unstable samples.

[0009] S2, the bus voltage disturbance trajectory under each operating condition is normalized and interpolated and resampled according to a preset time window and sampling interval, and the time dimension and node dimension are mapped to a two-dimensional grid to construct a voltage heatmap that characterizes the spatiotemporal evolution of the transient process; a dataset containing voltage heatmap and transient voltage stability category labels is established, and the dataset is divided into training set, validation set and test set according to a certain proportion.

[0010] S3 uses the least squares generative adversarial network LSGAN to perform targeted augmentation on samples marked as critically stable in the training set, and generates a critical sample augmentation set using the original critical samples as real samples.

[0011] S4 uses structural similarity indices to screen the generated critically stable samples for quality. Structural similarity scores are calculated based on brightness, contrast, and structural components. Samples with scores below the threshold or physically unreasonable are removed to form a high-quality training set with balanced distribution.

[0012] S5: Construct an adaptive hybrid spatiotemporal feature fusion network AHSTF-Net, using voltage heatmaps as input and transient voltage stability categories as output; AHSTF-Net internally incorporates a Transformer module to extract temporal features and a deformable residual convolution module to extract multi-spatial scale features; train the AHSTF-Net network model based on the training and validation sets.

[0013] S6. Input the test set data into the trained AHSTF-Net model to obtain the transient voltage evaluation results of the power system and evaluate the model performance.

[0014] Furthermore, in step S1, the transient simulation model of the power system is constructed as a multi-machine power system simulation model. The example system is an IEEE 3-machine 9-bus system. The operating mode is obtained by randomly generating several sets of generation and load conditions within a preset active and reactive power output fluctuation range. The fault scenarios include three-phase short-circuit faults on different buses and different fault durations. The golden section method is used to search for the fault clearing time t such that the voltage after fault clearing just meets the transient voltage qualification criterion. CCT Based on this, the transient voltage stability margin is calculated.

[0015] Furthermore, in step S2, the process of constructing the voltage heatmap includes: selecting a time window covering the disturbed trajectory before, during, and after the fault; interpolating and normalizing the disturbed voltage trajectory of each bus at a uniform sampling interval; mapping it to a two-dimensional grid according to the time dimension and node dimension to form a voltage amplitude matrix; and encoding the matrix into a grayscale image or pseudo-color image of a fixed size; and dividing the dataset into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0016] Furthermore, step S3 specifically involves: using a least-squares generative adversarial network (LSGAN) including a generator and a convolutional discriminator with convolutional-deconvolutional structures. The generator takes random noise and the original critical sample as input and outputs a new voltage heatmap sample. The discriminator simultaneously receives real samples and generated samples and outputs the true / false probabilities. The least-squares loss function is used to constrain the training of the generator and discriminator to improve the consistency between the generated samples and the real critical samples in terms of distribution and local texture.

[0017] Further, step S4 specifically involves: the structural similarity index SSIM calculates the structural similarity score between the generated sample and the corresponding real sample by comprehensively considering the luminance component, contrast component, and structure component. When the score is less than a preset threshold or when the voltage timing is detected to not meet physical constraints such as voltage extreme values ​​and recovery trends, the generated sample is judged as an unqualified sample and removed from the training set.

[0018] The Structural Similarity Index (SSIM) is calculated as follows:

[0019] (1)

[0020] Where x is the generated sample and y is the reference image. The mean, Standard deviation C1 and C2 are covariances, and constants used for numerical stability. Indicates brightness comparison, , These are the pixel mean values ​​of images x and y, respectively; the core of SSIM lies in its quantification of the similarity between two voltage heatmaps in the following three physical dimensions: Indicates contrast comparison, , y and x represent the standard deviations of the image, respectively. Indicates structural comparison, Let x and y be the covariances.

[0021] Luminance corresponds to the average value of the voltage waveform. This reflects the voltage recovery level of the system after a fault is cleared. Under critical steady-state conditions, the steady-state voltage value must be within a specific range. Defined as:

[0022] (2)

[0023] Contrast corresponds to the standard deviation of the voltage waveform, reflecting the amplitude range of voltage oscillations during transient processes. Critical stable samples require that the oscillation amplitude gradually decays and does not exceed a safe threshold. Defined as:

[0024] (3)

[0025] Structure corresponds to covariance: it reflects the shape similarity of voltage waveforms in the time domain, i.e., texture and edge features, that is, it quantifies the voltage change trend and spatiotemporal correlation pattern between nodes. Defined as:

[0026] (4)

[0027] The SSIM value range is [−1, 1]. A higher value indicates better physical consistency of the generated samples in the above three dimensions.

[0028] Furthermore, in step S5, the adaptive hybrid spatiotemporal feature fusion network AHSTF-Net includes: a feature extraction layer for low-level convolutional encoding of the input voltage heatmap; a Transformer temporal feature extraction module connected to the feature extraction layer for capturing voltage evolution dependencies across time scales; a dual-branch deformable residual convolution module connected in parallel with the feature extraction layer for adaptively adjusting the convolution sampling position to extract local voltage drop regions and cross-node spatial coupling features; and a classification layer for globally pooling the fused features output by the above modules and outputting transient voltage stability categories.

[0029] Furthermore, in step S5, the AHSTF-Net model is trained using the Adam optimization algorithm, with learning rate decay and early stopping strategies set.

[0030] Furthermore, step S6 specifically involves: acquiring real-time PMU voltage data within a time window that covers the time window of the fault occurrence and the disturbed trajectory after fault clearance. After normalization and spatial mapping, the PMU voltage data generates an online voltage heatmap, which is input into the trained AHSTF-Net model to obtain the power system transient voltage stability assessment results and evaluate the model performance.

[0031] This invention also discloses a transient voltage stability assessment system that integrates critical sample enhancement and spatiotemporal feature fusion, comprising:

[0032] The data generation module is used to set up power system operating modes and fault scenarios. It calls the transient simulation program to obtain the voltage disturbance trajectory of each bus and uses the golden section method to calculate the critical clearing time t for each operating condition. CCT The simulation samples are labeled as stable, critically stable, and unstable samples according to their stability category, forming the original sample dataset;

[0033] The critical sample augmentation module performs targeted augmentation on samples marked as critically stable in the training set, using the original critically stable samples as real samples to generate a critically stable sample augmentation set.

[0034] The quality screening module is used to calculate the structural similarity index of the generated samples and perform physical consistency detection, and remove samples with scores below the threshold or physical inconsistencies to form a high-quality training set with a balanced distribution.

[0035] The model building module is used to build and train the adaptive hybrid spatiotemporal feature fusion network AHSTF-Net model on the high-quality training set. It takes the voltage heatmap as input and the transient voltage stability category as output to obtain the AHSTF-Net model with optimal weight parameter configuration.

[0036] The transient stability assessment module is used to receive real-time voltage data collected by the PMU in the actual power system, construct a voltage heatmap and input it into the trained AHSTF-Net model to obtain the transient voltage stability assessment results of the power system and evaluate the model performance.

[0037] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method of the present invention.

[0038] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: Without requiring precise knowledge of the details of complex dynamic models, this invention relies solely on limited transient voltage data obtained through simulation and measurement. Through critical sample-oriented enhancement and SSIM sample-by-sample screening, it significantly enriches and purifies stability boundary samples, alleviating the problems of sample scarcity and class imbalance. An adaptive hybrid spatiotemporal feature fusion network, AHSTF-Net, is constructed, organically combining Transformer and deformable residual convolution. This strengthens the correlation mining of multi-bus voltages across time scales and highlights the local distortion and spatial coupling characteristics of voltage drops, improving deep feature extraction capabilities and robustness to noise and operating condition changes. Thus, while ensuring computational efficiency, it significantly improves the accuracy of transient voltage stability assessment and the reliability of online applications. Attached Figure Description

[0039] Figure 1 This is a flowchart of the transient voltage stability assessment method that integrates critical sample enhancement and spatiotemporal feature mining according to the present invention.

[0040] Figure 2 This is a topology diagram of the IEEE 3-machine 9-node system used in the method of this invention.

[0041] Figure 3 The above are the generated sample distributions of different sample augmentation methods in the embodiments of the present invention, where (a) is a generative adversarial network and (b) is a least-squares generative adversarial network.

[0042] Figure 4 This is a schematic diagram of the structure of the adaptive hybrid spatiotemporal feature fusion network AHSTF-Net used in the method of this invention. Detailed Implementation

[0043] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0044] Example 1

[0045] S1. Construct a transient simulation model of the power system. Based on the operating ranges of generators and loads in the system, set various operating modes and fault scenarios, perform time-domain simulation, and obtain the disturbance trajectory of voltage at each bus. Use the golden section method to search for the critical cut-off time t for each operating condition. CCT The samples are labeled as stable samples, critically stable samples, and unstable samples according to their stability category.

[0046] The IEEE 3-machine 9-node system is used as a simulation example, see... Figure 2The system's base capacity is 100 MVA. Multiple operating modes are constructed by adjusting the load level and load composition. Specifically, the load level is selected from three cases: 0.8, 1.0, and 1.2, and the active power output of each generator is adjusted accordingly to maintain power flow balance. The load adopts a comprehensive load model consisting of induction motors and constant impedance parallel connections. The proportion of induction motors varies between 20% and 80%, with a variation step of 10%, for a total of 7 combinations.

[0047] In terms of fault scenario setup, three-phase short-circuit faults were applied to the six AC lines in the system that do not contain transformers. The fault locations were set at 10% to 90% of the line length, with a variation step size of 10%, for a total of nine locations. The fault occurrence time was uniformly set to t0 = 2 s, the total simulation duration was 10 s, and the transient simulation step size was 0.01 s. The voltage disturbance trajectories of all buses before the disturbance, during the fault, and after the fault was cleared were recorded.

[0048] Under each combination of operating mode and fault scenario, the golden section method is used to iteratively search for the system's clearing time under critical stability within a preset fault clearing time search interval, and this time is taken as the critical clearing time t. CCT A total of 3×7×6×9=1134 sets of critical cut-off time data were generated, forming the critical stability surface of the system's transient voltage.

[0049] The samples under various operating conditions were divided into four categories: Referring to practical engineering criteria, a bus voltage consistently below 0.75 pu for more than 1 second after a disturbance was defined as voltage instability. This was based on the system's minimum complete dynamic observation time window (corresponding to a 60Hz system). =1 / 60s≈16.67ms), respectively defined as: stable sample ( ); Critical stable sample ( ); unstable samples ( ).

[0050] S2, the bus voltage disturbance trajectory under each operating condition is normalized and interpolated and resampled according to a preset time window and sampling interval, and the time dimension and node dimension are mapped to a two-dimensional grid to construct a voltage heatmap that characterizes the spatiotemporal evolution of the transient process; a dataset containing {voltage heatmap, transient voltage stability category label} is established, and the dataset is divided into a training set, a validation set and a test set according to a certain proportion.

[0051] A time window covering the short period before the fault, the fault duration, and the disturbed trajectory after the fault is cleared is selected to extract the voltage sequences of each bus. The simulation step size is kept at 0.01 s, and the linear interpolation is a fixed time sampling length.

[0052] Subsequently, the voltages of each bus are normalized, mapping the voltage amplitude to the [0,1] interval. The node dimensions are sorted by bus number or by electrical distance in the system topology, serving as the spatial axis. The sampling time is used as the time axis, mapping the voltage values ​​of each node to pixel intensity on the node-time two-dimensional coordinate system, constructing a voltage heatmap reflecting the spatiotemporal evolution characteristics of voltage drops and recovery for each bus. Each operating condition corresponds to one voltage heatmap, inheriting the stability category label for that condition, thus forming a dataset of {voltage heatmap, transient voltage stability category label}.

[0053] The entire dataset is divided into a training set, a validation set, and a test set in an 8:1:1 ratio. The training set is used for parameter learning, the validation set is used for parameter tuning and early stopping judgment, and the test set is used to evaluate the performance of the model.

[0054] S3 uses the least squares generative adversarial network LSGAN to perform targeted augmentation on samples marked as critically stable in the training set, and generates a critical sample augmentation set using the original critical samples as real samples.

[0055] Specifically, the voltage heatmaps corresponding to the critically stable samples in the training set are input into LSGAN. The generator adopts a convolution-deconvolution structure, and the discriminator adopts a multi-layer convolution downsampling structure to obtain generated samples with the same size as the real voltage heatmaps.

[0056] Through iterative training, new critical stability voltage heatmaps are generated in batches, forming a critical stability sample enhancement set. For example... Figure 3 As shown, compared with GAN, after enhancement by LSGAN, the distribution of critically stable generated samples in the feature space is more balanced, and the sample density near the stability boundary is significantly improved.

[0057] S4. The SSIM index is used to screen the generated critically stable samples. The structural similarity score is calculated based on brightness, contrast and structural components. Samples with scores below the threshold or physical defects are removed.

[0058] The structural similarity index (SSIM) is used to screen samples in the critical sample augmentation set for quality control. Taking the generated sample x and its corresponding real sample y as an example, the luminance, contrast, and structure components of the two samples are calculated, and the structural similarity score between the generated sample and the corresponding real sample is calculated comprehensively.

[0059] The Structural Similarity Index (SSIM) is calculated as follows:

[0060] (1)

[0061] Where x is the generated sample and y is the reference image. The mean, Standard deviation Let C1 and C2 be the covariances, and C1 and C2 be constants used for numerical stability. Indicates brightness comparison, , These are the pixel mean values ​​of images x and y, respectively; the core of SSIM lies in its quantification of the similarity between two voltage heatmaps in the following three physical dimensions: Indicates contrast comparison, , y and x represent the standard deviations of the image, respectively. Indicates structural comparison, Let x and y be the covariances.

[0062] Luminance corresponds to the average value of the voltage waveform. This reflects the voltage recovery level of the system after a fault is cleared. Under critical steady-state conditions, the steady-state voltage value must be within a specific range. Defined as:

[0063] (2)

[0064] Contrast corresponds to the standard deviation of the voltage waveform, reflecting the amplitude range of voltage oscillations during transient processes. Critical stable samples require that the oscillation amplitude gradually decays and does not exceed a safe threshold. Defined as:

[0065] (3)

[0066] Structure corresponds to covariance: it reflects the shape similarity of voltage waveforms in the time domain, i.e., texture and edge features, that is, it quantifies the voltage change trend and spatiotemporal correlation pattern between nodes. Defined as:

[0067] (4)

[0068] The SSIM value range is [−1, 1]. A higher value indicates better physical consistency of the generated samples in the above three dimensions.

[0069] S5, an adaptive hybrid spatiotemporal feature fusion network AHSTF-Net is constructed based on the high-quality training set, taking voltage heatmaps as input and transient voltage stability categories as outputs. AHSTF-Net includes: a feature extraction layer for low-level convolutional encoding of the input voltage heatmap; a Transformer temporal feature extraction module connected to the feature extraction layer to capture voltage evolution dependencies across time scales; a dual-branch deformable residual convolution module connected in parallel with the feature extraction layer to adaptively adjust the convolution sampling position to extract local voltage drop regions and cross-node spatial coupling features; and a classification layer for global pooling of the fused features output by the above modules and outputting transient voltage stability categories. AHSTF-Net is trained based on the training and validation sets to obtain the AHSTF-Net network hyperparameters.

[0070] The overall structure of the constructed adaptive hybrid spatiotemporal feature fusion network AHSTF-Net is shown in [link to full text]. Figure 4 The network input is a single-channel or multi-channel voltage heatmap, which yields transient voltage stability category labels and corresponding stability margins.

[0071] In terms of temporal feature extraction, the Transformer temporal feature extraction module is used to construct the relationship between different moments of the voltage heatmap on the time scale. The temporal location information is explicitly introduced through location encoding, enabling the network to capture the dynamic relationship between the moment of fault occurrence, the fault duration stage, and the recovery stage after fault clearance.

[0072] For spatial feature extraction, a dual-branch deformable residual convolution module is constructed: one branch uses standard convolution to extract global background features; the other branch uses deformable convolution kernels to adaptively learn sampling offsets for voltage drop regions and key busbars, thereby enhancing the characterization of local distortion patterns and cross-node spatial coupling effects. The outputs of the two branches are fused through residual connections, which helps alleviate the gradient vanishing problem in deep networks and improves training stability.

[0073] A fully connected classification layer is set at the top layer of the network to map the fused spatiotemporal features into multi-class output probabilities. During training, the cross-entropy loss function is used as the optimization objective, and the Adam algorithm is selected for optimization. Learning rate decay and early stopping strategies are combined to suppress overfitting. Performance evaluation metrics include classification accuracy (ACC), false negative rate (FNR), false positive rate (FPR), geometric mean (G-mean), and F1 score (F1-score), focusing on the ability to identify unstable and critical samples, as shown in Table 1.

[0074] Table 1 Performance Evaluation Indicators

[0075]

[0076] In Table 1, TP and TN represent the number of actual unstable and actual stable samples that were correctly identified, respectively. FP refers to samples that were actually stable but were misclassified as unstable, and FN refers to samples that were actually unstable but were missed as stable.

[0077] S6. Input the test set data into the trained AHSTF-Net model to obtain the transient voltage evaluation results of the power system and evaluate the model performance.

[0078] The results of the comparative experiments on different methods based on the test set data are shown in Table 2. Compared with traditional machine learning models such as SVM, CNN, LSTM, PCNN, Transformer, and DCN, the AHSTF-Net constructed in this embodiment achieves better results in terms of overall accuracy and critical sample recognition performance, and has higher evaluation accuracy and robustness.

[0079] Table 2 Comparison of different algorithms

[0080]

[0081] It can be seen that:

[0082] 1) Compared with the traditional machine learning model SVM, the proposed model improves ACC from 97.34% to 99.89%, reduces FNR from 2.31% to 0, reduces FPR from 6.64% to 0.16%, increases G-mean from 95.50% to 99.86%, and increases F1-score from 97.25% to 99.82%. Compared with typical deep learning models CNN, LSTM, PCNN, and Transformer, the proposed model also reaches or approaches the optimal in terms of ACC, FNR, FPR, G-mean, and F1-score. The reduction of FNR to 0 indicates that the proposed model is significantly better than traditional machine learning models and conventional deep learning models in terms of overall accuracy, boundary sample identification, and control of missed detection of unstable samples.

[0083] 2) When compared with the baseline model DCN, which has the best performance, the ACC of the model in this invention is further improved from 99.57% to 99.89%, the FPR is reduced from 0.47% to 0.16%, and the G-mean and F1-score are improved from 99.57% and 99.54% to 99.86% and 99.82%, respectively. This result shows that the adaptive hybrid spatiotemporal feature fusion network combining deformable convolution and Transformer in the structural layer can significantly enhance the discrimination ability of samples near the stability boundary, and achieve a more refined and reliable assessment of the transient voltage stability state of the power system while maintaining high overall accuracy.

[0084] Secondly, this invention provides a transient voltage stability assessment system that integrates critical sample enhancement and spatiotemporal feature mining, comprising:

[0085] Data generation module: Used to construct transient simulation models based on the basic parameters and operating constraints of the power system, perform time-domain simulations under various load levels, load compositions, fault types, and fault locations, record the disturbance trajectory of voltage at each bus, and calculate the critical clearing time t for each operating condition using the golden section method. CCT The simulation samples are labeled as stable, critically stable, and unstable samples according to their stability category.

[0086] Sample augmentation module: Used to augment samples marked as critically stable in the training set. It adopts the least squares generative adversarial network LSGAN, using the original critically stable samples as real samples to generate a critically stable sample augmentation set.

[0087] Sample screening module: Used to screen the generated critically stable samples for quality. It calculates the structural similarity score from three dimensions: brightness, contrast and structure, and removes samples with scores below the threshold or physical unreasonableness to form a high-quality training set with a balanced distribution.

[0088] Modeling Module: This module is used to build and train the adaptive hybrid spatiotemporal feature fusion network AHSTF-Net on the training set. It takes voltage heatmaps as input and transient voltage stability categories as output. This module includes convolutional coding layers, a Transformer temporal feature extraction module, a dual-branch deformable residual convolutional module, and a fully connected output layer. The AHSTF-Net model is trained using the training and validation sets.

[0089] Prediction module: During online operation, it generates voltage heatmaps from voltage data collected by the PMU in real time and inputs them into the trained AHSTF-Net model to obtain stable category prediction values, providing a basis for scheduling and control decisions.

[0090] It should be noted that the data generation module, sample enhancement module, sample screening module, modeling module and prediction module mentioned above correspond to steps S1 to S6 in the implementation method. The examples and application scenarios implemented by the above modules and corresponding steps are the same, but are not limited to the content disclosed above.

[0091] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute through the processor, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 Devices that specify functions in one or more boxes.

[0092] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes.

[0093] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps that specify the function in one or more boxes.

[0094] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A transient voltage stability assessment method integrating critical sample enhancement and spatiotemporal feature mining, characterized in that, Includes the following steps: S1. Construct a power system simulation model, setting various operating modes and fault scenarios based on the operating range of generator output and load power in the system. Obtain the bus voltage disturbance trajectory based on time-domain simulation; use the golden section method to search for the critical clearing time t for each operating condition. CCT The simulation samples are labeled as stable, critically stable, and unstable samples according to their stability category; S2, the bus voltage disturbance trajectory is resampled and normalized according to a preset time window and sampling interval, and converted into a voltage heatmap that characterizes the spatiotemporal evolution of the transient process. A dataset containing voltage heatmap and transient voltage stability category label is established, and the dataset is divided into training set, validation set and test set according to a certain ratio. S3 uses the least squares generative adversarial network LSGAN to perform targeted augmentation on the samples marked as critically stable in the training set, and uses the original critically stable samples as real samples to generate an augmented set of critically stable samples. S4. The generated critically stable samples are screened for quality using structural similarity index. Structural similarity scores are calculated based on brightness, contrast and structural components. Samples with scores below the threshold or physical inconsistencies are removed to form a high-quality training set with balanced distribution. S5. Construct an adaptive hybrid spatiotemporal feature fusion network AHSTF-Net, taking voltage heatmap as input and transient voltage stability category as output. The network internally sets a Transformer module to extract temporal features and a deformable residual convolution module to extract multi-spatial scale features. The AHSTF-Net model is trained based on training set and validation set data. S6. Input the test set data into the trained AHSTF-Net model to obtain the power system transient voltage stability assessment results and evaluate the model performance.

2. The transient voltage stability assessment method integrating critical sample enhancement and spatiotemporal feature mining according to claim 1, characterized in that, In step S1, a power system simulation model is constructed, and the example system is the IEEE 3-machine 9-bus system. The operation mode is obtained by randomly generating several sets of power generation and load conditions within the preset active and reactive power output fluctuation range. The fault scenarios include three-phase short-circuit faults on different buses and different fault durations. The golden section method is used to search for the fault clearing time that ensures the voltage meets the transient voltage qualification criterion after fault clearing, and this time is taken as the critical clearing time t. CCT Based on this, the transient voltage stability margin is calculated.

3. The transient voltage stability assessment method integrating critical sample enhancement and spatiotemporal feature mining according to claim 1, characterized in that, In step S2, the voltage heatmap construction process includes: selecting a time window covering the disturbance trajectory before, during, and after the fault is cleared; interpolating and normalizing the disturbance trajectory of each bus voltage at a uniform sampling interval; mapping it to a two-dimensional grid according to the time dimension and node dimension to form a voltage amplitude matrix; and encoding the matrix into a grayscale image or pseudo-color image of a fixed size; and dividing the dataset into a training set, a validation set, and a test set in a ratio of 8:1:

1.

4. The transient voltage stability assessment method integrating critical sample enhancement and spatiotemporal feature mining according to claim 1, characterized in that, Step S3 specifically involves using a least-squares generative adversarial network (LSGAN), which includes a generator and a convolutional discriminator with convolutional-deconvolutional structures. The generator takes random noise and the original critical sample as input and outputs a new voltage heatmap sample. The discriminator simultaneously receives real samples and generated samples and outputs the true and false probabilities. The least-squares loss function is used to constrain the training of the generator and discriminator to improve the consistency between the generated samples and the real critical samples in terms of distribution and local texture.

5. The transient voltage stability assessment method integrating critical sample enhancement and spatiotemporal feature mining according to claim 1, characterized in that, Step S4 is as follows: The structural similarity index SSIM calculates the structural similarity score between the generated sample and the corresponding real sample by comprehensively calculating the luminance component, contrast component and structure component. When the score is less than the preset threshold or the voltage timing is detected to not meet the physical constraints such as voltage extreme value and recovery trend, the generated sample is judged as an unqualified sample and removed from the training set. The Structural Similarity Index (SSIM) is calculated as follows: （1） Where x is the generated sample and y is the reference image. The mean, Standard deviation, C1 and C2 are covariances, and constants used for numerical stability. Indicates brightness comparison, , y and x are the pixel mean values ​​of the image, respectively; The core of SSIM lies in its quantification of the similarity between two voltage heatmaps in the following three physical dimensions: Indicates contrast comparison, , y and x represent the standard deviations of the image, respectively. Indicates structural comparison, Let x and y be the covariances. Luminance corresponds to the average value of the voltage waveform. This reflects the voltage recovery level of the system after a fault is cleared. Under critical steady-state conditions, the steady-state voltage value must be within a specific range. Defined as: （2） Contrast corresponds to the standard deviation of the voltage waveform, reflecting the amplitude range of voltage oscillations during transient processes. Critical stable samples require that the oscillation amplitude gradually decays and does not exceed a safe threshold. Defined as: （3） Structure corresponds to covariance: it reflects the shape similarity of voltage waveforms in the time domain, i.e., texture and edge features, that is, it quantifies the voltage change trend and spatiotemporal correlation pattern between nodes. Defined as: （4） The SSIM value range is [−1, 1]. A higher value indicates better physical consistency of the generated samples in the above three dimensions.

6. The transient voltage stability assessment method integrating critical sample enhancement and spatiotemporal feature mining according to claim 1, characterized in that, In step S5, the adaptive hybrid spatiotemporal feature fusion network AHSTF-Net includes: a feature extraction layer for low-level convolutional encoding of the input voltage heatmap; a Transformer temporal feature extraction module connected to the feature extraction layer for capturing voltage evolution dependencies across time scales; a dual-branch deformable residual convolution module connected in parallel with the feature extraction layer for adaptively adjusting the convolution sampling position to extract local voltage drop regions and cross-node spatial coupling features; and a classification layer for globally pooling the fused features output by the above modules and outputting transient voltage stability categories.

7. The transient voltage stability assessment method integrating critical sample enhancement and spatiotemporal feature mining according to claim 1, characterized in that, In step S5, the AHSTF-Net model is trained using the Adam optimization algorithm, with learning rate decay and early stopping strategies set to improve model convergence speed and suppress overfitting.

8. The transient voltage stability evaluation method integrating critical sample enhancement and spatiotemporal feature mining according to claim 1, characterized in that, The performance evaluation metrics for step S6 include classification accuracy (ACC), false negative rate (FNR), false positive rate (FPR), geometric mean (G-mean), and F1 score (F1-score), with a focus on the ability to identify unstable and critical samples.

9. A transient voltage stability assessment system based on fusion critical sample enhancement and spatiotemporal feature mining, used to implement the method of claim 1, characterized in that, include: The data generation module is used to call the transient simulation program to obtain the voltage disturbance trajectory of each bus based on the power system operation mode and fault scenario settings, calculate the critical clearing time under each operating condition using the golden section method, and mark the samples as stable, critically stable and unstable samples according to the critical clearing time to form the original sample dataset. The critically stable sample augmentation module is used to input the critically stable samples in the original sample dataset into a least-squares generative adversarial network for targeted sample augmentation, thereby obtaining a critically stable sample augmentation set. The quality screening module is used to calculate the structural similarity index of the enhanced critical samples and perform physical consistency detection to remove unqualified samples and obtain high-quality samples with balanced distribution. The model building module constructs a hybrid spatiotemporal feature fusion network AHSTF-Net, taking voltage heatmaps as input and transient voltage stability categories as output, to obtain the AHSTF-Net model with optimal weight parameter configuration; The transient stability assessment module inputs test set data into the trained AHSTF-Net model and evaluates the transient voltage stability state and stability margin of the power system based on the model output results.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method of claim 1.

Images