Improved TS-GAN-based power grid fault recording data enhancement method and system

Abstract

The invention relates to the technical field of data enhancement, in particular to a power grid fault recording data enhancement method and system based on an improved TS-GAN. The method comprises the following steps: acquiring small sample fault recording data; preprocessing the acquired small sample fault recording data; performing feature extraction on the small sample fault recording data by using an improved TS-GAN-based fault generation model; a fault generation model based on an improved TS-GAN is used to generate and discriminate faults for the extracted features; performing model training by taking small sample fault recording data as a sample; evaluating and optimizing the trained model; and performing data enhancement on the power grid fault recording data by using the optimized model. The TS-GAN is improved, and learning of local feature differences is enhanced by fusing a convolution self-attention layer and a multi-head self-attention layer, so that a generator is helped to more accurately and comprehensively grab local and global features of a time sequence.

Description

Technical Field

[0001] The present invention relates to the technical field of data augmentation, and particularly to a method and system for augmenting power grid fault recording data based on an improved TS-GAN. Background Art

[0002] As the structure and operation of the power system become more complex, once a power grid fault occurs, in the case where the fault cannot be quickly cleared, a simple fault may evolve into a more serious developing fault. Due to the small number of samples of these faults, it is relatively difficult to discriminate using conventional fault diagnosis methods. If these faults are not recognized and isolated in time, it may lead to misoperations of protection devices and circuit breakers, thereby causing a secondary impact on the power grid. In particular, three-phase short-circuit faults caused by single-phase grounding short-circuit faults, two-phase grounding short-circuit faults caused by single-phase grounding short-circuit faults, and three-phase short-circuit faults developed from two-phase short-circuit grounding faults. These small-sample faults usually refer to the evolution from a single fault to other types of faults, and it is often impossible to effectively discriminate the faults by only analyzing the waveforms when the faults occur. In addition, the existence of these small-sample data often leads to insufficient training of the diagnosis model and affects the discrimination rate of the model.

[0003] To solve the problem of low discrimination accuracy in the case of small-sample faults, there is an urgent need for a method for augmenting power grid fault recording data based on an improved TS-GAN. After preprocessing the fault recording data, it is input into a designed TS-GAN model for training to generate simulation samples for subsequent model training. Summary of the Invention

[0004] To solve the problems of small amount of fault sample data and data imbalance generated when faults occur in different power grids. The present invention provides a method and system for augmenting power grid fault recording data based on an improved TS-GAN. With the advantage of the time series generative adversarial network in sample augmentation, small-sample faults are augmented with data and used to train subsequent models.

[0005] In the first aspect, a method for augmenting power grid fault recording data based on an improved TS-GAN provided by the present invention adopts the following technical solution: A method for augmenting power grid fault recording data based on an improved TS-GAN includes: Obtain small-sample fault recording data; Preprocess the obtained small-sample fault recording data; Use a fault generation model based on an improved TS-GAN to extract features from the small-sample fault recording data; Use a fault generation model based on an improved TS-GAN to generate and discriminate faults for the extracted features; Use small-sample fault recording data as samples for model training; Evaluate and optimize the trained model; Use the optimized model to perform data augmentation on power grid fault recording data.

[0006] Furthermore, the preprocessing of the obtained small-sample fault recording data includes, for a real recording sequence with 2400 sampling points, gradually sampling each sampling point starting from the recording start point, so as to expand a single real fault data sequence into a set of training data sets containing multiple sampling points. Each segment corresponds to consecutive time points in the original sequence, which is used to increase the quantity and diversity of training data. Among them, in the fault recording data, the electrical quantity indicators form a time series process , where represents discrete time points, represents fault characteristic parameters, is the fault parameter space, and the trajectory sequence is a discrete set. The electrical quantity state can be expressed as . If the electrical quantity of the fault recording data generated by the model is , and it follows the probability distribution , then the probability distribution generated by the model needs to be close to the sequence distribution in the real scenario ; The performance in the data generation model is the parameter , represents the model parameter space, and satisfies: , where: represents the measure of the distance between different distributions.

[0007] Furthermore, the feature extraction of small-sample fault recording data using the improved TS-GAN fault generation model includes first integrating the global attention mechanism GLFM. The features of the sequence extracted by the multi-head attention layer are further learned through the feed-forward layer with a multi-layer perceptron to obtain a more complex non-linear mapping, enabling the model to learn a richer representation. Finally, the local correlation features of the sequence are extracted through the attention layer, and the information captured by each layer is fused through residual connections between the multi-head self-attention layer, the feed-forward layer, and the convolutional attention layer. Among them, in the first layer, the standard WGAN uses a random vector to generate a synthetic spectrogram, and in the second layer, by introducing conditional WGAN, the synthetic image generated in the first layer is transformed into an accurate synthetic time series.

[0008] Furthermore, the feature extraction of the small-sample fault recording data by using the improved TS-GAN fault generation model further includes enabling the model to have the ability of dynamic feature selection through an attention mechanism embedded in a lightweight architecture, and autonomously focusing on the key signal changes during the fault occurrence period. Among them, first, the query matrix is calculated and the key matrix to obtain the similarity. Subsequently, by dividing by , and through the function to perform normalization processing on it to obtain the score matrix ; then multiply the score matrix by the value matrix to calculate and obtain the attention matrix : The input of the attention matrix is linearly transformed to obtain the output formula of this layer as: where is the learnable network weight parameter.

[0009] Furthermore, the use of the improved TS-GAN-based fault generation model to generate and discriminate faults for the extracted features includes using a discriminator to evaluate the input hidden coding sequence and determine whether it comes from the true data distribution. Among them, the discriminator first scores the hidden coding sequence through the function , represents the probability that the sequence belongs to the true fault sequence, and outputs a binary classification result through feature comparison. Then, the random sequence is converted into a hidden coding by using the coding layer; secondly, the conditional distribution is calculated , and the temporal coherence of the generated sequence is ensured by modeling the Markov dependence of the hidden state, represents the temporal dynamic characteristics in the hidden coding sequence, and the second layer further approximates the true conditional distribution by recalculating to ensure that the generated hidden coding sequence is close to the distribution of the actual fault sequence.

[0010] Furthermore, the use of the improved TS-GAN-based fault generation model to generate and discriminate faults for the extracted features also includes using the generator for the original fault samples. Among them, first, the random sequence is converted into a hidden coding by using the coding layer; secondly, the conditional distribution is calculated , and the temporal coherence of the generated sequence is ensured by modeling the Markov dependence of the hidden state, so that the hidden coding of the generated fault sequence has probability characteristics similar to those of the true sequence.​​​ Represents the temporal dynamic characteristics in the latent encoding sequence. The second layer further approximates the true conditional distribution through re - calculation to ensure that the generated latent encoding sequence is close to the distribution of the actual fault sequence.

[0011] Furthermore, the model is trained using small - sample fault recording data as samples, including contrastive learning by calculating the similarity of samples to distinguish similar and dissimilar samples, promoting the model to obtain more discriminative feature representations. The training process of contrastive learning includes, in each iteration, randomly selecting n samples from the training set and applying a random augmentation strategy to generate n samples. For any sample, only its augmented version is regarded as a positive sample, while the remaining 2n - 2 samples in the batch are used as negative samples to construct a contrastive learning task; then all samples are encoded using the Generalized Linear Feature Map (GLFM) to obtain low - dimensional feature representations; finally, the loss function calculates the similarity of positive and negative sample pairs through the normalized temperature scaling strategy to calculate the contrastive loss to achieve discriminative learning in the feature space. The contrastive loss is defined as: , where: represents the contrastive learning coefficient, is the similarity function, using the cosine similarity z as the similarity function, which represents the dot product between and after L2 regularization.

[0012] Furthermore, the evaluation and optimization of the trained model include objectively analyzing the generated fault samples through three metrics, specifically including: quantifying the difference between the generated fault samples and the simulation samples using the L - distance; measuring the similarity between the two samples through brightness, contrast, and structural information; evaluating the signal quality of the generated samples using the peak signal - to - noise ratio of the generated samples. Among them, for two fault samples and , their structural similarity index is expressed as: , where: is the mean of sample ; is the mean of the generated sample ; is the variance of sample ; is the variance of the generated sample ; is and covariance.

[0013] Furthermore, the evaluation and optimization of the trained model also include analyzing the similarity between the real electrical quantities and the generated electrical quantities from two aspects: time domain and frequency domain. In the time domain, the parameters of the generated small-sample fault sequences are quantified for similarity using the Hellinger distance HD and the mean absolute error MAE. In the frequency domain, the local spectral differences between the generated data and the real data are calculated, and the average spectral error is calculated to analyze the similarity between each generated fault sequence and the real fault sequence, expressed as: , , where: represents the frequency domain representation of the generated data, represents the frequency domain representation of the real fault sequence, represents the spectrum of the real fault sequence that is most similar to the generated fault sequence , and are the serial numbers of the generated fault sequence and the real fault sequence respectively, is the spectrum segmentation serial number, represents the number of generated samples, is the number of real samples, is the sum of the number of spectrum segments.

[0014] Furthermore, the evaluation and optimization of the trained model also include using the t-SNE dimensionality reduction method to project 1000 randomly selected samples onto a two-dimensional plane while maintaining the topological relationship of the high-dimensional data. By analyzing the distribution coincidence degree of the samples generated by TS-GAN and the real samples in the two-dimensional space, the data distribution fitting ability of the generation model under the same training parameters is verified. Among them, within the given parameter space, a circular neighborhood with a radius of d is constructed centered on each sample point; then the number of sample points contained in each neighborhood is counted, and the average value of the statistical results of all sample points is taken; the average value is divided by the instance point density within the parameter interval to obtain the Rispley's function value at a distance of , expressed as: , where: is the area of the parameter interval; is the number of sample points in the parameter area; is the distance between different sample points; characterizes the positional relationship between the sample point and the circle.

[0015] In a second aspect, a power grid fault recording data enhancement system based on an improved TS-GAN includes: A data acquisition module configured to acquire small-sample fault recording data; A preprocessing module configured to preprocess the acquired small-sample fault recording data; A feature extraction module configured to extract features from the small-sample fault recording data by using a fault generation model based on an improved TS-GAN; A fault generation module configured to generate and discriminate faults from the extracted features by using a fault generation model based on an improved TS-GAN; A training module configured to perform model training by using the small-sample fault recording data as samples; An optimization module configured to evaluate and optimize the trained model; An enhancement module configured to perform data enhancement on the power grid fault recording data by using the optimized model.

[0016] In a third aspect, the present invention provides a computer-readable storage medium storing multiple instructions, and the instructions are suitable for being loaded and executed by a processor of a terminal device to perform the method for enhancing power grid fault recording data based on an improved TS-GAN.

[0017] In a fourth aspect, the present invention provides a terminal device including a processor and a computer-readable storage medium. The processor is used to implement each instruction; the computer-readable storage medium is used to store multiple instructions, and the instructions are suitable for being loaded and executed by the processor to perform the method for enhancing power grid fault recording data based on an improved TS-GAN.

[0018] In summary, the present invention has the following beneficial technical effects: By first analyzing the generation causes and properties of small-sample data, relying on the excellent ability of TS-GAN in sample generation and expansion, extracting and learning and fusing through a small-sample fault generation model for model training and obtaining a large number of small-sample faults. Secondly, by generating fault samples containing small-sample data based on TS-GAN and combining with subsequent models, the diagnostic ability of the fault diagnosis model is better improved, thereby improving the overall accuracy of the research method; The TS-GAN itself is improved by fusing a convolutional self-attention layer and a multi-head self-attention layer to strengthen the learning of local feature differences, so as to help the generator more accurately and comprehensively capture the local and global features of the time series, and ensure that the discriminator does not reduce the synthesis quality of a single variable in the channel while paying attention to the overall distribution features of the multi-variable time series, and can better test the diagnostic ability of the fault diagnosis model. Description of the Drawings

[0019] Figure 1 It is a schematic diagram of the improved TS-GAN model architecture in Embodiment 1 of the present invention.

[0020] Figure 2 It is a schematic diagram of small sample data generation based on TS-GAN in Embodiment 1 of the present invention.

[0021] Figure 3 It is a schematic diagram of fault generation in Embodiment 1 of the present invention.

[0022] Figure 4 It is the distribution of real fault sequences in Embodiment 1 of the present invention.

[0023] Figure 5 It is the generated fault sequence distribution diagram in Embodiment 1 of the present invention.

[0024] Figure 6 It is the Rispley's K curve diagram of the instance point distribution in Embodiment 1 of the present invention. Detailed implementation manner

[0025] The present invention will be further described in detail below with reference to the accompanying drawings.

[0026] Embodiment 1 Refer to Figure 1 , a method for enhancing power grid fault recording data based on an improved TS-GAN in this embodiment includes: In this embodiment, the actual fault recording data set of Dongfang Electronics is adopted to obtain three-phase current and three-phase voltage data. Four types of faults including A, B, and C phase grounding are simulated. Then the simulated data is combined with the real power grid fault recording data. 1853 small sample faults are extracted from the real fault recording data, and the small sample fault types are divided into 3 categories. In the real power grid fault occurrence scenario, the scenario instances with a fault recording duration within 400 ms account for more than 90%. For each actual occurrence instance, 2400 sample data points are collected within 200 milliseconds before and after the time point when the fault appears, so as to construct a data set of real small sample faults. If the data volume in the extracted fault recording is insufficient, it will be manually supplemented to 2400 data points according to the change trend of the sequence.

[0027] S1 Data processing, The fault record data in this dataset covers the electrical parameters at various time points. This data processing method ensures the integrity and accuracy of the dataset, and can effectively supplement it even when the amount of original data is insufficient, thus laying a foundation for the reliability of subsequent model training. In this way, researchers can more accurately capture the specific situation of power grid faults, and then improve the safety and stability of the power grid. When the sampling frequency is large, the electrical quantity sequence can be estimated from adjacent sample points. Therefore, taking time as a measure, the following data: , are used as the input data of TS-GAN.

[0028] In the fault recording data, the electrical quantity indicators constitute a time series process , where represents discrete time points, represents the fault characteristic parameters, is the fault parameter space. These parameters follow a probability distribution , and the trajectory sequence as a discrete time set realizes the discretized representation of a continuous process. Since the trajectory sequence is a discrete set, the electrical quantity state can be expressed as . If the electrical quantity of the fault recording data generated by the model is , and it follows a probability distribution , then the probability distribution generated by the model needs to be close to the sequence distribution in the real scenario . In the data generation model , the performance of the parameters , (representing the model parameter space) satisfies the following formula: , In the formula: represents the measure of the distance between different distributions.

[0029] In this paper, the sampling frequency of the input sequence of the generative model is set to 2000 Hz. For the real recorded wave sequence with 2400 sampling points of data, starting from the starting point of the recorded wave, each sampling point is gradually sampled, so as to expand a single real fault data sequence into a training data set containing multiple sampling points. Each segment corresponds to consecutive time points in the original sequence. By this method, the quantity and diversity of the training data can be significantly increased. This sliding window method not only improves the data utilization rate, but also enhances the model's ability to learn and predict under different conditions, thus better simulating the actual power grid fault situation. This method ensures that the model can more accurately understand and capture the time dynamic characteristics of the power grid fault, improves the robustness and generalization ability of the model, and ultimately helps to improve the safety and stability of the power grid.

[0030] Improvement of the TS-GAN model architecture, The basic architecture of the TS-GAN applied in this paper consists of two WGANs (Wasserstein generative adversarial networks): a standard WGAN and a conditional WGAN. In this paper, the WGAN uses the Wasserstein distance to measure the similarity between two probability distributions, which is used to simulate the real distribution of the synthetic data. Through this two-layer TS-GAN architecture, time series data can be effectively generated.

[0031] Improvements are proposed for the traditional TS-GAN structure model, such as Figure 3 As shown, first, the global attention mechanism GLFM is incorporated. The features of the sequence extracted by the multi-head attention layer are further learned through a feed-forward layer with a multi-layer perceptron to a more complex non-linear mapping, enabling the model to learn a richer representation. Finally, the local correlation features of the sequence are extracted through the attention layer, and the information captured by each layer is fused through residual connections between the multi-head self-attention layer, the feed-forward layer, and the convolutional attention layer.

[0032] Figure 1 The above is the improved TS-GAN model architecture.

[0033] Among them, in the first layer, the standard WGAN uses a random vector to generate a synthetic spectrogram. This process generates preliminary synthetic data, which, although somewhat similar to the real data, is not precise enough. In the second layer, by introducing a conditional WGAN, the synthetic image generated in the first layer is transformed into a more precise synthetic time series. The conditional WGAN plays a key role in this process. It further optimizes and adjusts the generated data according to specific conditions to make it closer to the actual time series data.

[0034] With this two-stage method, TS-GAN can not only generate high-quality synthetic time series data, but also significantly improve the accuracy and reliability of subsequent model recognition. This method enables researchers to better simulate and analyze power grid faults, thereby enhancing the security and stability of the power grid.

[0035] The embedding of the attention module endows the model with the ability of dynamic feature selection, enabling it to autonomously focus on the key signal changes during the fault occurrence period. This lightweight architecture has the following advantages compared to traditional deep networks: reducing the number of parameters and shortening the average diagnostic response time. Analyzing data only relying on the eigenvalues of the random matrix without considering the eigenvectors may lead to the neglect of the key information of the signal.

[0036] The multi-head attention used in the GLFM module is calculated as follows: First, calculate the query matrix and the key matrix of to obtain the similarity degree. Subsequently, divide it by and normalize it through the function to obtain the score matrix : (3-1) Then, multiply the score matrix by the value matrix to calculate the attention matrix : (3-2) The input of the attention matrix undergoes a linear transformation to obtain the output formula of this layer as follows, where are learnable network weight parameters.

[0037] (3-3) The calculation process of the convolutional self-attention layer is as follows, where represents a one-dimensional convolution with a convolution kernel size of . First, divide the input from the previous layer into multiple local blocks , and then use one-dimensional convolution operations to learn the of each local block . Among them: (3-4) In the formula: represents the weight at the position of the convolution kernel ; represents the convolution kernel size; represents the channel size of the input and output; Indicates the feature position on the local block of the feature tensor

[0038] During the experiment, select convolution kernels to focus on short-term temporal correlations. Such short-term correlations are more critical for the local details of the sequence

[0039] Calculate and the dot product of, and then divide by the square root of the dimension of the key vector and use function to normalize it and multiply with the value matrix to obtain the attention matrix between the chunks and the blocks : (3 - 5) The convolutional self-attention layer performs convolutional operations on the input sequence using groups of distinct convolution kernels and calculates the attention matrix , and is linearly transformed to obtain the final output result of this layer : (3 - 6) S3 Fault recording generation model based on improved TS-GAN In the data generation process, for the generator, first generate an image through a random noise vector (following a normal distribution) together with the text feature vector; then the generated image will be continuously improved during the training process and gradually approach the real image. The function of the discriminator is to judge whether the input image comes from a real sample (or an image generated by the generator).

[0040] The network structure includes: The input layer receives real images and generated images and combines the input text features; the convolutional layers extract image features through multiple convolutions and use the Leaky or ReLU activation function to avoid gradient vanishing; feature fusion combines text features and temporal features; they can be weighted and fused through concatenation operations or using attention mechanisms; after a series of convolutional and pooling layers, the fully connected layer outputs the final judgment result. Finally, an output layer outputs a probability value indicating that the sample is a real sample. The training process of TS-GAN is adversarial, and the generator and discriminator are alternately trained in two stages

[0041] The core of TS-GAN lies in the accurate modeling of the temporal conditional probability. For a given electrical quantity sequence , the model focuses on capturing the current state ​With historical data The conditional dependency . According to the probability model theory, the joint distribution can be decomposed into . The optimization objective of the generative model is to make the approximate conditional distribution at any time approach the true conditional distribution . The fault generation model satisfies: (3-7) To ensure the model's learning ability for the dynamic characteristics of the fault sequence, this paper follows the design of the latent coding space in the TS-GAN framework. This coding space can effectively characterize the conditional distribution characteristics of the fault sequence and provide the necessary temporal information modeling ability for the generative model. The basic structure is as Figure 2 shown. Multilevel gated recurrent units (Gated Recurrent Unit, GRU) are used to connect between modules. As a specialized architecture of the recurrent neural network, it is designed specifically for temporal data modeling. It dynamically regulates the information transfer path through an improved gating system (reset gate and update gate), effectively alleviating the long-term dependence learning difficulty problem of traditional RNNs and showing excellent temporal modeling ability in sequence analysis tasks.

[0042] Figure 2 The calculation process of the small-sample data generation model based on TS-GAN is shown in

[0043] As Figure 2 shown: - Loss function set; - Component loss; - Represent the discriminant outputs of the generated sequence and the true sequence respectively; - Reconstructed sequence; - Auxiliary discriminant parameter; - Represent the latent encodings of the generated and true data respectively. Through this design, the model can focus on the time dynamic characteristics of the fault sequence data without paying too much attention to static features. Moreover, the model can better simulate and predict the relationship between sample points in power grid faults. Each module uses multilevel GRUs to process sequence data, thus enhancing the model's understanding and processing ability of time series.

[0044] In the fault sample generation model, the core components of this generative model include: (1) Encoder : Establish a temporal feature space Mapping to the latent space The mapping to the latent space encodes the original sequence into a low-dimensional representation through a function; (3 - 8) (2) State recurrence calculation: The hidden state at each moment inherits historical information and fuses the current input; (3 - 9) (3) Decoder : It realizes the inverse transformation from the latent space to the feature space through calculation and outputs the reconstructed sequence. This bidirectional mapping mechanism completely preserves the temporal dynamic characteristics of the sequence.

[0045] (3 - 10) The calculation processes of the discriminator and the generator of the TS - GAN model are both completed in the latent coding space. The specific calculation steps are as follows: (1) The main task of the discriminator is to evaluate the input latent coding sequence and determine whether it comes from the real data distribution. The discriminator first scores the latent coding sequence through the function The output score ranges from 0 to 1, indicating the probability that the sequence belongs to the real fault sequence. The core function of this module is to perform discriminant analysis on the latent codings of the real fault sequence and the generated sequence, and output a binary classification result through feature comparison. This design realizes the quantitative evaluation of the spatial distribution difference between the two types of sequence codings.

[0046] (2) The generator adopts a two - layer cascade architecture, and its operation mechanism is as follows: First, the encoding layer converts the random sequence into the latent coding ; Secondly, the conditional distribution calculation , and the temporal coherence of the generated sequence is ensured by modeling the Markov dependence of the hidden state. This design enables the latent coding of the generated fault sequence to have probability characteristics similar to those of the real sequence. It represents the temporal dynamic characteristics in the latent coding sequence. The second layer further approximates the real conditional distribution through further calculation, ensuring that the generated latent coding sequence is as close as possible to the distribution of the actual fault sequence. Adopting this design not only ensures that the model can capture the temporal dynamic characteristics of the real fault sequence, but also improves the effect of model training.

[0047] In the model application stage, according to Figure 3For the shown structure, a data generation model is established, and time series sets corresponding to four different fault types are created. These data sets cover various fault conditions from normal samples to small fault samples, providing diverse samples for subsequent model training. After completing the model training, it enters the small sample data generation stage. In this stage, the trained model is used to generate new small sample fault sequences to verify the effectiveness and accuracy of the model. This method not only improves the model's ability to learn and predict under complex conditions but also ensures its adaptability and robustness to actual power grid fault data. Ultimately, this helps to enhance the security and stability of the power grid.

[0048] Figure 3 It is a schematic diagram for fault generation.

[0049] S4 Model training TS-GAN is a model for time series data generation. The following is an overview of the basic training steps of the TS-GAN model and related formulas. Training steps Data preparation Preprocess the time series data, including normalization, detrending, deperiodization, etc., to prepare for model training. Model structure design The Generator is responsible for generating sequences similar to real time series. The main function of the Discriminator is to distinguish real data from the data generated by the Generator, usually using a Convolutional Neural Network (CNN) or other classification networks. To improve the robustness of this model, the following learning methods are introduced in this paper: The principle of contrastive learning is to calculate the similarity of samples to distinguish similar samples from dissimilar samples. This drives the model to obtain more discriminative feature representations. The training process of contrastive learning is as follows: In each iteration, randomly select samples from the training set, and apply random augmentation strategies (such as geometric transformation, color perturbation, etc.) to each sample to generate samples ( original samples + augmented samples); Secondly, for any sample, only its augmented version is regarded as the positive sample, and the remaining 2n - 2 samples in the batch are used as negative samples to construct a contrastive learning task; Then, use the Generalized Linear Feature Mapping (GLFM) model to encode all samples to obtain their low-dimensional feature representations; Finally, use

[74] loss function to calculate the similarity of positive and negative sample pairs through the normalized temperature scaling strategy, and calculate the contrastive loss , which maximizes the similarity of positive sample pairs and minimizes the similarity of negative sample pairs at the same time to achieve discriminative learning in the feature space. The contrastive loss is defined as shown in formula (3-12): (3-12) In the formula: represents the contrast learning coefficient, is the similarity function. Using the cosine similarity z as the similarity function means that the dot product between and after L2 regularization is defined as shown in formula (3-13): (3-13) In the formula: represents norm.

[0050] The loss function is defined using the adversarial loss function. A typical loss function is as follows: The discriminator loss function is as shown in (3-14): (3-14) The generator loss is as shown in (3-15): (3-15) In the formula: is the discriminator's prediction of the real data, is the fake data generated by the generator, is the real data distribution.

[0051] Hyperparameter tuning adjusts hyperparameters such as the learning rate, batch size, and network structure according to the training effect to achieve the best generation effect. After generation and evaluation are completed, the generator is used to generate new time series data, and methods such as visualization and statistical tests (such as the KS test) are used to evaluate the authenticity and usability of the generated data. The training of TS-GAN involves an adversarial process between the generator and the discriminator. By alternately optimizing their loss functions, the generator can generate time series that are increasingly close to the real data. The specific structural parameters are shown in Tables 1 and 2.

[0052] Table 1 Model parameters of the generator part Neural network layer name Input vector dimension Output vector dimension Linear layer (16,100) (16,1500) Position encoding layer (16,1500) (16,150,10) Normalization layer - - Global melting feature layer (Global-Local Fusion Module) (16,150,10) (16,150,10) Random disconnection (Dropout = 0.7) - - Feed forward layer (Feed Forward Block) (16,11,50) (16,11,50) Convolution layer (Conv2d) (16,150,10) (16,3,1,150) Table 2 Model parameters of the discriminator part Neural network layer name Input vector dimension Output vector dimension Position encoding layer (Positional Em bedding) (16,31,150) (16,11,50) Multi-headed self-attention layer (16,150,10) (16,11,50) Normalization layer (LN) Random disconnection (Dropout = 0.7) - - Feed forward layer (Feed Forward Block) (16,11,50) (16,11,50) Classification head (16,11,50) (16,1) The model training adopts the steps of alternately training the generator and the discriminator: First, fix the generator and train the discriminator. Then, calculate the loss using the real data and the generated data and update the weights. Secondly, fix the discriminator and train the generator. Optimize the quality of the generated data through the generator's loss, and the discriminator will be unable to effectively distinguish the authenticity of the samples, thereby achieving the Nash equilibrium.

[0053] This section will analyze and verify the generated results in terms of authenticity and coverage, and compare them with the results generated by other baseline models to evaluate the credibility of the generated small-sample fault data.

[0054] Parameter settings The model parameters are updated by the Adam optimizer for hyperparameter tuning. And the F1 score is used as the evaluation criterion for the model parameters. The model training and comparative experiments in this paper are all carried out on the Windows system, and the programming environment uses the Python language based on the PyTorch framework. The experimental environment configuration is as follows: an RTX 3090 graphics card driven by CUDA 11.3 is used for GPU accelerated computing.

[0055] S5-1 Evaluation metrics for the generation model In the process of generating fault samples, manual discrimination is often restricted by subjective factors. Therefore, in order to more objectively evaluate the performance of the model in generating fault samples, it is necessary to make a quantitative comparison with the fault samples generated by the simulation model. This paper aims to objectively analyze the generated fault samples through three metrics: 1) The L-distance is used to quantify the difference between the generated fault samples and the simulation samples. Through this metric, the accuracy of the generated samples can be effectively evaluated. The smaller the L-distance, the closer the generated samples are to the simulation samples in terms of features; 2) Calculate the structural similarity (SSIM) between the two to measure the similarity degree between the two samples through brightness, contrast and structural information. The higher the SSIM value, the higher the similarity between the generated samples and the simulation samples in terms of structure, thus reflecting the good expression ability of the generated samples for fault features. 3) Calculate the peak signal-to-noise ratio (PSNR) of the generated samples to evaluate the signal quality of the generated samples. The higher the ratio of the signal to the noise, the better the performance of the generated samples in terms of anti-noise ability. A higher PSNR value represents that the generated samples inherit the characteristics of the real signal.

[0056] Let the input data sample , and the generated data sample be , the distance between them can be expressed by formula (3-16), The distance represents the generation effect of the sample: (3-16) In time series data, data at different time points can be regarded as similar to pixels in an image, and the three major elements of SSIM can also correspond to certain characteristics of time series data. Specifically: the mean of the time series data can be used to represent it, that is, the average value over the entire time period; the contrast can be measured by calculating the standard deviation of the data, reflecting the volatility of the data in the time series; the degree of structural similarity is measured by covariance. For two fault samples and , its structural similarity index is shown in Equation (3-17): (3-17) In the formula: - The mean of sample ; - The mean of production sample ; - The variance of sample ; - The variance of generated sample ; - and The covariance of

[0057] and are constants used to maintain stability. To avoid a denominator of 0, and are often taken. is the dynamic range of pixel values, and generally is taken. The closer is to 1, the more similar the two data samples are. When two time series data are exactly the same,

[0058] The peak signal-to-noise ratio is a commonly used objective criterion for evaluating the quality of generated data. The larger the value, the stronger the anti-noise ability of the generated sample number. Its expression is shown in Equation (3-15): (3-18) In the formula: - The maximum value of sequence information, - The mean square error of data points.

[0059] S5-2 Sample authenticity assessment, In the operation of a real power grid, various electrical quantities are collected and summarized in real time by a fault recorder. To ensure that the electrical quantities in the generated small-sample fault sequences still meet the constraints of real fault recording and that the generated fault sequences have a high similarity to the real fault sequences, an average similarity degree test and a frequency-domain analysis test are respectively carried out: (1) Average similarity degree, In this study, the similarity between real electrical quantities and generated electrical quantities was analyzed from two aspects: time domain and frequency domain. In the time domain, the parameters of the generated small-sample fault sequences were used to quantify the similarity by the Hellinger distance (HD) and the mean absolute error (MAE). As a measure of overall similarity, HD ranges from [0,1], and the smaller the value, the higher the similarity between the two probability distributions. MAE reflects the average similarity degree of each corresponding interval in multiple intervals, and its value is also in the range of [0,1], and the smaller the value, the higher the similarity. At the same time, small-sample fault sequences generated based on the Monte Carlo sampling (MCS) algorithm were used for comparison, and the similarity between the generated data and the real data was verified through the calculation of HD and MAE. This analysis method provides strong support for evaluating the performance of the generation model and its application in actual fault diagnosis.

[0060] (3-19) (3-20) In the formula: and are the electrical quantity distributions generated by different generation models, and are the electrical quantity sequences in a certain corresponding interval in the generated sequence.

[0061] The average similarity of the fault sequence samples generated by the two data augmentation methods , the difference degree of the sample points in the corresponding interval in the key time interval and the average difference degree of all key intervals of the fault sequence can be defined as: (3-21) (3-22) (3-23) In the formula: is the number of sequence parameters.

[0062] The analysis of the overall similarity experimental data shows that the electrical quantities generated by the TS-GAN model have a high similarity to the real sequence, and the The value reached 79.7%, exceeding 74.1% of the comparison model. The distribution similarity effects of the 8 parameters of the electrical quantity sequence generated by the TS-GAN model are all better than those of the comparison model, and the most obvious one is the morphological feature parameters of the current sequence volume. Moreover, the difference between the fault sequence generated by the TS-GAN model and the real fault sequence in some interval segments is also small. The minimum difference is 1.23%, which is less than 3.02% of the comparison model. Among them, the minimum difference of the samples generated by the improved TS-GAN model in the generated current quantity reaches 0.58%, showing high authenticity.

[0063] Table 3 Average Degree Similarity Analysis To screen out sequences with good similarity to the real electrical quantity sequence from the generated scenarios, the average Frechet distance between the generated fault sample sequence and its most similar real fault recording sequence is calculated as the maximum sequence difference. Calculate the average Frechet distance between the generated fault sequence and its most similar signal sequence in the fault recording sequence data generated based on the improved TS-GAN model: among them, the value based on the improved TS-GAN model is 0.11, which is 0.01 less than that of the comparison model. Experiments show that among the generated fault samples, more than 83.2% fall within the range of the most similar real fault sequence, which further proves the authenticity of the generated fault sequence.

[0064] (2) Frequency Domain Analysis To verify whether the improved TS-GAN model can effectively capture the dynamic change law of electrical quantities in the real power grid operation, this experiment conducts frequency domain analysis on the generated electrical quantities and compares them with the MCS algorithm. According to formula (3-21), the average spectral error between the generated fault sequence and the real fault sequence is calculated , so as to evaluate the overall similarity of its spectrum. Search for the most similar real fault sequence for each generated fault sequence. To calculate the local spectral difference between the generated data and the real data, formula (3-22) is introduced to calculate the average spectral error , and analyze the similarity between each generated fault sequence and the real fault sequence: (3-24) (3-25) In the formula: represents the frequency domain representation of the generated data, represents the frequency domain representation of the real fault sequence, represents the spectrum of the real fault sequence most similar to the generated fault sequence , and are the serial numbers of the generated fault sequence and the true fault sequence respectively, is the spectral segmentation serial number, represents the number of generated samples, is the number of true samples, is the sum of the number of spectral segments.

[0065] Table 4 Similarity of the generated sequences As shown in Table 4, the experimental data show that the generated electrical quantity sequences are and in the direction of the spectrum The error from the true fault sequence is controlled within 8%. The proposed fault sequence generation model can effectively learn the dynamic characteristics of the true electrical quantity sequence, and the generated simulation data has a high spectral consistency with the true fault data. Horizontal comparison shows that the spectral error of the generated sequence by the improved TS-GAN model is significantly lower than that of the MCS algorithm. Especially when best matching the true data, the error of TS-GAN further drops below 6.5%. This difference highlights the advantage of TS-GAN in the fault data generation task, which more precisely captures the distribution characteristics of the true data through the adversarial training mechanism as follows.

[0066] S5-3 Sample coverage evaluation, To verify the coverage effect of the generated samples and the true samples, this article uses the method of dimensionality reduction visualization to observe different electrical quantity sequences. As an important tool in the field of data visualization, t-SNE (t-distributed Stochastic Neighbor Embedding) realizes the effective conversion from high-dimensional data to low-dimensional representation through a dimensionality reduction process optimized by distribution. The generated two-dimensional or three-dimensional scatter plot can clearly reflect the topological structure of the original data, which greatly facilitates the exploratory analysis of complex data sets.

[0067] The t-SNE technology provides an effective visualization analysis method for the evaluation of the generation model through the joint dimensionality reduction mapping of two samples. The degree of distribution consistency between the generated samples and the true samples in the low-dimensional space objectively reflects the ability of the model to restore the characteristics of the original data, and is an important reference basis for model optimization.

[0068] First, randomly select samples and specify the sample size, set the shape as where , and by calculating the average value of each sample in the time dimension, is one-dimensional; next, use the t-SNE algorithm to process Reduce from 27 dimensions to 2 dimensions. The t-SNE dimensionality reduction technique was used in the research. Under the premise of maintaining the topological relationship of high-dimensional data, 1,000 randomly selected samples were projected onto a two-dimensional plane. As Figure 4 and Figure 5 shown by the experimental results, by analyzing the distribution overlap degree of the samples generated by TS-GAN and the real samples in the two-dimensional space, the data distribution fitting ability of the generation model under the same training parameters was effectively verified.

[0069] Figure 4 is the distribution of the collected real fault sequences, Figure 5 is the distribution map of the generated fault sequences.

[0070] As Figure 4 and Figure 5 shown, the timing relationship of each group of electrical quantity sequences is presented in the form of scatter points. By performing cluster analysis on these timing distributions and drawing probability density contour maps, it can be observed that the generated small-sample data shows a more extensive distribution characteristic in the parameter space than the real fault recording data. In addition, the generated data covers some fault scenarios not recorded in the real dataset, indicating its stronger diversity generation ability.

[0071] As shown in Equation (3-26), Rispley's function, as a spatial point pattern analysis method, its calculation process can be decomposed into the following steps: First, in the given parameter space, a circular neighborhood with a radius of d is constructed centered on each sample point; then, the number of sample points contained in each neighborhood is counted, and the statistical results of all sample points are averaged; finally, this average value is divided by the spatial density of the sample points in the research area to eliminate the influence of the sample size on the analysis results. This normalization process enables the K function to effectively evaluate the aggregation or dispersion characteristics of the point distribution at a specific spatial scale. Divide the average value by the instance point density within the parameter interval to obtain the Rispley's function value at a distance of , and the calculation formula is as follows: (3-26) In the formula: is the area of the parameter interval; is the number of sample points in the parameter area; is the distance between different sample points; characterizes the positional relationship between the sample point and the circle, and: (3-27) Select the value range of the radius to be ​, calculate the Rispley’s K function values corresponding to the sample point distributions generated by the improved TS-GAN model respectively. As Figure 6 shown, the Ripley's K function curves of three different datasets are studied and compared: (1) the sample point distribution of real fault recording data; (2) the scenario sample point distribution generated by the comparison model; (3) the reference sample points uniformly distributed in the parameter space. Through four groups of comparative analysis of the function curves, the differences in the spatial distribution characteristics of different datasets can be quantitatively evaluated.

[0072] Figure 6 The Rispley's K curve of the instance point distribution is shown as follows.

[0073] By Figure 6 it is found that within the same parameter range , the spatial aggregation degree of uniformly distributed instance points is the lowest, which conforms to the theoretical expectation. Further analysis reveals that: when takes a smaller value, the distribution of real fault data is more dispersed than that of the scenario instance points generated by the improved TS-GAN model; when takes a larger value, the distribution of the sample points generated by the improved TS-GAN model is more dispersed than that of the real fault sample points. In addition, the real fault data presents local aggregation characteristics in the parameter space and is only distributed in some areas, while the sample points generated by the improved TS-GAN are more uniformly distributed as a whole, but show higher aggregation in a small range. Compared with the MCS algorithm, the improved TS-GAN has obvious advantages in the uniformity of parameter space coverage.

[0074] Through the above analysis, compared with the real sample data, the samples generated by the TS-GAN model have a larger coverage range, and the samples generated by the improved TS-GAN data enhancement model present better uniform distribution characteristics in the parameter space. Compared with the real fault data, the model can generate more unseen samples, significantly expanding the scenario coverage range. Quantitative analysis shows that the and values of the generated sample data are on average 17.83% and 16.37% lower than those of the real data (average value), which indicates that under the same evaluation criteria, the fault data generated by the improved TS-GAN model better meets the preset rule requirements.

[0075] Table 5 Generation sequence distribution interval Electrical quantity parameter Parameter interval growth rate % Electrical quantity parameter Parameter interval growth rate % <![CDATA[I a > 7.82 <![CDATA[U a > 6.93 <![CDATA[I b > 23.73 <![CDATA[U b > 16.38 <![CDATA[I c > 6.00 <![CDATA[U C > 22.13 <![CDATA[I0]]> 3.57 <![CDATA[U0]]> 12.21 By comparing the distribution intervals of key fault sequence parameters between the generated data and the real data, it is found that the fault parameter intervals in the generated scenarios all show a significant expansion. The statistical results show that the maximum growth rate of the parameter distribution interval reaches 89.79%, and the average growth rate is 19.44%. This phenomenon indicates that the proposed generation method can effectively expand the distribution range of electrical quantity parameters, generate fault sequence variants not recorded in the real data, and thus significantly improve the diversity and scenario coverage ability of the generated data.

[0076] Embodiment 2 This embodiment provides a power grid fault recording data enhancement system based on an improved TS-GAN, including: A data acquisition module, configured to: A computer-readable storage medium, in which multiple instructions are stored, and the instructions are suitable for being loaded and executed by a processor of a terminal device for the described power grid fault recording data enhancement method based on an improved TS-GAN.

[0077] A terminal device, including a processor and a computer-readable storage medium, the processor is used to implement each instruction; the computer-readable storage medium is used to store multiple instructions, and the instructions are suitable for being loaded and executed by the processor for the described power grid fault recording data enhancement method based on an improved TS-GAN.

[0078] The above are all preferred embodiments of the present invention. Without limiting the protection scope of the present invention accordingly, therefore: all equivalent changes made according to the structure, shape, and principle of the present invention shall be covered within the protection scope of the present invention.

Claims

1. An improved TS-GAN-based method for enhancing power grid fault recording data, characterized in that Including: Obtain small-sample fault recording data; Preprocess the obtained small-sample fault recording data; Extract features from the small-sample fault recording data using a fault generation model based on improved TS-GAN; Generate and discriminate faults for the extracted features using a fault generation model based on improved TS-GAN; Use the small-sample fault recording data as samples for model training; Evaluate and optimize the trained model; Use the optimized model to perform data augmentation on power grid fault recording data.

2. The method for enhancing power grid fault recording data based on improved TS-GAN according to claim 1, wherein The preprocessing of the obtained small-sample fault recording data includes, for a real recording sequence with 2400 sampling points, gradually sampling each sampling point starting from the recording start point, so as to expand a single real fault data sequence into a training data set containing multiple sampling points. Each segment corresponds to consecutive time points in the original sequence, which is used to increase the quantity and diversity of training data. Among the fault recording data, the electrical quantity indicators form a time series process , where represents discrete time points, represents fault feature parameters, is the fault parameter space, and the trajectory sequence is a discrete set. The electrical quantity state can be expressed as . If the electrical quantity of the fault recording data generated by the model is , and it follows the probability distribution , then the probability distribution generated by the model needs to be close to the sequence distribution in the real scenario ; In the data generation model , it is represented by the parameter , represents the model parameter space, satisfying: ， Wherein: represents a measure of the distance between different distributions.

3. An improved TS-GAN-based power grid fault recording data enhancement method according to claim 2, characterized in that, The extracting features from the small-sample fault recording data using a fault generation model based on improved TS-GAN includes first integrating the global attention mechanism GLFM. The features of the sequence extracted by the multi-head attention layer are further learned through the feed-forward layer with a multi-layer perceptron to obtain more complex non-linear mappings, enabling the model to learn richer representations. Finally, the local correlation features of the sequence are extracted through the attention layer, and the information captured by each layer is fused through residual connections between the multi-head self-attention layer, the feed-forward layer, and the convolutional attention layer. Among them, in the first layer, the standard WGAN uses a random vector to generate a synthetic spectrogram, and in the second layer, by introducing conditional WGAN, the synthetic image generated in the first layer is transformed into an accurate synthetic time series.

4. An enhanced method for power grid fault recording data based on improved TS-GAN according to claim 3, characterized in that, The improved TS-GAN fault generation model is used to extract features from small sample fault recording data. The method also includes embedding a lightweight architecture attention mechanism to enable the model to have dynamic feature selection capabilities and autonomously focus on key signal changes during the fault occurrence period. First, the query matrix is calculated. With the key matrix of To obtain the similarity, we then divide by and through The function normalizes it to obtain the score matrix ; Then the score matrix and value matrix Multiply and calculate the attention matrix : ， Input of the attention matrix After linear transformation, the output formula of this layer is expressed as: ， Among them, are learnable network weight parameters.

5. A method for enhancing power grid fault recording data based on an improved TS-GAN according to claim 4, characterized in that, The discriminator in the improved TS-GAN-based fault generation model is utilized to evaluate the input latent code sequence and determine whether it comes from the true data distribution. Specifically, the discriminator first scores the latent code sequence through the function and outputs , which represents the probability that the sequence belongs to the true fault sequence. A binary classification result is obtained through feature comparison. Then, the random sequence is converted into a latent code by the encoding layer. Secondly, the conditional distribution is calculated. By modeling the Markov dependence of the hidden states, the temporal coherence of the generated sequence is ensured. represents the temporal dynamic characteristics in the latent code sequence. The second layer further approximates the true conditional distribution through re-computation to ensure that the generated latent code sequence is close to the distribution of the actual fault sequence.

6. A method for enhancing power grid fault recording data based on an improved TS-GAN according to claim 5, characterized in that, The model training using small - sample fault recording data as samples includes contrastive learning by calculating the similarity of samples to distinguish similar samples from dissimilar samples, promoting the model to obtain more discriminative feature representations. The training process of contrastive learning includes, in each iteration, randomly selecting samples from the training set, and applying a random augmentation strategy to each sample to generate samples. For any sample, only its augmented version is regarded as a positive sample, while the remaining 2n - 2 samples in the batch are used as negative samples, thus constructing a contrastive learning task; then encoding all samples using the Generalized Linear Feature Mapping (GLFM) to obtain low - dimensional feature representations; finally, using the loss function to calculate the similarity of positive and negative sample pairs through the normalized temperature scaling strategy, calculating the contrastive loss , and realizing discriminative learning in the feature space. The contrastive loss is defined as: ， Wherein: represents the contrastive learning coefficient, is the similarity function, Adopting the cosine similarity z as the similarity function means the dot product between and after L2 regularization.

7. A method for enhancing power grid fault recording data based on an improved TS-GAN according to claim 6, characterized in that The evaluation and optimization of the trained model includes objectively analyzing the generated fault samples through three metrics, specifically including: quantifying the difference between the generated fault samples and the simulation samples using the L distance; measuring the similarity between the two samples through brightness, contrast, and structural information; and evaluating the signal quality of the generated samples using the peak signal-to-noise ratio of the generated samples. Among them, for two fault samples and , its structural similarity index is expressed as: ， Wherein: is the mean of the sample ; is the mean of the production sample ; is the variance of the sample ; is the variance of the generated sample ; is the covariance of and .

8. A method for enhancing power grid fault recording data based on an improved TS-GAN according to claim 7, characterized in that, The evaluation and optimization of the trained model also includes analyzing the similarity between the real electrical quantities and the generated electrical quantities from two aspects: time domain and frequency domain. In the time domain, the parameters of the generated small-sample fault sequences are used with the Hellinger distance HD and the mean absolute error MAE to quantify the similarity. In the frequency domain, the local spectral differences between the generated data and the real data are calculated, and the average spectral error is calculated Analyze the similarity between each generated fault sequence and the real fault sequence, expressed as: ， ， Wherein: represents the frequency-domain representation of the generated data, represents the frequency-domain representation of the true fault sequence, represents the spectrum of the true fault sequence that is most similar to the generated fault sequence , and are the sequence numbers of the generated fault sequence and the true fault sequence respectively, is the spectrum segmentation sequence number, represents the number of generated samples, is the number of true samples, is the total number of spectrum segmentation numbers.

9. A method for enhancing power grid fault recording data based on an improved TS-GAN according to claim 8, characterized in that, The evaluation and optimization of the trained model also includes using the t-SNE dimensionality reduction method to project 1,000 randomly selected samples onto a two-dimensional plane while maintaining the topological relationship of the high-dimensional data. By analyzing the distribution overlap degree of the samples generated by the TS-GAN and the real samples in the two-dimensional space, the data distribution fitting ability of the generative model under the same training parameters is verified. Among them, within the given parameter space, a circular neighborhood with a radius of d is constructed centered on each sample point; then, the number of sample points contained in each neighborhood is counted, and the statistical results of all sample points are averaged; the average value is divided by the instance point density within the parameter interval to obtain the Rispley's function value at , expressed as: ， In the formula: is the area of the parameter interval; is the number of sample points in the parameter area; is the distance between different sample points; characterizes the positional relationship between the sample point and the circle.

10. A power grid fault recording data enhancement system based on an improved TS-GAN, characterized in that, Including: A data acquisition module configured to obtain small-sample fault recording data; A preprocessing module configured to preprocess the obtained small-sample fault recording data; A feature extraction module configured to extract features from the small-sample fault recording data using a fault generation model based on improved TS-GAN; A fault generation module configured to generate and discriminate faults for the extracted features using a fault generation model based on improved TS-GAN; A training module configured to use the small-sample fault recording data as samples for model training; An optimization module configured to evaluate and optimize the trained model; An augmentation module configured to use the optimized model to perform data augmentation on power grid fault recording data.

Images