A voltage sag data augmentation method based on diffusion probability model

Abstract

This application relates to a method and system for voltage sag data augmentation based on a diffusion probability model, comprising: acquiring voltage sag data with sag source labels, and normalizing the voltage sag data to obtain one-dimensional waveform data; converting the one-dimensional waveform data into two-dimensional image data; using the two-dimensional image data as training sample data to train a conditional diffusion probability model to obtain a target model; and inputting the sag source labels into the target model to obtain augmented sample data. Through this invention, voltage sag dataset augmentation can be achieved, thus realizing voltage sag data enhancement.

Description

Technical Field

[0001] This application relates to the field of transformer technology, specifically to a voltage sag data enhancement method based on a diffusion probability model. Background Technology

[0002] Voltage sags are a common power quality problem in power systems, attracting significant attention due to their highly destructive nature. In voltage sag analysis, machine learning algorithms are used for data classification and feature identification, enabling automated detection, diagnosis, tracing, and prediction of voltage sag events. The collection of large-scale power data is fundamental for utilizing artificial intelligence to analyze voltage sags; however, the lack of high-quality, field-sampled data limits the generalization ability and robustness of models. Simulated data often fails to simulate real-world harmonics, noise, oscillations, and other interference factors. Therefore, research is needed on voltage sag data generation methods to obtain highly realistic and diverse data, thereby enhancing voltage sag data. Summary of the Invention

[0003] The purpose of this application is to propose a voltage sag data augmentation method based on a diffusion probability model to expand the voltage sag dataset.

[0004] To achieve the above objectives, embodiments of this application propose a voltage sag data augmentation method based on a diffusion probability model, comprising:

[0005] Obtain voltage sag data with sag source tags, and normalize the voltage sag data to obtain one-dimensional waveform data;

[0006] The one-dimensional waveform data is converted into two-dimensional image data;

[0007] The two-dimensional image data is used as training sample data to train the conditional diffusion probability model to obtain the target model;

[0008] Input the temporary source label into the target model to obtain expanded sample data.

[0009] Optionally, the sag source label is the coded value corresponding to the sag source. The sag source includes a single sag source and a composite sag source. The single sag source includes induction motor starting, transformer excitation, and short circuit fault. The composite sag source includes multi-stage voltage sag caused by short circuit fault, short circuit fault and induction motor starting, short circuit fault and transformer switching, and induction motor starting and transformer switching.

[0010] Optionally, the voltage sag data is a sequence of effective voltage values, and the time and effective voltage values ​​of the one-dimensional waveform data correspond to the two-dimensional pixel coordinates of the image, respectively.

[0011] Optionally, the conditional diffusion probability model performs a convolution operation on the encoded value as a conditional input; the conditional diffusion probability model uses a fully convolutional model U-Net as the neural network structure, including a downsampling path and an upsampling path; the downsampling path includes features for extracting features from the two-dimensional image data to obtain a feature map and reducing the spatial dimension of the feature map; the upsampling path is used to restore the spatial dimension of the feature map.

[0012] Embodiments of this application also propose a voltage sag data enhancement system based on a diffusion probability model, comprising:

[0013] The data acquisition unit is used to acquire voltage sag data with sag source tags, and to normalize the voltage sag data to obtain one-dimensional waveform data.

[0014] A data conversion unit is used to convert the one-dimensional waveform data into two-dimensional image data;

[0015] The model training unit is used to train the conditional diffusion probability model to obtain the target model by using the two-dimensional image data as training sample data.

[0016] An expansion unit is used to input temporary source labels into the target model to obtain expanded sample data.

[0017] Optionally, the sag source label is the coded value corresponding to the sag source. The sag source includes a single sag source and a composite sag source. The single sag source includes induction motor starting, transformer excitation, and short circuit fault. The composite sag source includes multi-stage voltage sag caused by short circuit fault, short circuit fault and induction motor starting, short circuit fault and transformer switching, and induction motor starting and transformer switching.

[0018] Optionally, the voltage sag data is a sequence of effective voltage values, and the time and effective voltage values ​​of the one-dimensional waveform data correspond to the two-dimensional pixel coordinates of the image, respectively.

[0019] Optionally, the conditional diffusion probability model performs a convolution operation on the encoded value as a conditional input; the conditional diffusion probability model uses a fully convolutional model U-Net as the neural network structure, including a downsampling path and an upsampling path; the downsampling path includes features for extracting features from the two-dimensional image data to obtain a feature map and reducing the spatial dimension of the feature map; the upsampling path is used to restore the spatial dimension of the feature map.

[0020] The embodiments of this application have at least the following beneficial effects:

[0021] The embodiments of this application utilize a conditional diffusion probability model to generate voltage sag waveforms and expand voltage sag sample data, thereby enhancing voltage sag data. The expanded sample data generated by the conditional diffusion probability model provides more comprehensive feature information, which helps to achieve tasks such as sag source identification and voltage sag sensitive device evaluation.

[0022] Other features and advantages of embodiments of this application will be set forth in the following description. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart of a voltage sag data enhancement method based on a diffusion probability model, as described in an embodiment of this application.

[0025] Figure 2 This is a flowchart illustrating the training process of the conditional diffusion probability model in the embodiments of this application.

[0026] Figure 3 This is a sampling flowchart of the conditional diffusion probability model in the embodiments of this application. Detailed Implementation

[0027] The various exemplary embodiments, features, and aspects of this application will be described in detail below with reference to the accompanying drawings. Furthermore, numerous specific details are set forth in the following detailed embodiments to better illustrate this application. Those skilled in the art will understand that this application can be practiced without certain specific details. In some instances, means well-known to those skilled in the art have not been described in detail in order to highlight the main points of this application.

[0028] One embodiment of this application proposes a voltage sag data augmentation method based on a diffusion probability model, comprising the following steps:

[0029] Step S1: Obtain voltage sag data with sag source tags, and normalize the voltage sag data to obtain one-dimensional waveform data;

[0030] Specifically, the normalization process uses the Z-score normalization method:

[0031]

[0032] Where, x *The data is normalized, x is the data before normalization, μ is the sample mean, and σ is the sample variance. The processed data conforms to the standard normal distribution, that is, the mean is 0 and the standard deviation is 1. This process can improve the model accuracy and training convergence speed.

[0033] Step S2: Convert the one-dimensional waveform data into two-dimensional image data;

[0034] Step S3: Use the two-dimensional image data as training sample data to train the conditional diffusion probability model to obtain the target model;

[0035] Step S4: Input the temporary source label into the target model to obtain expanded sample data.

[0036] The method in this embodiment uses a conditional diffusion probability model to generate voltage sag waveforms and expand voltage sag sample data, thereby enhancing voltage sag data. The expanded sample data generated by the conditional diffusion probability model provides more comprehensive feature information, which helps to achieve tasks such as sag source identification and voltage sag sensitive device evaluation.

[0037] Furthermore, the sag source label is the coded value corresponding to the sag source. The sag source includes a single sag source and a composite sag source. The single sag source includes induction motor starting, transformer excitation, and short circuit fault. The composite sag source includes multi-stage voltage sag caused by short circuit fault, short circuit fault and induction motor starting, short circuit fault and transformer switching, and induction motor starting and transformer switching.

[0038] Furthermore, the voltage sag data is a sequence of effective voltage values, that is, multiple effective voltage values ​​arranged sequentially according to the sampling time, and the multiple effective voltage values ​​form one-dimensional waveform data; when the one-dimensional waveform data is converted into two-dimensional image data, the time and effective voltage values ​​of the one-dimensional waveform data correspond to the two-dimensional pixel coordinates of the image, respectively.

[0039] Furthermore, the conditional diffusion probability model performs a convolution operation on the encoded value as a conditional input; the conditional diffusion probability model adopts a fully convolutional model U-Net as the neural network structure, including a downsampling path and an upsampling path; the downsampling path includes features for extracting features from the two-dimensional image data to obtain a feature map and reducing the spatial dimension of the feature map; the upsampling path is used to restore the spatial dimension of the feature map.

[0040] Specifically, the downsampling path consists of several repeating modules, each consisting of two convolutional layers, one ReLU activation unit, and one pooling layer; the upsampling path uses similar repeating modules (also containing two convolutional layers, one ReLU activation unit, and one pooling layer) to restore the spatial dimension of the feature map, but each module of the upsampling path also contains a connection layer to connect the feature maps of the corresponding layers in the downsampling path.

[0041] Specifically, the training process of the conditional diffusion probability model is as follows: Figure 2 As shown, it includes:

[0042] Define noise control parameter β 1:T =[β1,β2,...,β T ];

[0043] Samples x0 to q(x) are selected from the training set;

[0044] Randomly select time step t from {1,2,…,T};

[0045] Take random noise

[0046] The neural network model is trained to minimize the distance between the actual noise and the model's predicted noise; furthermore, the distance can be FID distance (Fréchet Inception Distance) or Wasserstein distance.

[0047] Repeat the above steps until the model converges;

[0048] The diffusion probability model is trained using Lognorm as the loss function:

[0049]

[0050] Furthermore, the initial learning rate α of the diffusion probability model is 0.0005, the number of training rounds is 500, the size of each batch of training samples is 128, and the Adam stochastic gradient descent method is used to accelerate the update of model parameters.

[0051] Further, refer to Figure 3 The diffusion probability model sampling algorithm includes:

[0052] From random noise Set off;

[0053] Let t = T;

[0054] Take random noise

[0055] Neural network predicts noise ε θ ;

[0056] Calculate the variance of the inverse transition distribution

[0057] Calculate the first term

[0058] t = t-1, and repeat the above steps until t = 1.

[0059] Another embodiment of this application proposes a voltage sag data enhancement system based on a diffusion probability model, comprising:

[0060] The data acquisition unit is used to acquire voltage sag data with sag source tags, and to normalize the voltage sag data to obtain one-dimensional waveform data.

[0061] A data conversion unit is used to convert the one-dimensional waveform data into two-dimensional image data;

[0062] The model training unit is used to train the conditional diffusion probability model to obtain the target model by using the two-dimensional image data as training sample data.

[0063] An expansion unit is used to input temporary source labels into the target model to obtain expanded sample data.

[0064] Furthermore, the sag source label is the coded value corresponding to the sag source. The sag source includes a single sag source and a composite sag source. The single sag source includes induction motor starting, transformer excitation, and short circuit fault. The composite sag source includes multi-stage voltage sag caused by short circuit fault, short circuit fault and induction motor starting, short circuit fault and transformer switching, and induction motor starting and transformer switching.

[0065] Furthermore, the voltage sag data is a sequence of effective voltage values, and the time and effective voltage values ​​of the one-dimensional waveform data correspond to the two-dimensional pixel coordinates of the image, respectively.

[0066] Furthermore, the conditional diffusion probability model performs a convolution operation on the encoded value as a conditional input; the conditional diffusion probability model adopts a fully convolutional model U-Net as the neural network structure, including a downsampling path and an upsampling path; the downsampling path includes features for extracting features from the two-dimensional image data to obtain a feature map and reducing the spatial dimension of the feature map; the upsampling path is used to restore the spatial dimension of the feature map.

[0067] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and substitutions will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical applications, or technological improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A voltage sag data augmentation method based on diffusion probability model, characterized in that, The method comprises the following steps: obtain voltage sag data with a voltage sag source label, and normalize the voltage sag data to obtain one-dimensional waveform data; the voltage sag source label is an encoding value corresponding to a voltage sag source; the voltage sag source includes a single voltage sag source and a composite voltage sag source; the single voltage sag source includes induction motor starting, transformer excitation, and short circuit fault; the composite voltage sag source includes multi-stage voltage sag caused by short circuit fault, short circuit fault and induction motor starting, short circuit fault and transformer switching, and induction motor starting and transformer switching; the voltage sag data is a voltage effective value sequence; the time and voltage effective value of the one-dimensional waveform data correspond to the two-dimensional pixel coordinates of an image, respectively; convert the one-dimensional waveform data into two-dimensional image data; use the two-dimensional image data as training sample data to train a conditional diffusion probability model to obtain a target model; the conditional diffusion probability model performs convolution operation on the encoding value as a conditional input; the conditional diffusion probability model uses a full convolution model U-Net as a neural network structure, which includes a down-sampling path and an up-sampling path; the down-sampling path includes feature extraction of the two-dimensional image data to obtain a feature map and reduce the spatial dimension of the feature map; the up-sampling path is used to restore the spatial dimension of the feature map; input the voltage sag source label into the target model to obtain augmented sample data.

2. A voltage sag data augmentation system based on diffusion probability models, characterized in that, The method comprises the following steps: a data acquisition unit is configured to obtain voltage sag data with a voltage sag source label, and normalize the voltage sag data to obtain one-dimensional waveform data; the voltage sag source label is an encoding value corresponding to a voltage sag source; the voltage sag source includes a single voltage sag source and a composite voltage sag source; the single voltage sag source includes induction motor starting, transformer excitation, and short circuit fault; the composite voltage sag source includes multi-stage voltage sag caused by short circuit fault, short circuit fault and induction motor starting, short circuit fault and transformer switching, and induction motor starting and transformer switching; the voltage sag data is a voltage effective value sequence; the time and voltage effective value of the one-dimensional waveform data correspond to the two-dimensional pixel coordinates of an image, respectively; a data conversion unit is configured to convert the one-dimensional waveform data into two-dimensional image data; a model training unit is configured to use the two-dimensional image data as training sample data to train a conditional diffusion probability model to obtain a target model; the conditional diffusion probability model performs convolution operation on the encoding value as a conditional input; the conditional diffusion probability model uses a full convolution model U-Net as a neural network structure, which includes a down-sampling path and an up-sampling path; the down-sampling path includes feature extraction of the two-dimensional image data to obtain a feature map and reduce the spatial dimension of the feature map; the up-sampling path is used to restore the spatial dimension of the feature map; an augmentation unit is configured to input the voltage sag source label into the target model to obtain augmented sample data.

Images