A disturbance identification method based on gram angle difference field and transfer learning

Abstract

This invention relates to the field of power system disturbance identification technology, and particularly to a disturbance identification method based on Gram angle difference field and transfer learning. The method includes: acquiring time-series data of frequency and voltage amplitude corresponding to disturbances occurring at the IEEE 39 nodes of the power system, obtaining one-dimensional time-series frequency and voltage amplitude data, performing feature extraction on each and converting them into two-dimensional Gram angle difference field images, fusing the feature extraction results based on RGB image channels, inputting the feature two-dimensional image dataset into a preset transfer learning-AlexNet network model for training, and constructing a disturbance identification model based on the SE attention mechanism; inputting the acquired current PMU disturbance data of the power system into the disturbance identification model, and obtaining the disturbance identification result output by the disturbance identification model. This invention can improve the model's identification ability and noise resistance, further improving the identification accuracy for different types of disturbances in the power system.

Description

Technical Field

[0001] This invention relates to the field of power system disturbance identification technology, and in particular to a disturbance identification method based on Gram angle difference field and transfer learning. Background Technology

[0002] With the increasing penetration of renewable energy and the integration of distributed generation and microgrids, modern power systems exhibit high reliability and security, scalability and flexibility. However, due to their large scale, complex structure and diverse equipment, they also exhibit high complexity and nonlinearity, resulting in a wide variety of power system disturbances and uncertainties in their occurrence. Accurately identifying the types of power system disturbances helps to improve system protection and control strategies and enhance system robustness.

[0003] In existing technologies, distribution network synchronous phasor measurement devices are used for disturbance identification. This method is usually based on frequency domain analysis and time domain feature extraction. Such methods can directly obtain feature quantities with clear physical meaning. However, in practical applications, the presence of noise may cause the feature extraction results to be distorted. At the same time, the existing technology has insufficient feature representation for the nonlinear and non-stationary characteristics in disturbance data.

[0004] With the development of artificial intelligence technology, deep learning has been widely used in classification tasks. However, in the conventional deep neural network classification training process, a large amount of labeled sample data is usually required, and training the model from scratch requires a lot of resources. In addition, existing technologies lack efficient means to obtain sample resources. Furthermore, existing technologies have poor noise resistance, resulting in low accuracy of perturbation recognition. Summary of the Invention

[0005] The purpose of this invention is to provide a perturbation recognition method based on Gram angle difference field and transfer learning, which solves the problems of low efficiency in acquiring sample resources, poor noise resistance, and low accuracy in perturbation recognition in existing technologies.

[0006] To achieve the above objectives, this invention provides a perturbation identification method based on Gram angle difference field and transfer learning, comprising the following steps:

[0007] S1. Obtain the time series data of frequency and voltage amplitude corresponding to the disturbance at IEEE 39 node of the power system, and perform normalization preprocessing on the time series data of frequency and voltage amplitude to obtain one-dimensional time series frequency and voltage amplitude data.

[0008] S2. Extract features from the frequency and voltage amplitude data of the one-dimensional time series and convert them into two-dimensional Gram angle field images. Perform feature fusion on the feature extraction results based on the RGB image channels to obtain feature two-dimensional image data.

[0009] S3. Input the feature two-dimensional image dataset into the preset transfer learning-AlexNet network model for training, and build a perturbation recognition model based on the SE attention mechanism;

[0010] S4. Input the acquired current PMU disturbance data of the power system into the disturbance identification model to obtain the disturbance identification result output by the disturbance identification model.

[0011] In some embodiments of this application, S1, the normalization preprocessing of the frequency and voltage amplitude time series data includes:

[0012] S11, obtained from n Time series data of frequency and voltage amplitude values ​​at each moment. Normalization preprocessing is performed on it, and the expression is:

[0013] ;

[0014] Among them, the frequency and voltage amplitude data of the one-dimensional time series These are the data points in the original time series. For the data points in the new time series obtained after normalization, X for The frequency and voltage amplitude time series data are composed of values ​​at each moment.

[0015] In some embodiments of this application, S2, converting the frequency and voltage amplitude data of a one-dimensional time series into a two-dimensional Gram angular field image includes:

[0016] S21. Perform angle transformation on the data points in the new time series and encode the timestamps by radius. The expression is:

[0017] ;

[0018] in, The normalized time series, for Polar angle after inverse cosine function transformation for The polar radius of the point for The timestamp of the point This is a constant factor for regularizing the polar coordinate system;

[0019] Obtain the transformed time series ,in, These are the data points in the transformed time series;

[0020] S22. The Gram matrix is ​​calculated based on the phase difference of the transformed time series.G GADF The calculation formula is:

[0021] ;

[0022] For Gram matrix, For time series Point and j point angle difference;

[0023] Based on Gram matrix G GADF A two-dimensional gram angular field image is obtained.

[0024] In some embodiments of this application, in S2, feature fusion based on the feature extraction results using RGB image channels includes:

[0025] S23. Input the feature extraction results into the R and B channels of the RGB image for feature fusion. The feature extraction results include voltage and frequency data after the Gram angle difference field changes.

[0026] In some embodiments of this application, in step S3, constructing a perturbation recognition model based on the training results and the SE attention mechanism includes:

[0027] S31. Based on the SE attention mechanism, feature weights are assigned to the feature channels in the training results to obtain global information, expressed as:

[0028] ;

[0029] in, z c For the first c Global information for each channel H The height of the feature map, W The width of the feature map. For feature maps, C Number of channels;

[0030] S32. Based on the SE attention mechanism, the channel feature values ​​are calculated using the first fully connected layer and the ReLU activation function. The expression is:

[0031] ;

[0032] in, W 1 The weights of the first fully connected layer. b 1 The bias is for the first fully connected layer. z This is the global description vector for all channels in the first layer. s These are channel characteristic values;

[0033] S33. The channel weights are calculated based on the second fully connected layer using the SE attention mechanism and the sigmoid activation function, and the expression is:

[0034] ;

[0035] in, W 2 The weights of the second fully connected layer. b 2 For the bias of the second fully connected layer, σ It is the Sigmoid activation function. Channel weights;

[0036] S34. Optimize the feature channels of the training results based on channel weights, and construct a perturbation recognition model based on the optimization results.

[0037] In some embodiments of this application, S34 further includes:

[0038] S341. The improved cross-entropy loss function FocalLoss is used to optimize the sample weights. The expression is:

[0039] ;

[0040] in, p t The probability of the predicted target category; α t As a balance factor; γ As a regulating factor, An improved cross-entropy loss function;

[0041] S342. The CenterLoss algorithm is used to optimize the distance of intra-class features in the model. The expression is:

[0042] ;

[0043] in, For the first Features of each sample For the first The center of each sample category, N For batch values, The loss function is centered.

[0044] In some embodiments of this application, in step S4, the acquired current PMU disturbance data of the power system is input into the disturbance identification model, and the disturbance identification result output by the disturbance identification model is obtained, including:

[0045] S41. Obtain the voltage amplitude and frequency data of the current PMU disturbance in the power system;

[0046] S42. Normalize and preprocess the voltage amplitude data and frequency data of the current PMU disturbance, and then perform Gram angle difference field image conversion and feature fusion in sequence to obtain the image to be identified.

[0047] S43. Input the image to be identified into the perturbation recognition model, obtain the confidence score output by the perturbation recognition model, and obtain the perturbation recognition result based on the confidence score.

[0048] The advantages and beneficial effects of this invention compared to the prior art are:

[0049] 1. This invention converts one-dimensional time series data into two-dimensional Gram angular field images, which can more intuitively reflect the dynamic relationship and pattern between frequency and voltage amplitude data, better capture the features in the data, and thus provide more valuable information for subsequent feature extraction and model training, thereby improving the accuracy of disturbance identification. By combining the SE attention mechanism with the transfer learning-AlexNet network model to jointly construct a disturbance identification model and applying it to the disturbance identification task, it can make full use of the feature extraction advantages of the pre-trained model, pay more attention to the feature regions that are important for disturbance identification, thereby improving the model's identification ability and noise resistance, and further improving the identification accuracy of different disturbance types in the power system.

[0050] 2. This invention acquires different types of power system simulation data, such as simulation data of power grid node systems of different scales and simulation data under different operating conditions. It integrates and processes multiple types of data to construct a more comprehensive sample dataset, which can make full use of various data resources and improve the efficiency of sample resource acquisition.

[0051] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0052] Figure 1 This is a flowchart of a perturbation identification method based on Gram angle difference field and transfer learning according to the present invention;

[0053] Figure 2 This is a schematic diagram comparing the accuracy of disturbance identification with two other methods in this embodiment of the invention.

[0054] Figure 3 This is a schematic diagram comparing the loss rate of disturbance identification with two other methods in this embodiment of the invention;

[0055] Figure 4 This is a schematic diagram of the frequency and voltage amplitude of the PMU in an actual power system according to an embodiment of the present invention;

[0056] Figure 5This is a schematic diagram of a two-dimensional Gram angle field image in an embodiment of the present invention;

[0057] Figure 6 This is a schematic diagram of the RGB image after feature fusion in an embodiment of the present invention. Detailed Implementation

[0058] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product is in use. They are used only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," and "connect" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0059] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0060] like Figure 1 As shown, this invention provides a perturbation identification method based on Gram angle difference field and transfer learning, comprising the following steps:

[0061] S1. Obtain the time series data of frequency and voltage amplitude corresponding to the disturbance at IEEE 39 nodes of the power system, and perform normalization preprocessing on the time series data of frequency and voltage amplitude to obtain one-dimensional time series frequency and voltage amplitude data.

[0062] S2. Extract features from the frequency and voltage amplitude data of the one-dimensional time series and convert them into two-dimensional Gram angle field images. Perform feature fusion on the feature extraction results based on the RGB image channels to obtain feature two-dimensional image data.

[0063] S3. Input the feature two-dimensional image dataset into the preset transfer learning-AlexNet network model for training, and build a perturbation recognition model based on the SE attention mechanism.

[0064] Optionally, the basic structure of the AlexNet network is constructed, which includes 5 convolutional layers, 3 max pooling layers, 3 fully connected layers and 1 softmax output layer. On the basic structure of the AlexNet network, the SE attention mechanism is incorporated to adaptively assign higher weights to important feature channels.

[0065] S4. Input the acquired current PMU disturbance data of the power system into the disturbance identification model to obtain the disturbance identification result output by the disturbance identification model.

[0066] The advantages and beneficial effects of this invention compared to the prior art are:

[0067] 1. This invention converts one-dimensional time series data into two-dimensional Gram angular field images, which can more intuitively reflect the dynamic relationship and pattern between frequency and voltage amplitude data, better capture the features in the data, and thus provide more valuable information for subsequent feature extraction and model training, thereby improving the accuracy of disturbance identification. By combining the SE attention mechanism with the transfer learning-AlexNet network model to jointly construct a disturbance identification model and applying it to the disturbance identification task, it can make full use of the feature extraction advantages of the pre-trained model, pay more attention to the feature regions that are important for disturbance identification, thereby improving the model's identification ability and noise resistance, and further improving the identification accuracy of different disturbance types in the power system.

[0068] 2. This invention acquires different types of power system simulation data, such as simulation data of power grid node systems of different scales and simulation data under different operating conditions. It integrates and processes multiple types of data to construct a more comprehensive sample dataset, which can make full use of various data resources and improve the efficiency of sample resource acquisition.

[0069] In some embodiments of this application, S1, the normalization preprocessing of the frequency and voltage amplitude time series data includes:

[0070] S11, obtained from Time series data of frequency and voltage amplitude values ​​at each moment. Normalization preprocessing is performed on it, and the expression is:

[0071] ;

[0072] Among them, the frequency and voltage amplitude data of the one-dimensional time series These are the data points in the original time series. For the data points in the new time series obtained after normalization, X for The frequency and voltage amplitude time series data are composed of values ​​at each moment.

[0073] In some embodiments of this application, S2, converting the frequency and voltage amplitude data of a one-dimensional time series into a two-dimensional Gram angular field image includes:

[0074] S21. Perform angle transformation on the data points in the new time series and encode the timestamps by radius. The expression is:

[0075] ;

[0076] in, The normalized time series, for Angle after inverse cosine transformation for The polar radius of the point for The timestamp of the point This is a constant factor for regularizing the polar coordinate system;

[0077] Obtain the transformed time series ,in, These are the data points in the transformed time series;

[0078] S22. The Gram matrix is ​​calculated based on the phase difference of the transformed time series. G GADF The temporal correlation within different time intervals is represented by the trigonometric differences between each point, and the calculation formula is:

[0079] ;

[0080] For Gram matrix; angle difference This describes the relative positions of two points on the unit circle in a time series; while This reflects the influence of angle difference on the relationship between time series sample points, and its value fluctuates within [-1, 1]. Because... This matrix is ​​antisymmetric with zeros on its diagonal. During the GADF transformation, the data mapping is one-to-one at each stage, and the matrix construction depends on the angle of the sequence sample points, which is uniquely determined by the values ​​of the time series sample points. Therefore, the temporal information and nonlinear relationships of the perturbed time series are preserved.

[0081] Based on Gram matrix G GADF A two-dimensional gram angular field image is obtained.

[0082] In some embodiments of this application, in S2, feature fusion based on the feature extraction results using RGB image channels includes:

[0083] S23. Input the feature extraction results into the R and B channels of the RGB image for feature fusion. The feature extraction results include voltage and frequency data after the Gram angle difference field changes.

[0084] In some embodiments of this application, in step S3, constructing a perturbation recognition model based on the training results and the SE attention mechanism includes:

[0085] S31. Construct the basic structure of the AlexNet network. The basic structure of the AlexNet network includes 5 convolutional layers, 3 max pooling layers, 3 fully connected layers, and 1 softmax output layer.

[0086] Based on the basic structure of the AlexNet network, feature weights are assigned to the feature channels in the training results using the SE attention mechanism. Higher weights are adaptively assigned to important feature channels to obtain global information, expressed as:

[0087] ;

[0088] in, z c For the first c Global information for each channel H The height of the feature map, W The width of the feature map. For feature maps, C Number of channels;

[0089] S32. Based on the SE attention mechanism, the channel feature values ​​are calculated using the first fully connected layer and the ReLU activation function. The expression is:

[0090] ;

[0091] in, W 1 The weights of the first fully connected layer. b 1 The bias is for the first fully connected layer. z This is the global description vector for all channels in the first layer. s These are channel characteristic values;

[0092] S33. The channel weights are calculated based on the second fully connected layer using the SE attention mechanism and the sigmoid activation function, and the expression is:

[0093] ;

[0094] in, W 2 The weights of the second fully connected layer. b 2 This is the bias for the second fully connected layer; σ The Sigmoid activation function compresses the output to the range [0,1]. Channel weights;

[0095] S34. Optimize the feature channels of the training results based on channel weights, and construct a perturbation recognition model based on the optimization results.

[0096] In some embodiments of this application, S34 further includes:

[0097] Specifically, before performing step S341, the channel weights obtained above are... w The importance of each channel is "adjusted" by multiplying the output of each channel by its corresponding weight coefficient. The specific process is as follows:

[0098] ;

[0099] in, It is the attention coefficient, i.e., the channel. c Channel weights, X c These are channel characteristic values;

[0100] S341. Considering the problem of imbalanced samples, to increase the weight of hard-to-classify samples and decrease the weight of easy-to-classify samples, an improved cross-entropy loss function, FocalLoss, is adopted:

[0101] ;

[0102] in, p t It is the probability of the target category predicted by the model; α It is a balancing factor that adjusts the problem of class imbalance; γ It is a moderating factor used to reduce focus on simple samples, when γ When =0, FocalLoss degenerates into ordinary cross-entropy loss.

[0103] S342. To better constrain intra-class feature aggregation and optimize the feature space structure, the CenterLoss algorithm is used to optimize the distance between intra-class features in the model. The expression is:

[0104] ;

[0105] in, For the first Features of each sample For the first The center of each sample category, N For batch values, The loss function is centered.

[0106] During training, a batch training method was used with a batch size of 32. The dataset was randomly distributed into training and test sets. The optimization function was stochastic gradient descent with a momentum of 0.9. L2 regularization was set to prevent overfitting. The number of iterations was 25, and the initial learning rate was 0.0001. The learning rate was adjusted every 10 epochs, and each adjustment multiplied the current learning rate by 0.5. To suppress overfitting, the dropout probability was set to 0.5, which means 50% of the nodes were randomly dropped.

[0107] The AlexNet model trained on the massive ImageNet dataset is fine-tuned. The classification output layer is adjusted to suit the target task. The preceding structure is fine-tuned, and the model is trained according to the aforementioned structure and parameters to achieve the new image recognition task, thus obtaining the trained model.

[0108] In some embodiments of this application, in step S4, the acquired current PMU disturbance data of the power system is input into the disturbance identification model, and the disturbance identification result output by the disturbance identification model is obtained, including:

[0109] S41. Obtain the voltage amplitude and frequency data of the current PMU disturbance in the power system;

[0110] S42. Normalize and preprocess the voltage amplitude data and frequency data of the current PMU disturbance, and then perform Gram angle difference field image conversion and feature fusion in sequence to obtain the image to be identified.

[0111] S43. Input the image to be identified into the perturbation recognition model, obtain the confidence score output by the perturbation recognition model, and obtain the perturbation recognition result based on the confidence score.

[0112] To verify the effectiveness of this invention, it was compared with AlexNet (without transfer learning) and CNN (convolutional neural network). Under the same dataset and the same hyperparameters, the accuracy and loss curves of the validation set obtained from training the three models were compared. Figure 2 and Figure 3As shown, at the beginning of training, TL-AlexNet achieved an accuracy of 79.95%, CNN 51.04%, and AlexNet 48.96%. TL-AlexNet had a higher starting accuracy because the transfer-learning network already had initial weights and biases, eliminating the need for retraining from scratch. At epoch 5, TL-AlexNet's accuracy reached 98.69%, CNN 79.17%, and AlexNet 85.68%. Since AlexNet has deeper layers than CNN, it converged faster. At epoch 9, TL-AlexNet's accuracy stabilized at 99.48%, while CNN stabilized at 91.67% at epoch 15, and AlexNet stabilized at 92.45% at epoch 12. In summary, AlexNet outperformed CNN convolutional neural networks. Furthermore, the AlexNet network using transfer learning had a higher starting accuracy and higher convergence speed and value during iteration compared to the AlexNet network without transfer learning.

[0113] To verify the accuracy of this invention, the method was tested using a specific disturbance analysis. This disturbance originated from an actual power grid PMU, and its voltage amplitude and frequency data are as follows: Figure 4 As shown, it is normalized and converted to... Figure 5 The two-dimensional Gram difference field image shown is then placed into the R and G channels of an RGB image for feature fusion, as follows: Figure 6 As shown, the feature map is fed into the trained model for recognition, and the confidence scores are shown in Table 1. The highest confidence score of 0.982590 corresponds to the perturbation type of machine switching. The recognition result of the perturbation recognition method of the present invention is consistent with the actual perturbation type.

[0114] Table 1 Confidence Scores

[0115] ;

[0116] The beneficial effects of this invention are as follows: This invention adopts the above-mentioned perturbation identification method based on Gram angle difference field and transfer learning-AlexNet. Considering the small number of perturbation samples and the imbalance of perturbation sample categories, it proposes a perturbation identification method that can capture complex perturbation patterns, has good noise resistance, and strong generalization ability, thereby improving the perturbation identification accuracy and solving the problems of insufficient perturbation samples, high noise, and low calculation accuracy in the existing technology.

[0117] In this application, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. In case of any inconsistency, the meaning set forth in this specification or derived from the content described herein shall prevail. Furthermore, the terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit the scope of this application.

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A perturbation recognition method based on Gram angle difference field and transfer learning, characterized in that, Includes the following steps: S1. Obtain the time series data of frequency and voltage amplitude corresponding to the disturbance at IEEE 39 node of the power system, and perform normalization preprocessing on the time series data of frequency and voltage amplitude to obtain one-dimensional time series frequency and voltage amplitude data. S2. Extract features from the frequency and voltage amplitude data of the one-dimensional time series and convert them into two-dimensional Gram angle field images. Perform feature fusion on the feature extraction results based on the RGB image channels to obtain feature two-dimensional image data. S3. Input the feature two-dimensional image dataset into the preset transfer learning-AlexNet network model for training, and build a perturbation recognition model based on the SE attention mechanism; S4. Input the acquired current PMU disturbance data of the power system into the disturbance identification model to obtain the disturbance identification result output by the disturbance identification model; In step S2, converting the frequency and voltage amplitude data of the one-dimensional time series into a two-dimensional Gram angle field image includes: S21. Perform angle transformation on the data points in the new time series and encode the timestamps by radius. The expression is: ； in, The normalized time series, for Angle after inverse cosine transformation for The polar radius of the point for The timestamp of the point This is a constant factor for regularizing the polar coordinate system; Obtain the transformed time series ,in, These are the data points in the transformed time series; S22. The Gram matrix is ​​calculated based on the phase difference of the transformed time series. The calculation formula is: ； For Gram matrix, For time series Point and point angle difference; Based on Gram matrix Obtain a two-dimensional gram angular field image; In step S2, feature fusion based on the feature extraction results using RGB image channels includes: S23. Input the feature extraction results into the R and B channels of the RGB image for feature fusion. The feature extraction results include voltage and frequency data after the Gram angle difference field changes. In S3, the perturbation recognition model constructed based on the SE attention mechanism includes: S31. Based on the SE attention mechanism, feature weights are assigned to the feature channels in the training results to obtain global information, expressed as: ； in, For the first Global information for each channel The height of the feature map, The width of the feature map. For feature maps, Number of channels; S32. Based on the SE attention mechanism, the channel feature values ​​are calculated using the first fully connected layer and the ReLU activation function. The expression is: ； in, The weights of the first fully connected layer. The bias is for the first fully connected layer. This is the global description vector for all channels in the first layer. These are channel characteristic values; S33. The channel weights are calculated based on the second fully connected layer using the SE attention mechanism and the sigmoid activation function, and the expression is: ； in, The weights of the second fully connected layer. For the bias of the second fully connected layer, It is the Sigmoid activation function. Channel weights; S34. Optimize the feature channels of the training results based on channel weights, and construct a perturbation recognition model based on the optimization results; S34 also includes: S341. The improved cross-entropy loss function Focal Loss is used to optimize the sample weights. The expression is as follows: ； in, The probability of the predicted target category; As a balance factor; As a regulating factor, An improved cross-entropy loss function; S342. The CenterLoss algorithm is used to optimize the distance of intra-class features in the model. The expression is: ； in, For the first Features of each sample For the first The center of each sample category, For batch values, The loss function is centered.

2. The perturbation recognition method based on Gram angle difference field and transfer learning according to claim 1, characterized in that, In step S1, the normalization preprocessing of the frequency and voltage amplitude time series data includes: S11, obtained from Time series data of frequency and voltage amplitude values ​​at each moment. Normalization preprocessing is performed on it, and the expression is: ； Among them, the frequency and voltage amplitude data of the one-dimensional time series These are the data points in the original time series. For the data points in the new time series obtained after normalization, for The frequency and voltage amplitude time series data are composed of values ​​at each moment.

3. The perturbation recognition method based on Gram angle difference field and transfer learning according to claim 2, characterized in that, In step S4, the acquired current PMU disturbance data of the power system is input into the disturbance identification model, and the disturbance identification results output by the disturbance identification model include: S41. Obtain the voltage amplitude and frequency data of the current PMU disturbance in the power system; S42. Normalize and preprocess the voltage amplitude data and frequency data of the current PMU disturbance, and then perform Gram angle difference field image conversion and feature fusion in sequence to obtain the image to be identified. S43. Input the image to be identified into the perturbation recognition model, obtain the confidence score output by the perturbation recognition model, and obtain the perturbation recognition result based on the confidence score.

Images