CN113591960A - Voltage sag event type identification method based on improved generation countermeasure network - Google Patents

Abstract

The invention discloses a voltage sag event type identification method based on an improved generation countermeasure network, which comprises the following steps: converting the obtained sag three-phase voltage data into a two-dimensional trajectory curve based on a space vector to obtain SPM trajectories of different sag types; acquiring an improved identification model based on a preset identification model, and inputting SPM tracks of different sag types into the improved identification model to acquire a generated sample; data enhancement is realized by using the generated data with the sample characteristics consistent with the real sample characteristics and distribution; the method can effectively reduce the training data amount required by type identification while ensuring the accuracy of the type identification of the sag event, can still keep higher identification accuracy under different actual data scenes, and has great application value and prospect.

Description

Voltage sag event type identification method based on improved generation countermeasure network

Technical Field

The invention relates to a voltage sag event type identification method based on an improved generation countermeasure network, and belongs to the technical field of voltage sag identification.

Background

With the continuous development of the intelligent manufacturing industry, the number of sensitive users (enterprises such as semiconductor production and precision instrument processing) in the power grid is continuously increased, and the influence of voltage sag on the power grid and the users is more and more serious. In order to relieve the influence of voltage sag on a load side, particularly on key users with sensitive loads, a plurality of provincial power companies in China lay a large number of voltage sag monitoring devices in respective service ranges, and the purpose is to provide necessary data support for detailed analysis of sag events.

The method has the advantages that the voltage sag event type is accurately identified, and the main cause of the sag event is determined to be the basis for solving the sag problem. The voltage sag can be mainly divided into a random sag and a planned sag, wherein the random sag is a sag caused by short circuit (single-phase short circuit, two-phase short circuit and three-phase short circuit) faults, and the planned sag is a sag caused by manual operation and comprises two conditions of large-scale induction motor starting and transformer switching.

At present, two main ideas are provided for identifying the voltage sag type, one is to construct a sag sample feature matrix based on manual feature extraction, and train a classification model by using feature data to realize identification of the sag type; and the other method is to adopt a feature self-extraction network to replace signal processing analysis based on manual experience, continuously drive the adjustment of the parameters of the feature self-extraction network through loss calculation, and improve the self-adaptive capacity of the network to the sample, thereby reducing the influence of the feature extraction error on the final sag type identification result.

However, in practical situations, the distribution of sag samples has a great imbalance (the single-phase fault sag has a much higher proportion than other types), and the sag event distribution characteristics in different seasons and different regions are greatly different. Although the methods well avoid errors of artificial feature extraction and simplify the feature extraction process of the samples, a large amount of sample data is required for model training to support, when the number distribution of the samples is unbalanced or less, the feature self-extraction network cannot sufficiently learn various types of features, and the accuracy of the type identification cannot be guaranteed by temporarily reducing.

Disclosure of Invention

The invention aims to solve the problems, and provides a voltage sag event type identification method based on an improved generation countermeasure network, which integrates a CBAM module in a discriminator to improve the characteristic self-extraction capability of a judgment model, thereby improving the performance of the whole AC-GAN network.

A method of voltage sag event type identification based on an improvement generation countermeasure network, the method comprising:

converting the obtained sag three-phase voltage data into a two-dimensional trajectory curve based on a space vector to obtain SPM trajectories of different sag types;

and inputting SPM tracks of different sag types into a pre-established identification model to obtain the type of the voltage sag event, wherein a CBAM attention module is fused into the identification model.

Further, the method for converting the acquired sag three-phase voltage data into the two-dimensional track curve comprises the following steps:

acquiring three-phase instantaneous voltage data, and converting the three-phase instantaneous voltage data into a two-dimensional SPM waveform track through calculation, wherein the calculation formula is as follows:

in the formula V_a,V_b,V_cRespectively instantaneous sampled data of three-phase voltage, V_spmRepresenting the voltage value after the space vector change; alpha is an operator, and alpha is 1 & lt 120 deg.

Further, the identification model comprises a generator G and a discriminator D, the establishment of the identification model comprises the steps that label information of a generated sample is added in the generator G to guide the training of the generator G, meanwhile, the multi-type classification function of the sample is added, and the classification of the multi-type sample is realized on the basis of judging the truth of the sample;

and an attention module is fused in the discriminator D and is used for focusing detail characteristics in various sag SPM tracks.

Further, the attention module comprises a channel attention and space attention two-part network, average pooling and maximum pooling are used for aggregating full-channel information to generate two different space feature descriptions Favg and Fmax, then the space feature descriptions Favg and Fmax are respectively subjected to MLP, and the result is weighted to obtain a parameter distribution weight of each channel.

Further, the formula for calculating the weight of the parameter distribution of the channel is as follows:

M_C(F)＝σ(MLP(avgPool(F))+MLP(MaxPool(F)))

and F represents input convolution characteristics, sigma represents a sigmoid function, and MLP processing parameters of the maximum pooling result and the average pooling result are shared.

Further, taking a channel attention output result as an input of spatial attention, performing maximum pooling and average pooling on data of each position in a feature matrix of each channel, performing convolution dimension reduction aggregation on two pooled results, and finally obtaining spatial attention features through a sigmoid function, wherein a feature calculation expression is as follows:

in the formula, f represents dimension reduction convolution calculation, 7 multiplied by 7 convolution kernel dimension reduction convolution calculation is adopted, and the size of the characteristic matrix before and after convolution is unchanged by a 0 complementing mode.

Furthermore, the loss function of the identification model in the training process consists of two parts, namely the loss L for judging the true and false data source_SSecond, loss L of label category judgment_CThe expression is:

wherein P (C ═ C | X_real) And P (C ═ C | X)_fake) Respectively representing that a probability generator G which judges the true sample and the generated sample data label correctly optimizes an objective function in the training process as follows:

maxL(G)＝max(E[logP(C＝c|X_fake)]-E[logP(S＝fake|X_fake)])

the 'true and false' data source judgment optimization direction is subjected to negation processing and is consistent with the label judgment;

for the common optimization of the judgment of the true and false data sources and the label types in the training process of the discriminator G, the corresponding objective function is as follows:

maxL(D)＝max(L_S+L_C)

the results of the sample source and sample label determination are output by the determiner D, and the performance of the determiner D determines the overall performance of the improved recognition model.

Further, the types of the voltage sag event include six types, which are respectively: single-phase fault (C1), two-phase fault (C2), three-phase fault (C3), compound fault (C4), induction motor start (C5), and transformer switching (C6).

Compared with the prior art, the invention has the following beneficial effects: the invention provides a SPM-based sag voltage data conversion method, which unifies sag sample data into SPM two-dimensional track pictures, effectively distinguishes different types of sag characteristics, and reduces the data volume required to be analyzed for type identification.

Compared with the traditional model, the improved model can more accurately capture the characteristics of each sag sample, so that the identification capability of the whole model on the sag sample types is improved. Under different actual data scenes, the identification accuracy of the full sag type can still be kept to be more than 93%.

Drawings

FIG. 1 is an improved recognition model network architecture incorporating an attention model according to the present invention;

FIG. 2 is a graph of the RMS waveform measured during a sag event in a region of Jiangsu in accordance with the present invention;

FIG. 3 is an SPM trace of a voltage sag event of the present invention;

FIG. 4 is a schematic view of the channel attention of the present invention;

FIG. 5 is a schematic view of the invention with spatial attention;

FIG. 6 is the result of an iteration of the discriminant loss function of the present invention;

FIG. 7 is the classifier accuracy of the present invention;

FIG. 8 IS an IS result comparison between the conventional AC-GAN model and the improved AC-GAN model of the present invention;

FIG. 9 is a comparison of FID results between the conventional AC-GAN model and the improved AC-GAN model of the present invention;

FIG. 10 is a comparison of the improved AC-GAN model of the present invention with a conventional AC-GAN model feature extraction thermodynamic diagram;

FIG. 11 is a sample identification scenario of scenario one of the present invention;

FIG. 12 illustrates sample identification for scenario two of the present invention;

FIG. 13 is a sample identification scenario for scenario three of the present invention.

Detailed Description

In order to make the technical means, the creation characteristics, the achievement purposes and the effects of the invention easy to understand, the invention is further described with the specific embodiments.

As shown in fig. 1, the method for identifying the type of voltage sag event based on the improved generation countermeasure network according to the present invention includes the steps of:

converting voltage sag data based on a space vector, converting the acquired sag three-phase voltage data into a two-dimensional trajectory curve based on the space vector, and obtaining SPM trajectories of different sag types; fig. 2 shows an RMS waveform of actually measured data uploaded by a sag monitoring system in a city in Jiangsu, with a sampling rate of 12.8 kHz. Fig. 2(a) - (g) show single-phase fault, two-phase fault, three-phase fault, single-phase to three-phase fault, two-phase to three-phase fault, induction motor start and transformer switching, respectively. It can be known from the figure that the sag amplitude, the data length and the sag starting and stopping positions of the actually measured data of each sag event are different, and compared with simulation data, the actually measured data contain more interference of factors such as harmonic waves, noise and oscillation, and the difficulty of feature extraction is increased. Therefore, in order to make sample data more regular (only paying attention to the track, and avoiding the influence caused by different lengths of sampling data and different start and stop positions of a sag domain), improve the accuracy of feature self-extraction and simultaneously reduce the data calculation amount, the invention provides a data conversion method based on Space vector transformation (SPM), which converts three-phase instantaneous voltage data into a two-dimensional SPM waveform track, and has the following calculation formula:

in the formula V_a,V_b,V_cRespectively, three-phase voltage instantaneous sampling data, SPM transformation result is shown in FIG. 3, and the horizontal axis is V_SPMThe real part of (a) and the longitudinal axis of (b) is V_SPMThe imaginary part of (c). The order of the transform results is consistent with the order of the RMS waveforms.

As can be seen from fig. 3, the SPM traces of different dip types are clearly distinguished from each other. When no sag occurs, the SPM trace is a standard circle. When the single-phase fault occurs in an asymmetric condition, the voltage amplitude of the sag phase is compensated by the sag phase, and the transformation of the sag phase is approximate to a standard circle; when two-phase failure occurs, the SPM trace distorts into an ellipse. And the three-phase sag of the symmetrical condition and the normal voltage amplitude form a concentric circle together. The SPM tracks of the induction motor starting and the transformer switching type sag also have different characteristics respectively. In addition, the SPM transformation can well depict the fault evolution condition in the fault sag process, the conditions that a single-phase fault is evolved into a three-phase fault and a two-phase fault is evolved into a three-phase fault comprise the tracks of an ellipse and a concentric circle, the detail characteristics of a red shadow part represent the evolution process from a normal state to a three-phase short-circuit fault, and a foundation is laid for the accurate identification of the condition by a model. Based on this, the invention divides the sag types into 6 types, respectively: the method comprises the following steps of single-phase fault (C1), two-phase fault (C2), three-phase fault (C3), compound fault (C4), induction motor starting (C5) and transformer switching (C6).

And step two, identifying a model based on the voltage sag type of the improved AC-GA. The conventional generation countermeasure network (GAN) is composed of a generator (G) and a discriminator (D), and as the name suggests, the generator G generates sample data similar to or even consistent with a real sample according to real sample characteristics to supplement the sample types with insufficient quantity and balance sample distribution. The discriminator D is used for monitoring the quality of the generated samples and judging whether the generated samples are consistent with the characteristics of the real samples or not so as to achieve the purpose of expanding the number of the samples. For generator G, the input of data contains only the generation samples, so the objective function of generator G is:

wherein S represents a data source, X_fakeDenotes a sample generated by the generator, P (S ═ fake | X_fake) Representing the probability that the generated sample belongs to the source of the generated sample data, and E representing the desired calculation. As can be seen from equation (2), the optimization goal of the generator G is such that the smaller the probability that the generated sample is judged as a "false sample", the better, and the optimization goal is completed when the generated sample is sufficient to "falsely". For the discriminator D, the input of data contains both real samples and generated samples, and its objective function is shown in equation (3):

in the formula, X_realDenotes a real sample, P (S ═ real | X_real) Representing the probability that the real sample belongs to the real sample data source, and the optimization target is as follows: the final result can give a correct judgment regardless of whether the input is a real sample or a generated sample. GAN is optimized by alternation of two during trainingAnd (5) playing games mutually, and finally completing model training after Nash balance is achieved.

Furthermore, the invention provides an AC-GAN identification model, label information for generating samples is added in the traditional GAN generator G to guide the training of the generator G, meanwhile, the multi-type classification function of the samples is added, the multi-type sample classification is realized on the basis of judging the truth of the samples, and the application range of the generated countermeasure network is expanded. Therefore, the loss function of the AC-GAN recognition model in the training process is composed of two parts, namely loss L judged by a 'true and false' data source_SSecond, loss L of label category judgment_CThe formula is expressed as:

wherein P (C ═ C | X_real) And P (C ═ C | X)_fake) And respectively representing the probability of judging the true sample and the generated sample data label to be correct. For generator G, the optimization of the loss function in the training process does not contain real data, so the optimization objective function is:

maxL(G)＝max(E[logP(C＝c|X_fake)]-E[logP(S＝fake|X_fake)]) (5)

the 'true and false' data source judgment optimization direction is subjected to negation processing and is consistent with the label judgment. For the discriminator, the training process is the common optimization of the judgment of the true and false data sources and the label types, so the corresponding objective function is as follows:

maxL(D)＝max(L_S+L_C) (6)

the judgment results of the sample source and the sample label are output by a discriminator D, and the performance of the discriminator D determines the overall performance of the AC-GAN. Therefore, the invention adds CBAM (attention module) in the feature extraction network of the discriminator D to focus the detail features in various sag SPM tracks, thereby improving the feature capture capability of the discriminator network and leading the discriminator network to better calculate the training error.

The CBAM attention module can be divided into two parts of channel attention and space attentionThe sub-network first uses the average pooling and maximum pooling to aggregate the full-channel information to generate two different spatial profiles F_avg,F_maxThen, the result is weighted by a convolution sensing network including a single implicit layer to obtain a parameter distribution weight of each channel, and the calculation flow is shown in fig. 4. The calculation expression is:

M_C(F)＝σ(MLP(avgPool(F))+MLP(MaxPool(F))) (7)

and F represents input convolution characteristics, sigma represents a sigmoid function, and MLP processing parameters of the maximum pooling result and the average pooling result are shared.

The channel attention output results are used as the input of the spatial attention. As shown in fig. 5, the data at each position in the feature matrix of each channel is maximally pooled and averagely pooled, then convolution dimension reduction aggregation is performed on the two pooled results, and finally a sigmoid function is used to obtain a spatial attention feature.

The feature calculation expression is:

in the formula, f represents the dimensionality reduction convolution calculation, and a 7 × 7 convolution kernel dimensionality reduction convolution calculation is generally adopted (the size of the feature matrix before and after convolution is not changed by a 0 complementing mode).

As shown in fig. 5, the generator G network used by the model performs 2 times of upsampling and three times of convolution calculation, 0 filling of 1 row in horizontal direction and 1 column in vertical direction is performed on sample data before convolution calculation, the size of a convolution kernel is set to 3 × 3, the step size is 1, the size of an input pixel matrix is not changed after calculation is completed, and finally a 1 × 64 image with the same size as that of a real sample is output. And (3) adopting 4-layer convolution by a discriminator D network used by the model, carrying out 0 filling on sample data in 1 row horizontally and 1 column vertically before carrying out convolution, setting the size of a convolution kernel to be 3 x 3, and setting the step length to be 2. Meanwhile, to prevent overfitting, Dropout operation with a probability size of 0.25 is performed after each layer of convolution is completed. Each layer of convolution module fuses the CBAM attention module, giving higher weight to more valuable features. And D, inputting a two-dimensional SPM track picture with 64 × 64 data by the discriminator D, obtaining 128 channels for each sample after completing feature extraction, entering a feature matrix with 4 × 4 channels into a full connection layer, and selecting a sigmoid classifier for judging a data source, wherein a judgment error is transmitted back to a generation network so as to drive the updating of convolution kernel parameters of the generation network. For the judgment of the sag sample type label, softmax is selected as a multi-classifier, and the error gradient is used as the basis for self network training.

And step three, utilizing the data which is generated based on the method of the step one and the step two and is consistent with the characteristics and the distribution of the real sample, thereby realizing data enhancement.

In one embodiment, the invention improves the basic AC-GAN model in a pytorch by using a python 3.6-based deep learning framework of the pytorch.

In this embodiment, the same data set (140 training samples and 60 test samples for each sag type) is used to train the improved AC-GAN model and the conventional AC-GAN model at the same time, and after 300 times of iterative training, the loss function and the recognition accuracy of the test samples of the discriminator D are shown in fig. 6 and 7:

the CBAM attention module added in the improved model enables the discrimination network to focus on the detail features of each sag type, the feature points in the input picture are more accurately grasped, and the training error descent speed and the training error value are lower than those of the traditional model. Meanwhile, after 50 times of iteration, the identification accuracy of the model gradually enters a stable state, and the accuracy of the improved model is always higher than that of the traditional model in the process of completing the iteration.

Further, in order to evaluate the learning capability of the model for the features more intuitively, the invention introduces probability space score (IS) and feature space distance (FID) to evaluate the quality and diversity of the generated model generation samples.

The IS one of the most commonly used evaluation indexes for generating a model, and the indexes consider the quality of a sample generated by the sample and the distribution of the type of the sample based on the label judgment probability result of the generated sample. The calculation formula is as follows:

wherein the content of the first and second substances,

where P (y | x) is the probability distribution that the sample x belongs to each class, and P (y) is the edge distribution of the entire picture over all classes. KL represents a divergence calculation function to measure the distance between two probability density distributions, and the larger the value, the higher the discrimination between the two probability distributions. As can be seen from the formula, based on onehot coding setting, when the model performance IS superior, the P (y | x) vector IS more prominent at the class dimension, the rest IS close to 0 (the probability density result IS in a peak shape), and P (y) IS opposite, the better the model performance IS, the more the edge probability value IS, and therefore, the overall result of the IS should be increased with the increase of the number of training iterations. The IS results of the improved model compared to the conventional model are shown in fig. 7 as follows:

further, the invention uses FID to quantify the distance between the overall generated sample and the real sample feature distribution, and the calculation formula is as follows:

where g and r represent the generated image and the real image, respectively, mu_gAnd mu_rRepresenting the mean of the respective feature vectors, Σ_gSum-sigma_rAnd representing covariance matrixes of the respective eigenvectors, Tr represents the sum of diagonal elements of the matrixes, and if the result of the root-opening calculation is a complex number, only a real part value is taken. According to the formula, the closer the distance between the characteristic distribution of the generated sample and the characteristic distribution of the real sample is, the smaller the FID value is, the better the performance of the generated model is, and otherwise, the larger the FID value is.

Further, as shown in fig. 9, after the training is completed 50 times, the result of improving the FID of the model is 58.33, which is much lower than that of the conventional model, indicating that the training efficiency of the improved model is better than that of the conventional model. Secondly, after 200 iterations, the FID value of the traditional model has large oscillation, and the change of the improved model is relatively stable, so that the improved model basically completes parameter training after 200 iterations, the countermeasures of the generated network and the judgment network are approximately balanced, and after 500 times of training, the FID value of the improved model can reach 27.2.

As an embodiment, the invention utilizes Grad CAM to perform characteristic thermodynamic diagram visualization on the characteristic position of the important attention of the characteristic extraction network, and FIG. 10 is the comparison between the improved type and the traditional type of the extraction of the characteristic of an individual sample, wherein a red area is the area of the important attention of the network and is the basis for classifying the sample, P represents the probability value of softmax output, P _ C1-P _ C6 are sample prediction labels, and C1-C6 are sample actual labels. As shown in the figure, the feature extraction network of the conventional model does not capture the detail features of individual sag samples, which eventually results in an error in sag type determination. Like the single-phase short-circuit fault (C1) sample, the conventional model focuses on the blank area in the standard circle, while the improved model mainly extracts the features of the standard circle, and the corresponding P value and P _ C result also prove the deviation of the feature extraction of the conventional model (the improved value is 0.967, and the conventional value is only 0.678). For the samples of C3 and C4 types, the focus of attention of the improved model is accurately positioned in the detail part inside the standard circle, the difference between the samples of the type and other types is distinguished, and the finally obtained P value and P _ C result also correspond to the feature details of attention.

According to voltage sag monitoring statistics of Jiangsu power grid, the proportion of the events causing sag due to short-circuit faults is the largest in 6 sag types, wherein the proportion of single-phase short circuits exceeds more than 50%, and compared with the two-phase short circuits and the most serious three-phase short circuits, the occurrence frequency is low. And the planned temporary drop is that the percentage of the induction motor starting and transformer switching events is about 10 percent of the total number of faults. The invention respectively inspects the sag classification effect of the improved model under three different scenes.

Scene one: sag sample equilibrium distribution

In the ideal case, each sag type has 140 training samples, and 60 test samples, for 1200 samples.

Scene two: statistics of sag events in 6 months in a certain city

Scene two is measured data of sag events uploaded by a 10kV power distribution system in a certain area in south of the south of the earth.

Scene three: statistics of 9 months of sag events in a certain city.

Scene three is measured data of the sag events uploaded by a 10kV power distribution system in a certain city in North Suo in 9 months, and the number of the integral sag events is almost reduced to half of that in scene two because the number of days in sunny days in the time period is greatly increased compared with that in 6 months. Because local enterprises adopt a current-limiting soft start mode when a large-capacity motor is started, the rise of the starting current of the motor is well controlled, and only 12 temporary drop events caused by the starting of the induction motor in the whole month exist.

The sag training data distribution of the three scenarios is shown in table 1, and finally the test samples of the three scenarios are all consistent with the test sample of the first scenario.

TABLE 1 three scenarios sag training data distribution

Type of sample	Scene one	Scene two	Scene three
				C1	140	476	247
C2	140	233	112
				C3	140	172	89
C4	140	77	54
				C5	140	63	12
C6	140	56	46
				Total of	840	1077	560

In order to avoid the problem that the ROC tracks are in discrete zigzag due to insufficient number or uneven distribution of samples, the classification result is evaluated by adopting an improved continuous ROC curve, and the continuous expression form of a confusion matrix is shown in the following table 2:

TABLE 2 confusion matrix for improved continuous ROC curves

	Positive sample (actual)	Negative sample (actual)
			Positive sample (prediction)	α(1-Φ+(·))	(1-α)(1-Φ－(·))
Negative sample (prediction)	αΦ+(·)	(1-α)Φ－(·)
			Total of	α	(1-α)

Wherein

Phi denotes a gaussian distribution function, t denotes a threshold linear variation sequence,

respectively representing the prediction probability average value of positive and negative samples,

respectively representing the positive and negative sample prediction probability standard deviations. Alpha is a Positive and negative sample balance factor, and the numeric value in the text is 0.5, so that the influence of the imbalance of the Positive and negative samples on the ROC curve is avoided, and the calculation formulas of the abscissa, the False Positive Rate (FPR) and the ordinate True Positive Rate (TPR) of the ROC curve are as follows:

FPR＝1-Φ_-(·) (12)

with the change of the probability threshold t, the ROC result is changed, and as can be known from the formula, the more discrete the probability distribution of the model for judging the positive class and the negative class is, the larger the corresponding standard deviation is, the farther the final ROC curve is from the diagonal segment line, the better the performance of the reflected classifier is, and conversely, the smaller the probability distribution distance between the positive class and the negative class is, the worse the performance of the classification model is proved, and the closer the ROC curve result is to the diagonal line. The Area Under the ROC Curve (AUC) is usually used to quantify the classification result of the model for each type. The ROC results for the improved model for the three cases are shown in FIGS. 11-13.

The specific quantitative identification result is shown in the following table, wherein AUC represents the integral of the ROC curve on the abscissa, and ACC represents the classification accuracy of the class of samples:

TABLE 3 improved recognition results of AC-GAN model

In fig. 11 to 13, the ROC curve trajectories of the sag types are close to the (0,1) coordinate point, which indicates that the probability values determined by the positive samples are concentrated and clearly distinguished from the probability determination results of the negative samples. Under three scenes, the AUC result of each type of sag is greater than 0.95, and the identification accuracy can be maintained to be more than 93%. The improved model aims at the condition that the induction motor start sag training samples in the scene III are extremely deficient, and the generated sample data well make up the number of the samples, so that the classifier can fully learn the characteristics of the type, and the final judgment result is stable.

In order to further verify the superiority of the method provided by the invention, the K-SVD method (method I), the deep learning model fusion method (method II), the convolutional neural network method (method III), the deep confidence network method (method IV), the S transformation and the multi-dimensional fractal method (method V) are respectively adopted by the method to analyze the three scenes, and the training data and the test data are consistent with the scenes (the detailed calculation results of AUC and ACC are given in the appendix). The six identification methods show higher classification accuracy in three scenes under the scene with sufficient sample quantity and balanced distribution. For the fault type sag, the identification accuracy of the first, second and fifth methods for respectively carrying out phase-by-phase analysis on the three-phase voltage data is high. In contrast, the third and fourth methods have a certain loss of information during data processing, resulting in a lack of recognition capability.

The method is relative to C5 in the third scene, the classification accuracy rate is only 48.3%, in the misclassified samples, atom combinations contain more fault atoms, because the number of fault samples is greatly different from that of C5 and C6 sample data, the dictionary is more sufficient in learning fault modes, the characteristics of C5 and C6 are not sufficiently learned, and the atom modes in the dictionary library are limited, so that the capability of the algorithm for decomposing and matching the samples is deficient. And a fifth method adopts a signal processing mode based on S transformation to extract the characteristics of sag data, uses single-phase sag data as a discussion object (the sag lowest phase is usually selected), and synthesizes the three-phase processing results to output a final type result for comprehensive comparison processing results. The method has high accuracy in judging the single fault sag type. However, in the identification of the composite sag sample, the method cannot well express the characteristics of the fault evolution process, and if the composite sag sample has short duration and unobvious characteristics, the identification capability of the method is limited. The improved AC-GAN provided by the invention can be well adapted to the three different scenes, and the test results are superior to those of a comparison method.

Table 4 results comparing to prior art methods

The above description is only a preferred embodiment of the present invention, and it should be noted that, for those skilled in the art, several modifications and variations can be made without departing from the technical principle of the present invention, and these modifications and variations should also be regarded as the protection scope of the present invention.

Claims (8)

1. A method for identifying a type of voltage sag event based on an improvement of a generation countermeasure network, the method comprising:

converting the obtained sag three-phase voltage data into a two-dimensional trajectory curve based on a space vector to obtain SPM trajectories of different sag types;

and inputting SPM tracks of different sag types into a pre-established identification model to obtain the type of the voltage sag event, wherein a CBAM attention module is fused into the identification model.

2. The voltage sag event type identification method for generating the countermeasure network based on the improvement as claimed in claim 1, wherein the method for converting the acquired sag three-phase voltage data into a two-dimensional trajectory curve comprises the following steps:

acquiring three-phase instantaneous voltage data, and converting the three-phase instantaneous voltage data into a two-dimensional SPM waveform track through calculation, wherein the calculation formula is as follows:

in the formula V_a,V_b,V_cRespectively instantaneous sampled data of three-phase voltage, V_spmRepresenting the voltage value after the space vector change; alpha is an operator, and alpha is 1 & lt 120 deg.

3. The method for identifying the voltage sag event type based on the improved generation countermeasure network is characterized in that the identification model comprises a generator G and a discriminator D, the establishment of the identification model comprises the steps of adding label information of a generated sample in the generator G for guiding the training of the generator G, simultaneously adding a sample multi-type classification function, and realizing the multi-type sample classification on the basis of judging whether the sample is true or false;

and an attention module is fused in the discriminator D and is used for focusing detail characteristics in various sag SPM tracks.

4. The method as claimed in claim 3, wherein the attention module includes two part networks of channel attention and spatial attention, and the method generates two different spatial feature descriptions Favg, Fmax by aggregating full channel information with average pooling and maximum pooling, and then obtains the parameter distribution weight of each channel by including an MLP respectively and weighting the result.

5. The method for identifying the type of the voltage sag event based on the improved generation countermeasure network of claim 4, wherein the formula for calculating the weight of the parameter distribution of the channel is as follows:

M_C(F)＝σ(MLP(avgPool(F))+MLP(MaxPool(F)))

and F represents input convolution characteristics, sigma represents a sigmoid function, and MLP processing parameters of the maximum pooling result and the average pooling result are shared.

6. The method for identifying the type of the voltage sag event based on the improved generation countermeasure network of claim 5, wherein the channel attention output result is used as an input of spatial attention, the data at each position in the feature matrix of each channel are subjected to maximum pooling and average pooling, then the two pooled results are subjected to convolution dimension reduction aggregation, and finally the spatial attention feature is obtained through a sigmoid function, and then the feature calculation expression is as follows:

in the formula, f represents dimension reduction convolution calculation, 7 multiplied by 7 convolution kernel dimension reduction convolution calculation is adopted, and the size of the characteristic matrix before and after convolution is unchanged by a 0 complementing mode.

7. The method for identifying the type of voltage sag event based on the improved generation countermeasure network as claimed in claim 3, wherein the loss function of the identification model during the training process is composed of two parts, namely, the loss L for the judgment of the "true and false" data source_SSecond, loss L of label category judgment_CThe expression is:

wherein P (C ═ C | X_real) And P (C ═ C | X)_fake) The probability generator G respectively representing the judgment of the true sample and the generated sample data label is used for optimizing an objective function in the training process as follows:

maxL(G)＝max(E[logP(C＝c|X_fake)]-E[logP(S＝fake|X_fake)])

the 'true and false' data source judgment optimization direction is subjected to negation processing and is consistent with the label judgment;

for the common optimization of the judgment of the true and false data sources and the label types in the training process of the discriminator G, the corresponding objective function is as follows:

maxL(D)＝max(L_S+L_C)

the results of the sample source and sample label determination are output by the determiner D, and the performance of the determiner D determines the overall performance of the improved recognition model.

8. The method for identifying the type of voltage sag event based on the improvement generation countermeasure network of claim 1, wherein the type of voltage sag event comprises six types, respectively: single-phase fault (C1), two-phase fault (C2), three-phase fault (C3), compound fault (C4), induction motor start (C5), and transformer switching (C6).

Images