A power quality disturbance classification method based on multi-channel separable convolution

Abstract

The present application relates to the technical field of power quality disturbance classification, and particularly relates to a power quality disturbance classification method based on multi-channel separable convolution.The technical scheme comprises the following steps: a power quality disturbance mathematical simulation model is established, Gaussian white noise is added to the disturbance signal to obtain a noisy power quality disturbance signal S; the decomposition scale K and the penalty factor alpha of variational mode decomposition are optimized by using a particle swarm algorithm, the power quality disturbance signal is decomposed by using an improved variational mode decomposition model to obtain multiple intrinsic mode components; the Pearson correlation coefficient is calculated to set a threshold to screen the intrinsic mode components.The present application avoids the interference of redundant modes on classification, generates multi-channel image data in combination with a Markov transition field, retains the dynamic characteristics of the components, uses a multi-channel separable convolution residual network based on weight sharing for classification and identification, makes up for the shortcomings of traditional single-channel identification, and improves the classification accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power quality disturbance classification, and particularly relates to a power quality disturbance classification method based on multi-channel separable convolution. BACKGROUND

[0002] With the promotion of the "double carbon" target and the plan to build a new power system, power electronic devices such as wind energy, light energy, energy storage and converters are widely used to meet the large-scale access of clean energy. However, this extensive access leads to a large number of non-linear disturbance sources, which in turn causes various power quality problems. How to realize complex and massive power quality disturbance signal processing, accurate identification and classification is of great significance to ensure the high-quality and efficient operation of the new power system.

[0003] Most traditional power quality disturbance classification methods are based on one-dimensional time series, and are analyzed through three stages of feature extraction, feature selection and classifier design, which is relatively cumbersome and inefficient. In recent years, power quality disturbance classification technology based on deep learning has been continuously developed, and converting one-dimensional time series into images for identification has gradually become a new idea. The power quality disturbance classification method based on images avoids the process of manually extracting features in traditional methods, and improves the reliability of identification, but still has some problems. Actual power quality disturbance signals often contain noise, and directly converting time series into images will introduce interference information, which will cause the image features to be unable to accurately express, and the classification accuracy will be reduced; in addition, the mainstream single-channel network model is prone to image feature loss, which makes the extraction of target features not obvious.

[0004] Therefore, the present application provides a power quality disturbance classification method based on multi-channel separable convolution. SUMMARY

[0005] The purpose of the present application is to solve the problem of traditional power quality disturbance classification methods being cumbersome and inefficient in the background art, and to provide a power quality disturbance classification method based on multi-channel separable convolution.

[0006] The technical scheme of the present application: a power quality disturbance classification method based on multi-channel separable convolution, comprising the following steps:

[0007] S1, a power quality disturbance mathematical simulation model is established, and Gaussian white noise is added to the disturbance signal to obtain a noisy power quality disturbance signal ;

[0008] S2, the decomposition scale of variational mode decomposition (VMD) is optimized using a particle swarm algorithm and a penalty factor , through the improved variational mode decomposition model, the power quality disturbance signal is decomposed to obtain a plurality of intrinsic mode components; VMD is the abbreviation of variational mode decomposition, and VMD used in the following represents variational mode decomposition;

[0009] S3, a Pearson correlation coefficient is calculated, intrinsic mode components (IMC) are screened by setting a threshold, and a plurality of effective correlation mode components are obtained; IMC is the abbreviation of intrinsic mode component, and IMC used in the following represents intrinsic mode component.

[0010] S4, the plurality of effective correlation mode components are converted into a two-dimensional image by using a Markov transition field, and a plurality of channel power quality disturbance image data are generated;

[0011] S5, a power quality disturbance classifier is built based on a weight sharing multi-channel separable convolution residual network to perform identification and classification.

[0012] Optionally, in the S2, the following steps are specifically included:

[0013] S201: the decomposition scale of VMD and the penalty factor are determined based on a particle swarm algorithm , a particle swarm is initialized, the total number of particles is set to , the iteration number is , the maximum iteration number is , and the optimization variable is the decomposition scale and the penalty factor ;

[0014] S202: the fitness value is calculated according to the current position vector of each particle in the first generation , and the particle corresponding to the maximum fitness value is selected as the global optimal particle in the first generation , and an envelope entropy function is selected as the fitness function, and the envelope entropy function expression is:

[0015] wherein, is the envelope signal of the first component after decomposition of the power quality disturbance signal ; is the decomposition scale, and a total of components are represented; is the probability of each mode component, ​​​​​​​​​​The position corresponding to the minimum envelope entropy function value indicates the optimal decomposition effect of the perturbation signal;

[0016] The algorithm terminates when it reaches the maximum number of iterations, outputting the optimal fitness value and particle position, and correspondingly, the optimal decomposition scale is obtained. and penalty factor This leads to an improved VMD model;

[0017] S203: The power quality disturbance signal is decomposed using the improved VMD model to obtain K IMF components. , Indicates the first indivual Quantity, .

[0018] Optionally, S3 is specifically as follows:

[0019] S301: Calculate the modal components obtained from the improved VMD decomposition. = With power quality disturbance signal = The Pearson correlation coefficient between them is expressed as follows:

[0020]

[0021] in, For the first Modal components The One element, Power quality disturbance signal The One element, , The total number of elements.

[0022] S302: According to the formula:

[0023]

[0024] Solve for the threshold for selecting effective correlated mode components. , will each The corresponding Pearson correlation coefficient With threshold In comparison, those selected are greater than the threshold. The effective correlated mode components are obtained. One effective correlated modal component, , Indicates the first One effective correlated modal component, .

[0025] Optionally, in the S4, the following steps are included:

[0026] S401: For the obtained effective correlation modal component, each component is divided into intervals, and a state space is established. = = , represents the i-th interval in the state space The elements in correspond one-to-one with the elements in

[0027] S402: For intervals and , the transition probability formula between them is calculated as follows:

[0028]

[0029] where is the transition probability of intervals and ; , , are the i-th and j-th elements in component , , , , represents the probability of the next element being in interval under the condition that element is in interval ; represents belonging to

[0030] S403: Calculate the Markov dynamic probability transition matrix , the matrix dimension is , and the formula is as follows:

[0031]

[0032] The calculation formula of the element in the i-th row and j-th column of the matrix is as follows:

[0033] ​​​​​​​​​

[0034] wherein, , are the components of the first and the second element, . denotes the probability of the element to be in the interval under the condition that the element is in the interval ; ;

[0035] S404: performing steps S401 to S403 for each of the valid relevant modal components, and finally obtaining a multi-channel power quality disturbance image data set .

[0036] Optionally, in the S5, the steps specifically include the following steps:

[0037] S501: inputting the generated power quality disturbance image data set into the multi-channel separable convolution residual network, and the structure of each residual network includes parameters of convolution, pooling and other layers, which are consistent; S502: performing convolution calculation on the power quality disturbance image by using a deep separable convolution layer to extract feature information of the image; the deep separable convolution layer includes two parts of deep convolution and point convolution;

[0038] S503: performing batch normalization on the disturbance feature vector extracted by the deep separable convolution layer;

[0039] S504: applying a nonlinear activation function Swish function to the feature vector normalized by the BN layer;

[0040] S505: using a maximum pooling layer to highlight important parts in the power quality disturbance feature vector and reduce the data dimension at the same time;

[0041] S506: inputting the feature data after the pooling into the structure of the residual network to further speed up the training speed of the network and avoid gradient disappearance, and the structure of the residual network includes first to fourth residual layers, and the structure of each layer is the same;

[0042] S506: inputting the feature data after the pooling into the structure of the residual network to further speed up the training speed of the network and avoid gradient disappearance, and the structure of the residual network includes first to fourth residual layers, and the structure of each layer is the same;

[0043] ​​The first residual layer comprises a first convolutional layer, a first BN layer, a first ReLU layer, a second convolutional layer, a second BN layer, a Shortcut connection unit, a summation unit, a second ReLU layer, the first convolutional layer, the first BN layer, the first ReLU layer, the second convolutional layer, the second BN layer, the summation unit and the second ReLU layer are sequentially and orderly connected, an input end of the Shortcut connection unit is connected with an input end of the first convolutional layer, and an output end of the Shortcut connection unit is connected with an input end of the summation unit;

[0044] A Dropout layer is added between the first to fourth residual layers, and a dropout parameter probability is set to 0.2;

[0045] S507: Classify the multi-channel power quality disturbance features through a full connection layer and a Softmax function;

[0046] The full connection layer fuses the multi-channel image features extracted by the upper layer into a one-dimensional vector;

[0047] The Softmax function converts the output of the full connection layer into a category probability.

[0048] Compared with the prior art, the present application has at least one of the following beneficial technical effects:

[0049] 1. The present application decomposes the disturbance signal into multiple modal components through improved variational modal decomposition, and sets a threshold by calculating the Pearson correlation coefficient to screen effective correlation modal components, thereby avoiding the interference of redundant modalities on classification and enhancing the robustness of model recognition.

[0050] 2. The present application generates multi-channel image data by combining the obtained effective correlation modal components with a Markov transition field, thereby retaining the dynamic characteristics of the components and expanding the image generation method of the disturbance signal.

[0051] 3. The present application makes up for the shortcomings of single-channel recognition through a multi-channel residual network, the network structures of multiple channels are the same and the weights are shared, so that the features of each component of the disturbance signal can be fully extracted, and finally the multi-channel features are fused through a full connection layer, thereby improving the classification accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 It is a flowchart of a power quality disturbance classification method based on a multi-channel separable convolution. DETAILED DESCRIPTION

[0053] The technical solutions of the present application are further described below in combination with the drawings and specific embodiments.

[0054] EMBODIMENT

[0055] As Figure 1As shown, the power quality disturbance classification method based on multi-channel separable convolution is proposed, and each step is described in detail below.

[0056] S1: Establish a mathematical simulation model of power quality disturbance, and add Gaussian white noise to the disturbance signal to obtain a noisy power quality disturbance signal ;

[0057] In this embodiment, according to the IEEE Std.1159-2019 standard, 7 different types of disturbance signals are simulated and generated, labeled as C1-C7; 1000 samples of each type are generated, divided into training set and test set according to 8:2; the sampling rate is 6.4kHz, and 30dB Gaussian white noise is added to each sample;

[0058] S2: Use the particle swarm algorithm to optimize the decomposition scale of variational mode decomposition (VMD) and the penalty factor , decompose the power quality disturbance signal by the improved VMD model to obtain multiple intrinsic mode components (IMFs) ; Specifically as follows:

[0059] S201: Determine the decomposition scale of VMD based on the particle swarm algorithm and the penalty factor , initialize the particle swarm, set the total number of particles , the number of iterations is , the maximum number of iterations is , the optimization variable is the decomposition scale and the penalty factor , initialize the position and velocity vectors of the particles in the first generation particle swarm.

[0060] S202: Calculate the fitness value according to the current position vector of each particle in the first generation, and select the particle corresponding to the maximum fitness value as the global optimal particle in the first generation; select the envelope entropy function as the fitness function, and the envelope entropy function expression is shown in formula (1):

[0061] (1)

[0062] Where, is the envelope signal of the first IMF component after decomposition of the power quality disturbance signal ; is the decomposition scale, indicating that there are components in total ; ​probabilities for each modal component, perturbation signal , the position corresponding to the minimum envelope entropy function value represents the best decomposition effect of the perturbation signal.

[0063] When the algorithm reaches the maximum number of iterations, the loop is ended, and the best fitness value and particle position are output, and accordingly, the optimal decomposition scale is obtained and the penalty factor , to obtain an improved VMD model.

[0064] S203: decompose the power quality disturbance signal by using the improved VMD model to obtain IMF components, , representing the component, .

[0065] S3: screen the IMFs by setting a threshold value through calculation of the Pearson correlation coefficient to obtain effective correlation modal components;

[0066] In this embodiment, S3 is specifically as follows:

[0067] S301: calculate the Pearson correlation coefficient between the modal components obtained by the improved VMD decomposition = and the power quality disturbance signal = . The expression of the Pearson correlation coefficient is shown in formula (2):

[0068] (2)

[0069] wherein, is the element of the modal component, is the element of the power quality disturbance signal, , is the total number of elements. S302: solve the threshold value for selecting effective correlation modal components according to formula (3) , compare the Pearson correlation coefficient

[0070] of each with the threshold value , wherein the effective correlation modal component is selected to be greater than the threshold value , to obtain ​​​an effective relevant modal component, , an effective relevant modal component, an effective relevant modal component, .

[0071] (3).

[0072] S4: converting the screened out effective relevant modal components into two-dimensional images by using Markov transition field (MTF), to generate multi-channel power quality disturbance image data;

[0073] In the embodiment, the S4 is specifically as follows:

[0074] S401: for the obtained effective relevant modal components, dividing each component = into intervals to establish a state space = , wherein the th interval in the state space , there are intervals and the data amount in each interval is the same; the elements in correspond one by one to the elements in

[0075] S402: for the intervals and , calculating the transition probability therebetween as shown in formula (4):

[0076] (4)

[0077] wherein is the transition probability of the intervals and ; , , are the th and the th element in the component , , denotes the probability of the next element being in the interval under the condition that the element is in the interval ; denotes belonging to

[0078] S403: Calculate the Markov dynamic probability transition matrix The matrix dimension is As shown in equation (5), the element in the f-th row and h-th column of the matrix. The calculation formula is shown in equation (6):

[0079] (5)

[0080] (6)

[0081] in, , Components The first in and the One element, . Represents computed elements In the interval Under the condition, element In the interval The probability of; .

[0082] S4.4: For Each effective correlated modal component is processed through steps S401 to S403, ultimately yielding... Multi-channel power quality disturbance image data.

[0083] S5: A power quality disturbance classifier is built based on a weight-sharing multi-channel separable convolutional residual network for identification and classification.

[0084] In this embodiment, S5 is specifically as follows:

[0085] S501: The generated The power quality perturbation image dataset is input into the multi-channel separable convolutional residual network, each The structure of each residual network, including the parameters of convolutional, pooling, and other layers, is consistent.

[0086] S502: The depth-separable convolutional layer is used to perform convolution calculations on the power quality disturbance image to extract the image's feature information; the depth-separable convolutional layer includes two parts: depth convolution and point convolution, which greatly reduces the number of parameters and computational load.

[0087] S503: Perform batch normalization (BN) on the perturbation feature vectors extracted by the depthwise separable convolutional layer to accelerate the training process and improve the stability of the model;

[0088] S504: Applying the nonlinear activation function Swish to the normalized feature vector of the BN layer can provide better gradient propagation;

[0089] S505: highlight important parts in the power quality disturbance feature vector using the max pooling layer, reduce the data dimension, and improve the robustness of the model;

[0090] S506: input the pooled feature data into the residual network structure to further speed up the training of the network and avoid gradient disappearance. The residual network structure includes first to fourth residual layers, and each layer has the same structure. The first residual layer includes a first convolutional layer, a first BN layer, a first ReLU layer, a second convolutional layer, a second BN layer, a Shortcut connection unit, a summation unit, and a second ReLU layer. The first convolutional layer, the first BN layer, the first ReLU layer, the second convolutional layer, the second BN layer, the summation unit, and the second ReLU layer are sequentially connected. The input end of the Shortcut connection unit is connected with the input end of the first convolutional layer, and the output end thereof is connected with the input end of the summation unit. To prevent overfitting during model training, a Dropout layer is added between the first to fourth residual layers, and the parameter dropout probability is set to 0.2.

[0091] S507: classify the multi-channel power quality disturbance features through a fully connected layer and a Softmax function. The fully connected layer fuses the multi-channel image features extracted by the upper layer into a one-dimensional vector. The Softmax function converts the output of the fully connected layer into a class probability.

[0092] In summary, the power quality disturbance classification method based on the multi-channel separable convolution proposed in the embodiment avoids the interference of redundant modalities on classification by screening out the effective relevant modal components of the disturbance signal, generates multi-channel image data by combining the Markov transition field, retains the dynamic characteristics of the components, uses the multi-channel separable convolution residual network based on weight sharing for classification and recognition, makes up for the shortcomings of traditional single-channel recognition, and improves the classification accuracy.

[0093] The above specific embodiments are only several optional embodiments of the present application. Based on the technical solutions of the present application and the related inspiration of the above embodiments, those skilled in the art can make various alternative improvements and combinations on the above specific embodiments.

Claims

1. A power quality perturbation classification method based on multi-channel separable convolution, characterized in that, Includes the following steps: S1. Establish a mathematical simulation model of power quality disturbance and add Gaussian white noise to the disturbance signal to obtain a noisy power quality disturbance signal. ; S2. Optimize the decomposition scale of variational mode decomposition using particle swarm optimization. and penalty factor The power quality disturbance signal is decomposed by an improved variational mode decomposition model to obtain multiple intrinsic mode components. S3. Calculate the Pearson correlation coefficient, set a threshold, and filter the intrinsic mode components to obtain... One effective correlated modal component; S4. Using Markov transition fields to filter out... Convert each effective correlated mode component into a two-dimensional image, and generate Multichannel power quality disturbance image data; S5. A power quality disturbance classifier is built based on a weight-sharing multi-channel separable convolutional residual network for identification and classification; specifically, the following steps are included: S501: The generated The power quality perturbation image dataset is input into the multi-channel separable convolutional residual network, each The structure of each residual network, including the parameters of the convolution and pooling layers, is consistent; the structure of the residual network includes the first to fourth residual layers, and the structure of each layer is the same. The first residual layer includes a first convolutional layer, a first batch normalization (BN) layer, a first ReLU layer, a second convolutional layer, a second batch normalization (BN) layer, a shortcut connection unit, a summing unit, and a second ReLU layer. The first convolutional layer, the first batch normalization (BN) layer, the first ReLU layer, the second convolutional layer, the second batch normalization (BN) layer, the summing unit, and the second ReLU layer are connected sequentially. The input of the shortcut connection unit is connected to the input of the first convolutional layer, and its output is connected to the input of the summing unit. A Dropout layer is added between the first and fourth residual layers, with the probability of dropping parameters set to 0.2; S502: The power quality disturbance image is convolutionally calculated using a depthwise separable convolutional layer to extract the image's feature information; the depthwise separable convolutional layer includes two parts: depthwise convolution and pointwise convolution. S503: Perform batch normalization on the perturbation feature vectors extracted by the depthwise separable convolutional layer; S504: Apply the non-linear activation function Swish to the feature vectors normalized by the BN layer; S505: Utilizes max pooling layers to highlight important parts of the power quality perturbation feature vector while reducing data dimensionality; S506: Input the pooled feature data into the structure of the residual network to further accelerate the training speed of the network and avoid gradient vanishing; S507: Classifies multi-channel power quality perturbation characteristics using a fully connected layer and a Softmax function; The fully connected layer fuses the multi-channel image features extracted from the upper layer into a one-dimensional vector; The Softmax function converts the output of the fully connected layer into class probabilities.

2. The power quality perturbation classification method based on multi-channel separable convolution according to claim 1, characterized in that, S2 specifically includes the following steps: S201: Determining the Decomposition Scale of VMD Based on Particle Swarm Optimization and penalty factor Initialize the particle swarm and set the total number of particles. The number of iterations is The maximum number of iterations is The optimization variable is the decomposition scale. and penalty factor ; S202: According to the... Calculate the fitness value for each particle's current position vector, and select the particle with the highest fitness value as the [missing value]. For the globally optimal particle of the generation, the envelope entropy function is chosen as the fitness function. The expression for the envelope entropy function is: ,in, Power quality disturbance signal The decomposed first indivual The envelope signal of the component, ; To decompose the scale, representing the common indivual Quantity; The probability of each modal component, Disturbance signal The position corresponding to the minimum envelope entropy function value indicates the optimal decomposition effect of the perturbation signal; The algorithm terminates when it reaches the maximum number of iterations, outputting the optimal fitness value and particle position, and correspondingly, the optimal decomposition scale is obtained. and penalty factor This leads to an improved VMD model; S203: The power quality disturbance signal is decomposed using the improved VMD model to obtain... One IMF component, , Indicates the first indivual Quantity, .

3. The power quality perturbation classification method based on multi-channel separable convolution according to claim 1, characterized in that, S3 is specifically as follows: S301: Calculate the modal components obtained from the improved VMD decomposition. = With power quality disturbance signal = The Pearson correlation coefficient between them is expressed as follows: ,in, For the first Modal components The One element, Power quality disturbance signal The One element, , The total number of elements; S302: According to the formula: Solve for the threshold for selecting effective correlated mode components. , will each The corresponding Pearson correlation coefficient With threshold In comparison, those selected are greater than the threshold. The effective correlated mode components are obtained. One effective correlated modal component, , Indicates the first One effective correlated modal component, .

4. The power quality perturbation classification method based on multi-channel separable convolution according to claim 1, characterized in that, In step S4, the specific steps are as follows: S401: Regarding the obtained Each of the effective correlated modal components will be used to determine the effective correlated modal components. = Divided into Establish a state space for each interval. = , Representing the state space The first in Each interval There are a total of There are several intervals, and the amount of data in each interval is the same; elements in The elements in the database correspond one-to-one; S402: For intervals and The formula for calculating the transition probability between them is: ,in, For interval and The transition probability, ; , Components The first in and the One element, , Represents computed elements In the interval Under the condition, the next element In the interval The probability of; Indicates belonging to; S403: Calculate the Markov dynamic probability transition matrix The matrix dimension is The formula is: The first in the matrix Line 1 Column elements The calculation formula is: ,in, , Components The first in and the One element, ; Represents computed elements In the interval Under the condition, element In the interval The probability of; ; S404: For Each effective correlated modal component is processed through steps S4O1 to S4O3, ultimately yielding... Multi-channel power quality disturbance image data.

Images