Power quality disturbance classification method based on multi-channel separable convolution

Abstract

The invention relates to the technical field of power quality disturbance classification, in particular to a power quality disturbance classification method based on multichannel separable convolution. According to the technical scheme, the method comprises the following steps that a power quality disturbance mathematical simulation model is established, Gaussian white noise is added into a disturbance signal, and a noisy power quality disturbance signal S is obtained; optimizing a decomposition scale K and a penalty factor alpha of variational mode decomposition by using a particle swarm algorithm, and decomposing the power quality disturbance signal through an improved variational mode decomposition model to obtain a plurality of intrinsic mode components; and calculating a Pearson's correlation coefficient set threshold to screen the intrinsic mode components. According to the method, the interference of a redundant mode on classification is avoided, the multi-channel image data is generated in combination with the Markov conversion field, the dynamic characteristics of components are reserved, the multi-channel separable convolution residual network based on weight sharing is used for classification and recognition, the defects of traditional single-channel recognition are overcome, and the classification accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power quality disturbance classification, and in particular to a power quality disturbance classification method based on multi-channel separable convolution. Background Art

[0002] With the advancement of the "dual carbon" goal and the plan to build a new power system, power electronic equipment such as wind energy, solar energy, energy storage and converters are widely used to meet the large-scale access of clean energy. However, this widespread access has led to massive and nonlinear disturbance sources, which in turn has caused a variety of power quality problems. How to achieve 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. They are studied and analyzed through three stages: feature extraction, feature selection, and classifier design. The calculation process is relatively cumbersome and inefficient. In recent years, power quality disturbance classification technology based on deep learning has continued to develop, and converting one-dimensional time series into images for recognition has gradually become a new idea. The image-based power quality disturbance classification method avoids the process of manual feature extraction in traditional methods and improves the reliability of recognition, but there are still some problems. The actual power quality disturbance signal often contains noise. Directly converting the time series into an image will introduce interference information, resulting in the inability to accurately express the image features and reduce the classification accuracy. In addition, the mainstream single-channel network model is prone to the problem of image feature loss, which makes the extraction of target features unclear.

[0004] Therefore, this application proposes a power quality disturbance classification method based on multi-channel separable convolution. Summary of the invention

[0005] The purpose of the present invention is to propose a power quality disturbance classification method based on multi-channel separable convolution in order to solve the problem that the traditional power quality disturbance classification method in the background technology is cumbersome and inefficient.

[0006] The technical solution of the present invention is a method for classifying power quality disturbances based on multi-channel separable convolution, comprising the following steps:

[0007] 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 S;

[0008] S2. The particle swarm algorithm is used to optimize the decomposition scale K and penalty factor α of variational mode decomposition (VMD), and the power quality disturbance signal is decomposed by an improved variational mode decomposition model to obtain multiple intrinsic mode components; VMD is the abbreviation of variational mode decomposition, and VMD used in the following text means variational mode decomposition;

[0009] S3, calculate the Pearson correlation coefficient and set the threshold for the intrinsic mode component (IMF s ) to obtain P effective related modal components; IMF s It is the abbreviation of intrinsic mode component. IMF used below s represents the eigenmode component.

[0010] S4, using the Markov transition field to convert the screened P effective related modal components into a two-dimensional image, generating n×n×P multi-channel power quality disturbance image data;

[0011] S5. A power quality disturbance classifier is built based on a multi-channel separable convolutional residual network with weight sharing for identification and classification.

[0012] Optionally, S2 specifically includes the following steps:

[0013] S201: Determine the decomposition scale K and penalty factor α of VMD based on the particle swarm algorithm, initialize the particle swarm, set the total number of particles D, the number of iterations to e, the maximum number of iterations to E, and the optimization variables to the decomposition scale K and penalty factor α;

[0014] S202: Calculate the fitness value according to the current position vector of each particle of the e-th generation, and select the particle corresponding to the maximum fitness value as the global optimal particle of the e-th generation, and select the envelope entropy function as the fitness function. The envelope entropy function expression is:

[0015]

[0016] Among them, s i is the envelope signal of the i-th IMF component after the power quality disturbance signal S is decomposed, i = 1, 2, ..., K; K is the decomposition scale, indicating that there are K IMF components in total; g i is the probability of each modal component, H g is the envelope entropy of the disturbance signal S. The position corresponding to the minimum envelope entropy function value indicates that the decomposition effect of the disturbance signal is the best.

[0017] When the algorithm reaches the maximum number of iterations, the loop ends and the optimal fitness value and particle position are output. Accordingly, the optimal decomposition scale K and penalty factor α are obtained, and the improved VMD model is obtained;

[0018] S203: Decomposing the power quality disturbance signal using the improved VMD model to obtain K IMF components, {IMF k}:={IMF1, IMF2,…,IMF k}, IMF k represents the kth IMF component, k = 1, 2,…, K.

[0019] Optionally, the S3 is as follows:

[0020] S301: Calculate the modal component IMF obtained by improved VMD decomposition k ={x1, x2, ..., x n} and the power quality disturbance signal S = {s1, s2, ..., s n}, the Pearson correlation coefficient expression is:

[0021]

[0022] Among them, x i is the kth modal component IMF k The i-th element of s i is the i-th element of the power quality disturbance signal S, i=1,2,…,n, and n is the total number of elements.

[0023] S302: According to the formula:

[0024]

[0025] Solve the threshold μ for selecting effective related modal components and convert each IMF k The corresponding Pearson correlation coefficient R is compared with the threshold μ, and the one greater than the threshold μ is selected as the effective correlation modal component, and P effective correlation modal components are obtained, {IMF p}:={IMF1, IMF2,…,IMF P},IMF p represents the pth effective correlated modal component, p = 1, 2, …, P.

[0026] Optionally, in S4, the following specific steps are performed:

[0027] S401: For the obtained P effective correlated modal components, each component IMF p ={x1, x2, ..., x n} is divided into Q intervals, and the state space O = {q1, q2, ..., q Q}, q i represents the i-th interval in the state space O, i = 1, 2, ..., Q, there are Q intervals in total and the amount of data in each interval is the same; qi Elements in IMF p The elements in correspond to each other;

[0028] S402: For interval q i and q j , the formula for calculating the transition probability between them is:

[0029] w ij =w(x t ∈q j |x t-1 ∈q i )

[0030] Among them, w i,j For interval q i and q j The transition probability of i,j=1,2,…,Q; x t-1 , x t IMF p The t-1th and tth elements in , t=1,2,…,n, w represents the calculation element x t-1 In the interval q i Under the condition that t In the interval q j The probability of; ∈ means belonging to;

[0031] S403: Calculate the Markov dynamic probability transfer matrix M, the matrix dimension is n×n, and the formula is:

[0032]

[0033] The element m in the fth row and hth column of the matrix f,h The calculation formula is:

[0034]

[0035] Among them, x f , x h IMF p The fth and hth elements in m, f,h=1,2,…,n. f,h Represents the calculation element x f In the interval q i Under the condition that h In the interval q j The probability of; i, j = 1, 2, ..., Q;

[0036] S404: Steps S401 to S403 are performed on P effective related modal components, and finally n×n×P multi-channel power quality disturbance image data are obtained.

[0037] Optionally, S5 specifically includes the following steps:

[0038] S501: inputting the generated n×n×P power quality disturbance image data set into the multi-channel separable convolutional residual network, wherein the structure of each P residual network, including parameters of convolution, pooling and other layers, is consistent;

[0039] S502: performing convolution calculation on the power quality disturbance image using a depthwise separable convolution layer to extract feature information of the image; the depthwise separable convolution layer includes two parts: depthwise convolution and pointwise convolution;

[0040] S503: performing batch normalization on the disturbance feature vector extracted by the depthwise separable convolutional layer;

[0041] S504: Applying a nonlinear activation function Swish function to the feature vector normalized by the BN layer;

[0042] S505: using a maximum pooling layer to highlight important parts of the power quality disturbance feature vector while reducing the data dimension;

[0043] S506: inputting the pooled feature data into a residual network module to further speed up the network training speed and avoid gradient disappearance, wherein the residual network module includes first to fourth residual layers, and each layer has the same structure;

[0044] 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 summing unit, and a second ReLU layer, wherein the first convolutional layer, the first BN layer, the first ReLU layer, the second convolutional layer, the second BN layer, the summing unit, and the second ReLU layer are sequentially connected, an input end of the shortcut connection unit is connected to an input end of the first convolutional layer, and an output end thereof is connected to an input end of the summing unit;

[0045] Add a Dropout layer between the first to fourth residual layers, and set the probability of dropping parameters to 0.2;

[0046] S507: Classify multi-channel power quality disturbance features through a fully connected layer and a Softmax function;

[0047] The fully connected layer fuses the multi-channel image features extracted by the upper layer into a one-dimensional vector;

[0048] The Softmax function converts the output of the fully connected layer into class probabilities.

[0049] Compared with the prior art, the present invention has at least one of the following beneficial technical effects:

[0050] 1. The present invention decomposes the disturbance signal into multiple modal components by improving variational modal decomposition, and sets a threshold for screening by calculating the Pearson correlation coefficient to obtain effective relevant modal components, thereby avoiding the interference of redundant modes on classification and enhancing the robustness of model recognition;

[0051] 2. The present invention generates multi-channel image data by combining the obtained effective correlation modal components with the Markov transformation field, thereby retaining the dynamic characteristics of the components and expanding the image generation method of the disturbance signal;

[0052] 3. The present invention 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; finally, the features of multiple channels are fused using a fully connected layer to improve the classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 Flowchart of a power quality disturbance classification method based on multi-channel separable convolution. DETAILED DESCRIPTION

[0054] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0055] Example

[0056] like Figure 1 As shown, the present invention proposes a power quality disturbance classification method based on multi-channel separable convolution, and each step is described in detail below.

[0057] 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 S;

[0058] In this embodiment, according to the IEEE Std.1159-2019 standard, 7 different types of disturbance signals are simulated and generated, which are marked as C1-C7 respectively; 1000 samples are generated for each type, which are divided into training set and test set at 8:2; the sampling rate is 6.4kHz, and 30dB Gaussian white noise is added to each sample;

[0059] S2: Particle swarm optimization is used to optimize the decomposition scale K and penalty factor α of variational mode decomposition (VMD). The power quality disturbance signal is decomposed through the improved VMD model to obtain multiple intrinsic mode components (IMF s ); specifically as follows:

[0060] S201: Determine the decomposition scale K and penalty factor α of VMD based on the particle swarm algorithm, initialize the particle swarm, set the total number of particles D, the number of iterations to e, the maximum number of iterations to E, the optimization variables to the decomposition scale K and penalty factor α, and initialize the position and velocity vectors of the particles in the e-th generation particle swarm.

[0061] S202: Calculate the fitness value according to the current position vector of each particle of the e-th generation, and select the particle corresponding to the maximum fitness value as the global optimal particle of the e-th generation; select the envelope entropy function as the fitness function, and the envelope entropy function expression is shown in formula (1):

[0062]

[0063] Among them, s i is the envelope signal of the i-th IMF component after the power quality disturbance signal S is decomposed, i = 1, 2, ..., K; K is the decomposition scale, indicating that there are K IMF components in total; g i is the probability of each modal component, H g is the envelope entropy of the disturbance signal S. The position corresponding to the minimum envelope entropy function value indicates that the decomposition effect of the disturbance signal is the best.

[0064] When the algorithm reaches the maximum number of iterations, the loop ends and the optimal fitness value and particle position are output. Accordingly, the optimal decomposition scale K and penalty factor α are obtained, and the improved VMD model is obtained.

[0065] S203: Decomposing the power quality disturbance signal using the improved VMD model to obtain K IMF components, {IMF k}:={IMF1, IMF2,…,IMF k}, IMF k represents the kth IMF component, k = 1, 2,…, K.

[0066] S3: Filter IMFs by calculating the Pearson correlation coefficient and setting a threshold to obtain P effective correlated modal components;

[0067] In this embodiment, S3 is specifically as follows:

[0068] S301: Calculate the modal component IMF obtained by improved VMD decomposition k ={x1, x2, ..., x n} and the power quality disturbance signal S = {s1, s2, ..., s n The Pearson correlation coefficient between}. The Pearson correlation coefficient expression is shown in formula (2):

[0069]

[0070] Among them, x i is the kth modal component IMF k The i-th element of s i is the i-th element of the power quality disturbance signal S, i=1,2,…,n, and n is the total number of elements.

[0071] S302: According to equation (3), the threshold μ for selecting the effective correlated modal components is solved, and each IMF k The corresponding Pearson correlation coefficient R is compared with the threshold μ, and the one greater than the threshold μ is selected as the effective correlation modal component, and P effective correlation modal components are obtained, {IMF p}:={IMF1, IMF2,…,IMF P}, IMF p represents the pth effective correlated modal component, p = 1, 2, …, P.

[0072]

[0073] S4: using the Markov transition field (MTF) to convert the screened P effective related modal components into a two-dimensional image, generating n×n×P multi-channel power quality disturbance image data;

[0074] In this embodiment, S4 is specifically as follows:

[0075] S401: For the obtained P effective correlated modal components, each component IMF p ={x1, x2, ..., x n} is divided into Q intervals, and the state space O = {q1, q2, ..., q Q}, q i represents the i-th interval in the state space O, i = 1, 2, ..., Q, there are Q intervals in total and the amount of data in each interval is the same; q i Elements in IMF p The elements in correspond to each other;

[0076] S402: For interval q i and q j , the transition probability between them is calculated as shown in formula (4):

[0077] w ij =w(x t ∈q j |x t-1 ∈q i ) (4)

[0078] Among them, w i,j For interval q i and q jThe transition probability of i,j=1,2,…,Q; x t-1 , x t IMF p The t-1th and tth elements in , t=1,2,…,n, w represents the calculation element x t-1 In the interval q i Under the condition that t In the interval q j The probability of; ∈ means belonging to;

[0079] S403: Calculate the Markov dynamic probability transfer matrix M, the matrix dimension is n×n, as shown in formula (5). The element m in the fth row and hth column of the matrix f,h The calculation formula is shown in formula (6):

[0080]

[0081]

[0082] Among them, x f , x h IMF p The fth and hth elements in m, f,h=1,2,…,n. f,h Represents the calculation element x f In the interval q i Under the condition that h In the interval q j The probability of; i,j=1,2,…,Q.

[0083] S4.4: Steps S401 to S403 are performed for P effective related modal components, and finally μ×n×P multi-channel power quality disturbance image data is obtained.

[0084] S5: A power quality disturbance classifier is built based on a weight-sharing multi-channel separable convolutional residual network for identification and classification.

[0085] In this embodiment, S5 is specifically as follows:

[0086] S501: inputting the generated n×n×P power quality disturbance image data set into the multi-channel separable convolutional residual network, wherein the structure of each P residual network, including parameters of convolution, pooling and other layers, is consistent;

[0087] S502: Perform convolution calculation on the power quality disturbance image using a depthwise separable convolution layer to extract feature information of the image; the depthwise separable convolution layer includes two parts, namely, depthwise convolution and pointwise convolution, which greatly reduces the amount of parameters and calculations;

[0088] S503: performing batch normalization (BN) on the perturbation feature vector extracted by the depthwise separable convolutional layer to accelerate the training process and improve the stability of the model;

[0089] S504: applying a nonlinear activation function Swish function to the feature vector normalized by the BN layer to provide better gradient propagation;

[0090] S505: Use the maximum pooling layer to highlight the important parts of the power quality disturbance feature vector, while reducing the data dimension and improving the robustness of the model;

[0091] S506: Input the pooled feature data into the residual network module to further speed up the network training speed and avoid gradient disappearance. The residual network module includes the first to fourth residual layers, and the structure of each layer is the same. The first residual layer includes the first convolutional layer, the first BN layer, the first ReLU layer, the second convolutional layer, the second BN layer, the Shortcut connection unit, the summation unit, and the 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 connected in sequence. The input end of the Shortcut connection unit is connected to the input end of the first convolutional layer, and its output end is connected to the input end of the summation unit. In order to prevent overfitting during model training, a Dropout layer is added between the first to fourth residual layers, and the probability of discarding parameters is set to 0.2;

[0092] 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 category probability.

[0093] In summary, the power quality disturbance classification method based on multi-channel separable convolution proposed in this embodiment avoids the interference of redundant modes on classification by screening out the effective relevant modal components of the disturbance signal, combines the Markov transition field to generate multi-channel image data, retains the dynamic characteristics of the components, and uses a multi-channel separable convolution residual network based on weight sharing for classification and recognition, which makes up for the shortcomings of traditional single-channel recognition and improves the accuracy of classification.

[0094] The above specific embodiments are only several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant inspirations of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A method for classifying power quality disturbances based on multi-channel separable convolution, characterized in that: The following steps are involved: 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 S; S2. Using the particle swarm algorithm to optimize the decomposition scale K and penalty factor α of variational modal decomposition, decompose the power quality disturbance signal through the improved variational modal decomposition model to obtain multiple intrinsic modal components; S3, calculating the Pearson correlation coefficient and setting a threshold to screen the intrinsic modal components to obtain P effective correlated modal components; S4, using the Markov transition field to convert the screened P effective related modal components into a two-dimensional image, generating n×n×P multi-channel power quality disturbance image data; S5. A power quality disturbance classifier is built based on a multi-channel separable convolutional residual network with weight sharing for identification and classification.

2. According to claim 1, a method for classifying power quality disturbances based on multi-channel separable convolution is characterized in that: The S2 specifically includes the following steps: S201: Determine the decomposition scale K and penalty factor α of VMD based on the particle swarm algorithm, initialize the particle swarm, set the total number of particles D, the number of iterations to e, the maximum number of iterations to E, and the optimization variables to the decomposition scale K and penalty factor α; S202: Calculate the fitness value according to the current position vector of each particle of the e-th generation, and select the particle corresponding to the maximum fitness value as the global optimal particle of the e-th generation, and select the envelope entropy function as the fitness function. The envelope entropy function expression is: Among them, s i is the envelope signal of the i-th IMF component after the power quality disturbance signal S is decomposed, i = 1, 2, ..., K; K is the decomposition scale, indicating that there are K IMF components in total; g i is the probability of each modal component, H g is the envelope entropy of the disturbance signal S. The position corresponding to the minimum envelope entropy function value indicates that the decomposition effect of the disturbance signal is the best. When the algorithm reaches the maximum number of iterations, the loop ends and the optimal fitness value and particle position are output. Accordingly, the optimal decomposition scale K and penalty factor α are obtained, and the improved VMD model is obtained; S203: Decomposing the power quality disturbance signal using the improved VMD model to obtain K IMF components, {IMF k }:={IMF1, IMF2,…,IMF k }, IMF k represents the kth IMF component, k = 1, 2,…, K.

3. The method for classifying power quality disturbances based on multi-channel separable convolution according to claim 1, characterized in that: The S3 is as follows: S301: Calculate the modal component IMF obtained by improved VMD decomposition k ={x1, x2, ..., x n } and the power quality disturbance signal S = {s1, s2, ..., s n }, the Pearson correlation coefficient expression is: Among them, x i is the kth modal component IMF k The oth element of s i is the i-th element of the power quality disturbance signal S, i=1,2,…,n, and n is the total number of elements. S302: According to the formula: Solve the threshold μ for selecting effective related modal components and convert each IMF k The corresponding Pearson correlation coefficient R is compared with the threshold μ, and the one greater than the threshold μ is selected as the effective correlation modal component, and P effective correlation modal components are obtained, {IMF p }:={IMF1, IMF2,…,IMF P },IMF p represents the pth effective correlated modal component, p = 1, 2, …, P.

4. The method for classifying power quality disturbances based on multi-channel separable convolution according to claim 1, characterized in that: In said S4, the specific steps are as follows: S401: For the obtained P effective correlated modal components, each component IMF p ={x1, x2, ..., x n } is divided into Q intervals, and the state space O = {q1, q2, ..., q Q }, q i represents the i-th interval in the state space O, i = 1, 2, ..., Q, there are Q intervals in total and the amount of data in each interval is the same; q i Elements in IMF p The elements in correspond to each other; S402: For interval q i and q j , the formula for calculating the transition probability between them is: w ij =w(x t ∈q j |x t-1 ∈q i ) Among them, w i,j For interval q i and q j The transition probability of i,j=1,2,…,Q; x t-1 , x t IMF p The t-1th and tth elements in , t=1,2,…,n, w represents the calculation element x t-1 In the interval q i Under the condition that t In the interval q j The probability of; ∈ means belonging to; S403: Calculate the Markov dynamic probability transfer matrix M, the matrix dimension is n×n, and the formula is: The element m in the fth row and hth column of the matrix f,h The calculation formula is: Among them, x f , x h IMF p The fth and hth elements in m, f,h=1,2,…,n. f,h Represents the calculation element x f In the interval q i Under the condition that h In the interval q j The probability of; i, j = 1, 2, ..., Q; S404: Steps S401 to S403 are performed on P effective related modal components, and finally n×n×P multi-channel power quality disturbance image data are obtained.

5. The method for classifying power quality disturbances based on multi-channel separable convolution according to claim 1, characterized in that: The S5 specifically includes the following steps: S501: inputting the generated n×n×P power quality disturbance image data set into the multi-channel separable convolutional residual network, wherein the structure of each P residual network, including parameters of convolution, pooling and other layers, is consistent; S502: performing convolution calculation on the power quality disturbance image using a depthwise separable convolution layer to extract feature information of the image; the depthwise separable convolution layer includes two parts: depthwise convolution and pointwise convolution; S503: performing batch normalization on the disturbance feature vector extracted by the depthwise separable convolutional layer; S504: Applying a nonlinear activation function Swish function to the feature vector normalized by the BN layer; S505: using a maximum pooling layer to highlight important parts of the power quality disturbance feature vector while reducing the data dimension; S506: inputting the pooled feature data into a residual network module to further speed up the network training speed and avoid gradient disappearance, wherein the residual network module 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 summing unit, and a second ReLU layer, wherein the first convolutional layer, the first BN layer, the first ReLU layer, the second convolutional layer, the second BN layer, the summing unit, and the second ReLU layer are sequentially connected, an input end of the shortcut connection unit is connected to an input end of the first convolutional layer, and an output end thereof is connected to an input end of the summing unit; Add a Dropout layer between the first to fourth residual layers, and set the probability of dropping parameters to 0.2; S507: Classify 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 class probabilities.

Images