A method for identifying power quality disturbance signals based on machine learning

Abstract

The present invention provides a method for identifying power quality disturbance signals based on machine learning, including S1, applying DTCWT-LPP to the original disturbance signal, extracting corresponding feature quantities and performing dimensionality reduction processing; S2, obtaining corresponding feature quantities such as standard deviation, mean, maximum value, minimum value, etc. after the original disturbance signal is processed by DTCWT-LPP, inputting it as an input sample into local density clustering for classification and identification, and establishing a corresponding classification model for power quality disturbance identification; S3, when new data is input, classifying the power quality according to the power quality classification model after DTCWT-LPP transformation. The present invention takes active distribution network as the research background, effectively integrates DTCWT and LPP algorithms, uses DTCWT-LPP to extract the characteristics of various voltage quality disturbances and performs dimensionality reduction processing on the feature quantities, adopts a multi-level local density clustering algorithm to construct a classification model, classifies and identifies the power quality disturbance source, thereby solving problems such as power quality disturbance.

Description

Technical Field

[0001] The present invention belongs to the technical field of power distribution networks, and in particular relates to a method for identifying power quality disturbance signals based on machine learning. Background Art

[0002] As an important energy source for human production and life, the quality of electric energy has a huge impact on society. With the widespread use of power electronic devices and the continuous increase of nonlinear loads, the disturbance of power quality has further increased, bringing about a series of power quality problems. The effective management of power quality is particularly important and urgent, and the primary task of management is to effectively identify the source of power quality disturbance. Therefore, the effective identification of power quality problems will have important research significance.

[0003] So far, there have been many studies on power quality. The traditional research content and methods are to use relevant information such as voltage or current to construct the power quality waveform, and then perform mathematical transformation to obtain the power quality characteristic value. Feature extraction methods include adaptive transformation, short-time Fourier transform, inherent time scale decomposition algorithm, AdaBoost algorithm, HilbertHuang transform, etc. However, these methods have limitations such as lack of phase information and poor directionality. Machine learning algorithms include convolutional neural network algorithms, deep learning, support vector machines, etc. However, machine learning algorithms not only require a large amount of power quality data, but also have certain limitations.

[0004] Therefore, it is necessary to provide a method for identifying power quality disturbance signals based on machine learning. Taking the active distribution network as the research background, the DTCWT and LPP algorithms are effectively integrated, and the DTCWT-LPP is used to extract the characteristics of various voltage quality disturbances and reduce the dimension of the feature quantities to solve problems such as power quality disturbances. Summary of the invention

[0005] The purpose of the present invention is to make up for the shortcomings of the prior art, and to provide a new method for identifying power quality disturbance signals based on machine learning in order to solve the problems of increasingly serious power quality pollution and power quality disturbance caused by the access of a high proportion of new energy and the extensive use of nonlinear loads and power electronic components.

[0006] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0007] A method for identifying power quality disturbance signals based on machine learning comprises the following steps;

[0008] Step S1, applying DTCWT-LPP to the original disturbance signal, extracting the corresponding feature quantity and performing dimensionality reduction processing;

[0009] S11: Apply DTCWT decomposition to the original disturbance signal so that the real part and the imaginary part of the original disturbance signal meet the requirements of the Hilbert transform pair;

[0010] S12: Accurately extract the corresponding feature quantity, and apply LPP to reduce the dimension of the extracted feature quantity;

[0011] Step S2: After the original disturbance signal is processed by DTCWT-LPP, the corresponding standard deviation, mean, maximum value, minimum value and other characteristic quantities are obtained, which are input as input samples into the local density clustering for classification and identification, and the corresponding classification model for power quality disturbance identification is established;

[0012] Step S3: When new data is input, the power quality is classified according to the power quality classification model after DTCWT-LPP transformation.

[0013] Furthermore, in step S11, DTCWT is applied to the original disturbance signal to decompose the real part and the imaginary part of the original disturbance signal so that the real part and the imaginary part of the original disturbance signal meet the requirements of the Hilbert transform pair, which specifically includes the following steps:

[0014] S111, dividing the original disturbance signal into two parts, a real number and an imaginary number;

[0015] ψ(t)=ψ h (t)+iψ g (t) (1)

[0016] Among them, ψ h (t) is the real part of the original disturbance signal, ψ g (t) is the imaginary part of the original disturbance signal;

[0017] S112. According to wavelet transform theory, the wavelet coefficients of the real part are expressed as:

[0018]

[0019] The scale coefficient of the real part can be expressed as:

[0020]

[0021] The wavelet coefficients of the imaginary part can be expressed as:

[0022]

[0023] The scale factor of the imaginary part can be expressed as:

[0024]

[0025] The wavelet coefficients are obtained as follows:

[0026]

[0027] The scale factor is:

[0028]

[0029] S113, reconstructing the wavelet coefficients in the time domain obtained by wavelet transform:

[0030]

[0031] The scale coefficients in the time domain are obtained through wavelet transform and reconstructed:

[0032]

[0033] Furthermore, in S12, the corresponding feature quantities are accurately extracted, and LPP is applied to reduce the dimension of the extracted feature quantities, which specifically includes the following steps:

[0034] S121. Reconstruct the detailed feature vector of the decomposed original disturbance signal through the inverse transformation of DTCWT:

[0035]

[0036] S122, the obtained data is processed by LPP to reduce the dimension, and the sample point x is calculated. i =[x i1 ,x i2 ,…,x iD ] and x j =[x j1 ,x j2 ,…,x jD ]’s standard Euclidean distance:

[0037]

[0038] Where D is the data dimension, s is the component variance;

[0039] S123. Set the neighborhood value k0 of the initial sample, make further improvements to the standard Euclidean distance, and construct a neighborhood graph for each sample:

[0040]

[0041] Among them, m(x i ) and m(x j ) are sample points x i and x j The mean distance to other sample points;

[0042] S124, based on the neighborhood of each sample obtained in S123, calculate the sample x i and xj The weight value S ij :

[0043]

[0044] in

[0045]

[0046] Where σ is the Gaussian kernel function, which is taken as 1 here, and α ij ∈(0,1),α ij +1∈(1,2),α ij +2∈(2,3), so (α ij +1)α ij ∈(0,2),(α ij +2)α ij ∈(0,3), and α ij <(α ij +1)α ij <(α ij +2)α ij ;;

[0047] S125, weight matrix S obtained based on S124 ij , calculate the optimal transformation matrix W by the Lagrange multiplier method:

[0048] XL T W=λXDX T W (15)

[0049] Where L = DS, X is the training sample obtained by DTCWT transformation of the original input data;

[0050] S126. Calculate the low-dimensional data Y of the training sample:

[0051] Y=W T X (16).

[0052] Furthermore, in S2, the original disturbance signal is processed by DTCWT-LPP to obtain corresponding characteristic quantities such as standard deviation, mean, maximum value, minimum value, etc., which are input as input samples into local density clustering for classification and identification, and a corresponding classification model for power quality disturbance identification is established, which specifically includes the following steps:

[0053] S21, determine the cutoff distance dc, calculate the Euclidean distance between the sample points, and determine the corresponding cutoff distance based on the size of the sample points, which is mainly determined to be 2% of the minimum distance between the sample points;

[0054] S22. Calculate the local density of the sample points:

[0055]

[0056] Among them, d ij is the distance from sample point i to j, I s is the set of clustering sample points;

[0057] S23. The distance deviation formula of clustered sample points is as follows:

[0058]

[0059] S24, calculate the cluster center index γ based on the local density and distance deviation of each sample point i , determine the cluster center and the corresponding cluster, the calculation formula is as follows:

[0060] γ i =ρ i ×δ i (19)

[0061] S25. Repeat the above steps with the clustering results as new samples, perform re-clustering, effectively distinguish the power quality disturbance signals, analyze the relationship between the clustering results and the actual power quality problems, and establish a classification model for power quality disturbance identification.

[0062] The embodiments of the present invention bring the following beneficial effects:

[0063] (1) The present invention adopts DTCWT-LPP. DTCWT is a wavelet transform with the characteristics of anti-aliasing, approximate translation invariance, complete reconstruction, multi-directional selectivity, etc. The detail component obtained by DTCWT contains sufficient information and can be used as an effective eigenvalue for judging the disturbance signal. In addition, frequency aliasing and information loss are less likely to occur, pseudo-frequency components are reduced, and it has outstanding anti-aliasing characteristics. LPP can not only improve the dimensionality reduction performance, but also clear the redundant information between the projection components, and effectively reflect the structure of high-dimensional data.

[0064] (2) Based on the state features extracted by DTCWT-LPP, a multi-level local density clustering algorithm is used to construct a classification model to classify and identify the power quality disturbance sources. The center point parameter of the LDC clustering algorithm takes into account both the clustering sample density and the distance deviation of the sample points, which can effectively improve the effectiveness of the clustering results. The entire clustering process is simple and linear, and can be completed in one clustering. The multi-level local density clustering algorithm is simple, has high recognition accuracy, and fast computing speed, and provides a new method for the classification of power quality disturbance signals.

[0065] Other features and advantages of the present invention will be described in the following description, and some features will become apparent from the description or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0066] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 A flow chart of a method for identifying a power quality disturbance signal provided by an embodiment of the present invention;

[0068] Figure 2 A schematic diagram of the DTCWT decomposition process provided by an embodiment of the present invention;

[0069] Figure 3 A flow chart of the DTCWT-LPP-LDC algorithm provided by an embodiment of the present invention;

[0070] Figure 4 This is a power quality classification model after multi-level LDC clustering provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0071] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0072] The embodiment of the present invention provides a method for identifying power quality disturbance signals based on machine learning, which uses dual tree complex wavelet-locality preserving projection transform (DTCWT-LPP) to accurately extract feature quantities from power system transient signals and reduce the dimension of the extracted feature quantities; then the feature quantities after dimension reduction are input as input samples into local density clustering (LDC) for classification and identification, and a classification model for power quality disturbance identification is established. When new data is input, classification and identification are directly performed, including the following steps:

[0073] S1. Apply DTCWT-LPP to the original disturbance signal to accurately extract the corresponding feature quantity and perform dimensionality reduction processing;

[0074] S11, applying DTCWT decomposition to the original disturbance signal, so that the real part and the imaginary part of the original disturbance signal meet the requirements of the Hilbert transform pair; Figure 1As shown, the DTCWT decomposition process of this embodiment includes two parallel wavelet trees, namely: one is a real part tree corresponding to the real part of the complex wavelet, and the other is an imaginary part tree corresponding to the imaginary part of the complex wavelet;

[0075] S111, dividing the original disturbance signal into two parts, a real number and an imaginary number;

[0076] ψ(t)=ψ h (t)+iψ g (t) (1)

[0077] Among them, ψ h (t) is the real part of the original disturbance signal, ψ g (t) is the imaginary part of the original disturbance signal;

[0078] S112. According to wavelet transform theory, the wavelet coefficients of the real part are expressed as:

[0079]

[0080] The scale coefficient of the real part can be expressed as:

[0081]

[0082] The wavelet coefficients of the imaginary part can be expressed as:

[0083]

[0084] The scale factor of the imaginary part can be expressed as:

[0085]

[0086] The wavelet coefficients are obtained as follows:

[0087]

[0088] The scale factor is:

[0089]

[0090] S113, according to formula (8), the wavelet coefficients in the time domain obtained by wavelet transform of formula (2) and formula (4) are reconstructed:

[0091]

[0092] According to formula (9), the scale coefficients in the time domain obtained by wavelet transform in formula (3) and formula (5) are reconstructed:

[0093]

[0094] S12, accurately extracting corresponding feature quantities, and applying LPP to reduce the dimension of the extracted feature quantities;

[0095] S121. Reconstruct the detailed feature vector of the decomposed original disturbance signal through the inverse transformation of DTCWT:

[0096]

[0097] The above process reduces the loss of characteristic parameter information during the sampling process, can effectively reduce signal frequency aliasing, and enhance the accuracy of the power quality disturbance identification and classification model.

[0098] S122, the obtained data is processed by LPP to reduce the dimension, and according to formula (11), the sample point x is calculated i =[x i1 ,x i2 ,…,x iD ] and x j =[x j1 ,x j2 ,…,x jD ]’s standard Euclidean distance:

[0099]

[0100] Where D is the data dimension and s is the component variance.

[0101] S123. Set the neighborhood value k0 of the initial sample, and further improve the standard Euclidean distance according to formula (12) to construct a neighborhood graph for each sample:

[0102]

[0103] Where m(x i ) and m(x j ) are sample points x i and x j The improved distance can make the overall distribution of samples more uniform, thereby reducing the impact of the k value on the overall dimensionality reduction effect of the data.

[0104] S124. After obtaining the neighborhood of each sample through S123, if N ε (x i ) is x i Neighborhood, N ε (x j ) is x j The neighborhood of , the sample x can be calculated by formula (13) i and x j The weight value S ij :

[0105]

[0106] where α ij By calculating formula (14), we can get:

[0107]

[0108] The σ in the formula is the Gaussian kernel function, which is taken as 1 here. And α ij ∈(0,1),α ij +1∈(1,2),α ij +2∈(2,3), so (α ij +1)α ij ∈(0,2),(α ij +2)α ij ∈(0,3), and α ij <(α ij +1)α ij <(α ij +2)α ij , while ensuring that the weights are within a certain range, a good division is made for the sample points in the neighborhood.

[0109] S125, weight matrix S obtained according to step S124 ij , through the Lagrange multiplier method, the optimal transformation matrix W can be solved:

[0110] XL T W=λXDX T W (15)

[0111] Where L = DS, X is the training sample obtained by DTCWT transformation of the original input data.

[0112] S126. According to formula (16), obtain the low-dimensional data Y of the training sample:

[0113] Y=W T X (16)

[0114] S2, such as Figure 2 As shown in the figure, the original disturbance signal is processed by DTCWT-LPP to obtain the corresponding characteristic quantities such as standard deviation, mean, maximum value, minimum value, etc., which are input as input samples into the local density clustering (LDC) for classification and identification, and the corresponding classification model for power quality disturbance identification is established;

[0115] S21, determine the cutoff distance dc, calculate the Euclidean distance between the sample points, and determine the corresponding cutoff distance in combination with the size of the sample points, which is mainly determined to be 2% of the minimum distance between the sample points.

[0116] When there are fewer sampling points, you can adjust the corresponding cutoff distance yourself.

[0117] S22. Calculate the local density of the sample points. The calculation formula is as follows:

[0118]

[0119] Among them, dij is the distance (Euclidean distance) from sample point i to j, and Is is the cluster sample point set.

[0120] S23. The distance deviation formula of clustered sample points is as follows:

[0121]

[0122] From the distance deviation of sample point i, it can be seen that the local density is greater than the point closest to the center of the sample point. In general, the greater the density of the sample point, the greater the distance offset of the sample point. The distance deviation of each sample point is calculated.

[0123] S24, calculate the cluster center index γ based on the local density and distance deviation of each sample point i , determine the cluster center and the corresponding cluster, the calculation formula is as follows:

[0124] γ i =ρ i ×δ i (19)

[0125] The center point parameter of the LDC clustering algorithm takes into account both the clustering sample density and the distance deviation of the sample points, which can effectively improve the effectiveness of the clustering results, thereby further enhancing the accuracy of the power quality disturbance identification and classification model.

[0126] S25. Repeat the above steps with the clustering results as new samples, perform re-clustering, effectively distinguish the power quality disturbance signals, analyze the relationship between the clustering results and the actual power quality problems, and establish a classification model for power quality disturbance identification.

[0127] S3, such as Figure 3 As shown in the figure, the power quality classification model obtained after multi-level LDC clustering, F is the cluster label. When there is a power quality sample input, after the characteristic value of the corresponding power quality is obtained by DTCWT-LPP decomposition, the power quality is classified according to the power quality classification model after DTCWT-LPP transformation, and the final classification result of the power quality interference source is found by using the transitivity of the LDC clustering algorithm.

[0128] The corresponding relationship between the classification model and the interference classification in this embodiment is shown in Table 1 below:

[0129] Disturbance Type Perturbation Label Cluster labels Voltage sag C1 F1-1 Voltage swell C2 F1-2 Voltage interruption C3 F1-3 Voltage flicker C4 F2-1 Transient Oscillation C4 F2-2 harmonic C5 F3-1 Harmonics and voltage sags C6 F3-2 Harmonics and voltage swells C7 F3-3

[0130] Table 1 Correspondence between classification models and interference classification

[0131] It should be noted that, in this article, the terms "comprises", "includes" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or apparatus.

[0132] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying power quality disturbance signals based on machine learning, characterized in that: The following steps are included: Step S1, applying DTCWT-LPP to the original disturbance signal, extracting the corresponding feature quantity and performing dimensionality reduction processing; Step S11, applying DTCWT decomposition to the original disturbance signal, so that the real part and the imaginary part of the original disturbance signal meet the requirements of the Hilbert transform pair; Step S12, extracting corresponding feature quantities, and applying LPP to reduce the dimension of the extracted feature quantities; Step S2: based on the disturbance signal after dimensionality reduction processing, calculate the corresponding local density, distance deviation, cluster center index and other characteristic quantities, input them as input samples into the local density clustering for classification and identification, and establish the corresponding classification model for power quality disturbance identification; Step S3, based on the input of new input sample data, classify the power quality according to the power quality classification model after DTCWT-LPP transformation; In step S11, DTCWT is applied to the original disturbance signal to decompose the real part and the imaginary part of the original disturbance signal so that the real part and the imaginary part of the original disturbance signal meet the requirements of the Hilbert transform pair, which specifically includes the following steps: S111, dividing the original disturbance signal into two parts, a real number and an imaginary number; ψ(t)=ψ h (t)+iψ g (t) (1) Among them, ψ h (t) is the real part of the original disturbance signal, ψ g (t) is the imaginary part of the original disturbance signal; S112. According to the wavelet transform theory, the wavelet coefficients of the real part are calculated as: Calculate the scale factor of the real part as: Calculate the wavelet coefficient of the imaginary part as: Calculate the scale factor of the imaginary part as: The wavelet coefficients of the original disturbance signal are obtained as follows: The scale factor of the original disturbance signal is: S113, reconstructing the wavelet coefficients in the time domain obtained by wavelet transform: Reconstruct the scale coefficients in the time domain obtained by wavelet transform: In S12, the corresponding feature quantity is extracted, and LPP is applied to reduce the dimension of the extracted feature quantity, which specifically includes the following steps: S121. Reconstruct the detailed feature vector of the decomposed original disturbance signal through the inverse transformation of DTCWT: S122, the obtained data is processed by LPP to reduce the dimension, and the sample point x is calculated. i =[x i1 ,x i2 ,…,x iD ] and x j =[x j1 ,x j2 ,…,x jD ]’s standard Euclidean distance: Where D is the data dimension and S is the component variance; S123. Set the neighborhood value k0 of the initial sample, improve the standard Euclidean distance, and construct a neighborhood graph for each sample: Among them, m(x i ) and m(x j ) are sample points x i and x j The mean distance to other sample points; S124, based on the neighborhood of each sample obtained in S123, calculate the sample x i and x j The weight value S ij : in Where σ is the Gaussian kernel function, which is taken as 1 here, and α ij ∈(0,1),α ij +1∈(1,2),α ij +2∈(2,3), so (α ij +1)α ij ∈(0,2),(α ij +2)α ij ∈(0,3), and α ij <(α ij +1)α ij <(α ij +2)α ij ; S125, weight matrix S obtained based on S124 ij , calculate the optimal transformation matrix W by the Lagrange multiplier method: XLX T W=λXDX T W (15) Where L = DS, X is the training sample obtained by DTCWT transformation of the original input data; S126. Calculate the low-dimensional data Y of the training sample: Y=W T X。 (16) 2. The method for identifying power quality disturbance signals according to claim 1, characterized in that: In S2, based on the disturbance signal after dimensionality reduction processing, the corresponding local density, distance deviation, cluster center index and other characteristic quantities are calculated, and the characteristic quantities are input into the local density clustering as input samples for classification and identification, and the corresponding classification model for power quality disturbance identification is established, which specifically includes the following steps: S21. Determine the cutoff distance d c , calculate the Euclidean distance between sample points; wherein the Euclidean distance is 2% of the minimum distance between sample points; S22. Calculate the local density of the sample points: Among them, d ij is the Euclidean distance from sample point i to j, I s is the set of clustering sample points; S23. Calculate the distance deviation of cluster sample points: S24, calculate the cluster center index γ based on the local density and distance deviation of each sample point i , determine the cluster center and the corresponding cluster, the calculation formula is as follows: c i =ρ i ×d i (19) S25. Repeat the above steps with the clustering results as new samples, perform re-clustering, effectively distinguish the power quality disturbance signals, analyze the relationship between the clustering results and the actual power quality problems, and establish a classification model for power quality disturbance identification.

3. The method for identifying power quality disturbance signals according to claim 2, characterized in that: The greater the density of the sample points, the greater the distance offset of the sample points.

Images