A power quality disturbance identification method in high noise environment

Abstract

The application discloses a power quality disturbance identification method in a high-noise environment, and is implemented according to the following steps: adding Gaussian white noise to a simulation signal to simulate a high-noise environment, and collecting power quality disturbance signal data in the environment; performing feature extraction and feature fusion on the power quality disturbance signal data in the high-noise environment to obtain an optimal feature subset; inputting the optimal feature subset into a weighted random forest model based on mutual information to obtain an optimized random forest model, and inputting power quality disturbance signal data to be identified into the optimized random forest model to identify and classify the power quality disturbance signal; and the power quality disturbance identification method solves the problems that, in the prior art, a power quality disturbance signal identification method in a high-noise environment cannot effectively and accurately identify, and a traditional random forest has the same weight of each sub-decision tree, and repeated voting leads to a decrease in identification accuracy.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power quality disturbance identification, and particularly relates to a power quality disturbance identification method in a high-noise environment. BACKGROUND

[0002] Firstly, in the actual operation of the power system, the collection of massive power quality disturbance signals will be disturbed by different degrees of noise. Among them, high noise will have a serious impact on the power quality disturbance signal, not only reducing the working efficiency of the monitoring device, but also covering the disturbance characteristics of various disturbance signals, resulting in a decrease in the accuracy of power quality disturbance signal identification. Secondly, in the existing power quality disturbance identification method, some methods need to extract more features, and the features need to be selected manually, which is easy to cause feature redundancy and lose part of the feature information, resulting in a decrease in the recognition rate. Finally, the existing power quality classification method is generally based on the classification algorithm of machine learning theory to construct a classifier. In the case of composite disturbance and different signal-to-noise ratios, overfitting phenomenon is easy to occur, and it is affected by noise. Moreover, the traditional random forest algorithm has a high similarity in state evaluation, which causes repeated voting and reduces the accuracy of the model. SUMMARY

[0003] The purpose of the present application is to provide a power quality disturbance identification method in a high-noise environment, which solves the problem that the existing power quality disturbance signal identification method cannot effectively and accurately identify in a high-noise environment, and the traditional random forest has the same weight of each sub-decision tree, which causes repeated voting and reduces the recognition accuracy.

[0004] The technical solution adopted by the present application is a power quality disturbance identification method in a high-noise environment, which is implemented according to the following steps:

[0005] Step 1: Add Gaussian white noise to the simulation signal to simulate a high-noise environment, and collect power quality disturbance signal data in the environment;

[0006] Step 2: Extract and fuse the features of the power quality disturbance signal data in the high-noise environment to obtain an optimal feature subset;

[0007] Step 3: Input the optimal feature subset into a weighted random forest model based on mutual information to obtain an optimized random forest model, and input the power quality disturbance signal data to be identified into the optimized random forest model for power quality disturbance signal identification and classification.

[0008] The present application has the following characteristics:

[0009] The specific process of adding Gaussian white noise to the simulation signal in step 1 to simulate a high-noise environment is as follows: use the awgn function to add Gaussian white noise with a size of 5-15 db to the simulation signal to form a simulated high-noise environment.

[0010] The specific process of step 2 is as follows:

[0011] Step 2.1, feature extraction is performed on the power quality disturbance signal data in a high-noise environment, and three feature extraction methods are used for feature extraction, namely: fast Fourier transform, mel frequency cepstral coefficient and time domain feature extraction, to obtain three feature subsets;

[0012] Step 2.2, fuse any two feature subsets to obtain a first fused feature subset, and fuse the fused feature subset with the third feature subset to obtain an optimal feature subset.

[0013] The specific process of step 2.1 is as follows:

[0014] The power quality disturbance signal data in a high-noise environment is subjected to fast Fourier transform to obtain a frequency spectrum X[k]; and there is:

[0015] X[k] = H[k]E[k] (1)

[0016] Where X[k] is the frequency spectrum, H[k] and E[k] are the envelope and the details of the spectrum, k = 1, 2, 3,..., n;

[0017] The feature subsets obtained after fast Fourier transform include: the maximum value F1, the minimum value F2, the mean value F3, the variance F4 of the fundamental frequency amplitude curve, the maximum 4 peak values F5, F6, F7, F8 of the frequency amplitude envelope line, and the frequencies F9, F 10 , F 11 , F 12 ; Wherein,

[0018] The maximum value F1 of the fundamental frequency amplitude curve is represented as:

[0019] F1 = max(X[k]) (2)

[0020] The minimum value F2 of the fundamental frequency amplitude curve is represented as:

[0021] F2 = min(X[k]) (3)

[0022] The mean value F3 of the fundamental frequency amplitude curve is represented as:

[0023]

[0024] The variance F4 of the fundamental frequency amplitude curve is represented as:

[0025]

[0026] In the formula, X[k] is the spectrum after Fourier transform, where i = 1, 2, 3, ..., n, k = 1, 2, 3, ..., n, N s =1,2,3,....,n, s=1,2,3,....,n;

[0027] The Mel spectrum is obtained by passing the spectrum X[k] through a set of Mel filters:

[0028] log X[k]=log(Mel-Spectrum) (6)

[0029] The energy is obtained by taking the logarithm of the absolute value of the filter bank's output.

[0030] log||X[k]||=log||H[k]||+log||E[k]|| (7)

[0031] Perform a discrete cosine transform on its energy;

[0032] The 14th to 25th coefficients of the discrete cosine transform are taken as the characteristic subset of the Mel frequency cepstral coefficients, denoted as: F 13 F 14 F 15 F 16 F 17 F 18 F 19 F 20 F 21 F 22 F 23 F 24 ;

[0033] The time-domain feature subset includes the maximum value F of power quality disturbance signal data under high-noise environments. 25 Minimum value F 26 Average value F 27 Variance F 28 Standard deviation F 29 Root amplitude F 30 Effective value F 31 Peak factor F 32 Pulse factor F 33 Margin factor F 34 Waveform factor F 35 The calculation formula is:

[0034] F 25 =max(x(i))

[0035] F 26 =min(x(i))

[0036]

[0037]

[0038]

[0039]

[0040]

[0041]

[0042]

[0043]

[0044]

[0045] Where x(i) represents the power quality disturbance signal data under high noise environment.

[0046] Step 2.2 The specific process is as follows:

[0047] Merging any two feature subsets yields the first merged feature subset, expressed as:

[0048]

[0049] Z1 is the feature subset obtained after the first relevant discriminant analysis fusion. The feature subset is obtained by fast Fourier transform and transformation. It consists of the Mel frequency cepstral coefficients and the feature subset obtained after transformation;

[0050] The optimal feature subset is obtained by fusing the fused feature subset with another feature subset, expressed as:

[0051]

[0052] Z2 is the optimal feature subset obtained after the second feature fusion. This refers to the feature subsets after the first fusion and the transformation. It consists of time-domain features and a subset of features after transformation.

[0053] Step 3 is as follows:

[0054] Step 3.1: Divide the optimal feature subset into a 6:2:2 ratio, which will be the training set, the validation set, and the test set.

[0055] Step 3.2: Use the traditional random forest algorithm to obtain k sub-decision trees from the training set and determine the initial random forest;

[0056] Step 3.3: In the validation set, mutual information is used to evaluate the correlation between random subtrees. Among the subtrees with the highest correlation, the subtree with the highest evaluation accuracy is retained to simplify the random forest.

[0057] Step 3.4: Convert the evaluation accuracy into the weights used in voting for the new random forest model to obtain the random forest model. First, determine whether the weight matrix needs to be updated. If it does, calculate and update it based on the evaluation accuracy matrix. If it does not need to be updated, directly perform weighted voting and complete the classification to obtain the optimized random forest model.

[0058] Step 3.5: Input the power quality disturbance signal data to be identified into the optimized random forest model for power quality disturbance signal identification and classification.

[0059] Step 3.3 is as follows:

[0060] The mutual information is calculated as follows:

[0061] I(h i ,h k )=I(y i ,y k (10)

[0062] Where I(h) i ,h k (k≠i) represents the decision tree h i with h k Mutual information between them, and y i (i = 1, 2, 3, ..., K) represents the output state of the i-th decision tree, and I(h i ,h k The larger the value of ), the greater the correlation between the two selected trees and the more information they share.

[0063] Where I(h) i ,h k The formula for calculating ) is:

[0064]

[0065] Where X and Y are two sets of random variables representing two subtrees, the joint probability distribution of X and Y is represented by p(x,y), and the marginal probability distributions of X and Y are represented by p(x) and p(y), respectively. H(X) is the calculated information entropy of X, specifically expressed as:

[0066]

[0067] By calculating the information entropy of any two subtrees, the subtrees whose information entropy is greater than the information entropy threshold ε are combined.

[0068] The evaluation accuracy is calculated using the following formula:

[0069] I(h i ,y)=I(y i ,y) (13)

[0070] Where I(h) i ,y) represents the decision tree h i The mutual information value between the power quality disturbance type y and the power quality disturbance type y.

[0071] The specific process for determining whether the weight matrix needs to be updated in step 3.4 is as follows: if the power quality disturbance types to be judged are the same, then no update is needed; if the power quality disturbance types have changed, or have not been judged, then an update is needed, that is, the weight matrix needs to be recalculated.

[0072] The specific process of updating the weight matrix in step 3.4 is as follows:

[0073] After simplifying the random forest, the new random forest evaluation accuracy matrix ACC is:

[0074]

[0075] Where acc QP Let P represent the evaluation accuracy of the P-th subtree for the Q-th power quality disturbance signal, where P = 1, 2, 3, ..., p, Q = 1, 2, 3, ..., q, p is the number of subtrees after the random forest is simplified, and q is the type of power quality disturbance signal to be identified;

[0076] The voting weight matrix W can be obtained based on the evaluation accuracy:

[0077]

[0078] Where ω QP The weights of the P-th decision tree for the Q-th power quality disturbance signal are given by P = 1, 2, 3, ..., p and Q = 1, 2, 3, ..., q, where p is the number of decision trees after the random forest is simplified, and q is the type of power quality disturbance signal to be identified. The weights assigned to the new random forest model during voting are calculated as follows:

[0079] ω QP =acc QP (16).

[0080] The beneficial effects of this invention are:

[0081] This invention provides a method for identifying power quality disturbances in high-noise environments. It utilizes Gaussian white noise to simulate the power quality disturbance signal under high-noise conditions. Secondly, it employs a correlation discriminant analysis method, currently not used in power quality disturbance identification. Furthermore, this invention uses Fast Fourier Transform, Mel-frequency cepstral coefficients, and time-domain features to extract features from the signal. These methods extract features from multiple perspectives—frequency, cepstral, and time—to comprehensively utilize various features, achieving complementary advantages, reducing feature redundancy, and extracting the optimal subset. Unlike other methods that simply add or multiply feature subsets, this invention first analyzes the feature subsets before fusing them. Finally, the optimal feature subset obtained through correlation discriminant analysis is input into an improved weighted random forest based on mutual information to identify and classify the power quality disturbance signal. In summary, the method proposed in this patent can effectively extract perturbation features in high-noise environments and reduce feature redundancy. On the other hand, it can reduce the duplicate voting phenomenon of traditional random forests and improve the accuracy of perturbation recognition. Attached Figure Description

[0082] Figure 1 This is a flowchart of a power quality disturbance identification method under high noise environment according to the present invention;

[0083] Figure 2(a) is a waveform diagram of voltage sag and harmonics of a power quality composite disturbance signal with high noise.

[0084] Figure 2(b) is a waveform diagram of voltage rise and harmonics of a power quality composite disturbance signal with high noise.

[0085] Figure 2(c) shows the waveforms of voltage sag and transient oscillation in a power quality composite disturbance signal with high noise.

[0086] Figure 2(d) shows the waveforms of voltage sag and transient oscillation of a power quality composite disturbance signal with high noise.

[0087] Figure 2(e) shows the waveforms of harmonics and flicker in a power quality composite disturbance signal with high noise.

[0088] Figure 2(f) shows the waveforms of transient oscillations and flicker of a power quality composite disturbance signal with high noise.

[0089] Figure 2(g) shows the waveforms of power quality composite disturbance signals with high noise interruption and transient oscillation.

[0090] Figure 2(h) is a waveform diagram of power quality composite disturbance signal interruption and harmonics with high noise.

[0091] Figure 2(i) is a waveform diagram of voltage sag and flicker of a power quality composite disturbance signal with high noise.

[0092] Figure 2(j) shows the waveforms of transient pulses and flicker of the power quality composite disturbance signal with high noise.

[0093] Figure 3 This is a flowchart of the Mel frequency cepstral coefficient feature subset extraction method of the power quality disturbance identification method under high noise environment of the present invention;

[0094] Figure 4(a) is a schematic diagram showing the correlation between the feature subset of the Fast Fourier Transform (FFT) and the feature subset of the Mel frequency cepstral coefficients when they are first fused in this invention.

[0095] Figure 4(b) is a schematic diagram showing the correlation between the fused feature subset and the temporal feature subset after the first fusion in this invention;

[0096] Figure 5(a) is a graph showing the variation in feature subset recognition accuracy using Fast Fourier Transform (FFT) in the embodiment;

[0097] Figure 5(b) is a graph showing the change in accuracy of the feature subset identification using Mel frequency cepstral coefficients in the embodiment;

[0098] Figure 5(c) is a graph showing the change in recognition accuracy using temporal feature subsets in the embodiment;

[0099] Figure 5(d) is a graph showing the change in recognition accuracy using the fused feature subset method in this invention in the embodiment. Detailed Implementation

[0100] The present invention will now be described in detail with reference to specific embodiments.

[0101] This invention provides a method for identifying power quality disturbances in high-noise environments, such as... Figure 1 As shown, please follow these steps:

[0102] Step 1: Use the awgn function to add Gaussian white noise of magnitude 5 to 15 dB to the simulation signal to form a simulated high-noise environment, and collect power quality disturbance signal data under this environment;

[0103] Power quality disturbances in high-noise environments are difficult to identify. Therefore, it is necessary to consider how to effectively extract the characteristics of power quality disturbance signals and accurately identify them in high-noise environments. In this experiment, 5-15 dB Gaussian white noise was added to simulate power quality disturbance signals in high-noise environments. The specific composite disturbance waveform is shown in Figure 2. The sampling rate of the power quality disturbance waveform was 12.8 kHz, the sampling duration was 1 second, and there were a total of 12.8k sampling points. The high-noise environment included scenarios of 5 dB, 10 dB, 15 dB, and a mixture of 5-15 dB, and included standard signals, voltage dips, voltage swells, voltage interruptions, transient pulses, etc. Eight single signals, including transient oscillation, voltage harmonics, and voltage flicker; and ten composite disturbance signals, including voltage sag + harmonics (Figure 2(a)), voltage swell + harmonics (Figure 2(b)), voltage sag + transient oscillation (Figure 2(c)), voltage swell + transient oscillation (Figure 2(d)), harmonics + flicker (Figure 2(e)), transient oscillation + flicker (Figure 2(f)), interruption + transient oscillation (Figure 2(g)), interruption + harmonics (Figure 2(h)), voltage sag + flicker (Figure 2(i)), and transient pulse + flicker (Figure 2(j)), for a total of 18 types of PQD signals;

[0104] Step 2: Perform feature extraction and feature fusion on the power quality disturbance signal data under high noise environment to obtain the optimal feature subset; the specific process is as follows:

[0105] Step 2.1: Feature extraction is performed on the power quality disturbance signal data under high noise environment. Three feature extraction methods are used: Fast Fourier Transform (FFT), Mel frequency cepstral coefficients (MEL), and time-domain feature extraction. A total of 35 feature categories and labels are obtained, as shown in Table 1.

[0106] Table 1

[0107]

[0108] Among them, F1, F2, F3, F4, F5, F6, F7, F8, F9, F 10 F 11 F 12 The features obtained by Fourier transform are, in order, the maximum value, minimum value, mean, and variance of the fundamental frequency amplitude curve after FFT, the four maximum peaks of the frequency amplitude envelope, and the frequencies corresponding to the four maximum peaks. The calculation formula is as follows:

[0109] F1 = max(X[k]) (1)

[0110] F2=min(X[k]) (2)

[0111]

[0112]

[0113] The power quality disturbance signal data under high noise environment is subjected to Fast Fourier Transform to obtain the spectrum X[k]; and the following exists:

[0114] X[k]=H[k]E[k] (5)

[0115] In the formula, X[k] is the spectrum after Fourier transform, where i = 1, 2, 3, ..., n, k = 1, 2, 3, ..., n, N s =1,2,3,....,n, s=1,2,3,....,n;

[0116] F5, F6, F7, and F8 are the four largest peak values ​​of the frequency amplitude envelope E[k], while F9 and F... 10 F 11 F 12 The frequencies corresponding to the four maximum peak values;

[0117] The feature subset obtained from the Mel frequency cepstral coefficients is represented as: F 13 F 14 F 15 F 16 F 17 F 18 F 19 F 20 F 21 F 22 F 23 F 24 After taking values ​​from the following formulas, perform a Discrete Cosine Transform (DCT) on it. Take the 14th to 25th coefficients of H[k] after the DCT as the MFCC coefficients, which is F. 13 F 14 F 15 F 16 F 17 F 18 F 19 F 20 F 21 F 22 F 23 F 24 The value of .

[0118] The Mel spectrum is obtained by passing the spectrum X[k] through a set of Mel filters:

[0119] logX[k]=log(Mel-Spectrum) (6)

[0120] The energy is obtained by taking the logarithm of the absolute value of the filter bank's output.

[0121] log||X[k]||=log||H[k]||+log||E[k]|| (7)

[0122] Perform a discrete cosine transform on its energy, and take the 14th to 25th coefficients of H[k] after the discrete cosine transform as the characteristic subset of the Mel frequency cepstral coefficients;

[0123] The time-domain feature subset includes the maximum value F of the power quality disturbance signal data under high-noise conditions. 25 Minimum value F 26 Average value F 27 Variance F 28 Standard deviation F 29 Root amplitude F 30 Effective value F 31 Peak factor F 32 Pulse factor F 33 Margin factor F 34 Waveform factor F 35 The formulas for calculating its time-domain characteristics are shown in Table 2 below:

[0124] Table 2

[0125]

[0126]

[0127] In the formula, x(i) is the original power quality disturbance signal containing high noise.

[0128] The signal features extracted in this invention are extracted from multiple perspectives, including the frequency domain, cepstral domain, and time domain, taking into account various factors.

[0129] Step 2.2: The three feature subsets obtained are fused using discriminant correlation analysis. Any two feature subsets are fused to obtain the fused feature subset after the first fusion. This fused feature subset is then fused with another feature subset to obtain the optimal feature subset. The specific process is as follows:

[0130] Merging any two feature subsets yields the first merged feature subset, expressed as:

[0131]

[0132] Z1 is the feature subset obtained after the first relevant discriminant analysis fusion. The feature subset is obtained by fast Fourier transform and transformation. It consists of the Mel frequency cepstral coefficients and the feature subset obtained after transformation;

[0133] The optimal feature subset is obtained by fusing the fused feature subset with another feature subset, expressed as:

[0134]

[0135] Z2 is the optimal feature subset obtained after the second feature fusion. This refers to the feature subsets after the first fusion and the transformation. It consists of time-domain features and a subset of features after transformation.

[0136] Step 3: Input the optimal feature subset into a weighted random forest model based on mutual information to obtain an optimized random forest model. Input the power quality disturbance signal data to be identified into the optimized random forest model for power quality disturbance signal identification and classification. The specific process is as follows:

[0137] Step 3.1: Divide the optimal feature subset into a 6:2:2 ratio, which will be the training set, the validation set, and the test set.

[0138] Step 3.2: Use the traditional random forest algorithm to obtain k sub-decision trees from the training set and determine the initial random forest;

[0139] Step 3.3: In the validation set, mutual information is used to evaluate the correlation between random subtrees. The subtree with the highest correlation is selected, and the subtree with the highest evaluation accuracy is retained, thus streamlining the random forest. The specific process is as follows:

[0140] The mutual information is calculated as follows:

[0141] I(h i ,h k )=I(y i ,y k (10)

[0142] Where I(h) i ,h k (k≠i) represents the decision tree h i with h k Mutual information between them, and y i (i = 1, 2, 3, ..., K) represents the output state of the i-th decision tree, and I(h i ,h k The larger the value of ), the greater the correlation between the two selected trees and the more information they share.

[0143] Where I(h) i ,h k The formula for calculating ) is:

[0144]

[0145] Where X and Y are two sets of random variables representing two subtrees, the joint probability distribution of X and Y is represented by p(x,y), and the marginal probability distributions of X and Y are represented by p(x) and p(y), respectively. H(X) is the calculated information entropy of X, specifically expressed as:

[0146]

[0147] By calculating the information entropy of any two subtrees, the subtrees whose information entropy is greater than the information entropy threshold ε are combined.

[0148] By calculating the mutual information value between any two subtrees, the subtrees whose values ​​are greater than the threshold ε are combined. The evaluation accuracy is then calculated using the following formula:

[0149] I(h i ,y)=I(y i ,y) (13)

[0150] Where I(h) i ,y) represents the decision tree h i The mutual information value between the power quality disturbance type y and the power quality disturbance type y. The larger this value, the stronger the correlation.

[0151] Step 3.4: Convert the evaluation accuracy into the weights used in voting for the new random forest model to obtain the random forest model. First, determine whether the weight matrix needs to be updated. If it does, calculate and update it based on the evaluation accuracy matrix. If it does not need to be updated, directly perform weighted voting and complete the classification to obtain the optimized random forest model.

[0152] The specific process for determining whether the weight matrix needs to be updated is as follows: if the types of power quality disturbances to be judged are the same, then no update is needed; if the types of power quality disturbances have changed, or have not been judged, then an update is needed, that is, the weight matrix needs to be recalculated.

[0153] The specific process of updating the weight matrix is ​​as follows:

[0154] After simplifying the random forest, the new random forest evaluation accuracy matrix ACC is:

[0155]

[0156] Where acc QPLet P represent the evaluation accuracy of the P-th subtree for the Q-th power quality disturbance signal, where P = 1, 2, 3, ..., p, Q = 1, 2, 3, ..., q, p is the number of subtrees after the random forest is simplified, and q is the type of power quality disturbance signal to be identified;

[0157] The voting weight matrix W can be obtained based on the evaluation accuracy:

[0158]

[0159] Where ω QP The weights of the P-th decision tree for the Q-th power quality disturbance signal are given by P = 1, 2, 3, ..., p and Q = 1, 2, 3, ..., q, where p is the number of decision trees after the random forest is simplified, and q is the type of power quality disturbance signal to be identified. The weights assigned to the new random forest model during voting are calculated as follows:

[0160] ω QP =acc QP (16).

[0161] Step 3.5: Input the power quality disturbance signal data to be identified into the optimized random forest model for power quality disturbance signal identification and classification.

[0162] Example

[0163] In the simulation experiment, since the selection of a new random forest decision tree needs to be determined by a threshold ε, and the threshold ε needs to be determined experimentally, this invention conducts perturbation recognition and classification experiments on different threshold values ​​by inputting the same dataset under a high-noise environment. The classification accuracy under different thresholds is shown in Table 3.

[0164] Table 3

[0165]

[0166] As shown in Table 3, the new random forest model proposed in this invention has the highest recognition accuracy when the threshold is 0.85. Therefore, when classifying the model, the threshold should be set to 0.85 to achieve the best classification effect.

[0167] To analyze the higher recognition accuracy of the fusion feature method proposed in this invention compared to single features in high-noise environments, this invention inputs each type of single feature and mixed feature into the power quality recognition method in high-noise environments under single and mixed noise conditions of 5 to 15 dB. The recognition accuracy under different subsets can be obtained, and the accuracy under different feature subsets is shown in Table 4:

[0168] Table 4

[0169]

[0170] As shown in Table 4, after inputting the optimal feature subset and the single feature subset into the same improved random forest based on mutual information proposed in this invention, it can be seen that the classification effect of the subset obtained by the feature fusion method proposed in this invention is better than that of other methods. It has a significant improvement compared to various single features, especially for simple time-domain features, the highest accuracy improvement can reach 4.77%.

[0171] To analyze the correlation distribution of various features during feature fusion using three methods—Fast Fourier Transform, Mel frequency cepstral coefficients, and time-domain features—this invention analyzes the correlation of each type of feature in each feature subset during sub-feature fusion, as shown in Figures 4(a) and 4(b):

[0172] Figures 4(a) and 4(b) show the correlation distribution of each feature in feature fusion using the three methods of Fast Fourier Transform, Mel frequency cepstral coefficients, and time-domain features. The importance of each type of feature in the fusion process can be clearly seen.

[0173] To analyze the accuracy of the weighted random forest based on mutual information used in this invention for identifying power quality disturbance signals in high-noise environments, this invention compares its performance with several commonly used disturbance identification methods. The classification accuracy under different classification algorithms is shown in Table 5.

[0174] Table 5

[0175]

[0176] *DCA-RF is the algorithm proposed in this invention.

[0177] As can be seen from the data in Table 5, the improved random forest algorithm adopted in this invention has a higher overall classification accuracy than the traditional random forest algorithm, and it is also a significant improvement over the current algorithm, with a recognition rate that is basically stable at over 99%. This is of great significance for the identification of power quality disturbance signals in high-noise environments.

[0178] To analyze the ability of this invention to identify single perturbation features, experiments were conducted to measure the accuracy of single perturbation feature and fused feature recognition. The experimental results are shown in Figures 5(a), 5(b), 5(c), and 5(d).

[0179] As can be seen from the comparative analysis of Figures 5(a), 5(b), 5(c), and 5(d), when identifying single perturbation features and fused feature subsets, the method of the present invention can effectively extract the correlation between features and perform analysis and identification, and its accuracy is also the highest.

[0180] Experiments have demonstrated that the proposed power quality disturbance identification method in high-noise environments exhibits better classification performance compared to other power quality disturbance identification methods. It also takes into account high-noise interference and demonstrates good adaptability. Furthermore, it employs the DCA method, currently not used in the field of power quality disturbance identification, to fuse feature subsets obtained through FFT, Mel-frequency cepstral coefficients, and time-domain features. This effectively integrates multiple features, achieving complementary advantages and reducing feature redundancy, thus promoting the application of power quality disturbance identification technology in high-noise scenarios.

Claims

1. A method for identifying power quality disturbances in a high-noise environment, characterized in that, The specific steps are as follows: Step 1: Add Gaussian white noise to the simulation signal to simulate a high-noise environment, and collect power quality disturbance signal data under this environment; Step 2: Extract and fuse features from power quality disturbance signal data under high noise environment to obtain the optimal feature subset; Step 2 is as follows: Step 2.1: Extract features from the power quality disturbance signal data under high noise environment. Three feature extraction methods are used: Fast Fourier Transform, Mel frequency cepstral coefficients, and time domain feature extraction, resulting in three feature subsets. Step 2.1 involves performing a Fast Fourier Transform (FFT) on the power quality disturbance signal data under high-noise conditions to obtain the spectrum. X [ k ]; The feature subset obtained after the Fast Fourier Transform includes: the maximum value of the fundamental frequency amplitude curve. F 1. Minimum value F 2. Mean F 3. Variance F 4. The four maximum peak values ​​of the frequency amplitude envelope F 5. F 6. F 7. F 8 and the frequencies corresponding to the four largest peaks F 9. F 10 , F 11 , F 12 ; Mel frequency cepstral coefficients are: (The spectrum is...) X [ k The Mel spectrum is obtained by using a set of Mel filters, and the energy is obtained by taking the logarithm of the absolute value of the filter bank's output. A discrete cosine transform (DCT) is then performed on this energy, and the 14th to 25th coefficients of the DCT are taken as the characteristic subset of the Mel frequency cepstral coefficients, expressed as: , , , , , , , , , , , ; The time-domain feature subset includes the maximum values ​​of power quality disturbance signal data under high-noise environments. F 25 Minimum value F 26 ,average value F 27 ,variance F 28 Standard deviation F 29 Root amplitude F 30 Valid value F 31 Peak factor F 32 Pulse factor F 33 Margin factor F 34 Waveform factor F 35 ; Step 3: Input the optimal feature subset into the weighted random forest model based on mutual information to obtain the optimized random forest model. Input the power quality disturbance signal data to be identified into the optimized random forest model to identify and classify the power quality disturbance signal.

2. The method for identifying power quality disturbances in a high-noise environment according to claim 1, characterized in that, The specific process of adding Gaussian white noise to the simulation signal in step 1 to simulate a high-noise environment is as follows: use the awgn function to add Gaussian white noise of magnitude 5~15dB to the simulation signal to form a simulated high-noise environment.

3. The method for identifying power quality disturbances in a high-noise environment according to claim 1, characterized in that, Step 2 further includes step 2.2; Step 2.2 specifically involves: fusing any two feature subsets to obtain the first fused feature subset, and then fusing the fused feature subset with the third feature subset to obtain the optimal feature subset.

4. The method for identifying power quality disturbances in a high-noise environment according to claim 3, characterized in that, Step 2.1 The specific process is as follows: The spectrum is obtained by performing a Fast Fourier Transform on the power quality disturbance signal data under high noise environment. X [ k ]; and it exists: （1） in For the spectrum, For details of the envelope and spectrum, ; The feature subset obtained after the Fast Fourier Transform includes: the maximum value of the fundamental frequency amplitude curve. F 1. Minimum value F 2. Mean F 3. Variance F 4. The four maximum peak values ​​of the frequency amplitude envelope F 5. F 6. F 7. F 8 and the frequencies corresponding to the four largest peaks F 9. F 10 , F 11 , F 12 ;in, The maximum value of the fundamental frequency amplitude curve F 1 is represented as: （2） Minimum value of the fundamental frequency amplitude curve F 2 is represented as: （3） Mean of the fundamental frequency amplitude curve F 3 is represented as: （4） Variance of the fundamental frequency amplitude curve F 4 is represented as: （5） In the formula The spectrum is the result of the Fourier transform, where , , , ; Spectrum X [ k The Mel spectrum is obtained by passing a set of Mel filters: （6） The energy is obtained by taking the logarithm of the absolute value of the filter bank's output. （7） Perform a discrete cosine transform on its energy; The 14th to 25th coefficients of the discrete cosine transform are taken as the characteristic subset of the Mel frequency cepstral coefficients, expressed as: , , , , , , , , , , , ; The time-domain feature subset includes the maximum values ​​of power quality disturbance signal data under high-noise environments. F 25 Minimum value F 26 ,average value F 27 ,variance F 28 Standard deviation F 29 Root amplitude F 30 Valid value F 31 Peak factor F 32 Pulse factor F 33 Margin factor F 34 Waveform factor F 35 The calculation formula is: ； in, This refers to power quality disturbance signal data under high noise conditions.

5. The method for identifying power quality disturbances in a high-noise environment according to claim 1, characterized in that, Step 2.2 The specific process is as follows: Merging any two feature subsets yields the first merged feature subset, expressed as: （8） in This is the feature subset obtained after the first relevant discriminative analysis fusion. The feature subset is obtained by fast Fourier transform and transformation. It consists of the Mel frequency cepstral coefficients and the feature subset obtained after transformation; The optimal feature subset is obtained by fusing the fused feature subset with another feature subset, expressed as: （9） This is the optimal feature subset obtained after the second feature fusion. This refers to the feature subsets after the first fusion and the transformation. It consists of time-domain features and a subset of features after transformation.

6. The method for identifying power quality disturbances in a high-noise environment according to claim 1, characterized in that, Step 3 is as follows: Step 3.1: Divide the optimal feature subset into a 6:2:2 ratio, which will be the training set, the validation set, and the test set. Step 3.2: Use the traditional random forest algorithm to obtain k sub-decision trees from the training set and determine the initial random forest; Step 3.3: In the validation set, mutual information is used to evaluate the correlation between random subtrees. Among the subtrees with the highest correlation, the subtree with the highest evaluation accuracy is retained to simplify the random forest. Step 3.4: Convert the evaluation accuracy into the weights used in voting for the new random forest model to obtain the random forest model. First, determine whether the weight matrix needs to be updated. If it does, calculate and update it based on the evaluation accuracy matrix. If it does not need to be updated, directly perform weighted voting and complete the classification to obtain the optimized random forest model. Step 3.5: Input the power quality disturbance signal data to be identified into the optimized random forest model for power quality disturbance signal identification and classification.

7. The method for identifying power quality disturbances in a high-noise environment according to claim 1, characterized in that, Step 3.3 is as follows: The mutual information is calculated as follows: (10) in Representation of decision tree and Mutual information between them, and For the first The output state of each decision tree, and The larger the value, the greater the correlation between the two selected trees and the more information they share. in The calculation formula is: （11） in Two sets of random variables represent two subtrees. The joint probability distribution is expressed as ,and The marginal probability distribution is expressed as , yes The calculated information entropy is expressed as follows: （12） The information entropy obtained by calculating the information entropy of any two subtrees mentioned above is greater than the information entropy threshold. Combine the subtrees; The evaluation accuracy is calculated using the following formula: (13) in Representation of decision tree Types of power quality disturbances The mutual information value.

8. The method for identifying power quality disturbances in a high-noise environment according to claim 1, characterized in that, The specific process for determining whether the weight matrix needs to be updated in step 3.4 is as follows: if the power quality disturbance types to be judged are the same, then no update is needed; if the power quality disturbance types have changed, or have not been judged, then an update is needed, that is, the weight matrix needs to be recalculated.

9. The method for identifying power quality disturbances in a high-noise environment according to claim 1, characterized in that, The specific process of updating the weight matrix in step 3.4 is as follows: After simplifying the random forest, the new random forest evaluation accuracy matrix ACC is: (14) in Indicates the first Trees to the first The evaluation accuracy of various power quality disturbance signals, among which , , This represents the number of subtrees after the random forest has been streamlined. The types of power quality disturbance signals to be identified; The weight matrix of the votes can be obtained based on the evaluation accuracy. : (15) in No. decision trees for the first The weights of various power quality disturbance signals, among which , , This represents the number of decision trees after streamlining the random forest. The weighting method for the new random forest model in voting for the types of power quality disturbance signals to be identified is as follows: (16)。

Images