A power quality disturbance identification method using feature selection

Abstract

The application discloses a power quality disturbance identification method adopting feature selection, and aims at the situation that in most current power quality disturbance identification methods, the frequency domain feature analysis method is adopted more, the time-frequency analysis method needs to be calculated through multiple analysis sometimes, which can cause the speed to be reduced when the feature extraction is carried out, and the feature of the power quality disturbance signal cannot be extracted quickly, proposes a method combining the feature score and the genetic algorithm, extracts the optimal subset, reduces the operation speed, and adopts the adaptive IPSO-BP algorithm based on the dynamic weight, further optimizes the classification efficiency, and finally through experiments, it is known that the IPSO-BP power quality disturbance identification method based on the feature screening can identify the power quality composite disturbance more quickly and more accurately than the prior art.

Description

Technical Field

[0001] This invention belongs to the field of power quality disturbance identification technology, and specifically relates to a power quality disturbance identification method using feature selection. Background Technology

[0002] In today's society, the power system is constantly developing, and the equipment it contains is becoming increasingly diverse and complex. This trend makes it more difficult to determine the overall operating status of the system, and also slows down the overall identification rate. Power quality disturbances, as a manifestation of the operating status of the power system, need to be analyzed and determined in a timely manner.

[0003] However, most current power quality disturbance identification methods primarily employ frequency domain analysis, with fewer methods relying solely on time domain analysis. Frequency domain methods mainly involve frequency domain analysis of the power quality disturbance signal, lacking methods for calculating time domain features. Time domain features require extensive feature extraction during analysis, leading to excessively high dimensionality of the feature subsets to be processed during classification. This, in turn, increases the structural complexity of the classifier, makes it difficult to optimize model parameters, and ultimately reduces classification accuracy and recognition efficiency. Summary of the Invention

[0004] The purpose of this invention is to provide a power quality disturbance identification method using feature selection, which solves the problem in the prior art that the high dimensionality of the feature subset during classification leads to increased structural complexity of the classifier, difficulty in optimizing model parameters, and ultimately reduced classification accuracy and recognition efficiency.

[0005] The technical solution adopted in this invention is: a power quality disturbance identification method using feature selection, specifically implemented according to the following steps:

[0006] Step 1: Add Gaussian white noise to the power quality signal to obtain the power quality composite disturbance signal;

[0007] Step 2: Extract features from composite power quality disturbance signals;

[0008] Step 3: Input each feature into the support vector machine for recognition, obtain the classification accuracy of each feature, and sort the classification accuracy of each feature by the support vector machine and assign scores accordingly.

[0009] Step 4: Use a multi-population genetic algorithm to perform... The round of evolutionary computation seeks the optimal feature subset, and the optimal feature subset is used as the sample;

[0010] Step 5: Construct an improved particle swarm optimization-multilayer feedforward neural network model based on feature selection, and initialize the parameters of the improved particle swarm optimization-multilayer feedforward neural network model based on feature selection.

[0011] Step 6: Use the network parameter x from the first forward propagation of the neural network as the position of the initialized particle;

[0012] Step 7: Input the previously obtained samples and particles, update the particle states, and calculate the fitness of each particle;

[0013] Step 8: Adaptively adjust the inertia factor according to the particle's fitness.

[0014] Step 9: Update the individual extreme values ​​and the global extreme values ​​based on the fitness of each particle, and output the results to determine the global extreme value point.

[0015] Step 10: Update the dynamic adjustment coefficients and output the optimal parameters;

[0016] Step 11: Train and identify the model with the optimal feature subset as input to the optimal parameter to obtain the optimized particle swarm optimization-multilayer feedforward neural network model based on feature selection.

[0017] Step 12: Input the signal to be identified into the optimized feature-based improved particle swarm optimization-multilayer feedforward neural network model and output the classification result.

[0018] The invention is further characterized by:

[0019] Step 4 is as follows:

[0020] Step 4.1: First, initialize the entire subgroup;

[0021] Step 4.2: Calculate the size of each subpopulation. The calculation formula is as follows:

[0022]

[0023] The size of the subpopulation is represented as where It is by the most recent The highest score obtained from the feature subset (individuals) corresponds to the classification accuracy. This value and the individuals that obtained it are updated as needed in each round. It is the size of the entire population in the genetic algorithm; It is a coefficient that determines the advantage of the subpopulation with better highest classification accuracy;

[0024] Step 4.3: Perform crossover and mutation in each subpopulation;

[0025] Step 4.4: After crossover and mutation, the optimal feature subset is obtained.

[0026] Step 5, which initializes the parameters of the improved particle swarm optimization-multilayer feedforward neural network model based on feature selection, specifically involves initializing the parameters of the improved particle swarm optimization-multilayer feedforward neural network model based on feature selection, including the maximum and minimum values ​​of the inertia factor and the maximum value of the acceleration constant. Minimum value The acceleration constant is calculated as follows:

[0027]

[0028] Among them, the acceleration constant and These two parameters represent the weights of the statistical acceleration term for each particle.

[0029] The specific process of updating the particle state in step 7 is as follows:

[0030] Two parameters are used to represent the weight of the statistical acceleration term for each particle. By adjusting the proportion of the two acceleration parameters in the early and late stages, the specific expression is as follows:

[0031]

[0032] In the formula c max , c min and are the maximum and minimum values ​​of the acceleration constant, respectively;

[0033] The method for adjusting the two acceleration parameters introduces dynamic adjustment parameters, combining the particle swarm optimization algorithm and the gradient descent method to adjust the model, as shown below:

[0034]

[0035] in These are the overall network parameters of the model, while These are dynamically adjustable parameters. This is the gradient representation of the model in the gradient descent method. This represents the model learning rate. This indicates the model parameters determined by the algorithm.

[0036] In step 8, the inertia factor is adaptively adjusted according to the particle's fitness. The adaptive adjustment formula is:

[0037]

[0038] In the formula This represents the maximum value of the inertia factor. That is the minimum value. It refers to a linear function. It represents a nonlinear function. It is the expression for the minimum fitness of a particle in the algorithm. This represents the fitness of the current particle.

[0039] The beneficial effects of this invention are:

[0040] By combining feature scoring and a genetic algorithm, preliminary feature scores are calculated for various signal features extracted from power quality load disturbance signals, and these features are then sorted and filtered. The filtered features are further refined using a genetic algorithm through competition among multiple populations, ultimately obtaining a low-dimensional optimal feature subset. This significantly reduces the dimensionality of the feature subset, lowers the structural complexity of the classifier, reduces redundancy during classification, and improves classification accuracy. Subsequently, a dynamically adjusted adaptive IPSO-BP algorithm is employed to identify power quality disturbances, achieving rapid classification and further reducing the time required for classification. Attached Figure Description

[0041] Figure 1 This is a flowchart of a power quality disturbance identification method using feature selection according to the present invention.

[0042] Figure 2 This is a graph showing the feature scores of a power quality disturbance identification method using feature selection according to the present invention.

[0043] Figure 3(a) shows the individual crossover probability (p) of the power quality disturbance identification method using feature selection according to the present invention. j ) Classification accuracy plots under different values;

[0044] Figure 3(b) shows the individual mutation probability (p) of the power quality disturbance identification method using feature selection according to the present invention. t ) Classification accuracy plots under different values;

[0045] Figure 4 This is a waveform diagram of composite power quality disturbances in a power quality disturbance identification method using feature selection according to the present invention.

[0046] Figure 4 include Figure 4-1 to Figure 4-1 0. Detailed Implementation

[0047] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0048] This invention discloses a power quality disturbance identification method using feature selection, which is implemented according to the following steps:

[0049] Step 1: Add Gaussian white noise to the power quality signal to obtain the power quality composite disturbance signal;

[0050] Step 2: Extract features from composite power quality disturbance signals;

[0051] Step 3: Input each feature into the support vector machine for recognition, obtain the classification accuracy of each feature, and sort the classification accuracy of each feature by the support vector machine and assign scores accordingly.

[0052] Step 4: Use a multi-population genetic algorithm to perform... The round of evolutionary computation seeks the optimal feature subset, which is then used as the sample; the specific process is as follows:

[0053] Step 4.1: First, initialize the entire subgroup;

[0054] Step 4.2: Calculate the size of each subpopulation. The calculation formula is as follows:

[0055]

[0056] The size of the subpopulation is represented as where It is by the most recent The highest score obtained from the feature subset (individuals) corresponds to the classification accuracy. This value and the individuals that obtained it are updated as needed in each round. It is the size of the entire population in the genetic algorithm; It is a coefficient that determines the advantage of the subpopulation with better highest classification accuracy;

[0057] Step 4.3: Perform crossover and mutation in each subpopulation;

[0058] Step 4.4: After crossover and mutation, the optimal feature subset is obtained.

[0059] Step 5: Construct an improved particle swarm optimization-multilayer feedforward neural network model based on feature selection, and initialize the parameters of the improved particle swarm optimization-multilayer feedforward neural network model based on feature selection.

[0060] The specific process for initializing the parameters of the improved particle swarm optimization-multilayer feedforward neural network model based on feature selection is as follows: Initializing the parameters of the improved particle swarm optimization-multilayer feedforward neural network model based on feature selection includes the maximum and minimum values ​​of the inertia factor and the maximum value of the acceleration constant. Minimum value The acceleration constant is calculated as follows:

[0061]

[0062] Among them, the acceleration constant and These two parameters represent the weights of the statistical acceleration term for each particle.

[0063] Step 6: Use the network parameter x from the first forward propagation of the neural network as the position of the initialized particle;

[0064] Step 7: Input the previously obtained samples and particles, update the particle states, and calculate the fitness of each particle; the specific process for updating the particle states is as follows:

[0065] Updating the particle state involves updating its inertia factor, acceleration constant, and dynamic adjustment coefficient. The calculation methods for the acceleration constant and dynamic adjustment coefficient are as follows:

[0066] In the PSO algorithm, two crucial parameters are the acceleration constants and their weights. These parameters represent the weights of the statistical acceleration term for each particle. Appropriate values ​​help particles perform faster searches, enabling the classification of different energy quality perturbations and improving classification efficiency. This invention accelerates the classification speed by adjusting the proportions of the two acceleration constants in the early and late stages. The specific expressions are shown below:

[0067]

[0068] In the formula, c max , c min and are the maximum and minimum values ​​of the acceleration constant, respectively.

[0069] The expression for the traditional gradient descent method is as follows:

[0070]

[0071] In the improved BP neural network proposed in this invention, a dynamic adjustment parameter is introduced into the method for adjusting system data. This allows the PSO algorithm to be combined with the gradient descent method, enabling timely model adjustments and better adaptability to classification under different power quality disturbances, as detailed below:

[0072]

[0073] in These are the overall network parameters of the model, while These are dynamically adjustable parameters. This is the gradient representation of the model in the gradient descent method. This represents the model learning rate. This indicates the model parameters determined by the algorithm.

[0074] Step 8: Adaptively adjust the inertia factor according to the particle's fitness.

[0075] The inertia factor is adaptively adjusted according to the particle's fitness. The adaptive adjustment formula is as follows:

[0076]

[0077] In the formula This represents the maximum value of the inertia factor. That is the minimum value. It refers to a linear function. It represents a nonlinear function. It is the expression for the minimum fitness of a particle in the algorithm. This represents the fitness of the current particle.

[0078] Step 9: Update the individual extreme values ​​and the global extreme values ​​based on the fitness of each particle, and output the results to determine the global extreme value point.

[0079] Step 10: Update the dynamic adjustment coefficients and output the optimal parameters;

[0080] Step 11: Train and identify the model with the optimal feature subset as input to the optimal parameter to obtain the optimized particle swarm optimization-multilayer feedforward neural network model based on feature selection.

[0081] Step 12: Input the signal to be identified into the optimized feature-based improved particle swarm optimization-multilayer feedforward neural network model and output the classification result.

[0082] Example

[0083] First, data of the power quality composite disturbance signal to be judged is collected. This invention simulates the power quality disturbance signal that can be collected under normal conditions by adding 30-50 dB Gaussian white noise. The specific composite disturbance waveform can be seen in [the diagram]. Figure 4The sampling rate of the power quality disturbance waveform is 12.8kHz, the sampling duration is 1s, and there are a total of 12.8k sampling points. The high noise environment includes 5dB, 10dB, 15dB and mixed 5-15dB cases. There are 8 single signals: standard signal (C0), voltage sag (C1), voltage swell (C2), voltage interruption (C3), transient pulse (C4), transient oscillation (C5), voltage harmonic (C6), and voltage flicker (C7), as well as 10 composite disturbance signals: voltage sag + transient pulse (C8), voltage sag + harmonic (C9), voltage swell + transient oscillation (C10), voltage swell + flicker (C11), interruption + transient oscillation (C12), interruption + harmonic (C13), transient pulse + harmonic (C14), transient oscillation + harmonic (C15), transient oscillation + flicker (C16), and harmonic + flicker (C17), for a total of 18 types of PQD signals.

[0084] Input the power quality composite disturbance signal and extract the features required for screening in this invention. Finally, obtain a variety of time-domain features and entropy divided according to different time scales. The specific features are shown in Table 1 below.

[0085] Table 1

[0086]

[0087] To analyze the feature scores used in this invention and the elite feature subsets obtained after feature extraction and score calculation for 18 power quality disturbance signals containing 10 types of composite disturbances, the accuracy of each feature subset after passing through the classifier, and the score ranking of each feature are shown in Table 2:

[0088] Table 2

[0089]

[0090] As shown in Table 2, among the many features proposed in this invention at different time scales, the 19 features with high power quality disturbance recognition rates among the 18 features selected in this invention are listed in Table 1 and are ranked as follows. Therefore, the elite subset can be divided into {...} based on Table 2. F 24 , F 17 , F 6, F 1, F 49 , F 22 , F 35 , F 12 , F 43, F 50 , F 13 , F 22 , F 5, F 53 , F 44 , F 19 , F 20 , F 2, F 7}. The specific scores for each feature are as follows: Figure 2 As shown.

[0091] In the aforementioned process, this invention randomly selects parts of the subgroup for crossover and mutation. However, different selection ratios result in varying recognition accuracy of the optimal subset obtained through iteration. To find the optimal ratio, this invention experimentally determines the individual crossover probability p. j With individual mutation probability p t The values ​​are shown in Figure 3(a) and Figure 3(b).

[0092] As can be seen from Figures 3(a) and 3(b), at p j =0.8 and p t When p = 0.5, the classification accuracy of each feature subset is the highest. Therefore, in subsequent classifications of the corresponding types, p j With p t The values ​​are 0.8 and 0.5 respectively. Therefore, the optimal subset determined by this invention is the elite subset after p... j =0.8 and p t =0.5 is obtained after the change.

[0093] To analyze whether the optimal subset obtained by this invention achieves the highest accuracy not only with a single classifier but also with other classifiers, this invention proposes to input multiple feature subsets into the classifier for experiments to determine the accuracy of the optimal subset. The specific experimental results are shown in Table 3:

[0094] Table 3

[0095]

[0096] As can be seen from the data analysis in Table 3, the optimal feature subset obtained by the present invention based on the feature score of the support vector machine and the genetic algorithm can achieve the highest recognition accuracy among various power quality disturbance recognition algorithms.

[0097] The IPSO-BP algorithm proposed in this invention improves the classification speed compared to traditional algorithms, as shown in Table 4 below:

[0098] Table 4

[0099]

[0100] As can be seen from the analysis in Table 4, the improved particle swarm optimization-multilayer feedforward neural network power quality disturbance identification method based on feature screening proposed in this invention not only has a significantly improved classification speed compared to the traditional particle swarm optimization-multilayer feedforward neural network algorithm, but also has a significantly improved identification speed compared to other common power quality identification methods.

[0101] Through the above methods, the improved particle swarm optimization-multilayer feedforward neural network power quality disturbance identification method based on feature screening proposed in this invention addresses the issue that most current power quality disturbance identification methods rely heavily on frequency domain feature analysis, and time-frequency analysis methods sometimes require multiple analysis calculations, leading to a decrease in feature extraction speed and an inability to quickly extract features of power quality disturbance signals. This invention proposes a method combining feature scoring and genetic algorithms to extract the optimal subset, reducing processing speed. Furthermore, it employs an adaptive IPSO-BP algorithm based on dynamic weights to further optimize classification efficiency. Finally, experiments demonstrate that the IPSO-BP power quality disturbance identification method based on feature screening proposed in this invention can identify composite power quality disturbances faster and more accurately than existing methods.

Claims

1. A power quality disturbance identification method employing feature selection, characterized in that, The specific steps are as follows: Step 1: Add Gaussian white noise to the power quality signal to obtain the power quality composite disturbance signal; Step 2: Extract features from composite power quality disturbance signals; Step 3: Input each feature into the support vector machine for recognition, obtain the classification accuracy of each feature, and sort the classification accuracy of each feature by the support vector machine and assign scores accordingly. Step 4: Use a multi-population genetic algorithm to perform... The round of evolutionary computation seeks the optimal feature subset, and the optimal feature subset is used as the sample; Step 5: Construct an improved particle swarm optimization-multilayer feedforward neural network model based on feature selection, and initialize the parameters of the improved particle swarm optimization-multilayer feedforward neural network model based on feature selection. Step 6: Use the network parameter x from the first forward propagation of the neural network as the position of the initialized particle; Step 7: Input the previously obtained samples and particles, update the particle states, and calculate the fitness of each particle; The specific process of updating the particle state in step 7 is as follows: Two parameters are used to represent the weight of the statistical acceleration term for each particle. By adjusting the proportion of the two acceleration parameters in the early and late stages, the specific expression is as follows: In the formula c max , c min and are the maximum and minimum values ​​of the acceleration constant, respectively; The method for adjusting the two acceleration parameters introduces dynamic adjustment parameters, combining the particle swarm optimization algorithm and the gradient descent method to adjust the model, as shown below: in These are the overall network parameters of the model, while These are dynamically adjustable parameters. This is the gradient representation of the model in the gradient descent method. This represents the model learning rate. This indicates the model parameters determined by the algorithm; Step 8: Adaptively adjust the inertia factor according to the particle's fitness. Step 8 describes adaptively adjusting the inertia factor based on the particle's fitness. The adaptive adjustment formula is as follows: In the formula This represents the maximum value of the inertia factor. That is the minimum value. It refers to a linear function. It represents a nonlinear function. It is the expression for the minimum fitness of a particle in the algorithm. This represents the fitness of the current particle. Step 9: Update the individual extreme values ​​and the global extreme values ​​based on the fitness of each particle, and output the results to determine the global extreme value point. Step 10: Update the dynamic adjustment coefficients and output the optimal parameters; Step 11: Train and identify the model with the optimal feature subset as input to the optimal parameter to obtain the optimized particle swarm optimization-multilayer feedforward neural network model based on feature selection. Step 12: Input the signal to be identified into the optimized feature-based improved particle swarm optimization-multilayer feedforward neural network model and output the classification result.

2. The power quality disturbance identification method using feature selection according to claim 1, characterized in that, Step 4 is as follows: Step 4.1: First, initialize the entire subgroup; Step 4.2: Calculate the size of each subpopulation. The calculation formula is as follows: Represented as the size of the subpopulation, where It is by the most recent The highest score obtained from the feature subset corresponds to the classification accuracy. This value is updated and the individual corresponding to it is obtained in each round. It is the size of the entire population in the genetic algorithm; It is a coefficient that determines the advantage of the subpopulation with better highest classification accuracy; Step 4.3: Perform crossover and mutation in each subpopulation; Step 4.4: After crossover and mutation, the optimal feature subset is obtained.

Images