Method for optimizing identification data of primary frequency modulation system of thermal power generating unit by considering excitation characteristics

By calculating the Fischer information matrix condition number and using fuzzy C-mean and random forest algorithms, the sufficient excitation data segment in the primary frequency modulation system of the thermal power unit was screened, which solved the problem of insufficient consideration of excitation characteristics in the existing technology and improved the identification data accuracy and efficiency.

CN119046822APending Publication Date: 2024-11-29STATE GRID HEBEI ENERGY TECH SERVICE CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411128336.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-16
Publication Date
2024-11-29

AI Technical Summary

Technical Problem

In the prior art, it is difficult to effectively consider the excitation characteristics in the identification process of the primary frequency modulation system of thermal power units, resulting in poor conformity between the model output and the actual operation output, and manual screening of data is cumbersome, which affects efficiency.

Method used

The excitation characteristics are calculated using the Fischer information matrix conditional number, combined with the fuzzy C-mean algorithm and the random forest classification algorithm, a classification rule model is generated, and the identification data segments containing sufficient excitation are selected.

Benefits of technology

It improves the speed and accuracy of data screening, reduces the blindness of manual screening, saves data processing time, reduces repetitive work, and enhances the reliability of identification results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119046822A_ABST
    Figure CN119046822A_ABST
Patent Text Reader

Abstract

The invention relates to a thermal power generating unit primary frequency modulation system identification data optimization method considering excitation characteristics. The method comprises the following steps: S1, preliminarily selecting historical operation data; s2, acquiring a sample data segment; s3, acquiring excitation characteristics representing each segment of sample data; s4, performing data reconstruction on each segment of sample data to obtain a feature set containing excitation features; s5, generating a priori data set by using a fuzzy C-means algorithm; s6, generating a classification rule model by using a random forest classification algorithm according to the prior data set; s7, obtaining a classification rule model used for screening identification data; according to the method, a data segment which contains sufficient excitation and has a good identification output effect in historical operation data can be effectively optimized, the data screening speed is increased, a large amount of data processing time is saved, blindness and contingency of manual data selection are reduced, and an identification data screening basis is provided for system parameter identification; the data screening speed is improved, and a large amount of repeated work is reduced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a method for optimizing identification data of a primary frequency modulation system of a thermal power unit taking into account excitation characteristics, and belongs to the technical field of power system data processing. Background Art

[0002] With a large amount of new energy power connected to the grid, the energy structure of the power industry is constantly adjusting, posing a threat to the stability of the grid frequency. Thermal power units are important frequency regulation resources on the power supply side, and the primary frequency regulation of thermal power units is to adjust the power by the unit itself. They have the advantage of fast response speed, can alleviate the changes in grid frequency, and are an important guarantee for maintaining the stability of grid frequency.

[0003] Therefore, when analyzing the stability of the power system, the primary frequency regulation system of the thermal power unit is usually modeled and identified to understand the actual frequency regulation capability of the unit; the traditional system modeling methods mainly include mechanism modeling and system identification modeling. For thermal power units with complex physical structures and numerous subsystems, the mechanism modeling steps are numerous, the nonlinear link modeling is complex, and the mobility between units is poor. In addition, the parameter identification of the mechanism modeling usually does not take environmental factors into consideration enough, and the actual process is affected by environmental interference, resulting in poor consistency between the output of the mechanism model and the actual operation output data. Research on obtaining model identification data mainly focuses on model characteristic tests and historical operation data. At present, most of the field directly adopts the field step response test to establish the full-condition mathematical model of the primary frequency regulation of the unit to reflect the characteristics of the controlled object, but the field test steps are complicated, time-consuming, and affect the normal production operation. At present, the system parameter identification modeling process often requires high-quality input and output data. However, the method of adding long-term excitation signals to obtain test data in the production process is not conducive to normal unit operation. Compared with the data obtained from the identification test, the identification method based on massive historical operation data has the advantages of less impact on unit operation and low economic cost. However, when the unit is running relatively smoothly and the load changes slightly, the excitation characteristics of the input and output identification data are ignored, and the output of the identification result is quite different from the actual operation output. Considering that the current identification data screening work is mostly manually selected, and many factors such as the correlation between data variables, data change trends, and whether the identification data is sufficiently excited have an impact on the identification results, and it also causes great difficulties for manual selection of data. Therefore, it is of great significance to realize the screening of historical operation data considering the excitation characteristics for the identification research of the primary frequency regulation system of thermal power units. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a method for optimizing identification data of a primary frequency modulation system of a thermal power unit taking into account excitation characteristics, which provides a basis for screening identification data for system parameter identification.

[0005] The present invention adopts the following technical solution:

[0006] The method for optimizing identification data of a primary frequency modulation system of a thermal power plant unit considering excitation characteristics of the present invention comprises the following steps:

[0007] S1. Preliminary selection of historical operation data based on the identification system model structure;

[0008] S2. Divide the historical operation data into multiple data samples of the same length according to the time series, and obtain sample data segments;

[0009] S3, calculating the Fisher information matrix condition number of each segment of sample data, and obtaining the excitation characteristics that characterize each segment of sample data;

[0010] S4, reconstructing each segment of sample data to obtain a feature set containing excitation characteristics;

[0011] S5. Generate a priori data set using fuzzy C-means algorithm, including feature set and clustering results;

[0012] S6. Generate a classification rule model using a random forest classification algorithm based on a priori data set;

[0013] S7. According to the model evaluation index of the validation set, the classification result model is fixed to obtain a classification rule model for screening and identifying data.

[0014] The calculation of the Fisher information matrix condition number in step S3 of the present invention comprises the following steps:

[0015] S31. Calculate the Fisher information matrix F according to the input and output functions of the identification system model. The formula is as follows:

[0016]

[0017] Where f is the transfer function G(s), is an r-dimensional parameter vector, u t is the main steam flow d s , Unit load P e is the output; then the Jacobian matrix J corresponding to f is:

[0018]

[0019]

[0020]

[0021]

[0022] Let F = M T M, then

[0023]

[0024] Where f is an arbitrary function, is an r-dimensional parameter vector;

[0025] S32, calculating the maximum eigenvalue max(|λ(F)|) and the minimum eigenvalue min(|λ(F)|) of the Fisher information matrix F;

[0026] S33, calculating the Fisher information matrix condition number η according to the eigenvalue calculated in S32 data , characterizes the excitation characteristics of the sample data, and the calculation formula is as follows:

[0027]

[0028] Where λ(F) is the eigenvalue of the Fisher information matrix F, η data is the Fisher information matrix condition number.

[0029] Reconstructing the sample data in step S4 of the present invention includes: calculating data trend features, data correlation features and combining a feature set including excitation characteristics.

[0030] The formula for calculating the trend characteristics of data in the present invention is:

[0031]

[0032] Max X =max(X)-min(X)#

[0033]

[0034] Among them, c X For the data sequence X = [x1,…,x n ] is the coefficient of variation, σ X is the standard deviation, μ X is the mean difference; Max X is the range, max(X) and min(X) are the maximum and minimum values ​​in the data sequence X respectively; ΔM is the mode band of the data sequence X, n M is the data sequence X in the interval [μ X -kMa X ,μ X +kMax X ,], k is taken as 0.5.

[0035] The specific calculation formula of the data correlation feature of the present invention is as follows:

[0036]

[0037] Among them, r XY To identify the data correlation characteristics between the input and output variables of the system model, cov(X,Y) is the covariance of the variable sequences X and Y, and var(·) is the variance of the corresponding variable sequence.

[0038] The data feature set comprising the excitation characteristics described in the present invention specifically comprises the following steps:

[0039] S41. Using a normalization method to perform standard normalization processing on data trend characteristics and data correlation characteristics;

[0040] S42. Correspond the data trend and correlation feature set corresponding to each group of data samples to the Fisher information matrix condition number, add them to a column, and generate a data sample set containing excitation characteristics.

[0041] Step S5 of the present invention uses the fuzzy C-means algorithm to generate a priori data set, which includes the following steps:

[0042] S51. Define the objective function of the clustering algorithm. The calculation formula is as follows:

[0043]

[0044] Where J is the objective function; n is the number of samples; c is the number of cluster centers; m is any real number greater than 1; x i is the i-th sample point; u ij is the membership degree of the jth class center of the ith sample point; c j is the jth class center; ||x i -c j || represents x i to c j The Euclidean distance of

[0045] S52, select a data sample set containing the excitation characteristics as the input of the clustering algorithm, set the stop iteration threshold and the maximum number of iterations, and minimize the objective function by continuously iterating the membership degree and the cluster center. Obtain two clusters of data;

[0046] S53, by comparing the values ​​of the cluster center points of the two clusters of data in the direction of characterizing the excitation characteristics, the larger cluster of data is classified as suitable identification data 1, and the other cluster is classified as unsuitable identification data 0;

[0047] S54, a data sample set containing the excitation characteristics is matched one by one with the corresponding classification results to generate a priori data set.

[0048] Step S6 of the present invention uses the random forest classification algorithm to generate a classification rule model, which includes the following steps:

[0049] S61, dividing the standardized prior data set into a training set and a validation set according to a ratio of 80% and 20%;

[0050] S62, initializing and setting the number of basic classifiers of the random forest classification model parameters and the number of iterations of ensemble learning n_estimatiors;

[0051] S63, taking the data sample set in the training set as input and the classification result as output, performing iterative training to generate a classification rule model;

[0052] S64, calculating the fitness function according to the number of iterations, the specific calculation formula is as follows, obtaining the optimal number of basic classifiers and the number of iterations of ensemble learning n_estimatiors;

[0053]

[0054] Among them, TP and TN represent the number of samples predicted correctly for the positive and negative classes, and FP and FN represent the number of samples predicted incorrectly for the positive and negative classes.

[0055] S65. Input the validation set data into the generated classification rule model to evaluate the accuracy of the classification rule model.

[0056] The specific calculation formula for evaluating the accuracy of the classification rule model generated by the random forest classification algorithm in the present invention is as follows:

[0057] Recall: Precision: F-score:

[0058]

[0059] The present invention S41 adopts the Z-score method to make it conform to the standard normal distribution, that is, the mean is 0, the standard deviation is 1, and the data processing function is:

[0060]

[0061] Where x′ is the normalized data, μ is the sample mean, and σ is the sample standard deviation.

[0062] The positive effects of the present invention are as follows: the present invention calculates the Fisher information matrix condition number of historical operation data to characterize the excitation characteristics of identification data, and performs cluster analysis based on the fuzzy C-means algorithm to provide a priori data sample set, which not only realizes the dimensionality reduction and reconstruction of historical data, but also can better retain the dynamic information of historical operation data; secondly, the random forest classification algorithm is used to generate an identification data classification rule model, which can effectively select data segments in the historical operation data that contain sufficient excitation and have good identification output effects, improve the speed of screening data, save a lot of data processing time, reduce the blindness and contingency of artificially selected data, and provide a basis for screening identification data for system parameter identification; improve the speed of screening data, reduce a lot of repetitive work, and avoid the problem of poor identification effect caused by deviations generated in artificially screening and piecing together historical data. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Attached Figure 1 It is a flow chart of the present invention;

[0064] Attached Figure 2 This is a schematic diagram of the steam turbine work model structure;

[0065] Attached Figure 3 This is a schematic diagram of the steam turbine work model identification data structure of the present invention;

[0066] Attached Figure 4 Iterate the fitness function curve graph for the model;

[0067] Attached Figure 5 A comparison chart of classification results screened by the classification rule model of the present invention;

[0068] Attached Figure 6 This is a step response comparison diagram of the screening result data identification model of the present invention. DETAILED DESCRIPTION

[0069] As attached Figure 1 As shown, the present invention proposes a method for optimizing identification data of a primary frequency regulation system of a thermal power unit taking into account excitation characteristics. The primary frequency regulation system of a thermal power unit includes various adjustment systems. The present invention is described in detail using actual data of a steam turbine working system as an example.

[0070] Step 1: According to the identification system model structure, historical operation data is collected and segmented into multiple data samples of the same length; sample data segments are obtained;

[0071] The steam turbine power system is the main component of the primary frequency regulation model of a typical thermal power unit. The response capability of the frequency regulation action is closely related to the steam turbine power. The model structure is shown in the attached figure. Figure 2 As shown in the figure, in this model, the steam entering the turbine flows through the high-pressure cylinder, reheater, and medium- and low-pressure combined cylinders to expand and input mechanical power;s represents the main steam flow, t / h; p1 represents the pressure after the valve is adjusted, MPa; p2 represents the reheater pressure, MPa; P e represents the unit load, MW; T1 represents the volume time constant of the front steam chamber of the high-pressure cylinder, s; T2 represents the volume time constant of the reheater, s; F1 represents the power distribution coefficient of the high-pressure cylinder, F2 represents the power distribution coefficient of the medium and low-pressure combined cylinders, and λ represents the natural over-modulation coefficient of the high-pressure cylinder power; the example model of the present invention is based on the main steam flow d s As input, the unit load P e is the output; the transfer function G(s) is:

[0072]

[0073] in,

[0074] a0=1#

[0075] a1=T1+T2#

[0076] a2=T1T2#

[0077] b0=F1+F2=1#

[0078] b1=T2F1(1+λ)#;

[0079] The actual historical operation data of a 300MW thermal power unit in Handan is used, with a sampling time of 1s, and a total of 1,005,000 sets of data samples;

[0080] Step 2: Get the sample data segment; as shown in the attached Figure 3 As shown, the data is divided sequentially according to the length of the identification data of 5000, and 201 groups of sample data are obtained after preprocessing the data;

[0081] Step 3: Calculate the Fisher information matrix condition number between the input and output data of each set of sample data to characterize the excitation characteristics of the data in the data sample;

[0082] Considering that there is measurement noise in the collected historical data and the dimension of the parameter vector is smaller than the amount of data used for identification, it can be considered that there are identification data segments containing continuous and sufficient excitation in the historical data. Therefore, the Fisher information matrix condition number is considered as a criterion for determining whether the data used for identification contains continuous and sufficient excitation.

[0083] The Fisher information matrix condition number of the transfer function G(s) from the main steam flow to the unit load is calculated to evaluate the excitation characteristics of the data, and the window data length is 500;

[0084] S31. The input and output of the steam turbine power system identification process are as follows:

[0085]

[0086] Where f is the transfer function G(s), is an r-dimensional parameter vector, u t is the main steam flow d s , Unit load P e is the output, then the Jacobian matrix J corresponding to f is:

[0087]

[0088]

[0089] So it can be expressed as:

[0090]

[0091]

[0092] Let F = M T M, then

[0093]

[0094] Where f is an arbitrary function, is an r-dimensional parameter vector;

[0095] S32, calculating the maximum eigenvalue max(|λ(F)|) and the minimum eigenvalue min(|λ(F)|) of the Fisher information matrix F;

[0096] S33, according to the eigenvalue calculated in S32, calculate the Fisher information matrix condition number η data , characterizes the excitation characteristics of the sample data, and the calculation formula is as follows:

[0097]

[0098] Where λ(F) is the eigenvalue of the Fisher information matrix F, η data is the Fisher information matrix condition number.

[0099] The Fisher information matrix condition number between the main steam flow and the generator power of each data sample is calculated according to the above formula to characterize the data excitation characteristics and constitute the excitation characteristics of the sample data.

[0100] When the ratio of the maximum eigenvalue max(|λ(F)|) to the minimum eigenvalue min(|λ(F)|) of the matrix is ​​larger, it is more sensitive to the disturbance in the identification data and less robust to noise, which makes the accuracy of the identification parameters worse; on the contrary, when the ratio is smaller, that is, the condition number η data The smaller it is, the better the convergence and stability of the identification parameters.

[0101] Step 4: Calculate the data trend characteristics and data correlation characteristics for each data sample, and synthesize them into a data feature set containing excitation characteristics in the Fisher information matrix condition array.

[0102] Considering that the correlation and change trend between variables in historical data are also within the scope of consideration for manual screening of identification data, the data variable correlation, data change trend and data excitation characteristics are used as the feature set to characterize historical data, as shown in Table 1. The length of the identification data segment is fixed at r = 5000, and then all the features in the table are calculated and obtained. Each identification data sample is reconstructed into an identification data feature set consisting of f = 3 data features, which not only realizes the dimensionality reduction of the data, but also retains the temporal nature of the running data of this segment.

[0103] Trend describes the direction or pattern of change of data over time or other related factors; trend analysis is an important part of time series analysis, which helps to understand the long-term evolution of data and possible periodicity or regularity; therefore, statistics such as mean, standard deviation, coefficient of variation, mode band, etc. are used to characterize the trend characteristics of data;

[0104]

[0105] Max X =max(X)-min(X)#

[0106]

[0107] Among them, c X For the data sequence X = [x1,…,x n ] is the coefficient of variation, σ X is the standard deviation, μ X is the mean difference; Max X is the range, max(X) and min(X) are the maximum and minimum values ​​in the data sequence X respectively; ΔM is the mode band of the data sequence X, n M is the data sequence X in the interval [μ X -kMax X ,μ X +kMax X ,], k is usually taken as 0.5;

[0108] The correlation coefficient is mainly used to measure the linear correlation between two variables. n ] and Y=[y1,y2,…,y n ], and its correlation coefficient is:

[0109]

[0110] Among them, r XY To identify the data correlation characteristics between the input and output variables of the system model, cov(X,Y) is the covariance of the variable sequences X and Y, and var(·) is the variance of the corresponding variable sequence.

[0111] Table 1 Data characteristics

[0112]

[0113] S41. Normalize the data trend characteristics and data correlation characteristics, and use the Z-score method to make them conform to the standard normal distribution, that is, the mean is 0 and the standard deviation is 1. The data processing function is:

[0114]

[0115] Where x′ is the normalized data, μ is the sample mean, and σ is the sample standard deviation.

[0116] S42, corresponding the data trend and correlation feature set corresponding to each group of data samples to the Fisher information matrix condition number, adding them to a column, and generating a data sample set D containing the excitation characteristics 12×201 ;

[0117] Step 5: Use the fuzzy C-means algorithm to convert the data sample set D 12×201 As input data, generate a priori data set, which includes feature set and clustering results;

[0118] S51. Define the objective function of the clustering algorithm. The calculation formula is as follows:

[0119]

[0120] Where J is the objective function; n = 201 is the number of samples; c = 2 is the number of cluster centers; m is any real number greater than 1; x i is the i-th sample point; u ij is the membership degree of the jth class center of the ith sample point; c j is the jth class center; ||x i -c j || represents x i to c j The Euclidean distance of .

[0121] S52, select a data sample set containing the excitation characteristics as the input of the clustering algorithm, set the stop iteration threshold to 0.1 and the maximum number of iterations to 100, and minimize the objective function by continuously iterating the membership degree and the cluster center, and obtain two clusters of data results.

[0122] S53. By comparing the values ​​of the cluster centers of the two clusters of data in the direction of characterizing the excitation characteristics, the larger cluster of data is classified as suitable identification data "1", and the other cluster of data is classified as unsuitable identification data "0".

[0123] S54 , a data sample set containing the excitation characteristics is matched one by one with the corresponding classification results to generate a priori data set D.

[0124] Step 6: Divide the reconstructed sample set into a training set and a test set, and generate a classification rule model based on the random forest classification algorithm;

[0125] After reconstructing the data, 201 sets of sample data are obtained, of which 160 sets of sample data D train Used to generate classification rule models, validation set D test Samples: 41 groups;

[0126] As an integrated learning method based on decision trees, the basic principle of the random forest classification algorithm is to combine multiple basic learners and ultimately decide the final classification prediction result by voting to avoid the overfitting problem of a single decision tree. Since randomness is introduced when training a single decision tree, the classification prediction accuracy and generalization ability of the model are improved.

[0127] The feature set and feature identifier of each segment of the actual historical operation data are constructed respectively, and then the sample set D used to train the classification rule model is train Put it into the random forest network, set the number of basic learners N = 200, and the maximum feature tree M f = 10; the classification rule model learning training iteration fitness function curve is as shown in the attached Figure 4 ,When the number of basic learners N = 17, the highest classification accuracy score is 0.905, generating a classification rule model.

[0128] Step 7: Use the trained classification rule model to validate the set D test Verification is performed and the results are as follows Figure 5 As shown in the figure, the overlap between the actual verification sample classification identification and the predicted verification sample classification represents the effectiveness of the model. There are 19 groups of data segments that are actually suitable for identification, 22 groups of data segments that are not suitable for identification, and only one verification set sample is misclassified in the data segment that is suitable for identification.

[0129] Recall, precision and F-score are used to evaluate the effectiveness of the predicted classification results; the overall classification accuracy of the model is 0.97561, and the recall rate of suitable identification data is as high as 0.94737, the precision rate is 1, and the F-score is 0.96128. A preferential classification rule model for identification data is established. Its purpose is to non-artificially select data segments suitable for identification with sufficient incentives from longer historical operating data, rather than to screen out all data segments suitable for identification in historical data. Therefore, it is believed that the model accuracy of the present invention can meet the actual use requirements.

[0130] Attached Figure 6 The figure shows the identification model comparison diagram in which 4 groups of data samples are divided into data suitable for identification and 6 groups of data samples are divided into data unsuitable for identification; according to step 4, parameter identification is performed on each group of data samples, and a unit step signal is used as an input signal to output the response curve of each group of identification results, which is compared with the actual parameter model response curve obtained from the turbine work model characteristic test.

[0131] The classification results are 4 groups of data samples suitable for identification, and their identification response curves are shown in the left figure. The classification results are 6 groups of data samples unsuitable for identification, and their identification response curves are shown in the right figure. As shown in the figure, the unit step response curve of the identification model of the sample data whose classification results are suitable for identification is consistent with the actual response curve of the turbine work model, and the fitting degrees Fits are 90.8%, 89.99%, 81.5%, and 96.41%, respectively, all greater than 80%, which can better reflect the actual action response. In contrast, the unit step response curve of the identification model of the data samples whose classification results are too unsuitable for identification is quite different from the actual response curve, and data samples 3, 4, and 5 all have overshoot responses, and the Fits of the 6 groups of data samples are all lower than 30%.

[0132] To sum up, in order to screen out fully motivated identification data from historical operating data and reduce the blindness of manual data selection, the present invention takes into account the requirement that identification data needs to be fully motivated, combines the fuzzy C-means algorithm with the random forest classification algorithm, and uses historical operating data to screen out data segments suitable for parameter identification, thereby avoiding the blindness and contingency of manual screening, improving the speed of data screening, saving a lot of data processing time, reducing a lot of repetitive work, and avoiding the problem of poor identification effect caused by deviations in manual screening and piecing together historical data.

Claims

1. A method for optimizing identification data of a primary frequency modulation system of a thermal power unit considering excitation characteristics, characterized in that: The following steps are involved: S1. Preliminarily select historical operation data according to the identification system model structure; S2. Divide the historical operation data into multiple data samples of the same length according to the time sequence, and obtain sample data segments; S3, calculating the Fisher information matrix condition number of each segment of sample data to obtain the excitation characteristics that characterize each segment of sample data; S4, reconstructing each segment of sample data to obtain a feature set that includes the excitation characteristics; S5, using the fuzzy C-means algorithm to generate a priori data set, including a feature set and a clustering result; S6, using the random forest classification algorithm to generate a classification rule model based on the priori data set; S7. According to the model evaluation index of the validation set, the classification result model is fixed to obtain a classification rule model for screening and identifying data.

2. The method for optimizing identification data of a primary frequency modulation system of a thermal power plant considering excitation characteristics according to claim 1 is characterized in that: The calculation of the Fisher information matrix condition number in step S3 comprises the following steps: S31. Calculate the Fisher information matrix F according to the input and output functions of the identification system model. The formula is as follows: Where f is the transfer function G(s), is an r-dimensional parameter vector, u t is the main steam flow d s , Unit load P e is the output; then the Jacobian matrix J corresponding to f is: Let F = M T M, then Where f is an arbitrary function, is an r-dimensional parameter vector; S32, calculating the maximum eigenvalue max(|λ(F)|) and the minimum eigenvalue min(|λ(F)|) of the Fisher information matrix F; S33, calculating the Fisher information matrix condition number η according to the eigenvalue calculated in S32 data , characterizes the excitation characteristics of the sample data, and the calculation formula is as follows: Where λ(F) is the eigenvalue of the Fisher information matrix F, η data is the Fisher information matrix condition number.

3. The method for optimizing identification data of a primary frequency modulation system of a thermal power plant considering excitation characteristics according to claim 1 is characterized in that: The reconstructing of the sample data in step S4 includes: calculating data trend features, data correlation features and combining a feature set including excitation characteristics.

4. The method for optimizing identification data of a primary frequency modulation system of a thermal power plant considering excitation characteristics according to claim 3 is characterized in that: The formula for calculating the trend characteristics of data is: Max X =max(X)-min(X)# Among them, c X For the data sequence X = [x1,…,x n ] is the coefficient of variation, σ X is the standard deviation, μ X is the mean difference; Max X is the range, max(X) and min(X) are the maximum and minimum values ​​in the data sequence X respectively; ΔM is the mode band of the data sequence X, n M is the data sequence X in the interval [μ X -kMax X ,μ X +kMax X ,], k is taken as 0.

5.

5. The method for optimizing identification data of primary frequency modulation system of thermal power plant considering excitation characteristics according to claim 3 is characterized in that: The specific calculation formula of data correlation characteristics is as follows: Among them, r XY To identify the data correlation characteristics between the input and output variables of the system model, cov(X,Y) is the covariance of the variable sequences X and Y, and var(·) is the variance of the corresponding variable sequence.

6. The method for optimizing identification data of a primary frequency modulation system of a thermal power plant considering excitation characteristics according to claim 3 is characterized in that: The combination of the feature set containing the excitation characteristics specifically comprises the following steps: S41. Using a normalization method to perform standard normalization processing on data trend characteristics and data correlation characteristics; S42. Correspond the data trend and correlation feature set corresponding to each group of data samples to the Fisher information matrix condition number, add them to a column, and generate a feature set containing excitation characteristics.

7. The method for optimizing identification data of a primary frequency modulation system of a thermal power plant considering excitation characteristics according to claim 1, characterized in that: The step S5 generates a priori data set using the fuzzy C-means algorithm, comprising the following steps: S51. Define the objective function of the clustering algorithm. The calculation formula is as follows: Where J is the objective function; n is the number of samples; c is the number of cluster centers; m is any real number greater than 1; x i is the i-th sample point; u ij is the membership degree of the jth class center of the ith sample point; c j is the jth class center; ||x i -c j || represents x i to c j The Euclidean distance of S52, selecting a data sample set containing the excitation characteristics as the input of the clustering algorithm, setting the stop iteration threshold and the maximum number of iterations, and minimizing the objective function by continuously iterating the membership degree and the cluster center; and obtaining two clusters of data; S53, by comparing the values ​​of the cluster center points of the two clusters of data in the direction of characterizing the excitation characteristics, the larger cluster of data is classified as suitable identification data 1, and the other cluster is classified as unsuitable identification data 0; S54, making a one-to-one correspondence between a data sample set containing the excitation characteristics and a corresponding classification result to obtain a priori data set.

8. The method for optimizing identification data of a primary frequency modulation system of a thermal power plant considering excitation characteristics according to claim 1, characterized in that: The step S6 generates a classification rule model using a random forest classification algorithm, comprising the following steps: S61, dividing the standardized prior data set into a training set and a validation set according to a ratio of 80% and 20%; S62, initializing and setting the number of basic classifiers of the random forest classification model parameters and the number of iterations of ensemble learning n_estimatiors; S63, taking the data sample set in the training set as input and the classification result as output, performing iterative training to generate a classification rule model; S64, calculating the fitness function according to the number of iterations, the specific calculation formula is as follows, obtaining the optimal number of basic classifiers and the number of iterations of ensemble learning n_estimatiors; Among them, TP and TN represent the number of samples predicted correctly for the positive and negative classes, and FP and FN represent the number of samples predicted incorrectly for the positive and negative classes; S65. Input the validation set data into the generated classification rule model to evaluate the accuracy of the classification rule model.

9. The method for optimizing identification data of a primary frequency modulation system of a thermal power plant considering excitation characteristics according to claim 8, characterized in that: The specific calculation formula for evaluating the accuracy of the model in the classification rule model generated by the random forest classification algorithm is as follows: Recall: Precision: F-score:

10. The method for optimizing identification data of a primary frequency modulation system of a thermal power plant considering excitation characteristics according to claim 6, characterized in that: S41 uses the Z-score method to make it conform to the standard normal distribution, that is, the mean is 0 and the standard deviation is 1. The data processing function is: Where x′ is the normalized data, μ is the sample mean, and σ is the sample standard deviation.