Offshore wind farm unit grouping and flagship unit selection method for farm group control

Abstract

The invention discloses a method for grouping offshore wind farm units and selecting flagship units suitable for farm group control, which mainly relates to the field of wind power technology; the method comprises the following steps: S1, extracting wind turbine unit features; S2, clustering and grouping wind farms based on an improved Gaussian density distance clustering algorithm; S3, selecting flagship units based on correlation analysis; the invention utilizes an improved Gaussian density distance algorithm to group wind turbine units; adopts a correlation analysis method, follows the edge priority and redundancy principles to select flagship units with representativeness and full coverage, and specifically solves the problem of inaccurate measurement of input variables of wind turbine units in offshore wind farms, which is conducive to more accurate wind farm dispatching and control.

Description

Technical Field

[0001] The present invention relates to the field of wind power technology, and in particular to a method for grouping units in an offshore wind farm and selecting a flagship unit suitable for farm group control. Background Art

[0002] Large-scale wind farm systems are generally characterized by high latitude, strong time-varying, and strong nonlinearity, and are highly complex hybrid systems. In order to reduce the problem of increased dimensions due to the number of wind turbines, grouping wind turbines is an effective method. Although grouping can greatly reduce the scale and complexity of the optimization problem, it will also lead to a decrease in accuracy, resulting in a decrease in the optimization efficiency of the final goal or even reverse optimization. Therefore, the grouping of wind turbines needs to take into account both accuracy and complexity.

[0003] The existing methods for grouping wind turbines can be divided into the following three categories:

[0004] The first one is based on the input characteristics of wind turbines, which mainly studies the impact of input wind speed on the grouping of wind turbines. However, this method does not take other factors into consideration and has uncertainty.

[0005] The second method is based on the output characteristics of wind turbines. The main idea is to group wind turbines according to their output power. The principle of this method is simple and easy to implement, and the grouping results are usually certain. However, in theory, it has shortcomings such as the inability to control in real time.

[0006] The third grouping method comprehensively considers the factors of the first two, and groups the wind turbines after a comprehensive analysis of their input and output characteristics. It has the characteristics of accurate grouping and high reliability. However, the more factors are considered and the larger the amount of data, the more difficult the grouping becomes, and a certain grouping algorithm needs to be adopted to implement it.

[0007] Traditional grouping methods for offshore wind farms include grouping methods based on collector lines, grouping methods based on geographical location, grouping methods based on mechanical control system characteristics, etc. All of them have the problems of limited application scope and ignoring wake effects. Summary of the invention

[0008] The purpose of the present invention is to solve the problems existing in the prior art and to provide a method for grouping and selecting flagship units in offshore wind farms suitable for group control. The method comprehensively considers the inlet wind speed, output power, relative position and wake effect between wind turbines, and uses an improved Gaussian density distance algorithm to group wind turbines. The method adopts a correlation analysis method and follows the edge priority and redundancy principles to select flagship units with representativeness and full coverage, which specifically solves the problem of inaccurate measurement of input variables of wind turbines in offshore wind farms, and is conducive to more accurate wind farm scheduling and control.

[0009] In order to achieve the above object, the present invention is implemented through the following technical solutions:

[0010] The method for grouping and selecting flagship units of offshore wind farms applicable to farm group control comprises the following steps:

[0011] S1, wind turbine feature extraction;

[0012] S2. Clustering and grouping wind farms based on improved Gaussian density distance clustering algorithm;

[0013] S3. Select the flagship unit based on correlation analysis.

[0014] Preferably, in step S1, the wind speed at the wind inlet, the output power and the relative position of the unit are selected as clustering inputs.

[0015] Preferably, step S2 comprises the steps of:

[0016] S21, data preprocessing, data preprocessing includes screening and dimensionality reduction;

[0017] S22, calculating clustering algorithm related data, the related data including input matrix, mean and deviation vector of input matrix, Gaussian matrix, additive Gaussian vector, Gaussian mean vector, Gaussian deviation vector, distance matrix, distance mean vector, distance deviation vector;

[0018] S23, determining the cluster center point, by finding the maximum value of GPM from the unclustered data, and designating the point with the highest density as the cluster center;

[0019] S24, after the cluster center is determined, find all cluster members of the class;

[0020] S25, defining cluster member neighboring points and updating thresholds using the neighboring sample list and the neighbor search list;

[0021] S26. Grouping quality evaluation.

[0022] Preferably, in step S24, if the S between the two wind turbines overlap >0, that is, there is coupling, S overlap The calculation formula is:

[0023]

[0024]

[0025] Among them, r1 is the wake radius generated by the upwind wind turbine at the downwind wind turbine, r2 is the rotor diameter of the downwind wind turbine, and d is the distance between the two centers.

[0026] Preferably, the step S25 includes:

[0027] The neighbor sample list is formed using all samples in the space, and the neighbor search list contains only samples in the unclustered list;

[0028] Initially, the neighbor search list only stores the center point of all data, while the adjacent sample list stores the samples of each neighbor search list member. These samples are located in a circle with a radius of a fixed distance threshold, and there is no coupling between any samples. The fixed distance threshold uses the Euclidean distance, and the calculation formula is:

[0029]

[0030] The adjacent sample list determines the gradient density and distance changes in the region through the gradient distance threshold and Gaussian gradient density threshold;

[0031] The calculation formula of the gradient distance threshold is:

[0032]

[0033] Where, μ(GPD) is the average of all GPDs in the dataset;

[0034] The calculation formula of Gaussian gradient density threshold is:

[0035] GGDT = σ(SPLvariances);

[0036] If the GDT of the data of the neighborhood sample is too scattered, the sample point is excluded;

[0037]

[0038] Clustering criteria are introduced, including density criteria and distance criteria. In the density criteria, it is checked whether the Gaussian value of the sample is greater than or equal to the difference between FGDT and GGDT; in the distance criteria, it is checked whether the Euclidean distance of a sample is less than or equal to the sum of FDT and GDT; if the density criteria and distance criteria are met, the sample is included in the cluster;

[0039] The neighbor search list is then updated recursively.

[0040] Preferably, in step S26: the verification of the number of groups includes verifying the number of groups based on the silhouette coefficient and verifying the number of groups based on the Davies-Bouldin Index.

[0041] The method for verifying the number of groups based on the silhouette coefficient is:

[0042] Assuming the number of wind turbines is n, the contour coefficient expression is:

[0043]

[0044] Among them, the sample contour coefficient of the i-th wind turbine is:

[0045]

[0046]

[0047]

[0048] Assume that wind turbine i is clustered into group c, a(i) represents the average distance between wind turbine i and all other wind turbines in group c, and b(i) represents the minimum value of the average distance between wind turbine group i and all wind turbines in groups other than group c;

[0049] Among them, s i ∈[-1,1], when s i The closer it is to 1, the more reasonable the clustering of wind turbines is. i When it is small or even negative, it means that the clustering result of wind turbines is unreasonable;

[0050] The method for verifying the number of groups based on Davies-Bouldin Index is:

[0051] Define a value S representing the dispersion i , which indicates the dispersion of data points in the i-th category:

[0052]

[0053] Among them, T i represents the number of wind turbines in the i-th group, X j represents the jth wind turbine in the i-th group, A i represents the center of the i-th group; the value of q is 1 or 2. When q is 1, it represents the mean of the cluster from each point to the center; when q is 2, it represents the standard deviation of the distance from each point to the center;

[0054] Define a distance value M i,j , represents the distance between the i-th category and the j-th category:

[0055]

[0056] Define a similarity measure to indicate the similarity between the i-th category and the j-th category:

[0057]

[0058] For different clustering groups, take the maximum value of the similarity measure D i , D i =max j≠iR i,j ;

[0059] Finally, the mean value of the maximum similarity of all clusters is taken to obtain the DBI index:

[0060]

[0061] Preferably, the step S3 includes the steps:

[0062] S31. Formulate the selection principle of the flagship unit. The selection of the flagship machine needs to meet the representativeness principle, the edge priority principle, the full coverage principle, and the redundancy principle;

[0063] S32. Cluster and group the main wind directions, and select the flagship units through correlation analysis, with the number being N;

[0064] S33. Cluster and group the other wind directions, with the number being M;

[0065] S34. Comprehensively select the flagship units. When M = N, cluster and group the other wind directions with the flagship machine as the clustering center; when M < N, conduct correlation analysis on the N flagship units in the wind direction sector, and then select the first M units as the clustering center to conduct cluster grouping under the principle of edge priority.

[0066] Preferably, in the step S31, the correlation analysis method based on the Pearson correlation coefficient is used to select the flagship machine, and the calculation formula is:

[0067]

[0068] where n is the number of samples, and are the mean values of x i and y i respectively;

[0069] The representativeness principle means that the flagship unit is the most representative unit in the group, and the representativeness is defined by the average Pearson coefficient; the edge priority principle means that the wind turbines at the edge of the wind farm are superior to those inside the wind farm, because under different wind directions, the wind turbines at the edge of the wind farm are more likely to be unaffected by the wake effect and the measurement data is more reliable; the full coverage principle means that the role of the flagship machine needs to cover all wind directions; the redundancy principle means that the number of flagship machines ≥ the number of groups, and there is at least one flagship unit in each cluster to ensure that there are flagship machines in all clusters of the entire wind farm.

[0070] Preferably, in step S32, based on the historical wind measurement data of the wind farm, the wind turbines are clustered and grouped in the wind direction with the highest wind frequency, a correlation analysis is performed on the turbines in each group, the correlation analysis is sorted, and the top three wind turbines with the highest correlation coefficients are selected. Among these three wind turbines, if there is a marginal wind turbine, the marginal wind turbine is selected as the flagship unit; if not, the wind turbine with the highest correlation coefficient is selected as the flagship unit; the flagship unit selected at this time meets the representativeness principle and the edge priority principle.

[0071] Preferably, in step S33, for the full coverage principle and the redundancy principle, the number of flagship machines selected in the main wind direction is defined as N, and in other wind direction sectors, the contour coefficient and DBI in step S26 are used to perform cluster number analysis in the range of [2, N] to obtain the optimal cluster number M.

[0072] Compared with the prior art, the beneficial effects of the present invention are:

[0073] 1. The present invention comprehensively considers the wind speed, output power and relative position of the wind turbines, retains the operating characteristics of the wind farm to the greatest extent, and regards the wake effect as a coupling factor of the wind turbines. The influence of the wake effect is reduced during the grouping process, and the wind turbines are grouped using an improved Gaussian density distance clustering algorithm.

[0074] 2. To address the problem of inaccurate measurement of input variables of wind turbines in offshore wind farms, the present invention adopts a correlation analysis method, follows representativeness, edge priority, full coverage, and redundancy, and selects flagship units that can fully cover wind directions and are representative.

[0075] 3. The present invention overcomes the problem that the traditional grouping method only considers input characteristics or output characteristics and does not consider the impact of wake effect. It comprehensively considers the input and output characteristics of wind turbines and wake effect, and has the characteristics of accurate grouping and high reliability. At the same time, it introduces the concept of flagship units and selects the most representative wind turbines as flagship units, which specifically solves the problem of inaccurate input variable measurement of wind turbines in offshore wind farms, which is conducive to more accurate wind farm scheduling and control. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 It is a flow chart of the present invention

[0077] Figure 2 It is a flow chart of the present invention using the improved Gaussian density distance clustering algorithm to cluster and group wind farms;

[0078] Figure 3 It is a schematic diagram of the wake relationship of a wind farm;

[0079] Figure 4 It is a flow chart of the present invention using correlation analysis to select a flagship phone. DETAILED DESCRIPTION

[0080] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the application equally.

[0081] Grouping wind turbines requires accurate measurement of their variables, but at this stage, there are still large errors in the measurement of wind turbines. Flagship machines can effectively solve the problem of inaccurate measurements. The so-called flagship machine means that when building a wind farm, a large number of high-precision, low-latency equipment, such as lidar, communication devices, etc., are installed on some wind turbines in the wind farm to ensure the accuracy of measurement data and the timeliness of data transmission. Due to the huge economic cost of flagship machines, only a small number of flagship units can be installed.

[0082] Example: As attached Figure 1-4 As shown, the present invention is a method for grouping offshore wind farm units and selecting flagship units suitable for farm group control. The method comprehensively considers the wind speed, output power, relative position and wake effect between wind turbines, and uses an improved Gaussian density distance algorithm to group wind turbines. The correlation analysis method is used to select flagship units with representativeness and full coverage in accordance with the edge priority and redundancy principles.

[0083] Includes steps:

[0084] S1. Wind turbine feature extraction.

[0085] The most representative features are extracted and selected from the original data, and the extracted and selected features are used as the input of the clustering algorithm. Considering the wake effect and retaining the wind farm operation characteristics to the greatest extent, the inflow wind speed u, v, output power P, and position coordinates (x, y) of the wind turbine are selected as the input of the clustering. That is, the matrix X = [u, v, P, x, y].

[0086] S2. Clustering and grouping offshore wind farms based on an improved Gaussian density distance clustering algorithm, specifically including the following steps:

[0087] S21, data preprocessing;

[0088] The tasks performed in the preprocessing phase are as follows:

[0089] (1) Screening: Scan all the data in the data set. If any of the data has missing attribute information, it will be deleted from the data set.

[0090] (2) Dimensionality reduction: If the variance of a dimension vector is zero, then remove that dimension from the input dataset.

[0091] S22, calculating clustering algorithm related data;

[0092] Various data structures such as matrices and vectors are used in the calculation process. These data structures are as follows:

[0093] (1) Input matrix X n×d :

[0094]

[0095] Wherein, n represents the number of samples in the data set, and d represents the attribute dimension of the sample data. Specifically in this embodiment, n represents the number of wind turbines in the offshore wind farm, and d represents the dimension of the clustering input matrix X.

[0096] (2) The mean μ of the input matrix d With the deviation vector σ d

[0097] For the input matrix X n×d Calculate μ for each column d and σ d The values ​​of are required for the subsequent preprocessing of the input data and the calculation of the Gaussian distribution.

[0098] (3) Gauss Matrix (GM)

[0099] GM n×n Contains the discrete reciprocal effect of every two points in the data set. It is a non-negative, symmetric matrix whose diagonal contains the maximum Gaussian value of each point in the data set. The multidimensional Gaussian value is calculated as follows:

[0100]

[0101] (4) Additive Gaussian vector (AG), AG n is the matrix GM n×n The row summation of gives:

[0102]

[0103] (5) Gaussian Point Mean vector (GPM), GPM n is through the GM n×n The sum of all rows in the matrix except the diagonal elements is divided by (n-1) to obtain:

[0104]

[0105] (6) Gaussian Point Deviation vector (GPD), calculated using GPM in equation (3.4) j Value, calculate GM n×n The standard deviation of each row of the matrix is ​​used to obtain the GPD n .

[0106] (7) Distance matrix (DM). The elements of the distance matrix are the Euclidean distances between points. The matrix is ​​a non-negative symmetric matrix with zero diagonals. The calculation formula is:

[0107]

[0108] (8) Distance Point Mean vector (DPM), DPM n By using DM n×n The sum of all rows in the matrix except the diagonal elements is divided by (n-1) to obtain:

[0109]

[0110] (9) Distance Point Deviation vector (DPD), by using DPM j Value, calculate DM n×n The standard deviation of each row of the matrix is ​​used to obtain DPD n .

[0111] S23. Determine the cluster center point by finding the maximum value of GPM from the unclustered data and designate the point with the highest density as the cluster center.

[0112] S24, after the cluster center is determined, find all cluster members of the class; the specific steps are as follows:

[0113] Determine whether there is coupling between the member and the cluster center and members. The coupling between wind turbines is manifested by the wake effect. When a wind turbine is within the wake influence range of the upwind unit, it indicates that there is coupling between the two wind turbines. Figure 3 , if the S between two wind turbines overlap >0, that is, there is coupling. The calculation formula is:

[0114]

[0115]

[0116] Among them, r1 is the wake radius generated by the upwind wind turbine at the downwind wind turbine, r2 is the rotor diameter of the downwind wind turbine, and d is the distance between the two centers.

[0117] S25. Define the cluster member proximity points and update thresholds using the Samples in Proximity List (SPL) and Neighbor Search List (NSL). The specific steps are as follows:

[0118] SPL is formed using all samples in the space, while NSL contains only samples in the unclustered list. Initially, NSL only saves the center point of all data, while SPL saves samples of each NSL member, which are located within a circle with a radius of a fixed distance threshold (Fixed Distance Threshold, FDT), and there is no coupling between any samples. FDT uses Euclidean distance, calculated as:

[0119]

[0120] SPL determines the gradient density and distance change in the area through GDT and GGDT. The calculation formulas of GDT and GGDT are:

[0121]

[0122] GGDT = σ(SPLvariances);

[0123] Here, μ(GPD) is the average of all GPDs in the dataset.

[0124] FDT is used to determine clusters and cluster members. When a sample is in the neighborhood of a cluster, it should satisfy the Gaussian threshold equation and the Gradient Distance Threshold (GDT). GDT and Gaussian Gradient Density Thresholds (GGDT) represent the change in variance. If the value of the neighborhood sample meets the threshold conditions, these threshold equations are used to keep the neighborhood sample in a cluster.

[0125] Since each point in the neighborhood is used as a cluster center, a relative dynamic threshold is required to measure the convergence or divergence of the neighborhood samples. If the GDT of the neighborhood sample data is too scattered, the sample point is excluded.

[0126]

[0127] Clusters may have different shapes and connectivity. Therefore, the gradient threshold equation is introduced to represent clustering situations with arbitrary shapes and density changes. The calculation of the gradient threshold includes two steps:

[0128] 1) Calculate the variance of the current sample in the neighborhood;

[0129] 2) Calculate the mean and deviation of the previously calculated variance.

[0130] Each time a cluster is scanned, these dynamic values ​​are used to add samples to the cluster, thereby completing the inclusion of the neighborhood. GDD defines which sample belongs to which cluster. Two criteria must be met in this process: the density criterion (criterion A) and the distance criterion (criterion B). In the density criterion, check whether the Gaussian value of the sample is greater than or equal to the difference between FGDT and GGDT. In the distance criterion, check whether the Euclidean distance of a sample is less than or equal to the sum of FDT and GDT. If criteria A and B are met, the sample is included in the cluster. Criteria A and Criteria B are as follows:

[0131] Criterion A: Gauss.(centr.,s)≥FGDT-GGDT;

[0132] Criterion B: Dist.(centr.,s)≤FDT+GDT;

[0133] NSL is updated recursively. For example, in the first loop, only the center point is clustered into a cluster and saved in NSL. In the second loop, the neighborhood samples of the center point are added to NSL, the cluster is updated, and so on. The samples newly added to NSL form a whole with the original points in NSL as the center point of the next step, so that the neighborhood continues to expand until there are no samples that meet the requirements. When all samples are clustered, the algorithm stops automatically.

[0134] S26, grouping quality evaluation;

[0135] Verify the number of groups based on the silhouette coefficient. Assuming the number of wind turbines is n, the silhouette coefficient expression is:

[0136]

[0137] Among them, the sample contour coefficient of the i-th wind turbine is:

[0138]

[0139]

[0140]

[0141] Assume that wind turbine i is clustered into group c, a(i) represents the average distance between wind turbine i and all other wind turbines in group c, and b(i) represents the minimum value of the average distance between wind turbine group i and all wind turbines in groups other than group c. i ∈[-1,1], when s i The closer it is to 1, the more reasonable the clustering of wind turbines is. i When it is small or even negative, it means that the clustering result of wind turbines is unreasonable.

[0142] Verify the number of groups based on the Davies-Bouldin Index. Define a value S that represents the dispersion i , which indicates the dispersion of data points in the i-th category:

[0143]

[0144] Among them, T i represents the number of wind turbines in the i-th group, X j represents the jth wind turbine in the i-th group, A i The value of q is 1 or 2. When q is 1, it represents the mean of the cluster from each point to the center. When q is 2, it represents the standard deviation of the distance from each point to the center. Both can be used to measure the degree of dispersion. In this embodiment, q=2 is taken here.

[0145] Define a distance value M i,j , represents the distance between the i-th category and the j-th category:

[0146]

[0147] Define a similarity measure to indicate the similarity between the i-th category and the j-th category:

[0148]

[0149] For different clustering groups, take the maximum value of the similarity measure D i , D i =max j≠i R i,j . That is, the maximum value of the similarity between the i-th group and other groups. Finally, the average of the maximum similarities of all clusters is used to obtain the DBI index:

[0150]

[0151] is the average distance between the centers of all wind turbines in the i-th group.

[0152] S3. Select flagship phones based on clustering results. The specific steps are as follows:

[0153] S31. Principles for selecting flagship units;

[0154] The present invention adopts the correlation analysis method based on Pearson correlation coefficient to select the flagship machine, and the calculation formula is:

[0155]

[0156] Where n is the number of samples, and x i and i The mean of .

[0157] The selection of flagship phones needs to meet the following requirements:

[0158] (1) Representativeness principle: The flagship unit is the most representative unit in the group, and representativeness is defined by the average Pearson coefficient;

[0159] (2) Edge priority principle: Wind turbines at the edge of a wind farm are better than wind turbines inside the wind farm because wind turbines at the edge of a wind farm are less likely to be affected by the wake effect under different wind directions, and the measurement data are more reliable.

[0160] (3) Full coverage principle: The flagship device needs to cover all wind directions;

[0161] (4) Redundancy principle: the number of flagship units ≥ the number of clusters. There must be at least one flagship unit in each cluster to ensure that there are flagship units in all clusters of the entire wind farm.

[0162] S32, clustering and grouping of main wind direction, correlation analysis, selecting flagship units, the number is N;

[0163] According to the historical wind measurement data of the wind farm, the wind turbines are clustered and grouped in the wind direction with the highest wind frequency (main wind direction), and the units in each group are subjected to correlation analysis. The correlation analysis is sorted and the top three wind turbines with the highest correlation coefficients are selected. Among these three wind turbines, if there is a marginal wind turbine, the marginal wind turbine is selected as the flagship unit; if not, the wind turbine with the highest correlation coefficient is selected as the flagship unit. The flagship unit selected at this time meets the principle (1) representative principle and the principle (2) marginal priority principle.

[0164] S33, other wind direction clustering groups, the number is M;

[0165] Based on historical wind measurement data, an offshore wind farm can obtain a wind rose diagram, which divides the wind direction into several sectors according to sectors, and each sector contains the same range of wind direction angles. For principle (3) full coverage principle and principle (4) redundancy principle, the number of flagship machines selected in the main wind direction is defined as N, and in other wind direction sectors, the silhouette coefficient and DBI described in S26 are used to perform cluster number analysis in the range of [2, N] to obtain the optimal number of clusters as M.

[0166] S34, comprehensive selection of flagship units;

[0167] When M=N, the flagship aircraft is used as the cluster center to cluster other wind directions; when M

Claims

1. A method for grouping and selecting flagship units in an offshore wind farm suitable for farm group control, characterized in that: Includes steps: S1, select the wind speed, output power and relative position of the unit as clustering input; S2. Clustering and grouping wind farms based on improved Gaussian density distance clustering algorithm; The step S2 comprises the steps of: S21, data preprocessing, data preprocessing includes screening and dimensionality reduction; S22, calculating clustering algorithm related data, the related data including input matrix, mean and deviation vector of input matrix, Gaussian matrix, additive Gaussian vector, Gaussian mean vector, Gaussian deviation vector, distance matrix, distance mean vector, distance deviation vector; S23, determining the cluster center point, by finding the maximum value of GPM from the unclustered data, and designating the point with the highest density as the cluster center; wherein GPM is the Gaussian mean vector; S24, after the cluster center is determined, find all cluster members of the class; in the step S24, if the S between the two wind turbines overlap >0, that is, there is coupling, S overlap The calculation formula is: Among them, r1 is the wake radius generated by the upwind wind turbine at the downwind wind turbine, r2 is the rotor diameter of the downwind wind turbine, and d is the distance between the two centers; S25, using the adjacent sample list and the neighbor search list to define the cluster member adjacent points and update the threshold; the step S25 includes: The neighbor sample list is formed using all samples in the space, and the neighbor search list contains only samples in the unclustered list; Initially, the neighbor search list only stores the center point of all data, while the adjacent sample list stores the samples of each neighbor search list member. These samples are located in a circle with a radius of a fixed distance threshold, and there is no coupling between any samples. The fixed distance threshold uses the Euclidean distance, and the calculation formula is: Among them, FDT is the fixed distance threshold, DPM is the distance mean vector, GPM is the Gaussian mean vector, DPD is the distance deviation vector, GPD represents the Gaussian deviation vector, and Centr represents the center The adjacent sample list determines the gradient density and distance changes in the region through the gradient distance threshold and Gaussian gradient density threshold; The calculation formula of the gradient distance threshold is: Among them, μ(GPD) is the average of all GPDs in the data set, NSL represents the neighbor search list, SPL represents the adjacent sample list, AG(NSL) represents the additive Gaussian vector of the neighbor search list, AG(centr.) represents the additive Gaussian vector, and GPD(NSL) represents the Gaussian deviation vector of the neighbor search list; The calculation formula of Gaussian gradient density threshold is: GGDT = σ(SPLvariances); If the GDT of the data of the neighborhood sample is too scattered, the sample point is excluded; Clustering criteria are introduced, including density criteria and distance criteria. In the density criteria, it is checked whether the Gaussian value of the sample is greater than or equal to the difference between FGDT and GGDT, where FGDT represents the fixed Gaussian density threshold and GGDT represents the Gaussian gradient density threshold; in the distance criteria, it is checked whether the Euclidean distance of a sample is less than or equal to the sum of FDT and GDT; if the density criteria and distance criteria are met, the sample is included in the cluster; Then the neighbor search list is updated recursively; S26, grouping quality evaluation; in step S26: verification of the number of groups includes verification of the number of groups based on the silhouette coefficient and verification of the number of groups based on the Davies-Bouldin Index. The method for verifying the number of groups based on the silhouette coefficient is as follows: Assume the number of wind turbines is n, then the expression of the silhouette coefficient is: Among them, the sample silhouette coefficient of the i-th wind turbine is: Assume that after clustering, wind turbine i is divided into c groups. a(i) represents the average distance between wind turbine i and all other wind turbines belonging to the same c group, and b(i) represents the minimum value of the average distances between wind turbine i and all wind turbines in each group that is not in group c; Among them, s i ∈[-1,1], when s i The closer it is to 1, the more reasonable the clustering of wind turbines is. i When it is small or even negative, it means that the clustering result of wind turbines is unreasonable; The method for verifying the number of groups based on the Davies-Bouldin Index is as follows: Define a value S representing the dispersion i , which indicates the dispersion of data points in the i-th category: Among them, T i represents the number of wind turbines in the i-th group, X j represents the data point characterizing the j-th wind turbine in the i-th group, A i represents the center point of the i-th group of data; the value of q is 1 or 2. When q is 1, it represents the mean of the distance from each point to the center cluster; when q is 2, it represents the standard deviation of the distance from each point to the center; |X j -A i | represents the data point X representing the jth wind turbine in the i-th group j To the center point A of the i-th group of data i distance; Define a distance value M i,j , represents the distance between the i-th category and the j-th category: Define the similarity measurement index, which represents the similarity degree between the i-th class and the j-th class: For different clustering groups, take the maximum value of the similarity measure D i , D i =max j≠i R i,j ; Finally, taking the average of the maximum similarities of all clusters gives the DBI index: DBI stands for Davies-Bouldin Index; S3. Select the flagship units based on correlation analysis; The step S3 includes the following steps: S31. Formulate the principles for selecting flagship units. The selection of flagship units needs to meet the representativeness principle, the edge priority principle, the full coverage principle, and the redundancy principle; in the step S31, the correlation analysis method based on the Pearson correlation coefficient is used to select the flagship units, and the calculation formula is: Where n is the number of samples, and x i and i The mean of The representativeness principle means that the flagship unit is the most representative unit in the group, and the representativeness is defined using the average Pearson coefficient; the edge priority principle means that the wind turbines at the edge of the wind farm are superior to those inside the wind farm because, under different wind directions, the wind turbines at the edge of the wind farm are more likely to be unaffected by the wake effect and the measurement data is more reliable; the full coverage principle means that the role of the flagship unit needs to cover all wind directions; the redundancy principle means that the number of flagship units ≥ the number of groups, and there is at least one flagship unit in each cluster to ensure that there are flagship units in all clusters of the entire wind farm; S32. Cluster and group the main wind direction, and select the flagship units through correlation analysis, with the number being N; S33. Cluster and group other wind directions, with the number being M; S34. Comprehensively select the flagship units. When M = N, use the flagship units as the clustering centers to cluster and group other wind directions; when M < N, conduct correlation analysis on the N flagship units in this wind direction sector, and then select the first M units as the clustering centers to cluster and group under the principle of edge priority.

2. The offshore wind farm unit grouping and flagship unit selection method suitable for farm group control according to claim 1 is characterized in that: In the step S32, according to the historical wind measurement data of the wind farm, cluster and group the wind turbines in the wind direction with the highest wind frequency, conduct correlation analysis on the units in each group, sort the correlation analysis results, and select the top three wind turbines with the highest correlation coefficients. Among these three wind turbines, if there are edge wind turbines, then select the edge wind turbine as the flagship unit; if not, then select the wind turbine with the highest correlation coefficient as the flagship unit; the flagship units selected at this time meet the representativeness principle and the edge priority principle.

3. The offshore wind farm unit grouping and flagship unit selection method suitable for farm group control according to claim 1 is characterized in that: In the step S33, for the full coverage principle and the redundancy principle, define the number of flagship units selected in the main wind direction as N. In other wind direction sectors, use the silhouette coefficient and DBI described in step S26 to analyze the clustering number in the range of [2, N] to obtain the optimal clustering number M. DBI represents the Davies-Bouldin Index.

Images