Fan wind power data cleaning method

Abstract

The invention discloses a method for cleaning wind power data of a fan, which comprises the following steps of: calculating quantiles and quantile distances by applying a quartile method, and eliminating sparse dispersion points of the wind power data; according to the quartile and the quantile distance, determining an abnormal data boundary, and drawing a box diagram; by means of the similarity of the s gmoid function and a wind power curve, the image change condition of the s gmoid function after different parameters are added is analyzed, the distribution condition of abnormal points of a wind speed-power scatter diagram is combined, a s gmoid transformation model based on four parameters is established, and the abnormal points are removed; establishing an objective function based on an s igmoid transformation model, and solving parameters by applying an LM optimization algorithm until the objective function value is smaller than the allowable maximum deviation, so as to obtain four parameters, and further fitting a wind power curve; and obtaining a boundary curve of the abnormal scatter points by changing the four parameter values, identifying and removing the residual abnormal scatter points, and finally obtaining effective wind power scatter points to complete wind power data cleaning. According to the method, the wind power abnormal data cleaning quality is improved.

Description

Technical Field

[0001] The invention relates to the technical field of wind power data cleaning of a wind turbine set, and in particular to a method for cleaning wind power data of a wind turbine set. Background Art

[0002] Wind power data is an important performance indicator of wind turbines, reflecting the power generation performance and operating status of wind turbines. Data mining can provide important support for the inspection and maintenance of wind turbines and the operation and management of wind farms. Under ideal conditions, wind power data is relatively regular. At rated wind speed, the power increases with the increase of wind speed. After reaching the rated wind speed, the power is constant and there is no abnormal data. The wind speed-power curve can be directly fitted according to the IEC61400-12-1 standard. Actual wind power data often has abnormal scattered points. These abnormal scattered points are directly related to factors such as data collection, transmission, measurement problems, and substandard wind turbine power generation performance. If they are not identified and cleaned, it will affect the fitting of the wind power curve, resulting in inaccurate results such as power coefficient and power generation, which will have a great impact on power production, wind turbine power generation performance evaluation, scheduling, and later operation and maintenance.

[0003] Therefore, a wind power data cleaning method is proposed. Summary of the invention

[0004] The purpose of the present invention is to overcome the existing defects and provide a wind turbine wind power data cleaning method, thereby improving the cleaning quality of abnormal wind power data.

[0005] The technical solution to achieve the above purpose is:

[0006] A method for cleaning wind power data of a wind turbine, comprising:

[0007] Step S1, obtaining annual SCADA (Supervisory and Control System for Data Acquisition) operation data of the wind turbine generator set;

[0008] Step S2, extracting timestamp, wind speed and power data from SCADA operation data;

[0009] Step S3, drawing a wind speed-power scatter plot, and analyzing the distribution of abnormal points in the wind speed-power scatter plot;

[0010] Step S4, applying the quartile method, calculating the quantile and quantile range, determining the data elimination boundary, and eliminating sparse scattered points of wind power data;

[0011] Step S5, determining the abnormal data boundary through quartiles and quantile distances, and drawing a box plot;

[0012] Step S6, using the similarity between the sigmoid function and the wind power curve, analyzing the change of the sigmoid function image after adding different parameters, combining the distribution of abnormal points in the wind speed-power scatter plot, establishing a sigmoid transformation model based on four parameters, and removing abnormal points;

[0013] Step S7, based on the sigmoid transformation model, establish the objective function, and apply the LM (Levenberg-Marquardt) optimization algorithm to solve the parameters until the objective function value is less than the maximum allowable deviation, obtain four parameters, and then fit the wind power curve;

[0014] Step S8, by changing the four parameter values ​​to obtain the boundary curve of the abnormal scatter points, the remaining abnormal scatter points are identified and removed, and finally the effective wind power scatter points are obtained to complete the wind power data cleaning.

[0015] Preferably, in step S3, the abnormal point data of the wind speed-power scatter plot include: abnormal data caused by unplanned shutdown and planned shutdown, abnormal data caused by wind power curtailment or unit capacity reduction, abnormal data caused by signal propagation error, sensor failure, aging of wind turbine components, poor power generation performance or extreme weather, and abnormal data caused by failure of anemometer or communication error.

[0016] Preferably, in step S4, the quartile method is applied to calculate the quantile and the quantile range, including:

[0017] Let the first quantile be Q1, the second quantile be Q2, the third quantile be Q3, and the interquartile range be I QR ,

[0018] If a set of ascending sequence is x={x1,x2,x3,...x m}, then calculate Q1, Q2, Q3 and I QR The formula is as follows:

[0019]

[0020]

[0021]

[0022] I QR =Q3-Q1.

[0023] Preferably, in step S5, the outliers in the data set are determined by calculating the quantile and quantile range, and the outliers are less than Q1-1.5I QR Or greater than Q3+1.5I QR The data points are considered as outliers and a box plot is drawn.

[0024] Preferably, in step S6, the Sigmoid function is a commonly used binary classification function, and the function form is:

[0025]

[0026] The curve changes after adding d, c, b and a parameters to the Sigmoid function are as follows:

[0027] If the function is transformed as follows:

[0028]

[0029] The d value reflects the shift of the function graph on the x-axis. As the d value increases, the function shifts to the right, and as the d value decreases, the function shifts to the left.

[0030] If the function is transformed as follows:

[0031]

[0032] Among them, the c value reflects the edge inclination of the function image. The larger the c value, the steeper the image.

[0033] If the function is transformed as follows:

[0034]

[0035] Among them, the b value reflects the difference between the maximum and minimum values ​​of the function graph;

[0036] If the function is transformed as follows:

[0037]

[0038] Among them, the a value reflects the difference between the maximum and minimum values ​​of the function graph;

[0039] Through the analysis of the above parameters, the mathematical model based on the sigmoid function is established as follows, and the initial values ​​of a, b, c and d are set to obtain the sigmoid transformation model based on four parameters, namely:

[0040]

[0041] Preferably, in step S7, θ=[a, b, c, d] is set as the parameter to be determined, j i (θ) = y i -f(θ;x i ), the target formula is:

[0042]

[0043] In the formula, j i(θ) is the deviation function between the actual power value of the wind turbine and the power value of the wind turbine obtained by the sigmoid transformation model, f(θ; x i ) is the wind turbine power value required by the sigmoid transformation model, y i is the actual value of wind turbine power, m is the number of wind power data points, x i is the wind speed value of the wind power data point;

[0044] To minimize P(θ), the first step is to derive P(θ), set P′(θ) = 0, and perform Taylor expansion to obtain:

[0045] P′(θ+Δθ)≈P′(θ)+P″(θ)Δθ=0;

[0046] Where P′(θ) is the first derivative of the target formula, and P″(θ) is the second derivative of the target formula;

[0047] Then we get:

[0048] Δθ=-(P″(θ)) -1 P′(θ);

[0049] After two derivations, the Jacobian matrix K of θ with respect to j is obtained as follows:

[0050]

[0051] Wherein, j(θ) is the deviation function between the actual power value of the wind turbine and the power value of the wind turbine obtained by the sigmoid transformation model, j′(θ) is the differential matrix of the deviation function between the actual power value of the wind turbine and the power value of the wind turbine obtained by the sigmoid transformation model with respect to the parameters [a, b, c, d];

[0052] Gradient information g(θ) of P:

[0053]

[0054] In the formula, T is;

[0055] Then, the Hessian matrix of P is:

[0056]

[0057] Δθ=-(P″(θ))-1P′(θ)=-(K T K+S(θ) -1 K T j(θ);

[0058] In the formula, S(θ) is the high-order term of the Gauss-Newton method, j(θ) is the deviation function between the actual power value of the wind turbine and the power value of the wind turbine obtained by the sigmoid transformation model;

[0059] Among them, the amount of calculation of S(θ) is usually large, and the Gauss-Newton method generally ignores this term, and we get:

[0060] Δθ=-(P″(θ)) -1 P′(θ)=-(K T K) -1 K T j(θ);

[0061] In order to prevent the Gauss-Newton method from requiring full rank to solve the Jacobian matrix and causing instability or overfitting, P(θ) is added to the L2 regularization penalty term, and the result is:

[0062]

[0063] Δθ=-(K T K+βI) -1 K T j(θ);

[0064] Where β is the regularization parameter and I is the unit matrix;

[0065] It can be seen from the above formula that when β is large, the descent gradient is close to the steepest descent method, and when it is very small, it is close to the Gauss-Newton method. In order to obtain a suitable β, it is necessary to define an evaluation parameter ∈, let z = Δθ, then:

[0066]

[0067] That is, the closer ∈ is to 1, the more accurate the overall Taylor expansion is. If ∈ is less than 0, adjust the step size β, and continuously adjust the value of θ through gradient calculation to make P(θ) less than the set error value, thereby completing the four-parameter fitting of the wind power curve.

[0068] The beneficial effects of the present invention are as follows: first, the quartile method is used to identify and clean the sparse data of wind power, and a transformation model based on the sigmoid function is established, that is, a four-parameter model of the sigmoid function is adopted, a sigmoid curve fitting objective function based on four parameters is established, the LM optimization algorithm is applied to solve the parameters, and the maximum allowable deviation is set at the same time, and the four parameters of the model are optimized, so that an accurate wind power fitting curve can be obtained, and the remaining abnormal scattered points are eliminated by changing the values ​​of the four parameters, thereby realizing the cleaning of abnormal scattered points of wind power data. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is a flow chart of a method for cleaning wind power data of a wind turbine according to the present invention;

[0070] Figure 2 It is a typical distribution diagram of wind speed-power scatter points in the present invention;

[0071] Figure 3 It is the quartile box plot in the present invention;

[0072] Figure 4 is a scatter plot of the first group of wind power before cleaning by the quartile method in an embodiment of the present invention;

[0073] Figure 5 is the first group of wind power scatter plots after being cleaned by the quartile method in the embodiment of the present invention;

[0074] Figure 6 is a scatter plot of wind power of the second group before cleaning by the quartile method in an embodiment of the present invention;

[0075] Figure 7 is a second group of wind power scatter plots after being cleaned by the quartile method in an embodiment of the present invention;

[0076] Figure 8 is a scatter plot of wind power of the third group before cleaning by the quartile method in an embodiment of the present invention;

[0077] Fig. 9 is the third group of wind power scatter plots after being cleaned by the quartile method in the embodiment of the present invention;

[0078] Fig.10 is a curve change diagram after adding parameter d to the basic simoid function in an embodiment of the present invention;

[0079] Fig.11 is a curve change diagram after adding parameter c to the basic simoid function in an embodiment of the present invention;

[0080] Fig.12 is a curve change diagram after adding parameter b to the basic simoid function in an embodiment of the present invention;

[0081] Fig.13 is a curve change diagram after adding parameter a to the basic simoid function in an embodiment of the present invention;

[0082] Fig.14 is a first group of power curves fitted by a sigmoid transformation model in an embodiment of the present invention;

[0083] Fig.15 This is a schematic diagram of the first group after cleaning by the sigmoid transformation model in an embodiment of the present invention;

[0084] Fig.16 is a second group of power curves fitted by a sigmoid transformation model in an embodiment of the present invention;

[0085] Fig.17This is a schematic diagram of the second group after cleaning by the sigmoid transformation model in an embodiment of the present invention;

[0086] Fig.18 is a third group of power curves fitted by a sigmoid transformation model in an embodiment of the present invention;

[0087] Fig.19 This is a schematic diagram of the third group after cleaning by the sigmoid transformation model in an embodiment of the present invention. DETAILED DESCRIPTION

[0088] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. In the description of the present invention, it should be noted that the terms "center", "up", "down", "left", "right", "vertical", "horizontal", "inside", "outside" and the like indicate directions or positional relationships based on the directions or positional relationships shown in the accompanying drawings, which are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as a limitation on the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance.

[0089] The present invention will be further described below in conjunction with the accompanying drawings.

[0090] like Figure 1 As shown, a method for cleaning wind power data of a wind turbine includes:

[0091] Step S1, obtaining annual SCADA operation data of wind turbines.

[0092] Step S2, extracting the timestamp, wind speed and power data from the SCADA operation data.

[0093] Step S3, draw a wind speed-power scatter plot, and analyze the distribution of abnormal points in the wind speed-power scatter plot.

[0094] In the embodiment, the abnormal point data of the wind speed-power scatter plot includes: Figure 2 As shown in the mark ①, the abnormal data caused by unplanned and planned shutdowns, and because the wind turbine has self-consumption of electricity, as the wind speed increases, the power becomes less than or equal to zero; Figure 2 As shown in the mark ②, the abnormal data caused by wind curtailment or unit capacity reduction is manifested as a horizontally dense data band before the wind turbine reaches the rated power; Figure 2 As shown in the mark ③, the abnormal data caused by signal propagation error, sensor failure, aging of wind turbine components, poor power generation performance or extreme weather has a random degree of fluctuation; Figure 2As shown in the mark ④, abnormal data caused by failure of the wind speed and direction instrument or communication error will result in a large power point when the wind speed is too low; Figure 2 Marked ⑤ in the middle shows normal data.

[0095] Step S4, applying the quartile method, calculating the quantile and quantile range, determining the data elimination boundary, and eliminating sparse scattered points of wind power data.

[0096] In the embodiment, the quartile method is used to eliminate sparse abnormal data of wind power. The quartile method is a statistical analysis method that can be used for data cleaning. Its basic principle is to arrange data samples from small to large, divide the data samples into four parts evenly using three points, and then obtain the interquartile range and the upper and lower limits of the abnormal value. First, all data points are sorted from small to large, and the quartile method is applied to calculate the quantile and quantile range, including:

[0097] Let the first quantile be Q1, the second quantile be Q2, the third quantile be Q3, and the interquartile range be I QR ,

[0098] If a set of ascending sequence is x={X1,x2,x3,...x m}, then calculate Q1, Q2, Q3 and I QR The formula is as follows:

[0099]

[0100]

[0101]

[0102] I QR =Q3-Q1.

[0103] Step S5, determine the abnormal data boundary through quartiles and quantile distances, and draw a box plot.

[0104] In the embodiment, the outliers in the data set are judged by the calculated quantiles and quantile ranges, which are less than Q1-1.5I QR Or greater than Q3+1.5I QR The data points are considered as outliers, and then a box plot is drawn, such as Figure 3 As shown; the wind power scatter plots of the three groups after removing abnormally sparse wind power points by the quartile method are as follows Figure 4-9 shown.

[0105] Step S6, using the similarity between the sigmoid function and the wind power curve, analyzing the changes in the sigmoid function image after adding different parameters, combining the distribution of abnormal points in the wind speed-power scatter plot, establishing a sigmoid transformation model based on four parameters, and removing abnormal points.

[0106] In the embodiment, most of the scattered points, especially the power limit points, can be removed after the wind power curve is cleaned by the quartile method, but there are still some abnormal scattered points that need further processing. It is found through observation that the power curve under actual conditions is similar to the sigmoid curve. Therefore, the abnormal points are cleaned by curve fitting of the sigmoid transformation model.

[0107] In the embodiment, the Sigmoid function is a commonly used binary classification function, and the function form is:

[0108]

[0109] The curve changes after adding d, c, b and a parameters to the Sigmoid function are as follows:

[0110] If the function is transformed as follows:

[0111]

[0112] Among them, when d is 0.75, 1, 1.25, the function graph shows Fig.10 As shown in the changes, the d value reflects the shift of the function graph on the x-axis. As the d value increases, the function shifts to the right, and as the d value decreases, the function shifts to the left.

[0113] If the function is transformed as follows:

[0114]

[0115] When c is 0.5, 1, or 2, the function graph is as follows: Fig.11 The changes shown, the c value reflects the edge tilt of the function image. The larger the c value, the steeper the image.

[0116] If the function is transformed as follows:

[0117]

[0118] Among them, when b is 0.1, 0.25, and 0.75, the function graph appears as follows Fig.12 The change shown, the b value reflects the difference between the maximum and minimum values ​​of the function graph;

[0119] If the function is transformed as follows:

[0120]

[0121] Among them, when a is 10, 20, and 30, the function graph appears as follows Fig.13 The change shown, the a value reflects the difference between the maximum and minimum values ​​of the function graph;

[0122] Through the analysis of the above parameters, the mathematical model based on the sigmoid function is established as follows, and the initial values ​​of a, b, c and d are set to obtain the sigmoid transformation model based on four parameters, namely:

[0123]

[0124] Step S7, based on the sigmoid transformation model, establish the objective function, and apply the LM optimization algorithm to solve the parameters until the objective function value is less than the maximum allowable deviation, obtain four parameters, and then fit the wind power curve.

[0125] In the embodiment, since the sigmoid function has four parameters, it belongs to a four-parameter unconstrained optimization problem. In this paper, the LM (Levenberg-Marquardt) optimization algorithm is applied to solve the parameters to solve the characteristics that the gradient descent is insensitive to the initial value and the Gauss-Newton converges quickly near the optimal value.

[0126] The LM algorithm is an improvement on the Newton method. It is necessary to add perturbations or regularized expressions to the relevant equations to prevent algorithm instability and overfitting. First, the Newton method is used to solve the four-parameter fitting.

[0127] In the embodiment, let θ = [a, b, c, d] be the parameters to be determined, j i (θ) = y i -f(θ;x i ), the target formula is:

[0128]

[0129] In the formula, j i (θ) is the deviation function between the actual power value of the wind turbine and the power value of the wind turbine obtained by the sigmoid transformation model, f(θ; x i ) is the wind turbine power value required by the sigmoid transformation model, y i is the actual value of wind turbine power, m is the number of wind power data points, x i is the wind speed value of the wind power data point;

[0130] To minimize P(θ), the first step is to derive P(θ), set P′(θ) = 0, and perform Taylor expansion to obtain:

[0131] P′(θ+Δθ)≈P′(θ)+P″(θ)Δθ=0;

[0132] Where P′(θ) is the first derivative of the target formula, and P″(θ) is the second derivative of the target formula;

[0133] Then we get:

[0134] Δθ=-(P″(θ)) -1 P′(θ);

[0135] After two derivations, the Jacobian matrix K of θ with respect to j is obtained as follows:

[0136]

[0137] Wherein, j(θ) is the deviation function between the actual power value of the wind turbine and the power value of the wind turbine obtained by the sigmoid transformation model, j′(θ) is the differential matrix of the deviation function between the actual power value of the wind turbine and the power value of the wind turbine obtained by the sigmoid transformation model with respect to the parameters [a, b, c, d];

[0138] Gradient information g(θ) of P:

[0139]

[0140] In the formula, T is;

[0141] Then, the Hessian matrix of P is:

[0142]

[0143] Δθ=-(P″(θ)) -1 P′(θ)=-(K T K+S(θ) -1 K T j(θ);

[0144] In the formula, S(θ) is the high-order term of the Gauss-Newton method, j(θ) is the deviation function between the actual power value of the wind turbine and the power value of the wind turbine obtained by the sigmoid transformation model;

[0145] Among them, the amount of calculation of S(θ) is usually large, and the Gauss-Newton method generally ignores this term, and we get:

[0146] Δθ=-(P″(θ)) -1 P′(θ)=-(K T K) -1 K T j(θ);

[0147] In order to prevent the Gauss-Newton method from requiring full rank to solve the Jacobian matrix and causing instability or overfitting, P(θ) is added to the L2 regularization penalty term, and the result is:

[0148]

[0149] Δθ=-(K T K+βI) -1 K T j(θ);

[0150] Where β is the regularization parameter and I is the unit matrix;

[0151] It can be seen from the above formula that when β is large, the descent gradient is close to the steepest descent method, and when it is very small, it is close to the Gauss-Newton method. In order to obtain a suitable β, it is necessary to define an evaluation parameter ∈, let z = Δθ, then:

[0152]

[0153] That is, the closer ∈ is to 1, the more accurate the overall Taylor expansion is. If ∈ is less than 0, adjust the step size β, and continuously adjust the value of θ through gradient calculation to make P(θ) less than the set error value, thereby completing the four-parameter fitting of the wind power curve, as shown in Fig.14 , 16 , as shown in 18.

[0154] Step S8, by changing the four parameter values ​​to obtain the boundary curve of the abnormal scatter points, the remaining abnormal scatter points are identified and removed, and finally the effective wind power scatter points are obtained to complete the wind power data cleaning, such as Fig.15 , 17 , as shown in 19.

[0155] In the embodiment, the effectiveness of the method is verified for different forms of wind power data scatter points. After actual calculation, the optimal parameter tables under three different wind power data are obtained as shown in Table 1, Table 2, and Table 3, and the optimal parameter tables of the upper and lower boundary fitting curves for eliminating abnormal data are given.

[0156]

[0157]

[0158] Table 1 Parameters of the first group of power curves fitted by the sigmoid transformation model

[0159] category Parameter a Parameter b Parameter c Parameter d Upper Boundary 13.28 0.0065 0.8172 0.0667 Fitting curve 12.65 0.0065 0.8172 0.6585 Lower Boundary 12.27 0.0065 0.8172 1.6463

[0160] Table 2 The second group of power curve parameters fitted by sigmoid transformation model

[0161] category Parameter a Parameter b Parameter c Parameter d Upper Boundary 15.7485 0.0073 0.6915 0.2900 Fitting curve 15.2898 0.0073 0.6915 0.7251 Lower Boundary 14.5253 0.0073 0.6915 1.4502

[0162] Table 3 Parameters of the third group of power curves fitted by the sigmoid transformation model

[0163] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments may still be modified, or some or all of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for cleaning wind turbine wind power data, characterized in that: include: Step S1, obtaining annual SCADA operation data of wind turbines; Step S2, extracting timestamp, wind speed and power data from SCADA operation data; Step S3, drawing a wind speed-power scatter plot, and analyzing the distribution of abnormal points in the wind speed-power scatter plot; Step S4, applying the quartile method, calculating the quantile and quantile range, determining the data elimination boundary, and eliminating sparse scattered points of wind power data; Step S5, determining the abnormal data boundary through quartiles and quantile distances, and drawing a box plot; Step S6, using the similarity between the sigmoid function and the wind power curve, analyzing the change of the sigmoid function image after adding different parameters, combining the distribution of abnormal points in the wind speed-power scatter plot, establishing a sigmoid transformation model based on four parameters, and removing abnormal points; Step S7, based on the sigmoid transformation model, establish the objective function, and apply the LM optimization algorithm to solve the parameters until the objective function value is less than the maximum allowable deviation, obtain four parameters, and then fit the wind power curve; Step S8, by changing the four parameter values ​​to obtain the boundary curve of the abnormal scatter points, the remaining abnormal scatter points are identified and removed, and finally the effective wind power scatter points are obtained to complete the wind power data cleaning.

2. A method for cleaning wind turbine wind power data according to claim 1, characterized in that: In step S3, the abnormal point data of the wind speed-power scatter plot include: abnormal data caused by unplanned shutdown and planned shutdown, abnormal data caused by wind power curtailment or unit capacity reduction, abnormal data caused by signal propagation error, sensor failure, aging of wind turbine components, poor power generation performance or extreme weather, and abnormal data caused by failure of anemometer or communication error.

3. A method for cleaning wind turbine wind power data according to claim 1, characterized in that: In step S4, the quartile method is applied to calculate the quantile and quantile range, including: Let the first quantile be Q1, the second quantile be Q2, the third quantile be Q3, and the interquartile range be I QR , If an ascending sequence is x={x1,x2,x3,...x m }, then calculate Q1, Q2, Q3 and I QR The formula is as follows:

4. A method for cleaning wind turbine wind power data according to claim 3, characterized in that: In step S5, the outliers in the data set are determined by calculating the quantile and quantile range, and the outliers are smaller than Q1-1.5I QR Or greater than Q3+1.5I QR The data points are considered as outliers and a box plot is drawn.

5. A method for cleaning wind turbine wind power data according to claim 1, characterized in that: In step S6, the Sigmoid function is a commonly used binary classification function, and the function form is: The curve changes after adding d, c, b and a parameters to the Sigmoid function are as follows: If the function is transformed as follows: The d value reflects the shift of the function graph on the x-axis. As the d value increases, the function shifts to the right, and as the d value decreases, the function shifts to the left. If the function is transformed as follows: Among them, the c value reflects the edge inclination of the function image. The larger the c value, the steeper the image. If the function is transformed as follows: Among them, the b value reflects the difference between the maximum and minimum values ​​of the function graph; If the function is transformed as follows: Among them, the a value reflects the difference between the maximum and minimum values ​​of the function graph; Through the analysis of the above parameters, the mathematical model based on the sigmoid function is established as follows, and the initial values ​​of a, b, c and d are set to obtain the sigmoid transformation model based on four parameters, namely:

6. A method for cleaning wind turbine wind power data according to claim 5, characterized in that: In step S7, let θ = [a, b, c, d] be the parameters to be determined, j i (θ) = y i -f(θ;x i ), the target formula is: In the formula, j i (θ) is the deviation function between the actual power value of the wind turbine and the power value of the wind turbine obtained by the sigmoid transformation model, f(θ; x i ) is the wind turbine power value required by the sigmoid transformation model, y i is the actual value of wind turbine power, m is the number of wind power data points, x i is the wind speed value of the wind power data point; To minimize P(θ), the first step is to derive P(θ), set P'(θ) = 0, and perform Taylor expansion to obtain: P'(θ+Δθ)≈P'(θ)+P”(θ)Δθ=0; Where, P'(θ) is the first derivative of the target formula, and P"(θ) is the second derivative of the target formula; Then we get: Δθ=-(P”(θ)) -1 P'(θ); After two derivations, the Jacobian matrix K of θ with respect to j is obtained as follows: Wherein, j(θ) is the deviation function between the actual power value of the wind turbine and the power value of the wind turbine obtained by the sigmoid transformation model, j'(θ) is the differential matrix of the deviation function between the actual power value of the wind turbine and the power value of the wind turbine obtained by the sigmoid transformation model with respect to the parameters [a, b, c, d]; Gradient information g(θ) of P: In the formula, T is; Then, the Hessian matrix of P is: Δθ=-(P”(θ)) -1 P'(θ)=-(K T K+S(θ)) -1 K T j(θ); In the formula, S(θ) is the high-order term of the Gauss-Newton method, j(θ) is the deviation function between the actual power value of the wind turbine and the power value of the wind turbine obtained by the sigmoid transformation model; Among them, the amount of calculation of S(θ) is usually large, and the Gauss-Newton method generally ignores this term, and we get: Δθ=-(P”(θ)) -1 P'(θ)=-(K T K) -1 K T j(θ); In order to prevent the Gauss-Newton method from requiring full rank to solve the Jacobian matrix and causing instability or overfitting, P(θ) is added to the L2 regularization penalty term, and the result is: Δθ=-(K T K+βI) -1 K T j(θ); Where β is the regularization parameter and I is the unit matrix; It can be seen from the above formula that when β is large, the descent gradient is close to the steepest descent method, and when it is small, it is close to the Gauss-Newton method. In order to obtain a suitable β, it is necessary to define an evaluation parameter ∈, let z = Δθ, then: That is, the closer ∈ is to 1, the more accurate the overall Taylor expansion is. If ∈ is less than 0, adjust the step size β, and continuously adjust the value of θ through gradient calculation to make P(θ) less than the set error value, thereby completing the four-parameter fitting of the wind power curve.

Images