A wind turbine inflow wind speed estimation method based on GRA-XGBoost

Abstract

The application discloses a wind turbine inflow wind speed estimation method based on GRA-XGBoost, comprising the following steps: screening effective data based on SCADA time series data recorded by a wind turbine; dividing wind speed intervals according to the operation control strategy of the wind turbine to obtain sub-data sets under each wind speed interval; selecting characteristic values with high grey correlation degrees in each sub-data set and inputting the characteristic values into an XGBoost model; taking the real wind speed value of laser radar wind measurement data as a reference as a target vector of the model for training; adjusting and optimizing model parameters according to a set prediction model evaluation index to obtain a sub-model result meeting a condition and a GRA-XGBoost wind speed estimation combined model of a full wind speed section; and obtaining real-time effective inflow wind speed estimation values through online deployment of the model. The application not only can overcome the problem that the machine cabin transfer function has poor practicability, but also has high fitting precision and lower prediction error.

Description

Technical Field

[0001] This invention belongs to the field of wind speed prediction technology for wind turbines, specifically relating to a method for estimating inflow wind speed of wind turbines based on GRA-XGBoost. Background Technology

[0002] Wakes are one of the most important phenomena in wind farm research. Their presence reduces the power generation of downstream turbines and increases turbine load. To balance computational accuracy and efficiency, many researchers have proposed data-driven, parameter-corrected analytical wake models in recent years. However, the accuracy of wake models obtained by such methods largely depends on highly accurate input wind speed data. Accurately obtaining free-flow wind speed data is a necessary condition for ensuring the accuracy of such models.

[0003] Incoming wind speed is difficult to measure directly. Wind farms typically assume the nacelle wind speed is the same as the incoming wind speed, without distinguishing between the two. Currently, the nacelle wind speed values ​​recorded by the turbine's SCADA (Supervisory Control and Data Acquisition) system generally come from anemometers, usually installed at the rear of each turbine's nacelle. Due to factors such as rotor rotation and airflow distortion caused by the nacelle, the anemometer's measurements cannot accurately reflect the incoming wind speed, resulting in low accuracy of data-driven wake models built using nacelle wind speed. Therefore, to accurately and efficiently complete wake effect assessments and obtain accurate free-flow wind speeds for the turbine, it is necessary to correct the nacelle wind speed measured by the anemometer in the turbine's SCADA system, converting it into an equivalent environmental inflow wind speed.

[0004] Methods for correcting nacelle wind speed are mainly divided into two categories: theoretical correction methods and function fitting methods. Theoretical correction methods are generally based on aerodynamic quantification of the relationship between wind speed and relevant turbine performance parameters, but the calculations are complex and prone to significant deviations in practical applications. Function fitting methods, following the IEC standard (61400-12-2:2013), directly fit the functional relationship between the incoming wind speed and the nacelle wind speed, i.e., the nacelle transfer function (NTF), enabling large-scale application. However, the wind speed measured by the anemometer tower cannot be equated to the actual incoming wind speed at the turbine hub height, and the fitting method generally uses simple linear fitting, leading to calculation errors and poor practicality. Furthermore, these methods heavily rely on anemometer tower data, and in a single wind field, NTF fitting can only be performed on specific turbines near adjacent anemometer towers, and cannot be directly applied to other turbines, limiting their application. To address this, Chinese patent (publication number CN115510656A) utilizes flow field simulation and determines the NTF based on the average of the nacelle wind speed and the incoming wind speed in each wind speed range. This method can eliminate the reliance on wind measurement data from anemometers, but it has a high computational cost. Chinese patent (publication number CN112861429B) uses a neural network to deduce the theoretical power curve of the wind turbine, inputting actual wind power data from wind farms into the neural network to obtain the corresponding incoming wind speed value. Then, it performs linear function fitting on the nacelle wind speed and the corresponding incoming wind speed in high and low frequency regions. This method has a low computational cost, but it ignores the objective fact that the accuracy of the theoretical power curve of wind turbines is currently difficult to guarantee. Furthermore, it does not consider the impact of wind direction uncertainty on wake modeling, and therefore cannot guarantee the accuracy of free-flow wind speed derivation and wake effect assessment. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention aims to provide a method for estimating inflow wind speed in wind turbines based on GRA-XGBoost, thereby addressing the problems of large calculation errors and limited application scope in existing nacelle wind speed correction methods. To achieve this objective, the specific technical solution of this invention is as follows:

[0006] A method for estimating inflow wind speed of wind turbines based on GRA-XGBoost includes the following steps:

[0007] S1. Obtain historical time-series SCADA data of wind turbines equipped with lidar over a period of time, and perform preprocessing operations such as removing outliers, filling missing values, and filtering valid values ​​to improve data quality.

[0008] S2. Using wind speed data as the basis for dataset division, wind speed intervals are divided according to the wind turbine's operation control strategy, thereby obtaining subsets of data under each different wind speed interval.

[0009] S3. Within each subset of datasets, GRA is used as a tool for feature analysis and selection to reduce the input dimensionality. Feature values ​​with high gray relational degree are selected and input into the XGBoost model to improve the accuracy of training the model, while avoiding overfitting caused by high-dimensional input.

[0010] S4. Within each subset of data, lidar wind measurement data is used as a reference value for the actual inflow wind speed in front of the wind turbine rotor, and this value is used as the target vector for training and prediction of the model. The prediction model is then evaluated based on the established metrics: MSE, MAE, RMSE, and R². 2 Adjust and optimize the model parameters to obtain sub-model results that meet the conditions;

[0011] S5. By concatenating the sub-models obtained from each wind speed range according to wind speed magnitude, a combined GRA-XGBoost wind speed estimation model for the entire wind speed range is obtained. Finally, the model is deployed online. For the real-time output SCADA time-series data of the same type of wind turbine, after performing the corresponding data preprocessing operations, the data is input into the trained wind speed estimation model. After passing through a filtering module with appropriate bandwidth, the real-time effective wind speed estimate is obtained.

[0012] Furthermore, the SCADA historical time-series data of the wind turbine equipped with lidar acquired in step S1 over a period of time includes three information data: timestamp, lidar wind direction, and lidar wind speed. It also includes 28 relevant characteristic data points: 10-minute moving average wind speed, 10-minute average wind speed considering the effects of air temperature and turbulence, generator speed, wind turbine active power, grid B-phase voltage, generator front bearing temperature, FFT transform frequency in frequency range 2, high-speed shaft brake gearbox side temperature, transformer room temperature, hydraulic station oil temperature, main bearing gearbox side temperature, pitch angle 2, 10m pitch angle, nacelle Y-direction vibration acceleration, yaw hydraulic oil temperature, inverter cabinet temperature, gearbox oil pressure, frequency range 1, transformer top oil temperature, average wind direction, daily power generation, nacelle vibration RMS value, ultrasonic anemometer turbulence intensity, A3 control cabinet temperature, real-time turbulence intensity measured by mechanical anemometer, gearbox idle pump pressure, 10-minute average turbulence intensity, generator rear bearing temperature, and nacelle X-direction vibration acceleration.

[0013] Furthermore, the preprocessing operations in step S1, which involve removing outliers, filling in missing values, and filtering valid values, are as follows:

[0014] S11. Based on the timestamp, synchronize and align the lidar wind measurement data and related feature data, and then remove the shutdown data and power limitation data from the SCADA data according to the two pieces of information: generator speed and wind turbine active power.

[0015] S12. Perform missing value imputation. Use the missingno module tool to visualize the missing values ​​in the data, and use the k-nearest neighbor method, based on Euclidean distance, to find the k nearest sample points in the dataset of feature data with missing values, and use the mean of the non-empty values ​​of these sample points to impute the missing values.

[0016] S13. Further filter the wind direction sectors where the wind turbine is obstructed by obstacles under actual operating conditions, and remove the data of these sectors based on the lidar wind direction information.

[0017] Furthermore, the operating control modes of the wind turbine under different wind speeds in step S2 can be divided into the following four states:

[0018] Operating state 1: Constant speed section 1, when the wind speed is greater than the cut-in wind speed, the wind turbine starts to run at a constant speed;

[0019] Operating State 2: Variable Speed ​​Section. When the speed in State 2 is equal to the optimal speed corresponding to the current wind speed, the wind turbine enters the variable speed state to achieve maximum wind energy capture. At this time, the pitch angle is maintained at 0 degrees, and the electromagnetic torque acts on the system according to the optimal torque control strategy.

[0020] Operating state 3: Constant speed segment 2. When the speed of the maximum wind energy capture state is equal to the upper limit of the safe operating speed of the generator, the wind turbine will once again enter the constant speed operating state.

[0021] Operating status 4: Constant power section. When the wind speed is greater than the rated wind speed, the wind turbine rotor speed and generator speed of the wind turbine need to be maintained near their rated values.

[0022] Therefore, the entire wind speed range is divided into four ranges according to the different unit control strategies. All data in the dataset are sorted from low to high according to the size of the lidar wind speed data. The cut-in wind speed is taken as the starting critical value and the cut-out wind speed is taken as the ending critical value, thus obtaining the subsets of each different wind speed range.

[0023] Furthermore, in step S3, GRA, taking values ​​[0,1], represents the relative importance of a specific sequence or variable relative to a reference sequence. Based on the principle of GRA feature analysis, assuming the lidar wind speed sequence in the SCADA dataset is X0, representing the input sequence for GRA analysis, i.e., the target sequence for GRA analysis; the relevant feature value sequence is X... i Let represent the reference sequence for GRA analysis, then:

[0024] X0=(x 01 ,x 02 ,...,x 0n )

[0025] Xi =(x i1 ,x i2 ,...,x in ),

[0026] Then, the input sequence and the reference sequence can be represented by a matrix V, where i = 1, 2, ..., m;

[0027]

[0028] Normalize all values ​​in matrix V by scaling them proportionally to the range [0,1]. The normalized matrix is ​​represented as V′:

[0029]

[0030] Next, determine the input sequence X0 and the reference sequence X. i The absolute difference Δ between ik , is represented as:

[0031]

[0032] Where k = 1, 2, ..., m; ga represents Δ ik The minimum value, ha represents Δ ik The maximum value is obtained by evaluating the values ​​of ga and ha to obtain the relationship function r(x′) between the reference sequence and the input sequence. 0t ,x′ it ):

[0033] Where ρ is the discrimination coefficient, with a value of (0,1). In GRA analysis, it is generally considered that ρ is optimal at 0.5, so we take ρ = 0.5; then, the grey relational degree r(X0,X) is determined according to the following formula. i ):

[0034]

[0035] Finally, according to r(X0,X) i The value of ) will be used to sort the reference sequence Xi in descending order from smallest to largest and output it for subsequent selection of feature values ​​to be input into the XGBoost model.

[0036] Further, the evaluation indicators in step S4 are specifically: MAE represents Mean Absolute Error, MSE represents Mean Squared Error, RMSE represents Mean Squared Error, R 2 The goodness of fit is expressed by the following formula:

[0037]

[0038]

[0039]

[0040]

[0041] Where N is the total number of wind speed value samples recorded in the dataset, v i and It refers to the XGBoost estimated inflow wind speed value falling within this wind speed range and its corresponding actual value, i.e., the lidar wind speed value. The magnitude of all lidar wind speed values ​​within this range.

[0042] Compared with existing technologies, this invention has the following advantages: Based on SCADA time-series data recorded by wind turbines, this invention proposes a data preprocessing mechanism to filter effective data; it divides wind speed intervals according to the wind turbine's operation control strategy, thereby dividing the original dataset and obtaining subsets for each wind speed interval; within each subset, it uses GRA for feature analysis to reduce the input dimensionality and selects feature values ​​with high grey relational degree to input into the XGBoost model; then, it uses the real wind speed value with lidar wind measurement data as a reference as the target vector for training the model; and it uses the set prediction model evaluation indicators MSE, MAE, RMSE, and R... 2 The model parameters were adjusted and optimized to obtain sub-model results that met the conditions. These sub-models were then combined to obtain a GRA-XGBoost wind speed estimation combined model covering the entire wind speed range. Finally, the model was deployed online to obtain real-time and effective inflow wind speed estimates. Test results show that using this invention to correct inflow wind speed not only overcomes the poor practicality of the nacelle transfer function under the IEC standard, but also achieves high fitting accuracy with relatively simple linear fitting, exhibiting lower prediction errors than existing methods and demonstrating good applicability. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating the design of the wind turbine inflow wind speed estimation method based on GRA-XGBoost according to the present invention.

[0044] Figure 2 This is a visualization of the missing values ​​in the dataset after preprocessing, as shown in this embodiment of the invention.

[0045] Figure 3 This is a schematic diagram of the operating status of the Shanghai Electric SWT-4.0-146 wind turbine generator set according to the present invention;

[0046] Figure 4 This is a graph showing the fitting effect between the estimated values ​​and actual values ​​of each GRA-XGBoost sub-model trained in this embodiment of the invention.

[0047] Figure 5The figures show the fitting performance of the GRA-XGBoost wind speed estimation combined model and its online deployment effect in this embodiment of the invention. Detailed Implementation

[0048] To enable those skilled in the art to more clearly understand the technical solution of the present invention, the present invention will be further described below in conjunction with the accompanying drawings.

[0049] This invention addresses the problems of large calculation deviations and limited application scope of nacelle wind speed correction methods based on linear function fitting. In order to ensure the data quality of the incoming wind speed, the input variable of the data-driven wake model, this invention proposes a wind turbine inflow wind speed estimation method based on grey relational analysis-gradient boosting decision tree (GRA-XGBoost) for wake modeling.

[0050] like Figure 1 As shown, a method for estimating the inflow wind speed of a wind turbine based on GRA-XGBoost is presented. This method includes the following steps:

[0051] (1) Obtain SCADA second-level time-series data of a Shanghai Electric SWT-4.0-146 wind turbine equipped with lidar within a certain period of 30 days, including three information data: timestamp, lidar wind direction, and lidar wind speed; 10-minute moving average wind speed; 10-minute average wind speed considering the effects of air temperature and turbulence; generator speed; wind turbine active power; grid B-phase voltage; generator front bearing temperature; FFT transform frequency in frequency range 2; high-speed shaft brake gearbox side temperature; transformer room temperature; and hydraulic station oil temperature. The data includes 28 relevant characteristic data points: main bearing gearbox side temperature, pitch angle 2, 10m pitch angle, nacelle Y-direction vibration acceleration, yaw hydraulic oil temperature, inverter cabinet temperature, gearbox oil pressure, frequency range 1, transformer top oil temperature, average wind direction, daily power generation, nacelle vibration RMS value, ultrasonic anemometer turbulence intensity, A3 control cabinet temperature, real-time turbulence intensity measured by mechanical anemometer, gearbox idle pump pressure, 10-minute average turbulence intensity, generator rear bearing temperature, and nacelle X-direction vibration acceleration.

[0052] To improve data quality, preprocessing operations were performed to remove outliers, fill in missing values, and filter valid values. The specific process was as follows: First, the lidar wind measurement data and related feature data were synchronized and aligned based on timestamps. Then, based on the generator speed and active power of the wind turbine in the SCADA data, outage and power-limited data were removed. Next, missing value imputation was performed. The missing value situation was visualized using the `missingno` module tool, and a k-nearest neighbor (KNN) method was used, with Euclidean distance as the criterion. For feature data with missing values, the k nearest neighbor samples in the dataset were found, and the mean of the non-empty values ​​of these samples were used to impute the missing values. The missing value situation of the dataset after the above preprocessing operations is shown below. Figure 2 As shown, each feature data has a complete bar chart, and there are no missing values. Finally, the wind direction sectors where the wind turbine is obstructed by obstacles under actual operating conditions are filtered out based on the lidar wind direction information.

[0053] (2) Using wind speed data as the basis for dataset division, wind speed ranges are divided according to the wind turbine's operation control strategy. The operating control modes of modern large wind turbines under different wind speeds can be divided into the following four states:

[0054] Operating state 1: Constant speed section 1. When the wind speed is greater than the cut-in wind speed, the wind turbine starts to run at a constant speed.

[0055] Operating State 2: Variable Speed ​​Section. When the speed in State 2 is equal to the optimal speed corresponding to the current wind speed, the wind turbine enters the variable speed state to achieve maximum wind energy capture. At this time, the pitch angle is maintained at 0 degrees, and the electromagnetic torque acts on the system according to the optimal torque control strategy.

[0056] Operating state 3: Constant speed segment 2. When the speed of the maximum wind energy capture state is equal to the upper limit of the safe operating speed of the generator, the wind turbine will once again enter the constant speed operating state.

[0057] Operating State 4: Constant Power Section. When the wind speed is greater than the rated wind speed, the wind turbine rotor speed and generator speed of the wind turbine need to be maintained near their rated values.

[0058] like Figure 3 The diagram shown is a schematic of the operating status of the Shanghai Electric SWT-4.0-146 wind turbine in the embodiment. Therefore, the entire wind speed range is divided into four intervals (unit: m / s) according to the different control strategies of the unit: [3,6), [6,9), [9,10.5), and [10.5,25). All data in the dataset are sorted from low to high according to the size of the lidar wind speed data to obtain the subsets of each different wind speed range.

[0059] (3) Within each subset of datasets, GRA is used as a tool for feature analysis and selection to reduce the input dimensionality. Feature values ​​with high grey relational strength are selected and input into the XGBoost model. The characteristic of GRA is that it primarily represents the relative importance of a specific sequence or variable relative to a reference sequence within the value range [0,1]. As shown below, based on the principle of GRA feature analysis, assuming the lidar wind speed sequence in the SCADA dataset is X0, representing the input sequence for GRA analysis, i.e., the target sequence for GRA analysis; the relevant feature value sequence is X... i , indicating the reference sequence for GRA analysis.

[0060] X0=(x 01 ,x 02 ,…,x 0n )

[0061] X i =(x i1 ,x i2 ,…,x in ),

[0062] Then, the input sequence and the reference sequence can be represented by a matrix V, where i = 1, 2, ..., m.

[0063]

[0064] Normalize all values ​​in matrix V by scaling them proportionally to the range [0,1], as shown below. The normalized matrix is ​​represented as V′:

[0065]

[0066] Next, determine the input sequence X0 and the reference sequence X. i The absolute difference Δ between ik , is represented as:

[0067] Δ ik =|x′ 0i -x′ ik |

[0068] ga=min(Δ ik )

[0069] ha = max(Δ ik ),

[0070] Where k = 1, 2, ..., m, ga represents Δ ik The minimum value, ha represents Δ ikThe maximum value is shown below. The relationship function r(x′) between the reference sequence and the input sequence is obtained by evaluating the values ​​of ga and ha. 0t ,x′ it ):

[0071]

[0072] Where ρ is the discrimination coefficient, with a value of (0,1). In GRA analysis, a value of 0.5 is generally considered optimal for ρ. Therefore, we take ρ = 0.5, and then determine the grey relational degree r(X0,X) according to the following formula. i ):

[0073]

[0074] Finally, according to r(X0,X) i The reference sequence Xi is sorted in descending order from smallest to largest and output for subsequent selection of feature values ​​to be input into the XGBoost model. This improves the accuracy of model training and avoids overfitting caused by high-dimensional input.

[0075] (4) Within each subset of data sets, lidar wind measurement data is used as a reference value for the actual inflow wind speed in front of the wind turbine rotor, and is used as the target vector for training and prediction of the model. The prediction model is evaluated based on the set evaluation metrics MSE, MAE, RMSE, and R. 2 MAE represents Mean Absolute Error, MSE represents Mean Squared Error, RMSE represents Mean Squared Error, and R 2 This represents the goodness of fit; its formula is expressed as follows:

[0076]

[0077]

[0078]

[0079]

[0080] Where N is the total number of wind speed value samples recorded in the dataset, v i and It refers to the XGBoost estimated inflow wind speed value falling within this wind speed range and its corresponding actual value, i.e., the lidar wind speed value. The values ​​of all LiDAR wind speeds within this interval are calculated. Based on the above indicators, the parameters of the sub-models within each wind speed interval are adjusted. In this example, the adjusted XGBoost model parameters are objective(reg:squarederror\reg:squaredlogerror\reg:logistic), num_boost_round, max_depth, subsample, colsample_bytree, gamma, min_child_weight, lambda, and alpha. The fitting effect between the estimated values ​​and actual values ​​of the sub-models that meet the conditions is obtained as follows: Figure 4 As shown.

[0081] (5) By splicing the sub-models obtained from each wind speed range according to wind speed magnitude, a combined GRA-XGBoost wind speed estimation model for the entire wind speed range is obtained. The MSE, MAE, RMSE, and R of each sub-model and the combined GRA-XGBoost wind speed estimation model are then analyzed. 2 As shown in Table 1 below:

[0082] Table 1: Performance Indicators of Individual Sub-models and the Combined GRA-XGBoost Wind Speed ​​Estimation Model

[0083]

[0084] The online deployment model, after performing corresponding data preprocessing operations on the real-time SCADA time-series data output by the same type of wind turbine, is input into the trained wind speed estimation model. After passing through a filtering module with appropriate bandwidth, the real-time effective wind speed estimate is obtained; for example... Figure 5 The figure shown is a graph illustrating the fitting performance of the GRA-XGBoost wind speed estimation combined model obtained in this embodiment and the effect of its online deployment.

[0085] The nacelle wind speed information from the SCADA data of other wind turbines of the same type in the wind farm that are not equipped with lidar is input into the GRA-XGBoost wind speed estimation combination model, and the corrected inflow wind speed estimates of each wind turbine are output; then the estimated values ​​are input into the data-driven parameter correction analytical wake model.

[0086] As shown in Table 1, compared with the traditional linear function fitting method, the method proposed in this invention, which divides the wind speed range according to the wind turbine's operation control strategy, has higher fitting accuracy. Then, by combining the various sub-models, a GRA-XGBoost wind speed estimation combined model for the entire wind speed range is obtained, smoothing the switching disturbances caused by the piecewise linear function fitting method and improving the discontinuity problem of the traditional piecewise nacelle transfer function. Furthermore, it has good practicality for wind turbines of the same model within the same wind farm, has a wider range of applications, and can be used for inflow wind speed estimation calculations for common types of wind turbines in engineering practice, especially providing reliable data support for subsequent wind farm wake effect assessment.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for estimating inflow wind speed of a wind turbine based on GRA-XGBoost, characterized in that, Includes the following steps: S1. Obtain historical time-series SCADA data of wind turbines equipped with lidar over a period of time, and perform preprocessing operations such as removing outliers, filling missing values, and filtering valid values ​​to improve data quality. S2. Using wind speed data as the basis for dataset division, wind speed intervals are divided according to the wind turbine's operation control strategy, thereby obtaining subsets of data under each different wind speed interval. S3. Within each subset of datasets, GRA is used as a tool for feature analysis and selection to reduce the input dimensionality. Feature values ​​with high gray relational degree are selected and input into the XGBoost model to improve the accuracy of training the model, while avoiding overfitting caused by high-dimensional input. S4. Within each subset of data, lidar wind measurement data is used as a reference value for the actual inflow wind speed in front of the wind turbine rotor, and this value is used as the target vector for training and prediction of the model. The prediction model is then evaluated based on the established metrics: MSE, MAE, RMSE, and R². 2 Adjust and optimize the model parameters to obtain sub-model results that meet the conditions; S5. By splicing the sub-models obtained from each wind speed range according to the wind speed magnitude, a GRA-XGBoost wind speed estimation combined model for the entire wind speed range is obtained. Finally, the model is deployed online. For the real-time output SCADA time series data of the same type of wind turbine, after performing the corresponding data preprocessing operations, the data is input into the trained wind speed estimation model. After passing through a filtering module with appropriate bandwidth, the real-time effective wind speed estimate is obtained.

2. The wind turbine inflow velocity estimation method based on GRA-XGBoost according to claim 1, characterized in that, The SCADA historical time-series data acquired in step S1 for a certain period of time for the wind turbine equipped with lidar includes three information data: timestamp, lidar wind direction, and lidar wind speed. It also includes 28 relevant characteristic data points: 10-minute moving average wind speed, 10-minute average wind speed considering the effects of air temperature and turbulence, generator speed, wind turbine active power, grid B-phase voltage, generator front bearing temperature, FFT transform frequency in frequency range 2, high-speed shaft brake gearbox side temperature, transformer room temperature, hydraulic station oil temperature, main bearing gearbox side temperature, pitch angle 2, 10m pitch angle, nacelle Y-direction vibration acceleration, yaw hydraulic oil temperature, inverter cabinet temperature, gearbox oil pressure, frequency range 1, transformer top oil temperature, average wind direction, daily power generation, nacelle vibration RMS value, ultrasonic anemometer turbulence intensity, A3 control cabinet temperature, real-time turbulence intensity measured by mechanical anemometer, gearbox idle pump pressure, 10-minute average turbulence intensity, generator rear bearing temperature, and nacelle X-direction vibration acceleration.

3. The wind turbine inflow velocity estimation method based on GRA-XGBoost according to claim 1, characterized in that, The preprocessing operation in step S1, which involves removing outliers, filling in missing values, and filtering valid values, is as follows: S11. Based on the timestamp, synchronize and align the lidar wind measurement data and related feature data, and then remove the shutdown data and power limitation data from the SCADA data according to the two pieces of information: generator speed and wind turbine active power. S12. Perform missing value imputation. Use the missingno module tool to visualize the missing values ​​in the data, and use the k-nearest neighbor method, based on Euclidean distance, to find the k nearest sample points in the dataset of feature data with missing values, and use the mean of the non-empty values ​​of these sample points to impute the missing values. S13. Further filter the wind direction sectors where the wind turbine is obstructed by obstacles under actual operating conditions, and remove the data of these sectors based on the lidar wind direction information.

4. The wind turbine inflow velocity estimation method based on GRA-XGBoost according to claim 1, characterized in that, In step S2, the operating control modes of the wind turbine under different wind speeds can be divided into the following four states: Operating state 1: Constant speed section 1, when the wind speed is greater than the cut-in wind speed, the wind turbine starts to run at a constant speed; Operating State 2: Variable Speed ​​Section. When the speed in State 2 is equal to the optimal speed corresponding to the current wind speed, the wind turbine enters the variable speed state to achieve maximum wind energy capture. At this time, the pitch angle is maintained at 0 degrees, and the electromagnetic torque acts on the system according to the optimal torque control strategy. Operating state 3: Constant speed segment 2. When the speed of the maximum wind energy capture state is equal to the upper limit of the safe operating speed of the generator, the wind turbine will once again enter the constant speed operating state. Operating status 4: Constant power section. When the wind speed is greater than the rated wind speed, the wind turbine rotor speed and generator speed of the wind turbine need to be maintained near their rated values. Therefore, the entire wind speed range is divided into four ranges according to the different unit control strategies. All data in the dataset are sorted from low to high according to the size of the lidar wind speed data. The cut-in wind speed is taken as the starting critical value and the cut-out wind speed is taken as the ending critical value, thus obtaining the subsets of each different wind speed range.

5. The wind turbine inflow velocity estimation method based on GRA-XGBoost according to claim 1, characterized in that, In step S3, GRA, taking values ​​[0,1], represents the relative importance of a specific sequence or variable relative to a reference sequence. Based on the principle of GRA feature analysis, assuming the lidar wind speed sequence in the SCADA dataset is X0, representing the input sequence for GRA analysis, i.e., the target sequence for GRA analysis; the relevant feature value sequence is X... i Let represent the reference sequence for GRA analysis, then: X0=(x 01 ,x 02 ,...,x 0n ) X i =(x i1 ,x i2 ,...,x in ), Then, the input sequence and the reference sequence can be represented by a matrix V, where i = 1, 2, ..., m; Normalize all values ​​in matrix V by scaling them proportionally to the range [0,1]. The normalized matrix is ​​represented as V′: Next, determine the input sequence X0 and the reference sequence X. i The absolute difference Δ between ik , represented as: Δ ik =|x′ 0i -x′ ik | ga=min(Δ ik ) ha=max(Δ ik ), Where k = 1, 2, ..., m; ga represents Δ ik The minimum value, ha represents Δ ik The maximum value is obtained by evaluating the values ​​of ga and ha to obtain the relationship function r(x′) between the reference sequence and the input sequence. 0t ,x′ it ): Where ρ is the discrimination coefficient, with a value of (0,1). In GRA analysis, it is generally considered that ρ is optimal at 0.5, so we take ρ = 0.5; then, the grey relational degree r(X0,X) is determined according to the following formula. i ): Finally, according to r(X0,X) i The value of ) will be used to sort the reference sequence Xi in descending order from smallest to largest and output it for subsequent selection of feature values ​​to be input into the XGBoost model.

6. The wind turbine inflow velocity estimation method based on GRA-XGBoost according to claim 1, characterized in that, The specific evaluation indicators in step S4 are: MAE represents Mean Absolute Error, MSE represents Mean Squared Error, RMSE represents Mean Squared Error, and R... 2 The goodness of fit is expressed by the following formula: Where N is the total number of wind speed value samples recorded in the dataset, v i and It refers to the XGBoost estimated inflow wind speed value falling within this wind speed range and its corresponding actual value, i.e., the lidar wind speed value. The magnitude of all lidar wind speed values ​​within this range.

Images