A wind turbine power curve extraction method based on XAI method

Abstract

The present invention discloses a method for extracting a power curve of a wind turbine based on an XAI method. The method takes output power as output quantity and other wind turbine features as input quantity, trains a neural network, obtains an implicit relationship between output power and each wind turbine feature, and reconstructs a uniform, dense and smooth data set; based on the influence contribution of each wind turbine feature on the output power, screens out the main features with high influence contribution; calculates the local contribution cumulative curve of each main feature, divides the uniform, dense and smooth data set into several intervals, and the output power in different intervals is affected by different main features; performs symbolic regression and Pareto sorting screening on the data of each interval, respectively, and obtains the symbolic expression of the output power of the wind turbine in each interval; and outputs the power curve of the wind turbine under full incoming flow conditions. The present invention improves the calculation accuracy of the power curve.

Description

Technical Field

[0001] The invention belongs to the field of wind power generation, and in particular relates to a method for extracting a power curve of a wind turbine generator set based on an XAI method. Background Art

[0002] The wind turbine power curve is an important indicator for describing the operating performance of the unit. It gives the corresponding curve between the wind turbine output power and wind speed, with the horizontal axis representing the incoming wind speed at the wind turbine and the vertical axis representing the wind turbine output power. The factory theoretical power curve of the wind turbine is obtained by simulation calculation by the manufacturer, but under the actual wind field conditions, affected by factors such as wind conditions, transmission chain damping, control strategy and wind measurement accuracy, the actual power curve of the unit deviates from the theoretical power curve.

[0003] In order to obtain the real power curve of the wind turbine, it is necessary to extract it from the wind turbine operation data. The existing wind turbine power curve measurement method is: collect wind turbine operation SCADA data for more than one year to obtain the ten-minute average wind turbine operation time, wind turbine operation status, wind turbine incoming wind speed, wind turbine output power, remove the abnormal power generation status data, use the ten-minute average wind speed and ten-minute average power of the wind turbine to calculate the average wind speed and average power in each 0.5m / s wind speed interval as the data points of the power curve, and fit to obtain the actual power curve of the wind turbine.

[0004] However, the power curve obtained by the above method only describes the average output power of the wind turbine under the corresponding incoming wind speed conditions, and cannot reflect the power characteristics of the wind turbine output power changing with other operating parameters at the same wind speed. It is not conducive to the accurate prediction of the wind turbine output power under complex incoming flow conditions. At the same time, the simple wind speed small interval segmentation is not conducive to reflecting the impact of changes in the control strategy of the wind turbine in different wind speed intervals. Summary of the invention

[0005] The purpose of the present invention is to propose a method for extracting a power curve of a wind turbine generator set based on an XAI method. The technical solution for implementing the present invention is: a method for extracting a power curve of a wind turbine generator set based on an XAI method, comprising the following steps:

[0006] Step 1: Collect historical operation data sets of wind turbine SCADA system , including incoming wind speed, pitch angle, power coefficient, generator speed, blade speed, incoming wind direction, cabin temperature, cabin position and output power, remove data outliers and null values, align data dimensions and perform normalization to obtain the training set ;

[0007] Step 2: Take the output power as the output and other fan features as the input, and use the training set Train the neural network to obtain the implicit relationship between the output power and each wind turbine feature, perform uniform and dense value selection in each input feature distribution area, use the trained neural network to predict the output power, and reconstruct a uniform, dense and smooth data set , describing the power data of the wind turbine for all possible incoming flows;

[0008] Step 3: Analyze the dataset based on XAI method The contribution of each wind turbine feature to the output power is screened out to identify the main features with the highest contribution. ;

[0009] Step 4: Use the ALE method to calculate the local contribution cumulative curve of each main feature, taking the mutation point as the interval endpoint, and convert the data set Divided into several intervals , the output power in different intervals is affected by different main characteristics ;

[0010] Step 5: Use the symbolic regression method EQL to perform symbolic regression and Pareto sorting on the data in each interval to obtain the symbolic expression of the wind turbine output power in each interval: ;

[0011] Step 6: According to the symbolic expression of the wind turbine output power in each interval, the power curve of the wind turbine under full inflow conditions is obtained.

[0012] Furthermore, in step 1, the historical data set of the wind turbine SCADA system is collected , including incoming wind speed, pitch angle, power coefficient, generator speed, blade speed, incoming wind direction, cabin temperature, cabin position and output power, remove data outliers and null values, align data dimensions and perform normalization to obtain the training set , specifically:

[0013] Collect historical operation data sets of wind turbine SCADA system ,The data features include incoming wind speed, pitch angle, power coefficient, generator speed, blade speed, incoming wind direction, cabin temperature, cabin position and output power;

[0014] Running a historical dataset Preprocessing is performed, including removing null values ​​caused by wind turbine shutdown, removing outliers and outliers caused by power outages and failures, aligning data dimensions, normalizing each data feature to [0,1], and splitting it into training sets. With validation set .

[0015] Furthermore, in step 3, the data set is analyzed based on the XAI method The contribution of each wind turbine feature to the wind turbine power generation, and the main features with high contribution are selected , specifically:

[0016] The sensitivity method is used to calculate the contribution of each fan feature to the output power. The contribution of each fan feature is ranked to obtain the importance ranking of the contribution of different fan features to the output power. The fan features whose mean and variance are both less than the threshold are discarded. The remaining fan features are the main features that have an impact on the output power and are used as the feature dimensions for subsequent power curve prediction.

[0017] Furthermore, in step 4, the local contribution cumulative curve of each main feature is calculated using the ALE method, with the mutation point as the interval endpoint, and the data set is Divided into several intervals , the output power in different intervals is affected by different main characteristics , specifically:

[0018] The ALE method is used to calculate the local contribution cumulative curve of each main feature, and the curve gradient is calculated to obtain the local contribution of each main feature to the output power of the wind turbine. , when the local contribution is not zero, it means that the feature has an impact on the output power at that location;

[0019] The influence interval of each feature is determined based on whether the local contribution is zero, and the data set is divided into Divided into several intervals , the output power of wind turbines in each interval is affected by different main characteristics Influence.

[0020] Furthermore, in step 5, the symbolic regression method EQL is used to perform symbolic regression and Pareto sorting screening on the data in each interval to obtain the symbolic expression of the output power of the wind turbine in each interval: , specifically:

[0021] For interval Medium data, combined with main features is the independent variable, and the output power is the dependent variable. The symbolic algorithm EQL network is used to obtain the symbolic expression, which is used to characterize the power curve in this interval. The wind speed, pitch angle, power coefficient, generator speed, blade speed, wind direction, cabin temperature, cabin position and output power of the wind turbine are all continuous data. The symbolic activation function uses constants, multiplication, power function, cosine function or exponential function.

[0022] Using a validation set The test error is obtained by calculating the mean square error with the predicted value of the mathematical expression, and the expression complexity is calculated using the number of nodes in the EQL network structure. The symbolic regression method is run multiple times on the same interval data, and the test error and expression complexity are used for Pareto sorting to screen out a mathematical expression that takes into account both accuracy and simplicity.

[0023] Furthermore, in step 6, the power curve of the wind turbine under full inflow conditions is obtained according to the prediction formula of the wind turbine power generation sign in each interval, which is specifically:

[0024] The symbolic expression of the output power obtained by summarizing the data of each interval in step 5 is obtained to obtain the mathematical expression of the multi-dimensional power curve of the wind turbine under full inflow conditions, and the real-time power calculation of the wind turbine is performed by inputting the real-time data of the wind turbine.

[0025] A wind turbine power curve extraction system based on an XAI method implements the wind turbine power curve extraction method based on an XAI method to achieve wind turbine power curve extraction based on an XAI method.

[0026] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the wind turbine power curve extraction method based on the XAI method is implemented to realize the wind turbine power curve extraction based on the XAI method.

[0027] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the wind turbine power curve extraction method based on the XAI method is implemented to achieve the wind turbine power curve extraction based on the XAI method.

[0028] Compared with the prior art, the present invention has the following significant advantages: 1) The present invention obtains the multi-dimensional power curve of the wind turbine based on the XAI method, introduces other wind turbine operating parameters besides the incoming wind speed, improves the accuracy of the wind turbine power curve and can be used for wind turbine power prediction under different wind conditions and control conditions; 2) The present invention divides the wind speed interval according to the characteristic contribution based on the XAI method and obtains the power curve expression in each interval segment, reflecting the influence of different control strategies on the wind turbine power curve. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 The present invention is a flow chart of a method for extracting wind turbine power curve based on the XAI method.

[0030] Figure 2 Schematic diagram of the local cumulative effect curve with the main characteristics.

[0031] Figure 3This is a comparison chart between the symbolic regression results and the original data distribution. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0033] like Figure 1 As shown, the wind turbine power curve extraction method based on the XAI method includes the following steps:

[0034] Step 1: Collect historical data sets of wind turbine SCADA system , including incoming wind speed, pitch angle, power coefficient, generator speed, blade speed, incoming wind direction, cabin temperature, cabin position and output power, remove data outliers and null values, align data dimensions and perform normalization to obtain the training set ;

[0035] Collect historical operation data sets of wind turbine SCADA system The data features include information such as incoming wind speed, pitch angle, power coefficient, generator speed, blade speed, incoming wind direction, cabin temperature, cabin position and output power.

[0036] Running a historical dataset Preprocessing is performed, including removing null values ​​caused by wind turbine shutdown, removing outliers and outliers caused by power outages and failures, aligning data dimensions and normalizing each feature to [0,1] to obtain a data set, and then splitting it into a training set With validation set .

[0037] Step 2: Take the output power as the output and other fan features as the input, and use the training set Train the neural network to obtain the implicit relationship between the output power and each wind turbine feature, perform uniform and dense value selection in each input feature distribution area, use the trained neural network to predict the output power, and reconstruct a uniform, dense and smooth data set ;

[0038] Using the training set The neural network is trained with the output power as the output and other fan characteristics as the input, so that the neural network learns the implicit relationship between the output power and each fan characteristic. Uniform and dense values ​​are taken within the original range of each fan characteristic parameter, covering the distribution range of each parameter under normal operating conditions of the fan, and obtaining an input data set that is uniformly and densely distributed in the original parameter space. The trained neural network is used to predict the input data to obtain a reconstructed data set with uniform, dense and smooth distribution of the output power in the original parameter space. , describing the power data of the wind turbine for all possible incoming flows.

[0039] Step 3: Analyze the dataset based on XAI method The contribution of each wind turbine feature to the output power is screened out to identify the main features with the highest contribution. ;

[0040] The sensitivity method is used to calculate the contribution of each fan feature to the output power. The contribution of each fan feature is ranked to obtain the importance ranking of the contribution of different fan features to the output power. The fan features whose average and variance are both less than the threshold are discarded. The remaining fan features are the main features that have an impact on the output power and are used as the feature dimensions for subsequent power curve prediction.

[0041] Step 4: Use the ALE method to calculate the local contribution cumulative curve of each main feature, taking the mutation point as the interval endpoint, and convert the data set Divided into several intervals , the output power in different intervals is affected by different main characteristics ;

[0042] The ALE method is used to calculate the local contribution cumulative curve of each main feature, and the curve gradient is calculated to obtain the local contribution of each main feature to the output power of the wind turbine. When the local contribution is not zero, it means that the feature has an impact on the output power at that location.

[0043] Determine the influence interval of each feature based on whether the local contribution is zero, and divide the data into several intervals based on the influential feature combination , the output power of wind turbines in each interval is affected by different main characteristics Influence.

[0044] Step 5: Use the symbolic regression method EQL to perform symbolic regression and Pareto sorting on the data in each interval to obtain the symbolic expression of the wind turbine output power in each interval: ;

[0045] For interval Medium data, combined with main features The symbolic algorithm EQL network is used to obtain the symbolic expression, which is used to characterize the power curve in this interval. The wind speed, pitch angle, power coefficient, generator speed, blade speed, wind direction, cabin temperature, cabin position and output power of the wind turbine are all continuous data, so the symbolic activation function in the EQL method uses constants, multiplication, power function, cosine function, and exponential function.

[0046] Using a validation set The test error is obtained by calculating the mean square error with the predicted value of the mathematical expression, and the expression complexity is calculated using the number of nodes in the EQL network structure. The symbolic regression method is run multiple times on the same interval data, and the test error and expression complexity are used for Pareto sorting to screen out mathematical expressions that take into account both accuracy and simplicity.

[0047] Step 6: According to the prediction formula of the output power sign of the wind turbine in each interval, the power curve of the wind turbine under the full inflow condition is obtained.

[0048] The output power symbolic expression obtained by summarizing the data of each interval in step 5 is obtained to obtain the mathematical expression of the multi-dimensional power curve of the wind turbine under full inflow conditions. The real-time power calculation of the wind turbine can be performed by inputting the real-time data of the wind turbine.

[0049] A wind turbine power curve extraction system based on an XAI method implements the wind turbine power curve extraction method based on an XAI method to achieve wind turbine power curve extraction based on an XAI method.

[0050] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the wind turbine power curve extraction method based on the XAI method is implemented to realize the wind turbine power curve extraction based on the XAI method.

[0051] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the wind turbine power curve extraction method based on the XAI method is implemented to achieve the wind turbine power curve extraction based on the XAI method.

[0052] In summary, the present invention can provide a mathematical expression of a multi-dimensional power curve of a wind turbine generator set from historical data of the wind turbine generator set, thereby improving the calculation accuracy of the power curve.

[0053] Example

[0054] In order to verify the effectiveness of the solution of the present invention, the following experiment was conducted using wind turbine power collection data from a real wind farm.

[0055] 1) Collect data and perform preprocessing

[0056] A single wind turbine was selected for data collection in a real domestic wind farm. SCADA and LiDAR data were used as data sources to collect information such as the wind speed, pitch angle, power coefficient, generator speed, blade speed, wind direction, cabin temperature, cabin position and output power of the wind turbine. The data is the average data of ten minutes. The operation data of the wind turbine from November 2019 to April 2020 were collected.

[0057] For the collected data, null values ​​during downtime and maintenance were removed, outliers were removed through quartiles and normalized. The data had a total of nine dimensions. After processing, 24,453 groups of data were obtained, of which 70% were training sets, 20% were test sets, and 10% were validation sets.

[0058] 2) Use neural networks to learn and reconstruct wind turbine output power data

[0059] The neural network is trained using the training set. The number of hidden layers of the neural network is 3, and the number of training steps is 5000. Within the distribution range of the original data, each input quantity is uniformly and densely selected as the input set, and the trained neural network is input to obtain the output power prediction reconstruction value of the wind turbine under full inflow conditions.

[0060] The accuracy of the output power predicted by the neural network compared with the test set data is 91.48%, indicating that the neural network has well learned the potential relationship between the wind turbine output power data and other wind turbine characteristics.

[0061] 3) Screen the main features of the power curve and divide the data interval

[0062] Input the reconstructed data into the XAI algorithm module to obtain the Sensitivity result ranking and the feature importance curve of the ALE method as shown below: Figure 2 As shown. Based on the calculation and sorting of feature importance, the main features include incoming wind speed, generator speed and pitch angle. According to the cumulative effect diagram, the incoming wind speed has a positive contribution within 3-9m / s and no contribution within 9-14m / s; the blade speed has a positive contribution in the entire area; the pitch angle has a positive contribution in the entire area. Therefore, the reconstructed data can be divided into two intervals with 9m / s as the endpoint. The first interval corresponds to the incoming wind speed in the range of 3-9m / s, and the main features include incoming wind speed, blade speed, and pitch angle; the second interval corresponds to the interval above 9m / s, and the main features include blade speed and pitch angle. Observing the distribution of the original data, when the incoming wind speed is within 9-14m / s, the blade speed is stable at 14.8rpm and the output power is stable at 2230KW, so the power curve in this interval is a fixed value, and the main feature is the pitch angle.

[0063] 4) Obtain the power curve expression by symbolic regression and verify it

[0064] The EQL network is used to perform symbolic regression on the data in the two intervals, and the Pareto analysis is performed to screen the best mathematical expression based on the test error and expression complexity. The results are shown in the following formulas:

[0065]

[0066] In the formula is the output power of the wind turbine, is the intermediate variable, is the incoming wind speed, is the blade speed, is the pitch angle.

[0067] Figure 3 The comparison between the symbolic regression results and the original data distribution is shown. The prediction accuracy of the mathematical expression prediction formula compared with the validation set data reaches 86.28%, indicating that the obtained multi-dimensional power curve of the wind turbine can well predict the output power of the wind turbine under different wind conditions and control conditions.

[0068] The above embodiments show that the method proposed in the present invention provides a mathematical expression with good accuracy and simplicity for complex wind turbine output power data, explicitly expresses the mathematical laws learned by the neural network from the wind turbine power data, and obtains a multi-dimensional power curve that can be used for fast calculation of the wind turbine. It is suitable for calculating, verifying and predicting the output power characteristics in the wind turbine.

[0069] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0070] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A wind turbine power curve extraction method based on XAI method, characterized in that: The steps include: Step 1: Collect historical operation data sets of wind turbine SCADA system , including incoming wind speed, pitch angle, power coefficient, generator speed, blade speed, incoming wind direction, cabin temperature, cabin position and output power, remove data outliers and null values, align data dimensions and perform normalization to obtain the training set ; Step 2: Take the output power as the output and other fan features as the input, and use the training set Train the neural network to obtain the implicit relationship between the output power and each wind turbine feature, perform uniform and dense value selection in each input feature distribution area, use the trained neural network to predict the output power, and reconstruct a uniform, dense and smooth data set , describing the power data of the wind turbine for all possible incoming flows; Step 3: Analyze the dataset based on XAI method The contribution of each wind turbine feature to the output power is screened out to identify the main features with the highest contribution. ; Step 4: Use the ALE method to calculate the local contribution cumulative curve of each main feature, taking the mutation point as the interval endpoint, and convert the data set Divided into several intervals , the output power in different intervals is affected by different main characteristics ; Step 5: Use the symbolic regression method EQL to perform symbolic regression and Pareto sorting on the data in each interval to obtain the symbolic expression of the wind turbine output power in each interval: ; Step 6: According to the symbolic expression of the wind turbine output power in each interval, the power curve of the wind turbine under full inflow conditions is obtained.

2. The method for extracting wind turbine power curve based on XAI method according to claim 1, characterized in that: In step 1, collect the historical data set of the wind turbine SCADA system , including incoming wind speed, pitch angle, power coefficient, generator speed, blade speed, incoming wind direction, cabin temperature, cabin position and output power, remove data outliers and null values, align data dimensions and perform normalization to obtain the training set , specifically: Collect historical operation data sets of wind turbine SCADA system ,The data features include incoming wind speed, pitch angle, power coefficient, generator speed, blade speed, incoming wind direction, cabin temperature, cabin position and output power; Running a historical dataset Preprocessing is performed, including removing null values ​​caused by wind turbine shutdown, removing outliers and outliers caused by power outages and failures, aligning data dimensions, normalizing each data feature to [0,1], and splitting it into training sets. With validation set .

3. The method for extracting wind turbine power curve based on XAI method according to claim 1, characterized in that: In step 3, the dataset is analyzed based on the XAI method The contribution of each wind turbine feature to the wind turbine power generation, and the main features with high contribution are selected , specifically: The sensitivity method is used to calculate the contribution of each fan feature to the output power. The contribution of each fan feature is ranked to obtain the importance ranking of the contribution of different fan features to the output power. The fan features whose mean and variance are both less than the threshold are discarded. The remaining fan features are the main features that have an impact on the output power and are used as the feature dimensions for subsequent power curve prediction.

4. The method for extracting wind turbine power curve based on XAI method according to claim 1, characterized in that: In step 4, the local contribution cumulative curve of each main feature is calculated using the ALE method, with the mutation point as the interval endpoint, and the data set is Divided into several intervals , the output power in different intervals is affected by different main characteristics , specifically: The ALE method is used to calculate the local contribution cumulative curve of each main feature, and the curve gradient is calculated to obtain the local contribution of each main feature to the output power of the wind turbine. , when the local contribution is not zero, it means that the feature has an impact on the output power at that location; The influence interval of each feature is determined based on whether the local contribution is zero, and the data set is divided into Divided into several intervals , the output power of wind turbines in each interval is affected by different main characteristics Influence.

5. The method for extracting wind turbine power curve based on XAI method according to claim 1, characterized in that: In step 5, the symbolic regression method EQL is used to perform symbolic regression and Pareto sorting screening on the data in each interval to obtain the symbolic expression of the output power of the wind turbine in each interval: , specifically: For interval Medium data, combined with main features is the independent variable, and the output power is the dependent variable. The symbolic algorithm EQL network is used to obtain the symbolic expression, which is used to characterize the power curve in this interval. The wind speed, pitch angle, power coefficient, generator speed, blade speed, wind direction, cabin temperature, cabin position and output power of the wind turbine are all continuous data. The symbolic activation function uses constants, multiplication, power function, cosine function or exponential function. Using a validation set The test error is obtained by calculating the mean square error with the predicted value of the mathematical expression, and the expression complexity is calculated using the number of nodes in the EQL network structure. The symbolic regression method is run multiple times on the same interval data, and the test error and expression complexity are used for Pareto sorting to screen out a mathematical expression that takes into account both accuracy and simplicity.

6. The method for extracting wind turbine power curve based on XAI method according to claim 1, characterized in that: In step 6, the power curve of the wind turbine under full inflow conditions is obtained according to the prediction formula of the wind turbine power generation sign in each interval, which is: The symbolic expression of the output power obtained by summarizing the data of each interval in step 5 is obtained to obtain the mathematical expression of the multi-dimensional power curve of the wind turbine under full inflow conditions, and the real-time power calculation of the wind turbine is performed by inputting the real-time data of the wind turbine.

7. A wind turbine power curve extraction system based on XAI method, characterized in that: The wind turbine power curve extraction method based on the XAI method as described in any one of claims 1 to 6 is implemented to realize the wind turbine power curve extraction based on the XAI method.

8. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method for extracting a wind turbine power curve based on the XAI method according to any one of claims 1 to 6 is implemented to achieve the extraction of a wind turbine power curve based on the XAI method.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the wind turbine power curve extraction method based on the XAI method according to any one of claims 1 to 6 is implemented to achieve wind turbine power curve extraction based on the XAI method.

Images