Wind turbine generator gearbox oil pool temperature prediction method

Abstract

The invention discloses a wind turbine generator gearbox oil pool temperature prediction method. The method comprises the following specific steps: S1, acquiring basic information data of a wind turbine generator; s2, cleaning the original operation data of the wind turbine generator, removing the abnormal working condition operation data of the wind turbine generator, and improving the model prediction precision; s3, the distance correlation coefficient is utilized, the gearbox oil pool temperature serves as a target parameter, and parameters with high correlation with the target parameter are screened out to serve as input parameters of the oil temperature prediction model; s4, taking the screened parameters as input parameters of a prediction model, taking the oil pool temperature as output parameters, and adopting a light GBM algorithm to establish a gearbox oil pool temperature prediction model; and S5, optimizing prediction model parameter setting by using a TPE algorithm, and improving model prediction precision.

Description

Technical Field

[0001] The invention belongs to the technical field of wind turbine fault warning, and in particular relates to a method for predicting the temperature of an oil pool in a gear box of a wind turbine. Background Art

[0002] The gearbox is an important component of a wind turbine. Due to the influence of factors such as harsh operating environment and changeable working conditions, gearbox failures occur frequently. After a gearbox failure, the unit will be shut down for a long time and the maintenance cost will be high, which will have a great impact on the safety, reliability and operating efficiency of the wind farm.

[0003] When the wind turbine gearbox is running, the heat generated will be conducted between the gear and the lubricant due to factors such as frictional heat, ambient temperature fluctuations, load and speed changes. The wind turbine SCADA system collects the operating parameters of wind turbines, substations, meteorological stations and other equipment. Through the mining and analysis of SCADA data, the operating status of the wind turbine gearbox can be evaluated, which is an important basis for fault diagnosis and optimal scheduling of wind turbine gearboxes. Summary of the invention

[0004] The purpose of the present invention is to provide a method for predicting the oil pool temperature of a wind turbine gearbox, which uses the quartile method and a longitudinal filtering method based on data discreteness to clean the wind speed-power data, and performs feature selection through the distance correlation coefficient to construct a Bayesian optimized LightGBM oil pool temperature prediction model, aiming to fully explore the correlation between data and improve the reliability of wind turbine gearbox operating condition monitoring.

[0005] The technical solution adopted by the present invention is a method for predicting the temperature of the oil pool of a wind turbine gearbox, and the specific steps are as follows:

[0006] S1. Obtain basic information data of wind turbines;

[0007] S2. Clean the basic information data of the wind turbine generator set and remove the abnormal operating data of the wind turbine generator set;

[0008] S3. Calculate the correlation between the gearbox oil pool temperature and the candidate parameters using a distance correlation coefficient algorithm, and select parameters whose correlation with the gearbox oil pool temperature is greater than a threshold as input parameters;

[0009] S4, using the parameters screened in S3 as input parameters of the prediction model, using the gearbox oil pool temperature as the output parameter, and using the lightGBM algorithm to establish a gearbox oil pool temperature prediction model;

[0010] S5. Use the TPE algorithm to optimize the parameter settings of the gearbox oil pool temperature prediction model to improve the prediction accuracy of the gearbox oil pool temperature prediction model.

[0011] The present invention is also characterized in that:

[0012] The basic information data of wind turbines in S1 include wind turbine cut-in wind speed, rated wind speed, cut-out wind speed, power curve and unit SCADA data.

[0013] S2 is to eliminate the abnormal operating data of the wind turbine according to the wind speed-power curve distribution, and obtain the operating data of the wind turbine under normal operating conditions. The specific operation steps are as follows:

[0014] S2.1. Perform the first cleanup of wind speed-power based on the quartile method, as follows:

[0015] First, remove the data points with wind speed less than the cut-in wind speed and greater than the cut-off wind speed. According to the distribution of wind speed-power data, divide the wind speed axis into two wind speed segments: [cut-in wind speed, rated wind speed] and [rated wind speed, cut-out wind speed]. Divide the data of each wind speed segment into four parts according to the power value. The values ​​at the three split points are the lower quantile Q1, median Q2 and upper quantile Q3, respectively. For a sample X={x1,x2,…,x n-1 ,x n}, x in the sequence i <x i+1 , n is the total number of samples, i represents a point in the sequence,

[0016] First calculate the second median:

[0017]

[0018] Then calculate the lower quantile Q1 and the upper quantile Q3,

[0019] Q1 and Q3 represent the values ​​represented by the positions of 25% of the data points before and after the separation in sample X. When n = 2k (k = 0, 1, 2...),

[0020]

[0021] When n=4k+1(k=0,1,2…), we have

[0022]

[0023] Through the calculation of formula (2) and formula (3), the interquartile range I QR ,Right now:

[0024] I QR =Q3-Q1 (4)

[0025] In statistics, I QR Determine the inner limit range of sequence X, that is:

[0026] [F1,F u ]=[Q1-ω1I QR ,Q3+ω2I QR ] (5)

[0027] Where: F1 is the lower limit of X determined by the quartile method; F u is the upper limit of X; w1 and w2 are weights,

[0028] Take ω1=ω2=1.5, if Determined as abnormal data, removed; otherwise x i Retain, complete the first cleaning of wind speed-power based on the quartile method;

[0029] S2.2, use the longitudinal filtering method based on data discreteness for the second cleaning,

[0030] For the power value in each wind speed sub-interval, calculate its mean and variance, that is:

[0031]

[0032] In formula (6): i is the power mean square error of the i-th interval; N i is the scatter value of the ith interval; x j is the jth power value of the ith interval; μ i is the power mean of the ith interval,

[0033] When the unit is operating normally, the wind speed and power scatter points are normally distributed around the power band, that is, 95% of the data are within the confidence interval covered by the two mean square errors above and below the mean. Based on this, each wind speed interval is based on the power mean μ i With the power band as the center, discrete data around the power band can be effectively eliminated.

[0034] The specific methods of S3 are:

[0035] Assume that two variables are X and Y, and the correlation between the two variables is recorded as dCor(X,Y). When dCor(X,Y)=0, it means that the two variables X and Y are independent. The larger the dCor(X,Y), the stronger the correlation between the variables X and Y. The algorithm for the distance correlation coefficient is as shown in formula (7):

[0036]

[0037] Where: dCor(X,Y) is the distance correlation coefficient between X and Y; dCov(X,Y) is the distance covariance between X and Y; dVar(X) is the variance of X; dVar(Y) is the variance of Y;

[0038] The threshold is taken as 0.5, and the parameters with dCor(X,Y)>0.5 are taken as the input parameters of the prediction model.

[0039] The specific operation method of S4 is as follows:

[0040] LightGBM makes the GBDT algorithm more efficient through two improved technologies: unilateral gradient descent and mutually exclusive feature bundling. The unilateral gradient descent algorithm is as follows:

[0041] Assume that the training set data S = {(x1,y1),(x2,y2),…,(x n ,y n )}, where x is the input feature vector and y is the output data. The goal of the LightGBM model optimization process is to find a function Minimize the loss function L(y,f(x)). If the model prediction value is recorded as f(x), then the function It can be expressed as:

[0042]

[0043] Assuming that the number of decision trees in the model is M, since the model uses the forward distribution algorithm, the output value of the model at the Mth tree can be expressed as:

[0044]

[0045] At the mth tree, the loss function of the model is recorded as:

[0046]

[0047] In the LightGBM model, the function is approximated by Newton's method. After removing the constant term in the above formula, the loss function can be equivalent to:

[0048]

[0049] Among them, g i ,h i Represents the first-order and second-order gradients of the loss function. Since the model uses decision trees as the basis function, each tree can be expressed as: q(x) , q∈{1,2,…,J}, J represents the number of leaves in the decision tree, q represents the decision rule, and w represents the weight vector composed of leaf nodes. Therefore, formula (11) can be transformed into:

[0050]

[0051] For a tree of a specific structure, the optimal weight w of each leaf node is j * and the extreme value Γ of the loss function kIt can be solved as:

[0052]

[0053] in, It can be regarded as a scoring function to measure the quality of the tree structure. Finally, the loss function is divided into:

[0054]

[0055] I L , I R They are sample sets for the left and right branches respectively. By introducing the one-sided gradient descent algorithm, the model training speed is further improved. At the same time, the diversity of weak learners is increased, and the generalization ability of the model is further improved.

[0056] The main principle of the TPE algorithm in S5 is to model the objective function and use Bayes' theorem to continuously update the prior probability distribution so as to select the search point that is more likely to produce the best result in the next step. The main process is as follows:

[0057] The objective function f(x) is less than a certain threshold y * The expectation is used as the evaluation index of the algorithm, denoted as EI (Expected Improvement):

[0058]

[0059] Define the following two probability densities:

[0060]

[0061] Where g(x) is the objective function f(x i ) value is greater than or equal to the threshold y * The corresponding {x i}Probability density of the set; l(x) is the opposite, indicating that it is less than the threshold y * The corresponding {x i}Probability density of a set;

[0062] Since the probability of p(y|x) cannot be obtained directly, the Bayesian formula is used to convert the above formula into:

[0063]

[0064] Now the probability of being less than the threshold value y* is recorded as γ = p(y <y * ),but:

[0065] p(x)=∫p(x|y)p(y)dy=γl(x)+(1-γ)g(x) (19)

[0066] It is deduced that:

[0067]

[0068] It can be seen from formula (20) that the size of the evaluation index depends on the probability density function g(x) and l(x). The smaller the ratio of g(x) to l(x), the larger the evaluation index EI.

[0069] In the TPE algorithm, the probability density functions g(x) and l(x) are represented using a tree structure. This representation method is more convenient for sampling according to l(x). Through this method, the x that maximizes the evaluation index EI can be obtained in each iteration. * .

[0070] The beneficial effects of the present invention are:

[0071] (1) The wind turbine gearbox oil pool temperature prediction method provided by the present invention uses the quartile method and the longitudinal filtering method based on data discreteness to clean the wind speed-power data, and performs feature selection through the distance correlation coefficient to construct the Bayesian optimization LightGBM oil pool temperature prediction model, aiming to fully explore the correlation between the data and improve the reliability of wind turbine gearbox operating condition monitoring;

[0072] (2) The wind turbine gearbox oil pool temperature prediction method of the present invention can effectively mine the wind turbine SCADA data, quickly and accurately predict the gearbox oil pool temperature, monitor the gearbox operating status in real time, promptly detect abnormal conditions and take corresponding measures, reduce equipment downtime, and improve the operating efficiency of the unit; at the same time, reduce the workload of on-site operation and maintenance personnel, reduce labor costs, and thus improve the operating efficiency of the wind farm. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 The schematic diagram of the implementation process of the wind speed-power data cleaning method of the wind turbine generator set is as follows;

[0074] Figure 2 It is a flow chart of the wind turbine gearbox oil pool temperature prediction method. DETAILED DESCRIPTION

[0075] The present invention is described in detail below in conjunction with specific implementation modes.

[0076] The wind turbine gearbox oil pool temperature prediction method of the present invention is as follows: Figure 1 As shown in Figure 2, the wind speed-power data are cleaned by using the quartile method and the longitudinal filtering method based on data dispersion, and feature selection is performed by using the distance correlation coefficient, as shown in Figure 2. Figure 2As shown in the figure, the Bayesian optimization LightGBM oil pool temperature prediction model is constructed to fully explore the correlation between data and improve the reliability of wind turbine gearbox operating condition monitoring. The specific implementation steps are as follows:

[0077] S1. Obtain basic information data of wind turbines;

[0078] Among them, the basic information data includes wind turbine cut-in wind speed, rated wind speed, cut-out wind speed, power curve, unit SCADA data, etc.

[0079] S2. Clean the original operation data (basic information data) of the wind turbine set, remove the abnormal operation data of the wind turbine set according to the wind speed-power curve distribution, improve the model prediction accuracy, and obtain the operation data of the wind turbine set under normal conditions. The specific operation steps are as follows:

[0080] S2.1. Perform the first cleanup of wind speed-power based on the quartile method, as follows:

[0081] First, remove the data points where the wind speed is less than the cut-in wind speed and greater than the cut-off wind speed. According to the distribution of wind speed-power data, the wind speed axis is divided into two wind speed segments: [cut-in wind speed, rated wind speed] and [rated wind speed, cut-out wind speed]. The data of each wind speed segment is evenly divided into four parts according to the power value. The values ​​at the three split points are the lower quantile Q1, the median Q2 and the upper quantile Q3, respectively. For a sample X={x1,x2,…,x n-1 ,x n}, x in the sequence i <x i+1 , n is the total number of samples, i represents a point in the sequence,

[0082] First calculate the second median:

[0083]

[0084] Then calculate the lower quantile Q1 and the upper quantile Q3,

[0085] Q1 and Q3 represent the values ​​represented by the positions of 25% of the data points before and after the separation in sample X. When n = 2k (k = 0, 1, 2...),

[0086]

[0087] When n=4k+1(k=0,1,2…), we have

[0088]

[0089] Through the calculation of formula (2) and formula (3), the interquartile range I QR,Right now:

[0090] I QR =Q3-Q1 (4)

[0091] In statistics, I QR Determine the inner limit range of sequence X, that is:

[0092] [F1,F u ]=[Q1-ω1I QR ,Q3+ω2I QR ] (5)

[0093] Where: F1 is the lower limit of X determined by the quartile method; F u is the upper limit of X; w1 and w2 are weights,

[0094] Take ω1=ω2=1.5, if Determined as abnormal data, removed; otherwise x i Retain, complete the first cleaning of wind speed-power based on the quartile method;

[0095] S2.2, use the longitudinal filtering method based on data discreteness for the second cleaning,

[0096] For the power value in each wind speed sub-interval, calculate its mean and variance, that is:

[0097]

[0098] In formula (6): i is the power mean square error of the i-th interval; N i is the scatter value of the ith interval; x j is the jth power value of the ith interval; μ i is the power mean of the ith interval,

[0099] When the unit is operating normally, the wind speed and power scatter points are normally distributed around the power band, that is, 95% of the data are within the confidence interval covered by the two mean square errors above and below the mean. Based on this, each wind speed interval is based on the power mean μ i With the power band as the center, discrete data around the power band can be effectively eliminated.

[0100] S3. Calculate the correlation between the gearbox oil pool temperature and the candidate parameters using a distance correlation coefficient algorithm, and select parameters whose correlation with the gearbox oil pool temperature is greater than a threshold as input parameters;

[0101] When the wind turbine gearbox is running, due to factors such as frictional heat, ambient temperature fluctuations, load and speed changes, heat conduction will occur between the gear and the lubricating oil. The increase in oil temperature causes the oil to decompose and the viscosity to decrease, affecting the lubrication effect, which in turn causes local overheating, resulting in tooth surface bonding, and prone to gearbox oil temperature being too high, tooth surface wear and other faults. Therefore, the gearbox oil pool temperature is a key parameter to characterize the operating status of the gearbox. Taking the gearbox oil pool temperature as the target parameter, the distance correlation coefficient method is used to select parameters with strong correlation with it as the input parameters of the oil temperature prediction model;

[0102] The specific method is:

[0103] Assume that two variables are X and Y, and the correlation between the two variables is recorded as dCor(X,Y). When dCor(X,Y)=0, it means that the two variables X and Y are independent. The larger the dCor(X,Y), the stronger the correlation between the variables X and Y. The formula of the distance correlation coefficient is as follows:

[0104]

[0105] Where: dCor(X,Y) is the distance correlation coefficient between X and Y; dCov(X,Y) is the distance covariance between X and Y; dVar(X) is the variance of X; dVar(Y) is the variance of Y;

[0106] The threshold is taken as 0.5, and the parameters with dCor(X,Y)>0.5 are taken as the input parameters of the prediction model.

[0107] S4, using the parameters screened in S3 as input parameters of the prediction model, taking the oil pool temperature as the output parameter, and using the lightGBM algorithm to establish a gearbox oil pool temperature prediction model;

[0108] In the machine learning model, hyperparameters (such as decision tree depth, number of neural network neurons, etc.) need to be set in advance before starting training. These parameters will not be automatically updated as the model is trained. If the hyperparameters are set properly, the machine learning model will obtain better training and testing results. However, how to set appropriate parameters is not an easy task. Even for experts, adjusting parameters also consumes a lot of time costs. Therefore, the Bayesian optimization algorithm is used to optimize the prediction model parameters established in step S14 to improve the model prediction accuracy.

[0109] The specific operation method of S4 is as follows:

[0110] The LightGBM (light gradient boosting machine) model is a gradient boosting framework based on decision trees. Its calculation process follows the traditional GBDT (Gradient boosting decision tree) algorithm. It makes the GBDT algorithm more efficient through two improved techniques: one-sided gradient descent (GOSS) and mutually exclusive feature bundling (EFB). The one-sided gradient descent algorithm is as follows:

[0111] Assume that the training set data S = {(x1,y1),(x2,y2),…,(x n ,y n )}, where x is the input feature vector and y is the output data. The goal of the LightGBM model optimization process is to find a function Minimize the loss function L(y,f(x)). If the model prediction value is recorded as f(x), then the function It can be expressed as:

[0112]

[0113] Assuming that the number of decision trees in the model is M, since the model uses the forward distribution algorithm, the output value of the model at the Mth tree can be expressed as:

[0114]

[0115] At the mth tree, the loss function of the model is recorded as:

[0116]

[0117] In the LightGBM model, the function is approximated by Newton's method. After removing the constant term in the above formula, the loss function can be equivalent to:

[0118]

[0119] Among them, g i ,h i Represents the first-order and second-order gradients of the loss function. Since the model uses decision trees as the basis function, each tree can be expressed as: q(x) , q∈{1,2,…,J}, J represents the number of leaves in the decision tree, q represents the decision rule, and w represents the weight vector composed of leaf nodes. Therefore, formula (11) can be transformed into:

[0120]

[0121] For a tree of a specific structure, the optimal weight w of each leaf node is j * and the extreme value Γ of the loss functionk It can be solved as:

[0122]

[0123] in, It can be regarded as a scoring function to measure the quality of the tree structure. Finally, the loss function is divided into:

[0124]

[0125] I L , I R They are sample sets for the left and right branches respectively. By introducing the one-sided gradient descent algorithm, the model training speed is further improved. At the same time, the diversity of weak learners is increased, and the generalization ability of the model is further improved.

[0126] S5. Use TPE algorithm to optimize the prediction model parameter settings and improve the model prediction accuracy.

[0127] The TPE (Tree-structured Parzen Estimator) search algorithm is a hyperparameter optimization algorithm based on a Gaussian mixture model. The algorithm can use the performance of parameters that have been searched during training to infer the change trend of the next parameter, thereby reducing the search space and greatly improving the search efficiency. The main principle of the algorithm is to model the objective function and use Bayes' theorem to continuously update the prior probability distribution so as to select the search point that is more likely to produce the best result in the next step. The main process is as follows:

[0128] The objective function f(x) is less than a certain threshold y * The expectation is used as the evaluation index of the algorithm, denoted as EI (Expected Improvement):

[0129]

[0130] Define the following two probability densities:

[0131]

[0132] Where g(x) is the objective function f(x i ) value is greater than or equal to the threshold y * The corresponding {x i}Probability density of the set; l(x) is the opposite, indicating that it is less than the threshold y * The corresponding {x i}Probability density of a set;

[0133] Since the probability of p(y|x) cannot be obtained directly, the Bayesian formula is used to convert the above formula into:

[0134]

[0135] Now the probability of being less than the threshold value y* is recorded as γ = p(y <y * ),but:

[0136] p(x)=∫p(x|y)p(y)dy=γl(x)+(1-γ)g(x) (19)

[0137] It is deduced that:

[0138]

[0139] It can be seen from formula (20) that the size of the evaluation index depends on the probability density function g(x) and l(x). The smaller the ratio of g(x) to l(x), the larger the evaluation index EI.

[0140] In the TPE algorithm, the probability density functions g(x) and l(x) are represented using a tree structure. This representation method is more convenient for sampling according to l(x). Through this method, the x that maximizes the evaluation index EI can be obtained in each iteration. * .

[0141] Example 1

[0142] The wind turbine gearbox oil pool temperature prediction method of the present invention has the following specific implementation steps:

[0143] S1. Obtain basic information data of wind turbines, including wind turbine cut-in wind speed, rated wind speed, cut-out wind speed, power curve, unit SCADA data, etc.

[0144] S2. Clean the original operation data of the wind turbine set. According to the wind speed-power curve distribution, remove the abnormal operation data of the wind turbine set, improve the prediction accuracy of the model, and obtain the operation data of the unit under normal conditions. The specific operation steps are as follows:

[0145] S2.1. Perform the first cleanup of wind speed-power based on the quartile method, as follows:

[0146] First, remove the data points where the wind speed is less than the cut-in wind speed and greater than the cut-off wind speed. According to the distribution of wind speed-power data, the wind speed axis is divided into two wind speed segments: [cut-in wind speed, rated wind speed] and [rated wind speed, cut-out wind speed]. The data of each wind speed segment is evenly divided into four parts according to the power value. The values ​​at the three split points are the lower quantile Q1, the median Q2 and the upper quantile Q3, respectively. For a sample X={x1,x2,…,x n-1 ,x n}, x in the sequence i<x i+1 , n is the total number of samples, i represents a point in the sequence,

[0147] First calculate the second median:

[0148]

[0149] Then calculate the lower quantile Q1 and the upper quantile Q3,

[0150] Q1 and Q3 represent the values ​​represented by the positions of 25% of the data points before and after the separation in sample X. When n = 2k (k = 0, 1, 2...),

[0151]

[0152] When n=4k+1(k=0,1,2…), we have

[0153]

[0154] Through the calculation of formula (2) and formula (3), the interquartile range I QR ,Right now:

[0155] I QR =Q3-Q1 (4)

[0156] In statistics, I QR Determine the inner limit range of sequence X, that is:

[0157] [F1,F u ]=[Q1-ω1I QR ,Q3+ω2I QR ] (5)

[0158] Where: F1 is the lower limit of X determined by the quartile method; F u is the upper limit of X; w1 and w2 are weights,

[0159] Take ω1=ω2=1.5, if Determined as abnormal data, removed; otherwise x i Retain, complete the first cleaning of wind speed-power based on the quartile method;

[0160] S2.2, use the longitudinal filtering method based on data discreteness for the second cleaning,

[0161] For the power value in each wind speed sub-interval, calculate its mean and variance, that is:

[0162]

[0163] In formula (6): i is the power mean square error of the i-th interval; Ni is the scatter value of the ith interval; x j is the jth power value of the ith interval; μ i is the power mean of the ith interval,

[0164] When the unit is operating normally, the wind speed and power scatter points are normally distributed around the power band, that is, 95% of the data are within the confidence interval covered by the two mean square errors above and below the mean. Based on this, each wind speed interval is based on the power mean μ i With the power band as the center, discrete data around the power band can be effectively eliminated.

[0165] S3, using the distance correlation coefficient, taking the gearbox oil pool temperature as the target parameter, and selecting parameters with strong correlation with it as input parameters of the oil temperature prediction model;

[0166] When the wind turbine gearbox is running, due to factors such as frictional heat, ambient temperature fluctuations, load and speed changes, heat conduction will occur between the gear and the lubricating oil. The increase in oil temperature causes the oil to decompose and the viscosity to decrease, affecting the lubrication effect, which in turn causes local overheating, resulting in tooth surface bonding, and prone to gearbox oil temperature being too high, tooth surface wear and other faults. Therefore, the gearbox oil pool temperature is a key parameter to characterize the operating status of the gearbox. Taking the gearbox oil pool temperature as the target parameter, the distance correlation coefficient method is used to select parameters with strong correlation with it as the input parameters of the oil temperature prediction model;

[0167] S4, using the parameters screened in S3 as input parameters of the prediction model, taking the oil pool temperature as the output parameter, and using the lightGBM algorithm to establish a gearbox oil pool temperature prediction model;

[0168] In the machine learning model, hyperparameters (such as decision tree depth, number of neural network neurons, etc.) need to be set in advance before starting training. These parameters will not be automatically updated as the model is trained. If the hyperparameters are set properly, the machine learning model will obtain better training and testing results. However, how to set appropriate parameters is not an easy task. Even for experts, adjusting parameters also consumes a lot of time costs. Therefore, the Bayesian optimization algorithm is used to optimize the prediction model parameters established in step S14 to improve the model prediction accuracy.

[0169] S5. Use TPE algorithm to optimize the prediction model parameter settings and improve the model prediction accuracy.

[0170] Example 2

[0171] The wind turbine gearbox oil pool temperature prediction method of the present invention has the following specific implementation steps:

[0172] S1. Obtain basic information data of wind turbines;

[0173] S2. Clean the original operation data of the wind turbine set, remove the abnormal operation data of the wind turbine set according to the wind speed-power curve distribution, improve the prediction accuracy of the model, and obtain the operation data of the wind turbine set under normal conditions;

[0174] S3. Using the distance correlation coefficient, taking the gearbox oil pool temperature as the target parameter, select parameters with strong correlation as the input parameters of the oil temperature prediction model. The specific method is as follows:

[0175] Assume that two variables are X and Y, and the correlation between the two variables is recorded as dCor(X,Y). When dCor(X,Y)=0, it means that the two variables X and Y are independent. The larger the dCor(X,Y), the stronger the correlation between the variables X and Y. The formula of the distance correlation coefficient is as follows:

[0176]

[0177] Where: dCor(X,Y) is the distance correlation coefficient between X and Y; dCov(X,Y) is the distance covariance between X and Y; dVar(X) is the variance of X; dVar(Y) is the variance of Y.

[0178] The parameters with dCor(X,Y)>0.5 are taken as the input parameters of the prediction model.

[0179] S4. The parameters screened in S3 are used as input parameters of the prediction model, and the oil pool temperature is used as the output parameter. The lightGBM algorithm is used to establish a gearbox oil pool temperature prediction model. The specific operation method is as follows:

[0180] The LightGBM (light gradient boosting machine) model is a gradient boosting framework based on decision trees. Its calculation process follows the traditional GBDT (Gradient boosting decision tree) algorithm. It makes the GBDT algorithm more efficient through two improved techniques: one-sided gradient descent (GOSS) and mutually exclusive feature bundling (EFB). The one-sided gradient descent algorithm is as follows:

[0181] Assume that the training set data S = {(x1,y1),(x2,y2),…,(x n ,y n )}, where x is the input feature vector and y is the output data. The goal of the LightGBM model optimization process is to find a function Minimize the loss function L(y,f(x)). If the model prediction value is recorded as f(x), then the function It can be expressed as:

[0182]

[0183] Assuming that the number of decision trees in the model is M, since the model uses the forward distribution algorithm, the output value of the model at the Mth tree can be expressed as:

[0184]

[0185] At the mth tree, the loss function of the model is recorded as:

[0186]

[0187] In the LightGBM model, the function is approximated by Newton's method. After removing the constant term in the above formula, the loss function can be equivalent to:

[0188]

[0189] Among them, g i ,h i Represents the first-order and second-order gradients of the loss function. Since the model uses decision trees as the basis function, each tree can be expressed as: q(x) , q∈{1,2,…,J}, J represents the number of leaves in the decision tree, q represents the decision rule, and w represents the weight vector composed of leaf nodes. Therefore, formula (11) can be transformed into:

[0190]

[0191] For a tree of a specific structure, the optimal weight w of each leaf node is j * and the extreme value Γ of the loss function k It can be solved as:

[0192]

[0193] in, It can be regarded as a scoring function to measure the quality of the tree structure. Finally, the loss function is divided into:

[0194]

[0195] I L , I R They are sample sets for the left and right branches respectively. By introducing the one-sided gradient descent algorithm, the model training speed is further improved. At the same time, the diversity of weak learners is increased, and the generalization ability of the model is further improved.

[0196] S5. Use TPE algorithm to optimize the prediction model parameter settings and improve the model prediction accuracy.

[0197] Example 3

[0198] The wind turbine gearbox oil pool temperature prediction method of the present invention has the following specific implementation steps:

[0199] S1. Obtain basic information data of wind turbines;

[0200] S2. Clean the original operation data of the wind turbine set, remove the abnormal operation data of the wind turbine set according to the wind speed-power curve distribution, improve the prediction accuracy of the model, and obtain the operation data of the wind turbine set under normal conditions;

[0201] S3, using the distance correlation coefficient, taking the gearbox oil pool temperature as the target parameter, and selecting parameters with strong correlation with it as input parameters of the oil temperature prediction model;

[0202] S4, using the parameters screened in S3 as input parameters of the prediction model, taking the oil pool temperature as the output parameter, and using the lightGBM algorithm to establish a gearbox oil pool temperature prediction model;

[0203] S5. Use TPE algorithm to optimize the prediction model parameter settings and improve the model prediction accuracy.

[0204] The TPE (Tree-structured Parzen Estimator) search algorithm is a hyperparameter optimization algorithm based on a Gaussian mixture model. The algorithm can use the performance of parameters that have been searched during training to infer the change trend of the next parameter, thereby reducing the search space and greatly improving the search efficiency. The main principle of the algorithm is to model the objective function and use Bayes' theorem to continuously update the prior probability distribution so as to select the search point that is more likely to produce the best result in the next step. The main process is as follows:

[0205] The objective function f(x) is less than a certain threshold y * The expectation is used as the evaluation index of the algorithm, denoted as EI (Expected Improvement):

[0206]

[0207] Define the following two probability densities:

[0208]

[0209] Where g(x) is the objective function f(x i ) value is greater than or equal to the threshold y * The corresponding {x i}Probability density of the set; l(x) is the opposite, indicating that it is less than the threshold y * The corresponding {x i}Probability density of a set;

[0210] Since the probability of p(y|x) cannot be obtained directly, the Bayesian formula is used to convert the above formula into:

[0211]

[0212] Now the probability of being less than the threshold value y* is recorded as γ = p(y <y * ),but:

[0213] p(x)=∫p(x|y)p(y)dy=γl(x)+(1-γ)g(x) (19)

[0214] It is deduced that:

[0215]

[0216] It can be seen from formula (20) that the size of the evaluation index depends on the probability density function g(x) and l(x). The smaller the ratio of g(x) to l(x), the larger the evaluation index EI.

[0217] In the TPE algorithm, the probability density functions g(x) and l(x) are represented using a tree structure. This representation method is more convenient for sampling according to l(x). Through this method, the x that maximizes the evaluation index EI can be obtained in each iteration. * .

Claims

1. A method for predicting the temperature of a wind turbine gearbox oil pool, characterized in that: The specific steps are as follows: S1. Obtain basic information data of wind turbines; S2. Cleaning the basic information data of the wind turbine generator set to remove abnormal operating data of the wind turbine generator set; S3. Calculate the correlation between the gearbox oil pool temperature and the candidate parameters using a distance correlation coefficient algorithm, and select parameters whose correlation with the gearbox oil pool temperature is greater than a threshold as input parameters; S4, using the parameters screened in S3 as input parameters of the prediction model, using the gearbox oil pool temperature as the output parameter, and using the lightGBM algorithm to establish a gearbox oil pool temperature prediction model; S5. Use the TPE algorithm to optimize the parameter settings of the gearbox oil pool temperature prediction model to improve the prediction accuracy of the gearbox oil pool temperature prediction model.

2. The method for predicting the temperature of the oil pool of the wind turbine gearbox according to claim 1, characterized in that: The basic information data of the wind turbine set in S1 includes the cut-in wind speed, rated wind speed, cut-out wind speed, power curve and set SCADA data.

3. The method for predicting the temperature of the oil pool of the wind turbine gearbox according to claim 1, characterized in that: S2 is to eliminate the abnormal operating data of the wind turbine according to the wind speed-power curve distribution, and obtain the operating data of the wind turbine under normal operating conditions. The specific operation steps are as follows: S2.

1. Perform the first cleanup of wind speed-power based on the quartile method, as follows: First, remove the data points with wind speed less than the cut-in wind speed and greater than the cut-off wind speed. According to the distribution of wind speed-power data, divide the wind speed axis into two wind speed segments: [cut-in wind speed, rated wind speed] and [rated wind speed, cut-out wind speed]. Divide the data of each wind speed segment into four parts according to the power value. The values ​​at the three split points are the lower quantile Q1, median Q2 and upper quantile Q3, respectively. For a sample X={x1,x2,…,x n-1 ,x n }, x in the sequence i <x i+1 , n is the total number of samples, i represents a point in the sequence, First calculate the second median: Then calculate the lower quantile Q1 and the upper quantile Q3, Q1 and Q3 represent the values ​​represented by the positions of 25% of the data points before and after the separation in sample X. When n=2k(k=0,1,2…), we have When n=4k+1(k=0,1,2…), we have Through the calculation of formula (2) and formula (3), the interquartile range I QR ,Right now: I QR =Q3-Q1(4) In statistics, I QR Determine the inner limit range of sequence X, that is: [F1,F u ]=[Q1-ω1I QR ,Q3+ω2I QR ] (5) Where: F1 is the lower limit of X determined by the quartile method; F u is the upper limit of X; w1 and w2 are weights, Take ω1=ω2=1.5, if Determined as abnormal data, removed; otherwise x i Retain, complete the first cleaning of wind speed-power based on the quartile method; S2.2, use the longitudinal filtering method based on data discreteness for the second cleaning, For the power value in each wind speed sub-interval, calculate its mean and variance, that is: In formula (6): i is the power mean square error of the i-th interval; N i is the scatter value of the ith interval; x j is the jth power value of the ith interval; μ i is the power mean of the ith interval, When the unit is operating normally, the wind speed and power scatter points are normally distributed around the power band, that is, 95% of the data are within the confidence interval covered by the two mean square errors above and below the mean. Based on this, each wind speed interval is based on the power mean μ i With the power band as the center, discrete data around the power band can be effectively eliminated.

4. The method for predicting the temperature of the oil pool of a wind turbine gearbox according to claim 1, characterized in that: The specific method of S3 is: Assume that two variables are X and Y, and the correlation between the two variables is recorded as dCor(X,Y). When dCor(X,Y)=0, it means that the two variables X and Y are independent. The larger the dCor(X,Y), the stronger the correlation between the variables X and Y. The algorithm for the distance correlation coefficient is as shown in formula (7): Where: dCor(X,Y) is the distance correlation coefficient between X and Y; dCov(X,Y) is the distance covariance between X and Y; dVar(X) is the variance of X; dVar(Y) is the variance of Y; The threshold is taken as 0.5, and the parameters with dCor(X,Y)>0.5 are taken as the input parameters of the prediction model.

5. The method for predicting the temperature of the oil pool of the wind turbine gearbox according to claim 1, characterized in that: The specific operation method of S4 is as follows: LightGBM makes the GBDT algorithm more efficient through two improved technologies: unilateral gradient descent and mutually exclusive feature bundling. The unilateral gradient descent algorithm is as follows: Assume that the training set data S = {(x1,y1),(x2,y2),…,(x n ,y n )}, where x is the input feature vector and y is the output data. The goal of the LightGBM model optimization process is to find a function Minimize the loss function L(y,f(x)). If the model prediction value is recorded as f(x), then the function It can be expressed as: Assuming that the number of decision trees in the model is M, since the model uses the forward distribution algorithm, the output value of the model at the Mth tree can be expressed as: At the mth tree, the loss function of the model is recorded as: In the LightGBM model, the function is approximated by Newton's method. After removing the constant term in the above formula, the loss function can be equivalent to: Among them, g i ,h i Represents the first-order and second-order gradients of the loss function. Since the model uses decision trees as the basis function, each tree can be expressed as: q(x) , q∈{1,2,…,J}, J represents the number of leaves in the decision tree, q represents the decision rule, and w represents the weight vector composed of leaf nodes. Therefore, formula (11) can be transformed into: For a tree of a specific structure, the optimal weight w of each leaf node is j * and the extreme value Γ of the loss function k It can be solved as: in, It can be regarded as a scoring function to measure the quality of the tree structure. Finally, the loss function is divided into: I L , I R They are sample sets for the left and right branches respectively. By introducing the one-sided gradient descent algorithm, the model training speed is further improved. At the same time, the diversity of weak learners is increased, and the generalization ability of the model is further improved.

6. The method for predicting the temperature of the oil pool of a wind turbine gearbox according to claim 1, characterized in that: The main principle of the TPE algorithm in S5 is to model the objective function and use Bayes' theorem to continuously update the prior probability distribution so as to select the search point that is more likely to produce the best result in the next step. The main process is as follows: The objective function f(x) is less than a certain threshold y * The expectation is used as the evaluation index of the algorithm, denoted as EI (Expected Improvement): Define the following two probability densities: Where g(x) is the objective function f(x i ) value is greater than or equal to the threshold y * The corresponding {x i }Probability density of the set; l(x) is the opposite, indicating that it is less than the threshold y * The corresponding {x i }Probability density of a set; Since the probability of p(y|x) cannot be obtained directly, the Bayesian formula is used to convert the above formula into: Now the probability of being less than the threshold value y* is recorded as γ = p(y <y * ),but: p(x)=∫p(x|y)p(y)dy=γl(x)+(1-γ)g(x) (19) It is deduced that: It can be seen from formula (20) that the size of the evaluation index depends on the probability density function g(x) and l(x). The smaller the ratio of g(x) to l(x), the larger the evaluation index EI. In the TPE algorithm, the probability density functions g(x) and l(x) are represented using a tree structure. This representation method is more convenient for sampling according to l(x). Through this method, the x that maximizes the evaluation index EI can be obtained in each iteration. * .

Images