Design method of steel tower drum fan foundation based on machine learning

Abstract

The invention relates to the technical field of fan foundation design, and particularly provides a steel tower drum fan foundation design method based on machine learning, which comprises the following steps: selecting an input variable and an output parameter, and establishing a data set containing the input variable and the output parameter; screening of the machine learning model: delimiting a training set, and performing hyper-parameter optimization of the machine learning model; based on the prediction effect of each group of models in the test set, obtaining the predicted average statistic estimated value and variance of each group of models on the input variable, and establishing a comparison matrix of each group of models; screening out a GBDT algorithm model as an optimal machine learning model based on the weight coefficient and the normalization coefficient obtained by the comparison matrix; and the optimized GBDT algorithm model is improved, and the complete engineering quantity of the high-precision fan foundation is obtained. According to the method, rapid and accurate prediction of the wind turbine foundation engineering quantity is realized, and the method can be applied to rapid estimation of a bidding stage and engineering quantity checking of a design stage, so that a design scheme is optimized, and engineering time and cost are saved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of steel tower wind turbine foundation design, and particularly relates to a design method of a steel tower wind turbine foundation based on machine learning. BACKGROUND

[0002] As a key clean energy, wind power plays an important role in addressing global warming. In engineering practice, the traditional calculation method is mainly dominated by human factors, which is prone to errors and may result in insufficient or overestimated engineering quantity. In addition, during the project bidding stage, fast and accurate engineering quantity estimation is crucial.

[0003] Machine learning prediction is an efficient prediction method, which has not been widely applied in the field of wind turbine foundation design. According to its prediction accuracy and efficiency, it can be well applied to the review of wind turbine foundation design, fast estimation of engineering quantity during the project bidding stage, etc. Therefore, reasonable use of machine learning method can realize intelligent prediction and optimization of wind turbine foundation engineering quantity. SUMMARY

[0004] The present application aims to provide a design method of a steel tower wind turbine foundation based on machine learning, and the specific technical solutions are as follows:

[0005] A design method of a steel tower wind turbine foundation based on machine learning, comprising the following steps:

[0006] Step S1, selecting input variables and output parameters; the input variables include the wind turbine capacity of the steel tower wind turbine foundation to be built, the foundation bearing capacity, the seismic intensity, the impeller diameter, the tower height and the base load; the base load includes the horizontal bending moment M xy , vertical bending moment M z , horizontal force F xy and vertical force F z of the base when the wind turbine impeller is running and under the action of wind load, etc.; the output parameters include the foundation buried depth, the foundation bottom diameter, the concrete quantity and the steel bar quantity of the steel tower wind turbine foundation;

[0007] Step S2, establishing a data set containing input variables and output parameters; dividing the data set into a training set and a test set;

[0008] Step S3, screening of the machine learning model, specifically comprising: based on the training set demarcated in the data set, and through the combination of random search and Bayesian optimization to perform hyperparameter optimization of each group of machine learning models; based on the prediction effect of each group of machine learning models on the test set, obtaining the average statistical quantity estimate and the variance of the statistical quantity evaluation value of each group of machine learning models on the four input variables, and further establishing a comparison matrix of each group of machine learning models; based on the weight coefficient and the normalization coefficient obtained by the comparison matrix, screening to obtain the GBDT algorithm model as the preferred machine learning model;

[0009] Step S4, improving the GBDT algorithm model to obtain an improved GBDT algorithm model, specifically comprising: performing importance evaluation on the input variables through the GBDT algorithm model; performing monotonicity constraint processing on the input variables and the output parameters by the GBDT algorithm model; dynamically adjusting the tree depth of the input variables; introducing quantile loss function into the GBDT algorithm model; setting early stopping mechanism in the GBDT algorithm model;

[0010] Step S5, obtaining the output parameter based on the improved GBDT algorithm model; calculating the complete engineering quantity of the steel tower drum fan foundation based on the obtained output parameter.

[0011] Preferably, the screening of the machine learning model comprises the following steps:

[0012] Step 1, dividing the data set into a training set and a test set according to a ratio of 7:3;

[0013] Step 2, selecting 12 groups of machine learning models including Linear algorithm model, Lasso algorithm model, Ridge algorithm model, decision tree algorithm model, random forest algorithm model, AdaBoost algorithm model, DNN algorithm model, ANN algorithm model, XGBoost algorithm model, SVM algorithm model, GBDT algorithm model and CatBoost algorithm model; based on the hyperparameter setting of the training set in S1, performing hyperparameter optimization setting through the combination of random search and Bayesian optimization;

[0014] Step 3, obtaining the statistical quantity of each group of machine learning models, which includes correlation coefficient R, determination coefficient R 2 , mean absolute error MAE and root mean square error RMSE;

[0015] Step 4, based on the four statistical quantities obtained in step 3, obtaining the corresponding average statistical quantity estimate Ave and the variance of the statistical quantity evaluation value Std; based on the average statistical quantity estimate and the variance of the statistical quantity evaluation value, obtaining a comparison matrix;

[0016] Step 5, the comparison matrix is evaluated by weight coefficient and normalization value, and the GBDT algorithm model is screened as the preferred machine learning model.

[0017] Preferably, the correlation coefficient R, the determination coefficient R 2 , the mean absolute error MAE and the root mean square error RMSE are calculated according to the following formulas:

[0018]

[0019]

[0020] Wherein: X i is the actual observation value of the i-th sample, P i is the model prediction value of the i-th sample, is the average value of the observation value of the i-th sample, and n is the sample quantity;

[0021] The values of the correlation coefficient R and the determination coefficient R 2 are closer to 1, indicating that the prediction accuracy is higher; the values of the mean absolute error MAE and the root mean square error RMSE are closer to 0, indicating that the prediction accuracy is higher.

[0022] Preferably, the average statistical quantity estimate Ave is the average value of each statistical quantity value of the four output parameters obtained when each group of machine learning models performs fan foundation engineering quantity prediction; and the variance Std of the statistical quantity evaluation value is the variance of the four output parameters with respect to the average statistical quantity estimate Ave.

[0023] The comparison matrix is established based on the average statistical quantity estimate Ave and the variance Std of the statistical quantity evaluation value, and specifically: the average statistical quantity estimate Ave and the variance Std of the statistical quantity evaluation value obtained by each group of machine learning models are horizontally arranged in order and divided into 12 columns; a 12x13 matrix is obtained by arranging the average statistical quantity estimate Ave or the variance Std of the statistical quantity evaluation value obtained by each group of machine learning models under each column; finally, each value below the first row is divided by the first value of each column, and a 12x12 matrix with a diagonal line of 1 below the first row is obtained, that is, the comparison matrix is obtained.

[0024] Preferably, the weight coefficient is the product of the elements in each row of the comparison matrix, and is raised to the power of 12, as shown in the following formula:

[0025]

[0026] Wherein: S k is the weight coefficient of the k-th row in the comparison matrix; b kj represents the element in the j-th column of the k-th row in the comparison matrix;

[0027] The normalization coefficient is the sum of the weight coefficients of each row divided by the weight coefficient of each row, as shown in the following formula:

[0028]

[0029] wherein: Q k is the normalized coefficient of the kth row.

[0030] Preferably, the evaluation in S5 includes stability prediction evaluation and accuracy prediction evaluation;

[0031] The stability prediction evaluation is specifically: using the normalized coefficient Q Std for stability prediction evaluation to perform evaluation, as follows:

[0032] wherein: Q R-Std is the normalized value of the correlation coefficient R for stability prediction evaluation; is the normalized value of the determination coefficient R 2 for stability prediction evaluation; Q MAE-Std is the normalized value of the mean absolute error MAE for stability prediction evaluation; Q RMSE-Std is the normalized value of the root mean square error RMSE for stability prediction evaluation;

[0033] The accuracy prediction evaluation is specifically: using the normalized coefficient Q Ave for accuracy prediction evaluation to perform evaluation, as follows:

[0034]

[0035] wherein: Q R-Ave is the normalized value of the correlation coefficient R for accuracy prediction evaluation; is the normalized value of the determination coefficient R 2 for accuracy prediction evaluation; Q MAE-Ave is the normalized value of the mean absolute error MAE for accuracy prediction evaluation; Q RMSE-Ave is the normalized value of the root mean square error RMSE for accuracy prediction evaluation;

[0036] The greater the value of the normalized coefficient Q Std for stability prediction evaluation, the higher the stability of the machine learning model prediction; the greater the value of the normalized coefficient Q Ave for accuracy prediction evaluation, the higher the accuracy of the machine learning model prediction;

[0037] When performing evaluation, accuracy is considered first, and then stability.

[0038] Preferably, the improvement of the GBDT algorithm model further includes performing importance evaluation on the input variables by the GBDT algorithm model, which is specifically: performing elimination processing on parameters with little influence and increasing derivative characteristic parameters of parameters with significant influence;

[0039] The dynamic adjustment of the tree depth of the input variable is specifically: allowing the parameters with high importance degree to continue to split and using the basic depth for the parameters with low importance degree.

[0040] Preferably, the quantile loss function L τ (X i , P i ) is calculated as follows:

[0041]

[0042] Wherein: tau is the target quantile;

[0043] The maximum depth m max is obtained as follows:

[0044]

[0045] Wherein: m b is the reference depth of the tree depth, psi is the importance coefficient, psi0 is the reference importance coefficient, and alpha is the adjustment coefficient;

[0046] The monotonicity constraint processing of the GBDT algorithm model on the input variable and the output parameter is specifically: for the foundation bearing capacity, it is negatively correlated with the output variable; for the fan capacity, the seismic intensity, the impeller diameter, the tower drum height and the base load, they are positively correlated with the output variable.

[0047] Preferably, the complete engineering quantity of the steel tower drum fan foundation includes the foundation buried depth h, the foundation bottom diameter l, the concrete consumption, the steel bar consumption, the excavation amount W and the filling amount of the steel tower drum fan foundation.

[0048] The excavation amount W is calculated as follows:

[0049]

[0050] Wherein: h is the foundation buried depth of the steel tower drum fan foundation; l1 is the width of the foundation construction platform; l2 is the thickness of the foundation cushion; and beta is the foundation excavation broken angle.

[0051] The filling amount is the difference between the excavation amount W and the concrete consumption.

[0052] Preferably, the early stopping mechanism is specifically: setting the maximum iteration round; monitoring the performance index of the verification set in real time, and when the value exceeds 50 iterations without stopping, stopping the iteration to determine the optimal parameters.

[0053] The technical scheme of the application utilizes a machine learning model to realize rapid and accurate prediction of the engineering quantity of a fan foundation. In addition, a high-precision data set is first constructed, then a comparison matrix is constructed based on the mean and variance of the prediction results of four output quantities and based on multiple machine learning models, and a steel tower drum fan foundation prediction model with high prediction accuracy and best stability is optimized by combining a weight coefficient and normalization processing. The steel tower drum fan foundation prediction model output value is preferably used to calculate the complete engineering quantity of the steel tower drum fan foundation, which can be applied to rapid estimation in the bidding stage and engineering quantity checking in the design stage, effectively avoiding the problem of overestimation or underestimation of the engineering quantity, thereby optimizing the design scheme and significantly saving engineering time and cost.

[0054] In addition to the objects, features, and advantages described above, the present application has other objects, features, and advantages. The present application will be described in further detail below with reference to the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0055] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application, illustrate the preferred embodiments of the application, and assist in the explanation of the application. In the drawings, the same reference numbers represent the same elements throughout the several views of the drawings.

[0056] Figure 1 is a flowchart of the design method of the steel tower drum fan foundation based on machine learning in the embodiment;

[0057] Figure 2 is a structural schematic diagram of the steel tower drum fan foundation in the embodiment;

[0058] In the structural schematic diagram of the steel tower drum fan foundation, 1 represents the foundation bottom diameter, 2 represents the foundation burial depth, 3 represents the foundation excavation broken angle, 4 represents the foundation construction platform width, 5 represents the foundation cushion layer width, 6 represents the flange anchor bolt fixing area below the fan foundation, and 7 represents the concrete fan foundation contour. DETAILED DESCRIPTION

[0059] The embodiments of the present application will be described in detail below with reference to the drawings, but the present application can be implemented in various different ways as limited and covered by the claims.

[0060] Embodiment 1:

[0061] A design method of a steel tower drum fan foundation based on machine learning, and the final structural schematic diagram of the steel tower drum fan foundation is as shown in Figure 1 In the structural schematic diagram of the steel tower drum fan foundation, 1 represents the foundation bottom diameter, 2 represents the foundation burial depth, 3 represents the foundation excavation broken angle, 4 represents the foundation construction platform width, 5 represents the foundation cushion layer width, 6 represents the flange anchor bolt fixing area below the fan foundation, and 7 represents the concrete fan foundation contour.

[0062] The specific process of the design method is as shown in Figure 1 and specifically includes the following steps:

[0063] S1, select input variables and output variables, the input variables include fan capacity of the steel tower drum fan foundation to be built, foundation bearing capacity, seismic intensity, impeller diameter, tower drum height and base load; in the preferred embodiment, the base load includes horizontal bending moment M xy , vertical bending moment M z , horizontal force F xy and vertical force F z of the fan impeller when running and under the action of wind load; the output parameters include foundation burial depth, foundation bottom diameter, concrete consumption and steel consumption of the steel tower drum fan foundation.

[0064] S2, establish a data set containing all input variables and output parameters, and divide it into a training set and a test set.

[0065] S3, screening of machine learning models, specifically including: based on the training set divided in the data set, and through random search and Bayesian optimization to optimize the hyperparameters of each group of machine learning models; based on the prediction effect of each group of machine learning models on the test set, the average statistical value and the variance of the statistical value of each group of machine learning models are obtained on the four input variables, and a comparison matrix of each group of machine learning models is further established; based on the weight coefficient and the normalization coefficient obtained by the comparison matrix, the GBDT algorithm model is selected as the preferred machine learning model. In the embodiment: 12 groups of machine learning models including Linear algorithm model, Lasso algorithm model, Ridge algorithm model, decision tree algorithm model, random forest algorithm model, AdaBoost algorithm model, DNN algorithm model, ANN algorithm model, XGBoost algorithm model, SVM algorithm model, GBDT algorithm model and CatBoost algorithm model are selected; the evaluation of machine learning models is carried out through mathematical statistical correlation coefficient R, determination coefficient R 2 , mean absolute error MAE and root mean square error RMSE, and the prediction accuracy and stability of machine learning models are evaluated through the average statistical value Ave and the variance Std of the statistical value of the four output quantities. The comparison matrix is constructed based on the average statistical value and the variance of the statistical value; the weight coefficient and the normalization value are used to quantitatively evaluate the optimal machine learning model; finally, the Pearson coefficient and the importance coefficient are used for evaluation to determine the correlation between the input variables and to clarify the size of the influence of each input variable on the final output variable.

[0066] S4, the GBDT algorithm model is improved to obtain an improved GBDT algorithm model, specifically including: the importance of the input variables is evaluated by the GBDT algorithm model; the GBDT algorithm model performs monotonicity constraint processing on the input variables and the output parameters; the tree depth of the input variables is dynamically adjusted; the quantile loss function is introduced into the GBDT algorithm model; and the early stopping mechanism is set in the GBDT algorithm model.

[0067] S5, the output parameters are obtained based on the improved GBDT algorithm model; and the complete engineering quantity of the steel tower cylinder fan foundation is calculated based on the obtained output parameters, which is used for engineering bidding, design checking and other engineering practices.

[0068] Preferably, the foundation burial depth and the foundation diameter of the output variables are mainly used to judge whether the design meets the engineering land requirements and to calculate the excavation amount and the fill amount of the fan foundation.

[0069] Preferably, the data set for machine learning model training includes all input variables and all output parameters proposed in the embodiment; the design cases are not less than 200 groups to ensure the accuracy of the machine learning prediction results; and the design parameters of each group of design cases should be evenly distributed to avoid large faults.

[0070] Preferably, the machine learning model should select each type of machine learning model to select the model with the highest prediction accuracy of the fan foundation engineering quantity.

[0071] Preferably, the hyperparameter setting is the parameter setting of each group of machine learning models; the training set is divided into a training set and a test set according to 7:3 based on the established data set, the former is the training set, and the latter is the test set; and the hyperparameter setting is adjusted in the training set by the random search and Bayesian optimization methods to make the fitting degree of the machine learning model highest.

[0072] In this embodiment, the statistical quantity includes the correlation coefficient R, the determination coefficient R 2 , the mean absolute error MAE and the root mean square error RMSE. Preferably, the calculation formulas of the correlation coefficient R, the determination coefficient R 2 , the mean absolute error MAE and the root mean square error RMSE are as follows:

[0073]

[0074]

[0075] Wherein: X i is the actual observation value of the i-th sample, P i is the model prediction value of the i-th sample, ) is the average value of the observation values of the ith sample, and n is the sample number;

[0076] The correlation coefficient R and the determination coefficient R 2 The closer the value of the correlation coefficient R and the determination coefficient R

[0077] In this embodiment, the average statistical estimate Ave is the average value of each statistical value of the four output parameters obtained when each group of machine learning models performs fan foundation engineering quantity prediction; and the variance Std of the statistical estimate is the variance of the four output parameters with respect to the average statistical estimate Ave.

[0078] In this embodiment, the comparison matrix is established based on the average statistical estimate Ave and the variance Std of the statistical estimate, and specifically, the average statistical estimate Ave and the variance Std of the statistical estimate obtained by each group of machine learning models are horizontally arranged in order and divided into 12 columns; a 12x13 matrix is obtained by arranging the average statistical estimate Ave or the variance Std of the statistical estimate obtained by each group of machine learning models under each column; finally, each value below the first row is divided by the first value of each column, and a 12x12 matrix with a diagonal line of 1 below the first row is obtained, that is, the comparison matrix is obtained. Specifically, the Ave values or Std values obtained by the 12 groups of machine learning models are a1-a12. 12 Therefore, in the comparison matrix, the values of each column are as follows from top to bottom and from the first column to the twelfth column:

[0079]

[0080] wherein b k1 ,..., b k12 (k=1-12) are the values of each column in the comparison matrix.

[0081] The comparison matrix D is as follows:

[0082]

[0083] According to the evaluation process of this embodiment, eight comparison matrices need to be constructed.

[0084] The comparison matrix is evaluated by a weight coefficient and a normalized value, and the weight coefficient is the product of the elements of each row in the comparison matrix, raised to the power of 12, as shown in the following formula:

[0085]

[0086] wherein S k is the weight coefficient of the kth row in the comparison matrix; and b kjThis indicates comparing the element in the k-th row and j-th column of the matrix;

[0087] The normalization coefficient is the sum of the weight coefficients of each row and the sum of the weight coefficients of each row, as shown in the following formula:

[0088]

[0089] Among them: Q k is the normalization coefficient for the k-th row.

[0090] The evaluation described in this embodiment includes stability prediction evaluation and accuracy prediction evaluation;

[0091] The stability prediction assessment specifically involves using the normalized coefficient Q for stability prediction assessment. Std To conduct an evaluation, Q Std The normalized coefficients for the four statistics obtained based on the Std values ​​of the four output parameters are as follows:

[0092] Among them: Q R-Std This is the normalized value of the correlation coefficient R used for stability prediction and assessment; R is the coefficient of determination used for stability prediction assessment. 2 The normalized value of Q; MAE-Std Q is the normalized value of the mean absolute error (MAE) used for stability prediction assessment; RMSE-Std This is the normalized value of the root mean square error (RMSE) used for stability prediction assessment.

[0093] The accuracy prediction assessment specifically involves using the normalized coefficient Q for accuracy prediction assessment. Ave To conduct an evaluation, Q Ave The normalized coefficients for the four statistics obtained based on the Ave values ​​of the four output parameters are as follows:

[0094]

[0095] Among them: Q R-Ave This is the normalized value of the correlation coefficient R used for accuracy prediction and evaluation; R is the coefficient of determination used for accuracy prediction evaluation. 2 The normalized value of Q; MAE-Ave Q is the normalized value of the mean absolute error (MAE) used for accuracy prediction evaluation; RMSE-Ave This is the normalized value of the root mean square error (RMSE) used for accuracy prediction evaluation.

[0096] The normalized coefficient Q used for stability prediction assessment Std The larger the value of Q, the higher the predictive stability of the machine learning model; the normalization coefficient Q used for accuracy prediction evaluation AveThe larger the value, the higher the prediction accuracy of the machine learning model;

[0097] When conducting an evaluation, accuracy should be considered first, followed by stability.

[0098] In this embodiment, the importance coefficient of the input variable is determined based on the selected optimal machine learning model and is established according to the tree model; the larger the importance coefficient of the input variable, the greater its impact on the design of the wind turbine foundation.

[0099] In this embodiment, the complete engineering quantity of the steel tower wind turbine foundation should also include the excavation and filling volume of the wind turbine foundation. The excavation volume W is calculated using the following formula:

[0100]

[0101] Where: h is the foundation depth of the steel tower wind turbine foundation; l1 is the width of the foundation construction platform ( Figure 2 The standard number is 4), which is generally taken as 0.5m; l2 is the thickness of the foundation layer ( Figure 2 The standard number is 5), which is generally used for leveling the bottom of the foundation, and the thickness is generally taken as 0.1m; β is the foundation excavation angle ( Figure 2 The reference number is 3), which is generally 63° or 38°; 100 in the formula refers to the anchor bolt fixing area 6 under the wind turbine foundation flange, and the volume of this area does not exceed 100m³. 3 .

[0102] The fill volume is the difference between the excavation volume W and the concrete usage.

[0103] In this embodiment, the complete engineering quantity of the steel tower wind turbine foundation includes the foundation embedment depth h of the steel tower wind turbine foundation. Figure 1 (2) Foundation bottom diameter l ( Figure 2 The number of the sub-subjects is 1), the amount of concrete used, the amount of steel reinforcement used, the amount of excavation W, and the amount of filling.

[0104] Example 2:

[0105] Taking a specific engineering example from China Energy Engineering Group, machine learning prediction of wind turbine foundations is conducted. The specific implementation process is as follows: Figure 1 Consistent.

[0106] In this embodiment, the selected input variables are wind turbine capacity, foundation bearing capacity, seismic intensity, impeller diameter, tower height, and base load; the base load includes the horizontal bending moment M on the base when the wind turbine impeller is running and when wind loads are applied. xy Vertical bending moment M z Horizontal force F xy and vertical force F z .

[0107] The data set in this embodiment is established according to specific cases of the feasibility study, preliminary design and construction drawing stage of some projects of China Energy Construction Group. More than 200 groups of case data sets are established, and Table 1 shows some data sets:

[0108] Table 1 shows the data set of input variables and output parameters (partially shown)

[0109]

[0110] In this embodiment, the machine learning model includes 12 groups of Linear algorithm model, Lasso algorithm model, Ridge algorithm model, decision tree algorithm model, random forest algorithm model, AdaBoost algorithm model, DNN algorithm model, ANN algorithm model, XGBoost algorithm model, SVM algorithm model, GBDT algorithm model and CatBoost algorithm model. Random search and Bayesian optimization are combined to optimize the hyperparameters. In this embodiment, the accuracy of the first three linear algorithms (Linear algorithm model, Lasso algorithm model and Ridge algorithm model) is too low, and further discussion is not made; GBDT algorithm model and CatBoost algorithm model are more suitable for scenarios with relatively small data sets and relatively large data dispersion. The setting of the hyperparameters is shown in the English real representation of the code, and details are shown in Table 2:

[0111] Table 2 shows the setting statistics of the hyperparameters of the nine machine learning models

[0112]

[0113]

[0114] The average statistical value Ave and the variance Std of the statistical value are shown in Table 3:

[0115] Table 3 shows the statistical value table of the nine machine learning models

[0116]

[0117] When comparing the matrix, first, the Ave values or Std values obtained by the nine machine learning models (excluding three low-precision linear models) selected in this embodiment are arranged horizontally in order and divided into nine columns. Next, the Ave values or Std values obtained by the nine machine learning models are arranged in the columns below each column to obtain a 9x10 matrix. Finally, each value below the first row is divided by the first value of each column, and the 9x9 matrix with a diagonal line of 1 below the first row is the obtained comparison matrix; the comparison matrix is the Ave and Std comparison matrix of the determination coefficient R 2 ​

[0118]

[0119] The specific weight coefficients and normalization coefficients in this embodiment are shown in Table 4:

[0120] Table 4 Weight coefficient and normalization coefficient statistical table

[0121]

[0122]

[0123] According to the above analysis results, considering the prediction accuracy and stability, the GBDT algorithm model is selected in this embodiment for the prediction of the steel tower drum fan foundation design.

[0124] In this embodiment, the importance of each input variable is evaluated by the selected GBDT algorithm model to determine the importance ranking of each input variable on the output, and the corresponding importance coefficient of each input variable is matched. The greater the importance coefficient, the greater the influence on the design of the fan foundation, and the sum of the importance coefficients is 1. For the parameters with significant influence (i.e. the importance coefficient is greater than 0.15), the derived characteristic parameters are increased, and for the parameters with less influence (i.e. the importance coefficient is less than 0.05), they are removed to improve the prediction accuracy. Further optimization: for the prediction of foundation burial depth, the influence degree is the smallest (i.e. the importance coefficient is less than 0.05), respectively, the vertical bending moment M z and the horizontal force F xy , both less than 0.05, and the corresponding impeller diameter, seismic intensity, tower drum height, horizontal bending moment M xy and vertical force F z have the most significant influence. Therefore, the vertical bending moment M z and the horizontal force F xy with less influence are removed, and the derived characteristic parameters of the parameters with significant influence (i.e. the importance coefficient is greater than 0.15) are increased, and the ratio of impeller diameter / tower drum height is increased. The prediction of the foundation bottom diameter and the amount of concrete is similar, so it is not further described. For the amount of steel, the influence degree of the foundation bearing capacity, the vertical bending moment M z and the horizontal force F xy is the smallest, so they are removed in the prediction learning. The horizontal bending moment M xy (the importance coefficient is 0.577), the vertical force F z and the seismic intensity have greater influence, and no further derived parameters in the input variables are added.

[0125] Because there may be some extreme cases (such as excessive seismic fortification intensity), and the loss function in the traditional algorithm of GBDT only optimizes the mean, it will further amplify the error. Therefore, for specific fan foundation scenarios, a quantile loss function is introduced into the GBDT algorithm model, and its calculation expression is as follows:

[0126]

[0127] Wherein: τ is the target quantile, which is taken as 0.95 in this embodiment to ensure the safety of the project; X i is the actual observed value of the ith sample, P i is the model prediction value of the ith sample;

[0128] The GBDT algorithm model performs monotonicity constraint processing on the input variables and output parameters, that is, the model is forced to follow the physical law. For the bearing capacity of the foundation, it is negatively correlated with the output variable, and the constraint is supplemented as "-1" in the algorithm. For other variables, they are positively correlated with the output variable, and the constraint is supplemented as "+1" in the algorithm.

[0129] To improve the prediction accuracy of the algorithm model, the tree depth of the input variable is dynamically adjusted. Specifically, for variables with high importance, the tree depth is allowed to continue to split, while for variables with low importance, the tree depth adopts the basic depth. The maximum depth of the tree depth m max is obtained by using the following formula:

[0130]

[0131] Wherein: m b is the reference depth of the tree depth; ψ is the importance coefficient; ψ0 is the reference importance coefficient, which is the second-to-last importance coefficient of each input variable, that is, the input variable with too small influence degree is not considered in the algorithm, that is, the tree depth is not considered to be increased; α is the adjustment coefficient, which can be determined by grid search. In this case, for the prediction of the steel reinforcement consumption, α is taken as 8, and for the other three quantities, α is taken as 10. When the maximum tree depth calculated is not an integer, it is rounded to the nearest integer.

[0132] An early stopping mechanism is set in the GBDT algorithm model. Specifically, first, the maximum iteration round is set, which is recommended to be 1000. The performance indicator of the validation set is monitored in real time, mainly monitoring the RMSE value. When the value exceeds 50 iterations without stopping, the iteration is stopped, that is, the optimal parameters are determined. After early stopping, the optimal tree depth distribution can be analyzed in reverse to guide the subsequent hyperparameter setting. This can save iteration training time and avoid overfitting.

[0133] In actual engineering, all nine engineering quantities may not be obtained in the early stage. Through importance evaluation, it is known that, in the prediction of the four engineering quantities, the four variables with higher importance degree are mainly impeller diameter, earthquake intensity, horizontal bending moment M xy and vertical force F z . To this end, the improved GBDT algorithm can be used, and the four variables are input to preliminarily measure the engineering quantity of the fan foundation. The correlation coefficient of the predicted value obtained by the new model is also about 0.9.

[0134] The above only describes the preferred embodiments of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for designing a steel tower foundation for a wind turbine based on machine learning, characterized in that, The method comprises the following steps: Step S1, selecting input variables and output parameters; the input variables include fan capacity of the steel tower drum fan foundation to be built, foundation bearing capacity, earthquake intensity, impeller diameter, tower drum height and base load; the base load includes horizontal bending moment M xy , vertical bending moment M z , horizontal force F xy and vertical force F z of the base when the fan impeller is running and under the action of wind load; the output parameters include foundation burying depth, foundation bottom diameter, concrete consumption and steel bar consumption of the steel tower drum fan foundation; Step S2, establishing a data set containing input variables and output parameters; Divide the data set into a training set and a test set; Step S3, screening of the machine learning model, specifically comprising: based on the training set divided in the data set, and through the combination of random search and Bayesian optimization to optimize the hyperparameters of each group of machine learning models; based on the prediction effect of each group of machine learning models on the test set, the average statistical value estimate and the variance of the statistical value estimate of each group of machine learning models are obtained on the four input variables, and a comparison matrix of each group of machine learning models is further established; based on the weight coefficient and the normalization coefficient obtained by the comparison matrix, the GBDT algorithm model is screened as the preferred machine learning model; Step S4, improving the GBDT algorithm model to obtain an improved GBDT algorithm model, specifically comprising: performing importance evaluation on the input variables through the GBDT algorithm model; the GBDT algorithm model performs monotonicity constraint processing on the input variables and the output parameters; dynamically adjusting the tree depth of the input variables; introducing a quantile loss function into the GBDT algorithm model; setting an early stopping mechanism in the GBDT algorithm model; Step S5, obtaining the output parameters based on the improved GBDT algorithm model; calculating the complete engineering quantity of the steel tower cylinder fan foundation based on the obtained output parameters.

2. The method of designing a machine learning based steel tower foundation for a wind turbine as claimed in claim 1, wherein, The screening of the machine learning model comprises the following steps: Step ①, dividing the data set into a training set and a test set according to a ratio of 7:3; Step ②, selecting 12 groups of machine learning models including Linear algorithm model, Lasso algorithm model, Ridge algorithm model, decision tree algorithm model, random forest algorithm model, AdaBoost algorithm model, DNN algorithm model, ANN algorithm model, XGBoost algorithm model, SVM algorithm model, GBDT algorithm model and CatBoost algorithm model; based on the hyperparameter setting of the training set in S1, the hyperparameters are optimized and set through the combination of random search and Bayesian optimization; Step 3, obtaining the statistics of each group of machine learning models, the statistics including correlation coefficient R, determination coefficient R 2 , mean absolute error MAE and root mean square error RMSE; Step ④, based on the four statistics obtained in step ③, the average statistical value estimate Ave and the variance of the statistical value estimate Std are obtained; based on the average statistical value estimate and the variance of the statistical value estimate, a comparison matrix is obtained; Step ⑤, the comparison matrix is evaluated by weight coefficient and normalization value, and the GBDT algorithm model is screened as the preferred machine learning model.

3. The method of designing a machine learning based steel tower foundation for a wind turbine as claimed in claim 2, wherein, The correlation coefficient R, the determination coefficient R 2 The calculation formulas of the mean absolute error MAE and the root mean square error RMSE are as follows: where: X i is the actual observation value of the i-th sample, P i is the model prediction value of the i-th sample, is the average value of the observation value of the i-th sample, and n is the sample number; The closer the values of the correlation coefficient R and the determination coefficient R 2 to 1, the higher the prediction accuracy. The closer the values of the mean absolute error MAE and the root mean square error RMSE to 0, the higher the prediction accuracy.

4. The method of designing a machine learning based steel tower foundation for a wind turbine as claimed in claim 3, wherein, The average statistical value estimate Ave is the mean value of each statistical value of the four output parameters obtained when each group of machine learning models predicts the fan foundation engineering quantity; the variance of the statistical value estimate Std is the variance of the average statistical value estimate Ave of the four output parameters. The comparison matrix is established based on the average statistical value estimate Ave and the variance Std of the statistical value estimate, specifically: the average statistical value estimate Ave and the variance Std of the statistical value estimate obtained by each group of machine learning models are arranged horizontally in order and divided into 12 columns; a 12x13 matrix is obtained by arranging the average statistical value estimate Ave or the variance Std of the statistical value estimate obtained by each group of machine learning models under each column; finally, each value below the first row is divided by the first value of each column, and a 12x12 matrix with a diagonal line of 1 below the first row is obtained, that is, the comparison matrix is obtained.

5. The method of designing a machine learning based steel tower foundation for a wind turbine as claimed in claim 4, wherein, The weight coefficient is the product of the elements of each row in the comparison matrix, and is raised to the power of 12, as shown in the following formula: where: S k is the weight coefficient of the kth row in the comparison matrix; b kj denotes the element of the kth row and jth column in the comparison matrix; The normalization coefficient is the weight coefficient of each row divided by the cumulative weight coefficient of each row, as shown in the following formula: where: Q k is the normalized coefficient for the kth row.

6. The method of designing a machine learning based steel tower foundation for a wind turbine as claimed in claim 5, wherein, The evaluation in S5 includes stability prediction evaluation and precision prediction evaluation. The stability prediction evaluation is specifically: using a normalization coefficient Q for stability prediction evaluation Std The evaluation is performed as follows: wherein: Q R-Std is a normalized value of the correlation coefficient R for stability prediction evaluation; is a normalized value of the determination coefficient R 2 for stability prediction evaluation;Q MAE-Std is a normalized value of the mean absolute error MAE for stability prediction evaluation;Q RMSE-Std is a normalized value of the root mean square error RMSE for stability prediction evaluation; The accuracy prediction evaluation is specifically: using a normalization coefficient Q for accuracy prediction evaluation Ave The evaluation is performed as follows: wherein: Q R-Ave is a normalized value of the correlation coefficient R for precision prediction evaluation; is a normalized value of the coefficient of determination R 2 for precision prediction evaluation; Q MAE-Ave is a normalized value of the mean absolute error MAE for precision prediction evaluation; Q RMSE-Ave is a normalized value of the root mean square error RMSE for precision prediction evaluation; The normalization coefficient Q for stability prediction evaluation Std The greater the value, the higher the machine learning model predicts stability; the normalization coefficient Q for precision prediction evaluation Ave The greater the value, the higher the machine learning model predicts precision; When evaluating, consider the precision first, and then consider the stability.

7. The method of designing a machine learning based steel tower foundation for a wind turbine as claimed in claim 6, wherein, The improvement of the GBDT algorithm model also includes importance evaluation of input variables by the GBDT algorithm model, specifically: the parameters with little influence are removed, and the derivative characteristic parameters of the parameters with significant influence are increased; The tree depth of the input variable is dynamically adjusted, specifically: the parameters with high importance degree are allowed to continue to split, and the tree depth of the parameters with low importance degree is set to the basic depth.

8. The method of designing a machine learning based steel tower foundation for a wind turbine as claimed in any one of claims 4 to 7, wherein, The quantile loss function L τ (X i , P i ) is calculated as follows: Where: τ is the target quantile; Maximum depth of tree depth m max is obtained using the formula: wherein: m b is the reference depth of tree depth, ψ is the importance coefficient, ψ0 is the reference importance coefficient, and a is the adjustment coefficient. The GBDT algorithm model performs monotonicity constraint processing on the input variables and output parameters, specifically: for the bearing capacity of the foundation, it is negatively correlated with the output variable; for the fan capacity, the seismic intensity, the impeller diameter, the tower drum height and the base load, they are positively correlated with the output variable.

9. The method of designing a machine learning based steel tower foundation for a wind turbine as claimed in claim 1, wherein, The complete engineering quantity of the steel tower drum fan foundation includes the foundation burial depth h, the foundation bottom diameter l, the concrete consumption, the steel consumption, the excavation amount W and the filling amount of the steel tower drum fan foundation. The excavation amount W is calculated by the following formula: Where: h is the foundation burial depth of the steel tower drum fan foundation; l1 is the width of the foundation construction platform; l2 is the thickness of the foundation cushion; β is the foundation excavation corner angle; The filling amount is the difference between the excavation amount W and the concrete consumption.

10. The method of designing a machine learning based steel tower foundation for a wind turbine as claimed in claim 1, wherein, The early stopping mechanism is specifically: setting the maximum iteration round; monitoring the performance indicators of the validation set in real time, and stopping the iteration when the value exceeds 50 iterations without stopping, and determining the optimal parameters.

Images