Mountain wind turbine fatigue load and power prediction method based on optimized support vector machine
By combining gray correlation analysis and improved artificial bee colony algorithm in the support vector machine model, the problem of insufficient data complexity and model generalization ability in fatigue load and power prediction of mountain wind turbines is solved, and higher prediction accuracy and Efficiency, suitable for the prediction of wind turbine performance parameters of complex mountain wind farms.
Patent Information
- Application Number
- CN202410929040.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-11
- Publication Date
- 2024-11-12
AI Technical Summary
The existing mountain wind turbine fatigue load and power prediction models have problems such as incomplete and inaccurate data, poor generalization ability of the model, and difficulty in processing high-latitude characteristics when dealing with complex mountain wind farms, which cannot accurately reflect the complex characteristics of mountain wind farms. .
The grey correlation analysis and improved artificial bee colony algorithm were introduced into the support vector machine model. High correlation characteristics were screened through gray correlation analysis for input dimensionality reduction, and the hyperparameters were optimized using the Nelder-Mead simplex method to build an optimized support vector machine The model makes predictions.
Improve prediction accuracy and iterative efficiency. The optimized model can more accurately reflect the complex characteristics of mountain wind farms, avoid the problem of overfitting support vector machines on high-dimensional data, and improve the generalization ability and adaptability of the model. .
Smart Images

Figure CN118940129A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of wind power generation, and specifically relates to a method for predicting fatigue load and power of a mountain wind turbine based on an optimized support vector machine. Background Art
[0002] With the growth of global energy demand and the need for environmental protection, wind power generation has become one of the important renewable energy sources. Mountainous areas are rich in wind resources, but their complex terrain and wind field characteristics bring challenges to the design and operation of wind turbines. Existing prediction models have problems such as incomplete and inaccurate data, poor model generalization ability, and difficulty in handling high-latitude features in dealing with fatigue loads and power prediction of mountain wind turbines, and cannot accurately reflect the complex characteristics of mountain wind fields.
[0003] On the one hand, with the continuous development of machine learning technology and the continuous expansion of application scenarios, its application in the wind power industry is becoming more and more extensive. Support vector machine (SVM), as a powerful supervised learning model, is often used to solve various prediction problems in the field of flat wind turbines due to its good performance on nonlinear problems. However, mountain wind farms often have diverse and high-latitude characteristics, such as wind speed changes at different heights, blocking effects of terrain, etc. These characteristics may cause support vector machines (SVM) to overfit or fail to fully capture key features in high-dimensional space. Therefore, before using SVM to predict fatigue loads and power of mountain wind turbines, it is necessary to moderately reduce the dimensionality of the input of the data set to filter out redundant factors and retain important factors.
[0004] In addition, in machine learning training, parameter adjustment is an important means to improve model accuracy. Currently, commonly used parameter optimization methods include grid search method and swarm intelligence algorithm. The grid search method uses a traversal method to train the corresponding SVM prediction model for each parameter combination, and finally selects the parameter combination that can achieve the highest accuracy. This method traverses all parameter combinations in the search space, so it can often find better parameter settings. However, due to a large number of useless calculations, when the parameter dimension is high, the computing resource consumption and time cost are high; in addition, the performance of the grid search method also depends on the set search step size. If the search space and step size are set improperly, satisfactory accuracy may not be achieved.
[0005] Swarm intelligence algorithms simulate the foraging behavior of biological groups in nature, and through the collaboration and information exchange of individuals in the group, they can find the optimal solution to the problem at a faster speed. Each individual searches for the optimal solution independently in the search space, and individuals exchange information through competition, cooperation, or a combination of the two, so as to jointly find the optimal solution. Swarm intelligence algorithms have the characteristics of strong parallel processing capabilities and high efficiency, and have significant advantages in optimizing problems. The artificial bee colony algorithm has a strong global search capability, but it selects new nectar sources in a random manner during the scout bee stage, which leads to a slow convergence speed, increases the cost of later calculations, and makes it difficult to guarantee the accuracy during the local search process.
[0006] In order to improve the prediction accuracy and prediction efficiency, the present invention adds the grey correlation algorithm and the improved artificial bee colony algorithm to the underlying model of the support vector machine, and constructs an optimized support vector machine model to predict the equivalent fatigue load and power of the wind turbine. The grey correlation analysis method is used to measure the correlation between two data sequences. The basic idea is to draw curves for the two series respectively, and calculate the correlation of the data sequences according to the similarity between the curve shapes. In actual situations, due to the limitations of measurement methods and equipment accuracy, the completeness and reliability of the measured data are often in doubt, and the data distribution is often different from that in the ideal situation. This feature causes the conventional correlation analysis method to have requirements on the size of the data set, the characteristic distribution of the data, the correlation between the characteristic vectors, etc. In contrast, the grey correlation analysis method has no strict restrictions on the sample size and data distribution of the data set. Based on limited data, the degree of correlation between the independent variable and the dependent variable can be analyzed, and the importance of different parameters can be compared. Therefore, it has received widespread attention in the field of data analysis.
[0007] Therefore, it is of engineering significance to construct an optimized support vector machine model for fatigue load and power prediction of mountain wind turbines. The optimized SVM model is superior to the basic SVM model in both prediction accuracy and iteration efficiency, and achieves higher accuracy in predicting the performance parameters of mountain wind turbines while achieving satisfactory iteration efficiency. Summary of the invention
[0008] In view of this, the object of the present invention is to provide a method for predicting fatigue load and power of mountain wind turbines based on an optimized support vector machine, which can adapt to the characteristics of complex mountain wind fields and improve prediction accuracy.
[0009] In order to achieve the above object, the present invention provides the following technical solutions:
[0010] A method for predicting fatigue load and power of mountain wind turbines based on optimized support vector machine comprises the following steps:
[0011] Step 1: Data collection and preprocessing
[0012] 11) Collect the wind speed at the measuring point in the wind turbine rotor plane, the turbulence at the hub position, the pitch angle, the rotor speed, the fatigue load and power data of the mountain wind turbine to obtain the performance parameter database of the wind turbine under the mountain terrain;
[0013] 12) The grey correlation analysis method is used to analyze the correlation between input data and output data, and high correlation features are selected to achieve input dimension reduction; the input data includes the wind speed at the measuring point in the wind turbine rotor plane, the turbulence at the hub position, the pitch angle and the rotor speed; the output data includes the fatigue load and power data of the mountain wind turbine;
[0014] 13) performing normalization processing on the screened data, and dividing the normalized wind turbine performance parameter database under mountainous terrain into a training set and a test set;
[0015] Step 2: Construct a mountain wind turbine fatigue load and power prediction model
[0016] A fatigue load and power prediction model for mountain wind turbines was constructed based on support vector machine;
[0017] Step 3: Training mountain wind turbine fatigue load and power prediction model
[0018] The training set is used to train the mountain wind turbine fatigue load and power prediction model, and the artificial bee colony algorithm is used to optimize the hyperparameters of the support vector machine to obtain the mountain wind turbine fatigue load and power prediction model with the optimal hyperparameters.
[0019] Step 4: Test the mountain wind turbine fatigue load and power prediction model
[0020] The test set is used to test the prediction performance of the trained mountain wind turbine fatigue load and power prediction model;
[0021] Step 5: Mountain wind turbine fatigue load and power prediction
[0022] The input data collected in real time, including the wind speed at the measuring point in the wind turbine rotor plane, the turbulence at the hub position, the pitch angle and the rotor speed, are input into the mountain wind turbine fatigue load and power prediction model, and the predicted mountain wind turbine fatigue load and power are output.
[0023] Furthermore, in step 12), the method steps of using the grey correlation analysis method to analyze the correlation between the input data and the output data are as follows:
[0024] 121) Determine the evaluation sequence X and the reference sequence Y:
[0025] The evaluation series X includes the characteristics that need to be analyzed:
[0026] X=[X1 X2...Xm ]
[0027] X i =[x i (1)x i (2)...x i (n)] T (i=1,2,...,m)
[0028] Among them: m is the number of evaluation features; n is the number of samples under each feature;
[0029] The reference series are the research objects, including the statistical indicators of equivalent fatigue load and power:
[0030] Y=[Y1 Y2...Y l ]
[0031] Y j =[y j (1)y j (2)...y j (n)] T (j=1,2,...,l)
[0032] In the formula: l is the number of research objects; n is the number of samples under each feature;
[0033] 122) Data normalization processing:
[0034]
[0035] Where: x i,max and x i,min The sequence X i The maximum and minimum values in y j,max and j,min Sequence Y j The maximum and minimum values in ;
[0036] 123) Calculate the grey correlation coefficient:
[0037]
[0038] Among them: ξ is the grey relational coefficient; ρ is the resolution coefficient; and They are the minimum difference between the two levels and the maximum difference between the two levels respectively;
[0039] 124) Calculate the grey relational degree:
[0040]
[0041] Among them: r is the grey relational degree.
[0042] Furthermore, in step 3, the training method steps of the mountain wind turbine fatigue load and power prediction model are as follows:
[0043] 31) Randomly generate SN nectar sources in the parameter space and calculate the fitness of the initial nectar source;
[0044] 32) Hired bee stage: Hired bees search for new nectar sources in the neighborhood of the original nectar source, calculate the fitness of the new nectar source, and compare the fitness of the new nectar source with that of the original nectar source: if the fitness of the new nectar source is greater than that of the original nectar source, the nectar source is updated; if the fitness of the new nectar source is less than that of the original nectar source, the nectar source is not updated;
[0045] 33) Follower bee stage: Follower bees select employed bees to follow according to the fitness of the nectar source shared by employed bees using a roulette wheel method. Follower bees search for new nectar sources in the neighborhood of the original nectar source, calculate the fitness of the new nectar source, and compare the fitness of the new nectar source with that of the original nectar source: if the fitness of the new nectar source is greater than that of the original nectar source, the nectar source is updated; if the fitness of the new nectar source is less than that of the original nectar source, the nectar source is not updated;
[0046] 34) Scout bee stage: count the number of times the nectar source has not been updated. When the nectar source is updated in the employed bee stage and the follower bee stage, the number of nectar source collections, trail, is set to 0; otherwise, the number of nectar source collections, trail, is increased by 1;
[0047] 35) Determine whether the number of nectar source collection trail is greater than the threshold limit: if yes, execute step 36); if no, execute 37);
[0048] 36) For the nectar source whose collection times trail is greater than the threshold limit and its two nearby points, the Nelder-Mead simplex method is used to iterate to find a better nectar source. If a better nectar source is found, the nectar source is updated and the collection times trail of the nectar source is set to 0; if no better nectar source is found, the nectar source is abandoned and a new nectar source is randomly generated, and the collection times trail of the nectar source is set to 0;
[0049] 37) Select the optimal fitness of all current nectar sources, and determine whether the number of iterations for which the optimal fitness remains unchanged reaches the set threshold abc_stop; if so, execute 39); if not, execute step 38);
[0050] 38) Determine whether the current number of iterations reaches the set maximum number of iterations abc_iterations: if yes, execute step 39); if no, execute step 32);
[0051] 39) The current nectar source position with the best fitness is mapped to the optimal hyperparameters C and σ, and the optimal hyperparameters C and σ are applied to the support vector machine, and the support vector machine training is completed.
[0052] Furthermore, in step 36), the method steps of iterating using the Nelder-Mead simplex method to find a better nectar source are as follows:
[0053] 361) Initial simplex generation: Select n+1 vertices in n-dimensional space to form an initial simplex;
[0054] 362) Function value calculation and sorting: Calculate the objective function value of each vertex, sort the vertices from small to large, and determine the best vertex and the worst vertex;
[0055] 363) Calculate the reflection point of the worst vertex in a simplex relative to the centroid of the remaining vertices if:
[0056] If the objective function value of the reflection point is between the objective function values of the optimal vertex and the suboptimal vertex, the worst vertex is replaced;
[0057] If the objective function value of the reflection point is better than that of the optimal vertex, it is further extended in the reflection direction to obtain an extended point; if the objective function value of the extended point is better than that of the reflection point, the worst vertex is replaced by the extended point, otherwise the worst vertex is replaced by the reflection point;
[0058] If the objective function value of the reflection point is not as good as that of the worst vertex, the contraction point between the centroid and the worst vertex is calculated. If the objective function value of the contraction point is worse than that of the worst vertex, the worst vertex is replaced by the contraction point. Otherwise, the entire simplex is reduced to make all vertices close to the optimal vertex.
[0059] 364) Determine whether the change of the vertex function value of the simplex satisfies the preset convergence condition: if so, the iteration ends and step 366) is executed; if not, step 365) is executed;
[0060] 365) Determine whether the current number of iterations reaches the preset maximum number of iterations nm_iterations: if so, the iteration ends; if not, execute step 362);
[0061] 366) Take the current optimal vertex of the simplex as the new nectar source and determine whether the fitness of the new nectar source is better than the original nectar source: if so, a better nectar source is found; if not, no better nectar source is found.
[0062] The beneficial effects of the present invention are:
[0063] The present invention is based on a method for predicting fatigue load and power of mountain wind turbines based on an optimized support vector machine. The grey correlation algorithm is used to analyze the correlation between input and output, and high correlation features are screened to achieve input dimension reduction. The model prediction accuracy before and after dimension reduction is compared to determine whether to add the grey correlation algorithm to the optimization model. Then, an artificial bee colony algorithm based on the Nelder-Mead simplex method is used to optimize parameters. The obtained optimal parameters are used to set the hyperparameter settings of the support vector machine model to obtain the trained optimized support vector machine model. The prediction results show that the performance of the optimized support vector machine is better than that of the basic support vector machine, and a higher accuracy has been achieved in predicting the equivalent fatigue load and power of mountain wind turbines, and the operating efficiency is higher. Compared with the prior art, the present invention has the following advantages:
[0064] 1) Improved prediction accuracy: By introducing grey correlation analysis to reduce the input dimension, the features with high correlation with fatigue load and power are effectively screened out, thereby enhancing the prediction accuracy of the model.
[0065] 2) Optimized the parameter optimization process: The artificial bee colony algorithm based on the Nelder-Mead simplex method is used to make the parameter optimization process more efficient, solving the problems of large computational complexity and low efficiency of the traditional grid search method.
[0066] 3) Enhanced model adaptability: The optimized support vector machine model performs well under complex mountain wind field conditions and can more accurately reflect the complex characteristics of mountain wind fields.
[0067] 4) Improved the generalization ability of the model: Through effective feature selection and optimization algorithms, the overfitting problem of the support vector machine on high-dimensional data is avoided, and the generalization ability of the model is improved.
[0068] 5) Wide application: This method is not only suitable for fatigue load and power prediction of mountain wind turbines, but can also be extended to other wind power scenarios that require high-precision prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0069] In order to make the purpose, technical solution and beneficial effects of the present invention clearer, the present invention provides the following drawings for illustration:
[0070] Figure 1 It is a flow chart of the method for predicting fatigue load and power of mountain wind turbine based on optimized support vector machine of the present invention;
[0071] Figure 2 This is the principle diagram of support vector machine;
[0072] Figure 3 This is a flowchart for optimizing parameters of artificial bee colony algorithm based on Nelder-Mead simplex method;
[0073] Figure 4 This is the principle diagram of the Nelder-Mead simplex method;
[0074] Figure 5 This is the layout diagram of the wind speed measurement points. DETAILED DESCRIPTION
[0075] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.
[0076] like Figure 1 As shown, the fatigue load and power prediction method of mountain wind turbine based on optimized support vector machine includes the following steps:
[0077] Step 1: Data collection and preprocessing
[0078] 11) Collect the wind speed at the measuring point in the wind turbine rotor plane, the turbulence at the hub position, the pitch angle, the rotor speed, the fatigue load and power data of the mountain wind turbine to obtain the performance parameter database of the wind turbine under mountain terrain.
[0079] Specifically, in this embodiment, a wind turbine performance parameter database under mountainous terrain is obtained based on large eddy simulation and OpenFOAM, including flow information, wind turbine control parameters, and load and power parameters. Specifically:
[0080] Incoming flow information: wind speed and turbulence within the wind turbine rotor plane.
[0081] Control parameters: rotor speed and pitch angle of wind turbine.
[0082] Load and power parameters: Equivalent fatigue load and power statistical indicators of the wind turbine during the simulation time.
[0083] 12) The grey correlation analysis method is used to analyze the correlation between input data and output data, and high correlation features are screened to achieve input dimension reduction; the input data includes the wind speed at the measuring point in the wind turbine rotor plane, the turbulence at the hub position, the pitch angle and the rotor speed; the output data includes the fatigue load and power data of mountain wind turbines.
[0084] Specifically, the steps of using the grey correlation analysis method to analyze the correlation between input data and output data are as follows:
[0085] 121) Determine the evaluation sequence X and the reference sequence Y:
[0086] The evaluation series X includes the characteristics that need to be analyzed:
[0087] X=[X1 X2...X m ]
[0088] X i =[x i (1)x i (2)...x i (n)] T (i=1,2,...,m)
[0089] Among them: m is the number of evaluation features; n is the number of samples under each feature;
[0090] The reference series are the research objects, including the statistical indicators of equivalent fatigue load and power:
[0091] Y=[Y1 Y2...Y l ]
[0092] Y j =[y j (1)y j (2)...y j (n)] T (j=1,2,...,l)
[0093] In the formula: l is the number of research objects; n is the number of samples under each feature;
[0094] 122) Data normalization processing:
[0095]
[0096] Where: x i,max and x i,min The sequence X i The maximum and minimum values in y j,max and j,min Sequence Y j The maximum and minimum values in ;
[0097] 123) Calculate the grey correlation coefficient:
[0098]
[0099] Among them: ξ is the grey relational coefficient; ρ is the resolution coefficient; and They are the minimum difference between the two levels and the maximum difference between the two levels respectively;
[0100] 124) Calculate the grey relational degree:
[0101]
[0102] Among them: r is the grey relational degree.
[0103] Table 1 shows the correlation results between the variables. The larger the value, the higher the influence of the feature on the research object. The first column in the table is the input of the evaluation series, where "Wind1VelX" means the downwind wind speed at measuring point 1, "RotSpeed" and "BldPitch1" are the rotor speed and blade pitch angle respectively, and the value of 0.619 in the cross cell of the "Wind1VelX" row and the "lateral bending moment at the bottom of the tower" column means that the gray correlation degree of the downwind wind speed at measuring point 1 and the lateral equivalent fatigue load at the bottom of the tower is 0.619, and the rest is the same.
[0104] Table 1 Grey correlation analysis results (values with correlation greater than 0.68 are bolded)
[0105]
[0106]
[0107] According to the results in the above table, it can be seen that the factors affecting the load components and power statistics of different parts of the wind turbine are not the same. Therefore, it is necessary to select relevant inputs for each prediction target and establish machine learning prediction models respectively to improve the prediction accuracy and improve the calculation efficiency.
[0108] 13) The filtered data is normalized, and the normalized wind turbine performance parameter database under mountain terrain is divided into a training set and a test set. Specifically, the normalization method is:
[0109]
[0110] Among them: max(x) and min(x) are the maximum and minimum values in the data respectively.
[0111] Step 2: Construct a mountain wind turbine fatigue load and power prediction model
[0112] This embodiment constructs a mountain wind turbine fatigue load and power prediction model based on support vector machine.
[0113] Specifically, support vector machine is a powerful supervised learning model based on statistical learning theory. Its core principle is to minimize structural risk by dividing the prediction function set into multiple subsets, keeping the empirical risk of the function set fixed, and selecting the function with the smallest confidence interval. Figure 2 As shown, the principle of support vector is:
[0114] The core goal of linear support vector machine for regression prediction is to construct the optimal hyperplane, so as to minimize the total distance error between all sample points and the hyperplane. The expression of the hyperplane is as follows:
[0115] f(x)=wx+b
[0116] Where: w is the weight vector and b is the threshold.
[0117] When calculating the error, support vector regression allows an error of ε between the predicted value and the true value. If the true value falls within the range of f(x)+ε and f(x)-ε, it is considered to be error-free. The insensitive loss function of the model is expressed as follows:
[0118]
[0119] If you want to make as many sample points as possible fall within the decision boundary, you need to maximize the distance between the two decision boundaries, which can be expressed as max(2ε / ||w||). Under the condition of controlling the accuracy ε, it can be converted into solving min(||w|| 2 / 2). In addition, the model introduces a relaxation factor ζ i and And a certain penalty is added to the sample points that are beyond the range of ε but fall within the range of ε+ζ. From the above, the final optimization objective function of the support vector machine can be deduced as follows:
[0120]
[0121] Where: C is the penalty term coefficient of the relaxation factor, the purpose of which is to balance the sample regression accuracy with the maximization of the decision boundary. The larger the C value, the greater the penalty for sample points that exceed the ε range.
[0122] For convex quadratic programming problems, the function can be constructed through the Lagrange dual method to transform the constrained optimization problem into an unconstrained optimization problem. The extreme point can be obtained by differentiation. The expression is as follows:
[0123]
[0124] Where: α i and is the Lagrange multiplier.
[0125] In addition, the extreme points found should satisfy the KKT (Karush-Kuhn-Tucker) condition:
[0126]
[0127] By contradiction, we can see that when α i ≠0, Therefore, for the above equation, we always have Next, for the sample point α i The values of Similar discussions:
[0128] When α i= 0, we know that ζ i =0, which means that the sample point falls within the decision boundary of the model or is too far away from the decision boundary;
[0129] When 0<α i When ζ is less than C, we know i =0, then y i -f(x i )-ε=0, at this time the sample point is on the decision boundary;
[0130] When α i =C, y i -f(x i )-ε-ζ i = 0, at which point the sample point is outside the decision boundary and within the allowable range of the relaxation factor. For cases 2) and 3), the sample point is called a support vector.
[0131] From this we can find w and b:
[0132]
[0133] In summary, the function expression of linear support vector regression can be obtained:
[0134]
[0135] It can be seen that support vector machines can solve the classification problem of linearly separable data. For the classification problem of nonlinearly separable data, SVM is implemented by introducing kernel functions. The radial basis kernel function is called, and the vector calculation in the original data space is used to represent the mathematical method of the dot product result in the increased dimensional space. The specific principle is as follows:
[0136] Support vector machine can solve the classification problem of linearly separable data. For the classification problem of nonlinearly separable data, SVM is implemented by introducing kernel function.
[0137] The radial basis function (RBF) is selected as the kernel function, and its expression is:
[0138]
[0139] Where: ||x (i) -x (j) || is the Euclidean distance between two eigenvectors, and σ is the kernel width, which represents the influence of a single training sample.
[0140] The input data is mapped to a high-dimensional feature space through a radial basis kernel function. In a high-dimensional space, the optimization objective function of the support vector machine becomes:
[0141]
[0142] Where: φ(x i ) represents the high-dimensional space feature vector after mapping by the radial basis kernel function.
[0143] Applying the Lagrange dual method, the constrained optimization problem is transformed into an unconstrained optimization problem, and its dual form is:
[0144]
[0145] Where: α i is the Lagrange multiplier.
[0146] By solving the above dual problem, we can obtain the optimal Lagrange multiplier α i , and then calculate the decision function of the support vector machine as:
[0147]
[0148] Where: b is the bias term, which is obtained by the support vector that meets the KKT condition.
[0149] The support vector machine model is applied to the prediction task, and the new input data x is predicted to obtain the corresponding output result f(x).
[0150] Step 3: Training mountain wind turbine fatigue load and power prediction model
[0151] The training set is used to train the fatigue load and power prediction model of mountain wind turbines. The artificial bee colony algorithm is used to optimize the hyperparameters of the support vector machine to obtain the fatigue load and power prediction model of mountain wind turbines with the optimal hyperparameters.
[0152] Specifically, Figure 3 As shown, in this embodiment, the training method steps of the mountain wind turbine fatigue load and power prediction model are as follows:
[0153] 31) Parameter setting and initialization
[0154] Setting the initial parameters includes:
[0155] Number of nectar sources SN;
[0156] The upper and lower bounds of the hyperparameters C and σ are upper_bounds and lower_bounds;
[0157] Determine the iteration round limit for over-collection of nectar sources;
[0158] The total number of iterations of the ABC algorithm is abc_iterations;
[0159] The number of times the optimal value remains unchanged before the ABC algorithm terminates is abc_stop;
[0160] The number of iterations of the Nelder-Mead algorithm, nm_iterations;
[0161] The residual tol used to determine the convergence of the Nelder-Mead algorithm;
[0162] The reflection coefficient α, contraction coefficient β, and expansion coefficient γ in the Nelder-Mead algorithm;
[0163] Run the combining algorithm a number of times runtimes.
[0164] Initialization: Randomly generate SN nectar sources in the parameter space and calculate the fitness of the initial nectar source. In this embodiment, the number of nectar sources is equal to the sum of the number of employed bees and follower bees, and the nectar source occupied by each bee represents a feasible solution in the population.
[0165] Specifically, the method for randomly generating SN nectar sources in the parameter space is:
[0166]
[0167] Where: SN is the number of nectar sources; D is the dimension of the solution; x max,j and x min,j are the upper and lower limits of the search range of the variable in the jth dimension; X i is the location of the nectar source, i.e., the potential solution; rand(0,1) is a random number between (0,1).
[0168] The calculation method of nectar source fitness is:
[0169]
[0170] Among them: fit i is the fitness of the i-th nectar source; F(X i ) is the objective function value, the objective function F(X i ) is used to output the support vector machine model evaluation accuracy score. Score is the global maximum value, and the output of the support vector machine function is set to 1 / score. The artificial bee colony algorithm globally evaluates (1 / score) min (score>0) to search, thereby unifying the optimization direction of each algorithm. The fitness value is used to characterize the quality of the current solution. For solving the minimum value problem, the smaller the solution value, the greater the fitness value. Compare the fitness of the new nectar source and the original nectar source. If the fitness value of the new nectar source is higher, it will replace the original nectar source and enter the next round of iteration. In addition, when the new nectar source exceeds the boundary of the variable search range, the boundary value is taken.
[0171] 32) Hired bee stage: Hired bees search for new nectar sources in the neighborhood of the original nectar source, calculate the fitness of the new nectar source, and compare the fitness of the new nectar source with that of the original nectar source: if the fitness of the new nectar source is greater than that of the original nectar source, the nectar source is updated; if the fitness value of the new nectar source is less than that of the original nectar source, the nectar source is not updated.
[0172] Specifically, the principle of employed bees searching for new nectar sources in the neighborhood of the original nectar source is:
[0173] v i,j =x i,j +φ i,j (x i,j -x k,j ),k∈1,2,...,D,k≠i
[0174] Where: v i,j and x i,j They are new nectar source and original nectar source, x k,j is a randomly selected nectar source, φ i,j is a random number between -1 and 1, and j is a randomly selected dimension.
[0175] 33) Follower bee stage: Follower bees select employed bees to follow according to the fitness of the nectar source shared by the employed bees using a roulette wheel method. Follower bees search for new nectar sources in the neighborhood of the original nectar source, calculate the fitness of the new nectar source, and compare the fitness of the new nectar source with that of the original nectar source: if the fitness of the new nectar source is greater than that of the original nectar source, the nectar source is updated; if the fitness of the new nectar source is less than that of the original nectar source, the nectar source is not updated.
[0176] The probability that a follower bee chooses an employed bee to follow is:
[0177]
[0178] It can be seen that follower bees have a higher probability of choosing a nectar source with higher fitness. After selecting a nectar source, follower bees search for new nectar sources in the neighborhood of the original nectar source in the same way as employed bees.
[0179] 34) Scout bee stage: Count the number of times the nectar source has not been updated. When the nectar source is updated in the employed bee stage and the follower bee stage, the nectar source collection times trail is set to 0; otherwise, the nectar source collection times trail is increased by 1.
[0180] 35) Determine whether there is a nectar source whose collection times trail is greater than the threshold limit: if so, execute step 36); if not, execute 37).
[0181] 36) For the nectar sources where the collection times trail is greater than the threshold limit and the two points nearby, the Nelder-Mead simplex method is used for iteration to find a better nectar source. If a better nectar source is found, the nectar source is updated and the collection times trail of the nectar source is set to 0; if a better nectar source is not found, the nectar source is abandoned and a new nectar source is randomly generated, and the collection times trail of this nectar source is set to 0.
[0182] Specifically, as Figure 4 shown, the method steps for using the Nelder-Mead simplex method for iteration to find a better nectar source are as follows:
[0183] 361) Initial simplex generation: Select n + 1 vertices in the n-dimensional space to form an initial simplex.
[0184] 362) Function value calculation and sorting: Calculate the objective function values of each vertex, sort the vertices in ascending order, and determine the optimal vertex and the worst vertex. Let the vertex with the smallest objective function value, that is, the optimal vertex, be called X1, the sub-optimal vertex be X2, and the worst vertex be X3. Calculate the centroid X according to the following formula c :
[0185] X c =(X1 + X2) / 2
[0186] 363) Calculate the reflection point X c of the worst vertex X3 in the simplex with respect to the centroid X of the remaining vertices r :
[0187] X r =X c +α×(X c -X3), α > 0
[0188] Where: α is the reflection coefficient. In this embodiment, α = 1.
[0189] If:
[0190] (1) The objective function value of the reflection point is between the objective function values of the optimal vertex and the sub-optimal vertex, that is:
[0191] F(X1)<F(X r )<F(X2)
[0192] Then replace the worst vertex, that is, let X3 = X r .
[0193] (2) The objective function value of the reflection point is better than the objective function value of the optimal vertex, that is, F(X r )<F(X1), then further expand in the reflection direction and obtain the expansion point X e :
[0194] X e = X c + β×(X r - X c ), β > 1
[0195] Where: β is the expansion coefficient. In this embodiment, β = 2.
[0196] If the objective function value of the expansion point is better than that of the reflection point, i.e., F(X e ) < F(X r ), then replace the worst vertex X3 with the expansion point X e , otherwise replace the worst vertex X3 with the reflection point X r .
[0197] (3) If the objective function value of the reflection point is not as good as that of the worst vertex, then calculate the contraction point between the centroid of the simplex and the worst vertex. If the objective function value of the contraction point is better than that of the worst vertex, then replace the worst vertex with the contraction point; otherwise, perform a shrinking operation on the entire simplex to make all vertices approach the optimal vertex.
[0198] Specifically, if F(X2) < F(X r ) < F(X3), then perform outer contraction according to the following formula to calculate the outer contraction point X oc :
[0199] X oc = X c + γ×(X r - X c ), 0 < γ < 1
[0200] If F(X oc ) < F(X r ), replace X3 with X oc ; otherwise, perform inner contraction, i.e.:
[0201] If F(X3) ≤ F(X r ), then perform inner contraction according to the following formula to calculate the inner contraction point X ic , and replace X3 with X ic . Specifically, the inner contraction point X ic is:
[0202] X ic = X c + γ×(X3 - X c ), 0 < γ < 1
[0203] Where: γ is the contraction coefficient. In this embodiment, γ = 0.5;
[0204] If min{F(X3), F(Xr )} <F(X ic(oc) ), then the points are reduced according to the following formula:
[0205] X i '=(X i +X1) / 2,i=2,3
[0206] 364) Determine whether the change in the vertex function value of the simplex satisfies the preset convergence condition: if so, the iteration ends and step 366) is executed; if not, step 365) is executed.
[0207] 365) Determine whether the current number of iterations reaches the preset maximum number of iterations nm_iterations: if so, the iteration ends; if not, execute step 362).
[0208] 366) Take the current optimal vertex of the simplex as the new nectar source and determine whether the fitness of the new nectar source is better than the original nectar source: if so, a better nectar source is found; if not, no better nectar source is found.
[0209] 37) Select the optimal fitness of all current nectar sources, and determine whether the number of iterations for which the optimal fitness remains unchanged reaches the set threshold abc_stop; if so, execute 39); if not, execute step 38).
[0210] 38) Determine whether the current number of iterations reaches the set maximum number of iterations abc_iterations: if so, execute step 39); if not, execute step 32).
[0211] 39) The current nectar source position with the best fitness is mapped to the optimal hyperparameters C and σ, and the optimal hyperparameters C and σ are applied to the support vector machine, and the support vector machine training is completed.
[0212] Step 4: Test the mountain wind turbine fatigue load and power prediction model
[0213] The test set is used to test the prediction performance of the trained mountain wind turbine fatigue load and power prediction model.
[0214] Step 5: Mountain wind turbine fatigue load and power prediction
[0215] The input data collected in real time, including the wind speed at the measuring point in the wind turbine rotor plane, the turbulence at the hub position, the pitch angle and the rotor speed, are input into the mountain wind turbine fatigue load and power prediction model, and the predicted mountain wind turbine fatigue load and power are output.
[0216] In this embodiment, the following model is built to compare and evaluate the fatigue load and power prediction model of the mountain wind turbine in this embodiment:
[0217] Model ①: Basic support vector machine model, without adding any optimization algorithm, and the grid search method is used in the parameter optimization process;
[0218] Model ②: After dimension optimization based on grey correlation algorithm, support vector machine is used for prediction, and grid search method is used in parameter optimization process;
[0219] Model ③: Support vector machine is used for prediction, and the artificial bee colony algorithm is used in the parameter optimization process;
[0220] Model ④: Support vector machine is used for prediction, and the improved artificial bee colony algorithm is used in the parameter optimization process.
[0221] In this embodiment, nine wind speed measurement points in the wind rotor plane are selected, such as Figure 5 The grayscale correlation analysis results are shown in Table 1, where "Wind1VelX" means the downwind wind speed at measuring point 1, and "RotSpeed" and "BldPitch1" are the rotor speed and blade pitch angle, respectively.
[0222] Compare model ① and model ② to explore the impact of grey correlation algorithm on model prediction accuracy.
[0223] Compare model ③ and model ④ to explore the effect of the artificial bee colony algorithm in parameter optimization before and after improvement.
[0224] Before training, the random number seed is fixed to ensure that the model training results are not affected by the initial random number, so that the performance of different models and parameters can be better compared. The parameter settings of each algorithm are as follows:
[0225] (1) The grid search parameter combination is C∈[10 1 ,10 2 ,10 3 ,10 4 ,10 5 ,10 6 ],γ∈[10 -3 ,10 -2 ,10 -1 ,10 0 ,10 1 ], a total of 6×5=30 groups of parameter combinations;
[0226] (2) Grey correlation algorithm selects the grey correlation degree r i,j >0.68 influencing factors;
[0227] (3) The parameter space searched in the artificial bee colony algorithm is C∈[10 1 ,10 6 ],γ∈[10 -3 ,101 ], the convergence residual is 10 -5 ,To avoid accidental errors, the algorithm is repeated 3 times;
[0228] The mean square error RMSE, mean absolute percentage error MAPE, and determination coefficient R 2 The model performance is evaluated using indicators such as
[0229] Tables 2, 3, and 4 show the comparison of prediction accuracy of different models. The tables compare the prediction accuracy in detail and show the performance of the models on each characteristic component.
[0230] Table 2 Comparison of prediction accuracy of different models (RMSE)
[0231]
[0232]
[0233] Note: The unit of the bending moment value in the table is kN·m, the unit of the power mean value is kW, and the power pulsation intensity has no unit. The values in the table retain four significant digits, and some values retain more digits because four significant digits cannot show the error of the model.
[0234] Table 3 Comparison of prediction accuracy of different models (MAPE)
[0235]
[0236] Table 4 Comparison of prediction accuracy of different models (R 2 )
[0237]
[0238] Comparing the prediction accuracy of model ① and model ②, it can be seen that the dimension reduction of the grey correlation algorithm has uncertainty in improving the accuracy of the model: for the prediction model of the tower bottom lateral bending moment and power pulsation intensity, the addition of the grey correlation algorithm effectively improves the model accuracy, while for the prediction model of the blade root swing bending moment, blade root flapping bending moment, tower bottom forward bending moment and power mean, the addition of the grey correlation algorithm reduces the model accuracy. Therefore, it is necessary to perform grey correlation analysis to select appropriate input features before training the model. Model ③ introduces the artificial bee colony algorithm. Compared with the grid search method, one of the major advantages of this method is that it is not limited by the granularity of grid division and can autonomously search for the optimal parameter combination in the parameter space, so it can obtain relatively better prediction results. Model ④ adds the Nelder-Mead simplex method to the iteration of the artificial bee colony algorithm, and the prediction accuracy is improved compared with model ③. In summary, the optimized support vector machine model after parameter optimization using grey correlation analysis and the artificial bee colony algorithm based on the Nelder-Mead simplex method can achieve high-precision prediction for most characteristic components.
[0239] The above-described embodiments are only preferred embodiments for fully illustrating the present invention, and the protection scope of the present invention is not limited thereto. Equivalent substitutions or changes made by those skilled in the art based on the present invention are within the protection scope of the present invention. The protection scope of the present invention shall be subject to the claims.
Claims
1. A method for predicting fatigue load and power of mountain wind turbines based on optimized support vector machine, characterized in that: The steps include: Step 1: Data collection and preprocessing 11) Collect the wind speed at the measuring point in the wind turbine rotor plane, the turbulence at the hub position, the pitch angle, the rotor speed, the fatigue load and power data of the mountain wind turbine to obtain the performance parameter database of the wind turbine under the mountain terrain; 12) The grey correlation analysis method is used to analyze the correlation between input data and output data, and high correlation features are selected to achieve input dimension reduction; the input data includes the wind speed at the measuring point in the wind turbine rotor plane, the turbulence at the hub position, the pitch angle and the rotor speed; the output data includes the fatigue load and power data of the mountain wind turbine; 13) normalizing the screened data, and dividing the normalized wind turbine performance parameter database under mountainous terrain into a training set and a test set according to a ratio of 8:2; Step 2: Construct a mountain wind turbine fatigue load and power prediction model A fatigue load and power prediction model for mountain wind turbines was constructed based on support vector machine; Step 3: Training mountain wind turbine fatigue load and power prediction model The training set is used to train the mountain wind turbine fatigue load and power prediction model, and the artificial bee colony algorithm is used to optimize the hyperparameters of the support vector machine to obtain the mountain wind turbine fatigue load and power prediction model with the optimal hyperparameters. Step 4: Test the mountain wind turbine fatigue load and power prediction model The test set is used to test the prediction performance of the trained mountain wind turbine fatigue load and power prediction model; Step 5: Mountain wind turbine fatigue load and power prediction The input data collected in real time, including the wind speed at the measuring point in the wind turbine rotor plane, the turbulence at the hub position, the pitch angle and the rotor speed, are input into the mountain wind turbine fatigue load and power prediction model, and the predicted mountain wind turbine fatigue load and power are output.
2. The method for predicting fatigue load and power of mountain wind turbines based on optimized support vector machine according to claim 1 is characterized in that: In the step 12), the method steps of using the grey correlation analysis method to analyze the correlation between the input data and the output data are as follows: 121) Determine the evaluation sequence X and the reference sequence Y: The evaluation series X includes the characteristics that need to be analyzed: X=[X1 X2...X m ] X i =[x i (1)x i (2)...x i (n)] T (i=1,2,...,m) Among them: m is the number of evaluation features; n is the number of samples under each feature; The reference series are the research objects, including the statistical indicators of equivalent fatigue load and power: Y=[Y1 Y2...Y l ] Y j =[y j (1)y j (2)...y j (n)] T (j=1,2,...,l) In the formula: l is the number of research objects; n is the number of samples under each feature; 122) Data normalization processing: Where: x i,max and x i,min The sequence X i The maximum and minimum values in y j,max and j,min Sequence Y j The maximum and minimum values in ; 123) Calculate the grey correlation coefficient: Among them: ξ is the grey relational coefficient; ρ is the resolution coefficient; and They are the minimum difference between the two levels and the maximum difference between the two levels respectively; 124) Calculate the grey relational degree: Among them: r is the grey relational degree.
3. The method for predicting fatigue load and power of mountain wind turbine based on optimized support vector machine according to claim 1 is characterized in that: In step 3, the training method steps of the mountain wind turbine fatigue load and power prediction model are as follows: 31) Randomly generate SN nectar sources in the parameter space and calculate the fitness of the initial nectar source; 32) Hired bee stage: Hired bees search for new nectar sources in the neighborhood of the original nectar source, calculate the fitness of the new nectar source, and compare the fitness of the new nectar source with that of the original nectar source: if the fitness of the new nectar source is greater than that of the original nectar source, the nectar source is updated; if the fitness of the new nectar source is less than that of the original nectar source, the nectar source is not updated; 33) Follower bee stage: Follower bees select employed bees to follow according to the fitness of the nectar source shared by employed bees using a roulette wheel method. Follower bees search for new nectar sources in the neighborhood of the original nectar source, calculate the fitness of the new nectar source, and compare the fitness of the new nectar source with that of the original nectar source: if the fitness of the new nectar source is greater than that of the original nectar source, the nectar source is updated; if the fitness of the new nectar source is less than that of the original nectar source, the nectar source is not updated; 34) Scout bee stage: count the number of times the nectar source has not been updated. When the nectar source is updated in the employed bee stage and the follower bee stage, the number of nectar source collections, trail, is set to 0; otherwise, the number of nectar source collections, trail, is increased by 1; 35) Determine whether the number of nectar source collection trail is greater than the threshold limit: if yes, execute step 36); if no, execute 37); 36) For the nectar source whose collection times trail is greater than the threshold limit and its two nearby points, the Nelder-Mead simplex method is used to iterate to find a better nectar source. If a better nectar source is found, the nectar source is updated and the collection times trail of the nectar source is set to 0; if no better nectar source is found, the nectar source is abandoned and a new nectar source is randomly generated, and the collection times trail of the nectar source is set to 0; 37) Select the optimal fitness of all current nectar sources, and determine whether the number of iterations for which the optimal fitness remains unchanged reaches the set threshold abc_stop; if so, execute 39); if not, execute step 38); 38) Determine whether the current number of iterations reaches the set maximum number of iterations abc_iterations: if yes, execute step 39); if no, execute step 32); 39) The current nectar source position with the best fitness is mapped to the optimal hyperparameters C and σ, and the optimal hyperparameters C and σ are applied to the support vector machine, and the support vector machine training is completed.
4. The method for predicting fatigue load and power of mountain wind turbines based on optimized support vector machine according to claim 3 is characterized in that: In step 36), the method steps of iterating using the Nelder-Mead simplex method to find a better nectar source are as follows: 361) Initial simplex generation: Select n+1 vertices in n-dimensional space to form an initial simplex; 362) Function value calculation and sorting: Calculate the objective function value of each vertex, sort the vertices from small to large, and determine the best vertex and the worst vertex; 363) Calculate the reflection point of the worst vertex in a simplex relative to the centroid of the remaining vertices if: If the objective function value of the reflection point is between the objective function values of the optimal vertex and the suboptimal vertex, the worst vertex is replaced; If the objective function value of the reflection point is better than that of the optimal vertex, it is further extended in the reflection direction to obtain an extended point; if the objective function value of the extended point is better than that of the reflection point, the worst vertex is replaced by the extended point, otherwise the worst vertex is replaced by the reflection point; If the objective function value of the reflection point is not as good as that of the worst vertex, the contraction point between the centroid and the worst vertex is calculated. If the objective function value of the contraction point is worse than that of the worst vertex, the worst vertex is replaced by the contraction point. Otherwise, the entire simplex is reduced to make all vertices close to the optimal vertex. 364) Determine whether the change of the vertex function value of the simplex satisfies the preset convergence condition: if so, the iteration ends and step 366) is executed; if not, step 365) is executed; 365) Determine whether the current number of iterations reaches the preset maximum number of iterations nm_iterations: if so, the iteration ends; If not, execute step 362); 366) Take the current optimal vertex of the simplex as the new nectar source and determine whether the fitness of the new nectar source is better than the original nectar source: If so, a better nectar source has been found; if not, a better nectar source has not been found.
Citation Information
Patent Citations
Method for forecasting short term load under demand response
CN102930347A
Wind power forecast method and device based on self-adaptation bee colony algorithm
CN103871002A
Eagle policy and adaptive NM simplex-based photovoltaic model parameter identification method
CN106485075A
Air conditioner load short-term prediction method and system
CN113255223A
Offshore wind plant field level control strategy based on distributed rolling optimization
CN115333168A
Cited By
Complex mountain flow field simulation method based on improved GEKO turbulence model
CN120724850A
A complex mountain flow field simulation method based on an improved GEKO turbulence model
CN120724850B
Method and device for quickly evaluating fatigue damage of fan under typhoon working condition
CN122112778A
