Wind power plant power and fatigue load rapid prediction and optimization method based on machine learning
By adopting machine learning and multi-objective optimization algorithms in wind farms, a fully connected neural network model is built to predict the power and fatigue load of the wind turbine and adjust the yaw angle of the wind farm, the problem of power output and fatigue load balance in the wind farm is solved. The intelligent and efficient wind farm is achieved.
Patent Information
- Application Number
- CN202411090960.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-09
- Publication Date
- 2024-11-26
AI Technical Summary
The prior art is difficult to effectively balance the power output and fatigue load of the fan in a wind farm. Traditional methods usually focus on single-target optimization, fail to comprehensively consider the interaction between the two, and rely on simple models and manual input parameters, and cannot accurately adapt to complexity. environment.
Using machine learning-based rapid prediction and optimization methods of wind farm power and fatigue load, the power and fatigue load of wind turbines are predicted using a fully connected neural network model, and combined with multi-objective optimization The algorithm adjusts the fan yaw angle to achieve the optimal solution for maximizing power output and minimizing fatigue load.
The intelligent and automated control of the wind farm is realized, which can significantly improve the energy efficiency of the wind farm and extend the service life of the fan under real-time adaptation to changing environmental conditions, and significantly reduce the fatigue load caused by yaw control through multi-objective optimization.
Smart Images

Figure CN119026944A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a power and fatigue management technology for a wind farm, and in particular to a method for optimizing the yaw angle of a wind turbine in a wind farm by using machine learning and a multi-objective optimization algorithm to increase the power output of the entire wind farm and reduce the fatigue load of the wind turbine. Background Art
[0002] In wind farms, the interaction between wind turbines, especially the interaction caused by the wake effect, is a major challenge to improving the overall wind farm efficiency. Wake refers to the phenomenon of reduced wind speed and increased turbulence experienced by wind turbines located behind upstream wind turbines. This phenomenon can reduce the energy capture efficiency of downstream wind turbines and increase mechanical fatigue, thus affecting the performance and life of wind turbines.
[0003] Although there have been many attempts to manage the wake effect and optimize the use of wind energy by adjusting the yaw angle of wind turbines, these methods often fail to fully consider the balance between the overall power output of the wind farm and the fatigue load of the wind turbines. Traditional methods usually focus on the optimization of a single goal, such as increasing the total power or reducing the fatigue load, without considering the interaction and comprehensive optimization of the two.
[0004] CN116378897A discloses a method and device for controlling the yaw angle of a wind farm, wherein the method comprises: dividing the wind turbines in the wind farm into multiple wind turbine groups according to the wake interference relationship between any two wind turbines in the wind farm; generating multiple groups of random yaw angles corresponding to each wind turbine group for each wind turbine group, calculating the total power generation and average fatigue load of the wind turbine group when the yaw angles of each wind turbine group in the wind turbine group are each random yaw angle in each group of random yaw angles; inputting each group of random yaw angles corresponding to the wind turbine group and the total power generation and average fatigue load corresponding to each group of random yaw angles into an adaptive evaluation multi-objective particle swarm algorithm, outputting the target yaw angle of each wind turbine group; and adjusting the yaw angles of each wind turbine group in the wind turbine group using the target yaw angle. In this way, the total power generation of the wind farm is increased and the fatigue load of the wind turbine group is reduced.
[0005] CN 116739313 A discloses a wind farm layout method, device, equipment and medium considering wake control. The method comprises obtaining wind condition data and current wind turbine position coordinates in a target wind farm, calculating the wind farm power generation considering the wake effect, iteratively optimizing the current wind turbine position coordinates, and generating optimized wind turbine position coordinates; calculating the wind farm power generation considering the wake effect under the yaw state of the wind turbine based on the optimized wind turbine position coordinates, wind condition data and the current wind turbine yaw angle, iteratively optimizing the current wind turbine yaw angle, and generating an optimized wind turbine yaw angle; iteratively optimizing the optimized wind turbine position coordinates and the optimized wind turbine yaw angle, and generating target wind farm layout information. The present invention can further reduce the influence of the wake effect on the wind farm power generation and improve the power generation level of the wind farm.
[0006] CN 115544907 A discloses a method for predicting wind power and optimizing stock wind power based on wake effect, comprising the following steps: S1: firstly, importing environmental factors and the thrust coefficient of the most upstream wind turbine, and importing the predicted dynamic characteristics of the wind turbine. In the present invention, the operation function for single column data is realized through steps 1-4, and in the operation process, environmental factor data such as wind shear and turbulence intensity, as well as aerodynamic characteristics of the airfoil in the blade and dynamic characteristic data such as the pitch angle of each blade are added, and the Navier-Stokes equation controlling the entire flow field in the wind farm is numerically solved through the Jensen model, and the wind speed attenuation is calculated, thereby improving the accuracy of the operation result. In steps 5-7, the overall prediction function for wind power is achieved through the prediction and analysis of the overall data, and repeated verification is performed to ensure that the overall operation of the wind farm reaches the optimal wind power rate under the environmental factors.
[0007] CN 111881572 A discloses a wind farm coordinated yaw intelligent control method based on multi-objective optimization. According to the actual situation of the wind farm and the wind turbine dynamics simulation system, a wind farm coordinated yaw simulation system and a wind turbine generator set multi-dimensional working condition load database are established; the laser radar, vibration monitor, and SCADA carried by each unit in the wind farm are used to complete the calibration of the wind farm coordinated yaw simulation system and the multi-dimensional working condition load database parameters; an objective function that can achieve multi-objective performance optimization of the wind farm is established, and the optimal solution of the yaw angle of each wind turbine generator set in the wind farm under the measured incoming wind is searched through an intelligent optimization algorithm, and the optimal yaw control command is transmitted to each unit to achieve wind farm coordinated yaw intelligent control based on multi-objective optimization. The present invention comprehensively considers the influence of wake on the operation of the wind farm, reduces the load influence of wake on the downstream wind turbine generator set while improving the total output power of the wind farm, and achieves multi-objective optimization of the wind farm.
[0008] In addition, existing wind farm control strategies often rely on relatively simple models or require manual input of a large number of parameters, which limits their application effect and efficiency in complex environments. Due to the high dynamics and complexity of wind farm conditions, these methods often cannot accurately predict the specific response of each wind turbine, resulting in the inability to achieve optimal control.
[0009] Therefore, there is an urgent need for a technical solution that can automatically and accurately comprehensively consider and optimize the power output and fatigue load in the wind farm. The solution needs to be able to adapt to changing environmental conditions in real time and accurately adjust the yaw angle through intelligent algorithms to maximize the energy efficiency of the wind farm and extend the service life of the wind turbine. Summary of the invention
[0010] Purpose of the invention: The technical problem to be solved by the present invention is to provide a method for rapid prediction and optimization of wind farm power and fatigue load based on machine learning in view of the deficiencies of the prior art. The method can effectively reduce the fatigue load of the wind turbine while ensuring the maximization of the total power of the wind farm. In addition, the present invention aims to solve the problem that the prior art only optimizes power or fatigue load and lacks comprehensive consideration of the balance between the two. By integrating advanced machine learning models and multi-objective optimization algorithms, the intelligent and automated operation of wind farms is realized to meet the needs of the wind power industry for high efficiency and high reliability.
[0011] In order to solve the above technical problems, the present invention discloses a method for rapid prediction and optimization of wind farm power and fatigue load based on machine learning, comprising the following steps:
[0012] Step 1: Establish a high-fidelity simulation system to simulate the operation of the wind farm under various airflow conditions by setting background wind field parameters and wake effect parameters to accurately evaluate the interaction between wind turbines and the wake effect. The background wind field parameters include wind speed and turbulence; the wake effect parameters include the flow spacing between upstream and downstream wind turbines.
[0013] Step 2: According to the required working conditions, the high-fidelity multi-physics field coupling software FAST.Farm is used to calculate the wind turbine power and load under different background wind field parameters and wake effect parameters, wherein the load includes blade root load and tower base load; the blade root load includes the force in the blade root surface, the moment in the blade root surface, the force outside the blade root surface, and the moment outside the blade root surface; the tower base load includes the tower base pitch direction force, the tower base pitch direction moment, the tower base lateral force, and the tower base lateral moment; the above-mentioned continuously fluctuating time series load is converted into an equivalent fatigue load, which is also referred to as fatigue load in this article;
[0014] Step 3, taking the wind speed, turbulence, flow spacing, and yaw angle of the upstream wind turbine in the wind farm as input, and the power, blade root load, and tower base load corresponding to each wind turbine in the wind farm as output, a fully connected neural network model is constructed and trained; specifically, corresponding to each wind turbine, its output includes the fatigue load corresponding to the blade root resultant force, the fatigue load corresponding to the resultant moment, the fatigue load corresponding to the tower base resultant force, the fatigue load corresponding to the resultant moment, and the power corresponding to the wind turbine, that is, 4 fatigue loads and one power value. If there are 4 wind turbines, their output corresponds to 16 fatigue loads and 4 powers, 20 output values.
[0015] Step 4: Input the operating data of wind speed, turbulence, flow spacing, and yaw angle of the upstream wind turbine into the neural network model trained in step 3 to obtain the power of the wind turbine and the corresponding fatigue load, and use them as the mapping relationship between variables and functions in the optimization algorithm. The decision variable of the optimization algorithm is the yaw angle of the upstream wind turbine. The optimal solution of maximizing wind field power and minimizing fatigue load is obtained by optimizing the yaw angle of the upstream wind turbine.
[0016] Among them, in step 2, the fatigue load calculation method is:
[0017]
[0018] Where D total Represents the total damage; N i (﹒) represents the number of cycles of the load amplitude; n i represents the number of cycles of the i-th load amplitude; represents the load amplitude corresponding to the fixed load mean;
[0019] Load Amplitude The relationship between the number of cycles Ni(﹒) is as follows:
[0020]
[0021] Where, L ult is the ultimate design load of the corresponding part of the blade or tower base; L MF is the fixed load mean; m is index. The index depends on the material of the specific parts of the wind turbine. According to the specification, the tower base The index is generally 3-5, the leaves The index is generally 8-10.
[0022] Preferably, the load amplitude Correction is performed using the Goodman correction model:
[0023]
[0024] In the formula Indicates the mean load The load amplitude.
[0025] Specifically, the fatigue load calculation method for each time series is as follows:
[0026]
[0027] n STeq =f eq ·T (20)
[0028]
[0029]
[0030] Among them, D ST Indicates total damage;
[0031] f eq It represents the frequency of fatigue load, T represents the short-term load application time of a single time series. In conventional calculations, f is taken as 1, and f*T=n represents the total number of equivalent fatigue cycles. For example, each simulation time is 900s. According to the IEC standard, T is taken as the last 600s to obtain simulation data under the influence of wake. Under normal circumstances, 10 minutes is taken for fatigue load calculation. The first 300 seconds are discarded because the downstream wind turbine is not affected by the wake of the upstream wind turbine in the first 300s, and the data should be discarded. The simulation time discarded is different at different spacings. The closer the distance, the less time is discarded. It is completely sufficient to discard the first 300s within 8D;
[0032] n STeq Represents the equivalent number of fatigue cycles of a single time series;
[0033] N eq Indicates the equivalent number of cycles corresponding to the load amplitude on the SN curve;
[0034] DEL ST DEL for a single time series short-term load ST .
[0035] Specifically, the calculation method of the fatigue load corresponding to the resultant force on the inner and outer surfaces of the blade root surface, the fatigue load corresponding to the resultant moment on the inner and outer surfaces of the blade root surface, the fatigue load corresponding to the resultant force in the lateral pitching direction of the tower bottom, and the fatigue load corresponding to the resultant moment in the lateral pitching direction of the tower bottom is as follows:
[0036]
[0037] In the formula, DEL RootFxy Represents the fatigue load corresponding to the resultant force in the in-plane and out-plane directions of the blade; DEL RootFxRepresents the fatigue load corresponding to the shear force in the blade surface; DEL RootFy Represents the fatigue load corresponding to the shear force in the out-of-plane direction of the blade;
[0038] DEL RootMxy Represents the fatigue load corresponding to the resultant moment in the in-plane and out-plane directions of the blade; DEL RootMx Represents the fatigue load corresponding to the moment in the blade surface; DEL RootMy Represents the fatigue load corresponding to the moment in the out-of-plane direction of the blade;
[0039] DEL TwrBsFxy Represents the fatigue load corresponding to the resultant force in the lateral pitch direction of the tower bottom; DEL TwrBsFx Represents the fatigue load corresponding to the pitch shear force at the bottom of the tower; DEL TwrBsFy Represents the fatigue load corresponding to the lateral shear force at the bottom of the tower;
[0040] DEL TwrBsMxy DEL represents the fatigue load corresponding to the resultant moment in the lateral pitch direction of the tower bottom; TwrBsMx Represents the fatigue load corresponding to the lateral moment at the bottom of the tower; DEL TwrBsMy Represents the fatigue load corresponding to the pitching moment at the bottom of the tower.
[0041] In some embodiments, for a system with four wind turbines, the neural network model in step (3) adopts a 4-90-90-90-90-20 network structure, which has four inputs, namely wind speed, turbulence, flow spacing, and upstream wind turbine yaw angle; four hidden layers, each with 90 neurons; and 20 network outputs, including the power corresponding to the four wind turbines, the fatigue load corresponding to the blade root resultant force, and the fatigue load corresponding to the resultant moment. The activation function in the hidden layer is "tanh", and no activation function is used in the output layer.
[0042] The mean square error is used as the loss function, and the specific formula is:
[0043]
[0044] MSE stands for mean square error, where n is the total number of samples and y is i is the actual value of the ith sample, is the model's predicted value for the i-th sample;
[0045] After the network iteration is terminated, the following evaluation indicators are used to evaluate the network prediction performance:
[0046]
[0047]
[0048]
[0049] in represents the regression coefficient; y i and are the true value and the predicted value respectively; max and min are the maximum and minimum values of the true value respectively; n represents the number of samples; MAPE represents the mean absolute percentage error; NRMSE represents the normalized root mean square error.
[0050] In step 4, the wind speed, turbulence, and flow spacing in the wind farm are substituted as inputs into the trained neural network model, and the power corresponding to each wind turbine and the fatigue load corresponding to the blade root force and moment are obtained according to the neural network; the fatigue load corresponding to the tower base force and moment are obtained, and the input and output of the wind farm are used as the functional relationship between the functions in the optimization model.
[0051] Specifically, in step 4, the objective function of the optimization model is:
[0052]
[0053]
[0054]
[0055] In the formula, f 1 represents the objective function corresponding to the maximum power; the subscript i is the wind turbine number; θ is the yaw angle of the wind turbine; 0° means the yaw angle is 0°, that is, no yaw; P i (θ) represents the wind turbine power of the i-th wind turbine corresponding to the yaw angle θ; P i (0°) represents the wind turbine power of the i-th wind turbine when the yaw angle is 0°; f 2 represents the objective function corresponding to minimizing the resultant force; the subscript TF represents the tower base resultant force; i represents the wind turbine number; θ represents the yaw angle of the wind turbine; DEL TFi (θ) represents the fatigue load corresponding to the tower base resultant force of the i-th wind turbine when the yaw angle is θ; the subscript RF represents the blade root resultant force; DEL RFi (θ) represents the fatigue load corresponding to the blade root force of the i-th wind turbine when the yaw angle is θ; DEL TFi (0°) represents the fatigue load corresponding to the tower foundation force when the i-th wind turbine is not yawed; DEL RFi (0°) represents the fatigue load corresponding to the blade root force when the i-th wind turbine is not yawed; f 3 represents the objective function corresponding to the minimization of the resultant moment; the subscript TM represents the tower base resultant moment; i represents the wind turbine number; DEL TMi(θ) represents the fatigue load corresponding to the tower base moment of the i-th wind turbine when the yaw angle is θ; the subscript RM represents the blade root moment; DEL RMi (θ) represents the fatigue load corresponding to the blade root force of the i-th wind turbine when the yaw angle is θ; DEL TMi (0°) represents the fatigue load corresponding to the tower foundation moment when the i-th wind turbine is not yawed; DEL RMi (0°) represents the fatigue load corresponding to the blade root moment when the i-th wind turbine is not yawed;
[0056] After the objective function of the optimization model is determined, the optimization model is oriented to maximize f 1 direction, minimize f 2 、f 3 The Pareto frontier is obtained by optimizing the direction. The Utopia point is determined as the reference point, and the optimal solution is obtained by calculating the solution with the closest Euclidean distance to the Utopia point in the Pareto frontier. The Utopia point is defined as the maximum f in the Pareto frontier during each optimization process. 1 and minimum f 2 、f 3 .
[0057] The present invention further proposes a device for rapid prediction and optimization of wind farm power and fatigue load, comprising a memory and one or more processors communicatively connected thereto, wherein the memory stores computer instructions, and the processor executes the computer instructions to execute any of the above-mentioned methods for rapid prediction and optimization of wind farm power and fatigue load.
[0058] The present invention further proposes a computer-readable storage medium, on which computer instructions are stored, and the computer instructions are used to enable a computer to execute the above-mentioned method for rapid prediction and optimization of wind farm power and fatigue load.
[0059] Beneficial effects: The present invention provides a wind farm multi-objective optimization method that uses a neural network prediction and a joint optimization algorithm to perform multi-objective optimization control. The multi-objective optimization framework designed by the present invention can significantly reduce the fatigue load caused by yaw control while increasing power. BRIEF DESCRIPTION OF THE DRAWINGS
[0060] Figure 1 It is the background wind farm layout diagram of the present invention;
[0061] Figure 2 is a flow chart of the method of the present invention;
[0062] Figure 3 It is a diagram of the neural network structure;
[0063] Figure 4 It is a comparison chart between the predicted value of the neural network and the true value;
[0064] Figure 5 This is the effect diagram of power increase and load reduction after optimization. DETAILED DESCRIPTION
[0065] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent modifications of the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.
[0066] like Figure 1 As shown, the present invention proposes a method for rapid prediction and optimization of wind farm power and fatigue load based on machine learning, the method comprising the following steps:
[0067] Step 1: Set up a high-fidelity simulation system (i.e. Figure 2 The high-precision wind farm multi-physics field coupling model of the wind farm is used to simulate the operation status of the wind farm under various airflow conditions to accurately evaluate the interaction between wind turbines and the wake effect. Among them, the background wind field parameters include wind speed and turbulence; the wake effect parameters include the flow spacing between upstream and downstream wind turbines;
[0068] Step 2: According to the required working conditions, use the high-fidelity multi-physics field coupling software FAST.Farm to calculate the wind turbine power and load under different background wind field parameters and wake effect parameters, wherein the load includes blade root load and tower base load; the blade root load includes the force in the blade root surface, the moment in the blade root surface, the force outside the blade root surface, and the moment outside the blade root surface; the tower base load includes the tower base pitch direction force, the tower base pitch direction moment, the tower base lateral force, and the tower base lateral moment; convert the continuously fluctuating time series load into an equivalent fatigue load, hereinafter referred to as fatigue load;
[0069] Step 3: Taking the wind speed, turbulence, flow spacing, and yaw angle of the upstream wind turbine in the wind farm as input, and the power, blade root load, and tower base load corresponding to each wind turbine in the wind farm as output, a fully connected neural network model is constructed and trained;
[0070] Step 4: Input the operating data of wind speed, turbulence, flow spacing, and yaw angle of the upstream wind turbine into the neural network model trained in step 3 to obtain the power of the wind turbine and the corresponding fatigue load, and use them as the mapping relationship between variables and functions in the optimization algorithm. The decision variable of the optimization algorithm is the yaw angle of the upstream wind turbine. The optimal solution of maximizing wind field power and minimizing fatigue load is obtained by optimizing the yaw angle of the upstream wind turbine.
[0071] Among them, in step 1, the wind turbine power calculation formula is:
[0072]
[0073] Where: P represents wind turbine power; ρ represents air density, A represents rotor area, C P represents the thrust coefficient, V x represents wind speed; the subscript x represents axial wind speed; θ is the yaw angle.
[0074] In step 2, the fatigue load calculation method is:
[0075]
[0076] Where D total Represents the total damage; N i (﹒) represents the number of cycles of the load amplitude; n i represents the number of cycles of the i-th load amplitude; represents the load amplitude corresponding to the fixed load mean;
[0077] Load Amplitude The relationship between the number of cycles Ni(﹒) is as follows:
[0078]
[0079] Where, L ult is the ultimate design load of the corresponding part of the blade or tower base; L MF is the fixed load mean; m is index. The index depends on the material of the specific parts of the wind turbine. According to the specification, the tower base The index is generally 3-5, the leaves The index is generally 8-10.
[0080] The load amplitude Correction is performed using the Goodman correction model:
[0081]
[0082] In the formula Indicates the mean load The load amplitude.
[0083] Specifically, the fatigue load calculation method for each time series is as follows:
[0084]
[0085] n STeq =f eq ·T(36)
[0086]
[0087]
[0088] Among them, D ST Indicates total damage;
[0089] f eq It represents the frequency of fatigue load, T represents the short-term load application time of a single time series. In conventional calculations, f is taken as 1, and f*T=n represents the total number of equivalent fatigue cycles. For example, each simulation time is 900s. According to the IEC standard, T is taken as the last 600s to obtain simulation data under the influence of wake. Under normal circumstances, 10 minutes is taken for fatigue load calculation. The first 300 seconds are discarded because the downstream wind turbine is not affected by the wake of the upstream wind turbine in the first 300s, and the data should be discarded. The simulation time discarded is different at different spacings. The closer the distance, the less time is discarded. It is completely sufficient to discard the first 300s within 8D;
[0090] n STeq Represents the equivalent number of fatigue cycles of a single time series;
[0091] N eq Indicates the equivalent number of cycles corresponding to the load amplitude on the SN curve;
[0092] DEL ST DEL for a single time series short-term load ST .
[0093] Specifically, the calculation method of the fatigue load corresponding to the resultant force on the inner and outer surfaces of the blade root surface, the fatigue load corresponding to the resultant moment on the inner and outer surfaces of the blade root surface, the fatigue load corresponding to the resultant force in the lateral pitching direction of the tower bottom, and the fatigue load corresponding to the resultant moment in the lateral pitching direction of the tower bottom is as follows:
[0094]
[0095] In the formula, DEL RootFxy Represents the fatigue load corresponding to the resultant force in the in-plane and out-plane directions of the blade; DEL RootFx Represents the fatigue load corresponding to the shear force in the blade surface; DEL RootFy Represents the fatigue load corresponding to the shear force in the out-of-plane direction of the blade;
[0096] DEL RootMxy Represents the fatigue load corresponding to the resultant moment in the in-plane and out-plane directions of the blade; DEL RootMx Represents the fatigue load corresponding to the moment in the blade surface; DEL RootMy Represents the fatigue load corresponding to the moment in the out-of-plane direction of the blade;
[0097] DEL TwrBsFxy Represents the fatigue load corresponding to the resultant force in the lateral pitch direction of the tower bottom; DELTwrBsFx Represents the fatigue load corresponding to the pitch shear force at the bottom of the tower; DEL TwrBsFy Represents the fatigue load corresponding to the lateral shear force at the bottom of the tower;
[0098] DEL TwrBsMxy DEL represents the fatigue load corresponding to the resultant moment in the lateral pitch direction of the tower bottom; TwrBsMx Represents the fatigue load corresponding to the lateral moment at the bottom of the tower; DEL TwrBsMy Represents the fatigue load corresponding to the pitching moment at the bottom of the tower.
[0099] In some embodiments, for a system with four wind turbines, the neural network model in step (3) adopts a 4-90-90-90-90-20 network structure, which has four inputs, namely wind speed, turbulence, flow spacing, and upstream wind turbine yaw angle; four hidden layers, each with 90 neurons; and 20 network outputs, including the power corresponding to the four wind turbines, the fatigue load corresponding to the blade root resultant force, and the fatigue load corresponding to the resultant moment. The activation function in the hidden layer is "tanh", and no activation function is used in the output layer.
[0100] The mean square error is used as the loss function, and the specific formula is:
[0101]
[0102] MSE stands for mean square error, where n is the total number of samples and y is i is the actual value of the ith sample, is the model's predicted value for the i-th sample;
[0103] After the network iteration is terminated, the following evaluation indicators are used to evaluate the network prediction performance:
[0104]
[0105]
[0106]
[0107] in represents the regression coefficient; y i and are the true value and the predicted value respectively; max and min are the maximum and minimum values of the true value respectively; n represents the number of samples; MAPE represents the mean absolute percentage error; NRMSE represents the normalized root mean square error.
[0108] In step 4, the wind speed, turbulence, and flow spacing in the wind farm are substituted as inputs into the trained neural network model, and the power corresponding to each wind turbine and the fatigue load corresponding to the blade root force and moment are obtained according to the neural network; the fatigue load corresponding to the tower base force and moment are obtained, and the input and output of the wind farm are used as the functional relationship between the functions in the optimization model.
[0109] Specifically, in step 4, the objective function of the optimization model is:
[0110]
[0111]
[0112]
[0113] In the formula, f 1 represents the objective function corresponding to the maximum power; the subscript i is the wind turbine number; θ is the yaw angle of the wind turbine; 0° means the yaw angle is 0°, that is, no yaw; P i (θ) represents the wind turbine power of the i-th wind turbine corresponding to the yaw angle θ; P i (0°) represents the wind turbine power of the i-th wind turbine when the yaw angle is 0°; f 2 represents the objective function corresponding to minimizing the resultant force; the subscript TF represents the tower base resultant force; i represents the wind turbine number; θ represents the yaw angle of the wind turbine; DEL TFi (θ) represents the fatigue load corresponding to the tower base resultant force of the i-th wind turbine when the yaw angle is θ; the subscript RF represents the blade root resultant force; DEL RFi (θ) represents the fatigue load corresponding to the blade root force of the i-th wind turbine when the yaw angle is θ; DEL TFi (0°) represents the fatigue load corresponding to the tower foundation force when the i-th wind turbine is not yawed; DEL RFi (0°) represents the fatigue load corresponding to the blade root force when the i-th wind turbine is not yawed; f 3 represents the objective function corresponding to the minimization of the resultant moment; the subscript TM represents the tower base resultant moment; i represents the wind turbine number; DEL TMi (θ) represents the fatigue load corresponding to the tower base moment of the i-th wind turbine when the yaw angle is θ; the subscript RM represents the blade root moment; DEL RMi (θ) represents the fatigue load corresponding to the blade root force of the i-th wind turbine when the yaw angle is θ; DEL TMi (0°) represents the fatigue load corresponding to the tower foundation moment when the i-th wind turbine is not yawed; DEL RMi (0°) represents the fatigue load corresponding to the blade root moment when the i-th wind turbine is not yawed;
[0114] After the objective function of the optimization model is determined, the optimization model is oriented to maximize f1 direction, minimize f 2 、f 3 The Pareto frontier is obtained by optimizing the direction. The Utopia point is determined as the reference point, and the optimal solution is obtained by calculating the solution with the closest Euclidean distance to the Utopia point in the Pareto frontier. The Utopia point is defined as the maximum f in the Pareto frontier during each optimization process. 1 and minimum f 2 、f 3 .
[0115] Case Analysis
[0116] The present invention uses the Tensorflow platform to build a fully connected neural network and writes an optimization algorithm in Python. The wind farm is constructed using FAST.Farm. The layout of the wind farm is as follows: Figure 1 As shown, Turbsim is used to generate the background wind field, and the DWM model is used to generate the wake. The wind farm has a total of 4 wind turbines, arranged in a 2×2 layout, with two wind turbines upstream and two wind turbines downstream. The upstream wind turbines yaw in coordination, and the downstream wind turbines do not yaw. The span spacing is fixed to 4D, and the wind farm calculation parameters are shown in Table 1:
[0117] Table 1 Wind field parameters
[0118]
[0119] The neural network structure adopts a 4-90-90-90-90-20 network structure. The network has 4 inputs, namely wind speed, turbulence, flow spacing, and upstream wind turbine yaw angle; 4 hidden layers, each with 90 neurons; 20 network outputs, including the power corresponding to the four wind turbines, the fatigue loads corresponding to the blade root force and moment, and the fatigue loads corresponding to the tower base force and moment. The activation function in the hidden layer is "tanh", and no activation function is used in the output layer. The mean square error is used as the loss function, and the specific formula is:
[0120]
[0121] MSE stands for mean square error, where n is the total number of samples and y is i is the actual value of the ith sample, is the predicted value of the model for the i-th sample. The algorithm terminates when the number of iterations reaches 35,000. After the network iteration is terminated, the following evaluation indicators are used to evaluate the network prediction performance:
[0122]
[0123]
[0124]
[0125] in represents the regression coefficient; y i and are the true value and the predicted value respectively; max and min are the maximum and minimum values of the true value respectively; n represents the number of samples; MAPE represents the mean absolute percentage error; NRMSE represents the normalized root mean square error.
[0126] Based on this example, the method of the present invention is used to simulate the prediction accuracy of wind farm power and fatigue load under different evaluation indicators under different cases (see Table 2 for results), and the neural network structure framework is given (see Table 2 for results). Figure 3 ), and the comparison results between the true value and the predicted value when the proposed method is trained using a neural network are given (see the results in Figure 4 ), and also gives the power improvement and fatigue load reduction effects after optimizing the prediction results using a multi-objective optimization algorithm (see the results in Figure 5 ). The results show that after optimizing the prediction results using an optimization algorithm, the fatigue load can be significantly reduced while maintaining a high power boost effect.
[0127] Table 2 Prediction accuracy of neural network under different indicators
[0128]
[0129] Among them, the subscript "1" represents the two upstream wind turbines, and the subscript "2" represents the two upstream wind turbines; the subscript "RF" represents the blade root resultant force, and the subscript "RM" represents the blade root resultant moment; the subscript "TF" represents the tower base resultant force, and the subscript "TM" represents the tower base resultant moment.
[0130] The present invention uses a machine learning-based method to quickly predict the power and fatigue load of a wind turbine under yaw control and optimize the yaw. By combining data-driven with a multi-objective optimization algorithm, rapid prediction and optimization of the required working conditions are achieved.
[0131] The above description is only a specific implementation of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a technician familiar with the technical field within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention.
Claims
1. A method for rapid prediction and optimization of wind farm power and fatigue load based on machine learning, comprising the following steps: Step 1: Establish a high-fidelity simulation system to simulate the operation of the wind farm under various airflow conditions by setting background wind field parameters and wake effect parameters. Background wind field parameters include wind speed and turbulence; wake effect parameters include the flow spacing between upstream and downstream wind turbines; Step 2: According to the required working conditions, use the high-fidelity multi-physics field coupling software FAST.Farm to calculate the wind turbine power and load under different background wind field parameters and wake effect parameters, wherein the load includes blade root load and tower base load; the blade root load includes the force in the blade root surface, the moment in the blade root surface, the force outside the blade root surface, and the moment outside the blade root surface; the tower base load includes the tower base pitch direction force, the tower base pitch direction moment, the tower base lateral force, and the tower base lateral moment; convert the above-mentioned continuously fluctuating time series load into an equivalent fatigue load; Step 3: Taking the wind speed, turbulence, flow spacing, and yaw angle of the upstream wind turbine in the wind farm as input, and the power, blade root load, and tower base load corresponding to each wind turbine in the wind farm as output, a fully connected neural network model is constructed and trained to obtain the neural network model; Step 4: Input the operating data of wind speed, turbulence, flow spacing, and yaw angle of the upstream wind turbine into the neural network model trained in step 3 to obtain the power of the wind turbine and the corresponding fatigue load, and use them as the mapping relationship between variables and functions in the optimization algorithm. The decision variable of the optimization algorithm is the yaw angle of the upstream wind turbine. The optimal solution of maximizing wind field power and minimizing fatigue load is obtained by optimizing the yaw angle of the upstream wind turbine.
2. The method according to claim 1, characterized in that In step 2, the equivalent fatigue load is calculated as: Where D total Indicates total damage; N i (﹒) represents the number of cycles of the load amplitude; n i represents the number of cycles of the i-th load amplitude; represents the load amplitude corresponding to the fixed load mean; Load Amplitude The relationship between the number of cycles Ni(﹒) is as follows: Where, L ult is the ultimate design load of the corresponding part of the blade or tower base; L MF is a fixed load mean; m is index.
3. The method according to claim 2, characterized in that The load amplitude Correction is performed using the Goodman correction model: In the formula Indicates the mean load The load amplitude.
4. The method according to claim 1, characterized in that: The fatigue load for each time series is calculated as follows: n STeq =f eq ·T(5) Among them, D ST represents the total damage, f eq represents the frequency of fatigue load, T represents the short-term load application time n of a single time series STeq Represents the equivalent number of fatigue cycles of a single time series, N eq Indicates the equivalent number of cycles corresponding to the load amplitude on the SN curve, DEL ST DEL for a single time series short-term load ST .
5. The method according to claim 1, characterized in that The calculation methods for the fatigue load corresponding to the resultant force on the inner and outer surfaces of the blade root surface, the fatigue load corresponding to the resultant moment on the inner and outer surfaces of the blade root surface, the fatigue load corresponding to the resultant force in the lateral pitching direction of the tower bottom, and the fatigue load corresponding to the resultant moment in the lateral pitching direction of the tower bottom are as follows: In the formula, DEL RootFxy Represents the fatigue load corresponding to the resultant force in the in-plane and out-plane directions of the blade; DEL RootFx Represents the fatigue load corresponding to the shear force in the blade surface; DEL RootFy Represents the fatigue load corresponding to the shear force in the out-of-plane direction of the blade; DEL RootMxy Represents the fatigue load corresponding to the resultant moment in the in-plane and out-plane directions of the blade; DEL RootMx Represents the fatigue load corresponding to the moment in the blade surface; DEL RootMy Represents the fatigue load corresponding to the moment in the out-of-plane direction of the blade; DEL TwrBsFxy Represents the fatigue load corresponding to the resultant force in the lateral pitch direction of the tower bottom; DEL TwrBsFx Represents the fatigue load corresponding to the pitch shear force at the bottom of the tower; DEL TwrBsFy Represents the fatigue load corresponding to the lateral shear force at the bottom of the tower; DEL TwrBsMxy DEL represents the fatigue load corresponding to the resultant moment in the lateral pitch direction of the tower bottom; TwrBsMx Represents the fatigue load corresponding to the lateral moment at the bottom of the tower; DEL TwrBsMy Represents the fatigue load corresponding to the pitching moment at the bottom of the tower.
6. The method according to claim 1, characterized in that In step 3, the mean square error is used as the loss function in the neural network model training, and the specific formula is: MSE stands for mean square error, where n is the total number of samples and y is i is the actual value of the ith sample, is the model's predicted value for the i-th sample; After the network iteration is terminated, the following evaluation indicators are used to evaluate the network prediction performance: in represents the regression coefficient; y i and are the true value and the predicted value respectively; max and min are the maximum and minimum values of the true value respectively; n represents the number of samples; MAPE represents the mean absolute percentage error; NRMSE represents the normalized root mean square error.
7. The method according to claim 1, characterized in that In step 4, the wind speed, turbulence, and flow spacing in the wind farm are substituted as inputs into the trained neural network model, and the power corresponding to each wind turbine and the fatigue load corresponding to the blade root force and moment are obtained according to the neural network; the fatigue load corresponding to the tower base force and moment are obtained, and the input and output of the wind farm are used as the functional relationship between the functions in the optimization model.
8. The method according to claim 1, characterized in that In step 4, the objective function of the optimization model is: Where f1 represents the objective function corresponding to the maximum power; subscript i is the wind turbine number; θ is the yaw angle of the wind turbine; 0° means the yaw angle is 0°, that is, no yaw; P i (θ) represents the wind turbine power of the i-th wind turbine corresponding to the yaw angle θ; P i (0°) represents the wind turbine power of the i-th wind turbine when the yaw angle is 0°; f2 represents the objective function corresponding to the minimized resultant force; subscript TF represents the tower base resultant force; i represents the wind turbine number; θ represents the wind turbine yaw angle; DEL TFi (θ) represents the fatigue load corresponding to the tower foundation resultant force of the i-th wind turbine when the yaw angle is θ; Subscript RF represents the blade root force; DEL RFi (θ) represents the fatigue load corresponding to the blade root force of the i-th wind turbine when the yaw angle is θ; DEL TFi (0°) represents the fatigue load corresponding to the tower foundation force when the i-th wind turbine is not yawed; DEL RFi (0°) represents the fatigue load corresponding to the blade root force of the i-th wind turbine when it is not yawed; f3 represents the objective function corresponding to the minimized combined moment; subscript TM represents the tower base combined moment; i represents the wind turbine number; DEL TMi (θ) represents the fatigue load corresponding to the tower base moment of the i-th wind turbine when the yaw angle is θ; the subscript RM represents the blade root moment; DEL RMi (θ) represents the fatigue load corresponding to the blade root force of the i-th wind turbine when the yaw angle is θ; DEL TMi (0°) represents the fatigue load corresponding to the tower foundation moment when the i-th wind turbine is not yawed; DEL RMi (0°) represents the fatigue load corresponding to the blade root moment when the i-th wind turbine is not yawed; After the objective function of the optimization model is determined, optimization is performed in the direction of maximizing f1 and minimizing f2 and f3 to obtain the Pareto frontier. The method of determining the utopia point as the reference point is adopted to obtain the optimal solution by calculating the solution with the closest Euclidean distance to the utopia point in the Pareto frontier; the utopia point is defined as the maximum f1 and the minimum f2 and f3 in the Pareto frontier during each optimization process.
9. A device for rapid prediction and optimization of wind farm power and fatigue load, comprising a memory and one or more processors in communication therewith, wherein the memory stores computer instructions, and the processor executes the computer instructions to execute the method for rapid prediction and optimization of wind farm power and fatigue load according to any one of claims 1 to 6.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a computer to execute the method for rapid prediction and optimization of wind farm power and fatigue load according to any one of claims 1 to 5.
Citation Information
Cited By
Wind turbine generator aerodynamic performance real-time optimization method and device, medium and equipment
CN120867950A
