Wind farm primary frequency regulation control parameter optimization method and system considering fatigue load prediction

Abstract

The present application relates to the technical field of smart grid, and discloses a wind farm primary frequency modulation control parameter optimization method and system considering fatigue load prediction. The method takes the frequency safety constraint in the wind farm primary frequency modulation process, the fatigue load level of the wind turbine generator unit and the load distribution balance among the units into the control parameter optimization process, ensures that the minimum point of the system frequency meets the frequency constraint, and quickly obtains the equivalent fatigue load of the key components of each unit under different control parameters by pre-establishing a fatigue load prediction model. The virtual inertia control coefficient and the droop control coefficient of each unit are optimized differently by using dynamic grouping rolling optimization. The optimization method can avoid the uneven distribution of unit fatigue load caused by unified control parameters, reduce the overall fatigue load level of the wind farm, reduce the calculation overhead caused by repeated calling of high-fidelity simulation, and improve the frequency support effect and fatigue load optimization effect of the wind farm when participating in primary frequency modulation.

Description

Technical Field

[0001] This invention relates to the field of smart grid technology, and in particular to wind farm control and power system frequency support technology. Specifically, it relates to a method and system for optimizing primary frequency regulation control parameters of wind farms that takes fatigue load prediction into account. Background Technology

[0002] As the scale of new energy power generation connected to the grid continues to increase, the synchronous inertia and primary frequency regulation capability of the power system are relatively reduced, making the frequency stability problem of the system under high-power disturbances more prominent. As an important power source in a high-proportion new energy power system, wind farms need to participate in system frequency support through additional control methods to improve the system's frequency response performance after disturbances.

[0003] Wind turbines typically participate in primary frequency regulation using methods such as virtual inertia control and droop control. When the system experiences a power deficit, virtual inertia control provides additional active power support based on the rate of frequency change, while droop control provides additional active power support based on the frequency deviation. These control methods can improve the system's lowest frequency and rate of frequency change, but the control actions alter the dynamic load levels of critical components such as the wind turbine drivetrain, tower, and blades, potentially exacerbating fatigue damage.

[0004] Existing wind farm primary frequency regulation control parameters mostly adopt uniform or empirical parameter settings, typically focusing on frequency support performance, making it difficult to simultaneously consider differences in wind speed, load, and fatigue damage among different turbine units. For wind farms with multiple turbine units, directly optimizing the virtual inertia control coefficient and droop control coefficient for each unit results in a high dimensionality of optimization variables, and calculating fatigue loads through high-fidelity co-simulation for each candidate parameter incurs significant computational overhead. Therefore, there is an urgent need for a method that can quickly predict the fatigue loads corresponding to different control parameters while meeting system frequency safety constraints, and differentiate the control parameters of multiple turbine units within a wind farm, in order to reduce the overall fatigue load of the wind farm and improve the uniformity of fatigue load distribution among turbine units. Summary of the Invention

[0005] To address the shortcomings and problems of existing technologies, the present invention aims to provide a method for optimizing primary frequency regulation control parameters of wind farms that considers fatigue load prediction. This method solves the problem that existing primary frequency regulation control parameters for wind farms are unable to simultaneously ensure frequency safety, reduce overall fatigue load, and balance fatigue load among units. It reduces the overall fatigue load level of wind farms, reduces the computational overhead caused by repeated high-fidelity simulations, and thus improves the frequency support effect and fatigue load optimization effect when wind farms participate in primary frequency regulation.

[0006] According to a first aspect of the present invention, a method for optimizing primary frequency regulation control parameters of a wind farm considering fatigue load prediction is proposed, comprising:

[0007] Step 1: Using the virtual inertia control coefficient and droop control coefficient of each wind turbine as optimization variables, the total equivalent fatigue load of the wind farm and the variance of the equivalent fatigue load between turbines as optimization objectives, and setting the minimum frequency point constraint and the secondary frequency drop constraint of the system, an optimization model for the primary frequency regulation control parameters of the wind farm is established.

[0008] Step 2: For the given system power disturbance, the input wind speed of each wind turbine in the wind farm, and the preset frequency response constraints, perform a combined search within the preset range of virtual inertia control coefficient and droop control coefficient to ensure that all wind turbines in the wind farm use the same virtual inertia control coefficient and droop control coefficient as unified control parameters, and calculate the corresponding system frequency response; and when the system frequency response meets the preset frequency safety constraints, use this set of unified control parameters as the unified control benchmark scheme for subsequent differentiated optimization.

[0009] Step 3: Combine the unified control benchmark scheme or the candidate control parameters accepted in the previous round with the system power disturbance and the input wind speed of each wind turbine and input the pre-trained equivalent fatigue load prediction model to predict and output the equivalent fatigue load of each wind turbine.

[0010] Step 4: Determine the comprehensive equivalent fatigue load of the i-th wind turbine unit based on the equivalent fatigue load prediction results, and dynamically group the units in the wind farm according to the order from high to low, dividing the units in the wind farm into high equivalent fatigue load group, medium equivalent fatigue load group and low equivalent fatigue load group.

[0011] Step 5: Perform parameter adjustment pre-evaluation for each candidate group, calculate the changes in fatigue load, minimum frequency point, and load balance after adjustment, and determine the main path adjustment group based on the scoring results;

[0012] Step 6: Determine the discrete adjustment step size of the control parameters based on the current frequency safety margin and the recent candidate scheme acceptance rate, and perform a joint reduction control search of the virtual inertia control coefficient and droop control coefficient on the main path adjustment group to obtain candidate control parameter schemes;

[0013] Step 7: After each round of candidate control parameter schemes is accepted or single-variable backoff adjustment is performed to maintain the current solution, determine whether the rolling optimization has reached the termination condition. When the termination condition is reached, output the optimized virtual inertia control coefficients and droop control coefficients of each wind turbine in the wind farm, and send the virtual inertia control coefficients and droop control coefficients to the primary frequency controller of the corresponding wind turbine to perform primary frequency control.

[0014] According to a second aspect of the present invention, a computer system is provided, comprising:

[0015] One or more processors;

[0016] The memory stores operable instructions that, when executed by the one or more processors, cause the one or more processors to perform the wind farm primary frequency regulation control parameter optimization method considering fatigue load prediction as described in the foregoing embodiments.

[0017] The wind farm primary frequency regulation control parameter optimization method considering fatigue load prediction in the above embodiments of the present invention incorporates frequency safety constraints, wind turbine fatigue load levels, and load distribution balance among turbines into the control parameter optimization process. While ensuring that the lowest system frequency does not fall below a preset threshold, it rapidly obtains the equivalent fatigue loads of key components of each turbine under different control parameters using a pre-established fatigue load prediction model. Furthermore, a dynamic grouping rolling optimization method is employed to differentiate the virtual inertia control coefficients and droop control coefficients of each turbine within the wind farm. Therefore, the optimization method of the present invention avoids the problem of uneven fatigue load distribution caused by uniform control parameters, reduces the overall fatigue load level of the wind farm, and reduces the computational overhead caused by repeated high-fidelity simulations, thereby improving the frequency support effect and fatigue load optimization effect when the wind farm participates in primary frequency regulation.

[0018] It should be understood that all combinations of the foregoing concepts and the additional concepts described in more detail below may be considered part of the inventive subject matter of this disclosure, provided that such concepts do not contradict each other. Furthermore, all combinations of the claimed subject matter are considered part of the inventive subject matter of this disclosure.

[0019] The foregoing and other aspects, embodiments, and features of the teachings of the present invention will be more fully understood from the following description in conjunction with the accompanying drawings. Other additional aspects of the invention, such as features and / or beneficial effects of exemplary embodiments, will become apparent from the following description or may be learned through practice of specific embodiments according to the teachings of the present invention. Attached Figure Description

[0020] Figure 1 This is a flowchart of a method for optimizing primary frequency regulation control parameters of a wind farm, taking into account fatigue load prediction, according to an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the structure of the MLP equivalent fatigue load prediction model according to an embodiment of the present invention.

[0022] Figure 3 This is a schematic diagram of the two-stage rolling optimization process for control parameters according to an embodiment of the present invention.

[0023] Figure 4 This is a measured wind speed distribution map of the LY wind farm according to an embodiment of the present invention.

[0024] Figure 5 This is a diagram illustrating the optimization process under disturbance scenarios in the LY wind farm according to an embodiment of the present invention.

[0025] Figure 6 This is a schematic diagram comparing the comprehensive equivalent fatigue load of each unit before and after optimization under a disturbance scenario in the LY wind farm according to an embodiment of the present invention.

[0026] Figure 7 This is a distribution diagram of the virtual inertia control coefficients and droop control coefficients of each unit under the disturbance scenario in the LY wind farm according to an embodiment of the present invention.

[0027] Figure 8 This is a schematic diagram comparing the total power of the LY wind farm before and after optimization under a disturbance scenario, according to an embodiment of the present invention.

[0028] Figure 9 This is a schematic diagram showing the power comparison curves of representative units (units with the largest DEL change) before and after optimization in the disturbance scenario of the LY wind farm according to an embodiment of the present invention.

[0029] Figure 10 This is a schematic diagram comparing the system frequency response curves before and after optimization under disturbance scenarios in the LY wind farm according to an embodiment of the present invention.

[0030] Figure 11 This is a schematic diagram comparing the RoCoF curves of the system before and after optimization under a disturbance scenario in the LY wind farm according to an embodiment of the present invention.

[0031] Figure 12 This is a schematic diagram of the training process of the equivalent fatigue load prediction model based on MLP according to an embodiment of the present invention. Detailed Implementation

[0032] To better understand the technical content of the present invention, specific embodiments are described below in conjunction with the accompanying drawings.

[0033] Various aspects of the invention are described in this disclosure with reference to the accompanying drawings, which illustrate numerous illustrative embodiments. The embodiments of this disclosure are not necessarily intended to encompass all aspects of the invention. It should be understood that the various concepts and embodiments described above, as well as those described in more detail below, can be implemented in any of many ways, because the concepts and embodiments disclosed herein are not limited to any particular implementation. Furthermore, some aspects of the invention disclosed may be used alone or in any suitable combination with other aspects of the invention disclosed.

[0034] Example 1

[0035] In conjunction with embodiments of the present invention, addressing the problem in existing wind farm control and power system frequency support technologies that it is difficult to simultaneously consider frequency safety, overall fatigue load reduction, and fatigue load balance among units in the primary frequency regulation control parameters of wind farms, a wind farm primary frequency regulation control parameter optimization method considering fatigue load prediction is proposed. Based on OpenFAST-MATLAB / Simulink co-simulation, load data of key components under different operating conditions and primary frequency regulation control parameters are obtained, and an equivalent fatigue load prediction model (MLP) is established. A rapid evaluation interface for the equivalent fatigue load of wind farms is constructed, which can quickly predict the equivalent fatigue load of key components of each unit based on power disturbance, wind speed, virtual inertia control coefficient, and droop control coefficient.

[0036] Based on this, a control parameter optimization model is established with the total equivalent fatigue load of the wind farm and the variance of the equivalent fatigue load between units as optimization objectives and system frequency safety as constraint. Then, a unified control reference state that satisfies the frequency constraint and is close to the constraint boundary is searched. Subsequently, dynamic grouping is performed according to the comprehensive equivalent fatigue load of each unit, and the fatigue benefit, balance benefit, and frequency cost score of the candidate group are calculated respectively. Then, the virtual inertia control coefficient and droop control coefficient of the main path adjustment group with the highest score are adjusted simultaneously. When the joint adjustment is unacceptable, a single variable backoff is performed to adjust only the virtual inertia control coefficient or only the droop control coefficient. Finally, after the termination condition is met, the differentiated control parameters of each unit are output. Compared with existing technologies, the present invention has the following advantages: The present invention integrates frequency safety constraints, wind turbine fatigue load levels, and load distribution balance among units into the control parameter optimization process during the primary frequency regulation of wind farms. On the basis of ensuring that the lowest point of the system frequency is not lower than a preset threshold, the equivalent fatigue load of key components of each unit under different control parameters is quickly obtained through a pre-established fatigue load prediction model. Furthermore, a dynamic grouping rolling optimization method is used to differentiate the virtual inertia control coefficient and droop control coefficient of each unit in the wind farm, thereby avoiding the problem of uneven fatigue load distribution of units caused by uniform control parameters, reducing the overall fatigue load level of the wind farm, reducing the computational overhead caused by repeated calls to high-fidelity simulation, and thus improving the frequency support effect and fatigue load optimization effect when the wind farm participates in primary frequency regulation.

[0037] Combination Figure 1 The flowchart shown illustrates a method for optimizing primary frequency control parameters of a wind farm considering fatigue load prediction, according to an example of the present invention, which includes the following steps:

[0038] Step 1: Using the virtual inertia control coefficient and droop control coefficient of each wind turbine as optimization variables, the total equivalent fatigue load of the wind farm and the variance of the equivalent fatigue load between turbines as optimization objectives, and setting the minimum frequency point constraint and the secondary frequency drop constraint of the system, an optimization model for the primary frequency regulation control parameters of the wind farm is established.

[0039] Step 2: For the given system power disturbance, the input wind speed of each wind turbine in the wind farm, and the preset frequency response constraints, perform a combined search within the preset range of virtual inertia control coefficient and droop control coefficient to ensure that all wind turbines in the wind farm use the same virtual inertia control coefficient and droop control coefficient as unified control parameters, and calculate the corresponding system frequency response; and when the system frequency response meets the preset frequency safety constraints, use this set of unified control parameters as the unified control benchmark scheme for subsequent differentiated optimization.

[0040] Step 3: Combine the unified control benchmark scheme or the candidate control parameters accepted in the previous round with the system power disturbance and the input wind speed of each wind turbine and input the pre-trained equivalent fatigue load prediction model to predict and output the equivalent fatigue load of each wind turbine.

[0041] Step 4: Determine the comprehensive equivalent fatigue load of the i-th wind turbine unit based on the equivalent fatigue load prediction results, and dynamically group the units in the wind farm according to the order from high to low, dividing the units in the wind farm into high equivalent fatigue load group, medium equivalent fatigue load group and low equivalent fatigue load group.

[0042] Step 5: Perform parameter adjustment pre-evaluation for each candidate group, calculate the changes in fatigue load, minimum frequency point, and load balance after adjustment, and determine the main path adjustment group based on the scoring results;

[0043] Step 6: Determine the discrete adjustment step size of the control parameters based on the current frequency safety margin and the recent candidate scheme acceptance rate, and perform a joint reduction control search of the virtual inertia control coefficient and droop control coefficient on the main path adjustment group to obtain candidate control parameter schemes;

[0044] Step 7: After each round of candidate control parameter schemes is accepted or single-variable backoff adjustment is performed to maintain the current solution, determine whether the rolling optimization has reached the termination condition. When the termination condition is reached, output the optimized virtual inertia control coefficients and droop control coefficients of each wind turbine in the wind farm, and send the output virtual inertia control coefficients and droop control coefficients to the primary frequency controller of the corresponding wind turbine to perform primary frequency control.

[0045] As an optional implementation, in step 1, for the i-th wind turbine in the wind farm, its comprehensive equivalent fatigue load DEL i Determined according to the following formula:

[0046] DEL i =DEL TwrBsMx,i +DEL TwrBsMy,i +DEL RotTorq,i ;

[0047] Among them, DEL TwrBsMx,i DEL TwrBsMy,i and DEL RotTorq,i These represent the equivalent fatigue loads corresponding to the pitching moment, tilting moment, and transmission shaft torque of the i-th unit, respectively.

[0048] Construct the normalized synthesis objective function:

[0049]

[0050] In the formula, F represents the comprehensive objective function, DEL sum0 To uniformly control the total equivalent fatigue load of a wind farm under the reference state, DEL var0 To uniformly control the variance of equivalent fatigue loads between units under the baseline condition, α and β are the weighting coefficients corresponding to the total equivalent fatigue load of the wind farm and the variance of equivalent fatigue loads between units, respectively; DEL sum With DEL var Let V represent the total equivalent fatigue load of the wind farm and the variance of the equivalent fatigue load between units, respectively, and N represent the total number of wind turbine units in the wind farm; This represents the average value of the comprehensive equivalent fatigue load of all units in the wind farm;

[0051] The system's minimum frequency constraint and secondary frequency drop constraint are set as follows:

[0052] ;

[0053] In the formula, f min with f limit These represent the lowest point of the system frequency and the minimum limit of frequency drop, respectively, while RoCoF represents the rate of frequency change.

[0054] It should be understood that in this step, for each set of candidate control parameters, the DEL of key components of each unit is first obtained quickly through the equivalent fatigue load model pre-trained in step 1 (as a fast evaluation interface), and then the comprehensive DEL of each unit, the total DEL of the wind farm, and the variance of DEL between units are calculated.

[0055] As an optional implementation, in step 2, for a wind farm comprising N wind turbines, the optimization variable K is the virtual inertia control coefficient and droop control coefficient of each wind turbine, K=[K df,1 ,K pf,1 ,K df,2 ,K pf,2 ,…,K df,N ,K pf,N ], K df,N With K pf,N These represent the virtual inertia control coefficient and droop control coefficient of the Nth wind turbine, respectively.

[0056] Given the system power disturbance and the input wind speed of each wind turbine in the wind farm, N wind turbines are made to use a set of virtual inertia control coefficients and droop control coefficients as unified control parameters. The combination search is performed within the preset value range of the virtual inertia control coefficients and droop control coefficients, and the corresponding system frequency response is calculated respectively.

[0057] Based on the given system power disturbance scenario and the corresponding frequency response constraints, select the system frequency response that satisfies the frequency response constraints and whose lowest frequency point is closest to the minimum system frequency drop limit (f). limit A set of unified control parameters is used as a unified control benchmark scheme, which serves as the normalized benchmark for subsequent rolling optimization.

[0058] As an optional implementation, in step 3, the pre-trained equivalent fatigue load prediction model adopts an MLP neural network, and uses the wind speed, power disturbance, virtual inertia control coefficient, and droop control coefficient of each wind turbine in the wind farm as input X. i :

[0059] X i =[ΔP, v i K df,i K pf,i ];

[0060] The equivalent fatigue loads corresponding to the pitching moment, tilting moment, and transmission shaft torque of the i-th unit tower are used as the model output variable Y. i :

[0061] Y i =[DEL TwrBsMx,i ,DEL TwrBsMy,i ,DEL RotTorq,i ];

[0062] Where ΔP is the system power disturbance, v i Let K be the wind speed of the i-th wind turbine. df,i K represents the virtual inertia control coefficient for the i-th wind turbine. pf,i DEL is the droop control coefficient for the i-th wind turbine. TwrBsMx,i DEL TwrBsMy,i and DEL RotTorq,i These represent the equivalent fatigue loads corresponding to the pitching moment, tilting moment, and transmission shaft torque of the i-th unit, respectively.

[0063] In the embodiments of the present invention, load data of key components under different operating conditions and primary frequency regulation control parameters are obtained based on OpenFAST-MATLAB / Simulink co-simulation. Specifically, load data of key components of wind turbines are obtained under different wind speeds, power disturbances, virtual inertia control coefficients, and droop control coefficients, and the corresponding equivalent fatigue loads are calculated to form model training sample data. The training samples are used to train an MLP neural network model to establish a mapping relationship between operating conditions and primary frequency regulation control parameters and equivalent fatigue loads of key components. The trained equivalent fatigue load prediction model is used as a fast evaluation interface for equivalent fatigue loads. Based on the given system power disturbances, wind speeds, virtual inertia control coefficients, and droop control coefficients, the equivalent fatigue load (DEL) of key components of each unit can be quickly predicted.

[0064] To more clearly illustrate the acquisition and application of the MLP model in the embodiments of the present invention, we will... Figure 1 After describing the complete implementation process of the illustrated procedure, a more detailed description is given of the training process of the equivalent fatigue load prediction model based on MLP neural network.

[0065] It should be understood that during the dynamic grouping process in step 4 based on the comprehensive equivalent fatigue load of each unit, the dynamic grouping is not fixed. During the rolling optimization process, the comprehensive equivalent fatigue load of each wind turbine unit is recalculated according to the set iteration interval, and the grouping is reordered.

[0066] Furthermore, in step 5, parameter adjustment pre-evaluation is performed on the candidate groups obtained by sorting, the changes in fatigue load, the changes in the lowest frequency point, and the changes in load balance after adjustment are calculated, and the main path adjustment group is determined based on the scoring results.

[0067] As an optional implementation, in step 5, for any candidate group G g While keeping the other group control parameters unchanged, a one-step trial reduction control is implemented for the candidate group, and the total equivalent fatigue load DEL of the wind farm after the trial adjustment is calculated. sum,g The lowest frequency point f min,g and equivalent fatigue load variance DEL var,g ;

[0068] Then calculate the candidate group G according to the following formula. g The fatigue benefits, frequency costs, and load balance changes:

[0069] ;

[0070] Wherein, ΔDEL g , Δf g r g Representing candidate groups G respectivelyg Fatigue benefits, frequency costs, and load balance changes; DEL sum,base f min,base and DEL var,base These represent the total equivalent fatigue load, the lowest frequency point, and the variance of the equivalent fatigue load at the current rolling iteration starting point of the wind farm.

[0071] Candidate group G is determined by combining fatigue benefits, frequency costs, and load balance changes. g Group score g :

[0072] ;

[0073] Where ε is a very small positive number to prevent the denominator from being zero, and λ is the load balance reward / penalty weight.

[0074] It should be understood that each candidate group has a group score. g A higher score indicates that the candidate group has greater adjustment value in the current state.

[0075] As an optional implementation, in step 6, the aforementioned frequency safety margin is the difference between the lowest frequency point of the current scheme and the lowest frequency point constraint value. The recent candidate scheme acceptance rate is the proportion of schemes accepted in the most recent rounds of candidate schemes.

[0076] The discrete adjustment step size is adjusted based on the discrete quantization interval of the control parameters. When the frequency security margin is less than a preset security threshold or the recent candidate scheme acceptance rate is lower than a preset acceptance rate threshold, the discrete adjustment step size is reduced.

[0077] As an optional implementation, in step 6, if the candidate control parameter scheme after the joint reduction control search meets the frequency response constraint condition, and the total equivalent fatigue load of the wind farm is reduced or the comprehensive objective function is improved, then the control parameter scheme is accepted and the control parameters are updated; if not, the joint reduction control search is performed again after reducing the adjustment step size.

[0078] If the joint decreasing control search of the main path still fails to yield an acceptable solution after the step size is reduced to the minimum, then a univariate backoff adjustment is performed, including two methods:

[0079] Only reduce the virtual inertia control coefficient of the units within the main path adjustment group, while keeping the droop control coefficient unchanged;

[0080] Only reduce the droop control coefficient of the units within the main path adjustment group, while keeping the virtual inertia control coefficient unchanged.

[0081] Combination Figure 2As shown, after performing the single-variable rollback adjustment, the rollback adjustment result is accepted according to the aforementioned criteria, namely: whether the control parameter scheme after the single-variable rollback adjustment meets the frequency response constraint condition, and whether the total equivalent fatigue load of the wind farm is reduced or the comprehensive objective function is improved. If so, the control parameter scheme of the rollback adjustment is accepted and the control parameters are updated.

[0082] If the adjusted control parameter scheme does not meet the frequency response constraint or the scheme is unacceptable (the total equivalent fatigue load of the wind farm has not been reduced or the comprehensive objective function has not been improved), then the current solution (the current control parameter scheme) is maintained, and it is determined whether the rolling iteration stopping condition has been met.

[0083] As an optional implementation, in step 7, the following multiple conditions are used as criteria to terminate the rolling optimization iterative process, including:

[0084] (1) Reach the maximum number of iterations for rolling optimization;

[0085] (2) After the step size is reduced to the minimum value, there are no acceptable updates for three consecutive iterations;

[0086] (3) The variance of equivalent fatigue load between units is reduced to less than 5% of the initial value;

[0087] (4) The overall objective function has entered a plateau range and the improvement has been limited for several consecutive rounds;

[0088] When any of the conditions in (1)-(4) above are met, the rolling optimization is terminated, and the optimized virtual inertia control coefficient and droop control coefficient of each wind turbine in the wind farm are output.

[0089] Combination Figure 2 , Figure 12 As shown, the MLP neural network structure used in this invention is a multilayer perceptron regression structure with 4 inputs and 3 outputs. The input layer contains 4 input variables: power disturbance, wind speed, virtual inertia control coefficient, and droop control coefficient; the output layer contains 3 neurons, which output the DEL of the three key components respectively.

[0090] Combination Figure 2 For example, the hidden layers use a decreasing structure (100, 50, 25), meaning the first hidden layer has 100 neurons, the second hidden layer has 50 neurons, and the third hidden layer has 25 neurons. The hidden layers use the ReLU activation function to characterize the non-linear relationship between the input variables and DEL.

[0091] After inputting four variables, the MLP prediction process involves three consecutive feature transformations. First, the four input variables are standardized to bring the different dimensions and numerical ranges of the inputs to a scale suitable for network computation. Then, the standardized inputs enter the first hidden layer. The first hidden layer combines the four input features using weights and biases, and then performs a ReLU nonlinear transformation to extract the fundamental features between the input variables and the fatigue load. The second hidden layer further combines and compresses the features from the previous layer. The third hidden layer continues to extract higher-level nonlinear features. After feature transformations through these three hidden layers, the results finally enter the three output nodes, yielding the predicted values ​​for the three key components. Finally, the prediction results are destandardized to obtain the three DEL prediction results with actual physical dimensions.

[0092] The MLP neural network prediction model proposed in this invention is intended for building a rapid evaluation interface for equivalent fatigue loads, and is not the main focus or object of primary frequency regulation control parameter optimization. During pre-training, the MLP neural network prediction model utilizes sample data generated by OpenFAST-MATLAB / Simulink co-simulation to learn the nonlinear mapping relationship between wind turbine operating conditions, primary frequency regulation control parameters, and the equivalent fatigue load (DEL) of key components. After model training is complete, in subsequent control parameter optimization, it is unnecessary to repeatedly perform high-fidelity co-simulation and fatigue load post-processing for each set of candidate parameters; the DEL of each unit can be quickly obtained.

[0093] As an optional implementation method, combined with Figure 12 The MLP model training process shown is implemented using the OpenFAST-MATLAB / Simulink co-simulation platform for acquiring training samples. OpenFAST is used to describe the aerodynamic, structural, and transmission chain dynamic responses of the wind turbine, while MATLAB / Simulink is used to implement the primary frequency control system. Co-simulations are performed by changing the wind speed of the wind turbine, the system power disturbance, the virtual inertia control coefficient, and the droop control coefficient.

[0094] Each simulation yields load time-series data for key components of the tower and drivetrain. Fatigue load calculations are performed on the load time-series data to obtain the tower tilt moment (DEL), tower pitch moment (DEL), and drivetrain torque (DEL) under the corresponding working conditions. Thus, a set of supervised learning samples with "input parameters - DEL output" can be generated through a single simulation, as shown in Table 1 below as an example of input variable parameters.

[0095] Table 1 - Definition of input variables for MLP model training

[0096]

[0097] Based on the above examples of input parameter combinations, this embodiment generates 12,100 sets of frequency modulation simulation scenarios, each with a simulation duration of 200 seconds, to cover the frequency modulation transient response and recovery process. For each scenario, time-series response data of the corresponding load channel is collected, and rainflow counting and DEL calculation are performed to ultimately form a sample dataset with a one-to-one correspondence between input variables and target DELs. Simultaneously, 1,215 sets of independent external samples are constructed for model generalization performance verification. These external test samples do not participate in any form of parameter update during model training, hyperparameter search, and model selection; their role is solely for the final generalization ability evaluation. Furthermore, while the external test set maintains consistency with the training samples on the input parameter boundaries, the specific parameter combinations are not repeated, thus enabling a more objective examination of the model's predictive performance under unseen operating conditions.

[0098] Combination Figure 12 As shown, during model training, a hidden layer structure of (100, 50, 25), ReLU activation function, learning rate of 0.001, and regularization coefficient of 0.001 were used based on the previously determined parameter configuration. After model training was completed, the model's predictive ability was tested using data not used in training. Table 2 below shows examples of the training process parameters.

[0099] As an example, the specific steps of the training process are as follows:

[0100] The first step is to partition the data: the data is divided into training and test sets in an 80 / 20 ratio to ensure that the test data is not used during model training, thereby avoiding data leakage;

[0101] The second step is to perform model optimization and dynamic verification.

[0102] To avoid the adverse effects of significant differences in the magnitude of input and output data on the model training process, all input and output variables were standardized before training. Mean Squared Error (MSE) was used as the loss function during training to measure the difference between the model's predictions and the actual observations.

[0103] During training, the Adam optimizer is used to iteratively optimize the model parameters. This optimizer has the characteristic of adaptive learning rate, which helps to improve convergence efficiency and stability. Simultaneously, to improve the stability and convergence speed of the training process, an early stopping mechanism is introduced to dynamically monitor the MSE (Maximum Sequence Error) trend of the model on the test set. When the difference between the maximum and minimum values ​​of the MSE on the test set within 100 consecutive iterations is less than a preset threshold, the model is considered to have converged, the training process automatically terminates, and the model weights are restored to the state corresponding to the minimum test error. If this condition is not met, training continues until the preset maximum number of iterations is reached.

[0104] The third step involves global reconstruction and generalization testing. After locking in the optimal model hyperparameter configuration in the second stage, the MLP neural network is globally retrained using the complete initial dataset to fully absorb the distribution characteristics of the global data, constructing the final model version for the actual prediction task. Subsequently, an independent external dataset containing 1,215 samples is introduced to blind test the final model, evaluating and ensuring that the model still possesses excellent robustness and stable generalization prediction performance under unseen conditions.

[0105] Table 2 - Parameters for MLP Model Training Process

[0106]

[0107] After training, the MLP model no longer needs OpenFAST to participate in every prediction. When the subsequent optimization algorithm provides a new set of candidate control parameters, for the i-th unit, it is only necessary to input the current power disturbance, wind speed, and new control parameters into the trained MLP model for prediction. After standardization and hierarchical calculation by the hidden layer, the DEL of the three key components of the unit is output.

[0108] For example, for 48 wind turbine units, 48 ​​sets of variable features are input and predictions are made to quickly obtain the unit-level DEL results for the entire wind farm. Based on this, further differentiated optimization of the virtual inertia control coefficient and droop control coefficient of each unit in the wind farm is carried out through dynamic grouping, candidate group scoring, adaptive step size joint adjustment, and univariate backoff.

[0109] Example 2

[0110] In this embodiment, taking the LY wind farm, consisting of 48 wind turbine units, participating in primary frequency regulation as an example, the differentiated optimization process of the aforementioned wind farm primary frequency regulation control parameter optimization method considering fatigue load prediction is further elaborated.

[0111] The LY wind farm is connected to the IEEE 9-bus system, and the actual wind speed of each turbine is as follows: Figure 4 Therefore, the steady-state output power of the wind farm is 138.4 MW. The system power disturbance is assumed to be a moderate disturbance scenario, i.e., the power disturbance is ΔP. d =35 / 692 pu. Under this moderate disturbance scenario, when the wind farm does not participate in primary frequency regulation, the lowest system frequency is 49.65 Hz. Therefore, the constraint on the lowest system frequency is set to f. min ≥49.80 Hz. The following is based on... Figure 1 and Figure 3 The flowchart of the method of the present invention shown below illustrates the specific implementation of the present invention.

[0112] In this example, the operating conditions of a wind farm are constructed according to the aforementioned method, and a rapid evaluation interface for equivalent fatigue loads is built. The wind speed, power disturbance, virtual inertia control coefficient, and droop control coefficient variables of each wind turbine in the wind farm are used as inputs to the equivalent fatigue load prediction model, thereby quickly evaluating and outputting the equivalent fatigue load (DEL) of key components of each turbine.

[0113] In this example, the equivalent fatigue load prediction model employs a multilayer perceptron (MLP) neural network model. Implemented on the OpenFAST-MATLAB / Simulink co-simulation platform, load data for key components of the wind turbine are obtained under different wind speeds, power disturbances, virtual inertia control coefficients, and droop control coefficients. The corresponding equivalent fatigue loads are then calculated to form model training samples. These training samples are used to establish a mapping relationship between operating conditions, primary frequency control parameters, and the equivalent fatigue loads of key components. This allows the trained MLP model to serve as a rapid evaluation interface for the equivalent fatigue loads in subsequent algorithmic processing.

[0114] The equivalent fatigue load prediction model adopts a multilayer perceptron neural network model, using the wind speed, power disturbance, virtual inertia control coefficient, and droop control coefficient of each wind turbine in the wind farm as input X. i :

[0115] X i = [ΔP, v i K df,i K pf,i ];i=1,2,…,48.

[0116] Where ΔP is the system power disturbance, v i Let K be the wind speed of the i-th wind turbine. df,i K represents the virtual inertia control coefficient for the i-th wind turbine. pf,i Let be the droop control coefficient for the i-th wind turbine.

[0117] The equivalent fatigue load prediction model uses the equivalent fatigue loads corresponding to the pitching moment, tilting moment, and transmission shaft torque of the i-th unit as the model output variable Y. i :

[0118] Y i = [DEL TwrBsMx, i , DEL TwrBsMy, i , DEL RotTorq, i ];

[0119] Among them, DEL TwrBsMx, i DEL TwrBsMy, i and DEL RotTorq, iThese represent the equivalent fatigue loads (DEL) corresponding to the pitching moment, tilting moment, and transmission shaft torque of the i-th unit, respectively.

[0120] In this embodiment, wind speed data and power disturbance ΔP of 48 wind turbine units are obtained. d Virtual inertia control coefficient K df and droop control coefficient K pf The power disturbance is 35 / 692, and the steady-state output power of the wind farm is 138.4 MW. The input variable for each unit is represented as X. i =[ΔP d ,v i ,K df,i ,K pf,i ], where i = 1, 2, ..., 48. Input the above input variables into Figure 2 The MLP neural network equivalent fatigue load prediction model shown yields the tower tilt moment (DEL), tower pitch moment (DEL), and drive shaft torque (DEL) for each unit. The purpose of this step is to replace the OpenFAST-Simulink high-fidelity co-simulation and rainflow counting post-processing under each set of candidate control parameters with a fast prediction model, thereby enabling subsequent rolling optimization to quickly evaluate fatigue costs.

[0121] For the i-th wind turbine, its input variable is represented as X. i =[ΔP d ,v i ,K df,i ,K pf,i ], where i=1,2,…,48. By inputting the above input variables into the trained MLP model, the tower tilt moment DEL, tower pitch moment DEL, and drive shaft torque DEL corresponding to the i-th unit can be obtained.

[0122] In the subsequent rolling optimization process, for each group of candidate control parameters, the DEL of each unit is obtained directly through the equivalent fatigue load fast evaluation interface, thereby avoiding repeated OpenFAST-MATLAB / Simulink high-fidelity co-simulation and fatigue load post-processing, and providing fast fatigue load evaluation results for subsequent dynamic grouping, candidate group scoring and control parameter adjustment.

[0123] Step 1: Establish an optimization model for the primary frequency regulation control parameters of the wind farm.

[0124] The virtual inertia control coefficient and droop control coefficient of each wind turbine are used as optimization variables, and the total equivalent fatigue load of the wind farm and the variance of the equivalent fatigue load between turbines are used as optimization objectives. The system frequency minimum point constraint, secondary frequency drop constraint and control parameter boundary constraint are set.

[0125] In this embodiment, the optimization variables are the virtual inertia control coefficient and droop control coefficient of the 48 generator units, i.e., K=[K df,1 ,K pf,1 ,K df,2 ,K pf,2 ,…,K df,48 ,K pf,48 For each set of candidate control parameters, the DEL of key components of each unit is first obtained through the trained equivalent fatigue load fast evaluation interface, the comprehensive DEL of each unit is calculated, and then the total DEL of the wind farm and the variance of DEL between units are calculated.

[0126] In this embodiment, α=1 and β=0.5 are chosen to reflect the optimization principle of "total DEL priority and balance assistance".

[0127] Furthermore, in the frequency constraint condition, the minimum frequency drop value f limit Taking 49.80 Hz as an example, the frequency constraint is: the lowest point of the system frequency f min ≥49.80 Hz. Candidate control parameters must first meet this hard constraint. If a candidate scheme does not meet the frequency constraint, it will not be included in the fatigue index comparison.

[0128] Simultaneously, during the optimization of control parameters, it is necessary to ensure that they satisfy the engineering boundary constraints, i.e., K. df,i and K pf,i All must not exceed the set upper and lower limits.

[0129] Step 2: Search for a unified control reference state that satisfies the frequency constraints.

[0130] Given the system power disturbance, the input wind speed of each wind turbine in the wind farm, and the preset frequency response constraints, a combined search is performed within the preset range of virtual inertia control coefficients and droop control coefficients to ensure that all wind turbines in the wind farm use the same virtual inertia control coefficients and droop control coefficients, and the corresponding system frequency response is calculated. When the system frequency response meets the preset frequency safety requirements, this set of unified control parameters is used as the benchmark scheme for subsequent differentiated optimization.

[0131] In this embodiment, a unified control method is first adopted, meaning that all 48 wind turbines use the same virtual inertia control coefficient and the same droop control coefficient. Within the unified control parameter range, a parameter combination that satisfies the minimum frequency constraint is searched. Taking a moderate disturbance scenario as an example, the unified control parameter combination that satisfies the frequency constraint and whose minimum frequency is closest to 49.80 Hz is selected as the baseline state. The unified control baseline state is K. df =10, K pf =30. Under this unified control reference state, the lowest system frequency is 49.8025 Hz, satisfying f minConstraint of ≥49.80 Hz. Simultaneously, the DEL of key components of each unit (DEL under unified control baseline) is obtained through the aforementioned equivalent fatigue load rapid assessment interface, and the total DEL of the wind farm, the variance of DEL between units, and the comprehensive objective function value are calculated as a normalized baseline for subsequent differentiated optimization.

[0132] Step 3: Dynamically group the units according to their comprehensive equivalent fatigue load.

[0133] The unified control benchmark scheme or the candidate control parameters accepted in the previous round are combined with the system power disturbance and the input wind speed of each wind turbine and input to the pre-established equivalent fatigue load prediction model to obtain the equivalent fatigue load of each wind turbine; the turbines in the wind farm are dynamically grouped according to the equivalent fatigue load prediction results.

[0134] Under a unified control baseline, the aforementioned equivalent fatigue load rapid assessment interface is invoked to calculate the DEL of key components of 48 wind turbine units, and further obtain the comprehensive DEL of each unit.

[0135] Furthermore, the 48 generating units are sorted from highest to lowest by their overall DEL (Digital Efficiency Level), and divided into three groups of 16 units each. In this embodiment, the DEL prediction results under the current control parameters are reread in each round or every few rounds of iteration, and the high DEL group, medium DEL group, and low DEL group are updated accordingly, providing a basis for subsequent candidate group scoring.

[0136] Combination Figure 3 As shown, an exemplary two-stage rolling optimization process is illustrated, wherein the second stage includes dynamic grouping by DEL and dynamic division into high / medium / low DEL groups.

[0137] Step 4: Calculate the scores for each candidate group and determine the main path adjustment group.

[0138] For each candidate group, a parameter adjustment pre-evaluation is performed. The equivalent fatigue load of each unit after trial adjustment is calculated through the aforementioned equivalent fatigue load rapid evaluation interface. Furthermore, the changes in fatigue load, the changes in the lowest frequency point, and the changes in load balance are obtained. Based on the scoring results, the main path adjustment group is determined.

[0139] Step 5: Adjust the virtual inertia control coefficient and droop control coefficient together based on the adaptive step size.

[0140] The discrete adjustment step size of the control parameters is determined based on the current frequency safety margin and the acceptance rate of recent candidate schemes. A joint reduction control search of virtual inertia control coefficients and droop control coefficients is performed on the main path adjustment group to obtain candidate control parameter schemes.

[0141] As an optional approach, the virtual inertia control coefficient and droop control coefficient of the units within the main path adjustment group are simultaneously reduced to obtain candidate control parameter schemes. After the joint adjustment is completed, the system frequency response is calculated based on the system dynamic model, and the wind farm fatigue index is calculated through the aforementioned equivalent fatigue load rapid evaluation interface. If the adjusted minimum frequency still satisfies f min If the Hz is ≥49.80 Hz and the total DEL of the wind farm decreases or the comprehensive objective function improves, then the candidate scheme is accepted and the current control parameters are updated; otherwise, the step size is reduced and the joint adjustment is retried.

[0142] If the joint adjustment of the main path adjustment group in step 5 still fails to yield an acceptable solution after the step size is reduced to the minimum, then a single-variable rollback adjustment is performed. As mentioned above, only the virtual inertia control coefficient of the units within the main path adjustment group is reduced, while the droop control coefficient remains unchanged; only the droop control coefficient of the units within the main path adjustment group is reduced, while the virtual inertia control coefficient remains unchanged.

[0143] Combination Figure 2 As shown, after performing the single-variable rollback adjustment, the rollback adjustment result is accepted according to the aforementioned criteria, namely: whether the control parameter scheme after the single-variable rollback adjustment meets the frequency response constraint condition, and whether the total equivalent fatigue load of the wind farm is reduced or the comprehensive objective function is improved. If so, the control parameter scheme of the rollback adjustment is accepted and the control parameters are updated.

[0144] If the adjusted control parameter scheme does not meet the frequency response constraint or the scheme is unacceptable (the total equivalent fatigue load of the wind farm has not been reduced or the comprehensive objective function has not been improved), then the current solution (the current control parameter scheme) is maintained, and it is determined whether the rolling iteration stopping condition has been met.

[0145] Step 6: Determine the termination condition and output the optimized control parameters.

[0146] After each round of candidate solution acceptance or retention, this embodiment determines whether the rolling optimization has reached the termination condition. When the termination condition is reached, the optimized virtual inertia control coefficients and droop control coefficients of each wind turbine in the wind farm are output.

[0147] As an example, termination conditions include: reaching the maximum number of iterations; the joint adjustment step size has reached its minimum value and no effective improvement has been achieved for several consecutive rounds; the DEL variance has decreased to below a set proportion of the initial variance; or the improvement of the comprehensive objective function F in several consecutive rounds of iterations is less than a set threshold.

[0148] If the termination condition is met, output the final K corresponding to the 48 wind turbine units. df,i and K pf,i , which serves as the primary frequency regulation differential control parameter for wind farms under this moderate disturbance scenario.

[0149] Table 3 below shows a comparison of key indicators before and after optimization in the case of medium disturbance.

[0150] Table 3 - Comparison of key indicators before and after optimization in the medium-disturbance scenario

[0151]

[0152] Based on the comparison of key indicators before and after the optimization of the above control parameters, compared with the unified control benchmark scheme, the total DEL of the wind farm decreased from 81191.08 to 75893.65, a reduction of 5297.43, or 6.5%; the DEL variance decreased from 12334.82 to 583.53, a reduction of 95.27%; and the normalized comprehensive objective function decreased from 1.50 to 0.96, an improvement of 36.12%. The above comparison results show that the method proposed in this invention, while meeting frequency safety constraints, can not only effectively reduce the overall fatigue level of the wind farm, but also significantly improve the balance of fatigue load distribution among units.

[0153] In this embodiment, combined with Figures 5 to 11 This further illustrates the implementation process and optimization effect of the method of the present invention in the disturbance scenario of the LY wind farm.

[0154] Figure 5 This diagram illustrates the optimization process under disturbance scenarios in the LY wind farm. It characterizes the iterative changes of the method in this invention during the rolling optimization process, including the trends of the total equivalent fatigue load of the wind farm, the variance of the equivalent fatigue load, and the comprehensive objective function or related optimization indices as a function of iterations. Figure 5 As can be seen, the optimization index gradually improves and tends to stabilize as the rolling optimization proceeds, indicating that the dynamic grouping, candidate group scoring, main path joint adjustment and single variable backoff mechanism adopted in this invention can effectively promote the convergence of control parameters in the direction of reducing fatigue cost.

[0155] Figure 6 This figure shows the comparison curves of the comprehensive equivalent fatigue load of each wind turbine unit before and after optimization under disturbance scenarios in the LY wind farm. It is used to compare the distribution of the comprehensive equivalent fatigue load of each wind turbine unit under a unified control baseline state and an optimized control state. Figure 6 It can be seen that the load level of some units with originally high equivalent fatigue loads has been reduced after optimization, and the difference in comprehensive equivalent fatigue loads between different units has been reduced. This indicates that the method of the present invention can not only reduce the overall fatigue load level of the wind farm, but also improve the problem of uneven distribution of fatigue loads among units in the farm.

[0156] Figure 7This figure shows the distribution of virtual inertia control coefficients and droop control coefficients for each turbine in the LY wind farm under disturbance scenarios after optimization. This figure is used to illustrate the final differentiated primary frequency regulation control parameter configuration results output by the method of this invention. Figure 7 It can be seen that after optimization, the virtual inertia control coefficient and droop control coefficient corresponding to different units are no longer uniform, but are configured differently according to the wind speed conditions, fatigue load levels, and differences in contribution to frequency support of each unit. This result shows that the method of the present invention can avoid the problem of uneven fatigue load caused by using the same control parameters for all units in the wind farm, thereby achieving coordinated optimization of control parameters within the wind farm.

[0157] Figure 8 This figure shows a comparison of the total power output of the LY wind farm before and after optimization under disturbance scenarios. It illustrates whether the overall active power response of the wind farm still meets the primary frequency regulation support requirements after optimizing the control parameters. Figure 8 It can be seen that the overall trend of the total power response curve of the wind farm before and after optimization is similar, indicating that while reducing the frequency regulation control intensity of some units, the present invention does not significantly weaken the overall active power support capability of the wind farm, and can maintain the overall power response capability of the wind farm to system frequency disturbances.

[0158] Figure 9 This figure shows the power comparison curves of representative turbines before and after optimization under disturbance scenarios in the LY wind farm. The representative turbines are those with the largest change in comprehensive equivalent fatigue load during the optimization process. This figure is used to illustrate the impact of the method of this invention on the frequency regulation power response of a single representative turbine. Figure 9 It can be seen that the dynamic response of the active power of the unit changed after optimization, indicating that the unit was subject to differentiated control parameter adjustments during the rolling optimization process. This adjustment was not directly determined based on a single fatigue magnitude, but rather by a comprehensive judgment through a candidate grouping scoring mechanism, thereby reducing fatigue load while satisfying system frequency constraints.

[0159] Figure 10 This figure shows a comparison of the system frequency response before and after optimization under disturbance scenarios in the LY wind farm. This figure is used to verify whether the method of this invention still meets the system frequency safety constraints while reducing fatigue load. Figure 10 It can be seen that the system frequency response curves before and after optimization are generally similar, and the lowest point of the optimized system frequency is still not lower than the set frequency constraint value. This indicates that after the wind farm control parameter differential optimization, the present invention can still ensure the frequency safety of the system during the primary frequency regulation process.

[0160] Figure 11 This figure shows the RoCoF curves comparing the system frequency change rate before and after optimization under disturbance scenarios in the LY wind farm. This graph illustrates the difference in the dynamic rate of change of the system frequency before and after optimization. Figure 11It can be seen that the RoCoF curve of the optimized system is basically consistent with that before optimization, and there is no significant deterioration. This indicates that the method of the present invention reduces the fatigue load of the wind turbine without significantly weakening the dynamic support capability of the wind farm in the early stage of frequency change, and can take into account both the system frequency change rate constraint and the unit fatigue load suppression requirements.

[0161] In summary Figures 5 to 11 It can be seen that the rolling optimization method of control parameters proposed in this invention dynamically determines the main path adjustment group (i.e., the main path group) through the candidate group scoring mechanism, and combines the main path joint adjustment and single variable backoff. Under the premise of ensuring that the minimum frequency point and frequency change rate of the system meet the safety requirements, it can realize the differentiated optimization of the primary frequency regulation control parameters of the wind farm, reduce the overall equivalent fatigue load of the wind farm and improve the uniformity of fatigue load distribution among units.

[0162] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A method for optimizing primary frequency regulation control parameters of a wind farm considering fatigue load prediction, characterized in that, Includes the following steps: Step 1: Using the virtual inertia control coefficient and droop control coefficient of each wind turbine as optimization variables, the total equivalent fatigue load of the wind farm and the variance of the equivalent fatigue load between turbines as optimization objectives, and setting the minimum frequency point constraint and the secondary frequency drop constraint of the system, an optimization model for the primary frequency regulation control parameters of the wind farm is established. Step 2: For the given system power disturbance, the input wind speed of each wind turbine in the wind farm, and the preset frequency response constraints, perform a combined search within the preset range of virtual inertia control coefficient and droop control coefficient to ensure that all wind turbines in the wind farm use the same virtual inertia control coefficient and droop control coefficient as unified control parameters, and calculate the corresponding system frequency response; and when the system frequency response meets the preset frequency safety constraints, use this set of unified control parameters as the unified control benchmark scheme for subsequent differentiated optimization. Step 3: Combine the unified control benchmark scheme or the candidate control parameters accepted in the previous round with the system power disturbance and the input wind speed of each wind turbine and input the pre-trained equivalent fatigue load prediction model to predict and output the equivalent fatigue load of each wind turbine. Step 4: Determine the comprehensive equivalent fatigue load of the i-th wind turbine unit based on the equivalent fatigue load prediction results, and dynamically group the units in the wind farm according to the order from high to low, dividing the units in the wind farm into high equivalent fatigue load group, medium equivalent fatigue load group and low equivalent fatigue load group. Step 5: Perform parameter adjustment pre-evaluation for each candidate group, calculate the changes in fatigue load, minimum frequency point, and load balance after adjustment, and determine the main path adjustment group based on the scoring results; Step 6: Determine the discrete adjustment step size of the control parameters based on the current frequency safety margin and the recent candidate scheme acceptance rate, and perform a joint reduction control search of the virtual inertia control coefficient and droop control coefficient on the main path adjustment group to obtain candidate control parameter schemes; Step 7: After each round of candidate control parameter schemes is accepted or single-variable backoff adjustment is performed to maintain the current solution, determine whether the rolling optimization has reached the termination condition. When the termination condition is reached, output the optimized virtual inertia control coefficients and droop control coefficients of each wind turbine in the wind farm, and send the output virtual inertia control coefficients and droop control coefficients to the primary frequency controller of the corresponding wind turbine to perform primary frequency control.

2. The method for optimizing primary frequency regulation control parameters of a wind farm considering fatigue load prediction according to claim 1, characterized in that, In step 1, for the i-th wind turbine in the wind farm, its comprehensive equivalent fatigue load DEL i Determined according to the following formula: OF THE i =OF THE TwrBsMx,i +OF THE TwrBsMy,i +OF THE RotTorq,i ; Among them, DEL TwrBsMx,i DEL TwrBsMy,i and DEL RotTorq,i These represent the equivalent fatigue loads corresponding to the pitching moment, tilting moment, and transmission shaft torque of the i-th unit, respectively. Construct the normalized synthesis objective function: ； In the formula, F represents the comprehensive objective function, DEL sum0 To uniformly control the total equivalent fatigue load of a wind farm under the reference state, DEL var0 To uniformly control the variance of equivalent fatigue loads between units under the baseline condition, α and β are the weighting coefficients corresponding to the total equivalent fatigue load of the wind farm and the variance of equivalent fatigue loads between units, respectively; DEL sum With DEL var Let V represent the total equivalent fatigue load of the wind farm and the variance of the equivalent fatigue load between units, respectively, and N represent the total number of wind turbine units in the wind farm; This represents the average value of the comprehensive equivalent fatigue load of all units in the wind farm; The system's minimum frequency constraint and secondary frequency drop constraint are set as follows: ； In the formula, f min with f limit These represent the lowest point of the system frequency and the minimum limit of frequency drop, respectively, while RoCoF represents the rate of frequency change.

3. The method for optimizing primary frequency regulation control parameters of a wind farm considering fatigue load prediction according to claim 1, characterized in that, In step 2, for a wind farm comprising N wind turbines, the optimization variable K is the virtual inertia control coefficient and droop control coefficient for each wind turbine, K=[K df,1 ,K pf,1 ,K df,2 ,K pf,2 ,…,K df,N ,K pf,N ], K df,N With K pf,N These represent the virtual inertia control coefficient and droop control coefficient of the Nth wind turbine, respectively. Given the system power disturbance and the input wind speed of each wind turbine in the wind farm, N wind turbines are made to use a set of virtual inertia control coefficients and droop control coefficients as unified control parameters. The combination search is performed within the preset value range of the virtual inertia control coefficients and droop control coefficients, and the corresponding system frequency response is calculated respectively. Based on the frequency response constraints under the given system power disturbance scenario, a set of unified control parameters that satisfy the frequency response constraints and whose lowest frequency point is closest to the minimum frequency drop limit is selected as the unified control benchmark scheme, which serves as the normalized benchmark for subsequent rolling optimization.

4. The method for optimizing primary frequency regulation control parameters of a wind farm considering fatigue load prediction according to claim 1, characterized in that, In step 3, the pre-trained equivalent fatigue load prediction model employs an MLP neural network, using the wind speed, power disturbance, virtual inertia control coefficient, and droop control coefficient of each wind turbine in the wind farm as input X. i : X i =[ΔP,v i ,K df,i ,K pf,i ]; The equivalent fatigue loads corresponding to the pitching moment, tilting moment, and transmission shaft torque of the i-th unit tower are used as the model output variable Y. i : AND i =[OF THE TwrBsMx,i ,OF THE TwrBsMy,i ,OF THE RotTorq,i ]; Where ΔP is the system power disturbance, v i Let K be the wind speed of the i-th wind turbine. df,i K represents the virtual inertia control coefficient for the i-th wind turbine. pf,i DEL is the droop control coefficient for the i-th wind turbine. TwrBsMx,i DEL TwrBsMy,i and DEL RotTorq,i These represent the equivalent fatigue loads corresponding to the pitching moment, tilting moment, and transmission shaft torque of the i-th unit, respectively.

5. The method for optimizing primary frequency regulation control parameters of a wind farm considering fatigue load prediction according to claim 1, characterized in that, In step 4, during the rolling optimization process, the comprehensive equivalent fatigue load of each wind turbine is recalculated according to the set iteration interval, and the groups are reordered.

6. The method for optimizing primary frequency regulation control parameters of a wind farm considering fatigue load prediction according to claim 1, characterized in that, In step 5, for any candidate group G g While keeping the other group control parameters unchanged, a one-step trial reduction control is implemented for the candidate group, and the total equivalent fatigue load DEL of the wind farm after the trial adjustment is calculated. sum,g The lowest frequency point f min,g and equivalent fatigue load variance DEL var,g ; Then calculate the candidate group G according to the following formula. g The fatigue benefits, frequency costs, and load balance changes: ； Wherein, ΔDEL g , Δf g r g Representing candidate groups G respectively g Fatigue benefits, frequency costs, and load balance changes; DEL sum,base f min,base and DEL var,base These represent the total equivalent fatigue load, the lowest frequency point, and the variance of the equivalent fatigue load at the current rolling iteration starting point of the wind farm. Candidate group G is determined by combining fatigue benefits, frequency costs, and load balance changes. g Group score g : ； Where ε is a very small positive number to prevent the denominator from being zero, and λ is the load balance reward / penalty weight.

7. The method for optimizing primary frequency regulation control parameters of a wind farm considering fatigue load prediction according to claim 1, characterized in that, In step 6, the frequency safety margin is the difference between the lowest frequency point of the current scheme and the lowest frequency point constraint value; The recent candidate solution acceptance rate is the proportion of solutions accepted in the most recent rounds of candidate solutions; When the frequency security margin is less than a preset security threshold or the recent candidate scheme acceptance rate is lower than a preset acceptance rate threshold, the discrete adjustment step size is reduced.

8. The method for optimizing primary frequency regulation control parameters of a wind farm considering fatigue load prediction according to claim 1, characterized in that, In step 6, if the candidate control parameter scheme after the joint reduction control search meets the frequency response constraint and the total equivalent fatigue load of the wind farm is reduced or the comprehensive objective function is improved, then the control parameter scheme is accepted and the control parameters are updated; otherwise, the adjustment step size is reduced and the joint reduction control search is performed again. If the joint decreasing control search of the main path still fails to yield an acceptable solution after the step size is reduced to the minimum, then a univariate backoff adjustment is performed, including two methods: Only reduce the virtual inertia control coefficient of the units within the main path adjustment group, while keeping the droop control coefficient unchanged; Only reduce the droop control coefficient of the units within the main path adjustment group, while keeping the virtual inertia control coefficient unchanged.

9. The method for optimizing primary frequency regulation control parameters of a wind farm considering fatigue load prediction according to claim 1, characterized in that, In step 7, the iterative process of rolling optimization is terminated using the following multiple conditions as criteria: (1) Reach the maximum number of iterations for rolling optimization; (2) After the step size is reduced to the minimum value, there are no acceptable updates for three consecutive iterations; (3) The variance of equivalent fatigue load between units is reduced to less than 5% of the initial value; (4) The overall objective function has entered a plateau range and the improvement has been limited in several consecutive rounds; When any of the conditions in (1)-(4) above are met, the rolling optimization is terminated, and the optimized virtual inertia control coefficient and droop control coefficient of each wind turbine in the wind farm are output.

10. A computer system, characterized in that, include: One or more processors; A memory that stores operable instructions, which, when executed by the one or more processors, cause the one or more processors to perform the method according to any one of claims 1 to 9.

Images