A wind turbine full load condition load shedding optimization control method

Abstract

A kind of wind turbine full load operating condition load shedding optimization control method belongs to wind power generation control technical field, for wind turbine rated wind speed above full load operating condition, establish the dynamics model of wind turbine whole system, decoupling identification is carried out based on Hamers tangent structure, through radial basis function neural network fitting aerodynamic strong nonlinearity, through the autoregressive moving average model with external input fitting electromechanical dynamic characteristics and processing turbulent colored noise, construct the prediction model with high precision and physical explainability;Design cascade prediction inverse compensation model prediction control architecture, outer ring is based on linear model constraint rolling optimization to solve optimal aerodynamic torque reference, inner ring generates executable control instruction by adaptive inverse algorithm with physical gradient guardrail.The present application significantly improves the rated power tracking accuracy under high turbulence condition, effectively suppresses structural fatigue load, control real-time and robustness is strong, suitable for megawatt wind turbine rated wind speed above operating condition of safe and efficient operation control.

Description

Technical Field

[0001] This invention belongs to the field of wind turbine control, specifically involving nonlinear system identification, cascade model predictive control and fatigue load suppression technology for large horizontal axis wind turbine generators operating at wind speeds above rated speed. Background Technology

[0002] With the continuous increase in the scale of wind power installations and the capacity of individual units, large wind turbine generators operate for extended periods in complex turbulent wind fields. Their operating conditions can be categorized into partial load conditions below the rated wind speed and full load conditions above the rated wind speed, based on wind speed. Under full load conditions above the rated wind speed, the core objective of wind turbine control undergoes a fundamental shift: it no longer pursues maximizing wind energy capture, but rather, while strictly maintaining rated power output and meeting grid connection specifications, it aims to minimize torsional vibration of the transmission chain and tower structure caused by high turbulence, reduce the cumulative fatigue load on core components, and extend the service life of the unit.

[0003] Model predictive control (MDC) has become the mainstream approach for advanced control of wind turbines under full load conditions due to its advantages of multivariable coordinated control, explicit handling of multiple constraints, and rolling optimization with look-ahead prediction. Its control performance is highly dependent on the accuracy and structural rationality of the predictive model. Existing wind turbine MDC schemes suffer from the following core shortcomings: 1. Conventional linear time-invariant models can only maintain accuracy near specific operating points and cannot cover the strong nonlinear characteristics of aerodynamic systems under load conditions. When wind speed fluctuates, the model error increases sharply, leading to power tracking inaccuracy and load suppression failure. Although piecewise linear multi-models expand the operating condition coverage, they have the problem of control command jitter when switching models, which exacerbates component fatigue damage.

[0004] 2. Although pure data-driven nonlinear black-box models have nonlinear fitting capabilities, they lack physical interpretability, making engineering implementation and verification difficult. Furthermore, the online solution complexity of nonlinear model predictive control is extremely high, which cannot meet the millisecond-level real-time control requirements of industrial sites. At the same time, the model's generalization ability is highly dependent on the training data, and the accuracy is significantly reduced under high-turbulence untrained conditions.

[0005] 3. Existing wind turbine modeling schemes based on Hammerstein structures mostly employ joint parameter estimation of nonlinear and linear modules, which suffers from severe parameter coupling interference. This leads to biased parameter estimation of linear modules and insufficient model accuracy. Furthermore, linear modules often use autoregressive models with external inputs, which cannot handle colored noise interference from turbulent winds. Under full-load and high-turbulence conditions, the prediction error increases significantly.

[0006] 4. The existing single-loop nonlinear control architecture cannot simultaneously achieve efficient solutions for nonlinear decoupling and linear control. Local gradient singularities and gradient vanishing problems are prone to occur during the inverse solution of the aerodynamic model, leading to instability of inverse control. Under extreme wind conditions, it is impossible to strictly guarantee the compliance of actuator constraints and the safety of system operation.

[0007] In summary, existing wind turbine full-load operating condition control technologies cannot simultaneously meet the core requirements of "high model accuracy, strong physical interpretability, real-time online control, multi-objective collaborative optimization, and robustness under extreme operating conditions." The rated power tracking accuracy is insufficient under high turbulent wind conditions, and the structural fatigue load is too large, which seriously restricts the safe, long-term, efficient and stable operation of large wind turbine generator sets. Summary of the Invention

[0008] To overcome the shortcomings of existing technologies, this invention proposes a load reduction optimization control method for wind turbine generators under full-load conditions. This method aims to balance power tracking and fatigue load suppression. Specifically, for wind turbines operating at full load above rated wind speed, this invention establishes a dynamic model of the entire wind turbine system. Based on a static nonlinear-dynamic linear cascade structure, decoupling identification is performed. A radial basis function neural network is used to fit the strong aerodynamic nonlinearity, and an autoregressive moving average model with external input is used to fit the electromechanical dynamic characteristics and handle turbulent colored noise, constructing a predictive model with both high accuracy and physical interpretability. Furthermore, a cascade predictive inverse compensation model predictive control architecture is designed. The outer loop uses a linear model to achieve constrained rolling optimization to solve for the optimal aerodynamic torque reference, while the inner loop generates executable control commands through an adaptive inverse algorithm with physical gradient guardrails. This invention significantly improves rated power tracking accuracy under high turbulence conditions, effectively suppresses structural fatigue loads, and exhibits strong real-time control and robustness. It can be widely applied to the safe and efficient operation control of megawatt-level wind turbine generators operating above rated wind speed.

[0009] The technical solution adopted by this invention to solve its technical problem is: A method for optimizing load reduction control of a wind turbine generator under full-load conditions includes the following steps: Step 1: Full-system dynamic modeling of wind turbine under full load conditions: For full-load operation conditions above rated wind speed, establish a continuous domain dynamic model of the entire wind turbine system, including the aerodynamic system, transmission chain system, pitch actuator, and tower structure, and clarify the three core control objectives of "rated power tracking + fatigue load suppression + hard constraint compliance" under full load conditions; Step 2: Decoupling and Identification Modeling of Wind Turbine System Based on Cascade Structure: Construct a cascaded nonlinear prediction model that perfectly matches the physical architecture of the wind turbine. Use a radial basis function neural network as a static nonlinear module and independently pre-train it to fit the strong nonlinear characteristics of the wind turbine aerodynamic system. Use an autoregressive moving average model with external input as a linear dynamic module to fit the linear dynamic characteristics of the drive train, generator, and tower. At the same time, handle the colored noise interference caused by turbulent wind to complete the complete decoupling and identification of nonlinear and linear modules. Step 3: Construction of a cascade predictive inverse compensation control architecture matched with the cascade model: Construct a predictive inverse compensation model predictive control cascade control architecture. The outer loop is a constraint model predictive controller based on a linear dynamic model to solve for the optimal aerodynamic torque reference command. The inner loop is a nonlinear adaptive inverse compensator based on a radial basis function neural network to transform the abstract torque reference into physically executable pitch angle and generator torque control commands, realizing the fusion of nonlinear decoupling and linearized control. Step 4: Design of outer loop linear model predictive controller: The identified linear dynamic model is transformed into a minimum realization discrete state space equation. An integral disturbance rejection model is introduced to eliminate steady-state tracking error. A quadratic objective function is constructed that takes into account both generator speed tracking accuracy and control increment smoothness. Physical hard constraints of the unit are set. The optimal aerodynamic torque reference command is solved by rolling optimization through a quadratic programming algorithm. Step 5: Inner Loop Adaptive Inverse Compensator Design and Closed-Loop Control: The Jacobian adaptive inverse solution algorithm with physical gradient guardrail is adopted. Based on real-time wind speed, tip speed ratio and aerodynamic fitting model, the local aerodynamic sensitivity Jacobian matrix is ​​calculated. Gradient anomalies are corrected through physical boundary constraints. The outer loop aerodynamic torque reference command is converted into pitch angle control command. The generator torque control command is generated synchronously, sent to the actuator, and real-time operating data is collected to complete the closed-loop control.

[0010] The beneficial effects of this invention are as follows: 1) The cascaded nonlinear structure corresponds completely to the physical subsystem of the wind turbine. The radial basis function neural network accurately fits the strong nonlinearity of aerodynamic multivariates. The autoregressive moving average model with external input effectively handles the colored noise caused by turbulence. The decoupled identification completely avoids the coupling bias of joint parameter estimation. It can still maintain high model accuracy under high turbulence full load conditions, and at the same time has complete physical interpretability. 2) The cascade predictive inverse compensation control architecture avoids the high online computation of nonlinear model predictive control through the outer loop linear model predictive control, greatly reducing the solution complexity and ensuring the real-time control needs of industrial sites; the inner loop inverse compensation completely eliminates the model error caused by traditional linearization, and the inverse solution algorithm with physical gradient guardrail solves the pain points of local singularity and gradient vanishing in the neural network inverse model, ensuring control stability under extreme conditions. 3) Achieve precise tracking of rated power and rated speed under full load conditions, strictly meet grid connection specifications; at the same time, through control incremental constraints and dynamic optimization, effectively suppress torsional vibration of the transmission chain and forward and backward vibration of the tower, significantly reduce the cumulative fatigue damage of core components such as blades, transmission chain, and tower, and significantly extend the service life of the unit. 4) The entire process strictly follows the physical constraints of the wind turbine, setting hard constraint boundaries for the pitch angle amplitude / rate of change and the generator torque amplitude / rate of change to avoid the actuator from operating beyond the limit; the identification and control algorithm is modularly designed to adapt to the hardware architecture of the existing wind turbine main control system, so that the project can be implemented without large-scale hardware modification. Attached Figure Description

[0011] Figure 1 The diagram shows a cascaded nonlinear model structure according to a preferred embodiment of the invention, wherein the cascaded nonlinear model adopts a Hammerstein cascade structure. Figure 2 This is an overall block diagram of the predictive control architecture of the predictive inverse compensation cascade model described in this invention. Detailed Implementation

[0012] The present invention will be further described below with reference to the accompanying drawings.

[0013] Reference Figure 1 and Figure 2 An optimized control method for load reduction under full-load conditions of a wind turbine generator is disclosed. As an embodiment of the present invention, the cascaded nonlinear model adopts a Hammerstein cascade structure and is applied to a 5MW horizontal axis wind turbine generator. Its core rated parameters are: rated power 5MW, rated wind speed 11.6m / s, cut-in wind speed 3.2m / s, cut-out wind speed 25.3m / s, rotor radius 63.4m, gearbox transmission ratio 97.6:1, generator rated speed 1185rpm, rotor rated speed 12.5rpm, maximum pitch actuator change rate 9.3° / s, generator rated torque 4715kNm, and generator torque maximum change rate 156kNm / s.

[0014] The wind turbine generator set full-load condition load reduction optimization control method of this embodiment includes the following steps: Step 1: Full-system dynamic modeling of the wind turbine under full load conditions: For the full-load operation condition above the rated wind speed of the target unit, collect the wind turbine design parameters, real-time operating parameters, and wind speed data of the wind farm, and establish a continuous domain dynamic model of the entire wind turbine system, including the aerodynamic system, transmission chain system, pitch actuator, generator, and tower structure. The modeling process is as follows: 1.1) The wind turbine captures kinetic energy from the turbulent wind field and converts it into mechanical torque. The formulas for calculating aerodynamic power and aerodynamic torque are: ; In the formula, air density Wind turbine radius , The effective wind speed at the wheel hub. The wind energy utilization coefficient, The aerodynamic torque coefficient, For the tip speed ratio, The pitch angle is the propeller angle. This refers to the wind turbine rotation speed; 1.2) The torsional flexibility of the transmission chain is characterized by a two-mass torsional vibration model, and the dynamic equation is: ; In the formula, the moment of inertia of the wind turbine is... Generator rotational inertia Shaft stiffness coefficient =868547000 Nm / rad, shaft damping coefficient =6236000 Nm / (rad / s), gearbox transmission ratio =0.0102, Main spindle torque, The shaft torsion angle, For generator speed, This refers to the electromagnetic torque of the generator. 1.3) The electro-hydraulic pitch system is simplified to a first-order inertial model, and the dynamic equations are: ; In the formula, the inertial time constant of the pitch actuator is... =0.2s, This is a reference command for the pitch angle. 1.4) The generator model ignores millisecond-level electromagnetic transients and simplifies to a first-order inertial model. The dynamic equations are: ; In the formula, the generator inertial time constant =0.024s, generator efficiency =0.954, This is a reference command for generator torque. This refers to the generator's output power. 1.5) The forward and backward vibration of the tower is simplified to a single-degree-of-freedom spring-mass-damped system, and the dynamic equation is: ; In the formula, For the longitudinal displacement of the tower top, , , These are the equivalent mass, damping, and stiffness of the tower, respectively. This is the aerodynamic thrust coefficient. It is aerodynamic thrust; Based on the above dynamic model, the three core control objectives under full load conditions are defined as follows: a. The generator speed and output power strictly track the rated values ​​to meet the grid connection requirements; b. Suppress the torsional vibration of the transmission chain and the forward and backward vibration of the tower to reduce the fatigue load on the components; c. All control actions strictly comply with the physical hard constraints of the actuators to ensure the safe operation of the unit.

[0015] Step 2: System Decoupling Identification and Modeling Based on Hammerstein Cascade Structure: Based on the above wind turbine dynamics model, a Hammerstein cascade structure that perfectly matches the physical architecture is used for system decoupling identification. A radial basis function-autoregressive moving average wind turbine prediction model with external input is constructed, as follows: 2.1) Construct a 3-input, 2-output radial basis function neural network as a static nonlinear module of the Hammerstein structure, with the input vector being... This refers to real-time wind speed, blade pitch angle, and tip speed ratio; the output vector is... That is, the aerodynamic torque and aerodynamic thrust of the wind turbine.

[0016] The radial basis function neural network adopts a three-layer structure: an input layer, a hidden layer, and an output layer. The number of neurons in the hidden layer is... =50, the hidden Gaussian kernel mapping expression is: ; In the formula, For the output layer connection weights, Let be the center vector of the Gaussian kernel. The width of the basis functions; A full-condition operation dataset of the target unit under full load was collected and divided into a training set of 80% and a test set of 20%. The radial basis neural network was independently supervised and pre-trained to complete the fitting of the aerodynamic nonlinear mapping and completely isolate the nonlinear characteristics from the linear dynamic characteristics. 2.2) Construct a multi-input, single-output autoregressive moving average model with an external input as a linear dynamic module of the Hammerstein structure. The input is the aerodynamic torque output by the radial basis function neural network. Aerodynamic thrust and generator electromagnetic torque The output is the generator speed. and tower displacement ,Right now The expression for the discrete-domain difference equation is: ; In the formula, the shift operator satisfy Autoregressive polynomial order =4, that is Exogenous input polynomial order =[1,1,1]; Order of the moving average noise polynomial =3, that is Effectively handles colored noise interference from turbulent winds; transmission delay of all input channels. =1; Based on the same full-load working condition dataset, the aerodynamic torque and aerodynamic thrust output by the trained radial basis neural network are used as inputs to complete the parameter identification of the autoregressive moving average model with external input, and obtain the coefficient matrix of each polynomial. 2.3) The pre-trained radial basis static nonlinear module is cascaded with the identified autoregressive moving average linear dynamic module with external input to obtain the wind turbine Hammerstein prediction model, realizing the complete decoupling of aerodynamic nonlinearity and electromechanical dynamic characteristics.

[0017] Step 3: Construct the cascaded predictive inverse compensation model predictive control architecture, as follows: A cascaded predictive inverse compensation model predictive control architecture that perfectly matches the Hammerstein predictive model described above is constructed. This architecture consists of two interconnected control loops: an outer loop and an inner loop. The outer loop is a constrained model predictive controller based on an autoregressive moving average linear dynamic model with external input. Its core function is to solve for the optimal abstract aerodynamic torque reference command by rolling optimization based on the generator speed tracking target. ; The inner loop is a nonlinear adaptive inverse compensator based on a radial basis function neural network. Its core function is to convert the abstract aerodynamic torque reference command output by the outer loop into a physically executable pitch angle control command in real time. With generator torque control command Simultaneously, nonlinear decoupling and physical constraint limiting are achieved; This cascade architecture completely avoids the high online computational cost of traditional nonlinear model predictive control, while eliminating model errors caused by linearization, thus achieving a balance between control accuracy and real-time performance.

[0018] Step 4: Design of the outer-loop linear model predictive controller: Based on the identified autoregressive moving average model with external input, an outer-loop linear model predictive controller is constructed, as follows: 4.1) Using a minimum implementation algorithm, the identified autoregressive moving average difference equation with external input is transformed into a discrete-domain state-space model, eliminating redundant zero-pole dynamics, isolating unmeasurable noise components, and extracting the transitivity of deterministic physical inputs. The resulting state-space prediction model is as follows: ; In the formula, The system state vector is of minimum dimension. Given a deterministic input vector, the system output vector , , , , The state-space system matrix; Meanwhile, an integral output disturbance model is introduced into the state-space model to eliminate steady-state speed tracking error caused by wind speed disturbance. 4.2) Using the generator's rated speed as the tracking target and aerodynamic torque as the control variable, construct a constrained quadratic objective function: ; In the formula, the prediction time domain =15, control time domain =3, For the first Time for the first The predicted value of the generator speed at any given time. =122.9 rad / s is the rated speed of the generator. For tower displacement, The control increment for the aerodynamic torque reference command, weight matrix =10, =0.5, =1, balancing speed tracking accuracy and control smoothness, suppressing load fluctuations; 4.3) During the optimization process, hard constraints are set for unit operation, specifically expressed as follows: ; In the formula, and The first The pitch angle and its rate of change at time t. , and These are the minimum amplitude limit, maximum amplitude limit, and maximum pitch rate limit for the pitch angle, respectively. and The first The wind turbine speed and generator speed at any given time, , These are the cut-in speed and rated speed of the wind turbine, respectively. , These are the generator's cut-in speed and rated speed, respectively. and The first The generator electromagnetic torque at time t and its rate of change (control increment). and These are the rated torque and maximum torque change rate limits for the generator, respectively. For the first Generator output power at any given time This is the rated output power of the generator.

[0019] Meanwhile, proxy boundary constraints are set for the optimized variable aerodynamic torque reference command to limit its amplitude and rate of change, ensuring the physical executability of the control command; 4.4) Within each control cycle, based on the real-time operating status of the wind turbine, the minimum value of the objective function that satisfies the hard constraints is solved by a quadratic programming algorithm in the prediction time domain and the control time domain to obtain the optimal aerodynamic torque reference command sequence, and the command of the first control step is output to the inner loop inverse compensator.

[0020] Step 5, Inner Loop Radial Base Adaptive Inverse Compensator Design and Closed-Loop Control: The inner loop radial base nonlinear adaptive inverse compensator is used to convert the aerodynamic torque reference command output from the outer loop into physically executable control commands. The process is as follows: 5.1) Generate the feedforward pitch angle baseline based on real-time wind speed and full-load operating parameters. The local Jacobian matrix of the pre-trained radial basis aerodynamic model is calculated using the forward finite difference method. The local sensitivity of aerodynamic torque to pitch angle is characterized by the following formula: ; In the formula, =0.1° represents the step size of a small perturbation; 5.2) Based on the physical characteristics of the aerodynamic system, the aerodynamic torque under full load decreases monotonically with increasing propeller pitch angle; therefore, the Jacobian matrix must be strictly negative. The safety gradient matrix is ​​obtained by correcting the local Jacobian matrix through physical boundary constraints. The revised rule is as follows: ; In the formula, =-0.1MNm / ° is the maximum gradient value at the physical boundary, which completely avoids inverse control instability caused by gradient anomalies; 5.3) Based on the anchor point mapping using Taylor series expansion, and combined with the outer-loop optimal aerodynamic torque reference command, the final pitch angle control command is obtained, expressed as: ; In the formula, =4.3MNm is the aerodynamic torque anchor point value under full load conditions, which ensures complete matching with the steady-state operating point of the outer loop model predictive control design and eliminates steady-state deviation; 5.4) Under full-load steady-state conditions, the generator torque command is maintained at the rated value. =4715kNm; When transient conditions such as grid voltage drop or wind speed drop occur, smooth degradation protection is achieved based on generator speed feedback to avoid speed spike or stall. 5.5) After physical hard constraint limiting processing of the generated pitch angle control command and generator torque control command, they are sent to the pitch actuator and generator converter; at the same time, the operating parameters and wind speed data of the wind turbine are collected in real time and fed back to the model identification module, the outer loop model prediction controller and the inner loop inverse compensator to complete the full closed-loop full load condition optimization control.

[0021] In this embodiment, the proposed method was validated under high-turbulence full-load conditions by establishing a full-system dynamic simulation model of a 5MW unit. The results show that the decoupling identification method effectively overcomes the interference of colored noise from turbulence, and the cascade control architecture significantly reduces the online solution complexity of nonlinear model predictive control, fully meeting the requirements of real-time control. In particular, the inverse solution algorithm for the inner ring with physical gradient guardrails successfully avoids the problems of local singularities and gradient vanishing under extreme wind conditions, ensuring control stability. This method, while strictly adhering to the unit's physical hard constraints, accurately maintains the rated power output and significantly suppresses torsional vibration of the drive train and forward / backward vibration of the tower, greatly reducing the cumulative fatigue damage of core components.

[0022] The embodiments described in this specification are merely examples of implementations of the inventive concept and are for illustrative purposes only. The scope of protection of this invention should not be considered limited to the specific forms described in these embodiments; rather, it extends to equivalent technical means conceived by those skilled in the art based on the inventive concept.

Claims

1. A method for optimizing load reduction control of a wind turbine generator under full-load conditions, characterized in that, Includes the following steps: Step 1: For full-load operation at wind speeds above the rated wind speed, establish a continuous domain dynamic model of the entire wind turbine system, including the aerodynamic system, transmission chain system, pitch actuator, and tower structure. Step 2: Construct a cascaded nonlinear prediction model that perfectly matches the physical architecture of the wind turbine. Use a radial basis function neural network as a static nonlinear module to independently pre-train and fit the strong nonlinear characteristics of the wind turbine aerodynamic system. Use an autoregressive moving average model with external input as a linear dynamic module to fit the linear dynamic characteristics of the drive train, generator, and tower, while handling the colored noise interference caused by turbulent wind. Step 3: Construct a predictive inverse compensation model predictive control cascade control architecture. The outer loop is a constraint model predictive controller based on a linear dynamic model to solve for the optimal aerodynamic torque reference command. The inner loop is a nonlinear adaptive inverse compensator based on a radial basis function neural network to transform the abstract torque reference into physically executable pitch angle and generator torque control commands. Step 4: Transform the identified linear dynamic model into a minimum realization discrete state-space equation, introduce an integral disturbance rejection model to eliminate steady-state tracking error, construct a quadratic objective function that balances generator speed tracking accuracy and control increment smoothness, set physical hard constraints for the unit, and solve for the optimal aerodynamic torque reference command through a quadratic programming algorithm for rolling optimization. Step 5: Using the Jacobian adaptive inverse solution algorithm with physical gradient guardrail, based on real-time wind speed, tip speed ratio and aerodynamic fitting model, calculate the local aerodynamic sensitivity Jacobian matrix, correct gradient anomalies through physical boundary constraints, convert the outer loop aerodynamic torque reference command into pitch angle control command, synchronously generate generator torque control command, send it to the actuator and collect real-time operating data to complete closed-loop control.

2. The method for optimizing load reduction control of a wind turbine generator under full load conditions according to claim 1, characterized in that, The process of step 1 is as follows: 1.1) The wind turbine captures kinetic energy from the turbulent wind field and converts it into mechanical torque. The calculation formulas for aerodynamic power and aerodynamic torque are given. 1.2) The torsional flexibility of the transmission chain is characterized by a two-mass torsional vibration model, and the dynamic equations are given; 1.3) The electro-hydraulic pitch system is simplified to a first-order inertial model, and the dynamic equations are given; 1.4) The generator model ignores millisecond-level electromagnetic transients and simplifies to a first-order inertial model, giving the dynamic equations; 1.5) The forward and backward vibration of the tower is simplified into a single-degree-of-freedom spring-mass-damped system, and the dynamic equations are given.

3. The method for optimizing load reduction control of a wind turbine generator under full load conditions according to claim 2, characterized in that, In step 1, based on the above dynamic model, the three core control objectives under full load conditions are defined as follows: a. The generator speed and output power strictly track the rated values ​​to meet the grid connection requirements; b. Suppress torsional vibration of the transmission chain and longitudinal vibration of the tower, and reduce fatigue load on components; c. All control actions strictly adhere to the physical constraints of the actuators to ensure safe unit operation.

4. A method for optimizing load reduction control of a wind turbine generator under full load conditions according to any one of claims 1 to 3, characterized in that, The process of step 2 is as follows: 2.1) Construct a 3-input, 2-output radial basis function neural network as a static nonlinear module of the Hammerstein structure, with the input vector being... , For real-time wind speed, The pitch angle is the propeller angle. The tip speed ratio is ; the output vector is , For the aerodynamic torque of the wind turbine, It is aerodynamic thrust; The radial basis function neural network adopts a three-layer structure: an input layer, a hidden layer, and an output layer, with the number of neurons in the hidden layer set. The expression for the hidden Gaussian kernel mapping is given; the full-condition operation dataset of the target unit under full load is collected, divided into training set and test set, and the radial basis neural network is independently supervised pre-trained to complete the fitting of the aerodynamic nonlinear mapping, thus completely isolating the nonlinear characteristics from the linear dynamic characteristics.

5. The method for optimizing load reduction control of a wind turbine generator under full load conditions according to claim 4, characterized in that, Step 2 also includes: 2.2) Construct a multi-input, single-output autoregressive moving average model with an external input as a linear dynamic module of the Hammerstein structure. The input is the aerodynamic torque output by the radial basis function neural network. Aerodynamic thrust and generator electromagnetic torque The output is the generator speed. and tower displacement ,Right now The discrete domain difference equation expression is given; based on the same full-load working condition dataset, the aerodynamic torque and aerodynamic thrust output by the trained radial basis neural network are used as inputs to complete the parameter identification of the autoregressive moving average model with external input, and the coefficient matrix of each polynomial is obtained. 2.3) The pre-trained radial basis static nonlinear module is cascaded with the identified autoregressive moving average linear dynamic module with external input to obtain the wind turbine Hammerstein prediction model, realizing the complete decoupling of aerodynamic nonlinearity and electromechanical dynamic characteristics.

6. The method for optimizing load reduction control of a wind turbine generator under full load conditions according to claim 5, characterized in that, The process of step 3 is as follows: A cascaded predictive inverse compensation model predictive control architecture that perfectly matches the Hammerstein predictive model described above is constructed. This architecture consists of two interconnected control loops: an outer loop and an inner loop. The outer loop is a constrained model predictive controller based on an autoregressive moving average linear dynamic model with external input. Its core function is to solve for the optimal abstract aerodynamic torque reference command by rolling optimization based on the generator speed tracking target. ; The inner loop is a nonlinear adaptive inverse compensator based on a radial basis function neural network. Its core function is to convert the abstract aerodynamic torque reference command output by the outer loop into a physically executable pitch angle control command in real time. With generator torque control command It simultaneously achieves nonlinear decoupling and physical constraint limiting.

7. A method for optimizing load reduction control of a wind turbine generator under full load conditions according to any one of claims 1 to 3, characterized in that, The process of step 4 is as follows: 4.1) The minimum implementation algorithm is adopted to transform the identified autoregressive moving average difference equation with external input into a discrete domain state-space model, eliminate redundant zero-pole dynamics, isolate unmeasurable noise components, extract the transfer characteristics of deterministic physical input, and obtain a state-space prediction model. Meanwhile, an integral output disturbance model is introduced into the state-space model to eliminate steady-state speed tracking error caused by wind speed disturbance. 4.2) Using the generator's rated speed as the tracking target and aerodynamic torque as the control variable, construct a constrained quadratic objective function.

8. The method for optimizing load reduction control of a wind turbine generator under full load conditions according to claim 7, characterized in that, The process in step 4 also includes: 4.3) During the optimization process, hard constraints are set for unit operation. At the same time, proxy boundary constraints are set for the optimized variable aerodynamic torque reference command to limit its amplitude and rate of change, ensuring the physical executability of the control command. 4.4) Within each control cycle, based on the real-time operating status of the wind turbine, the minimum value of the objective function that satisfies the hard constraints is solved by a quadratic programming algorithm in the prediction time domain and the control time domain to obtain the optimal aerodynamic torque reference command sequence, and the command of the first control step is output to the inner loop inverse compensator.

9. A method for optimizing load reduction control of a wind turbine generator under full load conditions according to any one of claims 1 to 3, characterized in that, The process of step 5 is as follows: 5.1) Generate the feedforward pitch angle baseline based on real-time wind speed and full-load operating parameters. ; The local Jacobian matrix of the pre-trained radial basis aerodynamic model is calculated using the forward finite difference method. This characterizes the local sensitivity of aerodynamic torque to the pitch angle. 5.2) Based on the physical characteristics of the aerodynamic system, the aerodynamic torque under full load conditions decreases monotonically with increasing propeller pitch angle. Therefore, the Jacobian matrix must be strictly negative. By correcting the local Jacobian matrix through physical boundary constraints, the safety gradient matrix is ​​obtained. Provide correction rules; 5.3) Based on the anchor point mapping of Taylor series expansion, combined with the outer loop optimal aerodynamic torque reference command, the final pitch angle control command is obtained.

10. The method for optimizing load reduction control of a wind turbine generator under full load conditions according to claim 9, characterized in that, Step 5 also includes: 5.4) Under full-load steady-state conditions, the generator torque command is maintained at the rated value. When transient conditions such as grid voltage drop or sudden wind speed drop occur, smooth degradation protection is achieved based on generator speed feedback to avoid speed spikes or stall. 5.5) After physical hard constraint limiting processing of the generated pitch angle control command and generator torque control command, they are sent to the pitch actuator and generator converter; at the same time, the operating parameters and wind speed data of the wind turbine are collected in real time and fed back to the model identification module, the outer loop model prediction controller and the inner loop inverse compensator to complete the full closed-loop full load condition optimization control.

Images