Wind turbine control method based on cascaded nonlinear model and residual learning
Patent Information
- Application Number
- CN202610811695.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-05
- Publication Date
- 2026-09-29
AI Technical Summary
此类方案虽具备较强的非线性拟合能力，但存在无法规避的技术缺陷：其一，纯数据驱动模型属于黑箱模型，缺乏物理可解释性，无法保障控制策略的本质安全性，工程落地与运维验证难度极大；其二，模型的泛化能力高度依赖训练数据的规模与覆盖度，在有限运行数据下极易出现过拟合问题，在湍流、极端阵风等未训练的工况下，模型精度大幅下降，控制鲁棒性无法保障；其三，复杂的非线性黑箱模型会大幅提升模型预测控制滚动优化的在线求解复杂度，控制实时性难以满足工业现场毫秒级的控制周期要求
1) 级联非线性模型兼顾物理可解释性与非线性拟合能力，结合高斯过程回归残差前馈补偿，大幅降低部分负荷、满负荷全工况下的功率跟踪偏差，提升风能捕获效率；
Smart Images
Figure CN122834428A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of wind turbine control, and particularly relates to nonlinear system identification, model predictive control and fatigue load suppression technology for large wind turbine generator sets. Background Technology
[0002] With the continued advancement of the "dual-carbon" strategy, wind power, as a core component of renewable energy, has seen a continuous increase in installed capacity and unit capacity. Wind turbines operate in complex wind farm environments and are typical complex electromechanical systems characterized by strong nonlinearity, strong coupling, and multiple disturbances. Their operating conditions can be divided into a partial load zone below rated wind speed and a full load zone above rated wind speed. In the partial load zone, the core control objective is to achieve maximum power point tracking and maximize wind energy capture efficiency. In the full load zone, the core control objective is to maintain rated power output while suppressing fatigue loads on core components such as blades, drive trains, and towers, thereby extending the unit's service life. However, random disturbances such as turbulent winds and extreme gusts, as well as the vibration problems of flexible structures caused by large-scale wind turbines, place extremely high demands on the model accuracy, robustness, and multi-objective collaborative optimization capabilities of wind turbine control strategies.
[0003] Model predictive control (MDC) has become the mainstream technology for advanced wind turbine control due to its advantages in 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 generalization ability of the predictive model. Most existing wind turbine MDC schemes use linear time-invariant (LTI) models or piecewise linear multi-models as the predictive model. LTI models can only achieve high fitting accuracy near specific steady-state operating points and cannot cover the strong aerodynamic nonlinear characteristics of the wind turbine under all operating conditions. When wind speed fluctuates or the operating conditions deviate from the design operating point, the model prediction error increases sharply, resulting in insufficient power tracking accuracy and inability to achieve effective load suppression. While piecewise linear multi-models extend the operating condition coverage through multi-operating-point linearization, they suffer from control command jitter during model switching, which can exacerbate fatigue damage to turbine components. Furthermore, they still cannot fully fit the continuous strong nonlinear characteristics of the wind turbine's aerodynamic system and lack robustness under complex wind conditions.
[0004] To address the challenge of nonlinear modeling for wind turbines, existing technologies have proposed purely data-driven nonlinear modeling schemes based on neural networks and support vector machines. While these schemes possess strong nonlinear fitting capabilities, they suffer from unavoidable technical drawbacks: First, purely data-driven models are black-box models, lacking physical interpretability and failing to guarantee the inherent safety of the control strategy, making engineering implementation and operation verification extremely difficult. Second, the model's generalization ability is highly dependent on the scale and coverage of the training data, making it prone to overfitting under limited operational data. Under untrained conditions such as turbulence and extreme gusts, the model's accuracy drops significantly, compromising control robustness. Third, complex nonlinear black-box models significantly increase the online solution complexity of predictive control rolling optimization, making it difficult to meet the millisecond-level control cycle requirements of industrial sites.
[0005] Cascaded nonlinear models employ a cascaded structure of "static nonlinear modules + dynamic linear modules," combining nonlinear fitting capabilities with the interpretability and ease of solution of linear systems. The Hammerstein model is a typical example of a cascaded nonlinear structure, and existing technologies have attempted to apply it to wind turbine system modeling and control. However, existing wind turbine control schemes based on cascaded nonlinear models still have significant shortcomings: First, the fitting capability of the static nonlinear modules is insufficient, failing to accurately capture the strong nonlinear characteristics of multivariate coupling in the wind turbine aerodynamic system, including wind speed, pitch angle, and tip speed ratio, resulting in insufficient model accuracy. Second, the prediction residuals caused by unmodeled dynamics and random wind field disturbances are not considered, lacking online learning, uncertainty quantification, and feedforward compensation mechanisms for residuals. Under complex disturbances, model prediction errors cannot be effectively corrected, leading to significant performance degradation. Third, the controller design is not optimized for the core requirements of wind turbine load reduction, and the objective function weight tuning lacks a systematic optimal design method. It cannot collaboratively optimize power point tracking accuracy, actuator movement losses, and component fatigue load suppression. Furthermore, it lacks a robust mechanism for handling unit physical hard constraints, making it impossible to effectively guarantee operational safety under extreme conditions.
[0006] In summary, existing wind turbine control technologies cannot simultaneously meet the core requirements of "model physical interpretability, nonlinear fitting accuracy under all operating conditions, generalization ability under limited data, real-time online control, and synergistic optimization of power point tracking and load suppression." Under complex turbulent wind conditions and extreme operating conditions, the control robustness and safety are insufficient, and the fatigue load on the unit is too large, which seriously restricts the efficient, safe, and long-term stable operation of wind turbine generators. Summary of the Invention
[0007] To overcome the shortcomings of existing technologies, this invention proposes a wind turbine control method based on a cascaded nonlinear model and residual learning, aiming to balance power tracking and fatigue load suppression in wind turbine units. First, this invention establishes a dynamic model of the entire wind turbine system. System identification is performed based on a static nonlinear-dynamic linear cascaded structure. A fuzzy neural network is used to fit aerodynamic nonlinearity, and an autoregressive model with external input is used to fit mechanical dynamic characteristics, constructing a predictive model with both physical interpretability and nonlinear fitting capabilities. Then, Gaussian process regression is used to perform online learning and feedforward compensation of the model residuals, improving model accuracy and robustness. Finally, a model predictive controller is constructed based on the compensated model. Weights are tuned using linear quadratic regulator theory, and the optimal control command is obtained through rolling optimization. This invention significantly improves the power tracking accuracy of the wind turbine under all operating conditions, effectively suppresses unit fatigue load, exhibits strong robustness under turbulent wind conditions, and extends the service life of the unit.
[0008] The technical solution adopted by this invention to solve its technical problem is: A wind turbine control method based on a cascaded nonlinear model and residual learning includes the following steps: Step 1: Wind turbine full system dynamic modeling: Collect wind turbine design parameters, real-time operating parameters and environmental parameters, establish a wind turbine full system dynamic model including aerodynamic system, pitch angle servo system, transmission system, generator and tower, and clarify the input-output relationship and dynamic characteristics of each subsystem; Step 2: Identification and modeling of wind turbine system based on cascaded nonlinear structure: Construct a wind turbine prediction model with cascaded nonlinear structure, use a fuzzy neural network as a static nonlinear module to fit the strong nonlinear characteristics of the wind turbine aerodynamic system; use an autoregressive model with external input as a dynamic linear module to fit the mechanical dynamic characteristics of the transmission chain and generator, taking into account both the physical interpretability and nonlinear fitting ability of the model. Step 3: Residual modeling and feedforward compensation based on Gaussian process regression: For the unmodeled dynamics and external disturbances of the cascaded nonlinear prediction model, define the model output residual term, construct a Gaussian process regression model to learn and predict the residuals online, and superimpose the residual prediction values onto the model output to complete feedforward compensation, thus obtaining a global prediction model for wind turbines with residual compensation. Step 4: Linear Quadratic Regulator-Model Predictive Controller Design: A model predictive controller is constructed based on a global predictive model. The inverse module of the static nonlinear module is introduced to simplify the system linearization. The weight matrix is tuned in conjunction with the optimal control theory of the linear quadratic regulator. The objective function of multi-objective optimization is constructed, physical hard constraints of the unit are set, and the optimal control command is solved through rolling optimization.
[0009] Step 5, Closed-loop control execution and feedback: The optimal pitch angle and generator torque control commands are sent to the corresponding actuators, and the unit operation data is collected in real time and fed back to the model identification, residual prediction and controller modules to complete the closed-loop load reduction optimization control.
[0010] The beneficial effects of this invention are as follows: 1) The cascaded nonlinear model takes into account both physical interpretability and nonlinear fitting ability. Combined with Gaussian process regression residual feedforward compensation, it significantly reduces power tracking deviation under partial load and full load conditions, and improves wind energy capture efficiency. 2) By constraining the control increment through the objective function, the pitch action and drastic fluctuations in generator torque are limited, significantly reducing the fatigue load on blades, transmission chain, and tower, and extending the service life of the unit. 3) Gaussian process regression residual learning achieves good generalization ability under limited training data, while providing uncertainty quantification, and can still ensure control stability under turbulence and extreme wind conditions; hard constraint setting strictly limits the unit's operating range to avoid the risk of exceeding the limit.
[0011] 4) By transforming the cascaded nonlinear model into an inverse nonlinear block, the nonlinear system is transformed into a generalized linear system, which greatly reduces the solution complexity of model predictive control rolling optimization and improves the real-time performance and engineering feasibility of control. Attached Figure Description
[0012] Figure 1 This is a schematic diagram of the cascaded nonlinear model structure based on a fuzzy neural network with external input as described in this invention. Figure 2 This is the overall control block diagram of the wind turbine load reduction optimization control method described in this invention; Detailed Implementation
[0013] The present invention will be further described below with reference to the accompanying drawings.
[0014] Reference Figure 1 and Figure 2 A wind turbine control method based on cascaded nonlinear models and residual learning is proposed and applied to a 5MW horizontal axis wind turbine generator set. The core rated parameters of the turbine are shown in Table 1. This embodiment presents a wind turbine control method based on a cascaded nonlinear model and residual learning, including the following steps: Step 1: Full System Dynamics Modeling of Wind Turbine Generator: Collect the design parameters, real-time operating parameters, and environmental parameters of the target wind turbine generator, and establish a full system dynamics model of the wind turbine, including the aerodynamic system, pitch angle servo system, transmission system, generator, and tower. The modeling process is as follows: 1.1) The wind turbine captures wind energy and converts it into mechanical energy through an aerodynamic system. The formulas for calculating the aerodynamic torque and aerodynamic thrust of the wind turbine are: ; in, The torque coefficient, For thrust coefficient, For the tip speed ratio, The pitch angle is the propeller angle. Real-time wind speed; 1.2) The pitch servo system is simplified to a first-order inertial model, and its dynamic equations are: ; in, This is a reference command for the pitch angle. 1.3) The dynamic characteristics of the transmission system are characterized by a two-mass torsional vibration model, and its dynamic equation is: ; In the formula, The wind turbine rotation speed, For the aerodynamic torque of the wind turbine, Main spindle torque, The shaft torsion angle, For generator speed, This refers to the electromagnetic torque of the generator. 1.4) The generator is simplified to a first-order inertial system, and its dynamic equations are as follows: ; In the formula, This is a reference command for generator torque. This refers to the generator's output power. 1.5) Aerodynamic thrust drives the blades to rotate and is transmitted through the tower structure, generating a bending moment about the longitudinal axis, which is calculated as follows: ; In the formula, For the tower bending moment, For aerodynamic thrust, This refers to the tower height.
[0015] Step 2: Wind turbine system identification and modeling based on the Hammerstein structure: Based on the above-mentioned wind turbine full system dynamic model, the Hammerstein structure is used for system identification, and a wind turbine prediction model including static nonlinear modules and dynamic linear modules is constructed. The process is as follows: 2.1) A five-layer fuzzy neural network is used as a static nonlinear module, based on real-time wind speed. Tip speed ratio Pitch angle As input, the aerodynamic torque of the wind turbine Wind turbine aerodynamic thrust For output, fit the strongly nonlinear characteristics of the fan aerodynamic system; The calculation process of the fuzzy neural network is as follows: (a) The input layer receives input variables and passes them to the fuzzification layer; (b) The fuzzification layer transforms the input values into fuzzy values using a Gaussian membership function. The membership calculation formula for a single input and a single rule is as follows: ; In the formula, For the first The input variable for the first... The membership degree of a fuzzy rule. For the first The first moment One input variable, The central parameter of the Gaussian membership function is... The standard deviation of the Gaussian membership function; (c) The rule layer calculates the activation strength of each fuzzy rule, the first... The activation strength of a fuzzy rule is calculated as follows: ; (d) The normalization layer normalizes the activation intensity of all fuzzy rules, calculated as follows: ; In the formula, The total number of fuzzy rules, For fuzzy rule traversal index; (e) The inference layer performs a weighted summation of the normalized activation intensities to generate the final nonlinear output, thus completing the fitting of the aerodynamic nonlinear characteristics.
[0016] 2.2) An autoregressive model with external input is used as the dynamic linear module, with the output of the static nonlinear module and the generator torque as inputs, i.e. With the output power of the wind turbine generator For output, the dynamic characteristics of the wind turbine drive train and generator are fitted. The expression of the autoregressive model with external input is as follows: ; In the formula, , These represent the autoregressive order and the input order of the autoregressive model with external input, respectively, in this embodiment. , ; These are the autoregressive coefficients. To input dynamic weights, For the first Input variables at any given time; 2.3) By cascading the static nonlinear module and the dynamic linear module, the Hammerstein prediction model of the wind turbine is obtained, and its expression is: ; In the formula, Input to the model, As an intermediate variable, This represents the mapping relationship for static nonlinear modules. For model output, is the linear transfer function of the autoregressive model with external input.
[0017] Step 3: Residual Modeling and Feedforward Compensation Based on Gaussian Process Regression: For the unmodeled dynamics and external disturbances in the Hammerstein prediction model, a residual prediction model is constructed and feedforward compensation is performed. The process is as follows: 3.1) The difference between the actual output power of the wind turbine and the predicted power of the Hammerstein prediction model is defined as the residual term. ,Right now: ; In the formula, For the Hammerstein prediction model Predict output power in real time; 3.2) Construct a feature set using the historical input and output data of the wind turbine, and build a Gaussian process regression model, assuming that the residual term follows a Gaussian process with zero mean: ; In the formula, , For the input feature vector, For the kernel function, this embodiment uses the quadratic exponential kernel function, whose expression is: ; In the formula, For signal variance, The diagonal length scaling matrix is used; the hyperparameters of the Gaussian process regression model are optimized through maximum likelihood estimation, and the residual term is learned online. 3.3) Targeting new input features The mean of the residuals is predicted by the trained Gaussian process regression model. This is then superimposed on the output of the Hammerstein prediction model to complete feedforward compensation, resulting in a global prediction model for wind turbines with residual compensation. The final prediction output is: .
[0018] Step 4: Design of a linear quadratic regulator with residual compensation - Model predictive controller: Based on the above global predictive model, construct a model predictive controller and solve for the optimal control command. The process is as follows: 4.1) Introducing the inverse module of the Hammerstein structural static nonlinear module transforms the nonlinear wind turbine system into a generalized linear system, simplifying the solution process of model predictive control; 4.2) After receiving the real-time operating data of the wind turbine, the controller tunes the weight matrix using the optimal control theory of a linear quadratic regulator, constructs a multi-objective optimization objective function, and solves for the optimal control command through rolling optimization. The expression of the objective function is: ; In this embodiment, the prediction time domain is... Control time domain Weight matrix , , , The identity matrix is tuned using linear quadratic regulator theory to balance the requirements of power tracking performance, intermediate variable regularization, and control increment suppression. 4.3) During the optimization process, hard constraints are set for unit operation, specifically including: ; In the formula, the constraint terms are, in order, the pitch angle amplitude constraint. Pitch angle change rate constraint Wind turbine speed constraints Generator speed constraints Generator torque amplitude constraint Generator torque variation rate constraint Output power upper limit constraint ; 4.4) Within each control cycle, based on the real-time state of the wind turbine, the minimum value of the objective function that satisfies the above hard constraints is solved in both the prediction and control time domains to obtain the optimal pitch angle reference command and generator torque reference command.
[0019] Step 5: Closed-loop control implementation: The optimal pitch angle reference command obtained from the solution is sent to the pitch servo system, and the optimal generator torque reference command is sent to the generator converter. At the same time, the operating parameters and environmental parameters of the wind turbine are collected in real time and fed back to the system identification module, residual prediction module and controller to complete the closed-loop load reduction optimization control.
[0020] In this embodiment, a simulation model was established to verify the operation under full-load conditions. The results show that, under the condition that the physical dimensions of each module of the prediction model are strictly matched with the normalization range, the controller's rolling optimization solution process is stable and no extreme physical value deviations occur. The optimal control command output by this method effectively suppresses the drastic fluctuations in pitch and torque, significantly reduces the fatigue load on the core components of the unit while improving power tracking accuracy, and achieves the expected load reduction optimization objective.
[0021] 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 wind turbine control method based on a cascaded nonlinear model and residual learning, characterized in that, The method includes the following steps: Step 1: Collect wind turbine design parameters, real-time operating parameters and environmental parameters, and establish a dynamic model of the entire wind turbine system, including the aerodynamic system, pitch angle servo system, transmission system, generator and tower; Step 2: Construct a wind turbine prediction model with a cascaded nonlinear structure. Use a fuzzy neural network as a static nonlinear module to fit the strong nonlinear characteristics of the wind turbine aerodynamic system. Use an autoregressive model with external input as a dynamic linear module to fit the mechanical dynamic characteristics of the transmission chain and generator. Step 3: For the unmodeled dynamics and external disturbances of the cascaded nonlinear prediction model, define the model output residual term, construct a Gaussian process regression model to learn and predict the residuals online, and superimpose the residual prediction values onto the model output to complete feedforward compensation, thus obtaining a wind turbine global prediction model with residual compensation. Step 4: Construct a model predictive controller based on the global predictive model, introduce the inverse module of the static nonlinear module to complete the system linearization simplification, combine the optimal control theory of linear quadratic regulator to tune the weight matrix, construct the objective function of multi-objective optimization, set the physical hard constraints of the unit, and solve the optimal control command through rolling optimization. Step 5: Send the optimal pitch angle and generator torque control commands to the corresponding actuators, and at the same time collect the unit operation data in real time and feed it back to the model identification, residual prediction and controller modules to complete the closed-loop load reduction optimization control.
2. The wind turbine control method based on cascaded nonlinear models and residual learning according to claim 1, characterized in that, In step 1, the modeling process is as follows: 1.1) The wind turbine captures wind energy and converts it into mechanical energy through an aerodynamic system. The formulas for calculating the aerodynamic torque and aerodynamic thrust of the wind turbine are: ； in, The torque coefficient, For thrust coefficient, For the tip speed ratio, The pitch angle is the propeller angle. Real-time wind speed; 1.2) The pitch servo system is simplified to a first-order inertial model, and its dynamic equations are: ； in, This is a reference command for the pitch angle. 1.3) The dynamic characteristics of the transmission system are characterized by a two-mass torsional vibration model, and its dynamic equation is: ； In the formula, The wind turbine rotation speed, For the aerodynamic torque of the wind turbine, Main spindle torque, The shaft torsion angle, For generator speed, This refers to the electromagnetic torque of the generator. 1.4) The generator is simplified to a first-order inertial system, and its dynamic equations are as follows: In the formula, This is a reference command for generator torque. This refers to the generator's output power. 1.5) Aerodynamic thrust drives the blades to rotate and is transmitted through the tower structure, generating a bending moment about the longitudinal axis, which is calculated as follows: ； In the formula, For the tower bending moment, For aerodynamic thrust, This refers to the tower height.
3. The wind turbine control method based on cascaded nonlinear models and residual learning according to claim 1 or 2, characterized in that, The process of step 2 is as follows: 2.1) A five-layer fuzzy neural network is used as a static nonlinear module, based on real-time wind speed. Tip speed ratio Pitch angle As input, the aerodynamic torque of the wind turbine Wind turbine aerodynamic thrust The output is a fitted version of the strongly nonlinear characteristics of the fan aerodynamic system. 2.2) An autoregressive model with external input is used as the dynamic linear module, with the output of the static nonlinear module and the generator torque as inputs, i.e. With the output power of the wind turbine generator For output, the dynamic characteristics of the wind turbine drive train and generator are fitted. The expression of the autoregressive model with external input is as follows: ； In the formula, , These represent the autoregressive order and the input order of the autoregressive model with external input, respectively. These are the autoregressive coefficients. To input dynamic weights, For the first Input variables at any given time; 2.3) By cascading the static nonlinear module and the dynamic linear module, the Hammerstein prediction model of the wind turbine is obtained, and its expression is: ； In the formula, Input to the model, As an intermediate variable, This represents the mapping relationship for static nonlinear modules. For model output, is the linear transfer function of the autoregressive model with external input.
4. The wind turbine control method based on cascaded nonlinear models and residual learning according to claim 3, characterized in that, In section 2.1), the calculation process of the fuzzy neural network is as follows: (a) The input layer receives input variables and passes them to the fuzzification layer; (b) The fuzzification layer transforms the input values into fuzzy values using a Gaussian membership function. The membership calculation formula for a single input and a single rule is as follows: ； In the formula, For the first The input variable for the first... The membership degree of a fuzzy rule. For the first The first moment One input variable, The central parameter of the Gaussian membership function is... The standard deviation of the Gaussian membership function; (c) The rule layer calculates the activation strength of each fuzzy rule, the first... The activation strength of a fuzzy rule is calculated as follows: ； (d) The normalization layer normalizes the activation intensity of all fuzzy rules, calculated as follows: ； In the formula, The total number of fuzzy rules, For fuzzy rule traversal index; (e) The inference layer performs a weighted summation of the normalized activation intensities to generate the final nonlinear output, thus completing the fitting of the aerodynamic nonlinear characteristics.
5. The wind turbine control method based on cascaded nonlinear models and residual learning according to claim 1 or 2, characterized in that, In step 3, the process of constructing a residual prediction model and performing feedforward compensation for the unmodeled dynamics and external disturbances of the Hammerstein prediction model is as follows: 3.1) The difference between the actual output power of the wind turbine and the predicted power of the Hammerstein prediction model is defined as the residual term. ,Right now: ； In the formula, For the Hammerstein prediction model Predict output power in real time; 3.2) Construct a feature set using the historical input and output data of the wind turbine, and build a Gaussian process regression model, assuming that the residual term follows a Gaussian process with zero mean: ； In the formula, , For the input feature vector, For kernel functions; 3.3) Targeting new input features The mean of the residuals is predicted by the trained Gaussian process regression model. This is then superimposed on the output of the Hammerstein prediction model to complete feedforward compensation, resulting in a global wind turbine prediction model with residual compensation. The final prediction output is: 。 6. The wind turbine control method based on cascaded nonlinear models and residual learning according to claim 5, characterized in that, In section 3.2), the kernel function is the quadratic exponential kernel function, whose expression is: ； In the formula, For signal variance, The diagonal length scaling matrix is used; the hyperparameters of the Gaussian process regression model are optimized through maximum likelihood estimation, and the residual term is learned online.
7. The wind turbine control method based on cascaded nonlinear models and residual learning according to claim 1 or 2, characterized in that, In step 4, the process of constructing a model predictive controller based on the above global prediction model and solving for the optimal control command is as follows: 4.1) Introducing the inverse module of the Hammerstein structural static nonlinear module transforms the nonlinear wind turbine system into a generalized linear system, simplifying the solution process of model predictive control; 4.2) After receiving the real-time operating data of the wind turbine, the controller tunes the weight matrix using the optimal control theory of a linear quadratic regulator, constructs a multi-objective optimization objective function, and solves for the optimal control command through rolling optimization. The expression of the objective function is: ； In the formula, To predict the time domain, To control the time domain, the weight matrix , , , The identity matrix is tuned using linear quadratic regulator theory to balance the requirements of power tracking performance, intermediate variable regularization, and control increment suppression. 4.3) During the optimization process, hard constraints are set for unit operation, including: ； In the formula, the constraint terms are, in order, the pitch angle amplitude constraint. Pitch angle change rate constraint Wind turbine speed constraints Generator speed constraints Generator torque amplitude constraint Generator torque variation rate constraint Output power upper limit constraint ; 4.4) Within each control cycle, based on the real-time state of the wind turbine, the minimum value of the objective function that satisfies the above hard constraints is solved in both the prediction and control time domains to obtain the optimal pitch angle reference command and generator torque reference command.
8. The wind turbine control method based on cascaded nonlinear models and residual learning according to claim 1 or 2, characterized in that, In step 5, the optimal pitch angle reference command obtained by the solution is sent to the pitch servo system, and the optimal generator torque reference command is sent to the generator converter. At the same time, the operating parameters and environmental parameters of the wind turbine are collected in real time and fed back to the system identification module, residual prediction module and controller to complete the closed-loop load reduction optimization control.