Wind generating set model prediction control method based on Hammerstein structure

Abstract

The invention discloses a wind generating set model prediction control method based on a Hammerstein structure, and the method achieves the linear modeling of a system by introducing the Hammerstein structure to approximate the nonlinear static characteristics and linear dynamic behaviors of a wind generating set. On the basis, a multivariable model predictive control framework is established by adjusting the torque and the pitch angle of the generator at the same time. In order to solve the problem of multi-objective optimization between power tracking and fatigue load suppression, a quadratic cost function including a power error term, a control change rate term and a load suppression term is designed, so that smooth power output and structure service life prolonging of the wind turbine unit are realized.

Description

Technical Field

[0001] This invention relates to the field of wind turbine control, and more specifically to a model predictive control method for wind turbine generator sets based on the Hammerstein structure. Background Technology

[0002] Wind turbine generators are typically designed to adapt to varying wind speeds in order to capture more wind energy. However, under conditions of significant wind speed variation, the generator's output power can fluctuate, affecting power quality and system stability. Long-term power fluctuations and aerodynamic load variations can lead to fatigue damage to components such as blades, main shafts, and towers. Therefore, effectively mitigating fatigue loads on wind turbine generators has become a crucial technical challenge in operation and maintenance.

[0003] The fatigue load on wind turbine generators is mainly caused by fluctuating operating conditions resulting from the system's strong nonlinearity and high uncertainty. Furthermore, rotor aerodynamic imbalances caused by wind shear and tower shadow effects further increase the fatigue levels of structural components such as blades, main shaft, drive train, and tower. With the continuous expansion of wind farm scale, balancing grid power stability with the suppression of generator fatigue loads has become a significant technical challenge for wind turbine generator control strategies.

[0004] Existing wind turbine generators have significant nonlinear and time-varying characteristics. Traditional linearization modeling is only effective under local operating conditions and cannot accurately reflect the global dynamic characteristics of the system, resulting in insufficient model prediction accuracy and control performance that is easily affected by wind speed fluctuations.

[0005] While multi-model predictive control (MMPC) can improve the controller's adaptability to nonlinear conditions to some extent, frequent model switching causes control signal fluctuations and exacerbates actuator fatigue. Therefore, there is an urgent need for an optimized control method that can balance power tracking and fatigue load suppression in wind turbines. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a model predictive control method for wind turbine generators based on the Hammerstein structure, aiming to balance power point tracking and fatigue load suppression in wind turbine generators.

[0007] The wind turbine model predictive control method includes wind turbine model system identification and model predictive controller design. This method introduces a Hammerstein structure to approximate the nonlinear static characteristics and linear dynamic behavior of the wind turbine, thereby achieving linearized modeling of the system. Based on this, a multivariable model predictive control framework is established by simultaneously adjusting the generator torque and pitch angle. To address the multi-objective optimization problem between power point tracking and fatigue load suppression, a quadratic cost function is designed, incorporating power error, control rate of change, and load suppression terms, thereby achieving smooth power output and extended structural lifespan of the wind turbine.

[0008] The technical solution adopted by this invention to solve its technical problem is:

[0009] A predictive control method for wind turbine generator models based on the Hammerstein structure includes the following steps:

[0010] Step 1: Wind turbine model identification:

[0011] The Hammerstein structure is introduced to approximate the nonlinear static characteristics and linear dynamic behavior of wind turbine generators. This combined structure enables linear modeling of the wind power generation system, providing an accurate model basis for subsequent control.

[0012] Step 2: Model Predictive Controller Design:

[0013] Using the Hammerstein model as the prediction model, a model predictive controller (MPC) framework is constructed to predict the future operating state of the wind power generation system and generate the corresponding control input sequence, wherein the control input includes two control parameters: generator torque and pitch angle.

[0014] Step 3: Design of multi-objective cost function:

[0015] The design incorporates a multi-objective cost function that includes power tracking, load suppression, and smooth control actions. Under the premise of satisfying various operating constraints of the wind turbine generator, a rolling optimization solution is performed. The optimal control command obtained from the solution is then applied to the wind turbine generator to achieve optimal control of the system.

[0016] Furthermore, the specific process of step 1 is as follows:

[0017] 1.1) The Hammerstein model consists of a static nonlinear part and a dynamic linear part. The relationship between them is as follows: the intermediate variables are obtained by the input quantities through static nonlinear mapping, and the output quantities are calculated by the intermediate variables through the linear dynamic relationship. Among them, the input quantities and output quantities represent the control input parameters and operating state output parameters of the wind turbine generator, respectively. The sampling time is used to identify the system data at different times. The intermediate variables are parameters that cannot be directly measured within the model. The linear dynamic relationship contains two sets of feature matrices, which correspond to the historical feedback coefficients of the output quantities and the input transmission coefficients of the intermediate variables, respectively. The two sets of matrices correspond to the two key structural parameters of the model: the output order and the input order.

[0018] The model parameters are determined by solving an optimization problem. The goal is to keep the error between the estimated output of the model and the actual output of the system within an allowable range (not exceeding a set threshold). The optimization process must satisfy two basic constraints: a static nonlinear mapping relationship and a linear dynamic relationship. The estimated output value is the output prediction result calculated by the Hammerstein model, the estimated intermediate variable is the model's prediction of the internal unmeasurable parameters, and the estimated matrix is ​​the linear dynamic part feature matrix obtained through optimization.

[0019] 1.2) A multilayer perceptron (MLP) is used to approximate the static nonlinear part of the Hammerstein model. The input vector of the neural network consists of three key operating parameters: blade pitch angle, blade tip speed ratio, and wind speed. The output vector corresponds to the intermediate variables of the Hammerstein model, which contain core information such as speed-related parameters and force-related parameters.

[0020] The neural network's output calculation process is as follows: the final output value of the intermediate variables is obtained by adding the bias term of the network's output layer to the weighted sum of the outputs of each hidden layer node. The output of each hidden layer node is calculated using a nonlinear transfer function, and the input signal of each node is the sum of the node's bias term and the elements of the input vector, weighted accordingly. This neural network has a total of 10 hidden layer nodes, achieving a high-precision approximation of static nonlinear characteristics through this structure.

[0021] 1.3) The dynamic linear part of the Hammerstein model is approximated by a low-order autoregressive moving average model (ARX model). The input vector of this part consists of intermediate variables output from the static nonlinear part and disturbance-related parameters, while the output vector contains core operating state variables of the wind turbine, such as generator angular velocity, load parameters, speed parameters, and output power.

[0022] By constructing a dynamic equation between the input and output of the dynamic linear part using the time delay operator (which is essentially a linear relationship between the current and historical input quantities and the corresponding coefficients, and the time delay operator is used to characterize the impact of input at different times on the current output), and combining it with the output of the static nonlinear part, the complete output equation of the Hammerstein model is finally obtained. This equation can directly reflect the correspondence between the input quantity and the actual output parameters of the wind turbine generator.

[0023] Furthermore, the specific process of step 2 is as follows:

[0024] Based on the Hammerstein model obtained in step 1, at each control sampling time, the time-varying parameters of the model at the current operating point of the wind turbine are calculated. The Hammerstein model is linearized in real time through the time-varying parameters to obtain a linear prediction model suitable for the current operating conditions.

[0025] A linear prediction model is used to predict the system output over multiple future sampling periods, resulting in a predicted system output sequence. This predicted output sequence consists of three parts, and its mathematical logic can be simply described as "predicted output = step response coefficient × control increment + free response". Specifically, it includes: the step response coefficient of the dynamic linear part, the control increment sequence in the future control time domain (including pitch angle change and generator torque change), and the free response sequence determined by the current system state.

[0026] Terminal constraints and terminal cost functions are introduced into the controller framework. Terminal constraints are used to limit the convergence range of the future state of the system, and terminal cost functions are used to enhance the stability of the optimization process and ensure that the model predictive controller is stable and reliable throughout the entire operation.

[0027] Furthermore, the specific process of step 3 is as follows:

[0028] To achieve a balance between power point tracking accuracy, fatigue load suppression, and control smoothness, a quadratic multi-objective cost function is constructed. The optimization objective of this function is to minimize the overall control deviation. The core components of the cost function include: the deviation between the predicted generator angular velocity and the reference value, the deviation between the predicted load parameters and the target value, the deviation between the predicted speed parameters and the target value, the deviation between the predicted output power and the reference value, the pitch angle control increment, the generator torque control increment, and a relaxation factor term used to adjust the feasibility of optimization.

[0029] Each deviation term and control increment term is prioritized using weighting factors: the output weighting factor sets the importance of each operating state deviation, the input weighting factor limits drastic changes in control actions, and the penalty coefficient adjusts the influence of the relaxation factor. The relaxation factor allows for a very small range of deviations in output power, ensuring that the optimization problem still has a feasible solution under stringent constraints.

[0030] All weighting factors are organized in matrix form, divided into output weight matrix and input weight matrix. In the input weight matrix, the weight values ​​of pitch angle control increment and generator torque control increment are both set to 1. The specific parameters of the output weight matrix are as follows: generator angular velocity deviation weight is 3, load parameter deviation weight is 0.01, speed parameter deviation weight is 0.8, and output power deviation weight is 0.5. This weight allocation prioritizes ensuring the stability of generator operation.

[0031] The physical and operational constraints of the wind turbine generator set are set as constraints for the optimization problem, specifically including: output constraints: the output power does not exceed the rated power, the speed does not exceed the safety limit, and the torque fluctuates within the allowable range;

[0032] Input constraints: The pitch angle variation range is between the minimum and maximum angles allowed by the mechanical structure, and the rate of change of the generator torque does not exceed the equipment's tolerance limit;

[0033] Control increment constraints: Limit the maximum change in pitch angle and generator torque within each sampling period to ensure smooth control action.

[0034] Specifically, these include: output power not exceeding the rated power and not lower than zero; rotor speed between the cut-in speed and the rated speed; generator torque not exceeding the rated torque and not lower than zero; generator angular velocity between the minimum allowable value and the rated value; and pitch angle between the minimum limit value and the maximum limit value. These constraints ensure that all control variables and operating state variables are within physically permissible safe ranges, avoiding drastic fluctuations in pitch angle and generator torque.

[0035] The constructed multi-objective cost function is combined with constraints to transform it into a standard quadratic programming problem (which aims to minimize the cost function value and solve for the optimal control quantity within the given inequality constraints). At each sampling time, an efficient quadratic programming solver is used to solve the problem online in real time. After the solution is completed, only the first step of the optimized control input sequence is selected and applied to the wind turbine generator. Then, the next sampling period is entered, the system state is updated, and the above optimization process is repeated to achieve adaptive control based on the rolling time domain.

[0036] The design concept of this invention is as follows:

[0037] This method introduces a Hammerstein structure to approximate the nonlinear static characteristics and linear dynamic behavior of wind turbine generators, thereby achieving linearized modeling of the system. Based on this, a multivariable model predictive control framework is established by simultaneously adjusting the generator torque and pitch angle. To address the multi-objective optimization problem between power point tracking and fatigue load suppression, a quadratic cost function is designed, incorporating power error, control rate of change, and load suppression terms, thus achieving smooth power output and extended structural lifespan for the wind turbine generator.

[0038] The beneficial effects of this invention are as follows:

[0039] 1) The Hammerstein structure is used to jointly model the nonlinear static and linear dynamic characteristics of the wind turbine generator, which improves the modeling accuracy and control robustness of the system.

[0040] 2) Coordinated adjustment of generator torque and pitch angle enables multivariable predictive control, which can effectively smooth power output and enhance system stability;

[0041] 3) Design a multi-objective cost function that comprehensively considers power error, control smoothing and load suppression, so as to significantly reduce the fatigue load of the unit while ensuring power tracking accuracy. Attached Figure Description

[0042] Figure 1 Schematic diagram of a predictive control scheme for a wind turbine model based on the Hammerstein structure. Detailed Implementation

[0043] The present invention will be further described below with reference to the accompanying drawings.

[0044] like Figure 1 As shown, a predictive control method for a wind turbine generator model based on the Hammerstein structure is proposed, aiming to achieve multi-objective optimization of power point tracking and fatigue load suppression. The specific steps are as follows:

[0045] Step 1: Wind Turbine Model Identification

[0046] 1.1) The Hammerstein model consists of a static nonlinear part and a dynamic linear part:

[0047]

[0048]

[0049] in, and These represent the input and output of the wind turbine, respectively. Indicates the sampling time; It is an unmeasurable intermediate variable; and It is a matrix constructed from dynamic linear variables; and They are matrices and The corresponding model order;

[0050] Then, the Hammertein model is obtained by solving an optimization problem:

[0051]

[0052] in, This represents the estimated output value of the Hammertein model; These are estimates of unmeasured intermediate variables; and It is an estimation matrix constructed from dynamic linear elements; This represents the maximum permissible fitting error when the Hammerstein model approximates a real wind turbine. This indicates the number of sample data points used to train or validate the model.

[0053] 1.2) The static nonlinear part of the Hammerstein model is approximated using a multilayer perceptron (MLP). The input and output vectors of the neural network are defined as follows:

[0054]

[0055]

[0056] in, The pitch angle is the distance between the paddles. For the tip speed ratio, Wind speed. An intermediate variable in the Hammertein model, i.e., the output of the neural network. The equation is shown below:

[0057]

[0058] in, Represents a nonlinear transfer function. The network's first One output, Represents the network weights. Indicates the sampling time. By and the Number of nodes (total number of nodes) Associated input signals Calculations show that For the output layer of the neural network The bias value of each output node. Indicates the hidden layer number 1 The node to the output layer The weight coefficient of each node.

[0059]

[0060] Indicates the number of steps from the input layer to the hidden layer. The bias of each node. Indicates the input layer's first... The input is fed into the hidden layer. The weight of each node.

[0061] Therefore, the network output can be derived.

[0062]

[0063] 1.3) The dynamic linear part can be approximated as a low-order ARX model, whose input and output are defined as follows:

[0064]

[0065]

[0066] Based on the ARX form, the dynamic equation can be written as:

[0067]

[0068] in, It is an intermediate variable. It is a time delay operator. Indicates the number of input variables This indicates white noise interference. and Defined by the following formula,

[0069]

[0070] and This indicates the order of the dynamic linear part.

[0071] Therefore, the complete output equation of the Hammertein model is as follows.

[0072]

[0073] Step 2: Model Predictive Controller Design

[0074] Based on the Hammerstein model obtained in step one, at each sampling time... The model is linearized to calculate the predicted output; linearization is achieved by calculating the current operating point (the operating point is the input data at the current time k). Time-varying parameters under the determined system state and The calculation formula is as follows.

[0075]

[0076] Therefore, the predicted output sequence can be represented as:

[0077]

[0078] in Indicates the forecast period. This refers to the generator torque.

[0079] Finally, the complete system output prediction sequence can be obtained.

[0080]

[0081] matrix These are the step response coefficients of the linear part of the Hammerstein model at the corresponding operating point. For the aforementioned free response sequence, It is the sequence of control increments in the future control time domain, given by the following formula:

[0082]

[0083] in, This represents the predicted pitch angle change at time k, for the k-th step. Indicates control of the time domain. This represents the change in generator torque.

[0084] Step 3: Design of Multi-Objective Cost Function

[0085] To balance power tracking performance with fatigue load suppression, a quadratic cost function is constructed:

[0086]

[0087] in, This indicates the reference value for generator speed. This represents the predicted generator speed. This indicates the reference value for the tower's bending moment. This represents the predicted bending moment of the tower. This indicates the reference value for the spindle torque. This represents the predicted value of the spindle torque. Indicates the power reference value. This represents the predicted power value. This represents the change in pitch angle. This indicates the change in generator torque.

[0088] in , , and It is the output weighting factor, used to weight the deviation between the predicted output and the reference trajectory. and This is the input weighting factor, used to weight changes in the control input. Relaxation factor. This represents the minimum and maximum power output deviation that can be tolerated at each sampling time, ensuring that the optimization problem always has a better solution. Penalty coefficient. Used to adjust the degree of influence of slack variables.

[0089] The weighting factors are represented in the form of a weight matrix.

[0090]

[0091] in It is the output weight matrix. This is the input weight matrix. The value of the input weight matrix is... The output weight matrix value is

[0092]

[0093] To ensure system safety and actuator limitations, the constraints include:

[0094]

[0095] in Rated output power, and These are the cut-in rotor speed and the generator speed, respectively. and These are the rated rotor speed and the generator speed, respectively. For the minimum pitch angle, This is the maximum pitch angle.

[0096] These constraints ensure that the control quantities vary within physically permissible limits, avoiding drastic fluctuations in pitch and generator torque.

[0097] The aforementioned optimization problem is constructed as a standard quadratic programming (QP) problem, which is solved online through rolling time-domain optimization at each sampling time to obtain the optimal control input sequence. The first step control variable is selected and applied to the system, followed by rolling updates at the next time step to achieve real-time adaptive control.

[0098] Compared to traditional model predictive control (MPC) strategies, the method proposed in this invention exhibits superior robustness and load suppression capabilities under various turbulent conditions. Experimental results show that under low-turbulence conditions, this method reduces the power tracking error (RMSE) by approximately 70%, and reduces the fatigue damage equivalent load (DEL) of the main shaft torque and tower bending moment by approximately 15% and 24%, respectively. Even under strongly disturbed high-turbulence conditions, its RMSE is still reduced by 39.82%, and the fatigue loads of the two key components are reduced by 15.61% and 19.84%, respectively.

[0099] 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 predictive control method for a wind turbine generator model based on the Hammerstein structure, characterized in that, Includes the following steps: Step 1: Wind turbine model identification: The Hammerstein model structure is introduced to approximate the nonlinear static characteristics and linear dynamic behavior of wind turbine generators, thereby realizing the linearization modeling of the wind power generation system and providing an accurate model basis for subsequent control. Step 2: Model Predictive Controller Design: Using the Hammerstein model as the prediction model, a Model Predictive Controller (MPC) framework is constructed to predict the future operating state of the wind power generation system and generate the corresponding control input sequence, wherein the control input includes generator torque and pitch angle. Step 3: Design of multi-objective cost function: The design incorporates a multi-objective cost function that includes power tracking, load suppression, and smooth control actions. Under the premise of satisfying various operating constraints of the wind turbine generator, a rolling optimization solution is performed. The optimal control command obtained from the solution is then applied to the wind turbine generator to achieve optimized control of the wind power generation system.

2. The predictive control method for a wind turbine generator model based on the Hammerstein structure according to claim 1, characterized in that, The specific process of step 1 is as follows: 1.1) The Hammerstein model consists of a static nonlinear part and a dynamic linear part. The relationship between them is as follows: the intermediate variables are obtained by the input quantities through static nonlinear mapping, and the output quantities are calculated by the intermediate variables through linear dynamic relationships. The model parameters are determined by solving an optimization problem. The goal is to keep the error between the estimated model output and the actual system output within an acceptable range. The optimization process must satisfy two basic constraints: static nonlinear mapping relationship and linear dynamic relationship. 1.2) A multilayer perceptron (MLP) is used to approximate the static nonlinear part of the Hammerstein model. The input vector consists of three key operating parameters: blade pitch angle, tip speed ratio, and wind speed. The output vector corresponds to the intermediate variables of the Hammerstein model. The final output value of the intermediate variables is obtained by adding the bias term of the network output layer to the weighted sum of the outputs of each hidden layer node. The output of the hidden layer node is calculated through a nonlinear transfer function. The input signal of each node is composed of the bias term of that node and the elements of the input vector after being weighted by their respective weights. 1.3) The dynamic linear part of the Hammerstein model is approximated by a low-order ARX model. The input vector consists of intermediate variables and disturbance-related parameters output from the static nonlinear part, while the output vector includes generator angular velocity, load parameters, speed parameters, and output power. The dynamic equation between the input and output of the dynamic linear part is constructed by using the time delay operator. Combined with the output of the static nonlinear part, the complete output equation of the Hammerstein model is finally obtained, which directly reflects the correspondence between the input quantity and the actual output parameters of the wind turbine generator.

3. The predictive control method for a wind turbine generator model based on the Hammerstein structure according to claim 1, characterized in that, The specific process of step 2 is as follows: Based on the Hammerstein model obtained in step 1, at each control sampling time, the time-varying parameters of the model at the current operating point of the wind turbine are calculated. The Hammerstein model is then linearized in real time using the time-varying parameters to obtain a linear prediction model suitable for the current operating conditions. A linear prediction model is used to predict the output of the system over multiple future sampling periods, resulting in a predicted output sequence. This predicted output sequence consists of three parts: predicted output = step response coefficient × control increment + free response. Terminal constraints and a terminal cost function are introduced. The terminal constraints are used to limit the convergence range of the future state of the system, and the terminal cost function is used to enhance the stability of the optimization process, ensuring that the model predictive controller is stable and reliable throughout the entire operation.

4. The predictive control method for a wind turbine generator model based on the Hammerstein structure according to claim 1, characterized in that, The specific process of step 3 is as follows: To achieve a balance between power point tracking accuracy, fatigue load suppression, and control smoothness, a quadratic multi-objective cost function is constructed, with the optimization objective being to minimize the overall control deviation. The cost function includes: the deviation between the generator angular velocity prediction and the reference value, the deviation between the load parameter prediction and the target value, the deviation between the speed parameter prediction and the target value, the deviation between the output power prediction and the reference value, the pitch angle control increment, the generator torque control increment, and a relaxation factor term used to adjust the feasibility of optimization. Each deviation term and control increment term is prioritized through weighting factors: the output weighting factor is used to set the importance of each operating state deviation, the input weighting factor is used to limit drastic changes in control actions, and the penalty coefficient is used to adjust the influence of the relaxation factor. The relaxation factor allows for a very small range of deviation in output power, ensuring that the optimization problem still has a feasible solution under strict constraints. The physical and operational constraints of the wind turbine generator set are set as constraints for the optimization problem, specifically including: Output constraints: Output power shall not exceed rated power, speed shall not exceed the safety limit, and torque shall fluctuate within the allowable range; Input constraints: The pitch angle variation range is between the minimum and maximum angles allowed by the mechanical structure, and the rate of change of the generator torque does not exceed the equipment's tolerance limit; Control increment constraints: Limit the maximum change in pitch angle and generator torque within each sampling period to ensure smooth control action; The constructed multi-objective cost function is combined with constraints to transform it into a standard quadratic programming problem. At each sampling time, it is solved online in real time using an efficient quadratic programming solver. After the solution is completed, only the first step of the optimized control input sequence is selected and applied to the wind turbine generator. Then, the next sampling period is entered, the system state is updated, and the above optimization process is repeated to achieve adaptive control based on the rolling time domain.

Images