Intermittent output wind driven generator pitch angle time-varying neuron adaptive control method
By using an intermittent output switching state observer and time-varying neuron adaptive control, the problem of unstable pitch angle output of wind turbines was solved, reducing costs and improving information reliability, and adapting to complex wind conditions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-28
- Publication Date
- 2026-04-03
AI Technical Summary
Traditional wind turbine pitch angle control methods are difficult to adapt to the intermittent output caused by time-varying wind speed and turbulence, resulting in unstable output, affecting power quality and equipment lifespan, and neural network control is expensive.
By employing an intermittent output switching state observer and combining it with neural network estimation of disturbances and position dynamics, a time-varying neuron adaptive control method is designed. By dividing the observer into open-loop and closed-loop observers, the observation error is reduced and the use of neurons is optimized.
It improves the reliability of pitch angle output information, reduces calculation and hardware costs, adapts to sudden wind speed changes, reduces equipment wear, and conforms to engineering practice.
Smart Images
Figure CN121782102A_ABST
Abstract
Description
Technical Field
[0001] The time-varying neuron adaptive control method for intermittent output wind turbine pitch angle belongs to the field of wind turbine control. Background Technology
[0002] Wind power has become one of the core forces replacing traditional fossil fuels. However, with the increasing capacity of individual wind turbine units and their expansion into offshore and low-wind-speed areas, the complexity of wind conditions faced by these units has significantly increased. Factors such as random fluctuations in natural wind speed, turbulence effects, and gust impacts cause intermittent output in wind turbine pitch angle measurements. This output instability not only affects the power quality after wind power is integrated into the grid, causing voltage fluctuations and frequency deviations, but it can also exacerbate fatigue wear on key components such as the turbine's transmission system and tower, shortening equipment lifespan and severely restricting the large-scale application and efficient utilization of wind power.
[0003] As a core technology for power regulation and load optimization of wind turbine generators, pitch angle control directly determines the operating status of the unit under different wind conditions: at low wind speeds, the optimal pitch angle is maintained to maximize wind energy capture; above the rated wind speed, the angle between the blades and the airflow is adjusted to limit the amount of wind energy captured, ensuring that the output power remains stable near the rated value. However, traditional pitch angle control methods, such as PID control (e.g., Uehara, Akie and Pratap, Alokand Goya, Tomonori and Senjyu, Tomonobu and Yona, Atsushi and Urasaki, Naomitsuand Funabashi, Toshihisa. A coordinated control method to smooth wind power fluctuations of a PMSG-based WECS[J]. IEEE Transactions on energy conversion, 2011, 26(02), 550-558.) and sliding mode control methods (e.g., Zhu Yaoming, Zhang Lanhong, Chen Lulu. Vector control of doubly fed wind turbine based on sliding mode variable structure[J]. Southern Energy Construction, 2025, 12(01): 160-167. DOI: 10.16516 / j.ceec.2024-127.), are difficult to adapt to the time-varying wind speed, nonlinear aerodynamic parameters, and dynamic uncertainties caused by intermittent output. When faced with sudden power surges or drops caused by strong turbulence or gusts of wind, traditional controllers are prone to problems such as response lag, large overshoot, or insufficient robustness. They not only fail to quickly suppress power fluctuations caused by intermittent output, but may also exacerbate component wear and increase maintenance costs due to control inaccuracies. At the same time, to compensate for the performance defects of traditional control, higher-specification mechanical structures or energy storage devices are required, which further increases the equipment manufacturing cost and project investment cost. Meanwhile, with the rapid development of intelligent control theory, neural networks have shown significant advantages in the control of complex systems due to their powerful nonlinear mapping capabilities, self-learning, and adaptive characteristics. However, conventional neural control methods often employ fixed structures and parameters, leading to excessive neuron calls and increasing hardware and computational costs (e.g., Wang T, Zong G, Zhao X, et al. Data-driven-based sliding-mode dynamic event-triggered control of unknown nonlinear systems via reinforcement learning[J].Neurocomputing, 2024, 601:128176.), failing to fundamentally solve the cost optimization problem.
[0004] Meanwhile, with the rapid development of intelligent control theory, neural networks have shown significant advantages in the control of complex systems due to their powerful nonlinear mapping capabilities, self-learning and adaptive characteristics. However, conventional neural network control methods often use fixed structures and parameters, leading to excessive neural network calls, increasing hardware and computational costs, and failing to fundamentally solve the cost optimization problem. In summary, the traditional wind turbine pitch angle control scheme has the following problems: (1) discontinuous pitch angle output information caused by sudden changes in wind speed or communication interference. (2) high computational and hardware costs of pitch angle adaptive neural network control. (3) the problem that the observer observation error is difficult to conform to the specific engineering reality due to the boundedness of the intermittent output state observer error analysis. Summary of the Invention
[0005] The technical problem to be solved by this invention is to overcome the shortcomings of the prior art and provide an intermittent output switching state observer. This invention utilizes a neural network to estimate disturbances and position dynamics, thereby reducing the observer's observation error and improving the reliability of the system's pitch angle output information under sudden wind speed changes. This is a time-varying neuron adaptive control method for intermittent output wind turbine pitch angle.
[0006] The technical solution adopted by this invention to solve its technical problem is: the time-varying neuron adaptive control method for intermittently outputting wind turbine pitch angle, characterized by the following steps: Step 1: Establish a model of an intermittent wind turbine; Step 2: Based on the intermittent wind turbine model determined in Step 1, determine its control objective and lemma; Step 3: Construct a state observer for the intermittent output-switching structure of the wind turbine model; Step 4, Radial basis function neural network estimation based on time-varying neurons; Step 5: Construct a time-varying neuron adaptive trajectory tracking controller.
[0007] Preferably, in step 1: define state variables and The wind turbine pitch angle system is transformed into a second-order nonlinear system in the following state-space form, i.e., the expression for the intermittent wind turbine model is: in, The nominal nonlinear dynamic function of the system is known. , Let be the torque constant of the pitch motor. The equivalent rotational inertia of the pitch system, is the equivalent viscous friction coefficient of the pitch bearing and drive train. This refers to the input voltage or current control quantity. For lumped composite disturbance terms, It is the propeller pitch angle.
[0008] Preferably, in step 3: This addresses the unknown dynamics that arise in the intermittent wind turbine model determined in step 1. The following results were obtained by estimating it using an RBF neural network: Among them, approximation error satisfy Expected weight satisfy , These are the basis functions of the RBF neural network; Using the above formulas, the design of the intermittent output-switching structure state observer is determined as follows: in, State variables The estimated value, State variables The estimated value, and They are respectively and The first derivative, and For the time-varying correction gain of the observer, Expected weight The estimated value, , The input is represented as neural network activation functions, , Let be the torque constant of the pitch motor. The equivalent rotational inertia of the pitch system, This refers to the input voltage or current control quantity. Indicates time, Indicates the output error. Time period The output information cannot measure the start time.
[0009] Preferably, in step 4: define the time-varying neural network estimation function. for: in, Indicates the number of neurons. Indicates the first The weights of each neuron, Indicates the first The activation function corresponding to each neuron Let be the neuron change function, defined as ,in, For the floor function, Let be the time-varying function of the neuron, defined as: in, This is the upper bound of the time-varying function of the neuron. For the rate of change, Indicates time.
[0010] Preferably, in step 4: the neural network approximator is reconstructed using vector projection techniques. in, , , , Represents the weight vector With basis function vectors The angle between them, the optimal neural network approximator with time-varying neurons is: in, To minimize the estimation error, and satisfy the following: , For positive integers, The optimal weight is determined by these factors.
[0011] Preferably, in step 5: define the coordinate transformation. , in, For positional error, For speed error, For virtual control signals, For pitch angle tracking signal, State variables The estimated value, State variables The estimated value; The time-varying neuron adaptive continuous feedback control law is designed as follows: in, This is a continuously time-varying neuron adaptive trajectory tracking control law. and To control the gain, For the desired trajectory The first derivative, The number of neurons is The corresponding activation function, The weighted projection estimate, for The renewal law, Indicates in time Weight projection estimate, adaptive law parameters , , , Let be the torque constant of the pitch motor. The equivalent rotational inertia of the pitch system; The time-varying neuronal neural network was further identified as follows: in, For ideal weights, The activation function corresponding to the weight. estimation error, The nominal nonlinear dynamic function of the system is known. For lumped composite disturbance terms, This is a virtual control signal.
[0012] Compared with the prior art, the beneficial effects of this invention are: In the intermittent output wind turbine pitch angle time-varying neuron adaptive control method of this application, an intermittent output switching state observer is proposed. By dividing the design into open-loop and closed-loop observers, the neural network is used to estimate the disturbance and position dynamics, thereby reducing the observer's observation error and improving the reliability of the system's pitch angle output information under sudden wind speed changes.
[0013] This application addresses the problem of discontinuous pitch angle measurement or output information transmission caused by external environmental factors. It designs an intermittent output switching state observer that utilizes a radial basis function neural network to estimate external disturbances and nonlinear dynamics. This eliminates the need for a more precise wind turbine pitch system model and simultaneously addresses the adverse effects of unknown dynamics and disturbances on the observed values. The intermittency of the output information is characterized as unmeasurable and measurable periods. Closed-loop observation is performed during measurable periods, while open-loop observation is performed during unmeasurable periods. Correction terms are removed during open-loop periods to avoid correction errors that may be caused by the unmeasurability of the output information.
[0014] By intermittently switching the state observer's output, when pitch angle information is unmeasurable, the correction term is discarded, prioritizing the system's use of existing information for observation; during closed-loop periods, the neural network is immediately used for state observation. This method eliminates the need for state expansion or linearization, significantly improving the observer's ability to handle output discontinuities.
[0015] To address the high computational and hardware costs associated with neural network operation, a time-varying neuron method was designed, compared to traditional neural network control. Unlike traditional fixed-neuron neural networks, it has lower requirements for computational power and complexity. The time-varying neuron method relies on a bounded-growth time-varying function, allowing the number of neurons to increase steadily, avoiding the sharp estimation curve caused by an initially excessive number of neurons. Furthermore, the time-varying neuron method requires vector reconstruction techniques to design the neural network estimation law, achieving good estimation results.
[0016] The time-varying neuron method proposed in this application alleviates computational pressure and reduces the hardware cost of using neural network estimation. Compared with other control methods, neural network control is easier to design, does not require an accurate system model, and has a strong ability to cope with unknown dynamics and external disturbances.
[0017] Considering the practical engineering applications, the observation error of the observer cannot converge to 0. This invention improves upon the previous intermittent control convergence theorem and proves that, under the intermittent output information rule, the observation error of the pitch angle converges to a bounded region.
[0018] Compared to previous intermittent information analysis processes, the intermittent information requirements in this application are triggered holistically, without complexly characterizing each intermittent cycle. This reduces the demands on intermittent information and broadens the application scenarios of the intermittent output switching state observer. Furthermore, in stability analysis, the reduced constraints on intermittent information make the overall design framework more aligned with practical engineering applications. Attached Figure Description
[0019] Figure 1Flowchart of a time-varying neuron adaptive control method for intermittent output wind turbine pitch angle.
[0020] Figure 2 The graph shows the time-varying function curves of neurons under different parameters. Detailed Implementation
[0021] Figures 1-2 This is the preferred embodiment of the present invention, which is described below in conjunction with the accompanying drawings. Figures 1-2 The present invention will be further described below.
[0022] like Figure 1 As shown, the time-varying neuron adaptive control method for intermittently outputting wind turbine pitch angle includes the following steps: Step 1: Establish a model of an intermittent wind turbine; Based on the structure of wind turbines known in the art, state variables are defined. and (in (where the pitch angle is the angle between the blades), the wind turbine pitch angle system is converted into a second-order nonlinear system in the following state-space form: in, The nominal nonlinear dynamic function of the system is known. , Let be the torque constant of the pitch motor. The equivalent rotational inertia of the pitch system, is the equivalent viscous friction coefficient of the pitch bearing and drive train. This refers to the input voltage or current control quantity. This is a lumped composite disturbance term.
[0023] This application considers the problem of intermittent output information, and thus defines Time period The output information cannot measure the start time, among which... , .
[0024] definition Let this be the unmeasurable duration of the output. Define two sets as follows: and .
[0025] Intermittent output information is represented as: (2) The intermittent output information studied in this application meets the following conditions: There are two constants and , so that: (3) in, for The Lebesgue measure.
[0026] Step 2: Determine the control objectives for the wind turbine model; For the intermittent wind turbine model determined in step 1, the following control objectives are to be achieved: (1) Ensure all signals are bounded; (2) Under the condition of using only intermittent output information, the actual trajectory of the pitch angle. Able to track the desired trajectory Expected trajectory satisfy ,in, It is a normal number.
[0027] To achieve the above objectives, this application identifies the following lemma.
[0028] Lemma 1: Assumption It is a continuous unknown function.
[0029] in, It is a compact set The vectors in the matrix. Therefore, there exists a target weight matrix as... The RBFNN makes ,in, It is a weight vector.
[0030] It is the estimation error, which satisfies . is a vector of radial basis functions, where , . No. The Gaussian kernel function of each neuron is: And the width of the basis functions is .
[0031] Activation function It is bounded and satisfies Lipschitz continuity, that is... and .
[0032] Lemma 2: Let It is a positive definite function, defined by two sets. as well as The intersection is an empty set, and the union is an empty set. .
[0033] If there are two positive numbers and , and Bounded, satisfying: when hour, ;when hour, ,but ,in, .
[0034] Step 3: Construct a state observer for the intermittent output-switching structure of the wind turbine model; For the unknown dynamics that arise in the intermittent wind turbine model determined in step 1 Using an RBF neural network to estimate it, and combining it with Lemma 1 determined in step 2, we know that: (4) Among them, approximation error satisfy Expected weight satisfy , These are the basis functions of the RBF neural network.
[0035] Using formula (4), the intermittent output-switching structure state observer is designed as follows: (5) in, State variables The estimated value, State variables The estimated value, and They are respectively and The first derivative, and For the time-varying correction gain of the observer, Expected weight The estimated value, , The input is represented as neural network activation functions, , Let be the torque constant of the pitch motor. The equivalent rotational inertia of the pitch system, This refers to the input voltage or current control quantity. Indicates time, Indicates the output error. Time period The output information cannot measure the start time.
[0036] The core principle of a state observer is to utilize the output error. Constructing nonlinear feedback forces the estimated value and Approaching the true value and . The observer's dynamic gain is expressed as: (6) in, These represent the maximum and minimum values of the state observer gain, respectively. This is the error sensitivity parameter.
[0037] Step 4, Radial basis function neural network estimation based on time-varying neurons; Based on Lemma 1 determined in step 2, the time-varying neural network estimation function is defined. for: (7) in, Indicates the number of neurons. Indicates the first The weights of each neuron, Indicates the first The activation function corresponding to each neuron.
[0038] Unlike basic RBF neural networks, the number of neurons in the time-varying neuron method is no longer a constant. The neuron variation function is defined as... ,in, For the floor function, is a time-varying function of the neuron.
[0039] Time-varying function of a neuron Defined as: (8) in, This is the upper bound of the time-varying function of the neuron. The rate of change is used to regulate the growth rate of the number of neurons; an example curve is shown below. Figure 2 As shown, Figure 2 Curve a in the middle is , The curve under the given conditions; curve b is , The curve under the given conditions; curve c is , Curve under certain conditions.
[0040] The beginning time is At that time, the number of neurons was That is, in the initial stage, there are One neuron is involved in estimating an unknown nonlinear function. Represented as the number of neurons from become At that time, that is ,in, Since function (8) is bounded, the number of neurons can increase to a maximum of [number missing]. One. Within a time period Internally, vector projection techniques are used to reconstruct the neural network approximator: (9) in, , , . Represents the weight vector With basis function vectors The angle between them. Therefore, the optimal neural network approximator with time-varying neurons is: (10) in, To minimize the estimation error, and satisfy the following: , It is a positive integer. Define the optimal weight. for: (11) The time-varying function of a neuron is selected generally to meet the following conditions: (1) the time-varying function must be a bounded function; (2) the time-varying function has a non-decreasing property.
[0041] This time-varying function has the following advantages: (1) Boundedness avoids the surge in computational load and excessive memory usage caused by the unlimited expansion of the number of neurons, and also prevents redundant and ineffective neurons from increasing system redundancy, ensuring that the computational burden is always within a controllable range; (2) Non-decreasingness avoids the repeated oscillation of neural network parameters during training, ensuring a smoother network convergence process and reducing the risk of estimation divergence caused by parameter mutations.
[0042] Step 5: Construct a time-varying neuron adaptive trajectory tracking controller; Define coordinate transformation ,in, This refers to the position error (tracking error). This refers to speed error (virtual control error). For virtual control signals, This is the pitch angle tracking signal.
[0043] The time-varying neuron adaptive continuous feedback control law is designed as follows: (12) in, This is a continuously time-varying neuron adaptive trajectory tracking control law. and To control the gain, For the desired trajectory The first derivative, The number of neurons is The corresponding activation function, The weighted projection estimate, for The renewal law, Indicates in time Weight projection estimate, adaptive law parameters , , , Let be the torque constant of the pitch motor. This is the equivalent rotational inertia of the pitch system.
[0044] According to formula (11), the time-varying neuron neural network can be obtained (13) in, For ideal weights, The activation function corresponding to the weight. estimation error, The nominal nonlinear dynamic function of the system is known. For lumped composite disturbance terms, This is a virtual control signal.
[0045] The estimation error of a time-varying neuron neural network is defined as: ,in, for Estimated value.
[0046] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.
Claims
1. A time-varying neuron adaptive control method for intermittently outputting wind turbine blade pitch angle, characterized in that: Includes the following steps: Step 1: Establish a model of an intermittent wind turbine; Step 2: Based on the intermittent wind turbine model determined in Step 1, determine its control objective and lemma; Step 3: Construct a state observer for the intermittent output-switching structure of the wind turbine model; Step 4, Radial basis function neural network estimation based on time-varying neurons; Step 5: Construct a time-varying neuron adaptive trajectory tracking controller.
2. The time-varying neuron adaptive control method for intermittently outputting wind turbine pitch angle according to claim 1, characterized in that: In step 1: Define state variables and The wind turbine pitch angle system is transformed into a second-order nonlinear system in the following state-space form, i.e., the expression for the intermittent wind turbine model is: in, The nominal nonlinear dynamic function of the system is known. , Let be the torque constant of the pitch motor. The equivalent rotational inertia of the pitch system, is the equivalent viscous friction coefficient of the pitch bearing and drive train. This refers to the input voltage or current control quantity. For lumped composite disturbance terms, It is the propeller pitch angle.
3. The time-varying neuron adaptive control method for intermittently outputting wind turbine pitch angle according to claim 1, characterized in that: In step 3: This addresses the unknown dynamics that arise in the intermittent wind turbine model identified in step 1. The following results were obtained by estimating it using an RBF neural network: Among them, approximation error satisfy Expected weight satisfy , These are the basis functions of the RBF neural network; The intermittent output-switching structure state observer is designed as follows: in, State variables The estimated value, State variables The estimated value, and They are respectively and The first derivative, and For the time-varying correction gain of the observer, Expected weight The estimated value, , The input is represented as neural network activation functions, , Let be the torque constant of the pitch motor. The equivalent rotational inertia of the pitch system, This refers to the input voltage or current control quantity. Indicates time, Indicates the output error. Time period The output information cannot measure the start time.
4. The time-varying neuron adaptive control method for intermittently outputting wind turbine pitch angle according to claim 1, characterized in that: In step 4: Define the time-varying neural network estimation function. for: in, Indicates the number of neurons. Indicates the first The weights of each neuron, Indicates the first The activation function corresponding to each neuron Let be the neuron change function, defined as ,in, For the floor function, Let be the time-varying function of the neuron, defined as: in, This is the upper bound of the time-varying function of the neuron. The rate of change.
5. The time-varying neuron adaptive control method for intermittently outputting wind turbine pitch angle according to claim 4, characterized in that: In step 4: the neural network approximator is reconstructed using vector projection techniques. in, , , , Represents the weight vector With basis function vectors The angle between them, the optimal neural network approximator with time-varying neurons is: in, To minimize the estimation error, and satisfy the following: , For positive integers, The optimal weight is determined by these factors.
6. The time-varying neuron adaptive control method for intermittently outputting wind turbine pitch angle according to claim 4, characterized in that: In step 5: Define coordinate transformation , in, For positional error, For speed error, For virtual control signals, For pitch angle tracking signal, State variables The estimated value, State variables The estimated value; The time-varying neuron adaptive continuous feedback control law is designed as follows: in, This is a continuously time-varying neuron adaptive trajectory tracking control law. and To control the gain, For the desired trajectory The first derivative, The number of neurons is The corresponding activation function, The weighted projection estimate, for The renewal law, Indicates in time Weight projection estimate, adaptive law parameters , , , Let be the torque constant of the pitch motor. The equivalent rotational inertia of the pitch system; The time-varying neuronal neural network was further identified as follows: in, For ideal weights, The activation function corresponding to the weight. estimation error, The nominal nonlinear dynamic function of the system is known. For lumped composite disturbance terms, This is a virtual control signal.