Fan nonlinear load adaptive control method and device and storage medium

Abstract

The invention discloses a fan nonlinear load self-adaptive control method and device and a storage medium. The fan nonlinear load self-adaptive control method and device are used for efficiently achieving dynamic tracking control over fan nonlinear loads. The method comprises the steps of collecting operation state data of a fan system; constructing a dynamic neural network model containing a long short-term memory network-attention mechanism mixed structure, inputting the operation state data into the dynamic neural network model, and outputting a dynamic predicted value of the fan nonlinear load; constructing an adaptive model prediction controller based on the dynamic prediction value, and generating a control sequence; calculating a prediction residual error of the dynamic neural network model, and when the prediction residual error exceeds a standard threshold value, triggering a model updating mechanism; performing parameter updating on the dynamic neural network model by adopting an incremental learning algorithm with gradient constraint to obtain updated parameters; and the safety control quantity is output to a fan execution mechanism, and dynamic tracking control over the nonlinear load of the fan is achieved.

Description

Technical Field

[0001] This application relates to the field of wind turbine control technology, and in particular to a wind turbine nonlinear load adaptive control method, device and storage medium. Background Technology

[0002] With the surge in global demand for clean energy, wind power, as a crucial component of renewable energy, is playing an increasingly vital role in the power supply system. Wind power systems operate in complex and variable environments, and the nonlinear load on wind turbines is highly dynamic and uncertain due to factors such as wind speed fluctuations and pitch angle adjustments. Accurately predicting and effectively controlling the nonlinear load on wind turbines to ensure stable operation, improve power generation efficiency, and enhance system safety has become a critical technical challenge that urgently needs to be addressed in the wind power industry.

[0003] In existing technologies, wind turbine load control mostly employs traditional linear model predictive control methods. This involves establishing a simplified mathematical model of wind turbine operation and combining it with historical data for load prediction and control. Some studies have introduced artificial neural networks, utilizing their powerful nonlinear mapping capabilities to process wind turbine operating data and achieve load prediction.

[0004] However, existing wind turbine load control technologies suffer from insufficient processing capabilities for complex dynamic data. Traditional linear models struggle to accurately describe the dynamic characteristics of nonlinear wind turbine loads, and artificial neural networks lack the ability to deeply mine key data features. This results in low wind turbine load prediction accuracy and slow control response, failing to meet the requirements of modern wind power systems for efficient and stable operation, and severely impacting the reliability and economy of wind power generation. Summary of the Invention

[0005] This application discloses a method, device, and storage medium for adaptive control of nonlinear loads of wind turbines, which can efficiently realize dynamic tracking control of nonlinear loads of wind turbines.

[0006] The first aspect of this application discloses a nonlinear load adaptive control method for wind turbines, including: Real-time acquisition of wind turbine system operating status data, including wind speed, blade pitch angle, generator speed, bus voltage, and load torque; A dynamic neural network model with a hybrid structure of long short-term memory network and attention mechanism is constructed. The operating status data is input into the dynamic neural network model, and the dynamic predicted value of the nonlinear load of the wind turbine is output. An adaptive model predictive controller is constructed based on the dynamic predicted values, and a control sequence is generated. The prediction residual of the dynamic neural network model is calculated in real time, and a model update mechanism is triggered when the prediction residual exceeds a standard threshold. An incremental learning algorithm with gradient constraints is used to update the parameters of the dynamic neural network model to obtain updated parameters; The updated parameters are input into the adaptive model prediction controller, and the dynamic prediction values ​​are normalized by the Sigmoid function to obtain the safety control quantity. The safety control quantity is output to the wind turbine actuator to achieve dynamic tracking control of the wind turbine's nonlinear load.

[0007] Optionally, the step of constructing an adaptive model predictive controller based on the dynamic predicted values ​​and generating a control sequence includes: By utilizing the output of a dynamic neural network model and calculating the Jacobian matrix using the central difference method, the nonlinear wind turbine system is approximately transformed into a linear state-space model. Based on the linear state-space model, an MPC prediction model is constructed. Construct constraints and set safety boundaries for the dynamic predicted values ​​based on the constraints; Based on the MPC prediction model and the safety boundary, a control sequence covering the prediction time domain and the control time domain is generated through a quadratic programming algorithm.

[0008] Optionally, the prediction residual of the dynamic neural network model is calculated in real time. When the prediction residual exceeds a standard threshold, a model update mechanism is triggered, including: Calculate the residual vector between the dynamically predicted value and the operating status data; Calculate the standard deviation of the residuals based on historical residual data within the sliding time window; If the norm of the current residual exceeds the standard threshold, the confidence of the dynamic neural network model is determined to have decreased, triggering the model update mechanism.

[0009] Optionally, the step of using an incremental learning algorithm with gradient constraints to update the parameters of the dynamic neural network model and obtain updated parameters includes: Based on the running status data and the dynamic prediction value, the gradient of the loss function with respect to all parameters of the model is calculated. The loss function includes a prediction error term, a sensitivity regularization term, and an L2 regularization term. Apply constraints to the calculated gradient to obtain the constrained gradient; The gradient descent algorithm with momentum is used to iteratively update the model parameters using the constraint gradient, and the updated parameters are output.

[0010] Optionally, after employing the gradient descent algorithm with momentum, iteratively updating the model parameters using the constraint gradient, and outputting the updated parameters, the method further includes: A norm constraint is applied to the updated parameters to ensure that the overall change in the updated parameters does not exceed a preset threshold, which is set according to the dynamic characteristics of the wind turbine system.

[0011] Optionally, the expression for the Sigmoid function is:

[0012] Where M is the upper limit of the physical quantity, and k takes values ​​from 5 to 10. This is the original output of the neural network.

[0013] Optionally, the triggering conditions for the model update mechanism may also include: When the wind turbine's operating status switches from the rated power zone to the maximum wind energy capture zone, a model update is forcibly triggered.

[0014] A second aspect of this application provides a wind turbine nonlinear load adaptive control device, comprising: The data acquisition unit is used to acquire real-time operating status data of the wind turbine system, including wind speed, blade pitch angle, generator speed, bus voltage, and load torque. The construction unit is used to construct a dynamic neural network model containing a hybrid structure of long short-term memory network and attention mechanism, input the operating status data into the dynamic neural network model, and output the dynamic prediction value of the nonlinear load of the wind turbine. A generation unit is used to construct an adaptive model prediction controller based on the dynamic prediction values ​​and generate a control sequence. The triggering unit is used to calculate the prediction residual of the dynamic neural network model in real time, and trigger the model update mechanism when the prediction residual exceeds the standard threshold. The update unit is used to update the parameters of the dynamic neural network model using an incremental learning algorithm with gradient constraints to obtain updated parameters; The input unit is used to input the updated parameters into the adaptive model prediction controller, and to normalize the dynamic prediction values ​​using the Sigmoid function to obtain the safety control quantity. The implementation unit is used to output the safety control quantity to the wind turbine actuator to realize dynamic tracking control of the nonlinear load of the wind turbine.

[0015] A third aspect of this application provides a wind turbine nonlinear load adaptive control device, comprising: Processor, memory, input / output units, and bus; The processor is connected to memory, input / output units, and a bus; The memory holds a program, which the processor calls to execute, as in the first aspect and any optional method of the first aspect.

[0016] The fourth aspect of this application provides a computer-readable storage medium on which a program is stored, which, when executed on a computer, performs the methods of the first aspect and any optional methods of the first aspect.

[0017] As can be seen from the above technical solutions, the embodiments of this application have the following advantages: This application enhances the processing capability for complex dynamic data by constructing a dynamic neural network model with a hybrid structure of Long Short-Term Memory (LSTM) network and attention mechanism. The combination of LSM and attention mechanism enables the model to deeply mine key features in the operating state data, significantly improving the prediction accuracy of wind turbine nonlinear loads. An adaptive model predictive controller built based on dynamic prediction values, combined with real-time prediction residual monitoring and an incremental learning algorithm with gradient constraints, achieves dynamic updating of model parameters and adaptive adjustment of control strategies, effectively solving the problem of slow control response and enabling the system to quickly adapt to complex and changing operating conditions. Furthermore, the use of the Sigmoid function to normalize the dynamic prediction values ​​to obtain safe control quantities further ensures the stability and safety of the control process, improves the reliability and economy of the wind power generation system, and meets the stringent requirements of modern wind power systems for efficient and stable operation. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of an embodiment of the wind turbine nonlinear load adaptive control method of this application; Figure 2 This is a schematic diagram of an embodiment of a method for generating control sequences according to this application; Figure 3 This is a schematic diagram of an embodiment of a method for triggering a model update mechanism according to this application; Figure 4 This is a schematic diagram of an embodiment of a method for obtaining updated parameters according to this application; Figure 5 This is a schematic diagram of an embodiment of a wind turbine nonlinear load adaptive control device according to this application; Figure 6 This is a schematic diagram of another embodiment of the wind turbine nonlinear load adaptive control device of this application. Detailed Implementation

[0020] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0021] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0022] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0023] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0024] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0025] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not an embodiment," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0026] Based on this, this application discloses a wind turbine nonlinear load adaptive control method, device and storage medium, which can efficiently realize dynamic tracking control of wind turbine nonlinear load.

[0027] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0028] The method of this application can be applied to servers, devices, terminals, or other devices with logical processing capabilities; therefore, this application does not limit its application. For ease of description, the following description uses a system as the executing entity.

[0029] Please see Figure 1 This application provides an embodiment of a wind turbine nonlinear load adaptive control method, comprising: 101. Real-time acquisition of wind turbine system operating status data, including wind speed, blade pitch angle, generator speed, bus voltage, and load torque; 102. Construct a dynamic neural network model with a hybrid structure of long short-term memory network and attention mechanism, input the operating status data into the dynamic neural network model, and output the dynamic prediction value of the nonlinear load of the wind turbine. When constructing the dynamic neural network model, the input layer sequentially incorporates operating status data such as wind speed, pitch angle, rotational speed, and electrical parameters at time step t. The hidden layer first extracts temporal features through a 32-unit LSTM layer, then mines key information using Q / K / V transform and multi-head attention mechanisms, followed by a fully connected layer containing 16 neurons, ReLU activation, and drop. The output layer outputs predicted load dynamic prediction values ​​such as predicted torque and rotational speed change rate. The preprocessed operating status data is input into the model, passed through forward propagation, and processed collaboratively by the components of the hidden layer, ultimately outputting the dynamic prediction result of the wind turbine's nonlinear load. The calculation formula is as follows:

[0030] in, Let be the LSTM hidden state vector, with a dimension of 32, representing the internal state representation at time t; The sigmoid function maps the input to the (0,1) interval; This is the weight matrix input to the hidden layer, with dimensions of 32×5, connecting the input variables and the hidden state; The input vector at time t contains 5 elements (wind speed, blade pitch angle, rotational speed, voltage, and load torque). The weight matrix from hidden layer to hidden layer is 32×32, and the temporal dependency relationship is established. This is the hidden state vector from the previous time step, with a dimension of 32; This is the hidden layer bias vector, with a dimension of 32, and the activation threshold is adjusted accordingly. The context vector after the attention mechanism is weighted, with a dimension of 32; This represents the summation over the past T time steps, where T=10 represents the size of the time window; Here, is the attention weight coefficient, representing the degree of attention paid to historical time k at time t; is the dot product of the hidden state vectors, used to measure the similarity between the current state and the historical states; exp is the exponential function that converts the similarity into a non-negative value. As a normalization factor, it ensures that the sum of all weights is 1; The model predicts the output vector, which contains three elements (predicted torque, rate of change of speed, and power fluctuation). This is the weight matrix for the first fully connected layer, with a dimension of 16×32; is the bias vector of the first fully connected layer, with a dimension of 16; ReLU is the modified linear unit activation function, used to increase non-linear expressiveness; This is the weight matrix for the second fully connected layer, with a dimension of 3×16; This is the bias vector for the second fully connected layer, with a dimension of 3.

[0031] 103. Construct an adaptive model predictive controller based on dynamic predicted values ​​and generate control sequences; When constructing an adaptive model predictive controller, the dynamic prediction value of the nonlinear load of the wind turbine is first used as the output of the dynamic neural network. Combined with the physical constraints of the wind turbine actuator, a rolling optimization objective function with 5 steps in the prediction time domain and 3 steps in the control time domain is established. The goal is to minimize the deviation between the predicted load and the expected load. A quadratic programming algorithm is introduced to solve the problem.

[0032] The dynamic predicted values ​​are input into the controller in time series. Through online rolling optimization, control sequences such as pitch angle adjustment and torque control are generated in the corresponding control time domain. At the same time, the built-in model adaptive module of the controller is used to dynamically adjust the model parameters based on the real-time prediction residuals to ensure that the control sequence can effectively track the nonlinear load changes of the wind turbine.

[0033] 104. Calculate the prediction residuals of the dynamic neural network model in real time, and trigger the model update mechanism when the prediction residuals exceed the standard threshold; In practical applications, when calculating the prediction residuals of the dynamic neural network model in real time, the residual data is obtained by subtracting the dynamically predicted value of the wind turbine's nonlinear load output by the model from the actual load value fed back by the wind turbine actuator. The standard threshold is set according to the wind turbine model, operating conditions, and safety requirements. When the residuals for three consecutive sampling periods exceed this threshold, or when a single residual exceeds the threshold by 1.5 times, the model update mechanism is immediately triggered. This initiates an incremental learning algorithm with gradient constraints to update the parameters of the dynamic neural network model, ensuring that the model's prediction accuracy adapts to changes in operating conditions. The formula for the triggering mechanism is as follows:

[0034] in, The standard deviation of the residuals for the most recent 100 samples. The actual measured output value of the system. To predict residuals for the model, To predict residuals.

[0035] 105. Use an incremental learning algorithm with gradient constraints to update the parameters of a dynamic neural network model and obtain the updated parameters; 106. Input the updated parameters into the adaptive model predictive controller, and normalize the dynamic predicted values ​​using the Sigmoid function to obtain the safety control quantity; The model parameters, updated by the incremental learning algorithm with gradient constraints, are fed into the adaptive model prediction controller in real time. The controller calls the Sigmoid function to normalize the dynamic prediction values ​​of the wind turbine nonlinear load output by the dynamic neural network. The expression for the Sigmoid function is:

[0036] Where M is the upper limit of the physical quantity, and k takes values ​​from 5 to 10. This is the original output of the neural network.

[0037] This function maps the predicted values ​​to the target range, and, combined with the physical constraints of the wind turbine actuator, converts them into safety control quantities.

[0038] 107. Output safety control quantities to the fan actuator to achieve dynamic tracking control of the fan's nonlinear load.

[0039] After the safety control quantities are calculated, they are transmitted to the wind turbine actuators in real time via industrial Ethernet or fieldbus communication protocols. For pitch actuators, the control quantities are converted from digital to analog and used to drive servo motors to adjust the blade angle, resulting in reasonable response time control. For converters, the control signals directly participate in the PWM modulation of the power module to regulate the output torque.

[0040] During dynamic tracking control, the actual action parameters fed back by the actuator are collected in real time and compared with the safety control quantity. If the deviation exceeds the rated value, the feedback compensation mechanism is triggered to correct the subsequent control output, ensuring the dynamic tracking accuracy of the nonlinear load of the wind turbine, reducing the load fluctuation amplitude to within ±2%, and maintaining the stable and efficient operation of the wind turbine.

[0041] In this embodiment, a dynamic neural network model with a hybrid structure of long short-term memory network and attention mechanism is constructed to enhance the processing capability of complex dynamic data. The combination of long short-term memory network and attention mechanism enables the model to deeply mine key features in the operating state data, significantly improving the prediction accuracy of wind turbine nonlinear load. The adaptive model predictive controller built based on dynamic prediction values, combined with real-time prediction residual monitoring and incremental learning algorithm with gradient constraints, realizes dynamic updating of model parameters and adaptive adjustment of control strategy, effectively solving the problem of slow control response and enabling the system to quickly adapt to complex and changing operating conditions. In addition, the sigmoid function is used to normalize the dynamic prediction values ​​to obtain the safety control quantity, further ensuring the stability and safety of the control process, improving the reliability and economy of wind power generation system operation, and meeting the stringent requirements of modern wind power systems for efficient and stable operation.

[0042] Please see Figure 2 This application provides an embodiment of a method for generating control sequences, comprising: 201. Using the output of the dynamic neural network model, the Jacobian matrix is ​​calculated by the central difference method, and the nonlinear wind turbine system is approximately transformed into a linear state-space model. 202. Construct an MPC prediction model based on a linear state-space model; 203. Construct constraints and set safety boundaries for dynamic predicted values ​​based on the constraints; 204. Based on the MPC prediction model and safety boundary, a control sequence covering the prediction time domain and the control time domain is generated through a quadratic programming algorithm.

[0043] Step 201 first obtains the wind turbine load prediction result output by the dynamic neural network model, which reflects the nonlinear mapping relationship between the system input and output. Then, the central difference method is used to calculate the Jacobian matrix by applying small perturbations to the input variables and observing the output changes. The Jacobian matrix reflects the local linearity of the system at a specific operating point. Based on the Jacobian matrix, the originally complex nonlinear wind turbine system is approximately transformed into a linear state-space model. The linear state-space model describes the system's dynamic characteristics using state equations and output equations.

[0044] The formula for online linearization is:

[0045] in, This represents the predicted system output value at time k+1; This represents the predicted output of the neural network under the current state and control. This is the Jacobian matrix, representing the sensitivity of the output to the control input. This represents the partial derivative of the nth output with respect to the mth control input; n represents the number of output variables, which is 3 here; M represents the number of control variables, which is 2 here (pitch angle and torque control). This represents the control increment vector.

[0046] In steps 202 and 204, the MPC prediction model is constructed, and the formula for generating control sequences covering the prediction and control time domains using a quadratic programming algorithm is as follows:

[0047] in, This represents finding the minimum value of the control increment sequence Δu; This indicates the length of the prediction time domain; a value of 15 indicates the number of steps in the future prediction. This represents the predicted power output at step k; Indicates the power reference value (target value); This represents the weighted squared error between the output and the reference value; Q represents the output error weight matrix, a diagonal matrix diag(10,1,0.5), emphasizing the importance of different outputs; This indicates the length of the control time domain; a value of 5 indicates the number of control steps to be optimized. This represents the control increment at step k; R represents the weighted square norm of the control increment; R represents the control cost weight matrix with a value of 0.01I, which suppresses drastic changes in control actions.

[0048] The constraints in step 203 are:

[0049] in, This represents the rate of change of pitch angle, in degrees per second, and is constrained to ±5° / s. This represents the predicted load torque, in kN·m. and This indicates the lower and upper limits of the load torque, typically ranging from 0% to 120% of the rated value. Indicates the rotor speed; and This indicates the lower and upper limits of the rotational speed, typically 70% to 120% of the rated value. The nonlinear constraint function representing the predicted output of the neural network; This represents the predicted output of the neural network at step k.

[0050] Constraints are incorporated into the model in mathematical form to provide a basis for the generation of control sequences.

[0051] In this embodiment, the Jacobian matrix is ​​calculated using the central difference method, transforming the complex nonlinear wind turbine system into a linear state-space model. This preserves the key dynamic characteristics of the system while significantly reducing the complexity of the control algorithm, enabling model predictive control to be efficiently applied to wind turbine control scenarios and improving the feasibility and efficiency of control system design. Based on the linear state-space model, an MPC predictive model is constructed, combined with the high-precision predictive capabilities of dynamic neural networks, to achieve multi-step prediction and rolling optimization of wind turbine load changes. This significantly improves control response speed and tracking accuracy, especially under dynamic scenarios such as sudden wind speed changes and load fluctuations. By constructing physical constraints and safety boundaries, the limitations of wind turbine actuators (such as pitch angle range and torque change rate) and operational safety requirements (such as load fluctuation limits) are transformed into mathematical constraints, ensuring that the control sequence is always generated within a safe range. This effectively avoids the risk of exceeding limits and improves system reliability and equipment lifespan.

[0052] Please see Figure 3 This application provides an embodiment of a method for triggering a model update mechanism, comprising: 301. Calculate the residual vector between the dynamic predicted value and the operating status data; 302. Calculate the standard deviation of the residuals based on the historical residual data within the sliding time window; 303. If the norm of the current residual exceeds the standard threshold, the confidence of the dynamic neural network model is determined to have decreased, triggering the model update mechanism. 304. When the wind turbine's operating status switches from the rated power zone to the maximum wind energy capture zone, a model update is forcibly triggered.

[0053] In step 301, the system synchronously acquires load prediction data output by the dynamic neural network model and actual operating status data collected by sensors during real-time wind turbine operation. For each sampling moment, the difference between the predicted and actual values ​​is calculated point-by-point to form residual data. Specifically, if the model predicts the generator torque to be 950 kN·m, while the actual measured value is 980 kN·m, then the torque residual at that moment is -30 kN·m. To unify the residual measurement standards for different parameters, the residuals of each parameter are standardized to eliminate the influence of dimensions, resulting in directly comparable residual vectors.

[0054] In step 302, the system sets a sliding time window containing 100 sampling periods and continuously collects historical residual data within the window. For the residual sequence of each parameter, the statistical characteristics of the residual sequence fluctuation, i.e., the residual standard deviation, are calculated.

[0055] In step 303, the overall size of the residual vector at the current time, i.e., the residual norm, is calculated and compared with a standard threshold set based on the statistical characteristics of historical data. When the residual norm exceeds the standard threshold, the system determines that the model prediction confidence has decreased. The system automatically initiates a model update mechanism, employing an incremental learning algorithm with gradient constraints, to fine-tune the model parameters using the 500 most recently collected sample data, ensuring that the model continues to accurately predict load changes.

[0056] In step 304, a power threshold is preset in the wind turbine main control system. When the generator output power drops from the rated power range above the threshold to the maximum wind energy capture range below the threshold for more than 10 seconds, the system determines that the wind turbine operating state has switched. At this time, regardless of whether the residual exceeds the limit, the system will forcibly trigger a model update. During the update process, operating data similar to the current operating condition are extracted from the historical database first. An improved incremental learning algorithm is used to optimize the model, and constraints are introduced to prevent the model from over-adapting to new data and forgetting historical knowledge, ensuring that the model maintains good predictive performance under different operating states.

[0057] In this embodiment, by calculating the residual vector and standard deviation, real-time quantitative evaluation of the prediction accuracy of the dynamic neural network model is achieved, enabling timely detection of model performance degradation trends and providing a reliable basis for adaptive control. The calculation of residual standard deviation based on a sliding window allows the system to adapt to the model error characteristics under different operating conditions, maintaining stable control performance in scenarios such as sudden wind speed changes and load fluctuations. The dual triggering mechanism (residual norm exceeding limits and operating condition switching) avoids frequent and unnecessary model updates while ensuring timely model adjustments during critical operating condition changes, balancing control stability and computational efficiency. For the characteristic differences of different wind turbine operating areas (rated power range and maximum wind energy capture range), forced model updates are triggered, enabling the control system to maintain optimal performance across the entire wind speed range, improving wind energy conversion efficiency and system reliability.

[0058] Please see Figure 4 This application provides an embodiment of a method for obtaining updated parameters, comprising: 401. Based on the running status data and dynamic prediction values, calculate the gradient of the loss function with respect to all parameters of the model. The loss function includes the prediction error term, the sensitivity regularization term, and the L2 regularization term. 402. Apply constraints to the calculated gradient to obtain the constrained gradient; 403. Employ the gradient descent algorithm with momentum, use the constrained gradient to iteratively update the model parameters, and output the updated parameters; 404. Apply norm constraints to the updated parameters to ensure that the overall change in the updated parameters does not exceed the preset threshold. The preset threshold is set according to the dynamic characteristics of the wind turbine system.

[0059] In this embodiment, the formula for calculating the gradient of the loss function with respect to all parameters of the model is as follows:

[0060] Where L is the total loss function value, which is dimensionless and represents the training optimization objective; This is the weighting coefficient for the prediction error term, with a typical value of 0.6-0.8, balancing the importance of each term. To sum over the three output dimensions; The true label value is the actual measured value of the i-th output variable. The model predicts the value of the i-th output variable. This is the weighting coefficient for the sensitivity regularization term, with a value range of 0.05-0.2, and a suggested value of 0.1 in this paper; The partial derivatives of the output with respect to the control quantity (Jacobi matrix elements) represent the sensitivity; The control increment vector has the same dimension as the control input. This is the L2 regularization coefficient, typically 0.001-0.01, to prevent overfitting; This is the set of all parameters of the model, including all weights W and biases b; It is the square of the L2 norm of the parameters, that is, the sum of the squares of all parameters.

[0061] During the operation of the wind turbine control system, to ensure the model continuously and accurately predicts nonlinear loads, the parameters of the dynamic neural network model need to be updated. First, based on real-time collected operating status data and the model's dynamically predicted output values, the gradient of the loss function with respect to all parameters is calculated. The loss function comprehensively considers prediction error terms, sensitivity regularization terms, and L2 regularization terms to ensure model accuracy and generalization ability. Next, constraints are applied to the calculated gradients to limit the direction and magnitude of parameter updates, preventing drastic fluctuations that could lead to system instability. Then, a gradient descent algorithm with momentum is used to iteratively update the parameters using the constrained gradients, accelerating convergence and reducing oscillations by accumulating historical gradient information. Finally, a norm constraint is applied to the overall change magnitude of the updated parameters. This threshold is set according to the dynamic characteristics of the wind turbine system to ensure smooth parameter updates and maintain system control stability. Through this series of operations, the model parameters are updated safely and efficiently, enabling the control system to continuously adapt to changes in the wind turbine's operating status.

[0062] In this embodiment, the loss function, which comprehensively considers the prediction error term, sensitivity regularization term, and L2 regularization term, can improve model accuracy and generalization ability while reducing overfitting. Gradient constraints and norm limitations prevent drastic parameter fluctuations, ensuring stable system operation. The gradient descent algorithm with dynamic parameters accelerates convergence, reduces iterative oscillations, and improves parameter update efficiency. Overall, the model can quickly adapt to changes in wind turbine operating conditions, maintain high-precision predictions, enhance the robustness and reliability of the control system, and ensure efficient and stable wind turbine operation.

[0063] Please see Figure 5 This application provides an embodiment of a wind turbine nonlinear load adaptive control device, comprising: The acquisition unit 501 is used to acquire real-time operating status data of the wind turbine system, including wind speed, blade pitch angle, generator speed, bus voltage, and load torque. The building unit 502 is used to build a dynamic neural network model containing a hybrid structure of long short-term memory network and attention mechanism. The operating status data is input into the dynamic neural network model and the dynamic predicted value of the nonlinear load of the wind turbine is output. The generation unit 503 is used to construct an adaptive model predictive controller based on dynamic predicted values ​​and generate a control sequence. Specifically, generation unit 503 is also used for: By utilizing the output of a dynamic neural network model and calculating the Jacobian matrix using the central difference method, the nonlinear wind turbine system is approximately transformed into a linear state-space model. An MPC prediction model is constructed based on a linear state-space model. Construct constraints and set safety boundaries for dynamic predictions based on these constraints; Based on the MPC prediction model and safety boundary, a quadratic programming algorithm is used to generate control sequences covering the prediction time domain and the control time domain.

[0064] Trigger unit 504 is used to calculate the prediction residual of the dynamic neural network model in real time. When the prediction residual exceeds the standard threshold, the model update mechanism is triggered. Specifically, trigger unit 504 is also used for: Calculate the residual vector between the dynamically predicted values ​​and the operational status data; Calculate the standard deviation of the residuals based on historical residual data within the sliding time window; If the norm of the current residual exceeds the standard threshold, the confidence of the dynamic neural network model is determined to have decreased, triggering the model update mechanism.

[0065] Specifically, trigger unit 504 is also used for: When the wind turbine's operating status switches from the rated power zone to the maximum wind energy capture zone, a model update is forcibly triggered.

[0066] The update unit 505 is used to update the parameters of the dynamic neural network model using an incremental learning algorithm with gradient constraints, and obtain the updated parameters. Specifically, update unit 505 is also used for: Based on the running status data and dynamic prediction values, the gradient of the loss function with respect to all parameters of the model is calculated. The loss function includes the prediction error term, the sensitivity regularization term, and the L2 regularization term. Apply constraints to the calculated gradient to obtain the constrained gradient; The gradient descent algorithm with momentum is used to iteratively update the model parameters using the constrained gradient and output the updated parameters.

[0067] Specifically, it also includes a limiting unit 506, used for: Norm constraints are applied to the updated parameters to ensure that the overall change in the updated parameters does not exceed a preset threshold, which is set according to the dynamic characteristics of the wind turbine system.

[0068] Input unit 507 is used to input updated parameters into the adaptive model prediction controller, and normalize the dynamic prediction values ​​through the Sigmoid function to obtain the safety control quantity; The implementation unit 508 is used to output safety control quantities to the fan actuator to realize dynamic tracking control of the nonlinear load of the fan.

[0069] For detailed implementation methods, please refer to... Figures 1 to 4 Examples are not detailed here.

[0070] Please see Figure 6 This application provides a wind turbine nonlinear load adaptive control device, comprising: Processor 601, memory 602, input / output unit 603, and bus 604.

[0071] The processor 601 is connected to the memory 602, the input / output unit 603, and the bus 604.

[0072] The memory 602 stores a program, and the processor 601 calls the program to execute it, such as... Figure 1 , Figure 2 , Figure 3 and Figure 4 The method in the middle.

[0073] This application provides a computer-readable storage medium on which a program is stored, and when the program is executed on a computer, it performs the following... Figure 1 , Figure 2 , Figure 3 and Figure 4 The method in the middle.

[0074] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0075] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between apparatuses or units through some interfaces, and may be electrical, mechanical, or other forms.

[0076] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0077] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0078] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A nonlinear load adaptive control method for a wind turbine, characterized in that, include: Real-time acquisition of wind turbine system operating status data, including wind speed, blade pitch angle, generator speed, bus voltage, and load torque; A dynamic neural network model with a hybrid structure of long short-term memory network and attention mechanism is constructed. The operating status data is input into the dynamic neural network model, and the dynamic predicted value of the nonlinear load of the wind turbine is output. An adaptive model predictive controller is constructed based on the dynamic predicted values, and a control sequence is generated. The prediction residual of the dynamic neural network model is calculated in real time, and a model update mechanism is triggered when the prediction residual exceeds a standard threshold. An incremental learning algorithm with gradient constraints is used to update the parameters of the dynamic neural network model to obtain updated parameters; The updated parameters are input into the adaptive model prediction controller, and the dynamic prediction values ​​are normalized by the Sigmoid function to obtain the safety control quantity. The safety control quantity is output to the wind turbine actuator to achieve dynamic tracking control of the wind turbine's nonlinear load.

2. The wind turbine nonlinear load adaptive control method according to claim 1, characterized in that, The step of constructing an adaptive model predictive controller based on the dynamic predicted values ​​and generating a control sequence includes: By utilizing the output of a dynamic neural network model and calculating the Jacobian matrix using the central difference method, the nonlinear wind turbine system is approximately transformed into a linear state-space model. Based on the linear state-space model, an MPC prediction model is constructed. Construct constraints and set safety boundaries for the dynamic predicted values ​​based on the constraints; Based on the MPC prediction model and the safety boundary, a control sequence covering the prediction time domain and the control time domain is generated through a quadratic programming algorithm.

3. The wind turbine nonlinear load adaptive control method according to claim 1, characterized in that, The prediction residual of the dynamic neural network model is calculated in real time. When the prediction residual exceeds a standard threshold, a model update mechanism is triggered, including: Calculate the residual vector between the dynamically predicted value and the operating status data; Calculate the standard deviation of the residuals based on historical residual data within the sliding time window; If the norm of the current residual exceeds the standard threshold, the confidence of the dynamic neural network model is determined to have decreased, triggering the model update mechanism.

4. The wind turbine nonlinear load adaptive control method according to claim 1, characterized in that, The incremental learning algorithm with gradient constraints is used to update the parameters of the dynamic neural network model to obtain updated parameters, including: Based on the running status data and the dynamic prediction value, the gradient of the loss function with respect to all parameters of the model is calculated. The loss function includes a prediction error term, a sensitivity regularization term, and an L2 regularization term. Apply constraints to the calculated gradient to obtain the constrained gradient; The gradient descent algorithm with momentum is used to iteratively update the model parameters using the constraint gradient, and the updated parameters are output.

5. The wind turbine nonlinear load adaptive control method according to claim 4, characterized in that, After employing the gradient descent algorithm with momentum, iteratively updating the model parameters using the constrained gradient, and outputting the updated parameters, the method further includes: A norm constraint is applied to the updated parameters to ensure that the overall change in the updated parameters does not exceed a preset threshold, which is set according to the dynamic characteristics of the wind turbine system.

6. The wind turbine nonlinear load adaptive control method according to claim 1, characterized in that, The expression for the Sigmoid function is: Where M is the upper limit of the physical quantity, and k takes values ​​from 5 to 10. This is the original output of the neural network.

7. The wind turbine nonlinear load adaptive control method according to claim 1, characterized in that, The triggering conditions for the model update mechanism also include: When the wind turbine's operating status switches from the rated power zone to the maximum wind energy capture zone, a model update is forcibly triggered.

8. A wind turbine nonlinear load adaptive control device, characterized in that, include: The data acquisition unit is used to acquire real-time operating status data of the wind turbine system, including wind speed, blade pitch angle, generator speed, bus voltage, and load torque. The construction unit is used to construct a dynamic neural network model containing a hybrid structure of long short-term memory network and attention mechanism, input the operating status data into the dynamic neural network model, and output the dynamic prediction value of the nonlinear load of the wind turbine. A generation unit is used to construct an adaptive model prediction controller based on the dynamic prediction values ​​and generate a control sequence. The triggering unit is used to calculate the prediction residual of the dynamic neural network model in real time, and trigger the model update mechanism when the prediction residual exceeds the standard threshold. The update unit is used to update the parameters of the dynamic neural network model using an incremental learning algorithm with gradient constraints to obtain updated parameters; The input unit is used to input the updated parameters into the adaptive model prediction controller, and to normalize the dynamic prediction values ​​using the Sigmoid function to obtain the safety control quantity. The implementation unit is used to output the safety control quantity to the wind turbine actuator to realize dynamic tracking control of the nonlinear load of the wind turbine.

9. A wind turbine nonlinear load adaptive control device, characterized in that, include: The system includes a processor, a memory, an input / output unit, and a bus, wherein the processor is connected to the memory, the input / output unit, and the bus, the memory stores a program, and the processor calls the program to execute the wind turbine nonlinear load adaptive control method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains a program that, when executed on a computer, performs the wind turbine nonlinear load adaptive control method as described in any one of claims 1 to 7.

Images