Fan control parameter online optimization method and system based on digital twinning

Abstract

The application discloses a fan control parameter online optimization method and system based on digital twinning, which inputs real-time SCADA data into a fast digital twinning model, performs millisecond-level state estimation to meet the real-time response requirement of the control system. At the same time, the slow digital twinning model is driven by buffer data for deep calibration, a high-fidelity state log containing physical mechanisms is generated, and the knowledge distillation technology is used to migrate the physical knowledge of the slow model to the fast model for periodic correction of the fast model. Finally, the modified fast model is used for rolling optimization under the model predictive control framework to output the optimal control sequence. The scheme effectively overcomes the defects of the traditional physical model that the calculation time is too long, and the pure data-driven model that is easy to drift and lacks physical interpretability, while ensuring the millisecond-level response speed, and improving the physical consistency of model prediction and the robustness of long-term operation.

Description

Technical Field

[0001] This application relates to the field of wind power generation control technology, specifically to a method and system for online optimization of wind turbine control parameters based on digital twins. Background Technology

[0002] As a crucial component of the clean energy sector, wind power operates in an environment characterized by high randomness and time-varying variations. With the continuous increase in the capacity of individual wind turbine units and the increasing flexibility of blades, the challenges faced by control systems are becoming increasingly complex. To maximize wind energy capture while ensuring the structural safety of the unit, wind turbine units need to be able to sense their current operating status and adjust control parameters in real time, such as pitch angle setpoints and generator electromagnetic torque. Digital twin technology, by constructing a digital mapping of physical entities in virtual space, can extrapolate and predict the internal state of the wind turbine, providing an effective technical means for online optimization of control parameters.

[0003] However, in existing online optimization technologies for wind turbine control parameters, the construction and application of digital twin models present a contradiction between real-time computation and model accuracy. On the one hand, while high-fidelity physical models based on blade element momentum theory or multibody dynamics offer high prediction accuracy and good physical interpretability, their computational processes are complex and time-consuming, making it difficult to meet the millisecond-level real-time response requirements of wind turbine control systems, resulting in a significant lag in online decision-making. On the other hand, while data-driven models based on algorithms such as neural networks offer extremely fast computation speeds and can meet the needs of real-time control, these models are typically considered black-box models, lacking deep physical constraints. During long-term operation, as wind turbine components age or environmental conditions change, data-driven models are prone to model drift, leading to decreased prediction accuracy and even generating erroneous control commands that violate physical laws. Consequently, the lack of physical consistency guarantees affects the safety and stability of unit operation.

[0004] Therefore, an optimized online optimization scheme for wind turbine control parameters based on digital twins is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a method and system for online optimization of wind turbine control parameters based on digital twins.

[0006] According to one aspect of this application, a method for online optimization of wind turbine control parameters based on digital twins is provided, comprising: Acquire real-time SCADA data and buffered SCADA data; Real-time SCADA data is input into a fast digital twin model to obtain a fast model state estimate; Based on buffered SCADA data, a minute-level deep calibration is performed on the slow digital twin model to obtain the calibrated slow model physical parameters and high-fidelity status logs. Based on high-fidelity status logs and buffered SCADA data, the fast digital twin model is periodically corrected using knowledge distillation to obtain the corrected fast digital twin model. Using the fast model state estimate as the initial state, it is input into the modified fast digital twin model to obtain the final optimal control sequence; The control parameters of the first time step are extracted from the final optimal control sequence as the control command.

[0007] According to another aspect of this application, a wind turbine control parameter online optimization system based on digital twin is provided, comprising: The data acquisition module is used to acquire real-time SCADA data and buffered SCADA data; The fast model state estimation module is used to input real-time SCADA data into the fast digital twin model to obtain the fast model state estimate; The deep calibration module is used to perform minute-level deep calibration on slow digital twin models based on buffered SCADA data to obtain the calibrated physical parameters and high-fidelity status logs of the slow model. The periodic correction module is used to perform knowledge distillation-based periodic correction on the fast digital twin model based on high-fidelity status logs and buffered SCADA data to obtain the corrected fast digital twin model. The optimal control sequence acquisition module is used to take the fast model state estimate as the initial state and input it into the modified fast digital twin model to obtain the final optimal control sequence. The control command generation module is used to extract the control parameters of the first time step from the final optimal control sequence as control commands.

[0008] Compared with existing technologies, this invention inputs real-time SCADA data into a fast digital twin model for millisecond-level state estimation to meet the real-time response requirements of the control system. Simultaneously, buffered data drives a slow digital twin model for deep calibration, generating a high-fidelity state log containing physical mechanisms. Knowledge distillation is then used to transfer the physical knowledge from the slow model to the fast model, periodically revising it. Finally, the revised fast model is used for rolling optimization within a model predictive control framework to output the optimal control sequence. This scheme effectively overcomes the shortcomings of traditional physical models, such as excessive computation time and the tendency for pure data-driven models to drift and lack physical interpretability. While ensuring millisecond-level response speed, it improves the physical consistency of model predictions and long-term robustness. Attached Figure Description

[0009] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 This is a flowchart of a method for online optimization of wind turbine control parameters based on digital twins according to an embodiment of this application; Figure 2 This is a schematic diagram of the data flow of the online optimization method for wind turbine control parameters based on digital twins according to an embodiment of this application; Figure 3 This is a flowchart illustrating the process of performing minute-level deep calibration on a slow digital twin model based on buffered SCADA data to obtain the calibrated slow model physical parameters and high-fidelity status log, according to the online optimization method for wind turbine control parameters based on digital twins in this application embodiment. Figure 4 The flowchart illustrates the process of using high-fidelity status logs and buffered SCADA data to periodically modify a fast digital twin model based on knowledge distillation to obtain a modified fast digital twin model, according to the online optimization method for wind turbine control parameters based on digital twins in this application. Figure 5 This is a flowchart illustrating the online optimization method for wind turbine control parameters based on digital twins according to embodiments of this application, which uses a fast model state estimate as the initial state and inputs it into a modified fast digital twin model to obtain the final optimal control sequence. Figure 6 This is a block diagram of a digital twin-based online optimization system for wind turbine control parameters according to an embodiment of this application. Detailed Implementation

[0011] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0012] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0013] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0014] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0015] Existing online optimization techniques for wind turbine control parameters face the dilemma of balancing computational real-time performance with model accuracy. While high-fidelity physical models offer strong interpretability, their computational complexity is prohibitive and fails to meet real-time control requirements. Conversely, lightweight data-driven models, despite their rapid response, are prone to model drift due to a lack of physical constraints, leading to unreliable control decisions. Therefore, this application proposes an online optimization method for wind turbine control parameters based on digital twins. Specifically, this scheme balances response speed and prediction accuracy through a strategy of "fast-slow coordination and periodic correction." First, a fast digital twin model based on deep learning is used to perform millisecond-level state estimation of real-time SCADA data. This estimate is then input into the model predictive control framework, rapidly generating the optimal control sequence that satisfies the constraints, ensuring the millisecond-level response capability of the control system. Simultaneously, a slow digital twin model based on physical mechanisms is run in parallel, using buffered data for minute-level deep parameter calibration, generating a high-fidelity state log containing physical laws. Building upon this foundation, a bridge is constructed between the fast and slow models using knowledge distillation technology. This involves periodically transferring the physical knowledge from the slow model to the fast model, fine-tuning and correcting the network weights of the fast model. This mechanism allows the fast model to maintain its computational speed advantage while effectively overcoming model drift and achieving physical consistency. This enables precise, robust, and real-time online optimization of wind turbine control parameters under complex operating conditions.

[0016] Figure 1 This is a flowchart of a method for online optimization of wind turbine control parameters based on digital twins, according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow in the online optimization method for wind turbine control parameters based on digital twins according to an embodiment of this application. Figure 1 and Figure 2As shown, the online optimization method for wind turbine control parameters based on digital twin according to an embodiment of this application includes the following steps: S100, acquiring real-time SCADA data and buffered SCADA data; S200, inputting the real-time SCADA data into a fast digital twin model to obtain a fast model state estimate; S300, performing minute-level deep calibration on a slow digital twin model based on the buffered SCADA data to obtain calibrated slow model physical parameters and a high-fidelity state log; S400, performing knowledge distillation-based periodic correction on the fast digital twin model based on the high-fidelity state log and buffered SCADA data to obtain a corrected fast digital twin model; S500, using the fast model state estimate as the initial state, inputting it into the corrected fast digital twin model to obtain the final optimal control sequence; S600, extracting the control parameters of the first time step from the final optimal control sequence as control commands.

[0017] Specifically, in step S100, real-time SCADA data and buffered SCADA data are acquired. It is understood that the online control optimization of wind turbines requires both millisecond-level real-time response to cope with transient wind conditions and long-term historical data to capture slow-dynamic physical characteristics and calibrate model biases. Therefore, in the technical solution of this application, the current operating status data of the wind turbine is collected in real time, and a rolling buffer containing historical data from specific past time periods is maintained simultaneously. This provides differentiated data support for the instantaneous state estimation of the fast digital twin model and the deep parameter calibration of the slow digital twin model, respectively. This ensures that the control system obtains the latest observations for rapid decision-making while having sufficient historical information to invert physical parameters, thus balancing the real-time performance of control and the high physical fidelity of the model within a single architecture.

[0018] More specifically, in a concrete example of this application, the data acquisition process is implemented by establishing an industrial communication connection with the wind turbine's main controller. First, raw signals uploaded from the wind turbine's sensor network are directly read via industrial standard communication protocols, such as OPC UA or Modbus TCP. These signals cover key operating indicators including wind speed, generator speed, active power, pitch angle, and tower vibration acceleration. Subsequently, the read data stream is processed in two paths. The first path, as real-time SCADA data, is immediately timestamped and transmitted in vector form to a high-speed cache in memory for millisecond-level reading and deduction by the fast digital twin model, ensuring low latency in control command generation. The second path of data is sequentially written to a pre-defined circular buffer or first-in-first-out queue, which continuously stores continuous time-series data from the past few minutes, such as the operating records of the past ten minutes. Whenever a new data frame arrives, the buffer automatically removes the oldest frame, thus maintaining a constant time window length for buffering SCADA data. In addition, before storing the data in the buffer, the raw data is cleaned to remove null values ​​or outliers that deviate significantly from physical limits due to sensor malfunctions, ensuring the integrity and validity of the historical data used for subsequent slow digital twin model calibration.

[0019] Specifically, in step S200, real-time SCADA data is input into a fast digital twin model to obtain a fast model state estimate. It is understood that, because wind turbines operate in rapidly changing wind farm environments, their critical internal states, such as blade root loads or aerodynamic torques, are crucial for developing safe and efficient control strategies. However, these states are often difficult to measure directly in real-time using physical sensors, and online inversion calculations using complex physical mechanism models have significant time lags, failing to meet the stringent requirements of millisecond-level response in control systems. Therefore, in the technical solution of this application, real-time SCADA data, including various observations such as wind speed, rotational speed, and power, is further input into a pre-trained lightweight fast digital twin model. This model utilizes the efficient inference capabilities of neural networks to perform forward calculations, thereby enabling real-time estimation of the internal aerodynamic and structural states of the wind turbine, which are currently difficult to directly observe, within a very short time window. In this way, high-frequency and critical system state information can be obtained with extremely low computational latency, overcoming the shortcomings of traditional physical models that consume too much computation time. This provides an immediate and accurate initial state basis for subsequent online rolling optimization of control parameters, thereby effectively solving the problem of decreased control performance caused by state perception lag under complex operating conditions.

[0020] More specifically, in a specific example of this application, inputting real-time SCADA data into a fast digital twin model to obtain a fast model state estimate includes: performing real-time data preprocessing and temporal feature construction on real-time SCADA data and buffered SCADA data to obtain a SCADA input feature tensor; inputting the SCADA input feature tensor into a fast digital twin model to obtain a fast model state estimate, wherein the fast digital twin model is an LSTM unit.

[0021] Accordingly, real-time data preprocessing and temporal feature construction are performed on real-time SCADA data and buffered SCADA data to obtain SCADA input feature tensors. It is understandable that the raw SCADA data collected by wind turbines contains various physical quantities such as wind speed, power, and pitch angle. These physical quantities have significant differences in dimensions and orders of magnitude. Directly using them as input to deep learning models can lead to difficulties in model training convergence or unstable weight updates. Furthermore, instantaneous data at a single moment cannot reflect the dynamic changes and temporal dependencies of wind turbine operation. Therefore, in the technical solution of this application, standardized real-time data preprocessing is further performed on real-time SCADA data and buffered SCADA data, and temporal features containing historical information are constructed based on time windows. This eliminates the numerical influence caused by different physical dimensions and transforms discrete observation points into feature expressions with a time dimension. This significantly improves the computational stability and feature extraction efficiency of the fast digital twin model, ensuring that the model can accurately capture the dynamic laws of wind turbine state evolution over time.

[0022] More specifically, in a concrete example of this application, the data processing flow is executed using a sliding time window approach. First, upon receiving the latest real-time SCADA data vector, this vector is immediately appended to the end of the SCADA historical data buffer, while the oldest data vector at the head of the buffer is removed, thus maintaining a fixed-length time series window covering operational data over a continuous period. Next, normalization is performed on the complete data sequence within the buffer. Using statistical parameters pre-calculated based on offline datasets, feature values ​​of different dimensions, such as wind speed, rotational speed, and power, are uniformly mapped to a specific standard interval, eliminating dimensional differences. Finally, the normalized two-dimensional data sequence is dimensionally reshaped according to the input requirements of a deep neural network, constructing a SCADA input feature tensor with a three-layer structure of batch, time step, and feature dimension, providing a standardized input format for the efficient iterative computation of the subsequent fast digital twin model.

[0023] Accordingly, the SCADA input feature tensor is input into a fast digital twin model to obtain a fast model state estimate, wherein the fast digital twin model is an LSTM unit. It is understood that, due to the significant inertial characteristics and time-varying nonlinearity of wind turbine operation, its current internal state depends not only on the instantaneous input but also on the historical operating trajectory. Conventional feedforward neural networks struggle to effectively capture this long-term temporal dependency. Therefore, in the technical solution of this application, the constructed SCADA input feature tensor is further input into a fast digital twin model built based on a long short-term memory network unit. This utilizes the unique gating mechanism of the LSTM unit to recursively process the time-series data, accurately extracting the dynamic evolution patterns hidden in the historical data. In this way, within millisecond-level computational latency, the key physical states of the wind turbine that are currently difficult to measure directly can be accurately deduced from multi-dimensional sensor data, providing state feedback with both real-time performance and dynamic accuracy for the formulation of control strategies.

[0024] More specifically, in a concrete example of this application, the state estimation computation is implemented through forward propagation of a deep neural network. First, a preprocessed 3D feature tensor is loaded into the input layer of a fast digital twin model. This tensor is sequentially fed into a hidden layer containing multiple LSTM units along the time step dimension. In the computation at each time step, the forget gate, input gate, and output gate within the LSTM unit work together to selectively retain key historical trend information, such as the continuous increase in wind speed or the cumulative effect of tower vibration, based on the current input and the hidden state of the previous time step, while eliminating irrelevant noise interference, thereby updating the memory cell state within the unit. Subsequently, the final hidden state vector output by the LSTM network after processing the entire time series is extracted and passed to the parallel state estimation head and performance prediction head at the end of the model. The state estimation head maps the high-dimensional abstract temporal features back to the specific physical state space through matrix multiplication and bias addition operations, ultimately outputting a fast model state estimation vector containing estimates of key physical quantities such as blade root bending moment and transmission chain torque. The performance prediction head outputs predicted values ​​of external performance such as active power and generator speed, providing a basis for the fidelity term in subsequent composite losses.

[0025] Specifically, in step S300, based on buffered SCADA data, a minute-level deep calibration is performed on the slow digital twin model to obtain the calibrated slow model physical parameters and high-fidelity state log. It is understandable that during long-term operation, the aerodynamic efficiency, mechanical damping, and structural stiffness of wind turbines can slowly drift due to dust accumulation on the blade surface, changes in ambient temperature, or component wear and aging. This causes the statically set model parameters to fail to accurately reflect the current physical state, severely affecting the physical accuracy of the model-based analysis conclusions. Therefore, in the technical solution of this application, buffered SCADA data is further used to drive the slow digital twin model based on physical mechanisms to perform minute-level iterative simulation and deviation minimization optimization. This allows for online identification and calibration of key physical parameters within the slow model, and the regeneration of a high-fidelity state log that conforms to physical constraints based on the updated parameters. In this way, a physical benchmark with adaptive evolution capability can be established, ensuring that the obtained internal state data not only approximates the real observations numerically, but also strictly satisfies the energy conservation and dynamic constraints in terms of mechanism. This provides high-quality and physically interpretable knowledge data for subsequent elimination of cognitive biases in fast models.

[0026] Figure 3 This document describes a flowchart illustrating a method for online optimization of wind turbine control parameters based on digital twins, using buffered SCADA data to perform minute-level deep calibration on a slow digital twin model to obtain the calibrated slow model's physical parameters and high-fidelity status log. (See attached flowchart.) Figure 3 As shown, step S300 includes: S310, inputting buffered SCADA data into the slow digital twin model to obtain a simulation output sequence; S320, extracting the actual performance output sequence from the buffered SCADA data; S330, calculating the deviation sequence between the simulation output sequence and the actual performance output sequence; S340, performing model parameter identification and calibration based on an optimization algorithm on the deviation sequence to obtain calibrated slow model physical parameters; S350, after updating the internal physical parameters of the slow digital twin model to the calibrated slow model physical parameters, inputting buffered SCADA data into the updated slow digital twin model to obtain a high-fidelity status log.

[0027] In step S310, buffered SCADA data is input into the slow digital twin model to obtain a simulation output sequence. It is understood that since the parameter calibration of the slow digital twin model is essentially an inversion process based on a prediction-correction mechanism, it is necessary to first obtain the model's theoretical response to a specific historical operating condition under the current physical parameter configuration in order to reveal the deviation between the model parameters and the physical entity by comparing it with actual observations. Therefore, in the technical solution of this application, the environmental load data and control command data in the buffered SCADA data are further extracted and input as driving boundary conditions into the high-fidelity slow digital twin model to trigger the model to perform time-step simulation calculations based on physical mechanisms. This generates a simulation output sequence entirely based on the current parameter assumptions. This sequence quantifies the model's theoretical expectations for historical operating conditions, providing a necessary benchmark for subsequent calculation of the deviation sequence and guiding the optimization algorithm to correct the physical parameters.

[0028] More specifically, in a concrete example of this application, the generation process of the simulation output sequence is implemented through numerical simulation in a high-performance computing unit. First, historical data packets stored in a buffer are parsed to extract independent sequences of model driving variables. These sequences of driving variables precisely correspond to the input interfaces of the physical model, including actual measured environmental parameters such as wind speed, wind direction, and air density, as well as the actual blade pitch angle and generator electromagnetic torque commands issued by the wind turbine controller. Subsequently, the extracted sequences of driving variables are loaded into a slow digital twin model, which incorporates a set of differential equations (e.g., based on blade element momentum theory and multibody dynamics equations) describing the aerodynamic, structural, and transmission characteristics of the wind turbine. The simulation solver, based on a preset small time step, uses a numerical integration algorithm to solve the equations step-by-step in the time domain, calculating the dynamic response of the system at each moment. Finally, the active power, generator speed, and key component load data calculated during the simulation are reassembled and aligned in chronological order to form a continuous simulation output sequence.

[0029] In step S320, the actual performance output sequence is extracted from the buffered SCADA data. It is understood that since the parameter calibration of the slow digital twin model relies on the quantitative difference between simulated and observed values, it is necessary to establish the actual response of the physical entity within the same historical period as a comparison benchmark. Therefore, in the technical solution of this application, key performance index data representing the actual operating efficiency of the wind turbine are further screened and extracted from the buffered SCADA data to construct the actual performance output sequence used to evaluate the accuracy of the model. This provides a precise physical anchor point for subsequent calculation of the deviation sequence, ensuring that the parameter calibration process is guided by the goal of approximating the actual physical characteristics.

[0030] More specifically, in a concrete example of this application, the extraction of the true performance output sequence is implemented through data indexing and filtering operations. First, based on the performance index definition of the slow digital twin model, the corresponding feature column index is located in the multidimensional array of buffered SCADA data to identify key physical quantity channels, including generator active power, generator speed, and tower vibration amplitude. Then, according to the strict correspondence of timestamps, the data points within these channels are read out one by one, and data quality checks are performed to remove invalid points caused by communication packet loss or momentary sensor failures. If necessary, linear interpolation is used to fill in minor missing segments. Finally, the cleaned multidimensional data points are assembled into a standardized performance output sequence vector in chronological order, ensuring that its time length, sampling frequency, and physical dimensions are completely consistent with the simulation output sequence generated in the aforementioned steps, thus preparing the data for the point-by-point difference calculation between the two.

[0031] In step S330, the deviation sequence between the simulation output sequence and the actual performance output sequence is calculated. It is understood that the parameter optimization algorithm for slow digital twin models requires explicit numerical guidance to judge the quality of the current parameter settings, and a single cumulative error value may mask instantaneous feature differences in the dynamic process, failing to effectively guide the model to capture complex temporal evolution patterns. Therefore, in the technical solution of this application, point-by-point difference calculation is further performed on the simulation output sequence and the actual performance output sequence under the same time reference to construct a deviation sequence that can accurately reflect the distribution of model prediction deviation over time. This provides basic data support for the subsequent construction of a cost function containing least squares terms and regularization terms, transforming the problem of physical parameter identification into a solvable mathematical optimization problem, ensuring that the optimization direction accurately points to the gradient direction that reduces the behavioral differences between the model and the entity.

[0032] More specifically, in a concrete example of this application, the calculation process of the deviation sequence involves temporal alignment and weighted differential processing of multidimensional data. First, the high-fidelity simulation output sequence and the actual performance output sequence extracted from buffered SCADA data are placed in the same time axis coordinate system, and the timestamps are verified to ensure strict correspondence between the two at each sampling moment. Then, for each sampling point in the sequence, a vector subtraction operation is performed, that is, the actual performance observation value is subtracted from the simulation prediction value calculated by the model, thus obtaining the instantaneous residual vector at each time step. Based on this, in order to eliminate the numerical imbalance caused by the difference in dimensions between different physical quantities, such as megawatt-level power and revolutions per minute (RPM), the components in the above instantaneous residual vector are standardized and weighted using preset weighting coefficients. Finally, the residual vectors of all processed time steps are linked sequentially to form a complete deviation sequence matrix. This matrix not only records the amplitude difference but also retains the sign and trend information of the deviation, and is directly passed as input to the subsequent optimization algorithm module to construct the objective function.

[0033] In step S340, the deviation sequence is subjected to model parameter identification and calibration based on an optimization algorithm to obtain the calibrated physical parameters of the slow model. It is understood that simply calculating the deviation sequence is insufficient to directly eliminate the error between the model and the entity; a systematic optimization mechanism is necessary to deduce the root physical causes of these deviations, i.e., the specific parameter variations. Therefore, in the technical solution of this application, a parameter identification algorithm based on numerical optimization is further adopted. With the goal of minimizing the statistical norm of the deviation sequence, it automatically searches for and adjusts the key physical parameters of the slow model within a preset feasible region, thereby mapping the observed differences in external response back to the correction amounts of internal physical properties. In this way, a mathematically optimal and physically reasonable set of calibration parameters can be obtained, enabling the behavioral characteristics of the slow digital twin model to re-align with the actual operating state of the wind turbine.

[0034] More specifically, in a concrete example of this application, the parameter identification and calibration process is implemented as a closed-loop iterative optimization task. First, a composite objective function is constructed, comprising a data fitting term and a physical constraint term. The data fitting term quantifies the accuracy of the model's predictions by accumulating the mean squared error at each time step in the bias sequence, while the physical constraint term penalizes the parameter range in a regularized manner to ensure the physical reliability of the solution. Subsequently, a nonlinear optimization solver is initiated, employing, for example, a covariance matrix adaptive evolution strategy or a Bayesian optimization algorithm, to search within the defined parameter space. Based on the current parameter evaluation results, the solver calculates the gradient of the objective function or constructs a surrogate model, intelligently generating the next set of candidate parameter combinations with higher potential value. This cyclical process of parameter generation, error evaluation, and strategy updating continues until the decrease in the objective function falls below a preset threshold or the maximum number of iterations is reached. At this point, the final parameter combination is confirmed as the calibrated slow model physical parameters, representing the current true physical characteristic configuration of the wind turbine.

[0035] In step S350, after updating the internal physical parameters of the slow digital twin model to the calibrated slow model physical parameters, buffered SCADA data is input into the updated slow digital twin model to obtain a high-fidelity state log. It is understandable that the parameter optimization process in the preceding steps mainly focuses on minimizing the prediction error at the output end, and for computational efficiency, it typically only tracks key performance indicators without fully preserving the evolution trajectory of all internal physical states that conform to the final optimal parameter configuration. The subsequent correction of the fast model requires a set of full-state data that both conforms to the current physical reality and contains complete internal details as a teacher signal. Therefore, in the technical solution of this application, after updating the internal physical parameters of the slow digital twin model to the calibrated slow model physical parameters, buffered SCADA data is input into the updated slow digital twin model to obtain a high-fidelity state log, thereby performing a full-element physical reconstruction simulation and generating a complete time-series record containing all internal state variables that are difficult to measure directly. This ensures that the final output state log is completely consistent with the physical mechanism and the characteristics of the calibrated device, providing high-precision learning samples for the fast model and thus solving the problem of cognitive bias caused by the lack of physical constraints in the fast model.

[0036] More specifically, in a concrete example of this application, this step involves the sequential execution of model configuration updates and full data recording. First, the parameter configuration file or memory-mapped area of ​​the slow digital twin model is accessed, and all existing parameters to be identified, such as the aerodynamic airfoil coefficient, transmission chain damping coefficient, and material stiffness coefficient, are replaced with the calibrated values ​​output by the optimization algorithm, completing the update of the model's physical properties. Then, the physical simulation engine is invoked again, using historical environmental loads and control commands recorded in the buffered SCADA data as the driving source to perform time-domain simulation on the updated model. During this simulation, a full-state monitoring mode is enabled, calculating not only the conventional output indicators but also calculating and recording in real time all preset internal state variables, including blade root flapping moment, tower top displacement, main shaft torsion angle, and gearbox oil temperature. Finally, these internal state variables are organized according to time steps, synchronized with the original input data, and a formatted high-fidelity state log file is generated as high-quality training data for knowledge distillation.

[0037] Specifically, in step S400, based on high-fidelity state logs and buffered SCADA data, the fast digital twin model is periodically corrected using knowledge distillation to obtain a corrected fast digital twin model. It is understandable that, since the lightweight fast digital twin model is essentially a data-driven model based on statistical laws, it is susceptible to model drift due to biases in the distribution of training data during long-term online operation. Furthermore, lacking physical mechanism constraints, it may output state estimates that violate physical laws when facing complex operating conditions outside the training set, thereby reducing the reliability of control decisions. Therefore, in the technical solution of this application, the fast digital twin model is further periodically corrected using knowledge distillation based on high-fidelity state logs and buffered SCADA data to obtain a corrected fast digital twin model. This constructs a teacher-student learning framework, using the output of the slow model containing the true values ​​of physical mechanisms to guide the fast model in adjusting its network weights. In this way, the high-precision physical knowledge calibrated from the slow model can be effectively transferred to the fast model, allowing it to regain physical consistency and high accuracy while maintaining the advantage of millisecond-level inference speed, and ensuring that the online optimization system can continuously adapt to the time-varying evolution of the physical characteristics of the wind turbine.

[0038] Figure 4 This document describes a flowchart illustrating a method for online optimization of wind turbine control parameters based on digital twins, using high-fidelity status logs and buffered SCADA data. It outlines a process for periodically correcting a fast digital twin model based on knowledge distillation to obtain the corrected fast digital twin model. (The flowchart is not included in the provided text.) Figure 4As shown, step S400 includes: S410, constructing a knowledge distillation training dataset based on high-fidelity state logs and buffered SCADA data; S420, inputting the knowledge distillation training dataset into the fast digital twin model to obtain the gradient of the composite loss function; S430, fine-tuning and updating the network weights of the fast digital twin model based on the gradient of the composite loss function to obtain the corrected fast digital twin model.

[0039] In step S410, a knowledge distillation training dataset is constructed based on the high-fidelity state log and buffered SCADA data. It is understandable that, although the original buffered SCADA data and the generated high-fidelity state log describe the wind turbine's operation over the same time period, they originate from a real-time acquisition system and offline simulation calculations, respectively. This often results in differences in data structure or slight temporal deviations, and a lack of corresponding label relationships that can be directly read by deep learning frameworks. Therefore, in this application's technical solution, a knowledge distillation training dataset is further constructed based on the high-fidelity state log and buffered SCADA data. This rigorously aligns external observation data and internal physical state data in the time dimension, structuring them into one-to-one input-teacher label-real label sample pairs. This provides standardized supervision signals for the correction of the fast digital twin model, ensuring that the model can simultaneously learn the physical mechanisms of the slow model and the operational rules of the real world.

[0040] More specifically, in a concrete example of this application, the dataset is constructed through a multi-source time-series data fusion process. First, using the timestamps of the buffered SCADA data as a benchmark, the high-fidelity state log is indexed and matched in the time domain. Asynchronous sampling points are processed using interpolation or discarding strategies, generating two sequences that correspond precisely at each time step. Next, for each synchronized time step, feature engineering and label extraction are performed: wind speed, wind direction, and blade pitch angle are extracted from the SCADA data as model input features; actual power and rotational speed are extracted from the SCADA data as true performance labels; and simultaneously, blade root bending moment and tower vibration acceleration core internal state variables are extracted from the high-fidelity state log as teacher knowledge labels. Finally, the extracted input features, true performance labels, and teacher knowledge labels are sliced ​​and encapsulated into tensors according to the sequence length required by the fast digital twin model (e.g., the time window size of an LSTM), forming a standardized training dataset containing multiple batches, which can then be directly used by the subsequent gradient descent algorithm.

[0041] In step S420, the knowledge distillation training dataset is input into the fast digital twin model to obtain the gradient of the composite loss function. It is understood that since the network weight correction of the fast digital twin model is a supervised learning process based on error backpropagation, its core lies in accurately quantifying the multidimensional difference between the model's current predicted output and the ideal target, and transforming it into a mathematical gradient to guide weight updates. Therefore, in the technical solution of this application, the constructed knowledge distillation training dataset is further input into the fast digital twin model to perform forward propagation calculation, and a composite loss function containing physical consistency constraints and real data fitting constraints is constructed based on the output results to calculate the gradient, thereby establishing the direction and step size of the fast model parameter optimization. This ensures that subsequent weight update operations take into account both the inheritance of the physical mechanism of the slow model and the fitting of real environment observations, preventing the model from overfitting or physically failing under the drive of a single objective.

[0042] More specifically, in a concrete example of this application, the gradient calculation process is implemented within the deep learning training framework via an automatic differentiation mechanism. First, a batch of sample data is extracted from the knowledge distillation training dataset, and the input feature vector containing wind speed and control commands is input into the fast digital twin model. The model performs forward inference, outputting the internal state estimate and external performance prediction under the current parameter configuration. Subsequently, a composite loss function is constructed, consisting of two weighted parts: the first part is the distillation loss, which calculates the mean squared error between the internal state estimate output by the fast model and the corresponding high-fidelity state label of the slow model in the dataset, used to force the fast model to mimic the physical behavior of the slow model; the second part is the fidelity loss, which calculates the mean squared error between the performance prediction output by the fast model and the actual SCADA performance label, used to constrain the fast model to conform to objective observations. Finally, the composite loss function is differentiated using the backpropagation algorithm, and chain-law differentiation is used to obtain the partial derivatives of the loss function with respect to the weight matrix and bias vector of each LSTM unit in the fast model, thereby generating a gradient vector that indicates the direction of the fastest error descent.

[0043] In step S430, the network weights of the fast digital twin model are fine-tuned and updated based on the gradient of the composite loss function to obtain the corrected fast digital twin model. It is understandable that since the calculated gradient only indicates the direction and magnitude of the neural network parameters that need to be adjusted to minimize prediction error, without actual parameter updates, the model cannot internalize the physical laws and real-time data correction information transmitted by the slow digital twin model. Therefore, in the technical solution of this application, an optimizer is further used to fine-tune and update the network weights of the fast digital twin model based on the gradient of the composite loss function. This substantially maps physical consistency constraints and data fitting requirements into the parameter space of the fast model, correcting the cognitive biases caused by long-term operation. In this way, the fast digital twin model can dynamically adapt to changes in the current physical characteristics of the wind turbine while retaining its lightweight architecture, ensuring that its output state estimates have both physical interpretability and numerical accuracy in subsequent online control.

[0044] More specifically, in a concrete example of this application, the fine-tuning and updating of network weights is implemented through iterative optimization. First, an efficient stochastic gradient descent variant optimization algorithm, such as an adaptive moment estimator (IME), is configured, and a small learning rate parameter is set to ensure the stability of the fine-tuning process and avoid disrupting the model's original prior knowledge structure. The optimizer receives the gradient vectors for each LSTM unit and fully connected layer calculated by backpropagation, and adaptively calculates the update step size for each weight parameter by combining the statistical moment information of historical gradients. Subsequently, a parameter update operation is performed, subtracting the calculated update amount from the current weight matrix to generate a weight matrix containing the new knowledge. This calculation and update cycle is repeated multiple training rounds on the knowledge distillation training dataset until the composite loss function converges to a preset minimum range or reaches the maximum number of iterations. Finally, the converged model parameters are solidified, replacing the original online running model, thus completing the deployment of the modified fast digital twin model.

[0045] Specifically, in step S500, the fast model state estimate is used as the initial state and input into the modified fast digital twin model to obtain the final optimal control sequence. It is understood that, since a wind turbine is a complex system with large inertia, multivariable coupling, and nonlinear characteristics, simply relying on the current feedback error for adjustment often lags behind wind condition changes and cannot achieve optimal overall performance. Furthermore, the formulation of the control strategy must be strictly based on a precise understanding of the current internal state of the system and an accurate prediction of future dynamic evolution. Therefore, in the technical solution of this application, the fast model state estimate is further used as the starting anchor point in the prediction time domain and input into the fast digital twin model after physical consistency correction. Under the model predictive control framework, a constrained optimization problem is solved in a rolling manner to obtain the final optimal control sequence. This utilizes the computational speed and physical accuracy of the modified fast model to pre-simulate the system response to various future control behaviors within a finite control cycle. This ensures that the generated control strategy is not only safe at the current moment but also maximizes wind energy capture efficiency and minimizes mechanical fatigue load in the future prediction time domain, achieving global optimization of wind turbine operating efficiency.

[0046] Figure 5 This is a flowchart illustrating the online optimization method for wind turbine control parameters based on digital twins according to embodiments of this application. It uses a fast model state estimate as the initial state, inputs it into a modified fast digital twin model to obtain the final optimal control sequence. (See flowchart for example.) Figure 5 As shown, step S500 further includes: S510, constructing an MPC problem based on model state estimation, optimization objective, running constraint set, and prediction time domain; S520, inputting the MPC problem into the modified fast digital twin model to obtain a converged optimal control sequence; S530, extracting the final optimal control sequence from the converged optimal control sequence.

[0047] In step S510, an MPC problem is constructed based on model state estimation, optimization objective, operating constraint set, and prediction time domain. It is understandable that the multi-objective control requirements of wind turbines involve a dynamic trade-off between maximizing power generation and minimizing mechanical load, and the current control action will significantly affect the dynamic response over a future period due to system inertia. Simple single-point feedback regulation is insufficient to handle this cross-time domain coupling relationship and strict physical boundary constraints. Therefore, in the technical solution of this application, based on the real-time state estimation value output by the fast model, combined with a preset optimization objective function, a rigid operating constraint set, and a defined prediction time domain, the wind turbine control decision task is formally constructed as a standard model predictive control mathematical programming problem. This transforms the complex control requirements of the physical world into a constrained optimization model that can be accurately solved in numerical space. This provides a mathematically well-defined solution framework for subsequent optimization algorithms, ensuring that the final control strategy is the globally optimal solution for the comprehensive performance indicators of the wind turbine over a future time domain, under the premise of strictly satisfying all safety boundaries.

[0048] More specifically, in a concrete example of this application, the construction process of the MPC problem is implemented in the computational core of the controller through mathematical modeling and parameter assembly. First, the fast model state estimation vector obtained in the previous steps is directly set as the initial state condition at the start of the prediction time domain, ensuring that the predicted state evolution trajectory strictly begins at the current real physical instant of the wind turbine. Next, an optimization objective function is configured according to the current wind farm conditions and operating strategy. This function is constructed as a quadratic cost function accumulated over the entire prediction time domain, which includes terms for penalizing deviations between power generation and rated power, as well as terms for penalizing fatigue loads on key components such as blade root flapping moments and tower front-to-back vibrations. Adjustable weighting coefficients are used to quantify the priority between different control objectives. Simultaneously, the set of operating constraints is analyzed and transformed into a series of linear or nonlinear inequality constraints covering the entire prediction time domain. These constraints clearly define the allowable intervals of state variables and the physical limitations of control inputs. Finally, the specific length of the prediction time domain is determined, and the initial state, cost function, inequality constraints, and the modified fast digital twin model that serves as the system state transition equation are encapsulated as a whole and combined to form a nonlinear programming problem to be solved with the future control sequence as the decision variable.

[0049] In step S520, the MPC problem is input into the modified fast digital twin model to obtain the converged optimal control sequence. It is understood that since the constructed model predictive control problem is a non-convex optimization problem containing nonlinear system dynamics constraints and complex inequality constraints, a closed-form solution cannot be directly obtained analytically. Furthermore, the control system has extremely high real-time requirements, demanding that all calculations from state perception to optimal decision be completed within milliseconds. Therefore, in the technical solution of this application, an efficient numerical solution strategy based on sequential quadratic programming is further adopted. The nonlinear programming problem is transformed into a series of quadratic programming subproblems. The modified fast digital twin model is used to quickly deduce the state trajectory in each iteration and provide the gradient information required for linearization, thereby iteratively approximating the global or local optimal solution. This fully leverages the advantage of the fast model's rapid inference speed, and under the strict temporal constraints of the control cycle, calculates a converged optimal control sequence that satisfies all safe operating boundaries and maximizes the overall efficiency of the wind turbine within the prediction time domain.

[0050] More specifically, in a concrete example of this application, the solution process for the optimal control sequence is implemented through an iterative loop comprising linearization approximation, subproblem solving, and solution updating. First, the solver initializes a current candidate control sequence and performs forward prediction based on this sequence using a modified fast digital twin model, generating the corresponding current state trajectory. Then, the iterative loop begins. In the k-th iteration, for the operating point formed by the current candidate control sequence and the state trajectory, a Taylor series expansion of the modified fast digital twin model is performed to obtain the Jacobian matrix of the system dynamics. Simultaneously, the gradient and Hessian matrix of the objective function at this point are calculated, thus constructing a quadratic programming subproblem aimed at approximating the local characteristics of the original nonlinear problem. Next, the quadratic objective function matrix and linearization constraint matrix contained in this QP subproblem are input into the embedded QP solver. The solver calculates the optimal search direction that reduces the objective function and the corresponding step size factor using either the effective set method or the interior-point method. Finally, the current candidate control sequence is corrected and updated using the calculated search direction and step size factor, and the updated sequence is evaluated to see if it meets the preset convergence criteria, such as whether the rate of change of the objective function value is lower than the threshold. If it does not converge, the updated sequence is used as the basis for the next iteration and the above process is repeated until the convergence condition is met. Finally, a stable and converged control parameter sequence is output as the final result.

[0051] In step S530, the final optimal control sequence is extracted from the converged optimal control sequence. It is understood that since the numerical optimizer generates a series of intermediate candidate solutions during the iterative optimization process, only the final iteration result that satisfies the convergence criterion is the truly physically constrained and optimal solution. This solution is typically temporarily stored in the solver's internal volatile memory structure in the form of an abstract mathematical vector, requiring explicit reading, parsing, and solidification into a data format recognizable by the control system. Therefore, in the technical solution of this application, after confirming the convergence of the optimization process, the decision variable vector generated in the last iteration is immediately read directly from the solver's workspace, defined, and extracted as the final optimal control sequence. This transforms the mathematically optimal solution into executable policy data at the engineering level and ends the current optimization loop. This allows for a data transition from the computational domain to the execution domain, ensuring that the downstream execution stage receives a globally or locally optimal policy that has undergone rigorous convergence verification, while providing usable initial values ​​for the optimization solution of the next control cycle.

[0052] More specifically, in a concrete example of this application, the extraction process is implemented through a memory-mapped reading and data structured encapsulation process. First, the control algorithm program accesses the standard output interface of the numerical optimizer or a specific memory address to lock the array of final decision variables stored therein. This array numerically corresponds precisely to the optimal pitch angle setpoint and generator electromagnetic torque setpoint to be executed at each discrete moment within the entire prediction time domain. Then, the original numerical array is parsed according to a predefined time step sequence and variable dimensions, restoring it from a compact one-dimensional vector format within the solver to a multi-dimensional time series matrix with clear physical meaning. Finally, the extracted data undergoes a final boundary integrity check to confirm that all control parameters are within the physical safety range of the actuator. Finally, the verified time series data, along with the timestamp information of this calculation, is uniformly encapsulated into a standardized control sequence data package, which is output as the final optimal control sequence to the instruction distribution buffer, ready for subsequent single-step instruction extraction and execution operations.

[0053] Specifically, in step S600, the control parameters of the first time step are extracted from the final optimal control sequence as the control command. It is understandable that although model predictive control calculates multi-step optimal control strategies over a predictable time domain, the highly random and unpredictable nature of wind conditions, coupled with unavoidable minor biases in the model itself, means that directly executing the complete long sequence of commands in open loop would cause the actual trajectory to deviate from the optimal path, failing to respond promptly to subsequent immediate disturbances. Therefore, in the technical solution of this application, the basic principle of rolling time-domain control is further followed: only the control parameters corresponding to the first time step are extracted and retained from the calculated final optimal control sequence, while the prediction results of the remaining time steps in the sequence are discarded. This establishes the first time step as the only effective execution action in the current control cycle. This combines forward-looking decision-making based on long-term time-domain optimization with closed-loop correction based on current-moment feedback, ensuring that each control action not only meets long-term performance goals but also flexibly adapts to the latest environmental changes in the next moment.

[0054] More specifically, in a concrete example of this application, the extraction and issuance of control commands are implemented through a process of data index addressing and communication protocol conversion. First, the final optimal control sequence data packet stored in the memory buffer is parsed. This sequence is essentially a two-dimensional array arranged in time step order. The starting position of this array is located through indexing operations, and the set of control variable values ​​corresponding to the current time point is precisely read. This set of values ​​explicitly includes the optimal pitch angle setpoint and generator electromagnetic torque setpoint for the current moment. Subsequently, the read values ​​are converted into an industrial communication protocol format recognizable by the wind turbine main controller, such as CANopen or PROFINET messages, and subjected to double verification with safety limits. Finally, the encapsulated command message is sent to the underlying pitch drive and converter controller via an instant communication link, driving the actuators to move. Simultaneously, the optimization result of this time step is used as the initial value for the warm-start optimization calculation of the next control cycle, thus completing a complete online optimization control cycle.

[0055] In summary, the online optimization method for wind turbine control parameters based on digital twins according to the embodiments of this application is explained. It inputs real-time SCADA data into a fast digital twin model for millisecond-level state estimation to meet the real-time response requirements of the control system. Simultaneously, buffered data drives a slow digital twin model for deep calibration, generating a high-fidelity state log containing physical mechanisms. Knowledge distillation technology is then used to transfer the physical knowledge of the slow model to the fast model, periodically correcting the fast model. Finally, the corrected fast model is used for rolling optimization within a model predictive control framework to output the optimal control sequence. This scheme effectively overcomes the shortcomings of traditional physical models, such as excessive computation time and the tendency for pure data-driven models to drift and lack physical interpretability. While ensuring millisecond-level response speed, it improves the physical consistency of model predictions and the robustness of long-term operation.

[0056] Furthermore, an online optimization system for wind turbine control parameters based on digital twins is also provided.

[0057] Figure 6 This is a block diagram of a digital twin-based online optimization system for wind turbine control parameters according to an embodiment of this application. Figure 6 As shown, the online optimization system 100 for wind turbine control parameters based on digital twin according to an embodiment of this application includes: a data acquisition module 110 for acquiring real-time SCADA data and buffered SCADA data; a fast model state estimation module 120 for inputting real-time SCADA data into a fast digital twin model to obtain a fast model state estimate; a deep calibration module 130 for performing minute-level deep calibration on a slow digital twin model based on buffered SCADA data to obtain calibrated slow model physical parameters and a high-fidelity state log; a periodic correction module 140 for performing knowledge distillation-based periodic correction on the fast digital twin model based on the high-fidelity state log and buffered SCADA data to obtain a corrected fast digital twin model; an optimal control sequence acquisition module 150 for using the fast model state estimate as the initial state and inputting it into the corrected fast digital twin model to obtain a final optimal control sequence; and a control command generation module 160 for extracting the control parameters of the first time step from the final optimal control sequence as control commands.

[0058] The fast model state estimation module 120 includes a real-time data preprocessing and temporal feature construction unit, which is used to perform real-time data preprocessing and temporal feature construction on real-time SCADA data and buffered SCADA data to obtain SCADA input feature tensors; and a fast model state estimation unit, which is used to input the SCADA input feature tensors into a fast digital twin model to obtain fast model state estimation, wherein the fast digital twin model is an LSTM unit.

[0059] The periodic correction module 140 includes a knowledge distillation training dataset construction unit, used to construct a knowledge distillation training dataset based on high-fidelity state logs and buffered SCADA data; a composite loss function gradient acquisition unit, used to input the knowledge distillation training dataset into the fast digital twin model to obtain the gradient of the composite loss function; and a network weight fine-tuning and update unit, used to fine-tune and update the network weights of the fast digital twin model based on the gradient of the composite loss function to obtain the corrected fast digital twin model.

[0060] As described above, the online optimization system 100 for wind turbine control parameters based on digital twins according to embodiments of this application can be implemented in various types of computing devices or control units. For example, it can be deployed in an industrial personal computer at a wind farm, an intelligent edge computing gateway directly connected to the wind turbine main controller, or a high-performance simulation server integrated into a central control center's SCADA system. In one possible implementation, the online optimization system 100 for wind turbine control parameters based on digital twins according to embodiments of this application can be integrated into the computing device as a software module and / or a hardware module. For example, the online optimization system 100 for wind turbine control parameters based on digital twins can be an intelligent control module in the operating system of the computing device or edge node. This software module is configured to perform time-series feature construction of real-time SCADA data, fast model state estimation based on LSTM units, slow model parameter depth calibration based on physical mechanisms, and MPC rolling optimization solution based on the corrected fast model. Alternatively, it can be a dedicated wind turbine performance optimization control algorithm program developed for the computing device. Of course, the online optimization system 100 for wind turbine control parameters based on digital twins can also be one of the many hardware modules of the computing device or control unit, or it can be embedded in a field-programmable gate array circuit to accelerate the forward inference operation of deep neural networks and the iterative optimization process of quadratic programming solvers in parallel, or it can be an AI acceleration integrated circuit for a specific application.

[0061] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for online optimization of wind turbine control parameters based on digital twins, characterized in that, include: Acquire real-time SCADA data and buffered SCADA data; Real-time SCADA data is input into a fast digital twin model to obtain a fast model state estimate; Based on buffered SCADA data, a minute-level deep calibration is performed on the slow digital twin model to obtain the calibrated slow model physical parameters and high-fidelity status logs. Based on high-fidelity status logs and buffered SCADA data, the fast digital twin model is periodically corrected using knowledge distillation to obtain the corrected fast digital twin model. Using the fast model state estimate as the initial state, it is input into the modified fast digital twin model to obtain the final optimal control sequence; The control parameters of the first time step are extracted from the final optimal control sequence as the control command.

2. The online optimization method for wind turbine control parameters based on digital twins according to claim 1, characterized in that, Inputting real-time SCADA data into a fast digital twin model to obtain a fast model state estimate includes: Real-time data preprocessing and temporal feature construction are performed on real-time SCADA data and buffered SCADA data to obtain SCADA input feature tensors; The SCADA input feature tensor is input into a fast digital twin model to obtain a fast model state estimate, wherein the fast digital twin model is an LSTM unit.

3. The online optimization method for wind turbine control parameters based on digital twins according to claim 1, characterized in that, Based on buffered SCADA data, a minute-level deep calibration is performed on the slow digital twin model to obtain the calibrated slow model's physical parameters and high-fidelity state logs, including: Buffered SCADA data is input into the slow digital twin model to obtain the simulation output sequence; Extract the actual performance output sequence from buffered SCADA data; Calculate the deviation sequence between the simulation output sequence and the actual performance output sequence; The model parameters of the bias sequence are identified and calibrated based on an optimization algorithm to obtain the calibrated physical parameters of the slow model; After updating the internal physical parameters of the slow digital twin model to the calibrated slow model physical parameters, buffered SCADA data is input into the updated slow digital twin model to obtain a high-fidelity state log.

4. The online optimization method for wind turbine control parameters based on digital twins according to claim 1, characterized in that, Based on high-fidelity status logs and buffered SCADA data, a fast digital twin model is periodically revised using knowledge distillation to obtain a revised fast digital twin model, including: A knowledge distillation training dataset is constructed based on high-fidelity status logs and buffered SCADA data. Input the knowledge distillation training dataset into the fast digital twin model to obtain the gradient of the composite loss function; Based on the gradient of the composite loss function, the network weights of the fast digital twin model are fine-tuned and updated to obtain the corrected fast digital twin model.

5. The online optimization method for wind turbine control parameters based on digital twins according to claim 1, characterized in that, Using the fast model state estimate as the initial state, it is input into the modified fast digital twin model to obtain the final optimal control sequence, including: Based on model state estimation, optimization objective, running constraint set, and prediction time domain, construct the MPC problem; The MPC problem is input into the modified fast digital twin model to obtain the converged optimal control sequence; Extract the final optimal control sequence from the converged optimal control sequence.

6. The online optimization method for wind turbine control parameters based on digital twins according to claim 5, characterized in that, The set of operational constraints includes state constraints and control input constraints.

7. The online optimization method for wind turbine control parameters based on digital twins according to claim 5, characterized in that, The MPC problem is input into the modified fast digital twin model to obtain the converged optimal control sequence, including: Based on the MPC problem, the modified fast digital twin model, the current candidate control sequence, and the current state trajectory, determine the quadratic programming subproblem for the k-th iteration; The quadratic programming subproblem of the k-th iteration is input into the QP solver to obtain the optimal search direction and step size factor; Based on the optimal search direction and step size factor, the current candidate control sequence is updated and convergence is determined to obtain the converged optimal control sequence.

8. A wind turbine control parameter online optimization system based on digital twin, characterized in that, include: The data acquisition module is used to acquire real-time SCADA data and buffered SCADA data; The fast model state estimation module is used to input real-time SCADA data into the fast digital twin model to obtain the fast model state estimate; The deep calibration module is used to perform minute-level deep calibration on slow digital twin models based on buffered SCADA data to obtain the calibrated physical parameters and high-fidelity status logs of the slow model. The periodic correction module is used to perform knowledge distillation-based periodic correction on the fast digital twin model based on high-fidelity status logs and buffered SCADA data to obtain the corrected fast digital twin model. The optimal control sequence acquisition module is used to take the fast model state estimate as the initial state and input it into the modified fast digital twin model to obtain the final optimal control sequence. The control command generation module is used to extract the control parameters of the first time step from the final optimal control sequence as control commands.

9. The online optimization system for wind turbine control parameters based on digital twins according to claim 8, characterized in that, The fast model state estimation module includes: The real-time data preprocessing and temporal feature construction unit is used to perform real-time data preprocessing and temporal feature construction on real-time SCADA data and buffered SCADA data to obtain SCADA input feature tensors. The fast model state estimation unit is used to input the SCADA input feature tensor into the fast digital twin model to obtain the fast model state estimate, wherein the fast digital twin model is an LSTM unit.

10. The online optimization system for wind turbine control parameters based on digital twins according to claim 8, characterized in that, The periodic correction module includes: The knowledge distillation training dataset building unit is used to construct a knowledge distillation training dataset based on high-fidelity state logs and buffered SCADA data. The composite loss function gradient acquisition unit is used to input the knowledge distillation training dataset into the fast digital twin model to obtain the gradient of the composite loss function; The network weight fine-tuning and update unit is used to fine-tune and update the network weights of the fast digital twin model based on the gradient of the composite loss function to obtain the corrected fast digital twin model.

Images