Fan adaptive control method based on AI large model

Abstract

The invention discloses a fan adaptive control method based on an AI large model, and relates to a new energy power generation intelligent control technology, and the method comprises the following steps: S1, synchronously collecting fan body sensor data, laser radar / ultrasonic environment sensing data and a power grid dispatching signal, and carrying out the time-space alignment processing of multi-source heterogeneous data; and S2, inputting the preprocessed feature data into a wind power basic model based on a deep learning architecture, wherein the model is pre-trained through cross-wind-field historical data and is adapted through target unit data. According to the fan adaptive control method based on the AI large model, by means of the powerful learning and analysis capacity of the AI large model, complex data laws in the fan operation process can be deeply excavated, the fan operation state and environment changes can be accurately sensed, adaptive control is achieved, frequent manual intervention is not needed, the fan operation automation degree is greatly improved, and the fan operation efficiency is improved. The power generation efficiency is effectively improved, the energy loss is reduced, and the fault probability is reduced.

Description

Technical Field

[0001] This invention relates to intelligent control technology for new energy power generation, specifically to an adaptive control method for wind turbines based on an AI large model. Background Technology

[0002] In the current context of rapid development of renewable energy, wind power, as a crucial component of clean energy, plays a vital role in energy supply through its efficient and stable operation. Wind turbine control technology is a core element in ensuring the efficiency and safety of wind power generation. Traditional wind turbine control strategies primarily rely on the turbine's own operating data, such as wind speed, rotational speed, power, and temperature, achieving basic regulation functions through fixed control rules or a single model. However, as wind farm environments become increasingly complex, wind turbine operation requires comprehensive consideration of external environmental changes, including meteorological factors such as wind direction, turbulence intensity, and wind shear, as well as demand-side information such as grid dispatch instructions and electricity price signals. This places higher demands on the intelligence and adaptive capabilities of control strategies.

[0003] Existing wind turbine control methods are mostly based on preset rules or single small models, which have certain limitations in terms of data processing dimensions and decision-making capabilities. These methods struggle to effectively integrate multi-source heterogeneous data for accurate analysis, have limited accuracy in predicting short-term wind conditions, and often fail to achieve synergistic optimization of multiple objectives such as power generation improvement, load balancing, and lifespan extension when outputting control setpoints. Furthermore, traditional control strategies lack a continuous self-evolution mechanism, making it difficult to adapt to dynamic changes such as long-term climate variations in wind farms and the performance degradation of the wind turbines themselves. Their control performance under complex operating conditions needs further improvement. To address these issues, we propose an AI-based large-model adaptive wind turbine control method. Summary of the Invention

[0004] To address the aforementioned technical issues, an AI-based large-scale model-based adaptive control method for wind turbines is provided. This technical solution solves the problems of the control methods mentioned above, which rely on preset rules or a single small model, have limited data processing dimensions and decision-making capabilities, are difficult to effectively integrate multi-source heterogeneous data, have insufficient accuracy in short-term wind condition prediction, are difficult to achieve multi-objective collaborative optimization such as power generation improvement and load balancing, lack a continuous self-evolution mechanism, cannot adapt to dynamic changes, and have control effects that need to be improved under complex operating conditions.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The AI-based large-model adaptive control method for wind turbines includes the following steps: S1. Synchronously collect multi-source heterogeneous data including: wind turbine body sensor data, lidar / ultrasonic environmental perception data and power grid dispatch signals, and perform spatiotemporal alignment processing on the multi-source heterogeneous data; S2. Based on the spatiotemporal alignment of the multi-source heterogeneous data, preprocessing is performed, and the preprocessed feature data is input into the wind power basic model based on the deep learning architecture. The model is pre-trained with historical data across wind farms and adapted with target unit data to output wind condition prediction results and equipment status evaluation values. S3. Based on the wind condition prediction results and equipment status assessment values, generate a set of multi-objective optimization control setpoints through the basic model, including tip speed ratio curve, torque reference value, pitch angle sequence and yaw correction. S4. The optimized control setpoints are transformed into coordinated action commands for pitch, converter, and yaw systems using the Model Predictive Control (MPC) framework, with mechanical load constraints and dynamic compensation embedded in the transformation process. S5. Based on the feedback of actual control effect and wind condition prediction deviation, the parameters of the wind power basic model are updated through an online learning mechanism to achieve dynamic compensation for climate adaptability adjustment and equipment status changes, forming an adaptive control closed loop.

[0006] Preferably, the wind turbine body sensor data includes anemometer speed, generator power, gearbox vibration spectrum, main shaft torque, and bearing temperature; the environmental perception data includes feedforward wind speed profiles, three-dimensional turbulence intensity matrices, and wind shear gradients captured by lidar; the power grid dispatch signals cover frequency regulation commands, active / reactive power setpoints, and real-time marginal electricity prices in the electricity market; the multi-source data are fused using Kalman filtering to generate a spatiotemporally aligned feature tensor, with an adjustable sliding time window as the input window length; The length of the sliding time window is adaptively adjusted according to wind conditions: a long window is used to smooth noise in the stable wind speed range, and a short window is automatically switched to improve the response speed in the gust transition range; the process noise covariance matrix of the Kalman filter is dynamically corrected according to the sensor data quality, and the filtering intensity is automatically increased when the signal-to-noise ratio is lower than 20dB.

[0007] Preferably, the wind power foundation model adopts a hierarchical Transformer architecture, and the pre-training stage uses mask autoencoder technology to learn the aerodynamic-mechanical coupling relationship from petabyte-level cross-wind field data; The adaptation process employs a dual optimization mechanism: freezing the encoder layer weights, fine-tuning the multilayer sensor using specific unit labeled data, and using the weighted sum of the power tracking mean square error and the tower bending moment spectrum damage value as the loss function.

[0008] Preferably, the multi-objective optimization control setpoint is generated through Pareto optimization, and the objective function includes: a power generation maximization term, a key component fatigue damage accumulation rate suppression term, and a grid dispatch command tracking error term; the constraints include the pitch angle change rate, the generator torque fluctuation band rating, and the yaw system response lag time.

[0009] Preferably, the model predictive control (MPC) employs rolling time-domain quadratic programming, solved in each control cycle: ; In the formula, To control the amount of change in the input, For discrete time steps, To predict the length of the time domain, To control the length of the time domain, for Predicted power at time, for Reference power at time, This is the power tracking weight matrix. for The change in torque at any given time. This is the torque stability weighting matrix. This is the fatigue damage penalty coefficient. for The fatigue damage index at any given time.

[0010] Preferably, the online learning mechanism includes the following operations: The relative error between predicted power and actual power and the standard deviation of equipment load are calculated in real time through the closed-loop evolution layer. When the relative error between predicted power and actual power exceeds a preset percentage or the standard deviation of load rises above a predetermined threshold for a consecutive preset number of samples, the incremental learning process is triggered. An elastic weight fixation algorithm is adopted to lock the pre-trained weights of the Transformer encoder layer in the wind power foundation model and only update the weights of the fully connected layer of the model output layer. The updated model parameters are input into the digital twin, and the load fluctuation of the control command is simulated and verified. If the condition is met, the update is executed; otherwise, the original parameters are returned.

[0011] Preferably, an edge-cloud collaborative architecture is deployed, with embedded modules on the edge performing data preprocessing and lightweight model inference; the cloud training platform adopts a parameter server architecture, aggregates the operating data of multiple wind turbines to train the basic model in parallel, and protects data security through differential privacy technology; the model parameters are synchronized to the edge at a preset period via the OPC UA protocol.

[0012] Preferably, a robustness enhancement module is set up so that when the confidence level of the basic model output is lower than 90% or conflicts with the aeroelastic equation, it switches to a backup controller based on the deep deterministic policy gradient (DDPG). The backup controller uses a simplified physical model as an environment simulator and outputs control commands that satisfy the Lyapunov stability conditions.

[0013] Preferably, a life prediction subnetwork is coupled in the dynamic decision layer. This subnetwork receives the hidden layer features of the basic model and outputs the gearbox remaining life index and the blade fatigue damage index. The execution control layer, based on the dynamic relaxation power tracking target, initiates the load reduction mode at a certain time, limiting the torque command to 70% of the rated value.

[0014] Preferably, a multiphysics verification closed loop is established, and the actual load data is fed back to the digital twin. The twin includes an aeroelastic model, a transmission chain multibody dynamics model, and a power grid impedance model. The twin simulation results are used to correct the multi-objective weight coefficients of the basic model, forming a virtual-entity interactive optimization mechanism.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: The wind turbine adaptive control method based on an AI large model proposed in this invention leverages the powerful learning and analysis capabilities of AI large models to deeply mine complex data patterns during wind turbine operation, accurately perceive wind turbine operating status and environmental changes, and achieve adaptive control without frequent manual intervention. This greatly improves the automation level of wind turbine operation. It can automatically adjust the control strategy according to real-time operating conditions, ensuring that the wind turbine always operates in the optimal state, effectively improving power generation efficiency and reducing energy loss. The AI ​​large model has strong generalization capabilities, which can quickly adapt to different types and sizes of wind turbines, as well as diverse geographical environments and climatic conditions, enhancing the versatility and practicality of the method. It can predict potential wind turbine failures in advance, issue timely warnings, and take corresponding measures to reduce the probability of failure, extend the service life of wind turbines, and reduce maintenance costs. Through intelligent adaptive control, it can also reduce the impact of wind turbine operation on the power grid, improve grid stability, and provide strong support for the efficient utilization of renewable energy and the sustainable development of the energy system. Attached Figure Description

[0016] Figure 1 This is a diagram showing the stages of the method of the present invention; Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0018] Reference Figure 1 and Figure 2As shown, the AI-based large-model wind turbine adaptive control method simultaneously collects data from wind turbine sensors, lidar / ultrasonic environmental perception data, and grid dispatch signals, performing spatiotemporal alignment processing on the multi-source heterogeneous data. Specifically, the wind turbine sensor data comes from various monitoring devices installed at key locations on the turbine. Anemometers are typically deployed at the front of the hub to accurately capture incoming wind speed; speed signals come from the generator shaft encoder; the gearbox vibration spectrum is collected by an accelerometer on the casing surface; the main shaft torque is measured by shaft strain gauges; and the bearing temperature is monitored in real-time by built-in thermocouples. In the environmental perception data, lidar is generally installed at the top of the tower facing the incoming flow, capturing feedforward wind speed profiles, three-dimensional turbulence intensity matrices, and wind shear gradients hundreds of meters ahead, providing spatial distribution information for wind condition prediction. Grid dispatch signals are received in real-time through the power communication network, encompassing frequency regulation commands, active / reactive power setpoints, and real-time marginal electricity prices in the electricity market, ensuring the wind turbine responds to grid demands. Multi-source data is fused using Kalman filtering to generate a spatiotemporally aligned feature tensor. During the Kalman filtering process, the noise covariance matrix is ​​dynamically corrected based on the sensor data quality. When the signal-to-noise ratio (SNR) is below 20 dB, the filtering strength is automatically increased. This is because sensor measurement errors are large under low SNR, and enhanced filtering can effectively suppress noise interference and improve data reliability. The generated feature tensor is input to a sliding time window with an adjustable window length. The length of the sliding time window is adaptively adjusted according to wind conditions. In the stable wind speed range, a long window is used to smooth noise. This is because data fluctuations are gentle under stable wind speeds, and a long window can reduce the impact of random noise by accumulating samples. In the transitional gust range, a short window is automatically switched to improve response speed. This is because wind speed changes drastically during gusts, and a short window can quickly capture dynamic changes and avoid information lag caused by an excessively long window.

[0019] The preprocessed feature data is input into a wind power infrastructure model based on a deep learning architecture. This model is pre-trained using historical data from multiple wind farms and adapted to target turbine data, outputting wind condition prediction results and equipment status assessment values. The wind power infrastructure model adopts a hierarchical Transformer architecture. The Transformer was chosen because of its powerful ability to capture long-term time-series dependencies and model multi-feature interactions, effectively handling the complex spatiotemporal correlation characteristics in wind power data. During the pre-training stage, masking autoencoder technology is used to learn aerodynamic-mechanical coupling relationships from petabyte-level cross-wind farm data. Masking autoencoder randomly masks some input features and forces the model to learn and recover complete information, thereby enhancing the model's ability to extract and understand key features. The introduction of cross-wind farm data allows the model to learn common patterns under different regions and climatic conditions, improving its generalization ability. The adaptation process employs a dual optimization mechanism. The encoder layer weights are frozen because the encoder has already learned general wind power system feature representations during the pre-training stage. Freezing avoids overfitting caused by small-sample fine-tuning. At the same time, the multilayer perceptron is fine-tuned using labeled data of specific units to adapt the model to the individual characteristics of the target units. The loss function is a weighted sum of the power tracking mean square error and the tower bending moment spectrum damage value. This design aims to balance power generation performance and equipment lifespan. By adjusting the weights, different optimization objectives can be emphasized according to the wind farm's operational needs. For example, the damage value weight can be increased during the equipment aging stage to extend the lifespan.

[0020] Based on wind condition forecasts and equipment condition assessments, a set of multi-objective optimization control setpoints is generated using a basic model. This set includes the tip speed ratio curve, torque reference value, pitch angle sequence, and yaw correction. The multi-objective optimization control setpoints are generated through Pareto optimization. Pareto optimization is used because multiple objectives in wind power control often conflict; for example, maximizing power generation may lead to increased equipment fatigue damage. The Pareto optimal solution can find a non-dominated solution to the conflicting objectives, ensuring that improving the performance of one objective does not sacrifice the optimization space of other objectives. The objective function includes a term for maximizing power generation, a term for suppressing the cumulative rate of fatigue damage to key components, and a term for tracking error of grid dispatch instructions. These three terms correspond to the core requirements of wind power operation: improving economic efficiency, extending equipment life, and meeting grid requirements. The constraints include the pitch angle change rate, the rated value of the generator torque fluctuation band, and the yaw system response lag time. The pitch angle change rate limit is to avoid excessive mechanical shock caused by rapid movement of the pitch mechanism. The generator torque fluctuation band setting can protect the converter from severe current surges. The yaw system response lag time constraint takes into account the physical delay characteristics of the yaw drive mechanism to ensure that the generated control setpoints are executable in the actual system.

[0021] The Model Predictive Control (MPC) framework transforms the optimized control setpoints into coordinated action commands for the pitch, converter, and yaw systems, embedding mechanical load constraints and dynamic compensation during the transformation process. MPC employs rolling time-domain quadratic programming, solving for the desired result in each control cycle. ; In the formula, To control the amount of change in the input, For discrete time steps, To predict the length of the time domain, To control the length of the time domain, for Predicted power at time, for Reference power at time, This is the power tracking weight matrix. for The change in torque at any given time. This is the torque stability weighting matrix. This is the fatigue damage penalty coefficient. for The fatigue damage index at any given time. The rolling time domain design allows for continuous optimization within a finite prediction time domain, and, combined with real-time feedback information, corrects control decisions, effectively addressing uncertainties and dynamic changes in wind power systems. In the objective function, the power tracking weight matrix Q is used to emphasize the tracking accuracy of the reference power, ensuring that the wind turbine output meets grid dispatch requirements; the torque stability weight matrix R is used to suppress torque variations. Fluctuations, reducing mechanical stress in the transmission chain; fatigue damage penalty coefficient. The introduction of this factor balances the fatigue damage index. Prioritizing the optimization objectives by adjusting A reasonable balance can be achieved between power generation efficiency and equipment fatigue, ensuring that the control strategy is both efficient and safe.

[0022] Based on feedback from actual control performance and wind condition prediction deviations, the parameters of the wind power foundation model are updated through an online learning mechanism. This enables dynamic compensation for climate adaptability adjustments and equipment status changes, forming an adaptive control closed loop. The closed-loop evolution layer includes a model drift detection mechanism. When the relative error between the predicted power and the actual power exceeds a preset percentage for a consecutive preset number of samples, or when the load standard deviation rises above a predetermined threshold, an incremental learning process is triggered. This is because continuous prediction deviations or abnormal load fluctuations indicate that the model may no longer be able to accurately adapt to the current operating conditions, such as changes in wind characteristics due to climate change or performance degradation caused by equipment aging. An elastic weight solidification algorithm is adopted to lock important parameters and update only the weights of the fully connected layer. Elastic weight solidification can protect parameters that play a key role in core performance in the model, avoiding catastrophic forgetting caused by updates. Updating only the fully connected layer can significantly reduce computational overhead and ensure the real-time performance of online learning. After the update, the control stability is verified through digital twin simulation. The digital twin can simulate the actual operating environment of the wind turbine, testing the control effect of the updated model without affecting the field equipment, preventing equipment damage caused by control instability.

[0023] An edge-cloud collaborative architecture is deployed, with embedded modules on the edge performing data preprocessing and lightweight model inference. The edge devices are located close to the wind turbines, enabling low-latency processing of real-time data to meet the timeliness requirements of control command generation. The cloud training platform adopts a parameter server architecture, aggregating multi-wind turbine operating data to train the basic model in parallel. The aggregation of multi-wind farm data significantly improves the model's generalization ability, adapting it to the operating characteristics of different regions and turbine models. Differential privacy technology protects data security. Wind farm operating data contains commercially and technically sensitive information; differential privacy adds appropriate noise to the data to prevent data leakage without affecting model training performance. Model parameters are synchronized to the edge devices at preset intervals via the OPC UA protocol. OPC UA, as a universal communication protocol in industrial automation, possesses high reliability and compatibility, ensuring stable and accurate parameter transmission.

[0024] A robustness enhancement module is set up to switch to a backup controller based on Deep Deterministic Policy Gradient (DDPG) when the confidence level of the base model output is below 90% or conflicts with the aeroelastic equations. The 90% confidence threshold is based on extensive experimental verification. Below this value, the reliability of the model prediction decreases significantly. Conflicts with the aeroelastic equations indicate that the prediction results violate basic physical laws. Switching to the backup controller at this time can prevent control failure. The backup controller uses a simplified physical model as an environment simulator. The simplified model can reduce simulation complexity and improve controller training efficiency while ensuring basic physical characteristics. The output control commands satisfy the Lyapunov stability condition. The Lyapunov stability condition ensures that the control commands can bring the system state to a stable point, avoiding instability such as divergence or oscillation.

[0025] A life prediction subnetwork is coupled in the dynamic decision layer. This subnetwork receives the hidden layer features of the basic model and reuses the effective features extracted by the basic model to reduce redundant calculations and improve the efficiency and accuracy of life prediction. It outputs the gearbox remaining life index and the blade fatigue damage index. The execution control layer starts the load reduction mode when the life index is low or fatigue damage is aggravated, based on the life prediction results and the dynamic relaxation power tracking target. The torque command is limited to 70% of the rated value. The setting of 70% of the rated value takes into account the load reduction requirements and basic power generation guarantee. It can effectively reduce the equipment load and slow down fatigue accumulation, while maintaining a certain power generation benefit, balancing short-term power generation benefits and long-term equipment life.

[0026] A multiphysics verification closed loop is established, feeding back actual load data to a digital twin. The twin includes an aeroelastic model, a multibody dynamics model of the transmission chain, and a power grid impedance model. These models simulate the aerodynamic characteristics, mechanical transmission characteristics, and power grid interaction characteristics of the wind turbine, respectively, comprehensively covering the key physical processes of wind turbine operation. The twin simulation results are used to correct the multi-objective weight coefficients of the basic model. By comparing simulation and actual operating data, the priority of each optimization objective can be dynamically adjusted, making the control strategy more in line with actual operating requirements, forming a virtual-physical interactive optimization mechanism, and continuously improving control performance.

[0027] This embodiment is based on a 2.5MW permanent magnet direct-drive wind turbine in a wind farm. The specific parameters and operations are as follows: During the data acquisition phase, among the wind turbine's sensors, the anemometer sampling frequency was set to 10Hz (range 0-50m / s), the generator power sampling frequency was 5Hz (accuracy ±0.5%), the gearbox vibration spectrum was sampled by an accelerometer at 2kHz (range ±50g), the main shaft torque was acquired by strain gauges at 1Hz (range 0-30kN·m), and the bearing temperature thermocouple sampling frequency was 1Hz (accuracy ±0.3℃). The lidar was deployed at a height of 120m on the top of the tower, with a detection range of 200m, outputting a set of feedforward wind speed profiles every 10s (including 16 horizontal sampling points and 8 vertical sampling points), with a three-dimensional turbulence intensity matrix resolution of 5m×5m and a wind shear gradient calculation interval of 30s. The grid dispatch signal was updated every 2s via the OPC UA protocol, including a frequency regulation command response delay ≤100ms, an active power setpoint fluctuation range of ±50kW, and a real-time marginal electricity price update cycle of 15min. When multi-source data is fused by Kalman filtering, the filtering intensity is automatically increased by 40% when the signal-to-noise ratio of the anemometer drops to 18dB (below the 20dB threshold). The sliding time window is set to 30s in the stable section where the wind speed fluctuation is ≤2m / s, and automatically switches to 5s in the transition section of gusts (wind speed changes ≥5m / s within 10s).

[0028] The wind power basic model adopts an 8-layer hierarchical Transformer architecture. During the pre-training stage, 5 years of operational data from 10 cross-regional wind farms (a total of 8PB, including 28 types of features such as wind speed, power, and load) are input. The aerodynamic-mechanical coupling relationship is learned through mask autoencoder technology. When adapting to the target unit, the weights of the first 6 layers of encoders are frozen, and the last 2 layers of multilayer perceptron are fine-tuned using 3 months of labeled data (1.2TB, including 1200 sets of fault simulation samples) of the 2.5MW wind turbine. In the loss function, the power tracking mean square error weight is 0.6 and the tower bending moment spectrum damage value weight is 0.4. After adaptation, the model's wind condition prediction error is ≤4.8%, and the equipment condition assessment accuracy reaches 92.3%.

[0029] The MPC control loop is configured with a prediction time domain of Np = 10s and a control time domain of Nc = 3s. The Pareto optimization objective function weights are 0.5 for maximizing power generation, 0.3 for suppressing gearbox fatigue damage, and 0.2 for grid tracking. When the gearbox remaining life index output by the life prediction subnetwork drops to 0.6 (rated value 1.0), a load reduction mode is triggered, and the torque command is limited to 12.6 kN·m (70% of the rated torque of 18 kN·m for a 2.5MW wind turbine). The online learning trigger threshold is set as follows: the relative error between predicted and actual power for 15 consecutive samples is >5%, or the standard deviation of the main shaft load increases by >8%. After parameter updates, the control command fluctuation is verified by digital twin simulation, and the Lyapunov stability condition is met. After 3 months of field operation, the wind turbine's power generation efficiency is improved by 3.2% compared to traditional control, the gearbox fatigue damage rate is reduced by 18%, and the grid command tracking error is reduced to 2.1%.

[0030] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. An adaptive control method for wind turbines based on AI large-scale models, characterized in that, Includes the following steps: S1. Synchronously collect multi-source heterogeneous data including: wind turbine body sensor data, lidar / ultrasonic environmental perception data and power grid dispatch signals, and perform spatiotemporal alignment processing on the multi-source heterogeneous data; S2. Based on the spatiotemporal alignment of the multi-source heterogeneous data, preprocessing is performed, and the preprocessed feature data is input into the wind power basic model based on the deep learning architecture. The model is pre-trained with historical data across wind farms and adapted with target unit data to output wind condition prediction results and equipment status evaluation values. S3. Based on the wind condition prediction results and equipment status assessment values, generate a set of multi-objective optimization control setpoints through the basic model, including tip speed ratio curve, torque reference value, pitch angle sequence and yaw correction. S4. The optimized control setpoints are transformed into coordinated action commands for pitch, converter, and yaw systems using the Model Predictive Control (MPC) framework, with mechanical load constraints and dynamic compensation embedded in the transformation process. S5. Based on the feedback of actual control effect and wind condition prediction deviation, update the parameters of the wind power foundation model to achieve dynamic compensation for climate adaptability adjustment and equipment status changes, forming an adaptive control closed loop.

2. The adaptive control method for wind turbines based on AI large model according to claim 1, characterized in that, The wind turbine body sensor data includes anemometer speed, generator power, gearbox vibration spectrum, main shaft torque, and bearing temperature; the environmental perception data includes feedforward wind speed profiles, three-dimensional turbulence intensity matrices, and wind shear gradients captured by lidar; the power grid dispatch signals cover frequency regulation commands, active / reactive power setpoints, and real-time marginal electricity prices in the electricity market; the multi-source data are fused by Kalman filtering to generate a spatiotemporally aligned feature tensor, with an adjustable sliding time window as the input window length; The length of the sliding time window is adaptively adjusted according to wind conditions: a long window is used to smooth noise in the stable wind speed range, and a short window is automatically switched to improve the response speed in the gust transition range; the process noise covariance matrix of the Kalman filter is dynamically corrected according to the sensor data quality, and the filtering intensity is automatically increased when the signal-to-noise ratio is lower than 20dB.

3. The adaptive control method for wind turbines based on AI large model according to claim 1, characterized in that, The wind power basic model adopts a hierarchical Transformer architecture. During the pre-training stage, the aerodynamic-mechanical coupling relationship is learned from petabyte-level cross-wind field data using mask autoencoder technology. The adaptation process employs a dual optimization mechanism: freezing the encoder layer weights, fine-tuning the multilayer sensor using specific unit labeled data, and using the weighted sum of the power tracking mean square error and the tower bending moment spectrum damage value as the loss function.

4. The adaptive control method for wind turbines based on AI large model according to claim 1, characterized in that, The multi-objective optimization control setpoints are generated through Pareto optimization. The objective function includes: a power generation maximization term, a key component fatigue damage accumulation rate suppression term, and a grid dispatch command tracking error term. The constraints include the pitch angle change rate, the generator torque fluctuation band rating, and the yaw system response lag time.

5. The AI-based large-model adaptive control method for wind turbines according to claim 1, characterized in that, The Model Predictive Control (MPC) employs rolling time-domain quadratic programming, which is solved in each control cycle. ； In the formula, To control the amount of change in the input, For discrete time steps, To predict the length of the time domain, To control the length of the time domain, for Predicted power at time, for Reference power at time, This is the power tracking weight matrix. for The change in torque at any given time. This is the torque stability weighting matrix. This is the fatigue damage penalty coefficient. for The fatigue damage index at any given time.

6. The AI-based large-model adaptive control method for wind turbines according to claim 1, characterized in that, Step S5 updates the parameters of the wind power foundation model through an online learning mechanism, which includes the following operations: The relative error between predicted power and actual power and the standard deviation of equipment load are calculated in real time through the closed-loop evolution layer. When the relative error between predicted power and actual power exceeds a preset percentage or the standard deviation of load rises above a predetermined threshold for a consecutive preset number of samples, the incremental learning process is triggered. An elastic weight fixation algorithm is adopted to lock the pre-trained weights of the Transformer encoder layer in the wind power foundation model and only update the weights of the fully connected layer of the model output layer. The updated model parameters are input into the digital twin, and the load fluctuation of the control command is simulated and verified. If the condition is met, the update is executed; otherwise, the original parameters are returned.

7. The AI-based large-model adaptive control method for wind turbines according to claim 1, characterized in that, The deployment of an edge-cloud collaborative architecture involves embedded modules on the edge side performing data preprocessing and lightweight model inference; the cloud training platform adopts a parameter server architecture, aggregating multi-wind turbine operation data to train the basic model in parallel, and protecting data security through differential privacy technology. Model parameters are synchronized to the edge device via the OPC UA protocol at preset intervals.

8. The adaptive control method for wind turbines based on AI large model according to claim 1, characterized in that, A robustness enhancement module is set up so that when the confidence level of the basic model output is lower than 90% or conflicts with the aeroelastic equation, it switches to a backup controller based on the deep deterministic policy gradient (DDPG). The backup controller uses a simplified physical model as an environment simulator and outputs control commands that satisfy the Lyapunov stability conditions.

9. The AI-based large-model adaptive control method for wind turbines according to claim 1, characterized in that, A life prediction subnetwork is coupled in the dynamic decision layer. This subnetwork receives the hidden layer features of the basic model and outputs the gearbox remaining life index and the blade fatigue damage index. The execution control layer, based on the dynamic relaxation power tracking target, initiates the load reduction mode at the appropriate time, limiting the torque command to 70% of the rated value.

10. The AI-based large-model adaptive control method for wind turbines according to claim 1, characterized in that, A multiphysics verification closed loop is established, and the actual load data is fed back to the digital twin. The twin includes an aeroelastic model, a transmission chain multibody dynamics model, and a power grid impedance model. The twin simulation results are used to correct the multi-objective weight coefficients of the basic model, forming a virtual-entity interactive optimization mechanism.

Images