Wind turbine generator yaw error prediction control method and system based on deep learning

Abstract

The invention discloses a wind turbine generator yaw error prediction control method and system based on deep learning, and the method comprises the steps: firstly processing multi-dimensional original sensor data through a self-adaptive denoising module, separating out a pure physical signal, and obtaining an instantaneous difference signal; furthermore, the difference signal is modeled in combination with the operation context of the unit, and the expected noise characteristic under the current working condition is predicted, so that the attributive noise level estimation is generated, and the normal operation fluctuation and the abnormal interference are effectively distinguished. On the basis, de-noising input and noise level estimation are cooperatively input into a Bayesian prediction model, prediction yaw error distribution containing a mean value sequence and a variance sequence is output, and explicit quantization of prediction result uncertainty is achieved. Finally, control risk assessment is performed based on the distribution, and adaptive control parameters are dynamically synthesized to optimize yaw action, so that the accuracy and safety of a control decision in an uncertain environment are ensured while data noise interference is suppressed.

Description

Technical Field

[0001] This application relates to the field of wind power generation technology, and more specifically, to a method and system for predicting and controlling the yaw error of wind turbine generators based on deep learning. Background Technology

[0002] As a crucial component of renewable energy, wind power's operational efficiency directly depends on the wind turbine's ability to capture wind energy. The yaw system, a key subsystem of the wind turbine, is responsible for adjusting the nacelle angle to track wind direction changes in real time, thereby maximizing power generation and reducing structural loads caused by off-wind conditions. Achieving precise yaw control requires accurate and real-time acquisition of environmental parameters such as wind speed and direction, as well as the turbine's own operating status. However, wind turbines typically operate in harsh environments such as open fields and offshore, affected by turbulence, wake effects, mechanical vibrations, and extreme weather. The data collected by sensors often contains complex noise and significant uncertainties. These data quality issues directly interfere with the control system's judgment of the actual operating conditions. If the control strategy cannot effectively handle these uncertainties, it will lead to frequent or delayed yaw actions, increased mechanical wear, and reduced power generation.

[0003] While existing yaw control methods have begun to incorporate machine learning or deep learning techniques to improve their ability to predict wind condition changes, they still have limitations in handling non-ideal sensor data. Traditional filtering methods are prone to signal lag or over-smoothing, resulting in the loss of crucial transient information. Furthermore, most existing data-driven prediction models focus only on point prediction accuracy, lacking a quantification mechanism for the reliability of the prediction results themselves—meaning they cannot inform the control system how confident the current prediction value is. More critically, when dealing with data noise, existing technologies typically treat the difference between the original signal and the smoothed signal as homogeneous random noise, ignoring the significant impact of the wind turbine's operating context on noise patterns. For example, normal signal fluctuations generated when the turbine performs pitch or yaw maneuvers may have similar statistical characteristics to abnormal fluctuations caused by sensor malfunctions or sudden environmental changes, but their physical meanings are entirely different. Existing technologies struggle to distinguish whether such differences originate from normal operating conditions or abnormal disturbances, lacking attribution-based uncertainty analysis capabilities. This lack of understanding of noise sources and uncertainties makes it difficult for control systems to formulate robust control strategies that balance safety and economy when facing complex and ever-changing wind farm environments.

[0004] Therefore, an optimized deep learning-based yaw error prediction and control scheme for wind turbines is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a deep learning-based method and system for predicting and controlling the yaw error of wind turbine generators.

[0006] According to one aspect of this application, a deep learning-based method for predicting and controlling the yaw error of a wind turbine is provided, comprising: Acquire raw sensor data, including real-time wind speed, wind direction, unit power, yaw angle, yaw speed, blade pitch angle, and main shaft speed; The raw sensor data is input into the adaptive denoising module to obtain the denoised input. Based on the denoised input and raw sensor data, a noise level estimate is determined; The denoised input and noise level estimate are input into the Bayesian prediction model to perform uncertainty-aware yaw prediction and control decision-making to obtain yaw control commands.

[0007] According to another aspect of this application, a deep learning-based wind turbine yaw error prediction and control system is provided, comprising: The data acquisition module is used to acquire raw sensor data, which includes real-time wind speed, wind direction, unit power, yaw angle, yaw speed, blade pitch angle, and main shaft speed. The noise reduction input acquisition module is used to input the raw sensor data into the adaptive noise reduction module to obtain the noise reduction input. The noise level estimation and determination module is used to determine the noise level estimate based on the denoised input and the raw sensor data; The uncertainty-aware yaw prediction and control decision module is used to input the denoised input and noise level estimate into the Bayesian prediction model to perform uncertainty-aware yaw prediction and control decision to obtain yaw control commands.

[0008] Compared to existing technologies, this solution first utilizes an adaptive denoising module to process the acquired raw sensor data, separating high-fidelity physical signals. Then, it models signal differences based on the unit's operating context, generating a noise level estimate that includes predictive noise characteristics and residual uncertainties. Subsequently, the denoised input and the noise level estimate are jointly input into a Bayesian prediction model to obtain the probability distribution of future yaw errors, thereby explicitly quantifying the uncertainty of the prediction results. Based on this, the system performs risk assessment according to the statistical characteristics of the prediction distribution and dynamically adjusts control strategy parameters under safety constraints to generate the optimal yaw maneuver that balances performance and risk. This solution achieves adaptive management of control risks in noisy environments through end-to-end uncertainty quantification and propagation, ensuring the accuracy of yaw control commands and the robustness of system operation. 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 deep learning-based wind turbine yaw error prediction and control method according to an embodiment of this application; Figure 2 This is a schematic diagram of the data flow of a deep learning-based wind turbine yaw error prediction and control method according to an embodiment of this application. Figure 3 This is a flowchart illustrating the process of determining noise level estimation based on denoised input and raw sensor data in the deep learning-based wind turbine yaw error prediction and control method according to embodiments of this application. Figure 4 The flowchart shows the process of inputting denoised input and noise level estimate into a Bayesian prediction model to perform uncertainty-aware yaw prediction and control decision to obtain yaw control command in the deep learning-based wind turbine yaw error prediction and control method according to the embodiments of this application. Figure 5 This is a flowchart illustrating the process of performing control risk assessment and adaptive strategy parameterization on the predicted yaw error distribution based on control objectives and safety constraints to obtain risk measurement and adaptive control parameters according to the deep learning-based wind turbine yaw error prediction and control method of this application embodiment. Figure 6 This is a block diagram of a deep learning-based wind turbine yaw error prediction and control system 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] Currently, wind turbines operate in complex and variable environments, and sensor data is often accompanied by significant noise and uncertainty. Existing processing methods typically ignore the impact of operating context on noise patterns, making it difficult to effectively distinguish between normal operating condition fluctuations and abnormal disturbances. This results in the control system lacking the ability to perceive and attribute uncertainties, thus affecting the accuracy and stability of yaw control. Therefore, this application proposes a deep learning-based yaw error prediction and control method for wind turbines. Specifically, this method first uses an adaptive denoising module to process the raw sensor data, separating the pure physical signal while calculating the instantaneous difference between the raw and denoised signals. Subsequently, a context-aware noise pattern prediction mechanism is introduced to predict the expected noise characteristics based on the current operating state of the turbine, thereby further decomposing the instantaneous difference into interpretable background noise and residual uncertainty indicating anomalies, generating a noise level estimate with attributional significance. Based on this, the denoised physical characteristics and the noise level estimate are input into a Bayesian prediction model to obtain the probability distribution of future yaw errors, explicitly quantifying the confidence level of the prediction results. Ultimately, the control system performs risk assessment based on the variance characteristics of the predicted distribution and dynamically adjusts the aggressiveness of the control strategy under safety constraints to generate the optimal yaw control command that can adaptively balance power generation efficiency and load risk, thereby solving the control failure problem under noise interference.

[0016] Figure 1 This is a flowchart of a deep learning-based wind turbine yaw error prediction and control method according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow for a deep learning-based wind turbine yaw error prediction and control method according to an embodiment of this application. Figure 1 and Figure 2As shown, the deep learning-based yaw error prediction and control method for wind turbines according to an embodiment of this application includes the following steps: S100, acquiring raw sensor data, including real-time wind speed, wind direction, turbine power, yaw angle, yaw speed, blade pitch angle, and main shaft speed; S200, inputting the raw sensor data into an adaptive denoising module to obtain denoised input; S300, determining a noise level estimate based on the denoised input and the raw sensor data; S400, inputting the denoised input and the noise level estimate into a Bayesian prediction model to perform uncertainty-aware yaw prediction and control decision-making to obtain yaw control commands.

[0017] Specifically, in step S100, raw sensor data is acquired, including real-time wind speed, wind direction, turbine power, yaw angle, yaw speed, blade pitch angle, and main shaft speed. It is understood that since wind turbine yaw control is a nonlinear dynamic process affected by multivariate coupling, single-dimensional wind direction data cannot comprehensively and accurately reflect the real-time aerodynamic response characteristics and mechanical load state of the turbine in complex flow field environments. Furthermore, data from a single source is highly susceptible to environmental noise interference and measurement errors, leading to insufficient accuracy in state estimation. Therefore, in the technical solution of this application, raw sensor data is acquired, including real-time wind speed, wind direction, turbine power, yaw angle, yaw speed, blade pitch angle, and main shaft speed, to construct a full-dimensional state space encompassing external environmental excitation and internal turbine response. This fully utilizes the physical correlation between multi-source heterogeneous data to provide rich information redundancy for subsequent data processing, ensuring that the prediction model can accurately capture the intrinsic relationship between wind condition changes and turbine dynamic behavior, thereby laying a solid data foundation for achieving high-precision robust yaw optimization control.

[0018] More specifically, in a particular example of this application, the data acquisition operation begins by establishing a data interaction channel with the wind turbine main controller via an industrial fieldbus or industrial Ethernet communication protocol. This channel is then used to periodically read real-time measurement signals from physical sensors deployed at various key monitoring points of the turbine. For acquiring meteorological environmental parameters, analog voltage or digital signals from mechanical or ultrasonic anemometers and wind vanes mounted on the nacelle top are collected and interpreted into real-time wind speed and direction values ​​to characterize the external input conditions of the current flow field. For monitoring the turbine's energy conversion status, power data is directly read from the converter controller or power monitoring instruments on the generator side to reflect the turbine's current energy capture efficiency. Simultaneously, for capturing mechanical motion, the current yaw angle and yaw speed are obtained by reading absolute encoders or resolvers mounted on the yaw drive mechanism, the blade pitch angles of the three blades are obtained from angle sensors inside the pitch system, and the main shaft speed is obtained from speed sensors mounted on the drive train main shaft or gearbox high-speed shaft side. After completing the reading of the above multidimensional data, all the collected heterogeneous data are assigned a unified time reference stamp and strictly aligned and synchronized in time sequence, thereby integrating them into a time-synchronized raw sensor data sequence containing multidimensional physical quantities.

[0019] Specifically, in step S200, the raw sensor data is input into an adaptive denoising module to obtain denoised input. It is understandable that due to the extreme complexity and time-varying nature of the wind turbine operating environment, the raw data collected by sensors inevitably contains non-stationary noise generated by gust turbulence, tower vibration, electromagnetic interference, and sensor drift. These noise components can mask the physical signals reflecting the true aerodynamic state of the turbine. If directly used for control decisions, this can lead to malfunctions in the yaw system due to high-frequency interference, resulting in unnecessary mechanical wear and load impacts. Therefore, in the technical solution of this application, the raw sensor data is further input into an adaptive denoising module to obtain denoised input. This utilizes the nonlinear feature extraction capability of deep neural networks to intelligently separate and filter out random fluctuations unrelated to the current physical process from the contaminated time series, while retaining key dynamic features such as wind speed change trends and yaw error accumulation. This significantly improves the signal-to-noise ratio and physical fidelity of the input data, building a high-quality data foundation for subsequent accurate prediction and robust control, ensuring that control commands only respond to real operating condition changes rather than false noise disturbances.

[0020] More specifically, in a concrete example of this application, the adaptive denoising module is implemented based on a sequence-to-sequence Transformer deep learning architecture. This process first normalizes the synchronized multidimensional original sensor data sequence and maps it to a high-dimensional feature vector to meet the input requirements of the neural network. Then, an encoder incorporating a multi-head self-attention mechanism performs a global context scan of the input sequence. By calculating the attention weights between different time steps and different sensor channels, it dynamically identifies effective signal patterns with long-range dependencies and physical consistency in the data, while suppressing irrelevant random noise features. Next, the decoder layer reconstructs the extracted deep features into time-series data with the same dimension as the original input but with noise removed. In this process, the model utilizes parameters pre-trained on a large historical dataset containing real-world noise conditions to adaptively adapt to the specific noise distribution under the current wind conditions, ultimately outputting a denoised input sequence that accurately represents the current physical state of the wind turbine.

[0021] Specifically, in step S300, a noise level estimate is determined based on the denoised input and the original sensor data. It is understandable that the operating environment of wind turbines is extremely complex and rapidly changing. Wind speed, wind direction, turbulence intensity, and the turbine's own yaw and pitch maneuvers are all in continuous dynamic change. These changes in environmental and operating conditions directly alter the noise characteristics of the sensor data. For example, in highly turbulent wind fields or when the turbine performs rapid yaw operations, the inherent volatility of the sensor data is significantly enhanced, thus increasing the difference between the original signal and the denoised signal. Relying solely on simple statistical methods to process these differences will not effectively distinguish whether such differences are caused by normal and expected fluctuations strongly correlated with the current operating context, or by unexpected abnormal noise or potential sensor malfunctions. This context-independent difference estimation method treats all signal differences as homogeneous uncertainty, failing to capture the core, specific relationship—the complex dynamic causal relationship or strong correlation between the key operating parameters and environmental factors of the wind turbine and the noise and uncertainty characteristics in its sensor data.

[0022] Therefore, in the technical solution of this application, the instantaneous difference signal is further calculated based on the denoised input and the original sensor data, and the noise pattern prediction of the instantaneous difference signal is performed in combination with the operating context data to obtain the predicted noise characteristics and residual uncertainty. Then, the attribution uncertainty of the denoised input is integrated to determine the noise level estimate. In this way, context-aware noise pattern modeling and attribution uncertainty quantification are introduced. Instead of generalizing the difference between the original signal and the denoised signal, the solution actively understands and predicts the specific characteristics and degree of these differences in a specific operating context. This approach enables the establishment of explicit correlations between the wind turbine's operating context and sensor noise patterns. It allows for the explicit modeling and quantification of these deep correlations, thereby achieving attributable uncertainty quantification. This allows the system to provide refined and attributable clues when facing multiple and diverse factors, including inherent random measurement noise from sensor equipment, physical fluctuations induced by harsh environments, transient interference during turbine operation, and systematic biases caused by sensor malfunctions, aging, or drift. This effectively addresses the problem that traditional mechanisms fail to provide effective methods for identifying and attributing these different types and sources of differences, enhancing the entire predictive control system's deep understanding of real uncertainty and its robust decision-making capabilities.

[0023] Figure 3 This document describes a flowchart for determining noise level estimation based on denoised input and raw sensor data in a deep learning-based wind turbine yaw error prediction and control method according to embodiments of this application. (See attached flowchart.) Figure 3 As shown, step S300 includes: S310, calculating the instantaneous difference signal between the denoised input and the original sensor data; S320, based on the operating context data, performing noise pattern prediction on the instantaneous difference signal and the original sensor data to obtain the predicted noise characteristics and residual uncertainty; S330, based on the predicted noise characteristics and residual uncertainty, performing attribution uncertainty integration on the denoised input to obtain a noise level estimate.

[0024] In step S310, the instantaneous difference signal between the denoised input and the original sensor data is calculated. It is understood that to ensure the accuracy of the prediction results, it is necessary to effectively separate the core physical signal and quantify the instantaneous fluctuations in the original data. However, while simple denoising can extract the denoised input representing the pure physical state of the wind turbine, it often ignores the wind condition characteristics and equipment status information contained in the stripped noise components. This results in the system lacking an intuitive understanding of all the fluctuation components that the denoising model has not fully explained or eliminated. Therefore, in the technical solution of this application, the instantaneous difference signal between the denoised input and the original sensor data is further calculated, that is, the mathematical difference between the original sensor data and its denoised version is calculated simultaneously to precisely capture the instantaneous fluctuations in the original signal. This ensures that the difference is not just a simple residual, but a direct quantification of the unexplained or uncertain parts of the physical signal at the current time step. By accurately separating the pure signal and obtaining the instantaneous difference, a foundation is laid for subsequent context-aware analysis of uncertainty, making uncertainty no longer a fuzzy statistical quantity, but a traceable and analyzable fluctuation characteristic.

[0025] More specifically, in a particular example of this application, after receiving real-time synchronous sensor data from a wind turbine, the data is input into a pre-trained and deployed deep learning denoising model. This model, through its nonlinear mapping capabilities, identifies and suppresses various noise components in the signal while preserving the physically meaningful true signal to the maximum extent possible. Subsequently, point-by-point differential operations are performed on each sensor channel at each time step to generate these point-by-point differences that constitute a time series of instantaneous difference signals.

[0026] In step S320, noise pattern prediction is performed on the instantaneous difference signal and the original sensor data based on the operating context data to obtain the predicted noise characteristics and residual uncertainty. It is understood that the intrinsic causes of signal uncertainty cannot be fully understood solely based on the instantaneous difference, as the noise and fluctuations in sensor data are not static but have complex dynamic relationships with the operating conditions of the turbine and external environmental conditions. For example, mechanical vibrations generated during yaw and pitch control are expected background noise, while sudden signal jumps during stable operation may indicate a fault. Although the instantaneous difference amplitudes may be similar, their physical properties are completely different. Therefore, the technical solution of this application further aims to achieve attribution of uncertainty, that is, to construct a deep learning prediction model based on the operating context data, learn and explicitly model how the operating context affects the noise characteristics of the sensor data, and calculate the residual uncertainty accordingly, thereby establishing an explicit correlation between the wind turbine operating context and the sensor noise pattern. This enables the quantification of uncertainties with attribution, thereby distinguishing between normal fluctuations and abnormal signals. This provides a more refined and reliable attribution capability for subsequent prediction and control decisions, ensuring that the control system will not misjudge normal sensor fluctuations caused by high turbulent wind conditions as system faults, nor will it ignore abnormal small drifts under calm wind conditions.

[0027] More specifically, in a concrete example of this application, this step first encodes contextual information, acquiring and processing key auxiliary data that comprehensively characterizes the current operating status and environmental conditions of the wind turbine in real time, such as wind speed, wind direction, yaw angle, and power generation, transforming them into feature vectors that can be effectively processed by a deep learning model. For example, the current power level of the turbine, the nacelle vibration amplitude, and the environmental turbulence intensity transmitted from the external meteorological tower are encoded into high-dimensional context vectors. Subsequently, these encoded contextual features, along with a sequence of historical instantaneous differences within a preset sliding window, are used as input to a carefully trained noise pattern predictor. This predictor employs a long short-term memory network or a gated recurrent unit architecture to capture long-term dependencies in the time series. The loss function used during model training employs mean squared error loss, updating through backpropagation by minimizing the difference between the predicted noise characteristics and the actual observed noise statistics. The expected noise characteristics of each sensor channel under the current context conditions are predicted in real time, such as predicting the expected mean noise of each channel. and prediction noise variance .

[0028] In step S330, based on the predicted noise characteristics and residual uncertainty, the denoised input is integrated with attribution uncertainty to obtain a noise level estimate. It is understood that, to fully utilize the pure physical signals generated in the first two steps and the uncertainty information with attribution significance, isolated denoised signals or single residual statistics are insufficient to support complex decision-making logic. The final step aims to effectively fuse these heterogeneous information sources to construct a comprehensive and robust feature set. Therefore, in the technical solution of this application, the denoised input is further integrated based on the predicted noise characteristics and residual uncertainty to achieve explicit modeling of the specific relationship between wind turbine operating parameters and sensor noise, enabling the differentiation between data fluctuations that are inherent to the system and interpretable, and those that are unexpected anomalies. This ensures that the generated uncertainty assessment vector provides a multi-dimensional, attribution-based noise level assessment for the subsequent Bayesian deep prediction network. This allows the network to not only obtain the error value but also accurately quantify the confidence interval of the prediction when making yaw error predictions. This mechanism not only enhances the noise resistance and generalization performance of the prediction model but also makes the entire predictive control system more intelligent and robust at the decision-making level. For example, when the uncertainty prediction is high but attributed to the normal operating context, the control system can follow the prediction. When the residual uncertainty is high, indicating a potential anomaly, it can adaptively adopt a more conservative strategy or provide early warning. Thus, in challenging real-world industrial environments, the yaw control system of wind turbines can significantly improve operational safety, reliability, and adaptability while maximizing power generation efficiency.

[0029] More specifically, in a specific example of this application, this step first integrates the predicted noise variance, standardized residuals, and optional external sensor health indicators into a final attribution uncertainty vector, i.e., a noise level estimate, through feature concatenation.

[0030] Specifically, in step S400, the denoised input and noise level estimate are input into the Bayesian prediction model for uncertainty-aware yaw prediction and control decision-making to obtain yaw control commands. It is understandable that traditional deterministic predictive control methods can only output a single yaw error prediction value, failing to quantify the reliability of the prediction result. Especially when high environmental noise or abnormal residual uncertainty is identified in previous steps, if the control system blindly follows the point prediction value to execute yaw actions, it can easily lead to the unit tracking false wind direction fluctuations, resulting in frequent actuation of the yaw system and excessive fatigue of mechanical components. Therefore, in the technical solution of this application, the denoised input and noise level estimate are further input into the Bayesian prediction model for uncertainty-aware yaw prediction and control decision-making to obtain yaw control commands. This utilizes the probabilistic inference capability of the Bayesian deep learning network to map the input physical characteristics and the attributed uncertainty information into a probability distribution of future yaw errors, rather than a single numerical value. In this way, the mean and variance of the prediction results can be explicitly obtained, thereby enabling the control system to have risk perception capabilities. That is, when the prediction uncertainty is low, an aggressive strategy is adopted to maximize wind capture efficiency, while when the prediction uncertainty is high, an adaptive switch to a conservative strategy is made to avoid load risks, and finally the optimal yaw control command that takes into account both power generation performance and operational safety is generated.

[0031] Figure 4 This is a flowchart illustrating the process of inputting denoised input and noise level estimate into a Bayesian prediction model to perform uncertainty-aware yaw prediction and control decision-making to obtain yaw control commands, according to an embodiment of this application. Figure 4 As shown, step S400 includes: S410, inputting the denoised input and noise level estimate into the Bayesian prediction model to obtain the predicted yaw error distribution; S420, based on the control objective and safety constraints, performing control risk assessment and adaptive strategy parameterization on the predicted yaw error distribution to obtain risk metrics and adaptive control parameters; S430, based on the adaptive control parameters and safety constraints, performing robust control action optimization on the predicted yaw error distribution to obtain the optimal yaw action; S440, generating yaw control commands based on the optimal yaw action and risk metrics to obtain yaw control commands.

[0032] In step S410, the denoised input and noise level estimate are input into the Bayesian prediction model to obtain the predicted yaw error distribution. It is understood that due to the high randomness of the wind farm environment and the residual noise that may still exist in sensor data even after processing, a single deterministic point prediction cannot reflect the model's grasp of the future unit state. If the control system blindly executes yaw actions when the prediction confidence is low, it is highly likely that the unit will suffer additional mechanical loads or lose power generation due to misjudging the wind direction trend. Therefore, in the technical solution of this application, the denoised input and noise level estimate are further input into the Bayesian prediction model to obtain the predicted yaw error distribution. This utilizes a probabilistic deep learning framework to map the input physical characteristics and attribution uncertainty information into a probability density function of the future yaw error, rather than a single numerical estimate. This allows for the acquisition of a predicted mean sequence and variance sequence covering multiple future time steps, thereby not only predicting the evolution trend of the yaw error but also explicitly quantifying the reliability of the prediction result at different time scales, providing a rigorous statistical basis for subsequent risk assessment and robust decision-making.

[0033] More specifically, in a concrete example of this application, the execution flow of the Bayesian prediction model is based on a deep probabilistic inference mechanism. This process first receives a time-robust feature set composed of a denoised physical signal sequence and a noise level estimation vector. Next, a Long Short-Term Memory network or a Transformer encoder is used to extract the time dependencies and dynamic evolution patterns in the feature sequence, mapping them to a high-dimensional hidden state vector. To capture the uncertainty of the model and the data, the model employs a Monte Carlo Dropout mechanism or variational inference strategy during the inference phase. During the training phase, the model uses negative log-likelihood as the loss function to simultaneously optimize the accuracy of the prediction mean and the reasonableness of the variance. That is, while keeping the input features unchanged, multiple forward propagation operations are performed by randomly discarding some neuron connections in the neural network or sampling from the learned parameter distribution. Each forward propagation generates a possible future yaw error trajectory, and after a preset number of samplings, a prediction sample set containing multiple possibilities is formed. Finally, statistical analysis is performed on the sample set at each time point in the next N prediction time steps. The arithmetic mean of all samples is calculated as the prediction mean at that time, and the variance of the samples is calculated as the prediction uncertainty measure at that time. Thus, the prediction yaw error distribution containing the mean sequence and variance sequence of the next N time steps is output.

[0034] In step S420, based on the control objective and safety constraints, the predicted yaw error distribution is subjected to control risk assessment and adaptive strategy parameterization to obtain risk metrics and adaptive control parameters. It is understandable that while the simple predicted mean and variance reveal the future evolution trend and statistical uncertainty of the yaw error, they are not directly related to the specific operating limitations and multiple optimization objectives of the wind turbine. If the control system still uses fixed-weight control logic, it cannot automatically converge the control amplitude to avoid load exceeding limits when facing high-risk predictions, nor can it fully release the yaw potential to improve wind capture efficiency under low-risk conditions. Therefore, in the technical solution of this application, based on the control objective and safety constraints, the predicted yaw error distribution is further subjected to control risk assessment and adaptive strategy parameterization to obtain risk metrics and adaptive control parameters. This maps the abstract probability distribution information to specific engineering risk levels, and dynamically reconstructs the weight configuration and penalty coefficients within the control algorithm accordingly. In this way, the control system can adjust its decision-making tendency in real time according to the current prediction confidence and safety margin. When the prediction is accurate and the safety margin is sufficient, it adopts aggressive parameters to maximize power generation, while automatically switching to conservative parameters to ensure unit safety when the uncertainty diverges or approaches the safety boundary, thereby achieving adaptive and robust adjustment for complex operating conditions.

[0035] Figure 5 This document describes a flowchart illustrating the process of assessing control risk and parameterizing adaptive strategies based on control objectives and safety constraints in a deep learning-based wind turbine yaw error prediction and control method according to embodiments of this application. The flowchart describes the process to obtain risk metrics and adaptive control parameters by performing control risk assessment and adaptive strategy parameterization on the predicted yaw error distribution. Figure 5 As shown, step S420 further includes: S421, performing prediction uncertainty aggregation and risk level quantification on the predicted yaw error distribution to obtain the aggregated uncertainty measure and risk level; S422, calculating the safety margin deviation probability on the predicted yaw error distribution based on safety constraints to obtain the safety margin deviation probability; S423, dynamically synthesizing adaptive control parameters from the safety margin deviation probability, risk level, and control objective to obtain adaptive control parameters.

[0036] In step S421, the predicted yaw error distribution is subjected to prediction uncertainty aggregation and risk level quantification to obtain an aggregated uncertainty measure and risk level. It is understandable that, since the yaw error distribution output by the Bayesian prediction model contains a variance sequence of multiple future time steps, this high-dimensional temporal uncertainty information, while detailed, would increase the computational burden of the decision-making logic and make it difficult to form a unified risk perception if directly used for rapid adjustment of real-time control parameters. In particular, a sudden increase in uncertainty at a certain point in the prediction time domain might be averaged out and ignored, leading to a lack of vigilance against potential local prediction failures in the control system. Therefore, in the technical solution of this application, the predicted yaw error distribution is further subjected to prediction uncertainty aggregation and risk level quantification to obtain an aggregated uncertainty measure and risk level. This compresses the discrete uncertainty data distributed along the time axis into a scalar index and graded state characterizing the overall credibility of the current prediction period. This provides a global risk criterion for subsequent adaptive control, ensuring that the switching of control strategies is based on the worst-case scenario or comprehensive risk level within the entire prediction window, rather than fluctuations at a single moment, thereby grasping the risk attributes of the control task at a macroscopic level.

[0037] More specifically, in a concrete example of this application, this step first extracts the yaw error variance sequence for the future prediction time domain from the distribution information output by the Bayesian prediction model. This sequence quantifies the dispersion of the yaw error estimate at each prediction step. Then, to capture the maximum risk exposure within the prediction time domain, an aggregation operation is performed on this variance sequence, selecting the maximum value or weighted average value as the aggregated uncertainty measure. This measure intuitively reflects the uncertainty level or overall dispersion trend at the most unreliable moment of the model prediction over a future period. Next, the calculated aggregated uncertainty measure is numerically compared with a set of preset risk threshold intervals. Based on the interval range into which the measure value falls, it is mapped to a discrete risk level, such as classifying it as low-risk, medium-risk, or high-risk. This process effectively transforms continuous numerical uncertainty statistics into classification labels for logical decision-making, thereby clearly indicating whether the reliability of the prediction results under the current operating conditions is sufficient to support aggressive control actions, or whether defensive measures are needed due to prediction divergence.

[0038] In step S422, the safety margin deviation probability is calculated based on the predicted yaw error distribution according to safety constraints. It is understood that, due to the limitations of the wind turbine's yaw system by physical mechanical characteristics and operational safety regulations, there are strict yaw angle boundaries, maximum yaw rate limits, and allowable maximum yaw error thresholds. Simple variance statistics only reflect the dispersion of the prediction results, but fail to quantify the actual risk of the unit touching the safety limit by combining specific physical boundaries. Especially when the predicted mean approaches the safety limit, even with a small variance, the probability of exceeding the limit may be extremely high. Therefore, in the technical solution of this application, the safety margin deviation probability is further calculated based on safety constraints to obtain the safety margin deviation probability. This transforms abstract statistical uncertainty into a probability of exceeding the limit with clear physical meaning, i.e., the probability that the unit's state will exceed the allowable safe range in the future. This ensures that the control system not only focuses on achieving the control objective but also provides forward-looking warnings of potential violations, and forcibly intervenes with protection strategies when the risk probability increases, thereby preventing shutdown failures or structural damage caused by aggressive control.

[0039] More specifically, in a specific example of this application, calculating the safety margin deviation probability based on the predicted yaw error distribution based on safety constraints to obtain the safety margin deviation probability includes: calculating the point-by-point deviation probability for each predicted yaw error in the predicted yaw error distribution based on safety constraints to obtain the point-by-point deviation probability distribution; and selecting the maximum value in the point-by-point deviation probability distribution as the safety margin deviation probability.

[0040] More specifically, the process begins by retrieving a pre-defined set of wind turbine operating safety constraints from memory. This set clearly defines the permissible safe range for yaw error, such as the absolute alarm threshold for yaw misalignment angles or dynamic boundaries set to prevent yaw system overload. Subsequently, for the predicted yaw error distribution over N future time steps output by the Bayesian prediction model, a probability density function is constructed using the predicted mean and variance for each time step, and a point-by-point deviation probability calculation based on safety constraints is performed. This calculation process, for each discrete time point within the prediction time domain, calculates the area outside the safety constraint range of the yaw error probability density function at that moment through mathematical integration, thus obtaining the point-by-point deviation probability for that time point, forming a sequence of point-by-point deviation probability distributions covering the entire prediction time domain. Finally, to capture the most severe safety challenges within the entire prediction window, an extreme value search is performed on this sequence, and the maximum value in the point-by-point deviation probability distribution is selected as the safety margin deviation probability. This value represents the highest probability that a wind turbine will experience a safety overrun risk at any time during the future forecast period, and directly serves as a key quantitative basis for adjusting the conservatism of the control strategy in the subsequent risk decision-making module.

[0041] In step S423, adaptive control parameters are dynamically synthesized based on the safety margin deviation probability, risk level, and control objective to obtain adaptive control parameters. It is understood that since the control objective of wind turbines needs to seek a dynamic balance between maximizing annual power generation and minimizing structural fatigue load, fixed control parameters cannot adapt to operating conditions with drastic changes in prediction uncertainty. If aggressive parameters that prioritize power generation are maintained even when the predicted risk is extremely high, the control system will ignore potential safety boundary breach risks; conversely, excessive conservatism will lead to wasted wind energy. Therefore, in the technical solution of this application, adaptive control parameters are further dynamically synthesized based on the safety margin deviation probability, risk level, and control objective to obtain adaptive control parameters. This establishes a nonlinear mapping relationship between the control algorithm weight configuration and real-time risk measurement, transforming discrete risk levels and continuous probability indices into specific adjustment coefficients. This enables the control system to possess adaptive evolution capabilities, that is, when sensing an increase in the safety margin deviation probability or a deterioration in the risk level, it automatically increases the weight of the uncertainty penalty term or load constraint term in the cost function, thereby smoothly switching the control strategy from a performance-first mode to a safety-first mode.

[0042] More specifically, in a concrete example of this application, the parameter synthesis process first calls multiple sets of baseline parameter templates stored in the control knowledge base. These templates correspond to an aggressive strategy configuration centered on maximizing power generation and a conservative strategy configuration centered on minimizing load and maintaining the safety boundary. Subsequently, the adaptive parameter generator receives the currently calculated safety margin deviation probability and risk level, and fuses these two heterogeneous indicators into a normalized risk adjustment coefficient using a preset risk factor calculation function. The higher the value of this coefficient, the more severe the prediction uncertainty risk or safety limit risk faced by the system. Next, using this risk adjustment coefficient, a dynamic weighted interpolation operation is performed between the aggressive and conservative strategy parameter sets to synthesize a new set of adaptive control parameters in real time. For example, dynamically increasing the penalty weight for yaw action amplitude in the optimization objective function or tightening the relaxation variable of the safety constraint. This synthesized parameter set is then passed to the robust control optimizer, guiding it to generate an optimal yaw action sequence that conforms to both the current wind conditions and the risk control requirements in the following control cycle.

[0043] In step S430, based on adaptive control parameters and safety constraints, robust control action optimization is performed on the predicted yaw error distribution to obtain the optimal yaw action. It is understandable that since the final control execution requires transforming abstract strategy preferences into specific mechanical action commands, and risk level alone cannot directly drive the yaw motor to perform precise position adjustments, without rigorous mathematical optimization, it is difficult to find an accurate solution within the multi-dimensional constraint space that can both respond to wind direction changes and avoid high-risk prediction areas. Therefore, in the technical solution of this application, robust control action optimization is further performed on the predicted yaw error distribution based on adaptive control parameters and safety constraints to obtain the optimal yaw action. This constructs and solves a constrained optimization problem containing an uncertainty penalty term, integrating the distribution characteristics of Bayesian prediction and dynamically adjusted control weights into a unified mathematical model for solution. In this way, the control quantity that strictly satisfies physical boundary constraints and minimizes the objective function under the current risk tolerance can be calculated, ensuring that the final output action is a robust optimal solution considering all potential uncertainties.

[0044] More specifically, in a concrete example of this application, the optimization process employs a model predictive control framework. First, it receives the predicted mean and variance sequence of yaw error for the next N time steps generated by a Bayesian network, along with adaptive control parameters containing dynamic weights generated in previous steps. Then, a multi-objective cost function is constructed, comprising a yaw error term representing power generation loss, a yaw magnitude term representing mechanical wear, and an uncertainty penalty term representing risk aversion. The coefficients of each term are directly determined by the adaptive control parameters; for example, the coefficient of the uncertainty penalty term is significantly increased when the risk is high. Simultaneously, the physical constraints of the wind turbine, including maximum yaw speed, acceleration limits, and cable twist angle range, are set as hard constraints for the optimization problem. Next, a quadratic programming or nonlinear programming solver is invoked to perform a rolling time-domain solution to this constrained optimization problem, finding a set of optimal yaw speed sequences that minimize the cost function within the prediction time domain. Finally, the first control component in the sequence is selected as the optimal yaw action at the current moment. This action is essentially a robust control command after risk weighting and boundary truncation, which can directly guide the yaw drive mechanism to perform safe and efficient operation.

[0045] In step S440, yaw control commands are generated based on the optimal yaw action and risk metric to obtain yaw control commands. It is understood that since the optimal yaw action output by the optimization solver only represents the theoretical optimal solution at the mathematical level, the yaw rate or angle increment, usually existing in floating-point form, cannot be directly recognized and executed by the programmable logic controller or actuator at the bottom of the wind turbine. Furthermore, the simple action command lacks crucial risk context information. If the lower-level machine cannot perceive the uncertainty level behind the current command, it may fail to execute the correct fault-oriented safety logic in the event of communication packet loss or data anomalies. Therefore, in the technical solution of this application, yaw control commands are further generated based on the optimal yaw action and risk metric to obtain yaw control commands. This encapsulates the abstract control quantity into a standardized command frame conforming to the industrial communication protocol, and converts the risk metric into a confidence label that can be parsed by the downstream system and sends it along with the command. This ensures that the actuator of the wind turbine not only receives the accurate action target but also performs secondary safety verification based on the accompanying confidence information, thereby achieving a seamless and secure connection between upper-level intelligent decision-making and lower-level hardware execution.

[0046] More specifically, in a concrete example of this application, the instruction generation process first initiates an instruction formatting program, receiving the optimal yaw action value (e.g., target yaw rate or absolute yaw position) from the optimization module and a risk metric from the risk assessment module. Subsequently, based on the fieldbus communication protocol used by the wind turbine main control system, the optimal yaw action is mapped to a specific control register address and numerical encoding, while the risk metric is inversely mapped to a confidence score for the control instruction, or converted into a status bit indicating instruction priority. Next, this data is encapsulated into a complete control message data packet containing an instruction sequence number, execution timestamp, target action value, and confidence label. Finally, the yaw control instruction is sent to the controller of the yaw drive system via an industrial communication interface, enabling the underlying controller to verify whether the confidence score is higher than a preset execution threshold. Upon confirmation, the controller drives the yaw motor to perform the corresponding wind-fighting operation, completing the final stage of closed-loop control.

[0047] In summary, the deep learning-based yaw error prediction and control method for wind turbines according to the embodiments of this application is explained. It acquires real-time multi-dimensional sensor data from the wind turbine and uses an adaptive deep denoising module to separate clean physical signals from noisy data. Based on this, it calculates signal differences and performs noise pattern analysis in conjunction with the turbine's operating context to effectively identify normal operating condition fluctuations and abnormal interference, generating an attributable noise level estimate. Then, the denoised signal and noise estimate are fused and input into a Bayesian prediction model, outputting a yaw error prediction with confidence information in the form of a probability distribution. The control system then performs a risk assessment on the prediction distribution, dynamically adjusts strategy parameters based on the degree of uncertainty, and optimizes robust control actions under safety constraints. This concept achieves a closed loop from data cleaning and uncertainty attribution quantification to risk-aware control, solving the problems of low yaw control accuracy and poor robustness caused by sensor noise and uncertainty in complex environments.

[0048] Furthermore, a wind turbine yaw error prediction and control system based on deep learning is also provided.

[0049] Figure 6 This is a block diagram of a deep learning-based wind turbine yaw error prediction and control system according to an embodiment of this application. Figure 6 As shown, the deep learning-based yaw error prediction and control system 100 for wind turbines according to an embodiment of this application includes: a data acquisition module 110 for acquiring raw sensor data, including real-time wind speed, wind direction, turbine power, yaw angle, yaw speed, blade pitch angle, and main shaft speed; a denoising input acquisition module 120 for inputting the raw sensor data into an adaptive denoising module to obtain denoised input; a noise level estimation and determination module 130 for determining a noise level estimate based on the denoised input and the raw sensor data; and an uncertainty-aware yaw prediction and control decision module 140 for inputting the denoised input and the noise level estimate into a Bayesian prediction model to perform uncertainty-aware yaw prediction and control decision to obtain yaw control commands.

[0050] As described above, the deep learning-based wind turbine yaw error prediction and control system 100 according to the embodiments of this application can be implemented in various types of computing devices or control units. For example, it can be a programmable logic controller deployed in the wind turbine nacelle cabinet, an intelligent edge computing gateway assisting the PLC in complex calculations, or a high-performance industrial computer integrated into the digital twin platform of the wind farm control center. In one possible implementation, the deep learning-based wind turbine yaw error prediction and control system 100 according to the embodiments of this application can be integrated into the computing device as a software module and / or hardware module. For example, the deep learning-based wind turbine yaw error prediction and control system 100 can be an intelligent control function block in the operating environment of the computing device or PLC. This software module is configured to perform adaptive deep denoising and instantaneous difference calculation of raw sensor data, noise pattern prediction and attribution uncertainty integration based on operating context, yaw error distribution prediction based on a Bayesian deep learning model, and adaptive control parameter synthesis and robust action optimization based on safety margin deviation probability and risk level. Alternatively, it can be a dedicated wind turbine performance optimization control algorithm program developed for the computing device. Of course, the deep learning-based wind turbine yaw error prediction and control system 100 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 self-attention mechanism operation of the Transformer model and the Monte Carlo sampling inference process of the Bayesian network in parallel, or it can be a signal processing integrated circuit for a specific application.

[0051] 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 deep learning-based method for predicting and controlling the yaw error of a wind turbine, characterized in that, include: Acquire raw sensor data, including real-time wind speed, wind direction, unit power, yaw angle, yaw speed, blade pitch angle, and main shaft speed; The raw sensor data is input into the adaptive denoising module to obtain the denoised input. Based on the denoised input and raw sensor data, a noise level estimate is determined; The denoised input and noise level estimate are input into the Bayesian prediction model to perform uncertainty-aware yaw prediction and control decision-making to obtain yaw control commands.

2. The deep learning-based yaw error prediction and control method for wind turbines according to claim 1, characterized in that, The adaptive denoising module is a sequence-to-sequence Transformer model.

3. The deep learning-based yaw error prediction and control method for wind turbines according to claim 1, characterized in that, Based on the denoised input and raw sensor data, a noise level estimate is determined, including: Calculate the instantaneous difference signal between the denoised input and the raw sensor data; Based on runtime context data, noise pattern prediction is performed on instantaneous difference signals and raw sensor data to obtain predicted noise characteristics and residual uncertainties; Based on the predicted noise characteristics and residual uncertainty, the attribution uncertainty of the denoised input is integrated to obtain the noise level estimate.

4. The deep learning-based yaw error prediction and control method for wind turbines according to claim 1, characterized in that, The denoised input and noise level estimate are input into the Bayesian prediction model to perform uncertainty-aware yaw prediction and control decision-making to obtain yaw control commands, including: The denoised input and noise level estimate are input into the Bayesian prediction model to obtain the predicted yaw error distribution; Based on control objectives and safety constraints, control risk assessment and adaptive strategy parameterization are performed on the predicted yaw error distribution to obtain risk metrics and adaptive control parameters. Based on adaptive control parameters and safety constraints, robust control action optimization is performed on the predicted yaw error distribution to obtain the optimal yaw action. Yaw control commands are generated based on the optimal yaw action and risk metric.

5. The deep learning-based yaw error prediction and control method for wind turbines according to claim 4, characterized in that, The predicted yaw error distribution is a sequence of mean and variance over N future time steps.

6. The deep learning-based yaw error prediction and control method for wind turbines according to claim 4, characterized in that, Based on control objectives and safety constraints, the predicted yaw error distribution is used for control risk assessment and adaptive strategy parameterization to obtain risk metrics and adaptive control parameters, including: The predicted yaw error distribution is aggregated for prediction uncertainty and quantified for risk level to obtain aggregated uncertainty measure and risk level; Based on safety constraints, the safety margin deviation probability is calculated from the predicted yaw error distribution to obtain the safety margin deviation probability. Adaptive control parameters are obtained by dynamically synthesizing adaptive control parameters based on the safety margin deviation probability, risk level, and control objective.

7. The deep learning-based yaw error prediction and control method for wind turbines according to claim 6, characterized in that, Based on safety constraints, the safety margin deviation probability is calculated from the predicted yaw error distribution to obtain the safety margin deviation probability, including: Based on safety constraints, the point-by-point deviation probability is calculated for each predicted yaw error in the predicted yaw error distribution to obtain the point-by-point deviation probability distribution. The maximum value in the point-by-point deviation probability distribution is selected as the safety margin deviation probability.

8. A deep learning-based yaw error prediction and control system for wind turbines, characterized in that, include: The data acquisition module is used to acquire raw sensor data, which includes real-time wind speed, wind direction, unit power, yaw angle, yaw speed, blade pitch angle, and main shaft speed. The noise reduction input acquisition module is used to input the raw sensor data into the adaptive noise reduction module to obtain the noise reduction input. The noise level estimation and determination module is used to determine the noise level estimate based on the denoised input and the raw sensor data; The uncertainty-aware yaw prediction and control decision module is used to input the denoised input and noise level estimate into the Bayesian prediction model to perform uncertainty-aware yaw prediction and control decision to obtain yaw control commands.

9. The deep learning-based wind turbine yaw error prediction and control system according to claim 8, characterized in that, The noise level estimation and determination module includes: The instantaneous difference signal calculation unit is used to calculate the instantaneous difference signal between the denoised input and the original sensor data; The noise mode prediction unit is used to perform noise mode prediction on instantaneous difference signals and raw sensor data based on operating context data to obtain predicted noise characteristics and residual uncertainties. The attribution uncertainty integration unit is used to integrate the attribution uncertainty of the denoised input based on the predicted noise characteristics and residual uncertainty to obtain a noise level estimate.

10. The deep learning-based wind turbine yaw error prediction and control system according to claim 8, characterized in that, The uncertainty-perceived yaw prediction and control decision module includes: The unit for obtaining the predicted yaw error distribution is used to input the denoised input and noise level estimate into the Bayesian prediction model to obtain the predicted yaw error distribution. The control risk assessment and adaptive strategy parameterization unit is used to perform control risk assessment and adaptive strategy parameterization on the predicted yaw error distribution based on control objectives and safety constraints to obtain risk metrics and adaptive control parameters. The robust control action optimization unit is used to perform robust control action optimization on the predicted yaw error distribution based on adaptive control parameters and safety constraints to obtain the optimal yaw action; The yaw control command generation unit is used to generate yaw control commands based on the optimal yaw action and risk metric to obtain yaw control commands.

Images