A yaw instruction library-based wind farm wake optimization control method

Abstract

The application provides a yaw instruction library-based wind farm wake optimization control method, and relates to the technical field of wind farms. The method extracts typical conventional wind conditions and extreme wind conditions through cluster analysis, constructs a wind condition scenario library, and ensures the wide coverage and adaptability of the control strategy to diversified meteorological conditions. More importantly, in combination with the environmental contour method, a limit state surface model is constructed to determine the load safety boundary of the key components of the wind farm, and the optimization solution is performed in the offline stage. The yaw instruction library corresponding to each wind condition scenario cluster is constructed, and an online verification and protection mechanism is set before the execution of the instructions, thereby realizing the collaborative optimization of power gain and unit fatigue life, ensuring the safety and reliability of the control process, and maintaining stable operation even under data loss or extreme working conditions. The method has good engineering practicability and promotional value.

Description

Technical Field

[0001] This invention relates to the field of wind farm technology, and in particular to a wind farm wake optimization control method based on a yaw command library. Background Technology

[0002] With the development of wind power as a clean energy source, the scale of wind farm construction has expanded rapidly. However, traditional yaw control strategies that pursue the maximum power output of a single wind turbine fail to fully consider the mutual influence between multiple wind turbines, resulting in low overall power generation efficiency and increased operation and maintenance costs. Existing technologies include active yaw wake control strategies based on computational fluid dynamics simulation, genetic algorithms, and reinforcement learning. While these methods have improved power generation efficiency to some extent, they face problems in practical applications due to their reliance on real-time data and complex calculations, such as insufficient robustness, high computational complexity, poor adaptability, and difficulty in balancing power increase and fatigue life. Specifically, the high dependence of existing technologies on real-time data and low-latency communication makes the system perform poorly in extreme weather or data transmission delays. At the same time, the high computational cost and complex algorithm structure limit their large-scale application potential. Offline control technologies, on the other hand, cannot be updated in real time to adapt to changing weather conditions, resulting in rapid strategy aging and insufficient coverage. Summary of the Invention

[0003] The purpose of this invention is to provide a wind farm wake optimization control method based on a yaw command library, so as to solve the technical problems existing in the prior art and at least provide a beneficial option or create conditions.

[0004] The solution to the technical problem of this invention is as follows: This invention provides a wind farm wake optimization control method based on a yaw command library, comprising the following steps:

[0005] Historical operational data of wind farms are acquired, and data preprocessing and feature extraction are performed to obtain a wind farm feature dataset.

[0006] Cluster analysis was performed on the wind farm feature dataset to divide it into several clusters of normal wind conditions and extreme wind conditions, and a wind condition scenario library was constructed.

[0007] Based on the aforementioned extreme wind condition scenario cluster, the environmental contour line method is introduced to construct a limit state surface model and determine the load safety boundary of key components of the wind farm.

[0008] Based on the load safety boundary, wake model and power optimization model, the optimal yaw strategy and its corresponding power gain, fatigue cost and applicable operating condition range under each wind scenario cluster in the wind scenario library are solved offline, and a yaw command library is constructed.

[0009] Real-time acquisition of current wind conditions, generation of feature vectors and matching with the wind condition scenario library; when a match is successful, the corresponding yaw instruction in the yaw instruction library is invoked; when no match is found, a transition instruction is generated by interpolation based on the yaw instructions corresponding to the nearest clusters of wind condition scenarios in the wind condition scenario library.

[0010] The yaw command or transition command is subjected to load safety verification. If the verification is successful, it is issued for execution; otherwise, the protection mechanism is activated to reduce the control intensity or return to the normal wind-fighting operation mode.

[0011] Furthermore, the historical operating data includes incoming flow data, wind turbine power, yaw angle, and rotational speed; the incoming flow data is used to characterize wind conditions, including wind direction, wind speed, air density, and turbulence intensity; the data preprocessing includes missing value imputation, outlier removal, and noise filtering; the feature extraction refers to calculating representative statistical features of the historical operating data, including the incoming flow vector, based on the time series using a sliding time window method.

[0012] Furthermore, the clustering analysis of the wind farm feature dataset to divide it into several clusters of normal wind condition scenarios and extreme wind condition scenario clusters, and the construction of a wind condition scenario library, includes the following steps:

[0013] Principal component analysis was performed on the wind farm feature dataset to reduce its dimensionality and extract the main variation features, resulting in a wind condition distribution map in a low-dimensional feature space.

[0014] Based on the wind condition distribution map, the number and initial location of cluster centers are determined. Cluster analysis is performed using a Gaussian mixture model or a density-based clustering method to generate several wind condition scene clusters, including regular wind condition scene clusters and extreme wind condition scene clusters. Each wind condition scene cluster includes a cluster center mean vector, a covariance matrix, and weight coefficients, which are used to characterize the statistical characteristics of various wind conditions.

[0015] Based on the distribution characteristics of the aforementioned wind condition scene clusters, and combined with the frequency and coverage of historical wind conditions, a wind condition scene library for various wind conditions is constructed.

[0016] Furthermore, based on the extreme wind condition scenario cluster, the environmental contour line method is introduced to construct a limit state surface model and determine the load safety boundary of key components of the wind farm, including the following steps:

[0017] Based on the historical operating data, environmental contour lines under the target recurrence period are constructed in physical space, the quantile responses of each key component of the wind farm at the representative inflow vector are calculated, and the limit state surface model is constructed.

[0018] Regression analysis is performed on the limit state surface model and its narrow-band neighborhood to obtain the component-level load upper bound function as the load safety boundary.

[0019] Furthermore, the limit state surface model satisfies the following calculation formula:

[0020] ;

[0021] in, The limit state surface model represents the set of all incoming flow conditions under which the responses of each key component reach their quantile thresholds within the target recurrence period. This represents the representative inflow vector in the historical operational data. The first wind farm A key component represents the incoming flow vector. The quantile response at the point, For the first Key components during the target recurrence period The corresponding upper quantile value;

[0022] The load safety boundary satisfies the following calculation formula:

[0023] ;

[0024] in, Indicates the first A key component represents the incoming flow vector. The load safety boundary under the condition is the estimated maximum load that the component can withstand under extreme wind conditions; Representing the limit state surface model The narrowband neighborhood, representing the distance No more than small positive numbers All incoming flow vectors A set of values ​​used for robust regression and upper bound estimation within a local range; Denotes supremum, indicating the supremum of the set. All The minimum upper bound of the value, i.e. the maximum possible response value in this neighborhood, is used to conservatively estimate the ultimate load of the component.

[0025] Furthermore, based on the load safety boundary, wake model, and power optimization model, the optimal yaw strategy and its corresponding power gain, fatigue cost, and applicable operating condition range for each wind scenario cluster in the wind scenario library are solved offline to construct a yaw command library, including the following steps:

[0026] For each wind scenario in the wind scenario library, a wake model is used to describe the wake effect of the upstream wind turbine on the downstream wind turbine, and the wind speed distribution of the entire wind farm is calculated by superimposing the wake effects of each unit.

[0027] Based on the wind speed distribution, a power optimization model with fatigue penalty and load safety boundary constraints is established;

[0028] Based on the power optimization model, the optimal yaw angle distribution under each wind condition scenario cluster and its corresponding power gain, fatigue cost and applicable operating condition range are obtained offline, and a yaw command library is constructed.

[0029] Furthermore, the power optimization model satisfies the following calculation formula:

[0030] ;

[0031] in, This represents the objective function of the power optimization model, used to solve for the wind farm under a given inflow vector. and yaw angle The optimal yaw strategy is designed to maximize overall power gain while minimizing fatigue costs and yaw action costs. Indicates the first The yaw angle of a typhoon turbine needs to be determined during the optimization process to maximize the objective function by determining the optimal yaw angle for each turbine. Indicates based on and Normalized total power output of wind farms; This indicates the adjustment parameter, with a range of [value missing]. Used to balance the aggressiveness of strategies. A higher value indicates a greater tendency to pursue high power output, but also comes with higher fatigue costs and yaw costs. This represents the weighting coefficient used to adjust the contribution of the fatigue cost penalty term; Indicates based on and The normalized fatigue cost assessment value; This represents the weighting coefficient used to adjust the contribution of the yaw action cost penalty item; Indicates based on Normalized yaw maneuver cost metric.

[0032] Furthermore, the load safety boundary is incorporated into the feasible region of the power optimization model using a hard constraint method, satisfying the following calculation formula:

[0033] ;

[0034] in, For the first A key component represents the incoming flow vector. Given yaw angle Given the incoming flow vector The upper bound of the ultimate load under the condition, Indicates the first A key component represents the incoming flow vector. Load safety boundary under the given conditions;

[0035] The power optimization model must also satisfy the following constraints:

[0036] No. Yaw angle of typhoon turbine ,express It needs to be within the physically adjustable range of the yaw angle. , The design and structural limitations of the yaw drive system are designed to prevent mechanical overload or structural interference.

[0037] No. Yaw angle change of a typhoon turbine in a single control cycle , The maximum allowable single-step yaw angle change is used to limit the abrupt change in yaw action and avoid impacting the transmission system;

[0038] Within a unit of time, the first Yaw rate of typhoon turbine absolute value , This indicates the maximum permissible yaw rate, which is determined by the dynamic performance of the yaw motor and drive system, to prevent excessively fast rotation from causing structural vibration or motor overload.

[0039] No. Typhoon turbine in time Cable deflection angle inside , This indicates the maximum permissible cable deflection angle. Exceeding this value will trigger an automatic cable unwinding procedure to prevent damage caused by excessive cable twisting.

[0040] Furthermore, the transition command is generated by weighted interpolation using a kernel function based on Mahalanobis distance, satisfying the following calculation formula:

[0041] ;

[0042] in, Represents the current incoming flow vector Transitional yaw commands are used to achieve smooth control in scenarios without precise matching; Represents the current incoming flow vector The first wind condition scenario in the aforementioned wind condition scenario library The squared Mahalanobis distance between the centers of each wind condition scene cluster reflects the degree of similarity in their statistical characteristics. This indicates that the wind condition scenario library contains the vector of the current incoming flow. Number of the nearest wind condition scene clusters; The kernel function bandwidth parameter controls the decay rate of the influence of neighboring clusters during interpolation. The larger the value, the smoother the interpolation, but the more delayed the response. The smaller the value, the more sensitive the response, but this may introduce jitter. Indicates the first The optimal yaw angle set corresponding to each wind condition scenario cluster is stored in the yaw command library.

[0043] Furthermore, based on the newly added operational data and the evolution of wind conditions, the yaw command library and wind condition scenario library are dynamically updated. The dynamic update cycle is weekly, monthly, or quarterly. The update frequency is adaptively adjusted according to the climate characteristics and wind condition change rate of the wind farm's location to ensure that the yaw command library and wind condition scenario library continuously reflect the actual wind condition distribution.

[0044] The beneficial effects of this invention are as follows: This application provides a wind farm wake optimization control method based on a yaw command library. This method extracts typical conventional and extreme wind conditions through cluster analysis, constructing a wind condition scenario library to ensure the control strategy's broad coverage and adaptability to diverse meteorological conditions. More importantly, by combining environmental contour line methods, a limit state surface model is constructed to determine the load safety boundaries of key wind farm components. Optimization is performed offline to construct a yaw command library corresponding to each wind condition scenario cluster. An online verification and protection mechanism is set before command execution, achieving coordinated optimization of power gain and unit fatigue life, ensuring the safety and reliability of the control process. Especially under data loss or extreme operating conditions, it can still maintain stable operation, demonstrating good engineering practicality and promotional value.

[0045] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the description, claims and drawings. Attached Figure Description

[0046] The accompanying drawings are provided to further understand the technical solutions of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the technical solutions of the present invention, and do not constitute a limitation on the technical solutions of the present invention.

[0047] Figure 1 This is a flowchart of the wind farm wake optimization control method based on the yaw command library provided in this application;

[0048] Figure 2 This is a schematic diagram of the cluster analysis results based on wind speed, air density, and turbulence intensity provided in this application;

[0049] Figure 3This is a schematic diagram of the clustering analysis results based on wind speed and turbulence intensity provided in this application;

[0050] Figure 4 This is a schematic diagram of the clustering analysis results based on wind speed and wind direction provided in this application;

[0051] Figure 5 This is a schematic diagram of the limit state surface based on the environmental contour line method provided in this application;

[0052] Figure 6 This is a schematic diagram of the yaw command library for each wind turbine under different clusters provided in this application;

[0053] Figure 7 This is a schematic diagram of wind farm power gain under different clustering conditions based on the yaw command library provided in this application;

[0054] Figure 8 This is a schematic diagram showing the gain comparison of annual power generation of wind farms under different wind energy utilization hours provided in this application;

[0055] Figure 9 This is a schematic diagram of the fatigue damage monitoring results of the key components provided in this application. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0057] The present application will be further described below with reference to the accompanying drawings and specific embodiments. The described embodiments should not be considered as limitations on the present application, and all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of the present application.

[0058] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0059] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0060] As the global energy structure shifts towards cleaner and lower-carbon energy, wind power, as a crucial component of renewable energy, has experienced rapid development over the past two decades. The scale of onshore and offshore wind farms has continuously expanded, evolving from small-scale, distributed deployments to large-scale, intensive, and intelligent wind power bases. Against this backdrop, how to further improve the overall operating efficiency of wind farms and reduce the levelized cost of electricity (LCOE) has become a core concern for the industry.

[0061] In actual operation, significant aerodynamic interference effects, known as the "wake effect," exist between the individual wind turbines within a wind farm. When an upstream turbine captures wind energy and rotates, a wake region with reduced wind speed and enhanced turbulence forms downstream, causing a decrease in the effective wind speed received by the downstream turbine and significantly impacting its power output. Studies have shown that in densely packed wind farms, the wake effect can cause a loss of 10% to 20% in total power generation. Therefore, traditional control strategies that target only the maximum power of a single turbine are no longer sufficient to meet the demands of modern wind farms for maximizing overall energy efficiency.

[0062] To address this challenge, academia and industry have proposed various active wake control technologies. Among them, wake optimization control based on yaw angle adjustment has become a current research hotspot due to its advantages such as low implementation cost, no hardware modification required, and fast response speed. The basic principle of this technology is: by intentionally allowing the upstream wind turbine to slightly deviate from its facing direction (i.e., active yaw), its wake is shifted laterally, thereby reducing the impact on the downstream wind turbine. Although the upstream wind turbine's power output will decrease slightly due to yaw, if properly controlled, the total power output of the entire wind farm will achieve a net gain.

[0063] However, achieving efficient yaw and wake control faces numerous technical challenges, including strong multivariate coupling, severe system nonlinearity, complex and variable environmental conditions, and high risks associated with increased generator fatigue loads. Existing control methods largely rely on real-time data acquisition, high-precision modeling, and online optimization calculations, which have revealed a series of limitations in practical engineering applications. Therefore, a novel control architecture that balances computational efficiency, robustness, safety, and economy is urgently needed.

[0064] In existing technologies, wind farm yaw control has evolved from the traditional single-unit wind-following mode to various active wake optimization methods. Currently, practical wind farms generally adopt independent yaw control based on local wind vane feedback, which only pursues the maximum power output of a single unit and does not consider the wake effect. To improve the overall efficiency, researchers have proposed a number of improvement schemes: for example, patent CN120007508A proposes a control method based on yaw path constraints and the optimal sequence for the entire field, which solves for the optimal yaw sequence when the wind direction changes to reduce fatigue; patent CN120083648A uses a combination of offline learning and online iterative correction to generate yaw commands; patent CN119755011A introduces a hierarchical optimization strategy of "coarse search → fine search" to improve solution efficiency; patent CN118327883A uses wind direction turbulent kinetic energy to set a dynamic trigger interval to reduce invalid yaw actions; and patent CN120300942A partitions the wind farm through spectral clustering and jointly optimizes power and thrust fluctuations within sub-regions. In addition, there are studies based on intelligent optimization methods such as genetic algorithms and reinforcement learning that attempt to achieve globally optimal control.

[0065] However, the aforementioned existing technologies have several shortcomings: First, most advanced control methods heavily rely on real-time wind data and full-field communication, requiring high computing power and low-latency networks, and exhibiting poor robustness in the event of data delays or disconnections. Second, optimization algorithms (such as genetic algorithms and hierarchical optimization) have high computational complexity in large-scale wind farms, making it difficult to meet the requirements of second-level response and resulting in high engineering implementation costs. Third, existing offline control strategies are mostly based on historical averages and lack dynamic update mechanisms, making them unable to adapt to seasonal wind changes or sudden weather events, leading to strategy aging and insufficient coverage of operating conditions. In addition, most solutions focus on maximizing total power, neglecting the problem of increased fatigue load caused by yaw operations, which may accelerate the wear and tear of key components (such as bearings and gearboxes) and affect the lifespan of the unit. Finally, existing methods generally lack modeling and safety boundary control for extreme wind conditions, and have not established effective load warning and protection mechanisms, which can easily lead to overload risks under high-risk operating conditions, and overall fail to achieve synergistic optimization of power generation revenue and equipment reliability.

[0066] To address the aforementioned issues, this application provides a wind farm wake optimization control method based on a yaw command library. Its core lies in constructing a control architecture of "offline pre-simulation, online matching, dynamic updating, and a safe closed loop." This method first performs cluster analysis on historical operating data to extract typical conventional and extreme wind condition scenarios, constructing a comprehensive wind condition scenario library. Then, combining environmental contour lines, it constructs a limit state surface model to determine the load safety boundaries of key wind farm components. Based on this, it combines the wake model with a power optimization model with load constraints to solve the optimal yaw strategy for each scenario offline, forming a yaw command library that includes power gains, fatigue costs, and applicability. During runtime, it only needs to quickly match real-time wind conditions with the scenario library, directly calling or interpolating to generate control commands. A load safety verification mechanism ensures execution safety; otherwise, it automatically degrades or rolls back. This solution completely eliminates the reliance on complex real-time calculations, significantly improving response speed and system robustness. Simultaneously, through a periodic dynamic update mechanism, it maintains long-term adaptability to wind condition evolution, achieving a comprehensive optimization of power efficiency, equipment lifespan, and operational safety.

[0067] First, the wind farm wake optimization control method based on the yaw command library provided in this application will be described in detail below with reference to the accompanying drawings.

[0068] Reference Figure 1 The implementation process of the wind farm wake optimization control method based on yaw command library provided in this application embodiment includes, but is not limited to, the following steps.

[0069] Step S110: Obtain historical operating data of the wind farm, perform data preprocessing and feature extraction to obtain a wind farm feature dataset.

[0070] In step S110, historical data accumulated during the long-term operation of the wind farm is collected, including incoming wind speed, wind direction, air density, turbulence intensity, and key parameters such as wind turbine power, yaw angle, and rotational speed. This comprehensively reflects the dynamic behavior of the wind farm under different meteorological conditions and operating states. Since the raw data often contains missing values ​​due to sensor malfunctions, outliers caused by measurement errors, and high-frequency noise interference, systematic data preprocessing is necessary to ensure data integrity and accuracy. Based on this, feature extraction is further performed on the time series data using a sliding time window approach. Representative statistical characteristics such as the mean, variance, range, and trend changes of each variable within the time window are calculated. This transforms the original high-frequency time series data into physically meaningful low-dimensional feature vectors, forming a well-structured and information-condensed wind farm feature dataset to support subsequent wind condition clustering and scenario modeling.

[0071] Step S120: Perform cluster analysis on the wind farm feature dataset to divide it into several clusters of normal wind conditions and extreme wind conditions, and construct a wind condition scenario library.

[0072] In step S120, the complex and continuous actual wind condition space is divided into a finite number of representative typical operating condition categories, realizing the discretization and pattern representation of wind conditions. By performing cluster analysis on the feature dataset obtained in step S110, data groups with similar statistical characteristics in key parameters such as wind speed, wind direction, and turbulence intensity are identified. Each group constitutes a normal wind condition scenario cluster or an extreme wind condition scenario cluster. Each scenario cluster is characterized by its cluster center (mean vector), dispersion (covariance matrix), and frequency of occurrence (weighting coefficient), effectively describing the central trend, fluctuation range, and probability of occurrence of this type of wind condition. Finally, all identified normal and extreme wind condition scenario clusters are integrated into a structured wind condition scenario library, serving as a knowledge base for typical operating conditions of wind farms, providing clear input conditions for subsequent offline optimization of yaw strategies.

[0073] Step S130: Based on the extreme wind condition scenario cluster, the environmental contour line method is introduced to construct the limit state surface model and determine the load safety boundary of key components of the wind farm.

[0074] In step S130, by introducing the environmental contour line method, a multivariate joint probability distribution model is constructed based on long-term historical data of extreme wind condition scenario clusters. This determines the extreme wind condition boundary corresponding to a specific return period (e.g., a 50-year return period) in physical space, forming environmental contour lines. Based on this, combined with the aerodynamic-structural response model of the wind turbine, the ultimate load response of each key component (e.g., main shaft, tower, blade root) under extreme inflow conditions is calculated, and a limit state surface model is constructed, representing the set of all inflow conditions that cause the component response to reach the design quantile threshold. Furthermore, regression analysis is performed in the neighborhood of this surface to estimate the maximum load-bearing capacity of each component under extreme conditions, forming a "load safety boundary" function to ensure the applicability and safety of the control strategy in high-risk scenarios.

[0075] Step S140: Based on the load safety boundary, wake model and power optimization model, solve offline the optimal yaw strategy and its corresponding power gain, fatigue cost and applicable operating condition range for each wind scenario cluster in the wind scenario library, and construct the yaw command library.

[0076] In step S140, offline simulation and optimization are used to pre-calculate the optimal yaw control scheme for each wind condition in the wind condition scenario library. For each identified wind condition scenario cluster (including normal and extreme), an improved Jensen wake model or other engineering wake model is first used to simulate the wake impact of upstream wind turbines yawing on downstream wind turbines, and the overall wind speed distribution is calculated. Based on this, a comprehensive optimization objective function is established, with maximizing the total power of the wind farm as the core objective. Fatigue costs and yaw action costs are introduced as penalty terms, and the load safety boundary of key components is used as a hard constraint to ensure that the optimization result will not lead to unit overload. By solving this constrained optimization problem, the optimal yaw angle distribution of each wind turbine under the wind condition scenario is obtained, and the corresponding overall power gain, unit fatigue damage level, and applicable wind condition range of the strategy are recorded. Finally, the optimal yaw strategies and their performance indicators for all scenarios are organized into a structured yaw command library to realize the mapping relationship between "wind condition - control command - performance evaluation", providing a readily available control basis for online operation.

[0077] Step S150: Real-time acquisition of current wind conditions, generation of feature vectors and matching with the wind condition scenario library; if a match is successful, call the corresponding yaw instruction in the yaw instruction library; if no match is found, interpolate the yaw instructions corresponding to the nearest clusters of wind conditions in the wind condition scenario library to generate a transition instruction.

[0078] In step S150, the control strategy is transitioned from offline to online to ensure the system can quickly respond to changing wind conditions during actual operation. During real-time operation of the wind farm, incoming flow data such as wind speed, wind direction, and turbulence intensity are continuously collected, and a feature vector of the current wind condition is generated using the same feature extraction method as in step S110 (e.g., sliding time window statistics). Subsequently, this feature vector is matched for similarity with various scene clusters in the wind condition scene library, typically using statistical distance metrics such as Mahalanobis distance to determine whether the current wind condition falls within the distribution range of a particular scene cluster. If a match is successful, the optimal yaw command set corresponding to that scene is directly retrieved from the yaw command library and prepared for issuance. If the current wind condition is not precisely covered by any scene cluster (e.g., located at the boundary between two scenes), a hard switch is not used. Instead, several nearest-neighbor scenes are selected, and a smooth transitional yaw command is generated based on their statistical weights using kernel function weighted interpolation. This avoids abrupt changes in control actions, achieving continuous and smooth strategy switching and ensuring the stability and comfort of the control process.

[0079] Step S160: Perform load safety verification on the yaw command or transition command. If the verification passes, the command is issued for execution; otherwise, activate the protection mechanism to reduce the control intensity or revert to the normal wind-fighting operation mode.

[0080] Step S160 serves as the final safety barrier in the control process, ensuring that all yaw commands to be executed are within the safe operating range of the equipment. Before issuing the command, a pre-established load proxy model or rapid evaluation method is used to predict the ultimate load response of each key component under the current wind conditions and the proposed yaw command, and compare it with the load safety boundary determined in step S130. If the predicted load of all components does not exceed their safety threshold, the verification is deemed successful, and the command can be normally issued to the yaw control system of each wind turbine for execution. If the predicted load value of any component exceeds the safety boundary, the protection mechanism is immediately activated: measures such as reducing control aggression (e.g., reducing the yaw angle), limiting the yaw rate, or directly abandoning the optimization command and reverting to the traditional head-on wind operation mode are taken to prioritize the safety of the unit structure. This mechanism realizes a dynamic safety closed loop for the control strategy, effectively preventing overload risks caused by model errors, sudden changes in wind conditions, or extreme events, and ensuring that optimized control is carried out under safe conditions.

[0081] In some embodiments of this application, load safety verification is performed by a proxy model before yaw or transition commands are issued, and the upper limit of the load corresponding to the limit state surface model constructed from environmental contour lines is used as a unified gating boundary; if there is a risk exceeding the limit, protection or backoff strategies are automatically triggered. The final yaw or transition commands are issued to the wind turbine yaw system through a programmable logic controller (PLC) to achieve fast, low-cost wake optimization control without real-time complex calculations; when the gating is not satisfied, the system automatically reduces the strategy aggression or backoffs to the windward strategy, while using hysteresis and minimum hold time to suppress frequent switching.

[0082] In some embodiments of this application, historical operating data includes incoming flow data, wind turbine power, yaw angle, and rotational speed; incoming flow data is used to characterize wind conditions, including wind direction, wind speed, air density, and turbulence intensity; data preprocessing includes missing value imputation, outlier removal, and noise filtering; feature extraction refers to calculating representative statistical features of historical operating data, including incoming flow vectors, based on the time series using a sliding time window method.

[0083] Specifically, historical operational data encompasses multi-dimensional information accumulated during the long-term operation of wind farms, including inflow data, actual wind turbine output power, yaw angle, and rotational speed, among other key parameters. This data collectively constitutes a foundational information set reflecting the external meteorological conditions and internal turbine response behavior of the wind farm. Inflow data, as the core variable describing environmental wind conditions, includes wind direction, wind speed, air density, and turbulence intensity. Wind direction and speed directly determine the magnitude and direction of wind energy input, serving as the basic basis for analyzing wake effects, evaluating turbine wind performance, and constructing wind condition scenario models. Air density affects the aerodynamic thrust and power output of the turbine, and is particularly important in high-altitude areas or regions with significant seasonal temperature and humidity variations. Turbulence intensity characterizes the severity of wind speed fluctuations, directly affecting the turbine load spectrum and fatigue accumulation. Meanwhile, turbine power, yaw angle, and rotational speed reflect the actual operating status and control response of the turbine under specific wind conditions, providing crucial support for verifying model accuracy and retrieving control logic. By comprehensively collecting these parameters, a complete and accurate historical database can be established, providing a reliable data foundation for subsequent data processing and modeling analysis.

[0084] Data preprocessing is a crucial step in cleaning and improving the quality of raw historical data, aiming to eliminate data quality issues caused by factors such as sensor failure, communication packet loss, and electromagnetic interference. This process mainly includes three steps: First, missing value imputation is performed. For missing data caused by equipment power outages or signal interruptions, linear interpolation, spline interpolation, or machine learning-based methods (such as KNN and random forests) are used to reasonably fill in the gaps, ensuring the continuity of the time series. Second, outlier removal is performed. Statistical methods (such as the 3σ principle and box plots) or physical boundary judgments (such as high power readings even when wind speed exceeds the cutoff wind speed) are used to identify and remove outliers that significantly deviate from the normal range, preventing them from misleading subsequent modeling. Finally, noise filtering is performed. Techniques such as low-pass filters, moving averages, or wavelet transforms are used to suppress high-frequency measurement noise and retain trends reflecting the true physical processes. After systematic preprocessing, the completeness, consistency, and reliability of the raw data are significantly improved, providing high-quality input for feature extraction and cluster analysis.

[0085] Feature extraction is the process of transforming time-series data into low-dimensional feature vectors with physical meaning and statistical representativeness. Since the original data is typically sampled at high frequencies (seconds or minutes), direct use would lead to computational redundancy and difficulty in capturing the essential characteristics of the operating conditions. Therefore, a sliding time window method is used to perform time-domain statistical analysis on each variable. Specifically, the continuous time series is divided into multiple overlapping or non-overlapping time windows (e.g., 10 minutes per window). Within each window, representative statistical characteristics of various operating parameters are calculated, such as mean (reflecting the central trend), standard deviation (reflecting volatility), maximum / minimum values ​​(reflecting extreme levels), and rate of change (reflecting dynamic characteristics). This process achieves the transformation from "instantaneous observation" to "operating condition characterization," ensuring that each time window corresponds to a structured feature vector that can more stably and representatively describe the overall operating state within a certain time period. Particular attention must be paid to the inflow vector, which consists of wind speed and wind direction, requiring careful consideration of the periodicity of wind direction (e.g., 0° and 360° are equivalent) and the vector synthesis method (e.g., using sine / cosine decomposition followed by averaging and then inverse calculation) to avoid statistical bias. The resulting feature dataset not only compressed the data size but also highlighted the differences between various wind conditions, providing an ideal input format for subsequent cluster analysis.

[0086] In some embodiments of this application, the average wind speed within a certain time window satisfies the following calculation formula (1):

[0087] (1);

[0088] In formula (1), This represents the average wind speed within that time window. For the first Wind speed values ​​at each sampling point This represents the total number of sampling points within this time window.

[0089] In some embodiments of this application, the average wind direction can be calculated by trigonometric function averaging, satisfying the following calculation formula (2):

[0090] (2);

[0091] In formula (2), Average wind direction For the first Wind direction at each sampling point.

[0092] In some embodiments of this application, turbulence intensity The wind speed is characterized by the ratio of the standard deviation to the average wind speed, and the following calculation formula (3) applies:

[0093] (3).

[0094] In some embodiments of this application, the wind shear index is obtained by fitting the wind speed relationship at different heights, satisfying the following calculation formula (4):

[0095] (4);

[0096] In formula (4), The wind shear index, For height Wind speed at the location, For reference height Wind speed at that location.

[0097] In some embodiments of this application, step S120 involves performing cluster analysis on the wind farm feature dataset to divide it into several clusters of normal wind conditions and clusters of extreme wind conditions, thereby constructing a wind condition scenario library, which includes the following steps.

[0098] Step S210: Perform principal component analysis to reduce the dimensionality of the wind farm feature dataset in order to extract the main variation features and obtain the wind condition distribution map in the low-dimensional feature space.

[0099] In step S210, the wind farm feature dataset typically contains multiple highly correlated variables, such as wind speed, wind direction, turbulence intensity, air density, and temperature. These variables may exhibit significant linear or nonlinear correlations. Direct clustering in the original high-dimensional space is susceptible to the "curse of dimensionality," leading to unstable or distorted clustering results. Therefore, Principal Component Analysis (PCA) is introduced. By decomposing the eigenvalues ​​of the covariance matrix of the feature data, a new set of orthogonal coordinate axes (i.e., principal components) is found, such that the data has the largest variance in the first principal component direction, followed by the second principal component direction, and so on. By retaining the top few principal components with high contribution rates, the original high-dimensional feature vectors can be projected into a low-dimensional feature space, achieving data dimensionality reduction. This process not only significantly reduces computation but, more importantly, removes multicollinearity interference between variables, highlighting the most dominant driving modes of wind condition changes (e.g., wind speed-dominated changes, abrupt wind direction changes, and highly turbulent fluctuation changes). The "Wind Condition Distribution Map in Low-Dimensional Feature Space" generated on this basis intuitively shows the clustering trend of different wind condition samples in the main direction of change, providing a visual basis and mathematical foundation for determining the number of clusters and initializing cluster centers, ensuring that the clustering process can accurately capture the real structure of the wind condition space.

[0100] Step S220: Determine the number and initial location of cluster centers based on the wind distribution map, and perform cluster analysis using a Gaussian mixture model or a density-based clustering method to generate several wind condition scene clusters, including regular wind condition scene clusters and extreme wind condition scene clusters.

[0101] Among them, the wind condition scenario cluster includes the cluster center mean vector, covariance matrix and weight coefficients, which are used to characterize the statistical characteristics of various wind conditions.

[0102] Step S220, after data dimensionality reduction, is a crucial operation for substantially dividing the wind condition space. Its purpose is to identify statistically significant wind condition pattern groups from the low-dimensional feature space. Determining the number and initial locations of cluster centers is a critical prerequisite for successful clustering. The optimal number of clusters K can usually be determined by observing principal component contribution curves ("elbow rule"), silhouette coefficient analysis, or information criteria (such as AIC / BIC). The initial locations can be set based on density peaks in the distribution map or using intelligent initialization strategies such as K-means++ to avoid getting trapped in local optima.

[0103] Regarding the specific clustering algorithm selection, Gaussian Mixture Models (GMMs) assume that each wind condition cluster follows a multivariate normal distribution, outputting the cluster center mean vector, covariance matrix, and weight coefficients. This provides good probabilistic interpretation and is suitable for describing the statistical fluctuation characteristics of wind conditions. Density-based clustering methods (such as DBSCAN), on the other hand, can automatically identify cluster structures of arbitrary shapes and effectively eliminate noise points, making them suitable for scenarios with irregular wind distributions or outliers. By performing cluster analysis on the dimensionality-reduced wind condition samples using one of these two methods, multiple wind condition clusters are generated. Each cluster represents a typical operating state that is highly similar in combinations of parameters such as wind speed, wind direction, and turbulence. These clusters are not only the result of wind condition classification but also the basic unit for subsequent yaw strategy optimization.

[0104] Step S230: Based on the distribution characteristics of wind condition scene clusters and combined with the frequency and coverage of historical wind conditions, construct a wind condition scene library for various wind conditions.

[0105] In step S230, the scattered clustering outputs are integrated into a structurally complete and semantically clear knowledge base—the wind condition scenario database. Each wind condition scenario cluster is stored in the database in a structured form, containing three core elements: the cluster center mean vector, representing the typical state of this type of wind condition (e.g., "wind speed 8 m / s, wind direction 270°, turbulence intensity 0.12"); the covariance matrix, reflecting the dispersion of this type of wind condition across various feature dimensions and the correlation between variables, used to quantify the range of wind condition fluctuations; and the weight coefficient, usually determined by the proportion of the number of samples in this cluster to the total number of samples, characterizing the frequency or probability of this type of wind condition occurring in historical operations. By combining these three parameters, the wind condition scenario database not only records "what are the typical wind conditions," but also describes "what they look like," "how much they change," and "how often they occur." In addition, the construction process also needs to evaluate the coverage of each scenario cluster to ensure that it can cover the vast majority of historical wind condition samples (e.g., a cumulative probability of over 95%), avoiding control blind spots. The resulting wind scenario library serves as a static but updatable knowledge model, providing standardized input conditions for subsequent offline optimization. This allows for the independent solution of the optimal yaw strategy for each typical wind condition, thereby establishing a mapping relationship between "wind condition and control" and laying a solid foundation for achieving efficient and robust wind farm wake optimization control.

[0106] In some embodiments of this application, to reduce redundant information and highlight key features, principal component analysis is used to reduce the dimensionality of the wind condition feature vector. Let the wind condition features calculated within a time window be vectors. (i.e., the incoming flow vector), its sample mean is PCA maps it to a low-dimensional feature space. It satisfies the following formula (5):

[0107] (5);

[0108] In formula (5), This represents the projection matrix composed of the principal component directions.

[0109] In some embodiments of this application, step S130 involves introducing an environmental contour line method based on extreme wind condition scenario clusters to construct a limit state surface model and determine the load safety boundary of key components of the wind farm, including the following steps.

[0110] Step S310: Based on historical operating data, construct environmental contour lines in the physical space under the target recurrence period, calculate the quantile responses of each key component of the wind farm at the representative inflow vector, and construct the limit state surface model.

[0111] In step S310, a physical model reflecting the joint probability characteristics of extreme wind conditions is established based on long-term observation data, and the critical state of the wind farm structural response is defined based on this model. The environmental contour line method is an advanced extreme value statistical technique. Its core idea is to map the joint probability distribution of high-dimensional meteorological variables onto physical space, forming one or more closed curves (i.e., contour lines). All points on these curves have the same exceedance probability, representing the most severe but acceptable combination of environmental conditions that may occur within a specific return period (e.g., 1 year, 10 years, 50 years). The environmental contour lines constructed using this method not only consider extreme wind speeds but also integrate the synergistic effects of wind direction, turbulence intensity, and wind shear, avoiding the deficiency of traditional univariate extreme value analysis in ignoring the correlation between variables. Based on this, several "representative inflow vectors" on the contour lines are selected as input conditions. Combined with the wind turbine aerodynamic-structural coupling simulation model or field load measurement data, the ultimate load response of key components of the wind farm (such as the main shaft, gearbox, tower bottom, and blade root) under these extreme conditions is calculated, and its upper quantile value (such as the 98th percentile) under the target return period is determined. The set of all inflow conditions that cause the component response to reach this quantile threshold is defined as the "limit state surface model." This surface delineates the boundary from "safe" to "critical" in physical space, becoming the benchmark for subsequent load boundary modeling and safety verification. This process realizes the transformation from statistical probability to physical response, providing a scientific and quantifiable description for the ultimate bearing capacity assessment of wind farms.

[0112] Step S320: Perform regression analysis on the limit state surface model and its narrow-band neighborhood to obtain the component-level load upper bound function as the load safety boundary.

[0113] In step S320, the discrete limit state points are expanded into a continuous load safety boundary function that can be used for online evaluation, thereby supporting the safety verification of subsequent control strategies. Although the limit state surface model has defined the set of critical responses, it is itself a discrete structure composed of a finite number of simulation or measurement points, making it difficult to directly use for real-time load prediction under arbitrary inflow conditions. Therefore, this step sets a narrow neighborhood (i.e., "narrowband neighborhood") around the limit state surface model, collects the inflow vectors and corresponding component response values ​​of all high-load sample points within this region, and performs robust regression analysis (such as generalized additive model, Gaussian process regression, or support vector regression) within this local range to fit an upper bound function that takes the representative inflow vector as input and the maximum component response as output. This function is defined as the load safety boundary, representing the maximum load estimate that each component can withstand under the corresponding wind conditions. It is conservative and inclusive, ensuring that sufficient safety margin is provided even under model errors or operating condition fluctuations.

[0114] This load safety boundary not only reflects the structural strength limit of the component but also incorporates the effects of material fatigue, dynamic amplification effects, and uncertainties, becoming a hard constraint in the yaw optimization model. By establishing a component-level load safety boundary, the control system can quickly assess its impact on the unit's lifespan before command execution, fundamentally avoiding the risk of sacrificing equipment safety in pursuit of power generation gain.

[0115] In some embodiments of this application, the limit state surface model satisfies the following calculation formula (6):

[0116] (6);

[0117] In formula (6), The limit state surface model represents the set of all incoming flow conditions under which the responses of each key component reach their quantile thresholds within the target return period. This represents the representative flow vector in historical operational data. The first wind farm A key component represents the incoming flow vector. The quantile response at the point, For the first Key components during the target recurrence period The corresponding upper quantile value.

[0118] In some embodiments of this application, the load safety boundary satisfies the following calculation formula (7):

[0119] (7);

[0120] In formula (7), Indicates the first A key component represents the incoming flow vector. The load safety boundary under the condition is the estimated maximum load that the component can withstand under extreme wind conditions; Representing the limit state surface model The narrowband neighborhood, representing the distance No more than small positive numbers All incoming flow vectors A set of values ​​used for robust regression and upper bound estimation within a local range; Denotes supremum, indicating the supremum of the set. All The minimum upper bound of the value, i.e. the maximum possible response value in this neighborhood, is used to conservatively estimate the ultimate load of the component.

[0121] In some embodiments of this application, step S140 involves offline solving for the optimal yaw strategy and its corresponding power gain, fatigue cost, and applicable operating condition range under each wind scenario cluster in the wind scenario library, based on the load safety boundary, wake model, and power optimization model, and constructing a yaw command library, including the following steps.

[0122] Step S410: For each wind scenario in the wind scenario library, the wake model is used to describe the wake effect of the upstream wind turbine on the downstream wind turbine, and the wind speed distribution of the entire wind farm is calculated by superimposing the wake effects of each unit.

[0123] In step S410, a physical model of aerodynamic disturbances within the wind farm is established to accurately characterize the impact mechanism of upstream wind turbine yaw operation on downstream wind turbine wind energy input, providing precise input conditions for subsequent optimization. Under each identified wind condition scenario (including normal and extreme), initial meteorological conditions are first set based on the representative incoming flow vectors (such as wind speed and wind direction) of the scenario. Then, an engineered wake model (such as an improved Jensen model, Ainslie model, or Gaussian wake model) is used to simulate the wake trajectory, extent, and wind speed loss of each wind turbine under different yaw angles.

[0124] Since wind farms typically consist of multiple turbines arranged in a specific layout, downstream turbines are often simultaneously affected by the superimposed wakes of multiple upstream turbines. Therefore, it is necessary to perform spatial superposition calculations on the wake fields of all upstream turbines to obtain the true wind speed distribution map of the entire wind farm in three-dimensional space. This process not only considers the axisymmetric attenuation characteristics of the wake but also introduces the lateral offset effect of the wake caused by yaw angle—that is, when an upstream turbine actively yaws, its wake centerline shifts laterally, thereby reducing the shading of directly downstream turbines and increasing the coverage of turbines to the side or behind. Through this refined modeling, the changes in inflow wind speed of each turbine in the entire farm under different yaw combinations can be accurately assessed, thereby predicting its power output response and providing key input for building an optimization model aimed at maximizing the overall power of the farm. The accuracy of this step directly determines the effectiveness of subsequent optimization results and serves as a physical bridge connecting wind condition perception and control decisions.

[0125] Step S420: Based on wind speed distribution, establish a power optimization model with fatigue penalty and load safety boundary constraints.

[0126] In step S420, a comprehensive optimization objective function is constructed. While pursuing the increase of the total power of the wind farm, it actively considers the impact of control behavior on the fatigue life and structural safety of the units, avoiding the unsustainable mode of "trading power for life." Traditional yaw optimization models usually only take maximizing the total power of the wind farm as a single objective, ignoring the additional mechanical stress and fatigue damage caused by the yaw action itself. This step innovatively introduces a multi-objective trade-off mechanism, constructing a constrained nonlinear optimization model. Its objective function not only includes the normalized total power of the wind farm, but also introduces two penalty terms: one is a normalized fatigue cost penalty term, used to quantify the cumulative damage of key components (such as the main shaft, gearbox, and tower) within the rolling time window under a specific yaw strategy; the other is a normalized yaw action cost penalty term, reflecting the drive energy consumption and mechanical wear caused by frequent yaw system actions. By adjusting the weighting coefficients and aggressiveness parameters, the conservatism of the control strategy can be flexibly adjusted under different wind conditions.

[0127] More importantly, the model introduces the "load safety boundary" determined in step S130 as a hard constraint into the feasible region, requiring that the predicted ultimate load values ​​of each critical component must not exceed their safety upper limit under any combination of yaw angles. Furthermore, the model includes physical constraints such as yaw angle range, yaw rate, yaw change, and cable deflection angle, ensuring the optimization results are feasible in engineering. This comprehensive modeling achieves a shift from "simple power increase" to "power increase-lifetime-safety" synergistic optimization, laying a theoretical foundation for generating sustainable and implementable control strategies.

[0128] Step S430: Based on the power optimization model, the optimal yaw angle distribution under each wind condition scenario cluster and its corresponding power gain, fatigue cost and applicable operating condition range are obtained offline, and a yaw command library is constructed.

[0129] In step S430, the aforementioned modeling results are transformed into practically usable control knowledge assets, namely, constructing a complete and information-rich yaw command library. After completing the wake modeling and optimization model construction for each wind condition scenario, an efficient numerical optimization algorithm (such as sequential quadratic programming, genetic algorithm, or gradient descent) is used to solve the nonlinear constrained optimization problem, searching for the field-wide optimal yaw angle distribution that maximizes the objective function, i.e., the optimal yaw angle that each wind turbine should execute.

[0130] After the solution is obtained, not only is the optimal solution itself recorded, but also its corresponding overall power gain (relative to traditional wind-driven operation), fatigue damage level of key components, yaw system action intensity, and other performance indicators are output simultaneously. The applicable operating conditions of the strategy (such as wind speed range, wind direction sector, turbulence threshold, etc.) are also clearly defined, forming a complete control command record. The optimization results for all wind scenarios are summarized and organized according to the structure of "wind scenario → optimal yaw angle set → power gain → fatigue cost → applicable range," constructing a queryable and updatable yaw command library.

[0131] This instruction library is essentially a "control strategy knowledge graph," storing the optimal solutions for wind farms under various typical operating conditions. During subsequent online operation, the system does not need to recalculate; it only needs to match the closest scenario based on real-time wind conditions to quickly invoke the pre-stored optimal instructions, achieving a response time within seconds. Furthermore, because all instructions undergo rigorous security verification and performance evaluation offline, the reliability and consistency of online execution are ensured, truly achieving the engineering goal of "complex calculations offline and lightweight real-time control."

[0132] In some embodiments of this application, after determining the normal wind conditions and extreme wind conditions, it is necessary to perform wind turbine wake modeling and yaw angle optimization. This invention uses an improved Jensen wake model to describe the wake influence of the upstream unit on the downstream unit, satisfying the following calculation formula (8):

[0133] (8);

[0134] In formula (8), downstream of the wind turbine Wind speed loss at the location. For free-flowing wind speed, For thrust coefficient, The diameter of the wind turbine, is the wake diffusion coefficient.

[0135] In some embodiments of this application, the power optimization model satisfies the following calculation formula (9):

[0136] (9);

[0137] In formula (9), This represents the objective function of the power optimization model, used to solve for the wind farm under a given inflow vector. and yaw angle The optimal yaw strategy is designed to maximize overall power gain while minimizing fatigue costs and yaw action costs. Indicates the first The yaw angle of a typhoon turbine needs to be determined during the optimization process to maximize the objective function by determining the optimal yaw angle for each turbine. Indicates based on and Normalized total power output of wind farms; This indicates the adjustment parameter, with a range of [value missing]. Used to balance the aggressiveness of strategies. A higher value indicates a greater tendency to pursue high power output, but also comes with higher fatigue costs and yaw costs. This represents the weighting coefficient used to adjust the contribution of the fatigue cost penalty term; Indicates based on and The normalized fatigue cost assessment value; This represents the weighting coefficient used to adjust the contribution of the yaw action cost penalty item; Indicates based on Normalized yaw maneuver cost metric.

[0138] In some embodiments of this application, the load safety boundary is incorporated into the feasible region of the power optimization model using a hard constraint method, satisfying the following calculation formula (10):

[0139] (10);

[0140] In formula (10), For the first A key component represents the incoming flow vector. Given yaw angle Given the incoming flow vector The upper bound of the ultimate load under the condition, Indicates the first A key component represents the incoming flow vector. Load safety boundary under the given conditions.

[0141] In some embodiments of this application, the power optimization model also needs to satisfy the following constraints:

[0142] (1) No. Yaw angle of typhoon turbine ,express It needs to be within the physically adjustable range of the yaw angle. , The design and structural limitations of the yaw drive system are designed to prevent mechanical overload or structural interference.

[0143] (2) No. Yaw angle change of a typhoon turbine in a single control cycle , The maximum permissible single-step yaw angle change is used to limit the abrupt change in yaw action and avoid impacting the transmission system.

[0144] (3) Within a unit of time, the first Yaw rate of typhoon turbine absolute value , This indicates the maximum permissible yaw rate, which is determined by the dynamic performance of the yaw motor and drive system, to prevent excessively fast rotation from causing structural vibration or motor overload.

[0145] (4) No. Typhoon turbine in time Cable deflection angle inside , This indicates the maximum permissible cable deflection angle. Exceeding this value will trigger an automatic cable unwinding procedure to prevent damage caused by excessive cable twisting.

[0146] In some embodiments of this application, the new input sample (i.e., the current incoming flow vector) The similarity between a cluster and an existing cluster is measured by Mahalanobis distance, which satisfies the following calculation formula (11):

[0147] (11);

[0148] In formula (11), The current incoming flow vector Mahalanobis distance from the cluster center The mean vector of the cluster centers. This indicates the calculation of the cluster covariance matrix. Approximately obey The distribution (i.e., chi-square distribution) is used to determine the statistical threshold for "similar / dissimilar" relationships. , This indicates the confidence level (typically set to 0.95).

[0149] like That is, among all clusters in the wind condition scenario library, the samples If the squared Mahalanobis distance to the nearest cluster center is less than or equal to the threshold determined by the chi-square distribution, that is, the sample is sufficiently similar to a known wind condition scenario and can be successfully matched, the corresponding yaw command in the yaw command library is invoked.

[0150] like Then it represents a sample If the wind condition scenario library does not match, a smooth interpolation using the kernel weights of neighboring clusters is used to obtain a transitional scenario in order to avoid abrupt changes in control commands.

[0151] In some embodiments of this application, the transition command is generated by weighted interpolation using a kernel function based on Mahalanobis distance, satisfying the following calculation formula (12):

[0152] (12);

[0153] In formula (12), Represents the current incoming flow vector Transitional yaw commands are used to achieve smooth control in scenarios without precise matching; Represents the current incoming flow vector The first in the wind condition scene library The squared Mahalanobis distance between the centers of each wind condition scene cluster reflects the degree of similarity in their statistical characteristics. Represents the current incoming flow vector Number of clusters of similar wind conditions; The kernel function bandwidth parameter controls the decay rate of the influence of neighboring clusters during interpolation. The larger the value, the smoother the interpolation, but the more delayed the response. The smaller the value, the more sensitive the response, but this may introduce jitter. Indicates the first The optimal yaw angle set corresponding to each wind condition scenario cluster is stored in the yaw command library.

[0154] In some embodiments of this application, in order to achieve the safety and reliability of wind turbine operation, the present invention combines fatigue accumulation monitoring and extreme load early warning during the execution of the instruction library. The load signals of key components are statistically analyzed by the rainflow counting method for stress cycles, and fatigue damage is calculated by combining the Miner linear cumulative damage criterion, satisfying the following calculation formula (13):

[0155] (13);

[0156] In formula (13), For cumulative damage value, For the first Number of stress-like cycles This represents the number of cycles that can be withstood for this stress amplitude. This represents the number of stress types. When... When the value approaches or exceeds 1, it indicates that the component's lifespan is nearing its end and maintenance is required. Cycle life. It can be approximated by the SN curve of the material, satisfying the following calculation formula (14):

[0157] (14);

[0158] In formula (14), For stress amplitude, and This is a material constant. If the amplitude of a single stress exceeds the ultimate bearing capacity, the system will immediately trigger an extreme load warning.

[0159] In some embodiments of this application, the yaw command library and wind condition scenario library are dynamically updated based on newly added operational data and the evolution of wind condition distribution. The dynamic update cycle is weekly, monthly or quarterly. The update frequency is adaptively adjusted according to the climate characteristics and wind condition change rate of the wind farm's location to ensure that the yaw command library and wind condition scenario library continuously reflect the actual wind condition distribution.

[0160] Specifically, dynamically updating the yaw command library and wind condition scenario library is the core mechanism to ensure the long-term effective operation of the wake optimization control system. Wind conditions in wind farms are affected by seasonal changes, climate change, and the evolution of the surrounding environment, exhibiting significant time-varying characteristics. If the control strategy relies on an initially constructed static model for a long time, it will be difficult to adapt to the drift of actual wind conditions, leading to insufficient operating condition coverage, increased matching deviation, diminished optimization effects, and even the risk of miscontrol. Therefore, this invention introduces a periodic dynamic update mechanism. By continuously integrating newly added operational data, it re-analyzes wind condition distribution, corrects scenario division, and optimizes control strategies, ensuring that the wind condition scenario library and yaw command library always remain consistent with the current actual operating conditions, avoiding strategy "aging," and thus guaranteeing the system's power enhancement benefits and operational reliability throughout its entire lifecycle.

[0161] This dynamic update mechanism employs a hierarchical, multi-timescale strategy, including weekly, monthly, and quarterly updates, balancing response speed and model stability. Weekly updates focus on short-term fluctuations, using data from the most recent week to slightly adjust the mean vector and weight coefficients of each wind condition cluster to quickly respond to minor shifts in wind direction and speed while maintaining the cluster structure to prevent overfitting. Monthly updates address structural changes; when statistical tests reveal significant shifts in wind condition distribution, clustering results can be merged or split to optimize the rationality of scenario division and adapt to medium- to long-term wind flow field reconstruction. Quarterly updates address seasonal changes, comprehensively reconstructing principal component analysis projections, re-clustering, and solving for the optimal yaw strategy offline, generating a new instruction library version with seasonal characteristics that effectively incorporates seasonal extreme conditions such as typhoons and cold waves.

[0162] Furthermore, the update frequency can be adaptively adjusted according to the climate characteristics and wind condition change rate of the wind farm's location, enhancing the system's flexibility and applicability. For example, the update cycle can be appropriately extended in inland areas with stable wind conditions, while the update frequency can be increased in coastal or mountainous wind farms with variable wind conditions. Through this hierarchical, progressive, and intelligently adjustable update mechanism, this invention achieves continuous tracking and active learning of wind condition evolution, enabling the control system to possess "evolutionary capabilities." This not only solves the performance degradation problem caused by the lack of feedback in traditional offline methods but also ensures that the wake optimization strategy maintains high accuracy and strong adaptability at different time scales, truly achieving sustainable, stable, and efficient power boosting control throughout the year.

[0163] In some embodiments of this application, an onshore wind farm comprising 19 wind turbines in a double-row layout is used as the research object to verify the effectiveness and engineering applicability of the technical solution of this application. The total installed capacity of this wind farm is located in a typical complex terrain area, characterized by significant wind speed variations, turbulence intensity fluctuations, and seasonal wind direction changes. By collecting historical operating data and combining it with the environmental contour line method to construct a wind condition scenario library, an offline yaw command library is generated, and online matching and safety verification are performed, achieving synergistic optimization of the overall power gain and fatigue damage of key components of the wind farm.

[0164] First, historical operational data is acquired from the wind farm's SCADA system, including parameters such as incoming wind speed, wind direction, turbulence intensity, air density, turbine output power, yaw angle, and rotational speed. After missing value imputation, outlier removal, and noise filtering, the statistical characteristics of each variable are extracted using the sliding time window method, forming a structured feature dataset. For example... Figure 2 As shown, cluster analysis is performed on the feature data in three-dimensional space (wind speed-turbulence intensity-air density). After dimensionality reduction using principal component analysis (PCA), 10 clusters, namely wind condition scene clusters, are divided using Gaussian mixture model (GMM). Each cluster is characterized by its mean vector, covariance matrix and weight coefficients, forming a wind condition scene library. Figure 3 and Figure 4 The clustering results under two-dimensional projections of wind speed-turbulence intensity and wind speed-wind direction are presented, clearly reflecting the spatial distribution characteristics of different wind condition patterns and providing a foundation for subsequent modeling.

[0165] Building upon this, the environmental contour method was introduced, and a joint probability distribution model of wind speed (U), turbulence intensity (ti), and turbulence scale (Hs) was constructed based on historical extreme event data, under the target return period (e.g., once in 50 years). The calculated results are as follows: Figure 5The isosurface shown is the limit state surface model. The surface exhibits a typical nonlinear morphology. As wind speed increases, the surface is most sensitive to changes in turbulence intensity and scale in the moderate wind speed range (approximately 8–12 m / s), indicating that this region is most likely to trigger the ultimate load response of critical components. The overall surface delineates a closed safety domain: the interior of the surface represents acceptable operating condition combinations, while the exterior represents the danger zone exceeding design limits. By performing regression analysis on the limit state surface model and its neighborhood, the load safety boundaries for critical components such as the main shaft and the tower base are obtained.

[0166] Subsequently, for each normal and extreme scenario in the wind condition scenario library, an improved Jensen wake model was used to simulate the impact of upstream unit yaw on downstream, and the wake effects of each unit were superimposed to calculate the wind speed distribution across the entire field; a power optimization model with fatigue penalty terms and load safety boundary constraints was established to solve the optimal yaw angle distribution offline. Figure 6 The optimal yaw command values ​​(i.e., yaw command library) for each wind turbine under 10 wind conditions are shown. It can be seen that there are significant differences in the yaw strategy of wind turbines in different positions, especially in the middle row and downstream units, which exhibit the characteristics of active wake avoidance.

[0167] Based on the yaw command library, the power gain of wind farms under 10 wind conditions is as follows: Figure 7 As shown, Figure 7 The figure demonstrates significant differences in power gain under different wind conditions: Cluster 1 exhibits the highest power gain, reaching approximately 1.15 MW, indicating that optimizing the yaw strategy can achieve the greatest power generation improvement under these wind conditions; Clusters 8 and 9 also show relatively high gains, approaching 1.0 MW respectively, suggesting that these conditions also possess significant wake control potential; while Clusters 3, 4, 5, and 6 show lower gains, some approaching zero or showing only slight increases, reflecting that the turbine layout or wind energy distribution in these wind conditions results in minimal wake impact and limited optimization space. Overall, the figure visually demonstrates the effectiveness of the yaw optimization strategy under different wind conditions, validating the necessity of constructing a multi-scenario yaw command library, i.e., developing differentiated control schemes for different wind characteristics to achieve global optimization and continuous power gain across the entire field.

[0168] Figure 8 The annual power generation was compared under different wind energy utilization hours (100%, 4000h, 3000h, and 2000h). The results show that the optimized strategy can stably increase power generation under both full-load and low-load conditions, with a maximum gain of 8.6 GWh / year. Meanwhile, Figure 9The system demonstrated the spindle fatigue load time history curve and Miner's linear cumulative damage calculation results. It successfully identified multiple high fatigue damage events and triggered early warnings, ensuring that the control strategy pursues increased power without sacrificing equipment lifespan. The entire process achieved a complete control flow of "offline pre-simulation, online matching, dynamic updating, and safety closed loop," verifying the significant advantages of this invention in improving the operating efficiency and safety of wind farms.

[0169] In summary, the wind farm wake optimization control method based on yaw command library provided in this application has the following technical effects.

[0170] First, by constructing a wind condition scenario library, a systematic classification and modeling of the wind farm operating environment is achieved. The complex and ever-changing actual wind conditions are summarized into several typical scenarios, effectively solving the problem that traditional methods struggle to comprehensively cover control strategies due to the continuity and high dimensionality of wind conditions, significantly improving the adaptability and generalization ability of the control system. Second, for extreme wind conditions, an environmental contour line method is introduced and a load safety boundary is established, enabling accurate prediction and constraint of the extreme responses of key components. This ensures that the yaw optimization process pursues power generation gain without sacrificing unit structural safety, enhancing the robustness and engineering feasibility of the control strategy. Third, by solving the power optimization model with fatigue penalty and load constraints offline, a yaw command library containing optimal yaw angle distribution, power gain, fatigue cost, and applicable range is pre-generated. This achieves an efficient control mode of "offline pre-simulation and online invocation," significantly reducing the real-time computational burden and improving response speed and stability. Furthermore, by combining the Mahalanobis distance matching mechanism and dynamic update strategy, the system can quickly match the optimal control command based on real-time wind conditions and continuously optimize the command library content with seasonal changes and operational data accumulation, ensuring that control performance does not degrade during long-term operation. Finally, through fatigue accumulation monitoring and extreme load early warning mechanisms, closed-loop monitoring of the unit's health status was achieved, further improving the safety and economy of wind farm operation. Overall, this method achieves synergistic optimization between proactive control of wake effects, continuous improvement of power generation performance, and effective protection of equipment lifespan, providing a reliable technical path for the intelligent and refined operation control of large-scale wind farms.

[0171] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0172] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of the present invention.

Claims

1. A wind farm wake optimization control method based on a yaw command library, characterized in that, Includes the following steps: Historical operational data of wind farms are acquired, and data preprocessing and feature extraction are performed to obtain a wind farm feature dataset. Cluster analysis was performed on the wind farm feature dataset to divide it into several clusters of normal wind conditions and extreme wind conditions, and a wind condition scenario library was constructed. Based on the aforementioned extreme wind condition scenario cluster, an environmental contour line method is introduced to construct a limit state surface model and determine the load safety boundaries of key components in the wind farm, including the following steps: Based on the historical operating data, environmental contour lines under the target recurrence period are constructed in physical space, the quantile responses of each key component of the wind farm at the representative inflow vector are calculated, and the limit state surface model is constructed. Regression analysis is performed on the limit state surface model and its narrowband neighborhood to obtain the component-level load upper bound function as the load safety boundary; The limit state surface model satisfies the following calculation formula: ； in, The limit state surface model represents the set of all incoming flow conditions under which the responses of each key component reach their quantile thresholds within the target recurrence period. This represents the representative inflow vector in the historical operational data. The first wind farm A key component represents the incoming flow vector. The quantile response at the point, For the first Key components during the target recurrence period The corresponding upper quantile value; The load safety boundary satisfies the following calculation formula: ； in, Indicates the first A key component represents the incoming flow vector. The load safety boundary under the condition is the estimated maximum load that the component can withstand under extreme wind conditions; Representing the limit state surface model The narrowband neighborhood, representing the distance No more than small positive numbers All incoming flow vectors A set of values ​​used for robust regression and upper bound estimation within a local range; Denotes supremum, indicating the supremum of the set. All The minimum upper bound of the value, i.e. the maximum possible response value in this neighborhood, is used to conservatively estimate the ultimate load of the component. Based on the load safety boundary, wake model and power optimization model, the optimal yaw strategy and its corresponding power gain, fatigue cost and applicable operating condition range under each wind scenario cluster in the wind scenario library are solved offline, and a yaw command library is constructed. Real-time acquisition of current wind conditions, generation of feature vectors and matching with the wind condition scenario library; when a match is successful, the corresponding yaw instruction in the yaw instruction library is invoked; when no match is found, a transition instruction is generated by interpolation based on the yaw instructions corresponding to the nearest clusters of wind condition scenarios in the wind condition scenario library. The transition command is generated by weighted interpolation using a kernel function based on Mahalanobis distance, satisfying the following calculation formula: ； in, Represents the current incoming flow vector Transitional yaw commands are used to achieve smooth control in scenarios without precise matching; Represents the current incoming flow vector The first wind condition scenario in the aforementioned wind condition scenario library The squared Mahalanobis distance between the centers of each wind condition scene cluster reflects the degree of similarity in their statistical characteristics. This indicates that the wind condition scenario library contains the vector of the current incoming flow. Number of the nearest wind condition scene clusters; The kernel function bandwidth parameter controls the decay rate of the influence of neighboring clusters during interpolation. The larger the value, the smoother the interpolation, but the more delayed the response. The smaller the value, the more sensitive the response, but this may introduce jitter. Indicates the first The optimal yaw angle set corresponding to each wind condition scenario cluster is stored in the yaw command library; The yaw command or transition command is subjected to load safety verification. If the verification is successful, it is issued for execution; otherwise, the protection mechanism is activated to reduce the control intensity or return to the normal wind-fighting operation mode.

2. The wind farm wake optimization control method based on yaw command library according to claim 1, characterized in that, The historical operating data includes incoming flow data, wind turbine power, yaw angle, and rotational speed; the incoming flow data is used to characterize wind conditions, including wind direction, wind speed, air density, and turbulence intensity; the data preprocessing includes missing value imputation, outlier removal, and noise filtering; the feature extraction refers to calculating representative statistical features of the historical operating data, including the incoming flow vector, based on the time series using a sliding time window method.

3. The wind farm wake optimization control method based on yaw command library according to claim 1, characterized in that, The process of performing cluster analysis on the wind farm feature dataset to divide it into several clusters of normal wind conditions and extreme wind conditions, and constructing a wind condition scenario library, includes the following steps: Principal component analysis was performed on the wind farm feature dataset to reduce its dimensionality and extract the main variation features, resulting in a wind condition distribution map in a low-dimensional feature space. Based on the wind condition distribution map, the number and initial location of cluster centers are determined. Cluster analysis is performed using a Gaussian mixture model or a density-based clustering method to generate several wind condition scene clusters, including regular wind condition scene clusters and extreme wind condition scene clusters. Each wind condition scene cluster includes a cluster center mean vector, a covariance matrix, and weight coefficients, which are used to characterize the statistical characteristics of various wind conditions. Based on the distribution characteristics of the aforementioned wind condition scene clusters, and combined with the frequency and coverage of historical wind conditions, a wind condition scene library for various wind conditions is constructed.

4. The wind farm wake optimization control method based on yaw command library according to claim 1, characterized in that, Based on the load safety boundary, wake model, and power optimization model, the optimal yaw strategy and its corresponding power gain, fatigue cost, and applicable operating condition range for each wind scenario cluster in the wind scenario library are solved offline to construct a yaw command library, including the following steps: For each wind scenario in the wind scenario library, a wake model is used to describe the wake effect of the upstream wind turbine on the downstream wind turbine, and the wind speed distribution of the entire wind farm is calculated by superimposing the wake effects of each unit. Based on the wind speed distribution, a power optimization model with fatigue penalty and load safety boundary constraints is established; Based on the power optimization model, the optimal yaw angle distribution under each wind condition scenario cluster and its corresponding power gain, fatigue cost and applicable operating condition range are obtained offline, and a yaw command library is constructed.

5. The wind farm wake optimization control method based on yaw command library according to claim 4, characterized in that, The power optimization model satisfies the following calculation formula: ； in, This represents the objective function of the power optimization model, used to solve for the wind farm under a given inflow vector. and yaw angle The optimal yaw strategy is designed to maximize overall power gain while minimizing fatigue costs and yaw action costs. Indicates the first The yaw angle of a typhoon turbine needs to be determined during the optimization process to maximize the objective function by determining the optimal yaw angle for each turbine. Indicates based on and Normalized total power output of wind farms; This indicates the adjustment parameter, with a range of [value missing]. Used to balance the aggressiveness of strategies. A higher value indicates a greater tendency to pursue high power output, but also comes with higher fatigue costs and yaw costs. This represents the weighting coefficient used to adjust the contribution of the fatigue cost penalty term; Indicates based on and The normalized fatigue cost assessment value; This represents the weighting coefficient used to adjust the contribution of the yaw action cost penalty item; Indicates based on Normalized yaw maneuver cost metric.

6. The wind farm wake optimization control method based on yaw command library according to claim 5, characterized in that, The load safety boundary is incorporated into the feasible region of the power optimization model using a hard constraint method, satisfying the following calculation formula: ； in, For the first A key component represents the incoming flow vector. Given yaw angle Given the incoming flow vector The upper bound of the ultimate load under the condition, Indicates the first A key component represents the incoming flow vector. Load safety boundary under the given conditions; The power optimization model must also satisfy the following constraints: No. Yaw angle of typhoon turbine ,express It needs to be within the physically adjustable range of the yaw angle. , The design and structural limitations of the yaw drive system are designed to prevent mechanical overload or structural interference. No. Yaw angle change of a typhoon turbine in a single control cycle , The maximum allowable single-step yaw angle change is used to limit the abrupt change in yaw action and avoid impacting the transmission system; Within a unit of time, the first Yaw rate of typhoon turbine absolute value , This indicates the maximum permissible yaw rate, which is determined by the dynamic performance of the yaw motor and drive system, to prevent excessively fast rotation from causing structural vibration or motor overload. No. Typhoon turbine in time Cable deflection angle inside , This indicates the maximum permissible cable deflection angle. Exceeding this value will trigger an automatic cable unwinding procedure to prevent damage caused by excessive cable twisting.

7. The wind farm wake optimization control method based on yaw command library according to claim 1, characterized in that, Based on the newly added operational data and the evolution of wind conditions, the yaw command library and wind condition scenario library are dynamically updated. The dynamic update cycle is weekly, monthly, or quarterly. The update frequency is adaptively adjusted according to the climate characteristics and wind condition change rate of the wind farm's location to ensure that the yaw command library and wind condition scenario library continuously reflect the actual wind condition distribution.

Images