Methods, devices and electronic equipment for coordinated yaw wake control of wind farm clusters

By constructing a global state and decision model for joint training in a wind farm, the problem of reduced power generation caused by wake effect was solved, global collaborative optimization and independent rapid decision-making were achieved, and the overall power generation efficiency of the wind farm was improved.

CN123109944APending Publication Date: 2026-11-17HUADIAN ELECTRIC POWER SCI INST CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611270372.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In large wind farms, the wake effect leads to a decrease in the wind energy capture rate and power generation of downstream wind turbines. Existing control strategies ignore the aerodynamic coupling between wind turbines, resulting in a low total power generation for the entire farm.

Method used

By acquiring historical local observation data of each wind turbine, a global state of the wind farm is constructed. The decision-making model is then jointly trained using the global state and local observation data to build a strategy and value network, set a reward function, and achieve global collaborative optimization and independent rapid decision-making.

Benefits of technology

It improves the overall power generation efficiency of large wind farms. By participating in training with global information, the yaw action of wind turbines is coordinated and optimized in the direction of increasing the total power generation of the entire farm, which effectively improves power generation efficiency while suppressing wind turbine load and yaw amplitude.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123109944A_ABST
    Figure CN123109944A_ABST
Patent Text Reader

Abstract

This application relates to the field of wind power generation control technology, and discloses a method, device, and electronic equipment for coordinated yaw wake control of a wind farm cluster. The control method includes: acquiring historical local observation data of each wind turbine and constructing a global state of the wind farm based on this data; jointly training the decision models corresponding to each wind turbine using the global state and historical local observation data; after training, inputting the current local observation data of each wind turbine into the corresponding decision model, obtaining the output yaw action, and generating a yaw control command to be sent to the corresponding wind turbine. This application, through joint training of the global state and local observation data of the wind farm, enables the decision models of each wind turbine to learn global aerodynamic coupling information during the training phase, and to output yaw actions based solely on their own local observation data during the execution phase. This achieves the unification of global coordination during the training phase and independent decision-making during the execution phase, effectively improving the overall power generation efficiency of the wind farm.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of wind power generation control technology, specifically to a method, device, and electronic equipment for coordinated yaw wake control of wind farm clusters. Background Technology

[0002] As the installed capacity of newly built wind farms continues to expand and the wind turbine locations become increasingly dense, the upstream wind turbines capture wind energy and form a wake region with reduced wind speed downstream. This results in a decrease in the wind energy capture rate and a significant drop in power generation of downstream wind turbines located in this region, making the problem of low total power generation of the entire farm particularly prominent.

[0003] In related technologies, in order to reduce the impact of wake effect, wind farms are generally regarded as a whole. Existing wind farm control strategies mostly adopt "greedy control", that is, each wind turbine maximizes its own wind capture efficiency, ignoring the aerodynamic coupling between wind turbines, which further aggravates the loss of power generation of the whole farm. Summary of the Invention

[0004] This application provides a method, device, and electronic equipment for coordinated yaw wake control of wind farm groups to solve the technical problem of low total power generation in large wind farms due to wake effects.

[0005] In a first aspect, this application provides a method for coordinated yaw wake control of a wind farm cluster, comprising: acquiring historical local observation data of each wind turbine, and constructing a global state of the wind farm based on the historical local observation data of each wind turbine, wherein the historical local observation data includes at least the nacelle-side wind speed, wind direction, turbulence intensity, current yaw error angle, active power, and equivalent bending moment; jointly training the decision model corresponding to each wind turbine using the global state of the wind farm and the historical local observation data of each wind turbine; after training, acquiring the current local observation data of each wind turbine and inputting it into the decision model corresponding to each wind turbine, acquiring the yaw action output by each decision model; generating yaw control commands for each wind turbine based on each yaw action and issuing them to the corresponding wind turbine.

[0006] Beneficial Effects: This application acquires historical local observation data of each wind turbine and constructs a global state of the wind farm. Then, it uses the global state of the wind farm and the historical local observation data of each turbine to jointly train the decision-making models corresponding to each turbine. This allows each turbine's decision-making model to learn the state information of all turbines globally and the aerodynamic coupling relationships between them during the training phase. After joint training, the decision-making models no longer make decisions in isolation but incorporate global collaborative information during the training process. Therefore, after training, when each turbine can output yaw actions based solely on its current local observation data, the decision-making of each turbine is actually guided by the acquired global collaborative knowledge, thus achieving a unity between global collaborative optimization in the training phase and independent and rapid decision-making in the execution phase. By involving global information in the training, the yaw actions of each turbine are collaboratively optimized in the direction of increasing the total power generation of the entire wind farm, effectively improving the overall power generation efficiency of large-scale wind farms.

[0007] In one optional implementation, before acquiring the historical local observation data of each wind turbine, the method further includes: cleaning outliers from the initial data, wherein the outlier cleaning includes: removing unreasonable data points based on the theoretical power curve of each wind turbine, and filtering out abrupt spike data points in the time series using a filter; performing cubic spline interpolation on the initial data with different sampling frequencies to align the initial data to the same time axis; and normalizing the aligned data to map each state variable to the [0,1] interval.

[0008] Beneficial Effects: This application effectively removes invalid data generated by abnormal operating conditions such as sensor malfunctions, power outages, and shutdowns by cleaning outliers from the initial data, thus avoiding interference from noisy data on model training. Cubic spline interpolation aligns data from different sampling frequencies to the same time axis, solving the problem of multi-source data being unable to be directly fused due to different sampling frequencies. Min-Max normalization maps each state variable to the [0,1] interval, accelerating the convergence speed of the neural network and improving the stability of model training. These three factors work together to provide higher-quality, more uniformly formatted, and faster-converging input data for the subsequent joint training of the decision model, thereby improving the effectiveness and consistency of the final collaborative control effect.

[0009] In one optional implementation, the step of using the global state of the wind farm and historical local observation data of each wind turbine to perform correlation training on the decision model corresponding to each wind turbine includes: constructing a policy network and a value network corresponding to each wind turbine, wherein the policy network is used to output yaw actions based on the historical local observation data, and the value network is used to evaluate the yaw actions output by the policy network based on the global state of the wind farm; and updating the network parameters of the policy network based on the evaluation results of the value network.

[0010] Beneficial Effects: This application achieves a centralized training and decentralized execution approach by constructing a policy network and a value network corresponding to each wind turbine. The policy network inputs local observation data to output yaw actions, while the value network inputs the global state of the wind farm to evaluate these yaw actions. During training, the value network uses the global state of the wind farm to uniformly evaluate the yaw actions of each wind turbine, enabling the policy networks of each turbine to obtain feedback signals based on the overall field state when updating parameters. This achieves collaborative optimization of the decision-making models of each turbine. This training architecture allows the policy networks of each wind turbine to output yaw actions containing collaborative information by inputting only their own local observation data during the execution phase, providing a specific network structure and parameter update method for the joint training in claim 1.

[0011] In one optional implementation, the step of using the global state of the wind farm and historical local observation data of each wind turbine to perform correlation training on the decision model corresponding to each wind turbine further includes: setting a reward function, the reward function including a first term positively correlated with the total power generation of the wind farm, a second term negatively correlated with the load of each wind turbine, and a third term negatively correlated with the amplitude of each yaw action; calculating the reward value of the yaw action according to the reward function, and updating the network parameters of the value network and the policy network based on the reward value.

[0012] Beneficial Effects: This application unifies the three objectives of increasing overall power generation, reducing turbine fatigue load, and suppressing frequent yaws within a single optimization framework by setting a reward function that includes a first term positively correlated with overall power generation, a second term negatively correlated with the load of each turbine, and a third term negatively correlated with the amplitude of each yaw action. During training, the agent receives a comprehensive reward value based on this reward function for each yaw action output. This reward value reflects the combined impact of the action on overall power generation, turbine loads, and yaw wear. The value network and policy network are updated based on this reward value, causing the policy network to gradually converge to a decision strategy that balances the three objectives. This ensures that the yaw actions of each turbine automatically constrain yaw amplitude and load while pursuing an increase in overall power generation.

[0013] In one optional implementation, the first term is the difference between the total power generation of the wind farm and the reference power generation divided by the reference power generation, where the reference power generation is the estimated total power generation of the wind farm when all wind turbines are operating in wind conditions; the second term is determined based on the sum of the squares of the equivalent bending moments of each wind turbine; the reward function further includes a dynamic weight adjustment coefficient, which is determined based on the current turbulence intensity of the wind farm and is used to increase the weight of the second term when the turbulence intensity increases.

[0014] Beneficial Effects: This application specifically defines the first term of the reward function as the ratio of total power generation to the baseline power, ensuring that the agent's optimization objective directly corresponds to the actual control requirement of increasing total power generation. The second term is specifically defined as the sum of the squares of the equivalent bending moments of each wind turbine. Utilizing the mathematical properties of the squares, the penalty value is rapidly amplified when the load exceeds limits, effectively suppressing extreme load conditions and protecting the safety of critical unit structures. By introducing a dynamic weight adjustment coefficient based on turbulence intensity, the reward function can automatically adjust its optimization focus according to real-time wind conditions—prioritizing power gain under stable wind conditions and automatically increasing the load penalty weight when turbulence intensity increases. This mechanism enables the multi-objective reward function to have the adaptive capability to dynamically adjust its optimization direction according to environmental changes.

[0015] In one optional implementation, the step of using the global state of the wind farm and the historical local observation data of each wind turbine to perform correlation training on the decision model corresponding to each wind turbine further includes: performing weighted fusion of the historical local observation data of each wind turbine based on a spatial attention mechanism, and using the weighted fusion features as input to the value network; the spatial attention mechanism is used to determine the attention weight of each wind turbine according to the wake coupling relationship between each wind turbine, and dynamically adjust the attention weight of each wind turbine according to the real-time wind direction.

[0016] Beneficial Effects: This application introduces a spatial attention mechanism at the input of the value network to weightedly fuse historical local observation data for each wind turbine. This enables the value network to automatically identify affected turbines within the downstream fan-shaped impact zone based on the wake coupling relationship between turbines when evaluating their yaw behavior, and assign them higher attention weights. When the real-time wind direction changes, this attention mechanism can automatically switch the focus, ensuring that the value network's evaluation focus remains on the turbine most severely affected by the wake. This mechanism makes the value network's evaluation more accurate, strengthens the ability of each decision model to perceive the physical propagation characteristics of the wake, and avoids the communication and computational overhead caused by the indiscriminate evaluation of all turbines in traditional methods.

[0017] In one optional implementation, the step of using the global state of the wind farm and historical local observation data of each wind turbine to perform correlation training on the decision model corresponding to each wind turbine further includes: constructing a virtual environment using a wake simulation model, pre-training the decision model in the virtual environment; transferring the network parameters of the pre-trained decision model to the actual wind farm, and fine-tuning the decision model using real-time data collected on-site.

[0018] Beneficial Effects: This application utilizes a wake simulation model to construct a virtual environment for pre-training, enabling the decision-making model to quickly learn the basic physical laws of wind speed loss in a safe and low-cost environment. This avoids the high costs and potential safety risks associated with direct trial and error in actual wind farms using pure reinforcement learning models. Furthermore, the pre-trained network parameters are transferred to an actual wind farm, and the model is fine-tuned using real-time field data. This eliminates the differences between the simulation environment and the actual wind farm, allowing the model to quickly adapt to seasonal wind condition changes and atmospheric stability variations. This training mechanism effectively solves the problems of low exploration efficiency and slow convergence speed in the cold start phase of pure data-driven algorithms, significantly improving the model's adaptability and control performance during actual deployment.

[0019] In one optional implementation, generating yaw control commands for each wind turbine based on each yaw action includes: performing a first-order low-pass filter on each yaw action, and generating the yaw control command based on the filtered yaw action; and performing amplitude limiting on each yaw action, wherein the amplitude limiting includes limiting the adjustment amount of each yaw action to a preset maximum yaw step size range; generating a cable release control command when the cumulative yaw angle of any wind turbine exceeds a preset angle threshold; and forcibly setting the yaw control command of the corresponding wind turbine to a wind-facing mode command when the yaw motor temperature of any wind turbine exceeds a preset temperature threshold.

[0020] Beneficial Effects: This application establishes a physical safety filter between the yaw action output by the strategy network and the actual yaw control command issued to the wind turbine. By performing first-order low-pass filtering and amplitude limiting on the yaw action, it smooths out abrupt changes in the strategy network output and limits excessive amplitude commands, reducing the instantaneous impact on the yaw motor. When the cumulative yaw angle of the wind turbine exceeds a preset angle threshold, the system automatically generates a cable release control command to trigger mechanical cable release logic, keeping the cable within a safe operating angle range. When the yaw motor temperature exceeds a preset temperature threshold, the system automatically switches the control command for the corresponding wind turbine to wind-facing mode, ensuring the motor operates within a safe temperature range. Each safety barrier functions independently, collectively constraining the yaw command output by the AI ​​decision model within the safe operating boundaries of the equipment.

[0021] Secondly, this application provides a wind farm cluster coordinated yaw wake control system, comprising: an acquisition module, used to acquire historical local observation data of each wind turbine and construct a global state of the wind farm based on the historical local observation data of each wind turbine, wherein the historical local observation data includes at least nacelle-side wind speed, wind direction, turbulence intensity, current yaw error angle, active power, and equivalent bending moment; a training module, used to jointly train the decision model corresponding to each wind turbine using the global state of the wind farm and the historical local observation data of each wind turbine; an execution module, used to acquire the current local observation data of each wind turbine and input it into the decision model corresponding to each wind turbine, and acquire the yaw action output by each decision model; a generation module, used to generate yaw control commands for each wind turbine based on each yaw action; and a distribution module, used to distribute each yaw control command to the corresponding wind turbine.

[0022] Thirdly, this application provides an electronic device, including: a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the wind farm group coordinated yaw wake control method described in the first aspect or any corresponding embodiment.

[0023] Fourthly, this application provides a computer-readable storage medium storing computer instructions for causing a computer to execute the wind farm group coordinated yaw wake control method described in the first aspect or any corresponding embodiment. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of the first type of wind farm group coordinated yaw wake control method according to an embodiment of this application; Figure 2 This is a second flowchart illustrating the wind farm group coordinated yaw wake control method according to an embodiment of this application; Figure 3 This is a schematic diagram of the third process of the wind farm group coordinated yaw wake control method according to the embodiments of this application; Figure 4 This is a schematic diagram of the fourth process of the wind farm group coordinated yaw wake control method according to the embodiments of this application; Figure 5This is a structural block diagram of a wind farm group coordinated yaw wake control device according to an embodiment of this application; Figure 6 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0027] It is understood that before using the technical solutions disclosed in the various embodiments of this application, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this application in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained.

[0028] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0029] As an optional application scenario of this application embodiment, the wind farm group coordinated yaw wake control system may include a wind farm and a central control center. Multiple wind turbines are deployed in the wind farm, each communicating with the central control center, which houses a wind farm controller. For example, the wind farm may include wind turbine 1, wind turbine 2, ..., wind turbine N, and each wind turbine is connected to the central control center via a farm network.

[0030] The wind turbines can be either doubly-fed induction generators or permanent magnet direct-drive wind turbines. Each turbine is equipped with a yaw actuator, a SCADA data acquisition terminal, and a nacelle-mounted lidar wind measurement device. The control center can be a local server deployed within the wind farm's substation, or a remote cloud server or edge computing node. The wind farm controller can be a standalone physical server, a server cluster, or a distributed system. The site network can be a wired or wireless network, examples of which include, but are not limited to, fiber optic ring networks, industrial Ethernet, 4G / 5G cellular networks, and combinations thereof.

[0031] In some embodiments, the wind farm controller is used to execute the wind farm group coordinated yaw wake control method provided in the embodiments of this application. The wind farm controller acquires historical local observation data of each wind turbine, constructs the global state of the wind farm based on the historical local observation data of each wind turbine, and then uses the global state of the wind farm and the historical local observation data of each wind turbine to jointly train the decision model corresponding to each wind turbine. After training, the wind farm controller acquires the current local observation data of each wind turbine and inputs it into the decision model of each wind turbine, acquires the yaw action output by each decision model, generates yaw control commands based on each yaw action, and sends them to the corresponding wind turbine to control the yaw actuator of each wind turbine.

[0032] In some embodiments, the training process described above can also be completed by a training server independent of the wind farm controller. After training, the network parameters of each decision model are deployed to the wind farm controller, which only performs online inference and issues control commands. In some embodiments, the wind farm controller can be deployed in a distributed manner, for example, deploying the corresponding decision model in the local PLC of each wind turbine. The wind turbines share the global state information of the wind farm through the farm network to achieve decentralized collaborative control.

[0033] It should be noted that the above scenario is merely an example of an application scenario and does not limit the scope of protection of this application. Those skilled in the art will understand that in actual implementation, wind farms may include more or fewer wind turbines, and may also adopt other network topologies, such as ring topology, star topology, or mesh topology.

[0034] In related technologies, wind turbines in wind farms typically operate independently using a "greedy control" strategy, with each turbine maximizing its own wind capture efficiency while ignoring the aerodynamic coupling between turbines. This results in the total power generation of the entire farm failing to reach the global optimum. Some solutions have attempted to solve for the optimal yaw angle of each turbine by establishing a wind farm wake model and using optimization algorithms. However, these methods are limited by model fidelity and computational complexity, making it difficult to balance accuracy and real-time performance in practical applications.

[0035] To address this, this application provides a wind farm group collaborative yaw wake control method. By acquiring historical local observation data from each wind turbine to construct the global state of the wind farm, and using the global state and the historical local observation data of each turbine, the decision models corresponding to each turbine are jointly trained. This allows each decision model to fully learn the aerodynamic coupling information of the entire field during the training phase. After training, each turbine can output yaw action based solely on its current local observation data, balancing global collaborative optimization with real-time control response speed, effectively improving the overall power generation efficiency of the wind farm. The following detailed description of the wind farm group collaborative yaw wake control method provided in this application is in conjunction with the accompanying drawings.

[0036] According to an embodiment of this application, a method for coordinated yaw wake control of a wind farm group is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0037] This embodiment provides a method for coordinated yaw wake control of wind farm clusters, which can be used in the wind farm controller in the aforementioned centralized control center. Figure 1 This is a flowchart of a wind farm group coordinated yaw wake control method according to an embodiment of this application, such as... Figure 1 As shown, the process includes the following steps: Step S101: Obtain historical local observation data for each wind turbine, and construct the global state of the wind farm based on the historical local observation data for each wind turbine. The historical local observation data includes at least the nacelle-side wind speed, wind direction, turbulence intensity, current yaw error angle, active power, and equivalent bending moment.

[0038] Specifically, historical local observation data refers to the status data collected and recorded by sensors during the historical operation of each wind turbine.

[0039] In this embodiment, the local observation vector of the i-th wind turbine at time t Include:

[0040] in, For cabin side wind speed, For wind direction, For turbulence intensity, For the current yaw error angle, For active power and For the equivalent bending moment calculated based on fatigue load, the global state It is the set of all wind turbine states, i.e., s t ={o t 1 ,o t 2 ,…,o t N} where N is the total number of wind turbines in the wind farm. Historical local observation data for each of the above wind turbines can be obtained through the wind farm SCADA system, nacelle lidar wind measurement equipment, and vibration monitoring instruments.

[0041] The nacelle side wind speed refers to the incoming wind speed at the hub height, measured by an anemometer (such as an ultrasonic anemometer or a mechanical anemometer) installed on the top of the wind turbine nacelle; the wind direction refers to the angle of the incoming wind direction measured by the nacelle wind vane; turbulence intensity is the ratio of the root mean square component of wind speed fluctuations to the average wind speed within a certain time window, used to characterize the severity of turbulence in the wind field; the current yaw error angle is the angle between the wind turbine nacelle axis and the incoming wind direction, when this angle is 0°, it indicates that the wind turbine is completely aligned with the wind; active power refers to the current electrical power output of the wind turbine; equivalent bending moment refers to the equivalent fatigue load calculated based on measured strain data at the blade root or tower bottom using fatigue damage assessment methods such as rainflow counting, used to quantify the structural fatigue degree of the wind turbine. Historical local observation data for each of the above wind turbines can be obtained through wind farm SCADA systems, nacelle lidar anemometers, and vibration monitoring instruments. The global state of a wind farm refers to the overall state representation of the entire wind farm, composed of historical local observation data from each wind turbine. It includes the state information of all wind turbines within the wind farm, reflecting the current overall operating conditions and aerodynamic coupling relationships of the wind farm. For example, in a wind farm containing N wind turbines, the global state of the wind farm is the set of historical local observation data from each of the N wind turbines.

[0042] Step S102: Jointly train the decision-making models corresponding to each wind turbine using the global state of the wind farm and the historical local observation data of each wind turbine.

[0043] In this embodiment, the wind farm wake optimization control problem is constructed as a multi-agent Markov decision process, whose core elements are tuples. Description. Here, S is the global state space, composed of local observation data from each wind turbine; O is the local observation space; A is the continuous action space of each wind turbine, i.e., the adjustment amount of the yaw angle; P is the state transition probability; R is the reward function; and γd is the discount factor, used to balance the weights of current and future rewards.

[0044] Specifically, joint training refers to the synchronous training of the decision-making models of each wind turbine using the global state of the wind farm during the training process. This allows each turbine's decision-making model to obtain global feedback signals when updating parameters. In this process, each turbine's decision-making model not only learns the relationship between its own historical local observation data and yaw actions, but also learns the impact of its actions on other turbines across the entire wind farm through the global state of the wind farm. Joint training enables each turbine's decision-making model to share the common goal of maximizing the total power generation of the entire wind farm when updating parameters. Therefore, after training, the embedded strategies of each decision-making model already incorporate information about the entire wind farm's collaborative processes. For example, after joint training, the decision-making models of the upstream turbines can learn the impact of their yaw angle adjustments on the inlet wind speed of the downstream turbines, thus automatically taking into account the aerodynamic coupling relationship between the upstream and downstream turbines when outputting yaw actions.

[0045] Step S103: After training is completed, the current local observation data of each wind turbine is obtained and input into the decision model of each wind turbine, and the yaw action output by each decision model is obtained.

[0046] Specifically, the current local observation data refers to the current state data of each wind turbine acquired by the wind farm's real-time data acquisition system during the actual control phase after training. Its data dimensions are the same as the historical local observation data in step S101, including but not limited to current nacelle-side wind speed, wind direction, turbulence intensity, current yaw error angle, active power, and equivalent bending moment. The decision model refers to the neural network model obtained after joint training in step S102, capable of outputting corresponding yaw actions based on the input local observation data. Each wind turbine has its own corresponding decision model. During the online execution phase, the current local observation data of each wind turbine is input into its respective decision model. Each decision model independently performs forward propagation calculations and outputs a yaw action. This yaw action is a continuous value, representing the suggested adjustment amount of the wind turbine's yaw angle. Since each decision model has learned the overall collaborative information of the wind farm through the global state during the training phase, although each wind turbine makes independent decisions during online execution, the output yaw action already incorporates the results of global collaborative optimization.

[0047] Step S104: Generate yaw control commands for each wind turbine based on each yaw action and send them to the corresponding wind turbine.

[0048] Specifically, the yaw control command refers to the specific control signal issued to the yaw actuators of each wind turbine. This command is generated based on the yaw action output by the decision model. After receiving the yaw control command, the yaw actuator of each wind turbine drives the nacelle to yaw by a corresponding angle in the direction of the command, thereby changing the wind turbine's angle against the wind and optimizing the wake deflection and overall power generation. For example, the wind farm controller generates the yaw control command for a wind turbine based on the yaw action output by the decision model of that turbine. This command is sent to the PLC controller of that turbine through a fiber optic ring network. The PLC controller then controls the yaw motor to drive the nacelle to yaw by the corresponding angle.

[0049] The wind farm group collaborative yaw wake control method provided in this embodiment acquires historical local observation data of each wind turbine and constructs a global state of the wind farm. Then, it uses the global state of the wind farm and the historical local observation data of each wind turbine to jointly train the decision models corresponding to each wind turbine. This allows each wind turbine's decision model to learn the state information of all wind turbines globally and the aerodynamic coupling relationships between them during the training phase. After joint training, the decision models no longer make decisions in isolation but incorporate global collaborative information during the training process. Therefore, after training, when each wind turbine can output yaw actions based solely on its current local observation data, the decision of each wind turbine is actually made under the guidance of the acquired global collaborative knowledge, thus achieving a unity between global collaborative optimization in the training phase and independent and rapid decision-making in the execution phase. By involving global information in the training, the yaw actions of each wind turbine are collaboratively optimized in the direction of increasing the total power generation of the entire wind farm, effectively improving the overall power generation efficiency of large-scale wind farms.

[0050] This embodiment provides a method for coordinated yaw wake control of wind farm clusters, which can be used in the wind farm controller in the aforementioned centralized control center. Figure 2 This is a flowchart of a wind farm group coordinated yaw wake control method according to an embodiment of this application, such as... Figure 2 As shown, the process includes the following steps: Step S201: Obtain historical local observation data for each wind turbine, and construct the global state of the wind farm based on the historical local observation data for each wind turbine.

[0051] Specifically, step S201 above also includes: Step S2011: Perform outlier cleaning on the initial data. Outlier cleaning includes: removing unreasonable data points based on the theoretical power curves of each wind turbine, and using a Hamper filter to filter out abrupt spikes in the time series data points.

[0052] Specifically, the theoretical power curve refers to a reference curve provided by the wind turbine manufacturer or calibrated based on field measurement data, describing the theoretical output power of the wind turbine at different wind speeds. In actual operation, when the actual output power of the wind turbine deviates significantly from the theoretical power curve due to power outages, shutdowns, or sensor malfunctions, the corresponding data points are considered unreasonable. Using such data in model training will reduce the adaptability of the decision model to normal operating conditions, and therefore must be removed. The Hamper filter is an outlier detection method based on the median and absolute median difference. For each data point in the time series, the Hamper filter takes a fixed-length time window centered on that point and calculates the median and absolute median difference of the data within the window. If the deviation of the data point from the median exceeds a preset multiple of the absolute median difference, the point is determined to be a sudden spike in data and is filtered out. Compared to traditional mean filtering, this filter can more effectively handle non-Gaussian noise and sudden outliers.

[0053] Step S2012: Perform cubic spline interpolation on the initial data with different sampling frequencies to align the initial data to the same time axis.

[0054] Specifically, the sampling frequencies of various data sources in a wind farm may differ. For example, SCADA systems typically collect data at a frequency of seconds, while meteorological tower data is usually collected at a frequency of minutes. Data from different sampling frequencies cannot be directly fused. Cubic spline interpolation is a method that uses a cubic polynomial to interpolate between adjacent data points. It ensures that the interpolation curve has continuous first and second derivatives at the data points, thus generating smooth and highly accurate interpolation results. By using cubic spline interpolation, initial data from different sampling frequencies are uniformly interpolated onto the same time axis, ensuring that all state variables of each wind turbine have corresponding data values ​​at the same time. This provides a time-aligned data foundation for subsequent normalization processing and model training.

[0055] Step S2013: Normalize the aligned data and map each state variable to the [0,1] interval.

[0056] Specifically, normalization refers to scaling state variables of different dimensions and orders of magnitude to the same numerical range through a linear transformation. In this embodiment, for a certain state variable x, the normalization formula is:

[0057] Where x min and x max These are the minimum and maximum values ​​of the state variable in the training dataset, respectively. Normalization can eliminate the numerical scale differences between state variables caused by different units (such as wind speed in m / s, power in kW, and angle in degrees), and prevent certain features with larger values ​​from dominating the gradient update direction during neural network training, thereby accelerating the convergence speed of the neural network and improving the stability of model training.

[0058] Step S202 involves jointly training the decision-making models for each wind turbine using the global state of the wind farm and historical local observation data for each turbine. For details, please refer to [link to relevant documentation]. Figure 1 Step S102 of the illustrated embodiment will not be described again here.

[0059] Step S203: After training is complete, acquire the current local observation data of each wind turbine and input it into the corresponding decision model for each wind turbine to obtain the yaw action output by each decision model. For details, please refer to [link to relevant documentation]. Figure 1 Step S103 of the illustrated embodiment will not be described again here.

[0060] Step S204: Generate yaw control commands for each wind turbine based on each yaw action and send them to the corresponding wind turbine. For details, please refer to [link to relevant documentation]. Figure 1 Step S104 of the illustrated embodiment will not be described again here.

[0061] In this embodiment, by performing outlier cleaning, cubic spline interpolation alignment, and normalization on the initial data, invalid data generated by abnormal operating conditions such as sensor failure, power limit, and shutdown can be effectively removed. This solves the problem that multi-source data cannot be directly fused due to different sampling frequencies, and unifies each state variable to the same numerical scale, thereby providing higher quality, more uniform format, and faster convergence input data for the joint training of subsequent decision models.

[0062] This embodiment provides a method for coordinated yaw wake control of wind farm clusters, which can be used in the wind farm controller in the aforementioned centralized control center. Figure 3 This is a flowchart of a wind farm group coordinated yaw wake control method according to an embodiment of this application, such as... Figure 3 As shown, the process includes the following steps: Step S301: Obtain historical local observation data for each wind turbine, and construct the global state of the wind farm based on the historical local observation data for each wind turbine. For details, please refer to [link to relevant documentation]. Figure 1 Step S101 of the illustrated embodiment will not be described again here.

[0063] Step S302: Jointly train the decision-making model corresponding to each wind turbine using the global state of the wind farm and the historical local observation data of each wind turbine.

[0064] Specifically, step S302 above also includes: Step S3021: Construct the strategy network and value network corresponding to each wind turbine. The strategy network is used to output yaw actions based on historical local observation data, and the value network is used to evaluate the yaw actions output by the strategy network based on the global state of the wind farm.

[0065] Specifically, the policy network is a deep neural network whose input is historical local observation data of a single wind turbine, and whose output is the corresponding yaw action (i.e., the adjustment amount of the yaw angle) for that turbine. The policy network can adopt a multilayer perceptron architecture, including an input layer, several hidden layers, and an output layer. Nonlinear transformations can be introduced between the hidden layers using activation functions such as ReLU. Each wind turbine has its own corresponding policy network, but because the tasks of each turbine are similar, the network structures of all policy networks are the same, only the network parameters are independent. The role of the policy network is to map the local observation data of the wind turbine to specific yaw actions.

[0066] The value network is also a deep neural network. Its input is the global state of the wind farm, and its output is a scalar value used to evaluate the quality of the output actions of each strategy network under the current global state of the wind farm. During training, the value network learns to accurately evaluate the output actions of the strategy networks by fitting cumulative reward values, and feeds this evaluation result back to the strategy networks to guide their parameter updates. The role of the value network is to provide a unified evaluation standard based on the global state for the parameter updates of each strategy network, thereby enabling the update direction of each strategy network to collaboratively optimize towards the goal of increasing the total power generation of the entire wind farm.

[0067] The above training architecture is a centralized training and decentralized execution architecture. In the centralized training phase, the value network is updated using the global state of the wind farm. In the decentralized execution phase, each policy network outputs yaw action only based on the local observation data of the corresponding wind turbine.

[0068] Step S3022: Set the reward function, which includes a first term that is positively correlated with the total power generation of the wind farm, a second term that is negatively correlated with the load of each wind turbine, and a third term that is negatively correlated with the amplitude of each yaw action.

[0069] Specifically, the reward function is a quantitative function used to evaluate the quality of the yaw action output by the policy network. Its reward value is directly used to update the network parameters of the value network and the policy network during training. In the reinforcement learning framework, the setting of the reward function determines the agent's optimization objective and is a core factor guiding the convergence direction of the policy network.

[0070] The first term of the reward function is positively correlated with the total power generation of the wind farm, guiding the strategy network of each wind turbine to optimize towards increasing the total power generation of the entire farm. The total power generation is the sum of the active power of each wind turbine; a larger term indicates that the current yaw control strategy of each wind turbine is more beneficial to the total power generation of the entire farm, resulting in a larger reward value. The first term is the difference between the total power generation of the wind farm and the baseline power generation, divided by the baseline power generation. The baseline power generation is the estimated total power generation of the wind farm when all wind turbines are operating with the wind facing the wind. The baseline power generation refers to the estimated total power generation of the wind farm under the same inflow wind conditions when all wind turbines are set to 0° wind facing (i.e., the traditional greedy control state). Dividing the difference between the total power generation and the baseline power generation by the baseline power generation yields a normalized relative power gain value. Using a relative gain instead of an absolute power value ensures that the scale of the reward value remains stable under different wind conditions, avoiding large fluctuations in the reward value scale due to wind speed differences, thus facilitating stable training of the neural network.

[0071] The second term of the reward function is negatively correlated with the load of each wind turbine, and is used to constrain the policy network of each wind turbine to avoid yaw actions that would cause the wind turbine to bear excessive fatigue loads. The second term is determined based on the sum of the squares of the equivalent bending moments of each wind turbine, that is, the equivalent bending moment value of each wind turbine is divided by a preset bending moment threshold, the square is taken, and then the squares of the values ​​of each wind turbine are summed.

[0072] In this embodiment, the complete form of the equivalent bending moment square term penalty is:

[0073] Where ω2 is the load penalty weighting coefficient, M t i Let M be the equivalent bending moment of the i-th wind turbine at time t. limit A preset bending moment threshold is used. This design ensures that when the equivalent bending moment of any wind turbine exceeds the preset threshold, its penalty term increases rapidly by a squared multiple. This penalty effectively suppresses yaw movements that could lead to severe overload of a single wind turbine during model training.

[0074] The use of quadratic terms instead of linear terms is because quadratic functions have an amplification effect when the variable value is greater than 1. That is, when the equivalent bending moment of any wind turbine exceeds the preset bending moment threshold, the penalty term for that wind turbine will increase rapidly by a square factor. This design allows the policy network to prioritize avoiding yaw actions that would cause severe overload of any wind turbine during training, thereby effectively suppressing the occurrence of extreme load conditions.

[0075] The third term of the reward function is negatively correlated with the amplitude of each yaw action. This is used to suppress the policy network from outputting frequent or excessive yaw actions, thereby preventing the yaw actuator from wearing out due to repeated actions.

[0076] The third item |a t i | is a penalty for yaw action amplitude, which extends the maintenance cycle of the yaw drive system by suppressing ineffective and frequent yaws.

[0077] The reward function also includes a dynamic weight adjustment coefficient, which is determined based on the current turbulence intensity of the wind farm and is used to increase the weight of the second term when the turbulence intensity increases. The dynamic weight adjustment coefficient is a coefficient that automatically adjusts with changes in the current turbulence intensity of the wind farm, and it is multiplied by the second term of the reward function. When the turbulence intensity is low (e.g., in stable wind conditions), the dynamic weight adjustment coefficient is small, and the reward function is mainly dominated by the first term, with the strategy network prioritizing the increase of overall power generation. When the turbulence intensity increases, the dynamic weight adjustment coefficient increases accordingly, and the weight of the second term in the reward function increases, with the strategy network paying more attention to avoiding excessive fatigue loads on each wind turbine.

[0078] In this embodiment, the complete form of the composite reward function is:

[0079] Where, r t Let be the joint reward value at time t. Let P be the total power generation of the wind farm at time t. base For the estimated power across the entire field under baseline control (greedy yaw against the wind), M t i Let M be the equivalent bending moment of the i-th wind turbine at time t. limit To preset the bending moment threshold, a t i Let ω1, ω2, and ω3 be the yaw action output by the i-th wind turbine at time t, where ω1, ω2, and ω3 are multi-objective weighted hyperparameters. The first term is the normalized relative gain between the total power generation and the reference power generation, the second term is the squared penalty for overload, and the third term is the penalty for yaw action amplitude.

[0080] Step S3023: Calculate the reward value of the yaw action according to the reward function, and update the network parameters of the value network and policy network based on the reward value.

[0081] Specifically, during training, after each time the policy network outputs a yaw action based on historical local observation data, this yaw action is applied to the simulation environment, causing a state transition and returning a reward value. This reward value is calculated using the reward function in step S3022, comprehensively reflecting the combined impact of the yaw action on the total power generation, the load on each wind turbine, and the wear of the yaw actuator. The value network fits an evaluation of the yaw action based on this reward value, while the policy network adjusts its parameters using a policy gradient algorithm based on the reward value, causing the yaw action output by the policy network to move towards the optimization direction that yields a larger reward value. In this embodiment, the policy network uses a truncation mechanism of the near-end policy optimization algorithm for parameter updates to limit the magnitude of a single update and avoid training instability caused by drastic changes in the policy during training.

[0082] Specifically, during the update process, the advantage of the current action relative to the average level is first assessed using generalized advantage estimation (GAE):

[0083] in, As a discount factor, For smoothing parameters, δ t For timing difference error: , The policy network update employs a truncation mechanism from the proximal policy optimization algorithm, with the following loss function:

[0084] in, For importance sampling ratio, The truncation threshold (e.g., 0.2) is used to limit the magnitude of a single policy update to maintain training stability.

[0085] The value network is updated using the mean squared error loss function:

[0086] in, G t To achieve the target value in multiple steps, V ψ ( s t ) represents the value network for the current global state s of the wind farm. t The evaluation value is obtained by minimizing the above loss function. The value network gradually converges to an accurate evaluation of the policy network's output action.

[0087] Step S3024: Based on the spatial attention mechanism, the historical local observation data of each wind turbine are weighted and fused, and the weighted and fused features are used as the input of the value network. The spatial attention mechanism is used to determine the attention weight of each wind turbine according to the wake coupling relationship between each wind turbine, and to dynamically adjust the attention weight of each wind turbine according to the real-time wind direction.

[0088] Specifically, the spatial attention mechanism refers to a mechanism that can automatically learn and allocate attention weights in a spatial dimension. Its core idea is to allow the value network to selectively focus on the turbines most severely affected by the wake under the current wind conditions when evaluating the yaw behavior of each wind turbine, rather than treating the state of all turbines equally. In this embodiment, based on the spatial positional relationship between each wind turbine (i.e., the latitude and longitude coordinates or relative coordinates of each turbine) and the current real-time wind direction, the upstream and downstream relationships and the degree of wake influence between each turbine can be determined. For example, when the incoming wind direction is northerly, a turbine located on the south side is within the wake region of a turbine on the north side, and thus the south-side turbine is affected by the wake of the north-side turbine. When the wind direction changes to southerly, the upstream and downstream relationships and the affected turbines are reversed. The spatial attention mechanism assigns attention weights to each wind turbine according to the above wake coupling relationship, with turbines more severely affected by the wake (i.e., those more severely affected downstream) receiving higher attention weights. A higher attention weight indicates a greater contribution of the wind turbine's state information to the evaluation result when the value network evaluates the policy network's output action. Then, the historical local observation data of each wind turbine are multiplied by their respective attention weights and fused together, with the fused feature vector serving as the input to the value network. When the real-time wind direction changes, this spatial attention mechanism can recalculate the wake coupling relationship between wind turbines, automatically switching higher attention weights from the previously affected turbines to the newly affected ones, ensuring that the value network's evaluation focus remains on the turbine most severely affected by the wake.

[0089] In some optional implementations, the above spatial attention mechanism adopts a multi-head attention architecture, which includes multiple parallel attention heads. Each attention head independently calculates a set of attention weights, and then the outputs of each attention head are concatenated or weighted averaged to obtain the final output, so as to improve the expressive power of the attention mechanism for different feature subspaces.

[0090] Step S3025: Construct a virtual environment using a wake simulation model, and pre-train the decision model in the virtual environment.

[0091] Specifically, a wake simulation model refers to a numerical calculation model used to simulate the propagation and superposition effects of wind turbine wakes within a wind farm. It can employ either the FLORIS analytical model or the Large Eddy Simulation (LES) model. Constructing a virtual environment using a wake simulation model involves building a wind farm simulation platform within simulation software that is aerodynamically equivalent to a real wind farm, based on the actual wind turbine layout, terrain data, and historical wind condition data. Within this virtual environment, the decision-making models for each wind turbine can perform the top 10 calculations without safety risks and at low cost. 5 Through repeated training iterations, the system quickly learns basic physical laws such as wind speed loss and wake deflection, accumulating a wealth of experience in coordinated yaw control.

[0092] Step S3026: Transfer the network parameters of the pre-trained decision model to the actual wind farm and fine-tune the decision model using real-time data collected on-site.

[0093] Specifically, migration refers to directly loading the network parameters of each decision model obtained after training in the virtual environment into the actual wind farm controller as the initial model for actual deployment. Fine-tuning refers to, in the actual deployment phase after migration, using real-time operational data collected on-site by the wind farm SCADA system, continuing to train and iterate a small number of times on the pre-trained model with a small learning rate, so that the model parameters gradually adapt to the unique wind conditions and turbine characteristics of the actual wind farm, thereby eliminating the differences between the simulation environment and the physical entity.

[0094] Step S303: After training is complete, acquire the current local observation data of each wind turbine and input it into the corresponding decision model for each wind turbine, then obtain the yaw action output by each decision model. For details, please refer to [link to relevant documentation]. Figure 1 Step S103 of the illustrated embodiment will not be described again here.

[0095] Step S304: Generate yaw control commands for each wind turbine based on each yaw action and send them to the corresponding wind turbine. For details, please refer to [link to relevant documentation]. Figure 1 Step S104 of the illustrated embodiment will not be described again here.

[0096] In this embodiment, a strategy network and a value network are constructed, and the value network is used to uniformly evaluate the yaw actions of each wind turbine based on the global state of the wind farm. This allows the strategy networks of each wind turbine to obtain feedback signals based on the overall field state when updating parameters, thereby achieving collaborative optimization of the decision models of each wind turbine. By setting a reward function that includes a positive correlation between overall power generation, a negative correlation between the load of each wind turbine, and a negative correlation between the amplitude of each yaw action, the three optimization objectives of increasing overall power generation, reducing wind turbine fatigue load, and suppressing frequent yaw are unified in the same reinforcement learning framework. Furthermore, by using the equivalent bending moment square term penalty and a dynamic weight adjustment coefficient based on turbulence intensity, a dynamic balance between power generation revenue and unit fatigue life is achieved. By introducing a spatial attention mechanism at the input of the value network, the value network can automatically identify downstream wake-affected wind turbines based on the spatial positional relationship between wind turbines and real-time wind direction, and assign them higher attention weights, thereby improving the physical perception capability and adaptability to dynamic wind conditions of the decision model. Furthermore, by using a wake simulation model to construct a virtual environment for pre-training, the decision-making model can quickly learn the basic physical laws of wind speed loss in a safe and low-cost environment. Then, by transferring the pre-trained network parameters to the actual wind farm and fine-tuning them using real-time data from the field, the difference between the simulation environment and the actual wind farm is eliminated, significantly improving the model's adaptability and control performance in actual deployment.

[0097] This embodiment provides a method for coordinated yaw wake control of wind farm clusters, which can be used in the wind farm controller in the aforementioned centralized control center. Figure 4 This is a flowchart of a wind farm group coordinated yaw wake control method according to an embodiment of this application, such as... Figure 4 As shown, the process includes the following steps: Step S401: Obtain historical local observation data for each wind turbine, and construct the global state of the wind farm based on the historical local observation data for each wind turbine. For details, please refer to [link to relevant documentation]. Figure 1 Step S101 of the illustrated embodiment will not be described again here.

[0098] Step S402 involves jointly training the decision-making models for each wind turbine using the global state of the wind farm and historical local observation data for each turbine. For details, please refer to [link to relevant documentation]. Figure 1 Step S102 of the illustrated embodiment will not be described again here.

[0099] Step S403: After training is complete, acquire the current local observation data of each wind turbine and input it into the corresponding decision model for each wind turbine, then obtain the yaw action output by each decision model. For details, please refer to [link to relevant documentation]. Figure 1 Step S103 of the illustrated embodiment will not be described again here.

[0100] Step S404: Generate yaw control commands for each wind turbine based on each yaw action and send them to the corresponding wind turbine.

[0101] Specifically, step S404 above also includes: Step S4041: Perform first-order low-pass filtering on each yaw action, generate yaw control commands based on the filtered yaw actions; and perform amplitude limiting on each yaw action, the amplitude limiting process including limiting the adjustment amount of each yaw action to a preset maximum yaw step size range.

[0102] Specifically, first-order low-pass filtering refers to smoothing the yaw action sequence output by the policy network, and its basic form is as follows:

[0103] Where γ cmd,t γ is the actual yaw angle command issued at time t. cmd,t-1 The yaw angle command issued at the previous moment, γ current Given the current yaw angle of the wind turbine, a tThe yaw action (i.e., yaw angle adjustment) output by the strategy network is represented by τ, which is a smoothing coefficient ranging from (0,1). This filtering process effectively suppresses abrupt changes in the output commands of the strategy network, preventing frequent starts and stops of the yaw motor or excessive instantaneous impact, thus extending the service life of the yaw actuator. The limiting process refers to truncating the yaw action adjustment output by the strategy network to the preset maximum yaw step size when it exceeds this size, preventing impact damage to the yaw motor and transmission structure caused by excessively large single yaw actions.

[0104] In this embodiment, the hard constraint of the yaw action can be expressed as:

[0105] Where a t i Let Δγ be the yaw action of the i-th wind turbine at time t. max The maximum yaw step size is preset. Through the above limiting process, it is ensured that the adjustment amount of each yaw action is always within the safe range allowed by the equipment.

[0106] Step S4042: If the cumulative yaw angle of any wind turbine exceeds a preset angle threshold, a cable release control command is generated.

[0107] Specifically, the cumulative yaw angle refers to the total angle of yaw in the same direction accumulated from the initial position of the wind turbine. During wind turbine operation, due to continuous changes in wind direction, the turbine may need to perform frequent yaw movements. Over time, this accumulation can lead to damage to the internal cables of the turbine due to excessive tangling. The preset angle threshold is a pre-set safe angle value (e.g., ±720°). When the cumulative yaw angle of any wind turbine exceeds this threshold, the system automatically generates a cable unwinding control command, triggering mechanical unwinding logic to cause the wind turbine to yaw in the opposite direction to release the tangled cable and restore the cable to a safe tangling angle range.

[0108] Step S4043: If the temperature of the yaw motor of any wind turbine exceeds the preset temperature threshold, the yaw control command of the corresponding wind turbine is forcibly set to the wind-fighting mode command.

[0109] Specifically, the yaw motor temperature refers to the real-time operating temperature of the yaw drive motor, which is monitored in real time by temperature sensors inside or on the surface of the motor. When the yaw motor temperature exceeds the preset temperature threshold, it indicates that the motor may be operating under abnormal conditions such as overload or poor heat dissipation. If the yaw action is continued to be executed according to the instructions of the strategy network, the motor may burn out. In this case, the system automatically forces the yaw control command of the corresponding fan to the windward mode command, so that the fan returns to the 0° windward state, reduces the load on the yaw motor, and resumes the yaw control of the strategy network after the motor temperature drops to a safe range.

[0110] In this embodiment, by setting up a physical safety filter between the yaw action output by the strategy network and the actual yaw control command issued to the wind turbine, the yaw action is subjected to first-order low-pass filtering and amplitude limiting processing. This smooths out abrupt changes in the command output by the strategy network and limits excessive amplitude commands, reducing the instantaneous impact on the yaw motor. When the cumulative yaw angle of the wind turbine exceeds a preset angle threshold, the system automatically generates a cable unwinding control command to trigger mechanical cable unwinding logic, keeping the cable within a safe operating angle range. When the yaw motor temperature exceeds a preset temperature threshold, the system automatically switches the control command for the corresponding wind turbine to wind-facing mode, ensuring the motor operates within a safe temperature range. Each safety barrier functions independently, collectively constraining the yaw command output by the AI ​​decision model within the safe operating boundaries of the equipment.

[0111] This embodiment also provides a wind farm group coordinated yaw wake control device, which is used to implement the above embodiments and preferred embodiments, and will not be repeated as already described. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0112] This embodiment provides a wind farm group coordinated yaw wake control device, such as Figure 5 As shown, it includes: The acquisition module 501 is used to acquire historical local observation data of each wind turbine and construct the global state of the wind farm based on the historical local observation data of each wind turbine. The historical local observation data includes at least the nacelle side wind speed, wind direction, turbulence intensity, current yaw error angle, active power and equivalent bending moment. Training module 502 is used to jointly train the decision models corresponding to each wind turbine using the global state of the wind farm and the historical local observation data of each wind turbine. The execution module 503 is used to obtain the current local observation data of each wind turbine after training is completed and input it into the decision model of each wind turbine to obtain the yaw action output by each decision model. The generation module 504 is used to generate yaw control commands for each wind turbine based on each yaw action; The sending module 505 is used to send each yaw control command to the corresponding wind turbine.

[0113] In some optional implementations, the acquisition module 501 further includes: The first processing unit is used to clean outliers from the initial data. Outlier cleaning includes: removing unreasonable data points based on the theoretical power curves of each wind turbine, and using a Hamper filter to filter out abrupt spikes in the time series data. The second processing unit is used to perform cubic spline interpolation on the initial data with different sampling frequencies to align the initial data to the same time axis. The third processing unit is used to normalize the aligned data and map each state variable to the [0,1] interval.

[0114] In some alternative implementations, the training module 502 further includes: The network construction unit is used to construct the strategy network and value network corresponding to each wind turbine. The strategy network is used to output yaw actions based on historical local observation data, and the value network is used to evaluate the yaw actions output by the strategy network based on the global state of the wind farm. The reward setting unit is used to set the reward function, which includes a first term that is positively correlated with the total power generation of the wind farm, a second term that is negatively correlated with the load of each wind turbine, and a third term that is negatively correlated with the amplitude of each yaw action. The parameter update unit is used to calculate the reward value of the yaw action according to the reward function, and update the network parameters of the value network and policy network based on the reward value; The attention fusion unit is used to perform weighted fusion of historical local observation data of each wind turbine based on the spatial attention mechanism, and uses the weighted fused features as input to the value network. The spatial attention mechanism is used to determine the attention weight of each wind turbine according to the wake coupling relationship between each wind turbine, and dynamically adjust the attention weight of each wind turbine according to the real-time wind direction. The pre-training unit is used to construct a virtual environment using the wake simulation model and pre-train the decision model in the virtual environment; The fine-tuning unit is used to transfer the network parameters of the pre-trained decision model to the actual wind farm and fine-tune the decision model using real-time data collected on-site.

[0115] In some optional implementations, in the reward setting unit, the first term is the difference between the total power generation of the wind farm and the benchmark power generation divided by the benchmark power generation, where the benchmark power generation is the estimated total power generation of the wind farm when each wind turbine is running against the wind; the second term is determined based on the sum of the squares of the equivalent bending moments of each wind turbine; the reward function also includes a dynamic weight adjustment coefficient, which is determined based on the current turbulence intensity of the wind farm and is used to increase the weight of the second term when the turbulence intensity increases.

[0116] In some alternative implementations, the generation module 504 further includes: The filtering and limiting unit is used to perform first-order low-pass filtering and limiting processing on each yaw action, and generate yaw control commands based on the processed yaw actions; the limiting processing includes limiting the adjustment amount of each yaw action to a preset maximum yaw step size range. The unwind control unit is used to generate an unwind control command when the cumulative yaw angle of any wind turbine exceeds a preset angle threshold. The temperature protection unit is used to force the yaw control command of the corresponding fan to be set to the wind-fighting mode command when the temperature of the yaw motor of any fan exceeds the preset temperature threshold.

[0117] The wind farm group coordinated yaw wake control device provided in this application can execute the wind farm group coordinated yaw wake control method provided in any embodiment of this application, and has the corresponding functional modules and beneficial effects of the method. Further functional descriptions of the above modules and units are the same as those in the corresponding embodiments above, and will not be repeated here.

[0118] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0119] The following is a detailed reference. Figure 6 This diagram illustrates a suitable structural schematic for implementing the electronic device described in the embodiments of this application. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 601, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 602 or a program loaded from memory 608 into random access memory (RAM) 603. The RAM 603 also stores various programs and data required for the operation of the electronic device. The processor 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0120] Typically, the following devices can be connected to I / O interface 605: input devices 606 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 607 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 608 including, for example, magnetic tapes, hard disks, etc.; and communication devices 609. Communication device 609 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 6 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0121] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 609, or installed from a memory 608, or installed from a ROM 602. When the computer program is executed by the processor 601, it performs the functions defined in the wind farm group coordinated yaw wake control method of embodiments of this application.

[0122] Figure 6 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0123] This application also provides a computer-readable storage medium. The methods described in this application can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the wind farm group coordinated yaw wake control method shown in the above embodiments is implemented.

[0124] A portion of this application can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to this application through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0125] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and all such modifications and variations fall within the scope defined by the appended claims.

Claims

1. A method for coordinated yaw wake control of a wind farm cluster, characterized in that, include: Historical local observation data of each wind turbine is obtained, and the global state of the wind farm is constructed based on the historical local observation data of each wind turbine. The historical local observation data includes at least the nacelle side wind speed, wind direction, turbulence intensity, current yaw error angle, active power and equivalent bending moment. The decision-making models corresponding to each wind turbine are jointly trained using the global state of the wind farm and the historical local observation data of each wind turbine. After training is completed, the current local observation data of each wind turbine is obtained and input into the decision model of each wind turbine, and the yaw action output by each decision model is obtained. Based on each yaw action, a yaw control command is generated for each wind turbine and sent to the corresponding wind turbine.

2. The method according to claim 1, characterized in that, Before acquiring the historical local observation data of each wind turbine, the following steps are also included: Outlier cleaning is performed on the initial data. The outlier cleaning includes: removing unreasonable data points based on the theoretical power curves of each wind turbine, and filtering out abrupt spike data points in the time series using filters. Cubic spline interpolation is performed on the initial data with different sampling frequencies to align the initial data to the same time axis; The aligned data is normalized, and each state variable is mapped to the [0,1] interval.

3. The method according to claim 1, characterized in that, The step of using the global state of the wind farm and historical local observation data of each wind turbine to perform correlation training on the decision model corresponding to each wind turbine includes: Construct a strategy network and a value network for each wind turbine. The strategy network is used to output yaw action based on the historical local observation data, and the value network is used to evaluate the yaw action output by the strategy network based on the global state of the wind farm. The network parameters of the policy network are updated based on the evaluation results of the value network.

4. The method according to claim 3, characterized in that, The step of using the global state of the wind farm and historical local observation data of each wind turbine to perform correlation training on the decision model corresponding to each wind turbine also includes: A reward function is set, which includes a first term that is positively correlated with the total power generation of the wind farm, a second term that is negatively correlated with the load of each wind turbine, and a third term that is negatively correlated with the amplitude of each yaw action. The reward value for the yaw action is calculated based on the reward function, and the network parameters of the value network and the policy network are updated based on the reward value.

5. The method according to claim 4, characterized in that, The first item is the difference between the total power generation of the wind farm and the reference power generation divided by the reference power generation, where the reference power generation is the estimated power generation of the entire wind farm when each wind turbine is running in the wind. The second term is determined based on the sum of the squares of the equivalent bending moments of each of the aforementioned wind turbines; The reward function also includes a dynamic weight adjustment coefficient, which is determined based on the current turbulence intensity of the wind farm and is used to increase the weight of the second term when the turbulence intensity increases.

6. The method according to claim 3, characterized in that, The step of using the global state of the wind farm and historical local observation data of each wind turbine to perform correlation training on the decision model corresponding to each wind turbine also includes: The historical local observation data of each wind turbine are weighted and fused based on the spatial attention mechanism, and the weighted and fused features are used as the input of the value network. The spatial attention mechanism is used to determine the attention weight of each wind turbine according to the wake coupling relationship between each wind turbine, and to dynamically adjust the attention weight of each wind turbine according to the real-time wind direction.

7. The method according to claim 1, characterized in that, The step of using the global state of the wind farm and historical local observation data of each wind turbine to perform correlation training on the decision model corresponding to each wind turbine also includes: A virtual environment is constructed using a wake simulation model, and the decision model is pre-trained in the virtual environment; The network parameters of the pre-trained decision model are transferred to an actual wind farm, and the decision model is fine-tuned using real-time data collected on-site.

8. The method according to claim 1, characterized in that, The process of generating yaw control commands for each wind turbine based on each yaw action includes: Each of the yaw actions is subjected to a first-order low-pass filter, and the yaw control command is generated based on the filtered yaw actions; and each of the yaw actions is subjected to a limiting process, the limiting process including limiting the adjustment amount of each of the yaw actions to a preset maximum yaw step size range. If the cumulative yaw angle of any of the wind turbines exceeds a preset angle threshold, a cable release control command is generated. If the yaw motor temperature of any of the aforementioned wind turbines exceeds a preset temperature threshold, the yaw control command for the corresponding wind turbine will be forcibly set to the wind-fighting mode command.

9. A wind farm cluster coordinated yaw wake control system, characterized in that, include: The acquisition module is used to acquire historical local observation data of each wind turbine and construct the global state of the wind farm based on the historical local observation data of each wind turbine. The historical local observation data includes at least the nacelle side wind speed, wind direction, turbulence intensity, current yaw error angle, active power and equivalent bending moment. The training module is used to jointly train the decision models corresponding to each wind turbine using the global state of the wind farm and the historical local observation data of each wind turbine. The execution module is used to acquire the current local observation data of each wind turbine and input it into the decision model corresponding to each wind turbine, and acquire the yaw action output by each decision model; The generation module is used to generate yaw control commands for each of the wind turbines based on each of the yaw actions. The sending module is used to send the yaw control commands to the corresponding wind turbines.

10. An electronic device, characterized in that, include: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the wind farm group coordinated yaw wake control method according to any one of claims 1 to 9.