Wind power plant real-time yaw cooperative control system and method fused with machine learning
By constructing a virtual lidar simulation environment and combining it with hybrid deep neural networks and measured data, along with multi-agent reinforcement learning, the real-time and accuracy problems of wind farm wake control under complex terrain were solved, achieving efficient collaborative control of wind farms and maximizing power generation benefits.
Patent Information
- Application Number
- CN202511934832.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-21
- Publication Date
- 2026-04-07
AI Technical Summary
Existing engineering wake models cannot accurately describe the flow field distortion caused by complex terrain, while high-precision CFD simulations are computationally intensive and cannot meet the requirements of real-time control. Data-driven machine learning methods have poor generalization ability under extreme wind conditions and are difficult to achieve efficient collaborative control of wind farms.
A simulation environment for virtual lidar is constructed to generate a prior knowledge base. Wake prediction is performed using a hybrid deep neural network combined with measured data. A collaborative control strategy is generated through multi-agent reinforcement learning to achieve real-time yaw control of the wind farm.
While ensuring real-time performance, the accuracy and robustness of wake control are significantly improved, maximizing the overall power generation benefits of the wind farm, reducing equipment fatigue, and increasing the power generation of downstream units.
Smart Images
Figure CN121806456A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wind power generation control technology, and in particular to a real-time yaw collaborative control system and method for wind farms that integrates machine learning. Background Technology
[0002] As wind power development extends into complex terrains (such as mountains and hills), the wake effect problem of large wind farms is becoming increasingly prominent. During the operation of upstream wind turbines, a wake region with reduced wind speed and increased turbulence intensity forms behind them. This not only leads to a significant decrease in power generation from downstream turbines but also increases the fatigue load on downstream units, shortening equipment lifespan. To alleviate this problem, "active yaw coordinated control," which involves adjusting the yaw angle of upstream units to deflect the wake and thus avoid the downstream units, has become a hot research topic in the industry.
[0003] However, implementing efficient real-time collaborative control in real-world wind farms with complex terrain faces significant technical challenges. On one hand, existing engineering wake models (such as the Jensen or Gaussian models) are overly simplified and cannot accurately describe the flow field distortions caused by terrain undulations. While high-precision computational fluid dynamics (CFD) simulations are accurate, their computational demands are extremely high, failing to meet the millisecond-level real-time control requirements. On the other hand, while purely data-driven machine learning methods are computationally fast, they heavily rely on the quality and coverage of the training data. In actual wind farm operations, wake interaction data under extreme wind conditions or specific wind directions are extremely difficult to capture, resulting in sparse samples. This leads to poor generalization ability of models trained directly using SCADA data, making it difficult to accurately predict and control wakes in complex and variable real-world conditions, and ultimately failing to maximize the overall benefits of the wind farm. Summary of the Invention
[0004] The main objective of this invention is to provide a real-time yaw collaborative control system and method for wind farms that integrates machine learning, in order to solve the problems raised in related technologies.
[0005] To achieve the above objectives, according to one aspect of the present invention, a real-time yaw collaborative control system and method for wind farms integrating machine learning are provided, comprising the following steps:
[0006] S1. Construct a prior knowledge base: acquire wind farm topographic data, construct a flow field simulation model using computational fluid dynamics (CFD), simulate the aerodynamic characteristics of the wind turbine using an actuator line model; set up a virtual lidar in the simulation environment to simulate the scanning method of a physical lidar, and generate a batch of simulation datasets containing environmental parameters, unit status, and virtual lidar wind measurement data.
[0007] S2. Construct and train a virtual-real fusion prediction model: Establish a hybrid deep neural network, which includes a main prediction network and a residual correction network; use the simulation dataset to pre-train the main prediction network offline to master the physical laws of the terrain flow field; use the measured SCADA data and physical lidar data of the wind farm to train the residual correction network online, and fine-tune the main prediction network to output the predicted inflow wind speed of the downstream unit.
[0008] S3. Generate a cooperative control strategy: Establish a multi-agent reinforcement learning model, treating each wind turbine as an agent, and using the virtual-real fusion prediction model as the interaction environment; design a reward function that includes single-unit power, total power of downstream units, and yaw action penalty, and train the agents to learn a cooperative yaw strategy.
[0009] S4. Closed-loop real-time control: Real-time acquisition of wind farm operation data, input of a trained multi-agent reinforcement learning model, generation of yaw angle control commands for all units in the field and issuance for execution.
[0010] Further, in step S1, the line-of-sight velocity of the virtual lidar The calculation formula is:
[0011]
[0012] in, The velocity vector in the Cartesian coordinate system calculated by CFD. For the half-angle of the conical scan, The scanning azimuth angle is used; the data generation process of the virtual lidar includes volume weighting of the velocity of the grid nodes around the sampling point on the beam path.
[0013] Further, in step S2, the output of the virtual-real fusion prediction model is: ;
[0014] in, The output of the main prediction network is virtual lidar data and terrain slope factor. The output of the residual correction network is the measured SCADA data and physical lidar data; the loss function of the online training is based on the actual inflow wind speed inferred from the downstream wind turbine SCADA data. The difference is constructed.
[0015] Further, in step S3, the reward function The expression is:
[0016]
[0017] in, This represents the current unit power. This is the sum of the power of downstream units affected by the wake. As a cooperating factor, Adjust the yaw angle increment. This is the penalty coefficient.
[0018] Furthermore, step S4 also includes a safety protection mechanism: when the communication interruption between the software system and the SCADA system exceeds a preset threshold, the cooperative control command is automatically revoked and the unit is restored to automatic wind-following mode; before issuing the command, the load proxy model is used to evaluate the predicted load, and if it exceeds the safety threshold, the yaw angle command is reduced.
[0019] On the other hand, the present invention also provides a real-time yaw cooperative control system for wind farms that integrates machine learning, characterized in that it includes:
[0020] The data acquisition and interaction module is used to connect the wind farm SCADA system and the nacelle-type lidar to collect unit operation data and wind measurement data;
[0021] The high-performance computing and processing module is used to run the computational fluid dynamics simulation engine to generate virtual lidar data and to run the deep learning framework to train and infer hybrid deep neural networks.
[0022] The collaborative control execution module is used to generate yaw control commands based on a multi-agent reinforcement learning strategy and write the commands to the wind turbine main control PLC through the OPCUA interface; the system is configured to execute the method described in any of the above-mentioned steps.
[0023] Secondly, a computer-readable storage medium is provided, on which a computer program is stored, characterized in that, when the program is executed by a processor, it implements the method as described in any one of claims 1 to 5.
[0024] Compared with the prior art, the present invention has the following beneficial effects:
[0025] This invention can effectively integrate high-precision simulation data with field measurement data, and significantly improve the accuracy and robustness of wake control under complex terrain while ensuring real-time control.
[0026] This invention constructs a simulation environment based on computational fluid dynamics and virtual lidar, generating a prior knowledge base rich in spatial flow field information. It then utilizes a hybrid deep neural network architecture to combine the physical laws learned from simulation with residual corrections learned from measured data. This approach effectively overcomes the shortcomings of purely physical models (long computation time) and purely data-driven models (poor generalization ability under sparse samples). The system possesses both CFD-level spatial flow field perception capabilities and the ability to correct terrain and environmental errors online using measured data, achieving millisecond-level high-precision wake prediction.
[0027] Furthermore, this invention employs a collaborative decision-making mechanism based on multi-agent reinforcement learning. By designing a reward function that incorporates overall power gain and unit load penalties, it overcomes the limitations of traditional single-unit control. The system can guide upstream units to proactively sacrifice some of their own power to deflect the wake, thereby significantly increasing the power output of downstream units. This maximizes the overall power generation efficiency of the wind farm while ensuring equipment safety. Attached Figure Description
[0028] Figure 1 This is a schematic diagram of the hardware architecture and data flow of the wind farm real-time yaw coordinated control system provided in an embodiment of the present invention.
[0029] Figure 2 The overall flowchart of the wind farm real-time yaw collaborative control method integrating machine learning provided in the embodiments of the present invention is shown.
[0030] Explanation of reference numerals in the attached diagram: 100, Data acquisition and interaction module; 200, High-performance computing and processing module; 300, Collaborative control and execution module. Detailed Implementation
[0031] It should be noted that, unless otherwise specified, the embodiments and features described in this embodiment can be combined with each other. The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0032] Example 1:
[0033] This embodiment provides a real-time yaw collaborative control system for wind farms that integrates machine learning. Physically, this system is typically deployed in the wind farm's central control server or edge computing nodes, aiming to address power generation losses and equipment fatigue caused by wake effects in large wind farms operating in complex terrain.
[0034] like Figure 1 As shown, the control system mainly includes:
[0035] 1. Data Acquisition and Interaction Module 100
[0036] This module is responsible for connecting the wind farm equipment in the physical world. Specifically, it communicates with the wind farm's existing SCADA system via OPCDA / UA or ModbusTCP protocol to acquire second-level operating data (including power, speed, pitch angle, nacelle position, etc.) of all turbines at a frequency of 1Hz. Simultaneously, it connects to the nacelle-mounted lidar installed on the turbines via TCP / IP protocol to acquire measured data of inflow wind speed and direction. The collected raw data undergoes moving average filtering (e.g., using a 10-minute time window) to eliminate instantaneous fluctuations caused by gusts, and is then stored in a time-series database such as MySQL or InfluxDB.
[0037] 2. High-performance computing and processing module 200
[0038] As the "brain" of the system, this module is built on a Linux operating system environment and is equipped with a multi-core CPU and a dedicated GPU accelerator card. Internally, it integrates the OpenFOAM fluid dynamics simulation engine and a Python-based deep learning framework. This module primarily performs virtual simulation calculations, deep neural network training and inference, and reinforcement learning policy generation.
[0039] 3. Collaborative Control Execution Module 300
[0040] This module is responsible for translating the calculated optimal control strategy into specific execution instructions. It connects to the main control PLC of each wind turbine generator via the OPCUA interface, writing the optimized yaw angle into the PLC's "yaw deviation setpoint" register. Simultaneously, this module incorporates a built-in safety protection mechanism, including a communication heartbeat monitoring unit and a load constraint limiting unit, to ensure the safety of the control instructions.
[0041] Example 2:
[0042] Based on the above system architecture, this embodiment elaborates on a real-time yaw collaborative control method for wind farms that integrates machine learning. This method deeply integrates physical simulation with artificial intelligence, and the entire implementation process is as follows: Figure 2 As shown, it includes the following four steps:
[0043] S1: Constructing a prior knowledge base for complex terrain flow fields based on virtual lidar
[0044] This step aims to address the lack of extreme wind condition samples in actual operation through high-fidelity simulation.
[0045] S1.1 Terrain and Mesh Processing: Obtain the digital elevation model (DEM) and surface roughness data of the target wind farm. Use OpenFOAM's snappyHexMesh tool to generate a body-fitted unstructured mesh. To accurately capture wake features, mesh refinement is performed at the wind turbine hub height and in the expected wake region. The background mesh size is set to 20m, while the refined area mesh size is less than 5m.
[0046] S1.2 Aerodynamic Modeling: Unlike the traditional Actuated Disk Model (ADM), this embodiment uses an actuation line model. The wind turbine blades are discretized into several airfoil sections, and lift and drag are calculated separately and coupled in reverse to the Navier-Stokes equations as volume force source terms. This method can accurately simulate the generation of tip vortices and the rotational effect of the wake, significantly improving the simulation accuracy in the near-wake region.
[0047] S1.3 Virtual LiDAR Scanning: This is the crucial link between simulation and reality. In the CFD computational domain, instead of directly extracting the full-field velocity, a post-processing program is written to simulate the conical scanning method of the physical lidar (such as PPI or RHI mode). Assuming the virtual lidar is installed on the top of the cabin, and the half-angle of the scanning cone is... (For example ), and set on the beam path A distance gate.
[0048] For any sampling point on the beam path Its line-of-sight velocity The calculation formula is as follows:
[0049]
[0050] in, The velocity vector of this point in the Cartesian coordinate system calculated by OpenFOAM. The beam unit direction vector, To scan the azimuth angle, considering the probe volume effect of the physical radar, the system performs a Gaussian weighted average of the velocities of the grid nodes around the sampling point, and outputs the final virtual measurement value.
[0051] S1.4 Data Generation: Set up a working condition matrix covering different inflow wind speeds (3-25m / s), wind directions (16 sectors), and turbulence intensities (5%-20%), run simulations in batches, and build a structured dataset containing {environmental parameters, upstream yaw angle, virtual LiDAR data, downstream effective wind speed, and total power}.
[0052] S2: Constructing a wake prediction model based on "virtual and real data fusion"
[0053] This step establishes a precise mapping from "radar wind measurement data + unit status" to "downstream inflow wind speed".
[0054] S2.1 Hybrid Network Architecture Design: Employs a hybrid deep neural network, comprising two sub-networks:
[0055] The main prediction network is responsible for learning the laws of fluid physics. Its inputs are feature vectors generated by a virtual LiDAR and terrain slope factors, and its output is the theoretical downstream inflow wind speed. This network is pre-trained offline using DatasetA generated in step S1.
[0056] Residual correction network: Responsible for capturing transient disturbances and model errors that cannot be simulated by the simulation. Inputs are measured SCADA data (such as upstream power and pitch angle) and measured LiDAR data; output is the wind speed correction. .
[0057] S2.2 Online Transfer Learning: The model is fine-tuned using real-world data during actual system operation. The loss function is defined as follows:
[0058]
[0059] in, These are the predicted values from the physical model. This is a correction value. It utilizes the power data from the downstream wind turbine's SCADA system. Through the inverse function of the unit power curve The "true" inflow wind speed is obtained by backpropagation. The ResNet weights are updated in real time through backpropagation, and the BaseNet is fine-tuned with a very small learning rate, enabling the model to adapt to changes in the current season and atmospheric stability.
[0060] S3: Collaborative Decision Making Based on Multi-Agent Reinforcement Learning
[0061] The high-precision prediction model trained in step S2 is used as the interactive environment to solve for the optimal yaw strategy.
[0062] S3.1 Problem Modeling: Model wind farm control as a Markov Decision Process (MDP). Treat each wind turbine as an intelligent agent.
[0063] state space ( ): Includes the power of the unit and its upstream neighboring units, the current yaw angle, the inflow wind speed measured by LiDAR, and the turbulence intensity.
[0064] Action space ( ): Yaw angle adjustment increment The range of a single adjustment is limited to Between, and the cumulative yaw deviation does not exceed .
[0065] S3.2 Reward Function Design: To achieve group collaboration, the first... Taiwan machine group rewards The design is as follows:
[0066]
[0067] In the formula, This refers to the power of the machine. This is the sum of the power of all downstream units affected by the wake of this unit; The synergy factor (ranging from 0 to 1, for example 0.5) forces upstream units to weigh downstream interests when making decisions. This is a penalty for yaw actions, used to suppress frequent or large yaw actions and protect the gearbox.
[0068] S3.3 Policy Training: Employs a multi-agent deep deterministic policy gradient algorithm. Offline training is conducted in a virtual environment, using an Actor-Critic structure to enable each agent to learn a collaborative strategy of "sacrificing personal interests (active yaw) for the greater good (improving overall field power)."
[0069] S4: Closed-loop real-time control and safety strategy
[0070] S4.1 Real-time Control: The system generates yaw commands at a period of 1-5 minutes (adjustable). The current real-time state is input into the trained Actor network, which outputs the optimal yaw angle matrix for the entire air crew in milliseconds and sends it to the air crew via OPCUA.
[0071] S4.2 Security Protection:
[0072] Communication heartbeat: If the communication between the software system and SCADA is interrupted for more than 10 seconds, the cooperative control will be automatically cancelled and the wind turbine will be instructed to resume the default maximum power tracking automatic wind mode.
[0073] Load reduction: The system has an internal load proxy model. If the yaw angle output by the algorithm is expected to cause the unit load to exceed the safety threshold, the system will automatically reduce the instruction to prioritize equipment safety.
[0074] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.
Claims
1. A real-time yaw cooperative control method for wind farms integrating machine learning, characterized in that, Includes the following steps: S1. Construct a prior knowledge base: acquire wind farm topographic data, construct a flow field simulation model using computational fluid dynamics (CFD), simulate the aerodynamic characteristics of the wind turbine using an actuator line model; set up a virtual lidar in the simulation environment to simulate the scanning method of a physical lidar, and generate a batch of simulation datasets containing environmental parameters, unit status, and virtual lidar wind measurement data. S2. Construct and train a virtual-real fusion prediction model: Establish a hybrid deep neural network, which includes a main prediction network and a residual correction network; use the simulation dataset to pre-train the main prediction network offline to master the physical laws of the terrain flow field; use the measured SCADA data and physical lidar data of the wind farm to train the residual correction network online, and fine-tune the main prediction network to output the predicted inflow wind speed of the downstream unit. S3. Generate a cooperative control strategy: Establish a multi-agent reinforcement learning model, treating each wind turbine as an agent, and using the virtual-real fusion prediction model as the interaction environment; design a reward function that includes single-unit power, total power of downstream units, and yaw action penalty, and train the agents to learn a cooperative yaw strategy. S4. Closed-loop real-time control: Real-time acquisition of wind farm operation data, input of a trained multi-agent reinforcement learning model, generation of yaw angle control commands for all units in the field and issuance for execution.
2. The wind farm real-time yaw cooperative control method integrating machine learning according to claim 1, characterized in that, In step S1, the line-of-sight velocity of the virtual lidar The calculation formula is: ； in, The velocity vector in the Cartesian coordinate system calculated by CFD. For the half-angle of the conical scan, The scanning azimuth angle is used; the data generation process of the virtual lidar includes volume weighting of the velocity of the grid nodes around the sampling point on the beam path.
3. The wind farm real-time yaw cooperative control method integrating machine learning according to claim 1, characterized in that, In step S2, the output of the virtual-real fusion prediction model is: ; in, The output of the main prediction network is virtual lidar data and terrain slope factor. The output of the residual correction network is the measured SCADA data and physical lidar data; the loss function of the online training is based on the actual inflow wind speed inferred from the downstream wind turbine SCADA data. The difference is constructed.
4. The wind farm real-time yaw cooperative control method integrating machine learning according to claim 1, characterized in that, In step S3, the reward function The expression is: ； in, This represents the current unit power. This is the sum of the power of downstream units affected by the wake. As a cooperating factor, Adjust the yaw angle increment. This is the penalty coefficient.
5. The wind farm real-time yaw cooperative control method integrating machine learning according to claim 1, characterized in that, Step S4 also includes a safety protection mechanism: when the communication interruption between the software system and the SCADA system exceeds a preset threshold, the coordinated control command is automatically revoked and the unit is restored to automatic wind-following mode; before issuing the command, the load proxy model is used to evaluate the predicted load, and if it exceeds the safety threshold, the yaw angle command is reduced.
6. A real-time yaw collaborative control system for wind farms integrating machine learning, characterized in that, include: The data acquisition and interaction module is used to connect the wind farm SCADA system and the nacelle-type lidar to collect unit operation data and wind measurement data; The high-performance computing and processing module is used to run the computational fluid dynamics simulation engine to generate virtual lidar data and to run the deep learning framework to train and infer hybrid deep neural networks. A collaborative control execution module is used to generate yaw control commands based on a multi-agent reinforcement learning strategy and write the commands into the wind turbine main control PLC through the OPCUA interface; the system is configured to execute the method as described in any one of claims 1 to 5.
7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 5.