Wind turbine generator yaw error identification and correction method and system based on graph attention network and embedded decoupling

Abstract

The invention discloses a wind turbine generator yaw error identification and correction method and system based on a graph attention network and embedded decoupling, belongs to the technical field of wind turbine generator control, and solves the problem of how to design a method and system for automatically and synchronously decoupling high-precision static errors and dynamic errors from mixed historical data. The problem of accurate yaw correction under all working conditions is solved. The method comprises the steps of collecting historical data of a wind turbine generator, and performing preprocessing; constructing a heterogeneous double-branch deep neural network model, wherein the heterogeneous double-branch deep neural network model comprises a graph attention network module and a learnable embedding module; training the heterogeneous double-branch deep neural network model by adopting a loss function; and performing one-time static compensation on the wind turbine generator, performing real-time dynamic error compensation by adopting a graph attention network module, and optimizing yaw control. The method is suitable for wind turbine generator error identification and correction scenes.

Description

A method and system for identifying and correcting yaw error in wind turbines based on graph attention networks and embedding decoupling. Technical Field

[0001] This invention belongs to the field of wind turbine control technology, specifically relating to a method for identifying and correcting yaw error in wind turbines. Background Technology

[0002] Against the backdrop of a global energy structure transition towards cleaner and lower-carbon energy, wind energy, as a key renewable energy source, is experiencing unprecedented growth in its development and utilization. As the core equipment for wind energy capture, the operating efficiency and total power generation of wind turbines directly depend on advanced control strategies. Among the three major control components of wind turbines—yaw, pitch, and torque—the performance of the yaw control system is particularly critical. The core task of this system is to adjust the turbine's orientation in real time, ensuring that the rotor's swept surface is always perpendicular to the incoming wind. This is a crucial prerequisite for achieving maximum power point tracking (MPPT) and reducing asymmetric fatigue loads on the turbine. However, the accuracy of yaw control highly depends on the wind direction measurement data provided by the anemometer mounted on top of the nacelle. Extensive industry practice and research have shown that, due to its installation location, the anemometer's measurement data generally suffers from significant and complex errors, making it unable to accurately reflect the free-flowing wind direction.

[0003] This total wind alignment error is generally considered to be a complex superposition of two different types of errors. The first is static error, also known as the inherent deviation of the wind vane. This error mainly originates from mechanical negligence during the on-site installation and commissioning of the turbine, sensor zero-point calibration deviation, or zero-point drift after long-term operation. Its core characteristic is that, for a single turbine, this error value is a basically constant physical deviation that does not change with operating conditions; each turbine is different. The second is dynamic error, also known as systematic error. The causes of this error are complex and varied, mainly due to the operating state of the wind turbine itself (such as the disturbance of the airflow ahead by blade rotation, and the obstruction effect of the nacelle on the wind) and mutual interference within the wind farm, especially the impact of the wake generated by upstream turbines on downstream turbines. Therefore, dynamic error is a non-linear, time-varying quantity, and its amplitude is closely related to real-time wind speed, wind direction, atmospheric turbulence intensity, and the operating conditions of other turbines in the field (such as power and pitch angle).

[0004] To correct these errors and improve power generation, existing technologies typically attempt to introduce high-precision reference data, such as nacelle-based LiDAR or ground-based anemometers, to identify the total error by comparing the reference wind direction with the nacelle anemometer readings. However, existing technologies generally employ a step-by-step identification strategy when dealing with these two types of errors. This strategy suffers from serious and insurmountable technical flaws, the core problem being the coupling and contamination of errors. Specifically, traditional methods attempt to first identify the static error. The logic is to select data under what are considered "ideal" operating conditions, such as high wind speeds, specific wind directions, and theoretically no upstream wake interference, assuming that the dynamic error is zero at these times, and then estimating the static error through simple averaging. However, this assumption is almost impossible to hold in the real, complex turbulent wind farm environment, and the selected data inevitably still carries unknown dynamic error components. This directly results in the static error value identified in the first step being contaminated and inaccurate. More seriously, this initial identification bias will be propagated and amplified step by step. When technicians attempt to build dynamic error models (e.g., using physical wake models or conventional neural networks) on a (faulty) static error baseline, the model is forced to fit residual data that already incorporates the faulty baseline. This not only significantly reduces the accuracy of the dynamic model but may even cause it to learn incorrect physical dependencies. Furthermore, dynamic errors (especially wakes) have strong spatiotemporal correlations. A wind farm is essentially a complex graph structure, where the error of a single turbine depends on the state and relative position of all turbines in the field (especially upstream). Existing single-unit models (such as LSTM and ANN) or simplified analytical models cannot effectively capture these nonlinear topological effects that change in real-time with wind direction across all turbines. Finally, most existing solutions terminate after identifying an inaccurate static error and performing a one-time compensation, completely ignoring online compensation for real-time dynamic errors. This results in turbines operating at suboptimal yaw angles for most of their operating time, continuously losing power generation. Therefore, the core technical problem to be solved in this application is how to design a method that can automatically and synchronously "decouple" high-precision static and dynamic errors from mixed historical data to achieve accurate yaw correction under all operating conditions.

[0005] Against this backdrop, Luan Fuming et al. disclosed a "Wind Turbine Yaw Error Testing Method Based on Feedforward Technology" in invention patent CN202211245137.4. This method utilizes lidar feedforward to measure wind direction and, based on the average value and dispersion of the measured yaw error, classifies the operating conditions into four types, thereby selecting between static and dynamic correction strategies. However, the core of this method is a rule-based strategy "selection," rather than error "identification." It does not provide an algorithm that can automatically "decouple" static installation errors and dynamic wake errors from mixed data. Its static correction (dependent on average values) identification results are still severely affected by dynamic errors (discretion). Furthermore, its dynamic correction heavily relies on expensive lidar for real-time feedforward and does not construct a dynamic prediction model capable of learning the full-field spatiotemporal wake effect and operating solely on SCADA data, limiting its application cost and scope.

[0006] On the other hand, Yu Kexi et al. disclosed a "method for calculating the yaw error of wind turbines based on goodness of fit" in invention patent CN202411753715.4. This method only uses SCADA data, finely divides wind speed and yaw angle into binaries, and selects binomial or trinomial fitting based on the statistical distribution characteristics of power in each bin (such as KS normal detection) to find the yaw angle corresponding to the maximum power point as the error. However, this method is essentially a "static" error identification technique based on traditional power curve fitting, and its accuracy is highly dependent on the data preprocessing (such as operating condition exclusion) and the rationality of the binning strategy. This method completely ignores the existence of dynamic system errors. It cannot handle dynamic deviations caused by complex spatiotemporal factors such as upstream wake, which change in real time with operating conditions, and it cannot provide a model for online compensation of dynamic errors. The GAT module of this invention can capture this dynamic spatiotemporal correlation, and the Embedding decoupling mechanism also avoids the dependence on ideal data screening. Summary of the Invention

[0007] This invention provides a method and system for identifying and correcting yaw error of wind turbines based on graph attention networks and embedded decoupling. The purpose is to solve the problem of how to design a system that can automatically and synchronously decouple high-precision static and dynamic errors from mixed historical data, and achieve accurate yaw correction under all operating conditions.

[0008] Firstly, the present invention aims to provide a method for identifying and correcting yaw error of wind turbine generators based on graph attention network and embedded decoupling, comprising the following steps: S1: Collecting historical data of all wind turbine generators in the wind farm and preprocessing it to construct a database; S2: Constructing a heterogeneous dual-branch deep neural network model, wherein the heterogeneous dual-branch deep neural network model includes a graph attention network module and a learnable embedded module, wherein the graph attention network module is used to output dynamic yaw error and the learnable embedded module is used to output static yaw error; S3: Training the heterogeneous dual-branch deep neural network model using a loss function until the model converges, and synchronously updating the parameters of the graph attention network module and the learnable embedded module through backpropagation to achieve automatic decoupling of dynamic and static errors; S4: Extracting and solidifying the static error value of the learnable embedded module from the trained model to perform one-time static compensation of the wind turbine generators, and using the graph attention network module for real-time dynamic error compensation to optimize yaw control, thereby completing the identification and correction of yaw error of the wind turbine generators.

[0009] Furthermore, a preferred solution is provided: the historical data includes: reference wind direction, wind direction measured in the nacelle, and unit operating data.

[0010] Furthermore, a preferred solution is provided: the preprocessing includes: data cleaning and data normalization of historical data.

[0011] Furthermore, a preferred embodiment is provided: the output of the heterogeneous dual-branch deep neural network model is the sum of the dynamic yaw error and the static yaw error.

[0012] Furthermore, a preferred solution is provided: the graph attention network module uses a multi-layer stacked graph attention network for spatiotemporal feature propagation and aggregation, and the acquisition of dynamic yaw error includes the following steps: calculating the attention coefficients of the nodes in the first layer that receive information from their neighbors; normalizing the attention coefficients to obtain attention weights; updating the feature representation of the nodes through weighted aggregation; obtaining the final node representation after multi-layer stacking, and obtaining the dynamic yaw error using a multilayer perceptron regression head.

[0013] Furthermore, a preferred solution is provided: the construction steps of the learnable embedding module include: defining a learnable embedding matrix, determining the matrix dimension based on the total number of units in the wind farm; the learnable embedding module outputs the static yaw error through an index lookup operation.

[0014] Furthermore, a preferred embodiment is provided: the loss function is defined as the mean square error between the predicted and actual total yaw error, wherein the total yaw error is the sum of the dynamic yaw error and the static yaw error.

[0015] Secondly, the purpose of this invention is to propose a method for identifying and correcting yaw error of wind turbines based on graph attention network and embedded decoupling, as described in any one or more of the above-mentioned schemes. The system includes: a data acquisition and processing unit for acquiring historical data from all wind turbines in the wind farm, preprocessing it, and constructing a database; and a neural network model construction unit for constructing a heterogeneous dual-branch deep neural network model, which includes a graph attention network module and a learnable embedded module. The graph attention network module is used to output dynamic yaw error... The learnable embedding module is used to output static yaw error; the model training unit is used to train the heterogeneous dual-branch deep neural network model using a loss function until the model converges, and to synchronously update the parameters of the graph attention network module and the learnable embedding module through backpropagation to achieve automatic decoupling of dynamic and static errors; the error correction unit is used to extract and solidify the static error value of the learnable embedding module from the trained model to perform one-time static compensation for the wind turbine, and to perform real-time dynamic error compensation using the graph attention network module to optimize yaw control and complete the identification and correction of the yaw error of the wind turbine.

[0016] Thirdly, the present invention aims to provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, and when the processor runs the computer program stored in the memory, the processor executes a method for identifying and correcting yaw error of wind turbines based on graph attention network and embedded decoupling, according to any one or more of the above-described schemes.

[0017] Fourthly, the present invention aims to provide a computer-readable storage medium for storing a computer program that executes the method for identifying and correcting yaw error of wind turbines based on graph attention networks and embedded decoupling, as described in any one or more of the above-described schemes.

[0018] Compared with existing technologies, the advantages of this invention are as follows: 1. Automatic decoupling and high identification accuracy: The core of the method described in this invention is that, through the heterogeneous fusion architecture of GAT + Embedding and end-to-end training, static errors and dynamic errors are automatically and synchronously decoupled in complex data, which completely solves the fundamental defect of mutual contamination of the two types of errors in existing technologies, leading to inaccurate identification.

[0019] II. Powerful dynamic spatiotemporal modeling: The method described in this invention uses a graph attention network (GAT), which can adaptively learn the complex wake effects between units in a wind farm (edge ​​weights of the graph). Its ability to capture nonlinear, time-varying, and spatial topological relationships far exceeds that of traditional wake models or time series models such as LSTM.

[0020] 3. Zero cost of static calibration: Compared with the "S-curve" method that requires active power abandonment, the method described in this invention can identify high-precision static installation errors using only historical operating data (SCADA and LiDAR) without any additional downtime or power loss.

[0021] IV. Real-time Compensation Under All Operating Conditions: This invention not only provides a one-time static correction value, but also an online-running GAT model for real-time compensation of dynamic errors. This ensures that the unit can be aligned with the incoming wind to the greatest extent possible under any wind speed, wind direction, and wake conditions, significantly improving the unit's power curve and total power generation.

[0022] V. Strong model adaptability: GAT's attention mechanism enables it to automatically adapt to different wind directions (changes in wake propagation direction) and atmospheric conditions, exhibiting strong robustness and generalization ability.

[0023] This invention is applicable to scenarios involving error identification and correction in wind turbine generators. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0025] Figure 1 is a schematic diagram of the system architecture of a wind turbine yaw error identification and correction method based on graph attention network and embedded decoupling according to a specific embodiment of the present invention; Figure 2 is a schematic diagram of the internal structure of the dynamic decoupling identification model according to a specific embodiment of the present invention; Figure 3 is a flowchart of the implementation of the yaw error identification and correction method according to a specific embodiment of the present invention. Detailed Implementation

[0026] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application can also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0027] The technical solutions in the embodiments of this application 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 this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0028] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0029] This implementation method provides a method for identifying and correcting yaw error in wind turbines based on graph attention networks and embedded decoupling, including the following steps: Step 1, Data Acquisition and Preprocessing: Step 1.1, Data Acquisition: The first step of the method described in this implementation is to construct a comprehensive, time-synchronized multi-source database. This database covers all sources within the wind farm. Typhoon turbine ( In continuous Each time step ( Historical data within a sampling period of 1 second, 1 minute, or 10 minutes. The historical data includes: (1) high-precision reference wind direction. This data, as a truth label, usually comes from a feedforward laser wind measuring radar installed on the nacelle or a wind measuring tower calibrated on-site; (2) Wind direction measured in the nacelle (2) The original reading of the nacelle wind vane; (3) The crew operating data (SCADA), including at least the following: Time of the first Active power of the Taiwanese generator set Pitch angle Rotor speed And the cabin orientation (absolute bearing). (4) Unit unique identifier That is, a discrete integer ( ), used to uniquely designate each unit.

[0030] Step 1.2, Data Preprocessing: First, calculate the target ground truth value required for model training, i.e., the total wind error. The calculation formula is as follows: .

[0031] Simultaneously, it is necessary to handle wind direction data crossing... To address the issue of time jumps and ensure the continuity of error calculation.

[0032] Secondly, regarding SCADA operating condition data Perform data cleaning to remove obvious abnormal data points caused by shutdown, power limitation, communication interruption, or sensor failure.

[0033] Finally, to eliminate the influence of different physical dimensions and accelerate neural network convergence, all input SCADA operating condition data need to be normalized, i.e., Z-Score standardization is adopted. Scale it to or Within the range.

[0034] Step 2, Dynamic Decoupling Identification Model Construction: The core of the method described in this embodiment lies in designing a heterogeneous dual-branch deep neural network model, which can achieve simultaneous identification and decoupling of two different types of errors under a unified architecture.

[0035] The constructed heterogeneous dual-branch deep neural network model includes a graph attention network branch and a learnable embedding branch.

[0036] The purpose of the graph attention network branch is to capture complex, nonlinear, and spatiotemporally correlated dynamic system errors caused by wake, terrain, etc. First, The wind farm at any given time is constructed as a graph .in yes Node set of typhoon generator units; yes The node feature matrix ( (For SCADA feature dimensions) The adjacency matrix can be set as a fully connected graph (allowing GAT to learn dependencies automatically) or a sparse graph based on physical distance. This module uses a multi-layer stacked graph attention network (GATv2) for spatiotemporal feature propagation and aggregation. For nodes in layer 1... From the neighbors Attention coefficient of received information The calculation is as follows: ,in It is a node The feature vectors of the first layer, and It is a learnable linear transformation matrix. It is a learnable attention vector. The attention coefficients are obtained through... The function is normalized to obtain the final attention weights. : , Dynamically represented in Time, node To what extent should the dynamic error focus on the nodes? The operating conditions were then updated via weighted aggregation. Feature representation: To improve robustness, the following approach is adopted. Multiple parallel multi-head attention will The outputs of each item are concatenated or averaged. After stacking layers of GAT, the final node representation containing the entire spatiotemporal conditions is obtained. This data is fed into a multilayer perceptron (MLP) regression head to output the dynamic error scalar value at that moment: Due to its input It is time-varying; the output of the GAT module. It will inevitably take time And the dynamic changes in the overall operating conditions.

[0037] The purpose of the learnable embedding branch is to identify the inherent, constant static anemometer error of each unit. This implementation innovatively introduces a learnable embedding layer, which is essentially a lookup table. A learnable embedding matrix is ​​defined. Its dimensions are ,in This represents the total number of turbines in the wind farm. The matrix represents the [missing information]. OK (A scalar) is designed as a trainable model parameter, which is related to the first... Taiwanese machine group Unique binding. When the model processing unit In any When the data is accessed at a given time, this module performs a simple index lookup operation: The key is that The output depends only on the unit , and with SCADA operating data at any time Completely irrelevant. Therefore, this parameter Throughout the training process, a constant value is forced to be learned, which represents the first... The inherent deviation component that exists in all operating conditions of the unit.

[0038] The final predicted total error in this step This is explicitly defined as the sum of the outputs of two independent branches; this summation structure is key to achieving decoupling. It is represented as: ,in, It is the output of the Graph Attention Network (GAT) branch and depends on the real-time operating conditions of the entire field. The predicted value of dynamic system error; and It is output by the learnable embedding branch and depends only on the unit. And it does not change with time. The changing static wind vane error prediction values. These two modules are jointly trained under the supervision of the same loss function.

[0039] Step 3, Model Training and Error Decoupling: Step 3.1, Loss Function Definition: The training objective of the model is to minimize the total prediction error. Total measurement error compared to the actual value The difference between them. This step uses Mean Squared Error (MSE) as the loss function. The computation is performed on a mini-batch of B samples: Substituting the overall model architecture defined in step 2 into the loss function, we obtain the core optimization objective of the model: Substituting the outputs of the two branches further, we obtain the complete loss function expression: ,in It is the GAT module for the unit exist The final characteristic representation of the operating condition at any given time.

[0040] Step 3.2, Backpropagation and Decoupling: The automatic decoupling mechanism of this method is implemented during backpropagation. This occurs when the Adam optimizer is used to calculate the loss function. When applying gradients to model parameters, the gradient is propagated back along two independent branches: (1) Dynamic gradient path: loss For GAT network parameters (including) gradient of weights Because the input to GAT is time-varying... The highly variable gradient path will force the GAT module to learn and fit. All components that fluctuate over time are considered dynamic system errors.

[0041] (2) Static gradient path: Loss For a specific entry in the Embedding matrix gradient .because It is a step in all time. The optimizer adjusts the shared, constant parameters when minimizing the global loss. Absorb The mean bias that does not change over time.

[0042] In this way, the GAT module is forced to handle variance, while the Embedding module is forced to handle mean. The single loss function and dual-branch architecture ultimately achieve automatic separation and identification of the two coupled errors in a single training session.

[0043] Step 4, Error Correction: Step 4.1, Static Error Correction: After the model training converges, the weights of the Embedding module are extracted and fixed from the trained model, i.e. Embedding matrix The first of the matrix OK That is, the first one identified The high-precision static installation error of the unit. This value should be written into the first... The main control software parameters of the generator set include permanent zero-point offset compensation for the wind vane, thereby completing static calibration.

[0044] Step 4.2, Dynamic Error Correction: Simultaneously, the trained GAT module (containing all GAT layers and the MLP regression head) is deployed to the online central control server or edge computing unit of the wind farm. This model subscribes to the entire wind farm in real time (e.g., every minute). SCADA operating data of the unit It then performs forward inference computation. The model outputs data in real time. One dynamic error prediction value The real-time changing dynamic compensation value is sent to the main controller of the corresponding unit. The main control software superimposes it on the (statically corrected) wind direction reading to generate the final yaw target value, thereby achieving real-time dynamic wind optimization under all operating conditions.

[0045] Implementation Method Two: This implementation method provides a wind turbine yaw error identification and correction system based on graph attention networks and embedded decoupling. This system is typically deployed in the central control server or edge computing gateway of a wind farm. Based on the wind turbine yaw error identification and correction method based on graph attention networks and embedded decoupling as described in Implementation Method One, as shown in Figure 1, the system includes: a data acquisition module, which connects to the main controllers and nacelle lidar (not shown) of all wind turbines in the wind farm via an industrial bus or OPC UA protocol, for real-time and batch acquisition of SCADA operating condition data, nacelle wind vane measurements, and high-precision reference wind direction values; and a data preprocessing module, connected to the data acquisition module, responsible for timestamping the acquired raw data, removing outliers, normalizing operating condition data, and calculating the total error label for model training. The dynamic decoupling identification model, the core execution unit, receives data from the data preprocessing module. As shown in Figure 2, the model integrates a heterogeneous dual-branch architecture, comprising a GAT dynamic error branch and an Embedding static error branch. The GAT dynamic error branch receives real-time operating condition data from all units in the site. As a graph feature input, its graph attention network structure is used to capture the spatiotemporal wake effect and output dynamic error. The static error branch is the unique identifier of the receiver group. As input, by finding a trainable embedding matrix Output the constant static error of the unit. An adder sums the outputs of two branches. This total predicted value is used during the training phase to compare with... Calculate the loss. The correction signal output module, connected to the model and the wind farm main control system, is responsible for extracting the static error matrix from the Embedding static error branch after model training is complete. It then sends this information to the main control system for a one-time update of the unit's parameters; simultaneously, while running online, this module acquires the output of the GAT dynamic error branch in real time. It is then sent to the main control system as a dynamic compensation signal.

[0046] Implementation Method 3 is a further illustrative example of the wind turbine yaw error identification and correction method based on graph attention network and embedded decoupling described in Implementation Method 1.

[0047] First, the data acquisition and preprocessing steps are performed. The data acquisition module collects data from N wind turbines (e.g., N wind turbines) within the wind farm. The unit's historical data at T time points (e.g., 1 year) is processed by the data preprocessing module, which cleans the data and calculates the total error according to the method in step 1.2. Next, the model building and training steps are executed. The system constructs the dynamic decoupling identification model shown in Figure 2, where the GAT dynamic error branch is specifically set to include... A multi-layered GATv2 network, with each layer employing A single attention head is used to fully capture complex nonlinear spatiotemporal dependencies; the embedding static error branch is specifically set as a... dimensional trainable embedding matrix Subsequently, the model is trained offline end-to-end using the MSE loss function and AdamW optimizer defined in step 3.1. During backpropagation, the system, according to the decoupling mechanism described in step 3.2, uses the loss gradient to simultaneously update the network weights of the GAT branch and the Embedding branch. Parameters. After the model training converges, the deployment and calibration steps are executed. First, the calibration signal output module extracts the converged parameters from the model. (e.g., the value in row 3) ), and wrote it as a static bias compensation value into the main control system of wind turbine No. 3 in one go, completing all Static calibration of the generator set. Then, the model weights of the GAT dynamic error branch are solidified and deployed to the online system, which receives real-time data on the entire field's operating conditions. The module calculates and outputs the dynamic error in real time. (like The dynamic value is fed to the main control system, which then superimposes it onto the statically corrected wind direction to achieve real-time closed-loop optimization across all operating conditions.

[0048] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A method for identifying and correcting yaw error in wind turbines based on graph attention networks and embedded decoupling, characterized in that, Includes the following steps: S1: Collect historical data of all wind turbines in the wind farm, preprocess the data, and build a database; S2: Construct a heterogeneous dual-branch deep neural network model, which includes a graph attention network module and a learnable embedding module. The graph attention network module is used to output dynamic yaw error, and the learnable embedding module is used to output static yaw error. S3: Train the heterogeneous dual-branch deep neural network model using a loss function until the model converges. Then, update the parameters of the graph attention network module and the learnable embedding module synchronously through backpropagation to achieve automatic decoupling of dynamic and static errors. S4: Extract and solidify the static error values ​​of the learnable embedding module from the trained model to perform one-time static compensation for the wind turbine, and use the graph attention network module for real-time dynamic error compensation to optimize yaw control and complete the identification and correction of the yaw error of the wind turbine.

2. The method for identifying and correcting yaw error of wind turbines based on graph attention networks and embedded decoupling as described in claim 1, characterized in that, The historical data includes: reference wind direction, wind direction measured in the engine room, and unit operating data.

3. The method for identifying and correcting yaw error of wind turbines based on graph attention networks and embedded decoupling as described in claim 1, characterized in that, The preprocessing includes: data cleaning and data normalization of historical data.

4. The method for identifying and correcting yaw error of wind turbines based on graph attention networks and embedded decoupling as described in claim 1, characterized in that, The output of the heterogeneous dual-branch deep neural network model is the sum of the dynamic yaw error and the static yaw error.

5. The method for identifying and correcting yaw error of wind turbines based on graph attention networks and embedded decoupling as described in claim 1, characterized in that, The graph attention network module uses a multi-layer stacked graph attention network for spatiotemporal feature propagation and aggregation. The acquisition of dynamic yaw error includes the following steps: calculating the attention coefficients of the nodes in the first layer that receive information from their neighbors; normalizing the attention coefficients to obtain attention weights; updating the feature representation of the nodes through weighted aggregation; obtaining the final node representation after multi-layer stacking; and obtaining the dynamic yaw error using a multilayer perceptron regression head.

6. The method for identifying and correcting yaw error of wind turbines based on graph attention networks and embedded decoupling as described in claim 1, characterized in that, The construction steps of the learnable embedding module include: defining a learnable embedding matrix and determining the matrix dimension based on the total number of units in the wind farm; the learnable embedding module outputs the static yaw error through an index lookup operation.

7. The method for identifying and correcting yaw error of wind turbines based on graph attention networks and embedded decoupling as described in claim 1, characterized in that, The loss function is defined as the mean square error between the predicted and actual total yaw error, where the total yaw error is the sum of the dynamic and static yaw errors.

8. A wind turbine yaw error identification and correction system based on graph attention network and embedded decoupling, characterized in that, The system is based on a wind turbine yaw error identification and correction method based on graph attention network and embedded decoupling as described in any one of claims 1-7. The system includes: a data acquisition and processing unit for collecting historical data of all wind turbines in the wind farm, preprocessing the data, and constructing a database; a neural network model construction unit for constructing a heterogeneous dual-branch deep neural network model, which includes a graph attention network module and a learnable embedded module, wherein the graph attention network module outputs dynamic yaw error and the learnable embedded module outputs static yaw error; a model training unit for training the heterogeneous dual-branch deep neural network model using a loss function until the model converges, and synchronously updating the parameters of the graph attention network module and the learnable embedded module through backpropagation to achieve automatic decoupling of dynamic and static errors; and an error correction unit for extracting and solidifying the static error value of the learnable embedded module from the trained model to perform one-time static compensation for the wind turbine, and using the graph attention network module for real-time dynamic error compensation to optimize yaw control and complete the identification and correction of wind turbine yaw error.

9. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes a method for identifying and correcting yaw error of wind turbines based on graph attention networks and embedded decoupling, according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program that executes a method for identifying and correcting yaw error of a wind turbine based on graph attention network and embedding decoupling, as described in any one of claims 1-7.

Images