Wind turbine generator rapid modeling method and system based on data driving

Abstract

The method comprises the following steps: acquiring operation data of an SCADA (Supervisory Control And Data Acquisition) system of the wind turbine generator and preprocessing the operation data to form a time sequence data set; a mutual information and random forest fusion feature selection method is adopted, double screening thresholds are set, and key feature variables are reserved; constructing a fan dynamic characteristic agent model based on a gating circulation unit, and training by adopting a Huber loss function; introducing physical constraints such as power-wind speed monotonicity, rotating speed-power characteristics and a pitch angle-power relation, and fusing a loss function through a penalty term; model parameters are updated in real time by using a recursive least square algorithm with a forgetting factor; a standard interface service is constructed through a gRPC framework, and integration with a wind field control system is realized; and establishing a three-level verification system of real-time residual monitoring, rolling cross validation and monthly model auditing. According to the method, high-precision, self-adaptive and deployable rapid modeling of the wind turbine generator is realized, and the modeling period is remarkably shortened.

Description

Technical Field

[0001] This invention belongs to the field of wind turbine modeling technology, and specifically relates to a data-driven method and system for rapid modeling of wind turbines. Background Technology

[0002] Currently, wind turbine modeling primarily relies on physical principle-based white-box or gray-box models, such as using specialized software like Bladed and FAST to build high-precision simulation models. This type of approach is typically used in the wind turbine design phase, constructing mathematical models based on aerodynamics, structural mechanics, and control theory, and then verifying their performance through simulation.

[0003] However, existing technologies have the following drawbacks: First, traditional methods have long modeling cycles, relying on detailed physical parameters and complex simulations, making the modeling process time-consuming and unable to meet the needs of real-time control and rapid operation and maintenance in engineering. Second, they have weak adaptability to operating conditions; classic models have fixed parameters, making it difficult to adaptively adjust to dynamic operating conditions such as equipment aging and environmental changes, resulting in a decrease in prediction accuracy over time. Finally, deploying traditional models is very complex; high-fidelity models often require specific hardware and software environments, making seamless integration with existing wind turbine control systems and data platforms difficult, resulting in high deployment costs and poor flexibility.

[0004] In summary, the technical problem to be solved by this invention is: 1. How to quickly build a high-precision digital model and shorten the modeling cycle by utilizing actual operating data of wind turbine units; 2. How to achieve dynamic self-calibration of model parameters to adapt to various operating conditions such as unit aging and environmental changes; 3. How to design standardized interfaces and verification mechanisms to support the rapid engineering deployment and application of the model in existing control systems. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a data-driven method and system for rapid modeling of wind turbines. By integrating multi-source operating data with lightweight machine learning algorithms, high-precision, adaptive digital modeling of the aerodynamic, transmission and control characteristics of wind turbines can be achieved.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A data-driven rapid modeling method for wind turbines, comprising the following steps: S1: Collect operating data from the SCADA system of the wind turbine, preprocess the operating data, and form a high-quality time series dataset; S2: A feature selection method that combines mutual information and random forest is adopted to calculate the correlation between each operating parameter and the core state variables of the wind turbine in the operating data, and a dual screening threshold is set to retain key feature variables; S3: A wind turbine dynamic characteristic proxy model is constructed based on a gated recurrent unit. The wind turbine dynamic characteristic proxy model contains multiple gated recurrent units and a fully connected output layer. The Huber loss function is used for training. S4: The parameters of the wind turbine dynamic characteristic proxy model are updated in real time using a recursive least squares algorithm with a forgetting factor. The forgetting factor is initially set to a value within a preset range. When the prediction error of the wind turbine dynamic characteristic proxy model exceeds a set threshold in multiple consecutive cycles, the forgetting factor is reduced. S5: Introduce physical constraints during the training process of the wind turbine dynamic characteristic proxy model, and add them to the Huber loss function in the form of a penalty term; S6: A standardized interface service is built through the gRPC framework. The standardized interface service provides prediction, parameter update and status query functions for the wind turbine dynamic characteristic proxy model, realizing the integration of the wind turbine dynamic characteristic proxy model with the wind farm control system. S7: Establish a three-level verification system that includes real-time residual monitoring, rolling cross-validation, and monthly model auditing. When the performance of the wind turbine dynamic characteristic proxy model continues to decline, initiate the incremental retraining process to update the wind turbine dynamic characteristic proxy model.

[0007] This invention integrates data-driven modeling, online parameter updates, physical mechanism constraints, and standardized engineering to achieve high-precision, adaptive, and deployable rapid modeling of wind turbines, significantly shortening the modeling cycle and improving the stability and generalization ability of the model under complex working conditions.

[0008] Preferably, the dual screening thresholds in step S2 include a mutual information value threshold and a random forest feature importance score threshold, and the number of key feature variables is 15 to 20.

[0009] This invention effectively improves the robustness and interpretability of feature selection, reduces redundant input, lowers model complexity, and improves training efficiency by setting a dual selection mechanism of mutual information and random forest, and limiting the range of key feature variables.

[0010] Preferably, in step S3, the wind turbine dynamic characteristic proxy model is trained using the AdamW optimizer, and a Dropout layer is set between the gated recurrent unit and the fully connected layer to prevent overfitting.

[0011] This invention employs the AdamW optimizer to improve the convergence speed and stability of model training, and combines it with the Dropout layer to suppress overfitting, thereby enhancing the generalization performance of the wind turbine dynamic characteristic surrogate model under small sample or noisy data.

[0012] Preferably, the mutual information value threshold is 0.2, and the random forest feature importance score threshold is 0.05.

[0013] This invention sets a screening threshold of 0.2 for mutual information and 0.05 for random forest feature importance score, which can effectively remove irrelevant or weakly correlated variables while retaining key operating parameters, thereby improving the quality of model input and ensuring modeling accuracy.

[0014] Preferably, in step S4, the recursive least squares algorithm with a forgetting factor sets an upper limit on the parameter update magnitude when updating the parameters of the wind turbine dynamic characteristic proxy model, to prevent the single parameter adjustment from exceeding the set range and causing the output of the wind turbine dynamic characteristic proxy model to be unstable.

[0015] This invention limits the magnitude of parameter updates to prevent drastic changes in model parameters due to abnormal data or sudden changes in operating conditions, thereby ensuring the numerical stability of the online learning process and improving the robustness of the model in dynamic environments.

[0016] Preferably, the physical constraints in step S5 include power-wind speed monotonicity constraints, rotational speed-power characteristic constraints, and pitch angle-power relationship constraints.

[0017] This invention introduces three types of physical constraints: power-wind speed, rotational speed-power, and pitch angle-power. This ensures that the model output conforms to the actual operating laws of wind turbines, reduces abnormal predictions that violate physical laws, and improves the model's credibility and engineering usability.

[0018] Preferably, the power-wind speed monotonicity constraint requires that the output power of the wind turbine unit remains monotonically constant as the wind speed increases under the same operating conditions; the speed-power characteristic constraint requires that the power value output by the wind turbine dynamic characteristic proxy model falls within a reasonable range of the wind turbine unit characteristic curve; and the pitch angle-power relationship constraint requires that in the region above the rated wind speed, the output power of the wind turbine unit decreases as the pitch angle increases.

[0019] This invention clarifies the specific mathematical and engineering meanings of three types of physical constraints, ensuring that the model's behavior in the critical operating range conforms to actual physical characteristics and improving the engineering credibility of the prediction results.

[0020] Preferably, the penalty term for physical constraints in step S5 adopts an adaptive weight adjustment strategy. The system monitors the satisfaction status of each physical constraint in real time. When any physical constraint is violated, the weight coefficient of the corresponding penalty term is increased. The system dynamically adjusts the weight of physical constraints according to the operating conditions of the wind turbine. The weight of physical constraints is reduced during the start-up, shutdown or change of operating conditions of the wind turbine.

[0021] This invention achieves a model that strictly follows physical laws under steady-state conditions and maintains flexibility during dynamic transitions by adaptively adjusting the physical constraint penalty weights, thus balancing model accuracy and adaptability.

[0022] Preferably, in step S6, the standardized interface service adopts a dual-buffered data access architecture. The real-time data channel accesses the memory mapping area of ​​the wind turbine SCADA system through shared memory, and the batch processing data channel connects to the real-time database through ODBC. Data synchronization between the two channels is achieved using a lock-free circular buffer.

[0023] This invention achieves efficient parallel access to high-frequency real-time data and large-volume historical data through a dual-caching architecture and a lock-free circular buffer design, ensuring data consistency and low latency, and improving the model service response speed and system throughput.

[0024] Preferably, the data acquisition latency of the real-time data channel is less than 10 milliseconds.

[0025] This invention controls the real-time data channel latency to within 10 milliseconds, meeting the application requirements of real-time control and rapid response of wind turbine units, and ensuring the timeliness of model prediction results.

[0026] Preferably, in step S7, real-time residual monitoring calculates the prediction error of the wind turbine dynamic characteristic surrogate model with a fixed time granularity, and issues an early warning when the prediction error of multiple consecutive monitoring cycles exceeds a set threshold; rolling cross-validation uses a sliding time window to evaluate the generalization ability of the wind turbine dynamic characteristic surrogate model; and monthly model auditing comprehensively evaluates the long-term performance of the wind turbine dynamic characteristic surrogate model.

[0027] This invention achieves full lifecycle monitoring of the model through a three-level verification system, promptly detects performance degradation, ensures long-term model availability, and reduces operational risks.

[0028] Preferably, the time granularity of real-time residual monitoring is 1 minute, and an early warning is issued when the prediction error of three consecutive monitoring cycles exceeds a set threshold; the rolling cross-validation adopts a 24-hour sliding time window.

[0029] This invention sets up a 1-minute monitoring granularity, a continuous 3-cycle early warning mechanism, and a 24-hour sliding verification window, which can effectively identify model performance anomalies while ensuring response speed and avoiding false alarms and missed alarms.

[0030] Preferably, the three-level verification system in step S7 adopts an intelligent threshold adjustment strategy. The system adjusts the alarm threshold of the prediction error according to the wind field turbulence intensity. When the wind field turbulence intensity increases, the error threshold is increased; when the wind field turbulence intensity decreases, the error threshold is decreased.

[0031] This invention dynamically adjusts the error alarm threshold based on the turbulence intensity of the wind field, thereby avoiding frequent false alarms under high turbulence conditions and improving the adaptability and practicality of the model verification system.

[0032] Preferably, the incremental retraining process retains important samples from historical data when updating the parameters of the wind turbine dynamic characteristic proxy model, and uses a memory playback mechanism to prevent the wind turbine dynamic characteristic proxy model from forgetting historical knowledge.

[0033] This invention preserves important historical samples and employs a memory replay mechanism to prevent the model from forgetting early operating conditions during incremental learning, thus preventing catastrophic forgetting and ensuring the model's long-term learning ability.

[0034] Preferably, the wind turbine dynamic characteristic proxy model is deployed on a Kubernetes container orchestration platform to achieve automatic scaling and fault recovery of standardized interface services through containerization.

[0035] This invention enables containerized deployment of model services using Kubernetes, improving system scalability, high availability, and ease of operation and maintenance, while reducing engineering integration costs.

[0036] A data-driven rapid modeling system for wind turbines includes: The data preprocessing module is used to collect the operating data of the wind turbine SCADA system and preprocess the operating data to generate a high-quality time series dataset. The feature selection module is used to calculate the correlation between each operating parameter and the core state variable of the wind turbine in the SCADA system operation data of the wind turbine using a feature selection method that combines mutual information and random forest. Based on the dual screening conditions of mutual information value threshold and random forest feature importance score threshold, it outputs key feature variables. The model building module is used to build a wind turbine dynamic characteristic proxy model based on gated recurrent units. The wind turbine dynamic characteristic proxy model contains multiple layers of gated recurrent units and a fully connected output layer, and is trained using the Huber loss function. The online learning module is used to update the parameters of the wind turbine dynamic characteristic surrogate model in real time using a recursive least squares algorithm with a forgetting factor. The forgetting factor is initially set to a value within a preset range. When the prediction error of the wind turbine dynamic characteristic surrogate model exceeds a set threshold for multiple consecutive cycles, the forgetting factor is reduced. The physical constraint embedding module is used to introduce physical constraints during the training process of the wind turbine dynamic characteristic proxy model and add them to the Huber loss function in the form of a penalty term. The interface service module is used to build standardized interface services through the gRPC framework. The standardized interface services provide prediction, parameter update and status query functions for the wind turbine dynamic characteristic proxy model. The verification and validation module is used to perform real-time residual monitoring, rolling cross-validation, and monthly model auditing. When the performance of the wind turbine dynamic characteristic proxy model continues to decline, the incremental retraining process is initiated to update the wind turbine dynamic characteristic proxy model. All modules are deployed on the Kubernetes container orchestration platform through a microservice architecture, achieving high availability and elastic scaling of the system.

[0037] This invention constructs a scalable, easily integrated, and adaptive rapid modeling system for wind turbines through modular and microservice design, realizing full-process automation from data acquisition to model deployment and verification, and supporting large-scale wind farm applications.

[0038] The present invention can achieve the following beneficial effects: 1. This invention significantly shortens the modeling cycle through automated data preprocessing, automatic selection of key parameters using feature selection algorithms, and a lightweight GRU network structure. Practical applications show that this method can complete high-quality model training within hours, significantly outperforming traditional physical modeling methods.

[0039] 2. This invention uses a recursive least squares algorithm with a forgetting factor to achieve real-time parameter updates. When the prediction error continues to increase, the forgetting factor can be automatically reduced to enhance the response capability to new data, enabling the model to effectively adapt to changes in the operating status of wind turbine units.

[0040] 3. This invention introduces physical constraints during training, such as power-wind speed monotonicity, rotational speed-power characteristics, and propeller pitch angle-power relationship, and incorporates a penalty term into the loss function, effectively reducing abnormal outputs that violate physical laws. Implementation data shows that the probability of model output violations significantly decreases after adding physical constraints.

[0041] 4. This invention constructs a standardized interface service through gRPC and adopts a dual-caching data access architecture. The real-time data channel accesses the SCADA system memory mapping area through shared memory, and the data acquisition latency is less than 10 milliseconds, which meets the response speed requirements of real-time control of wind turbine units. Attached Figure Description

[0042] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a time-series response comparison diagram of the wind speed prediction model of this invention. Detailed Implementation

[0043] Preferred solutions include Figures 1 to 2 As shown, a data-driven rapid modeling method for wind turbines is presented, with the following specific steps: S1 collects historical and real-time operating data recorded by the wind turbine's SCADA system, including wind speed, turbine speed, output power, pitch angle, generator torque, ambient temperature, and humidity.

[0044] Wind turbine: 2.5MW, has been in continuous operation for 3 years.

[0045] Wind field environment: annual average wind speed 7.5 m / s, turbulence intensity level B.

[0046] Data source: Unit SCADA system, sampling frequency is 1Hz.

[0047] Using historical data from the past six months as the training foundation, the data includes 12 core parameters such as wind speed, turbine speed, output power, blade pitch angle, generator torque, ambient temperature, and nacelle vibration. Outliers were identified and removed using the 3σ criterion or box plot method, and missing data were filled using linear interpolation or time series smoothing methods. All continuous variables were Z-score standardized to form a time-aligned, high-quality dataset, providing reliable input for subsequent modeling.

[0048] S2 employs a nonlinear feature selection method based on mutual information (MI) to calculate the correlation strength between each operating parameter and the core state variables of the wind turbine. Simultaneously, it combines the random forest algorithm to calculate feature importance scores, setting a double screening threshold: MI > 0.2 and importance score > 0.05, ultimately retaining 15-20 key feature variables.

[0049] S3 constructs a proxy model for the dynamic characteristics of wind turbines based on gated recurrent units (GRUs). The network structure consists of two layers of GRU units (64 / 32 neurons) and a fully connected output layer. The AdamW optimizer (learning rate 0.001) and Dropout (0.2) are used to prevent overfitting. The Huber loss function is adopted to balance the sensitivity of MSE to outliers.

[0050] S4 employs a recursive least squares algorithm with a forgetting factor to achieve real-time updates of model parameters. This method adjusts model parameters online through recursive calculation to adapt to system changes. The specific implementation process includes: first, constructing a regression vector containing historical input and output data; then, calculating the Kalman gain matrix; next, updating the parameter estimates; and finally, adjusting the covariance matrix. The forgetting factor in the algorithm is set between 0.95 and 0.99 to control the rate of weight decay of historical data. When a sustained increase in prediction error is detected, the system automatically reduces the forgetting factor to 0.85 to enhance responsiveness to new data. Simultaneously, the system monitors the parameter update magnitude, automatically limiting the adjustment step size when a single update exceeds a set threshold to prevent parameter mutations from causing system instability. This online learning process is executed automatically every 15 minutes to ensure the model can promptly track changes in the wind turbine's operating status.

[0051] S5 introduces physical constraints based on the wind turbine's operating mechanism to ensure that the model output conforms to actual physical laws. These constraints include a monotonicity constraint on power growth with wind speed, requiring the model's power output to monotonically increase with wind speed under the same operating conditions; a speed-power characteristic constraint, ensuring that the model's output power value falls within a reasonable range of the wind turbine's characteristic curve; and a pitch angle-power relationship constraint, ensuring that power decreases with increasing pitch angle in the region above rated wind speed. These constraints are added to the loss function in the form of penalty terms and optimized synchronously during model training.

[0052] S6 is designed as a high-performance interface service based on the gRPC framework, enabling seamless integration of the model with existing wind farm control systems. The interface service provides three core functionalities: model prediction, parameter updates, and status queries. It adopts the Protocol Buffers data exchange format. The deployment environment uses the Kubernetes container orchestration platform, supporting automatic scaling and fault recovery.

[0053] S7 establishes a comprehensive three-tiered model validation system. The first tier is real-time residual monitoring, calculating prediction error at a 1-minute granularity and issuing an alert when the error exceeds a threshold for three consecutive periods. The second tier is rolling cross-validation, using a 24-hour sliding time window to evaluate the model's generalization ability and calculate statistical indicators such as the coefficient of determination. The third tier is monthly model auditing, comprehensively evaluating the model's long-term performance. Validation indicators include mean absolute percentage error, root mean square error, and coefficient of determination.

[0054] Furthermore, in step S4, the online learning employs an intelligent forgetting mechanism, dynamically adjusting the forgetting factor based on the magnitude of the prediction error. When the prediction error is small, a high forgetting factor is maintained to fully utilize historical data and maintain estimation stability; when the prediction error increases, the forgetting factor is automatically reduced to quickly track system changes. Simultaneously, a parameter update constraint mechanism is introduced to ensure the numerical stability of the learning process.

[0055] Furthermore, in step S5, the physical constraints employ an adaptive weight adjustment strategy. The system monitors the satisfaction status of each constraint condition in real time. When a constraint is violated, the weight coefficient of the corresponding penalty term is automatically increased, forcing the model output to adjust in a direction that satisfies physical laws. Simultaneously, the constraint weights are dynamically adjusted according to different operating conditions, appropriately relaxing constraint restrictions during wind turbine start-up and shutdown or changes in operating conditions to improve model adaptability.

[0056] Furthermore, in step S6, the interface service adopts a dual-buffered data access architecture. The real-time data channel directly accesses the memory-mapped area of ​​the SCADA system via shared memory, ensuring a data acquisition latency of less than 10 milliseconds. The batch processing data channel connects to the real-time database via ODBC, supporting asynchronous loading and processing of large amounts of historical data. A lock-free circular buffer is used between the two channels to achieve data synchronization, ensuring data consistency and integrity.

[0057] Furthermore, in step S7, the verification system employs an intelligent threshold adjustment strategy. The system automatically adjusts the error threshold based on the wind turbulence intensity, using a strict threshold when wind conditions are stable and appropriately relaxing the threshold requirement when turbulence intensity is high to avoid false alarms. When model performance continues to decline, the system automatically triggers the model retraining process, using incremental learning to update model parameters while retaining historically important samples, thus avoiding catastrophic forgetting problems.

[0058] 1. The modeling efficiency of this invention is greatly improved, including the following aspects: ① Automated data preprocessing processes reduce manual intervention; ② Feature engineering algorithms automatically select key parameters; ③ Lightweight network structure accelerates the training process. Practical applications show that the new modeling method reduces manual labor time by 85% and computational resource consumption by 60%.

[0059] 2. The adaptability of this invention is significantly enhanced. Through an adaptive learning mechanism, the model can adapt well to different operating conditions: within a turbulence intensity variation range of ±0.15, the model prediction error fluctuation is less than 2%; when the unit performance degrades by 20%, the model accuracy degradation is controlled within 5% through adaptive parameter adjustment. This solves the pain point of traditional fixed-parameter models being unable to adapt to long-term operational changes.

[0060] The quantitative data records for the key steps are shown in the table below: Table 1. Quantitative Data Recording Table for Key Steps

[0061] 3. Significantly reduced engineering deployment costs: ① Plug-and-play interfaces reduce system modification work; ② Containerized deployment reduces environmental dependencies; ③ Automated verification processes reduce debugging time. Meanwhile, maintenance costs are reduced by 50%, mainly due to the model's self-updating capabilities, which reduce manual intervention.

[0062] 4. Compare and verify the data: To highlight the comprehensive advantages of this invention, the method (Solution A) is compared with traditional physical modeling methods (Solution B, using Bladed software) and a single GRU data-driven model that does not include online learning and physical constraints (Solution C). The comparison results are shown in the table below:

[0063] The image shows a comparison curve between predicted and actual values, indicating that the prediction error fluctuation is less than 2%.

[0064] Alternatively, the present invention can also employ other methods: 1. In addition to the mutual information-random forest fusion method, this invention can also employ: ① the maximum correlation minimum redundancy (mRMR) algorithm, which selects features by optimizing the correlation and redundancy between them; ② feature importance ranking based on XGBoost, using a gradient boosting framework to evaluate feature contribution; ③ principal component analysis (PCA) for feature dimensionality reduction, retaining principal components with more than 95% variance. These methods can also achieve the purpose of feature selection, but their computational complexity and effectiveness differ slightly.

[0065] 2. Algorithm replacement The GRU network can be replaced by: ① Temporal Convolutional Network (TCN), which uses dilated convolution to capture long-term dependencies and is more computationally efficient; ② Transformer encoder, which processes temporal data through a self-attention mechanism and is suitable for long sequence modeling; ③ Gaussian Process Regression (GPR), which provides estimation of prediction uncertainty and is suitable for small sample scenarios. These alternatives each have their advantages in different application scenarios.

[0066] 3. Alternatives to system integration solutions The gRPC framework can be replaced by: ① RESTful API, based on the HTTP protocol, offering better compatibility; ② MQTT protocol, suitable for lightweight communication in IoT scenarios; ③ OPCUA standard, a universal interface in the field of industrial automation. These alternatives can all achieve system integration in different application environments, but their performance and functional characteristics differ.

[0067] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A data-driven rapid modeling method for wind turbine generators, characterized in that, Includes the following steps: S1: Collect operating data from the SCADA system of the wind turbine, preprocess the operating data, and form a time-series dataset; S2: A feature selection method that combines mutual information and random forest is adopted to calculate the correlation between operating parameters and core state variables of wind turbine units in the operating data, and a dual screening threshold is set to retain key feature variables; S3: A wind turbine dynamic characteristic proxy model is constructed based on a gated recurrent unit. The wind turbine dynamic characteristic proxy model contains multiple gated recurrent units and a fully connected output layer. The Huber loss function is used for training. S4: The parameters of the wind turbine dynamic characteristic proxy model are updated in real time using a recursive least squares algorithm with a forgetting factor. The forgetting factor is initially set to a value within a preset range. When the prediction error of the wind turbine dynamic characteristic proxy model exceeds a set threshold in multiple consecutive cycles, the forgetting factor is reduced. S5: Introduce physical constraints during the training process of the wind turbine dynamic characteristic proxy model, and add them to the Huber loss function in the form of a penalty term; S6: A standardized interface service is built through the gRPC framework. The standardized interface service provides prediction, parameter update and status query functions for the wind turbine dynamic characteristic proxy model, realizing the integration of the wind turbine dynamic characteristic proxy model with the wind farm control system. S7: Establish a three-level verification system that includes real-time residual monitoring, rolling cross-validation, and monthly model auditing. When the performance of the wind turbine dynamic characteristic proxy model continues to decline, initiate the incremental retraining process to update the wind turbine dynamic characteristic proxy model.

2. The data-driven rapid modeling method for wind turbines according to claim 1, characterized in that, In step S2, the dual screening thresholds include a mutual information value threshold and a random forest feature importance score threshold, with 15 to 20 key feature variables.

3. The data-driven rapid modeling method for wind turbines according to claim 1, characterized in that, In step S3, the wind turbine dynamic characteristic proxy model is trained using the AdamW optimizer, and a Dropout layer is set between the gated recurrent unit and the fully connected layer to prevent overfitting.

4. The data-driven rapid modeling method for wind turbines according to claim 2, characterized in that, The mutual information value threshold is 0.2, and the random forest feature importance score threshold is 0.

05.

5. The data-driven rapid modeling method for wind turbines according to claim 1, characterized in that, In step S4, the recursive least squares algorithm with a forgetting factor sets an upper limit on the parameter update magnitude when updating the parameters of the wind turbine dynamic characteristic proxy model, to prevent the single parameter adjustment from exceeding the set range and causing the output of the wind turbine dynamic characteristic proxy model to be unstable.

6. The data-driven rapid modeling method for wind turbines according to claim 1, characterized in that, The physical constraints in step S5 include power-wind speed monotonicity constraints, rotational speed-power characteristic constraints, and pitch angle-power relationship constraints.

7. The data-driven rapid modeling method for wind turbines according to claim 6, characterized in that, The power-wind speed monotonicity constraint requires that, under the same operating conditions, the output power of the wind turbine unit remains monotonically constant as the wind speed increases; the speed-power characteristic constraint requires that the power value output by the wind turbine dynamic characteristic proxy model falls within a reasonable range of the wind turbine unit's characteristic curve; the pitch angle-power relationship constraint requires that, in the region above the rated wind speed, the output power of the wind turbine unit decreases as the pitch angle increases.

8. The data-driven rapid modeling method for wind turbines according to claim 1, characterized in that, In step S5, the penalty term for physical constraints adopts an adaptive weight adjustment strategy. The system monitors the satisfaction status of each physical constraint in real time. When any physical constraint is violated, the weight coefficient of the corresponding penalty term is increased. The system dynamically adjusts the weight of physical constraints according to the operating conditions of the wind turbine. The weight of physical constraints is reduced during the start-up, shutdown or change of operating conditions of the wind turbine.

9. The data-driven rapid modeling method for wind turbines according to claim 1, characterized in that, In step S6, the standardized interface service adopts a dual-buffered data access architecture. The real-time data channel accesses the memory mapping area of ​​the wind turbine SCADA system through shared memory, and the batch processing data channel connects to the real-time database through ODBC. Data synchronization between the two channels is achieved through a lock-free circular buffer.

10. A data-driven rapid modeling method for wind turbine generators according to claim 9, characterized in that, The data acquisition latency of the real-time data channel is less than 10 milliseconds.

11. The data-driven rapid modeling method for wind turbine generators according to claim 1, characterized in that, In step S7, real-time residual monitoring calculates the prediction error of the wind turbine dynamic characteristic surrogate model at a fixed time granularity. When the prediction error exceeds a set threshold for multiple consecutive monitoring periods, an early warning is issued. Rolling cross-validation uses a sliding time window to evaluate the generalization ability of the wind turbine dynamic characteristic surrogate model. Monthly model auditing provides a comprehensive evaluation of the long-term performance of the wind turbine dynamic characteristic surrogate model.

12. The data-driven rapid modeling method for wind turbines according to claim 11, characterized in that, The real-time residual monitoring time granularity is 1 minute, and an early warning is issued when the prediction error exceeds the set threshold for three consecutive monitoring cycles; the rolling cross-validation adopts a 24-hour sliding time window.

13. The data-driven rapid modeling method for wind turbines according to claim 1, characterized in that, In step S7, the three-level verification system adopts an intelligent threshold adjustment strategy. The system adjusts the alarm threshold of the prediction error according to the wind field turbulence intensity. When the wind field turbulence intensity increases, the error threshold is increased; when the wind field turbulence intensity decreases, the error threshold is decreased.

14. The data-driven rapid modeling method for wind turbines according to claim 1, characterized in that, The incremental retraining process retains important samples from historical data when updating the parameters of the wind turbine dynamic characteristic proxy model, and uses a memory playback mechanism to prevent the wind turbine dynamic characteristic proxy model from forgetting historical knowledge.

15. The data-driven rapid modeling method for wind turbine generators according to claim 1, characterized in that, The wind turbine dynamic characteristic proxy model is deployed on the Kubernetes container orchestration platform, which realizes automatic scaling and fault recovery of standardized interface services through containerization.

16. A data-driven rapid modeling system for wind turbine generators, characterized in that, The system employs a data-driven rapid modeling method for wind turbines according to any one of claims 1-15, comprising: The data preprocessing module is used to collect the operating data of the wind turbine SCADA system and preprocess the operating data to generate a high-quality time series dataset. The feature selection module is used to calculate the correlation between each operating parameter and the core state variable of the wind turbine in the SCADA system operation data of the wind turbine using a feature selection method that combines mutual information and random forest. Based on the dual screening conditions of mutual information value threshold and random forest feature importance score threshold, it outputs key feature variables. The model building module is used to build a wind turbine dynamic characteristic proxy model based on gated recurrent units. The wind turbine dynamic characteristic proxy model contains multiple layers of gated recurrent units and a fully connected output layer, and is trained using the Huber loss function. The online learning module is used to update the parameters of the wind turbine dynamic characteristic surrogate model in real time using a recursive least squares algorithm with a forgetting factor. The forgetting factor is initially set to a value within a preset range. When the prediction error of the wind turbine dynamic characteristic surrogate model exceeds a set threshold for multiple consecutive cycles, the forgetting factor is reduced. The physical constraint embedding module is used to introduce physical constraints during the training process of the wind turbine dynamic characteristic proxy model and add them to the Huber loss function in the form of a penalty term. The interface service module is used to build standardized interface services through the gRPC framework. The standardized interface services provide prediction, parameter update and status query functions for the wind turbine dynamic characteristic proxy model. The verification and validation module is used to perform real-time residual monitoring, rolling cross-validation, and monthly model auditing. When the performance of the wind turbine dynamic characteristic proxy model continues to decline, the incremental retraining process is initiated to update the wind turbine dynamic characteristic proxy model. All modules are deployed on the Kubernetes container orchestration platform through a microservice architecture, achieving high availability and elastic scaling of the system.

Images