Wind turbine generator full-link self-adaptive control method based on dynamic perception and coevolution DRL

Abstract

The invention provides a wind turbine generator set full-link self-adaptive control method based on dynamic perception and coevolution DRL, and relates to the technical field of wind power generation control. S1, the running state and environmental wind regime data of a wind turbine generator set are collected through a multi-mode sensor array, the collected running state and environmental wind regime data are preprocessed, and a wind turbine generator set full-link self-adaptive control result is obtained; and outputting the multi-source data passing the verification. And S2, inputting the multi-source data passing the verification into a physical knowledge enhanced convolution encoder, embedding hydrodynamic constraints to carry out feature extraction, generating a working condition type, a dynamic feature parameter and confidence, and synthesizing an extreme working condition sample through a generative adversarial network. On-line co-evolution of a model and a control strategy can be realized through multi-source data fusion, lightweight deep reinforcement learning and a stability constraint mechanism.

Description

Technical Field

[0001] This invention relates to the field of wind power generation control technology, and in particular to a full-link adaptive control method for wind turbine generators based on dynamic perception and collaborative evolutionary DRL. Background Technology

[0002] With the continued growth in global demand for renewable energy, wind power has been widely used as a clean and efficient energy source. However, natural wind conditions are highly random and dynamic, with rapid switching between various operating conditions such as stable wind, turbulent wind, and extreme wind. Furthermore, dynamic factors such as wind speed turbulence, wind shear, and tower shadow effects directly affect the pitch and yaw response of wind turbine units.

[0003] In the prior art, patent number CN112483334B discloses a wind turbine intelligent control method based on edge computing. The technical solution collects real-time operating data through the processor on the wind turbine side, performs FFT analysis, fault diagnosis and control parameter adjustment based on expert strategy after preprocessing and screening, and finally interacts with the SCADA system and cloud platform. However, this technical solution still has obvious defects: it relies on preset preprocessing conditions and expert strategies to adjust control parameters, which is essentially a static optimization mode. It does not consider the rapid switching of the working condition from stable wind to turbulent wind, and cannot match the dynamic response requirements of dynamic factors such as wind speed turbulence, wind shear, and tower shadow effect to the pitch and yaw links in real time, which makes the control commands prone to lag or over-adjustment. At the same time, the priority rules of its adaptive strategy are preset and fixed, and no dynamic adjustment mechanism for the priority of control commands under dynamic disturbances is set. In dynamic disturbance scenarios such as extreme wind conditions and instantaneous equipment failure, it is prone to local convergence or strategy instability, lacks global optimization capabilities, and cannot balance the relationship between strategy exploration and stable control in real time. It may stick to the local optimal solution or core commands and non-core commands compete for transmission resources, which may lead to mechanical load impact. Summary of the Invention

[0004] To address the shortcomings of the existing technologies, the technical problem to be solved by this invention is to provide a wind turbine end-to-end adaptive control method based on dynamic perception and co-evolutionary DRL. This method can achieve online co-engineering of the model and control strategy through multi-source data fusion, lightweight deep reinforcement learning, and stability constraint mechanisms. It solves the technical problems encountered in the construction process, such as the inaccuracy of static models in depicting dynamic working conditions and the lack of stability of adaptive algorithms under random disturbances.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: The present invention provides a wind turbine end-to-end adaptive control method based on dynamic perception and cooperative evolutionary DRL, comprising the following steps: S1. Collect the operating status and environmental wind condition data of the wind turbine generator set through a multi-modal sensor array, preprocess the collected operating status and environmental wind condition data, and output the verified multi-source data. S2. Input the verified multi-source data into a physics-enhanced convolutional encoder, embed fluid dynamics constraints for feature extraction, generate working condition types, dynamic feature parameters and confidence levels, and synthesize extreme working condition samples through a generative adversarial network. S3. Input the dynamic feature parameters and confidence scores into the pre-trained deep reinforcement learning DRL model deployed in the AI ​​co-controller. The deep reinforcement learning DRL model integrates the regularization mechanism driven by dual critics and TD error, embeds Lyapunov stability constraints, adaptively switches parameter configurations according to the working condition type, and outputs variable pitch PID coefficients and yaw angle control commands. S4. Correct the pitch PID coefficient and yaw angle control command through a self-punishment mechanism. S5. Using an event-driven communication framework, the corrected pitch PID coefficients and yaw angle control commands are transmitted to the main control PLC for execution according to priority, triggering the wind turbine to perform pitch and yaw angle adjustments. S6. Collect actual operating data of the actuator and feed it back to the deep reinforcement learning (DRL) model in step S3 to achieve online iterative optimization of the strategy.

[0006] In the preferred embodiment, step S1 includes a multimodal sensor array consisting of a high-frequency lidar, an inertial measurement unit, an acoustic sensor, a vibration sensor, an accelerometer, an inclinometer, a gyroscope, and an ultrasonic probe. Operational status and environmental wind condition data include SCADA system data, vibration sensor data, meteorological sensor data, wind speed profile, blade vibration, tower displacement, and aerodynamic load data; When the wind turbine generator completes its startup self-test or the preset operating condition monitoring cycle is reached, the multi-modal sensor array is triggered to start data acquisition.

[0007] In the preferred embodiment, the specific operating logic of the multimodal sensor array in step S1 is as follows: S11. When the wind turbine generator starts its self-test, it triggers sensor fault diagnosis, sends test signals to each sensor, and receives feedback response time < This is normal; S12. After the self-test passes, the accelerometer detects that the gearbox speed is greater than... Start Sampling, the inclinometer is used when the blade pitch angle is greater than... Start Sampling is initiated when the gyroscope is activated during yaw system startup. Sampling, ultrasonic probe under tower vibration > Start sampling; S13. When any sensor detects an abnormality in its own data, it triggers the activation of redundant sensors and sends a "sensor abnormality" warning to the AI ​​co-controller. in, This is expressed as the maximum tolerable response time of the sensor feedback; This represents the rotational speed threshold at which the accelerometer starts sampling; This represents the angle threshold at which the inclinometer starts sampling; This represents the vibration threshold at which the ultrasonic probe initiates sampling. This is expressed as the sampling frequency of the accelerometer; This is expressed as the sampling frequency of the inclinometer; This is expressed as the sampling frequency of the gyroscope; This is expressed as the sampling frequency of the ultrasonic probe.

[0008] In the preferred embodiment, the specific steps of step S2 are as follows: S21. Based on the verified multi-source data, it is determined that there is data heterogeneity, triggering the physical knowledge-enhanced convolutional encoder to start the feature extraction process; obtain the data alignment module, input the collected data into the data alignment module, and unify the low-resolution data into the preset high-resolution data through linear interpolation. S22. The attention mechanism module is used to extract physical constraint features by embedding fluid dynamics constraints. At the same time, the attention mechanism module is triggered to calculate the weight coefficients based on the importance scores of each sensor data output by S1 and dynamically allocate data weights. The final output includes the working condition type, dynamic feature parameters and confidence level. The working condition type includes steady wind, turbulent wind and extreme wind. The dynamic feature parameters include wind speed change rate and load fluctuation amplitude. S23. When the working condition classification module collects extreme working condition samples and the generator of the generative adversarial network is in a ready state, the generator is triggered to start sample synthesis, and the synthesized samples are added to the sample library.

[0009] In the preferred embodiment, step S3 consists of the following steps: S31. Based on the dynamic feature parameters and confidence level, the AI ​​co-controller triggers the loading of the pre-trained deep reinforcement learning (DRL) model. S32. Read the working condition type based on the loaded deep reinforcement learning (DRL) model. If the working condition type changes, trigger the model parameter configuration switch. S33. Subsequently, the model integrates a dual-criterion and TD error-driven regularization mechanism, embedding Lyapunov stability constraints; whereby, when the deep reinforcement learning DRL model performs Q-value calculation, the dual-criterion architecture is triggered to run in parallel: Critics 1 calculates based on dynamic feature parameters. ,in, Represented as the Q-function of commentator 1; This is represented as the parameter matrix for commentator 1; commentator 2 calculates based on the same parameters. ,in, Represented as the Q function of commentator 2; Represented as the parameter matrix of commentator 2; Pick and The minimum value is taken as the target Q value, and then the TD error is calculated. Let the threshold of TD error be set. ,like This triggers the regularization mechanism, applying an L2 regularization term to the policy network parameters. According to the regularization term Adjust the learning rate from Down to ,in, This is represented as the learning rate after regularization; Set a threshold number of iterations for TD error stability. ,when And maintain In the next iteration, the regularization mechanism is turned off, and the initial learning rate is restored; S34. Input dynamic characteristic parameters for inference, and output variable pitch PID coefficients and yaw angle control commands.

[0010] In the preferred embodiment, step S4 consists of the following steps: S41. Safety assessment of triggering the self-penalty mechanism based on the pitch PID coefficient and yaw angle control command; S42. Call the preset safety threshold library. Safety thresholds include pitch angle ≤ yaw rate ≤ , load ≤ The deviation between the calculation instruction and the threshold. ; S43, if Then a penalty signal is generated. and will issue a punishment signal Feedback is sent to the deep reinforcement learning DRL model of S3 via the backpropagation channel; S44. When the deep reinforcement learning (DRL) model receives a penalty signal When this happens, the strategy network parameters are updated, and the corrected pitch PID coefficients and yaw angle control commands are re-output. S45, if continuous If the output pitch PID coefficient and yaw angle control command still exceed the safety domain, the model will be loaded with the historical optimal strategy until the output pitch PID coefficient and yaw angle control command fall into the safety domain, and the final pitch PID coefficient and yaw angle control command with the "safety verification passed" flag will be generated.

[0011] In the preferred embodiment, step S5 is as follows: S51. If the final pitch PID coefficient and yaw angle control command with the "safety verification passed" flag is received, and the communication module detects that the OPC UA Pub / Sub channel status is normal, the command priority determination is triggered; if the pitch PID coefficient and yaw angle control command type is "emergency adjustment", a real-time control channel is allocated; if the pitch PID coefficient and yaw angle control command type is "normal adjustment", a normal control channel is allocated. S52, then the communication module is triggered to transmit instructions to the main control PLC according to priority, and at the same time the instruction transmission confirmation mechanism is started. After the main control PLC receives the instruction... The system will send a "command received" signal within a specified time; if the timeout occurs, a retransmission will be triggered. When the main control PLC receives the command and has no faults in its self-test, it will trigger the actuator to start the pitch and yaw angle adjustment actions.

[0012] In the preferred embodiment, step S6 is as follows: When the data acquisition module receives the actual operating data fed back by the main control PLC actuator, the actual operating data includes the actual pitch angle, actual yaw angle, and generator power, and the data timestamp is less than the S5 command execution time deviation. When the deviation calculation module is activated, it is triggered. The time deviation threshold for data correlation; Calculate the relative deviation between the actual data and the S3 output command value. Set the deviation threshold ,like If the biased data is marked as a "valid training sample", it is input into the experience pool of the deep reinforcement learning DRL model in S3; when the number of samples in the experience pool is ≥ The bar or model iteration counter reaches At this time, the deep reinforcement learning (DRL) model is triggered to start online iterative optimization, updating the policy parameters of the DRL model; among which... This represents the number of samples in the experience pool. This represents the number of iterations.

[0013] In the preferred embodiment, in step S32, when the deep reinforcement learning (DRL) model training iteration counter shows that it has reached the specified value... The experience pool detected that the number of stored samples was ≥ When a condition is met, the experience replay module is triggered to randomly draw from the experience pool. Samples, sample format is ,in, This represents the operating parameters output by S2 at time t; This represents the control command output by the model at time t; This represents the reward value fed back by S6 at time t; This represents the operating parameters output by S2 at time t+1; The number of iterations; To store the number of samples; The number of randomly selected samples; When sample extraction is complete and the target network detects that its parameters have not been updated ≥ At this time, the target network parameters are updated, and the current network parameters are updated. Copy to target network ,in, This is represented as the parameter matrix of the current network; This is represented as the parameter matrix of the target network; This represents the number of times its own parameters have not been updated. Then, the target Q-value is calculated, and the network Q-value is compared with the target Q-value. Calculate the loss and update the current network parameters using gradient descent; where, This is represented as the Q-function of the current network; Set reward weight coefficient threshold When extreme wind conditions are detected, trigger Adjusted to Prioritize ensuring stable rotational speed.

[0014] In the preferred embodiment, a threshold number of times the TD error stabilizes is set in step S33. TD error decay trigger threshold Threshold for the number of TD error fluctuations The trigger threshold for increased exploration rate ; When the absolute value of TD error is continuous less than At this time, the exploration rate decay mechanism of the deep reinforcement learning (DRL) model is triggered: the exploration rate decreases from... linear decay to The decay step size is / iterations, where This is expressed as the exploration rate after decay; This represents the decay step size of the exploration rate; When the absolute value of TD error is continuous Next > At that time, the exploration rate increase mechanism is triggered: the exploration rate increases from... linear increase to Increase step size to / iteration, simultaneously triggering empirical pool sample diversity verification; if sample coverage < Uncovered operating condition types > If a sample is found, it will trigger the download of supplementary samples from the cloud; in, This represents the step size for increasing the exploration rate; This is expressed as a threshold for sample coverage. The threshold is represented by the uncovered operating condition type; supplementary samples include extreme wind conditions and fault condition samples.

[0015] This invention provides a wind turbine end-to-end adaptive control method based on dynamic sensing and co-evolutionary DRL. Through the coordination of the above-mentioned structures, it has the following advantages compared with existing methods: First, it can accurately capture dynamic operating conditions and match control requirements in real time, effectively solving the problem of dynamic response lag in traditional static optimization models. By collecting dynamic data such as wind speed turbulence, wind shear, and tower shadow effect in real time through a multimodal sensor array, and combining it with physical knowledge-enhanced feature extraction to uncover the essential changes in operating conditions, it dynamically outputs pitch and yaw control parameters through a co-evolutionary DRL model. Therefore, it can adapt to rapid operating condition switching from stable wind to turbulent wind in real time, avoiding control command lag or over-adjustment, and ensuring that pitch and yaw are always synchronized with dynamic wind conditions.

[0016] Secondly, it can dynamically balance strategy exploration and stable control. By introducing a dual-commentator mechanism and TD error regularization into the DRL model, and combining a self-penalty mechanism to dynamically adjust the priority of control commands, instead of using a preset and fixed priority rule, it can avoid the strategy from getting stuck in local convergence and sticking to the local optimum under dynamic disturbances such as extreme wind conditions and instantaneous equipment failures. It can also prevent core commands and non-core commands from competing for transmission resources, reduce mechanical load impact, and ensure the stability of unit operation.

[0017] Third, it achieves synergy between edge real-time control and cloud-based global optimization to improve the overall operating efficiency of wind farms. The co-controller acts as an edge node to undertake local DRL inference and real-time control, while the cloud aggregates data from multiple wind turbines through federated learning to iteratively train the meta-controller. This ensures both the real-time control of individual units and the optimization of model parameters using global data from the wind farm, thereby achieving cluster-level control performance improvement.

[0018] Fourth, by adopting standardized communication interfaces and modular architecture, there is no need to modify the original PLC main control system hardware and core program. AI intelligent control functions can be integrated simply by adding a co-controller, rather than replacing the entire control hardware. Therefore, it can significantly reduce the risk of retrofitting existing wind farms and downtime, while being compatible with wind turbines of different power levels, significantly improving the flexibility and economy of technology implementation. Attached Figure Description

[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the invention; Figure 2 This is a structural diagram of the method framework of the present invention. Detailed Implementation

[0020] To better understand the purpose, system architecture, and functional implementation of this embodiment, the embodiments and features in the embodiments of this application can be combined with each other without conflict. The exemplary embodiments disclosed in this application will be described below with reference to the accompanying drawings, which include specific technical details disclosed in this embodiment to aid understanding; however, these details should be considered exemplary rather than restrictive. Therefore, those skilled in the art should understand that various improvements and adjustments can be made to the embodiments described herein without departing from the scope and core ideas of the invention. Similarly, for clarity, detailed descriptions of well-known technologies, functions, and structures (such as standard image processing algorithms and common communication protocols) are omitted in the following description.

[0021] In the field of wind power control technology, intelligent control methods are widely used to improve the operating efficiency and stability of wind turbines. By analyzing and processing real-time operating data, control parameters can be optimized and adjusted. In complex wind conditions, control schemes combining edge computing and artificial intelligence algorithms can significantly reduce data transmission latency and improve adaptability to dynamic environments. For example, related technologies have developed wind turbine control methods based on edge processors for FFT analysis and fault diagnosis. By performing data preprocessing and local decision-making at the turbine level, reliance on cloud platforms is reduced.

[0022] With the development of wind power generation technology, wind turbine control is gradually evolving from static parameter control to adaptive intelligent control to cope with the strong randomness and dynamism of natural wind conditions. Control schemes integrating deep reinforcement learning (DRL) and multimodal sensing have become a development trend. These schemes attempt to achieve dynamic adjustment of core aspects such as pitch and yaw through model training and real-time data-driven approaches. However, dynamic factors such as wind speed turbulence, wind shear, and tower shadow effects, as well as dynamic disturbances such as extreme wind conditions and instantaneous equipment failures, place higher demands on the real-time response and global optimization capabilities of the control model, making it difficult for traditional adaptive strategies to fully adapt.

[0023] When implementing intelligent control of wind turbines, related technologies face several limitations. Firstly, control models often rely on fixed PID parameters or static optimization logic. Even with the introduction of AI algorithms, issues such as unreasonable computing power allocation and rigid model architecture prevent them from adapting to the dynamic response changes of core control components under rapid changes in operating conditions. Secondly, the priority rules of adaptive strategies are often preset and fixed, failing to consider the dynamic adjustment of control command priorities under dynamic disturbances, making them prone to local convergence or strategy instability. Therefore, the control methods in these technologies have certain limitations and are ill-suited to the comprehensive requirements of wind turbine control accuracy, stability, and global optimization capabilities under complex dynamic wind conditions.

[0024] Example 1 like Figures 1-2 As shown, a wind turbine end-to-end adaptive control method based on dynamic perception and co-evolutionary DRL includes the following steps: S1. When the wind turbine generator completes its startup self-test or the preset operating condition monitoring cycle. Upon arrival, the multimodal sensor array is triggered to start data acquisition. Following preset acquisition rules, the multimodal sensor array collects data on the wind turbine's operating status and ambient wind conditions. After acquisition is complete, the data preprocessing module performs preliminary verification of the collected operating status and ambient wind condition data. The verification standard is data integrity ≥ Furthermore, there were no outliers exceeding the physical range, and the output was multi-source data that passed verification; among which, This is the threshold for data integrity.

[0025] Specifically, the self-test results showed that all sensor communication was normal, the main control PLC had no fault codes, and the power grid connection status was normal. The multimodal sensor array includes a high-frequency lidar, an inertial measurement unit, acoustic sensors, vibration sensors, accelerometers, inclinometers, gyroscopes, and ultrasonic probes. The accelerometers are deployed on the gearbox housing, the inclinometers are deployed at the blade roots, the gyroscopes are deployed on the top of the nacelle, and the ultrasonic probes are deployed in the middle of the tower.

[0026] Operational status and environmental wind condition data include SCADA system data, vibration sensor data, meteorological sensor data, wind speed profile, blade vibration, tower displacement, and aerodynamic load data; High-frequency lidar collects wind speed profiles, inertial measurement units collect blade vibrations, acoustic sensors collect aerodynamic noise, and vibration sensors collect tower displacement. In this embodiment, the specific operating logic of the multimodal sensor array is as follows: S11. When the wind turbine generator starts its self-test, it triggers sensor fault diagnosis, sends test signals to each sensor, and receives feedback response time < This is normal; S12. After the self-test passes, the accelerometer detects that the gearbox speed is greater than... Start Sampling, the inclinometer is used when the blade pitch angle is greater than... Start Sampling is initiated when the gyroscope is activated during yaw system startup. Sampling, ultrasonic probe under tower vibration > Start sampling; S13. When any sensor detects an abnormality in its own data, it triggers the activation of redundant sensors and sends a "sensor abnormality" warning to the AI ​​co-controller. in, This is expressed as the maximum tolerable response time of the sensor feedback; This represents the rotational speed threshold at which the accelerometer starts sampling; This is expressed as the sampling frequency of the accelerometer; This represents the angle threshold at which the inclinometer starts sampling; This is expressed as the sampling frequency of the gyroscope; This is expressed as the sampling frequency of the inclinometer; This represents the vibration threshold at which the ultrasonic probe initiates sampling. This is expressed as the sampling frequency of the ultrasonic probe; This represents the threshold for abrupt changes in accelerometer data. It should be noted that the AI ​​co-controller is an independent control unit specifically designed to support AI intelligent algorithms in wind turbine generators and assist traditional main control PLCs (Programmable Logic Controllers). It serves as both the AI ​​computing power carrier and the collaborative control hub. It neither replaces the basic safety logic control functions of the main control PLC nor relies on the computing power resources of the main control PLC. Instead, it forms a "division of labor and collaboration" dual control system with the main control PLC through standardized interaction. The AI ​​co-controller adopts an industrial-grade embedded computer architecture, integrating GPUs (such as NVIDIA Jetson AGX Xavier) or TPUs (such as Google Coral Dev Board) with high parallel computing capabilities. It is specifically designed to support computationally intensive tasks such as deep reinforcement learning (DRL) model inference, multimodal data feature extraction, and federated learning local training. This solves the problem that traditional main control PLCs, which are primarily focused on logic control and have limited CPU computing power, cannot efficiently run AI algorithms.

[0027] The AI ​​co-controller has a built-in real-time computing power monitoring unit that can dynamically collect the utilization of CPU, GPU / TPU, and avoid resource contention caused by multiple concurrent AI tasks through a preset dynamic computing power allocation strategy.

[0028] The AI ​​co-controller also has a built-in "self-punishment mechanism module" that specifically verifies the safety of control commands output by the AI ​​model: it calls a preset safety threshold library to calculate the deviation between the command and the threshold; if the command exceeds the safety domain, it generates a penalty signal to feed back to the DRL model to correct the command until it falls into the safety domain; in extreme scenarios (N1 consecutive command exceedances), it automatically loads the best historical strategy to prevent the main control PLC from executing dangerous commands.

[0029] The AI ​​co-controller interacts with the main PLC in a non-intrusive co-control mode. It communicates with the main PLC through the OPC UAPub / Sub standardized interface and only sends the "modified final control instruction" (with a "safety verification passed" flag) to the main PLC. It does not change the original basic logic of the PLC, thus solving the compatibility risk of adding AI functions to the PLC program in traditional solutions. The AI ​​co-controller also receives instruction execution data from the main control PLC for online iterative optimization of the DRL model.

[0030] The AI ​​co-controller can serve as an edge node for federated learning in wind farms. After training the DRL model locally, it only uploads the feature extraction layer parameters (without disclosing the complete model or sensitive data) to the cloud meta controller, thus protecting data privacy. It receives the lightweight model after distillation from the cloud, updates the local model, and realizes the collaboration from cloud-based global optimization to edge-based real-time control, so as to better match the optimization needs of wind farm clusters.

[0031] Specifically, based on a sampling mechanism triggered by both sensor status and equipment operating conditions, redundant data or missing key data caused by traditional fixed-frequency sampling is avoided, and the reliability of the perception layer is improved by combining redundancy switching. S2. When the verification of the multi-source data output by S1 is received, and the data heterogeneity detection module detects that the data is heterogeneous, the physical knowledge-enhanced convolutional encoder is triggered to start the feature extraction process. First, the collected data is input into the data alignment module, and the low-resolution data is unified to the preset high resolution through linear interpolation. The linear interpolation formula is shown in the following formula (1): (1); In equation (1), Represented as the target time; , This represents the adjacent sampling times of the S1 output; , Represented as , The collected data value at that moment; Subsequently, fluid dynamic constraints are embedded to extract physical constraint features. The Bernoulli equation is shown in equation (2) below: (2); In equation (2), Expressed as fluid pressure; This is expressed as real-time air density, collected by meteorological sensors. This represents the wind speed output by S1; Expressed as gravitational acceleration; This is represented as the sensor installation height; Represented as a constant; Simultaneously, the attention mechanism module is triggered, and the weight coefficients are calculated based on the importance scores of each sensor data output by S1. The data weights are dynamically allocated, and the weight coefficients are calculated as shown in the following formula (3): (3); In equation (3), Represented as the first Weighting coefficients for sensor-like data; Represented as the first Importance rating of sensor data; Represented by the total number of sensor data categories; vibration data score ≥ SCADA data ≥ ; The final output includes the operating condition type, dynamic characteristic parameters, and confidence level; the operating condition type includes steady wind, turbulent wind, and extreme wind; the dynamic characteristic parameters include the wind speed change rate and load fluctuation amplitude.

[0032] When the working condition classification module collects extreme working condition samples and the generator of the generative adversarial network (GAN) is in a ready state (without training labels), the generator is triggered to start sample synthesis. The generator loss function of GAN is shown in the following equation (4): (4); In equation (4), Represented as generator loss value; This is expressed as the discrimination probability function of the discriminator; This represents the synthesized sample output by the generator; The discriminator loss function is shown in equation (5) below: (5); In equation (5), Represented as the discriminator loss value; This represents a real extreme working condition sample collected by S1; This represents the probability that the discriminator classifies a real sample; This represents the discrimination probability of the discriminator against the synthesized sample; it is then added to the sample database after synthesis. Specifically, the operational logic for "embedding fluid dynamics constraints for feature extraction" in step S2 is as follows: S221. When the data alignment module of S2 outputs data with a uniform resolution, and the physical constraint module detects that the fluid dynamics parameters have been updated to the latest version, the parameter substitution of the Bernoulli equation is triggered. S222, Extract real-time air density from the meteorological sensor data of S1. Sensor installation height Calculate fluid pressure With wind speed The associated features; S223. Subsequently, the associated features are concatenated with the vibration and load data from S1, and then input into a convolutional encoder for feature extraction. The kernel size of the convolutional encoder is [missing information]. The activation function is ReLU; This is expressed as the size parameter of the convolution kernel; S224. When the encoder output feature map dimension is... And the characteristic variance < At that time, the attention mechanism module is triggered to assign weights, where, This is represented as the dimension of the feature map; The threshold is represented as the feature variance; S225. Finally, output the standardized working condition feature vector; Specifically, based on the heterogeneity of multimodal data, physical knowledge is dynamically triggered for embedding, and the physical rationality of features is enhanced through fluid dynamics constraints, thus solving the problem of feature distortion caused by neglecting physical laws in traditional data fusion. S3. Load the pre-trained intelligent control model, input dynamic feature parameters to generate optimized control commands, and correct the commands through a safety verification mechanism. The specific steps are as follows: S31. When the dynamic feature parameters and confidence level output by S2 are received, and the AI ​​co-controller detects that its own CPU utilization is < When GPU / TPU resources are idle (no other high-priority tasks), the pre-trained deep reinforcement learning (DRL) model is loaded. The S32 and DRL models are DQN models based on graph neural networks (GNN). After loading, the operating condition type output by S2 is read first. If the operating condition type changes from steady wind to turbulent wind, the model parameter configuration is switched. S33. Subsequently, the model integrates a regularization mechanism driven by dual critics and TD error, and embeds Lyapunov stability constraints. The loss function is shown in the following equation (6): (6); In equation (6), This is expressed as the total loss value of the model; Represented as the policy loss value; Represented as stability weights (initial value is...) ); Represented as Lyapunov function values; S34. Input dynamic characteristic parameters for inference, and output variable pitch PID coefficients ( ) and yaw angle control commands; Specifically, the operational logic of the "DRL model experience replay mechanism" in step S32 is as follows: S321. When the DRL model training iteration counter of S3 reaches... The experience pool detected that the number of stored samples was ≥ When a condition is met, the experience replay module is triggered to randomly draw from the experience pool. Samples, sample format is ,in, This represents the operating parameters output by S2 at time t; This represents the control command output by the model at time t; This represents the reward value fed back by S6 at time t; This represents the operating parameters output by S2 at time t+1; The number of iterations; To store the number of samples; The number of randomly selected samples; S322, When sample extraction is complete and the target network detects that its own parameters have not been updated ≥ times At this time, the target network parameters are updated, and the current network parameters are updated. Copy to target network ,in, This is represented as the parameter matrix of the current network; This is represented as the parameter matrix of the target network; This represents the number of times its own parameters have not been updated. S323. Then calculate the target Q value, as shown in the following formula (10): (10); In equation (10), Represented as the target Q value; This is represented as the reward value at time t; Represented as a discount factor; Represented as candidate actions at time t+1; Represented as the Q-function of the target network; These are the operating parameters at time t+1; S324. Based on the target Q-value and the current network Q-value Calculate the loss and update the current network parameters using gradient descent; where, This is represented as the Q-function of the current network; The reward function is shown in equation (11) below: (11); In equation (11), Represented as reward value; , This is represented as a reward weighting coefficient; This is expressed as the actual power generated by the S6 feedback; This is represented as a preset power loss threshold; This is expressed as the actual rotor speed fed back by S6; This represents the preset wind turbine speed setting value; S325, Preset reward weight coefficient threshold When extreme wind conditions are detected, trigger Adjusted to Prioritize ensuring stable rotational speed; The specific operational logic of "dual critics and TD error regularization" in step S33 is as follows: S331. When the DRL model performs Q-value calculation, the dual-commentator architecture is triggered to run in parallel: Commentator 1 calculates based on dynamic feature parameters ,in, Represented as the Q-function of commentator 1; This is represented as the parameter matrix for commentator 1; commentator 2 calculates based on the same parameters. ,in, Represented as the Q function of commentator 2; Represented as the parameter matrix of commentator 2 Pick and The minimum value of is taken as the target Q value, and the calculation formula is shown in the following formula (16): (16); In equation (16), Represented as the target Q value; S332. Then calculate the TD error, and the calculation formula is shown in the following formula (17): (17); In equation (17), This is expressed as TD error; This is represented as the Q-value of the current network; Set the threshold for TD error ,like If so, the regularization mechanism will be triggered; The L2 regularization term is applied to the policy network parameters, and the calculation formula is shown in equation (18) below: (18); In equation (18), Represented as a regularization term; Represented as the regularization coefficient; It is represented as the parameter matrix of the policy network; Adjust the learning rate from Down to ,in, This is represented as the learning rate after regularization; S333, when And maintain In the next iteration, the regularization mechanism is disabled, and the initial learning rate is restored. This is expressed as the threshold number of iterations for TD error stabilization; In this embodiment, the specific steps in step S332 are as follows: S3321. Set a threshold number of times the TD error stabilizes. TD error decay trigger threshold Threshold for the number of TD error fluctuations The trigger threshold for increased exploration rate ; When the absolute value of TD error is continuous less than At this time, the exploration rate decay mechanism of the DRL model is triggered: the exploration rate decreases from... linear decay to The decay step size is / iterations, reducing unnecessary strategy exploration, where This is expressed as the exploration rate after decay; This represents the decay step size of the exploration rate; S3322, When the absolute value of the TD error is continuous Next > At that time, the exploration rate increase mechanism is triggered: the exploration rate increases from... linear increase to Increase step size to / iteration, simultaneously triggering empirical pool sample diversity verification; if sample coverage < Uncovered operating condition types > If a certain number of samples are found, it will trigger the download of supplementary samples from the cloud to enhance the global search capability of the strategy. in, This represents the step size for increasing the exploration rate; This is expressed as a threshold for sample coverage. The threshold is represented by the uncovered operating condition type; supplementary samples include extreme wind conditions and fault operating condition samples. Specifically, the exploration rate is dynamically adjusted based on the TD error trend, and combined with sample diversity verification, to avoid premature strategy convergence or resource waste caused by a fixed exploration rate, and to achieve a dynamic balance between exploration and utilization.

[0033] In this embodiment, the specific operating logic of the "Lyapunov stability constraint" in step S33 is as follows: when the DRL model embeds the stability constraint, the Lyapunov function calculation is triggered: based on the load fluctuation amplitude output in S2. Speed ​​deviation Construct the function formula: ,in, Represented as Lyapunov function values; , Represented as weighting coefficients; This is expressed as the load fluctuation amplitude output by S2; The wind turbine speed deviation is represented by the output of S2. ; If the calculation yields If the system is unstable, then the system is unstable. Expressed as the rate of change of the Lyapunov function; It is expressed as the rate of change of the system state vector; Represented as a system state vector; If it is represented as a positive definite matrix, then the stability weights are triggered. from Upgraded to ,in, This represents the stability weight when the system is unstable, while also limiting the scope of policy exploration, allowing exploration only within the set of historical security actions; when And maintain Triggered during the next iteration Restore to Remove the restrictions on the scope of exploration; among them, This represents the threshold number of iterations required for system stability.

[0034] S4. The specific steps for correcting the variable pitch PID coefficient and yaw angle control command through the self-penalty mechanism are as follows: S41. When the variable pitch PID coefficient and yaw angle control command output by S3 are received, and the safety assessment module detects that the command does not have the "safety verification passed" flag, the self-punishment mechanism module is triggered to start the safety assessment. S42. First, call the preset safety threshold library. Safety thresholds include pitch angle ≤ yaw rate ≤ , load ≤ The deviation between the calculation instruction and the threshold is shown in the following formula (7): (7); In equation (7), This is expressed as the deviation between the instruction value and the safety threshold; This represents the instruction value output by S3; This is represented as a preset safety threshold; S43, if Then a penalty signal is generated. As shown in equation (7): (8); In equation (8), Represented as a penalty signal value; This is represented as a penalty coefficient, taken under extreme wind conditions. , taken under normal wind conditions ; and punishment signal Feedback is sent to the deep reinforcement learning DRL model of S3 via the backpropagation channel; S44. When the model receives a penalty signal, it triggers the update of the strategy network parameters and re-outputs the corrected pitch PID coefficients and yaw angle control commands. S45, if continuous If the output pitch PID coefficient and yaw angle control command still exceed the safety domain, the model will be loaded with the historical best strategy until the output pitch PID coefficient and yaw angle control command fall into the safety domain, and the final pitch PID coefficient and yaw angle control command with the "safety verification passed" flag will be generated. In this embodiment, the specific operational logic of the "security assessment of the self-punishment mechanism" in step S4 is as follows: when the self-punishment mechanism module generates a punishment signal... When the data is fed back to the DRL model in S3, the model's policy update module detects the "urgency" flag of the penalty signal, where extreme wind conditions are marked as "high urgency" and normal wind conditions are marked as "low urgency". If the condition is "high emergency," the model will pause regular exploration and focus on correcting the pitch PID coefficients and yaw angle control commands, thereby reducing the exploration rate. ,in, This is represented as the strategy exploration rate under high-urgent conditions; If it is "low urgency", then maintain the exploration rate. At the same time, the weight of the load constraint term in the loss function is increased; After the model outputs the corrected pitch PID coefficient and yaw angle control command, the secondary verification of S4 is triggered. If the verification passes, a "safety verification passed" flag is generated; otherwise, the penalty and correction process is repeated. When the number of penalties detected is ≥ At this time, the AI ​​co-controller sends a "model abnormality" warning to the main control PLC and switches to the backup control strategy. By constructing a three-tiered safety control closed loop of "punishment, correction, and fault tolerance," the exploration strategy is dynamically adjusted according to the degree of urgency to avoid the rigidity of the strategy caused by a single correction mechanism. At the same time, backup strategies are introduced to ensure the continuity of control in extreme scenarios.

[0035] S5. Using an event-driven communication framework, the corrected pitch PID coefficients and yaw angle control commands are transmitted to the main control PLC for execution according to priority, triggering the wind turbine to perform pitch and yaw angle adjustments. The specific steps are as follows: S51. If the final pitch PID coefficient and yaw angle control command with the "safety verification passed" flag is received, and the communication module detects that the OPC UA Pub / Sub channel status is normal, the command priority determination is triggered. If the pitch PID coefficient and yaw angle control command type are "emergency adjustment", then a real-time control channel is allocated. If the pitch PID coefficient and yaw angle control command type are "normal adjustment", then a normal control channel will be assigned. S52, then the communication module is triggered to transmit instructions to the main control PLC according to priority, and at the same time the instruction transmission confirmation mechanism is started. After the main control PLC receives the instruction... The system will send a "command received" signal within a specified time; if the timeout occurs, a retransmission will be triggered. When the main control PLC receives the command and its self-test is fault-free, the actuator will be triggered to start the pitch and yaw angle adjustment actions. In this embodiment, the specific operating logic of the "event-driven communication framework" in step S5 is as follows: when the communication module triggers instruction transmission, if a sudden blockage of the OP CUA Pub / Sub channel is detected, the channel switching mechanism is triggered to switch to the backup wireless channel. The AI ​​co-controller has a pre-installed OpenFAST simulation module. Simultaneously triggering the OpenFAST simulation module simulates the execution effect of commands based on the S2 operating parameters. If the simulation result shows the load ≤ If so, the command is confirmed to be valid and marked as "Simulation verification passed". This is expressed as the load safety threshold verified by simulation. When the main control PLC receives the final pitch PID coefficient and yaw angle control command with "Safety Verification Passed" and "Simulation Verification Passed" indicators, it triggers the actuator to start closed-loop control, collects execution data in real time, and checks the deviation from the command value. Time fine-tuning, among which, This is represented as the deviation fine-tuning threshold for closed-loop control. If the communication link is interrupted > This triggers the AI ​​co-controller to temporarily store instructions. Once the link is restored, the stored instructions will be transmitted first. This represents the maximum tolerance time for link interruption; Specifically, the above embodiments involve parallel triggering of instruction transmission and simulation verification, combined with channel switching and breakpoint resume mechanism, to solve the defect of traditional communication that only focuses on transmission reliability and ignores the effectiveness of instruction execution; Furthermore, the specific operational logic of "dynamic bandwidth allocation" in step S5 is as follows: when the communication module detects that the total bandwidth utilization of the wind farm network is greater than 100%, the system will automatically allocate bandwidth to the wind farm network. At that time, a bandwidth priority reallocation is triggered: the proportion of real-time control channel bandwidth changes from... Upgraded to Model update channel from Down to State feedback channel maintained ; Simultaneously trigger temporary storage of non-critical data, pending bandwidth utilization < Transmit again later; When the command transmission delay of a single wind turbine is detected to be greater than When this occurs, the wind turbine is triggered to exclusively occupy a portion of the bandwidth (extra allocation). ), to ensure that instructions arrive in a timely manner; in, This is expressed as the redistribution threshold for bandwidth utilization; This represents the proportion of the real-time control channel after bandwidth reallocation. This represents the proportion of channels updated in the initial model. This represents the proportion of the updated channel in the model after bandwidth reallocation; This is represented as a fixed percentage of the state feedback channel; This represents the bandwidth idle threshold for non-critical data transmission. This represents the tolerance threshold for instruction transmission delay; This represents the percentage of additional bandwidth allocated. Specifically, this embodiment focuses on dynamic bandwidth scheduling based on network load and instruction priority, avoiding the delay of critical instructions caused by traditional fixed bandwidth allocation, while temporarily storing non-critical data to balance bandwidth requirements.

[0036] S6. When the data acquisition module receives the actual operating data fed back by the main control PLC actuator, including the actual pitch angle, actual yaw angle, and generator power, and the data timestamp deviates from the execution time of the S5 command by less than 1 / 3. When this happens, the deviation calculation module is activated; Calculate the relative deviation between the actual data and the S3 output command value. As shown in equation (9): (9); In equation (9), This represents the relative deviation between the actual operating data and the command value; "Measured value" represents the actual operating data fed back by the main control PLC actuator; "Command value" represents the control command value output by S3. Set deviation threshold ,like If the biased data is marked as a "valid training sample", it is input into the experience pool of the deep reinforcement learning DRL model in S3; when the number of samples in the experience pool is ≥ The bar or model iteration counter reaches At this time, the DRL model is triggered to start online iterative optimization and update the model policy parameters; among which... This represents the number of samples in the experience pool. This represents the number of iterations.

[0037] In this embodiment, step S7 is also included: when the number of DRL model iterations in S6 reaches... Once, or weather sensor continuously When extreme wind conditions are detected over a long period of time, the dynamic model switching module is triggered, in which... This is represented by the number of model iterations. This is expressed as the monitoring duration; First, the differential transformation module is called to establish the system state equation based on the dynamic characteristic parameters of S2, as shown in equation (12) below: (12); In equation (12), This is represented as a system state vector, which includes the wind turbine speed, key component loads, and tower displacement. Represented as time; Represented as a state matrix; Predicting the future Duration-frequency response, where, This is expressed as the predicted duration of the frequency response; If the predicted frequency response exceeds the safety threshold This triggers a switch from "normal power generation control mode" to "frequency support mode," with the pitch rate changing from... Adjusted to Yaw accuracy requirements from ± Relaxed to ± ; When the frequency response returns to the safe threshold and remains After a certain period of time, the system will switch back to "normal power generation control mode" and upload the mode switch record to the cloud log. in, This is expressed as the safety threshold for frequency response; This represents the pitch rate in normal mode. This is expressed as the pitch rate in frequency support mode; This represents the yaw accuracy in normal mode. This is expressed as yaw accuracy in frequency support mode; This is expressed as the duration of stable frequency response; Specifically, the early mode switching based on frequency response prediction avoids the control instability caused by the reliance on lag feedback in traditional mode switching, while dynamically adjusting the control accuracy index to balance safety and power generation efficiency. In this embodiment, step S8 is also included: when the wind farm cloud management platform detects ≥ The S3 model of the typhoon generator has reached the number of iterations. Secondly, the network connectivity of each wind turbine AI co-controller is normal (latency < When this occurs, the federated learning framework is triggered to start; in, This represents the threshold number of wind turbines that trigger federated learning. This represents the threshold number of iterations for a single wind turbine model; This represents the maximum tolerable network latency. First, the cloud sends a "parameter upload command" to each AI co-controller. After receiving the command, the AI ​​co-controller triggers the extraction of feature extraction layer parameters (not complete model parameters). Policy layer parameters need to be excluded to protect the local optimization strategy, and then uploaded to the cloud through an encrypted channel (AES-256 encryption). When the cloud has received all edge node parameters (reception completion rate ≥ When, among them, This is represented as the parameter reception completion rate threshold, triggering a two-stage single-round training: The first stage starts the spatiotemporal memory-enhanced autoencoder. The encoder formula is shown in equation (13) below: (13); In equation (13), This is represented as the feature vector output by the encoder; Represented as an encoding function; This is represented as local feature data input; Represented as the ReLU activation function; Represented as an encoding weight matrix; Represented as an encoding bias vector; Global fault features are extracted using the decoder formula, as shown in equation (14) below: (14); In equation (14), This represents the reconstructed data output by the decoder; Represented as a decoding function; Represented as a decoded weight matrix; Represented as a decoding bias vector; The second stage trains the meta-controller strategy, using a federated averaging algorithm to aggregate the parameters of each node; after training, the model distillation module is triggered, and the knowledge distillation loss function is shown in the following equation (15): (15); In equation (15), Expressed as distillation loss value; , Represented as loss weight coefficient; Represented as the cross-entropy loss function; This is represented as the output probability distribution of the student model (lightweight model); Represented as real labels (based on the fault sample library); Represented as the KL divergence function; This is represented as the output probability distribution of the teacher model (meta-controller); The meta-controller strategy is compressed into a lightweight model with fewer than 100 parameters. ,in, The parameter thresholds for the lightweight model are distributed to each edge node. When an edge node receives a lightweight model and detects that the local model version is lower than that in the cloud, it triggers a model update and backs up the old model to the local cache. Specifically, this embodiment combines federated learning and knowledge distillation in a cloud-edge co-evolution model. It uploads only feature layer parameters to protect data privacy, and uses lightweight model compression to adapt to edge computing power, thus resolving the contradiction between privacy protection and real-time edge performance that traditional cloud training cannot balance.

[0038] Furthermore, the specific operational logic of "cloud meta controller deployment" in step S8 is as follows: When the cloud receives the feature parameters uploaded by each edge node, it triggers Kubernetes (K8s) containerization scheduling: First, create the meta controller Deployment and set the initial number of Pod replicas. One, of which This represents the initial number of Pod replicas; resource configuration CPU. Core, memory ,in, This is expressed as the number of CPU cores in a single Pod; This is expressed as the memory capacity of a single Pod. When Pod CPU utilization > continued Over a certain period of time, HPA (Horizontal Pod Autoscaler) will be triggered to automatically scale up to [the specified size]. One copy; When an edge node receives a lightweight model from the cloud, it triggers a local computing power check. If the CPU utilization is < Then immediately initiate the Deep Reinforcement Learning (DRL) model update; if > The update task will then be added to the queue and executed when computing power is available. After the deep reinforcement learning (DRL) model is updated, the edge nodes are triggered to send a "successful update" confirmation to the cloud. The cloud records the update status of each node, forming a closed loop. in, The expansion threshold is expressed as the CPU utilization of the Pod. This represents the duration of time the utilization rate exceeds the limit. This represents the number of Pod replicas after the expansion. This represents the idle computing power threshold for edge nodes to initiate model updates; The innovation of this step lies in containerized deployment based on dynamic scheduling of computing power, combined with a queue mechanism to avoid competition for resources between model updates and local control, thereby achieving lightweight and efficient cloud-edge collaboration.

[0039] In this embodiment, the specific operational logic of "model distillation" in step S8 is as follows: After the cloud-based meta-controller has been trained, the teacher model (meta-controller) is loaded, and its output probability distribution formula is extracted: ,in, This is represented as the output probability distribution of the teacher model; This is represented as the logits output of the teacher model; This is expressed as a temperature parameter; subsequently, the student model (lightweight model) is loaded, and the same global features are input to obtain the student probability distribution formula: ,in, This is represented as the output probability distribution of the student model; This is represented as the logits output of the student model; Formula for calculating the loss from triggered distillation: The meanings of each parameter are the same as those described in formula (15); Optimize student model parameters using gradient descent; when the difference in fault identification accuracy between the student model and the teacher model is < At that time, the distillation process is terminated, the student model parameters are compressed, and quantization techniques are used to extract them from the data. 2-bit floating-point conversion This ensures that the number of parameters is reduced compared to the teacher model. ; in, The threshold representing the difference in accuracy; This represents the number of bits in the parameter before compression; This is expressed as the number of bits in the compressed parameter; This is expressed as the threshold percentage for the reduction in the number of parameters. Specifically, this step involves knowledge distillation with temperature parameters adaptively, combined with quantization compression to achieve model lightweighting, and at the same time, controlling the distillation quality through an accuracy difference threshold to solve the problem of losing key strategy information in traditional distillation.

[0040] When the oil contamination level detected by the oil sensor deployed in the gearbox oil sump is > ,in, This represents the warning threshold for oil contamination levels, and is continuous. When the sampling results are consistent, the sampling interval ,in, This represents the threshold number of times the sampling is consistent. This is represented as the sampling interval time, triggering early fault warning: First, a fault code "GearBox_Oil_Pollution" is generated, along with oil temperature and contamination values, and transmitted to the main control PLC and cloud via the S5 communication framework; at the same time, the federated learning special sample labeling of S8 is triggered, marking the fault data as a "high-value sample" and prioritizing its inclusion in the cloud meta-controller training to improve the global model's ability to identify oil faults; The steps in this embodiment involve the linkage labeling of fault data and federated learning. By prioritizing training with high-value samples, the insufficient fault identification accuracy caused by the homogeneity of samples in traditional federated learning is solved.

[0041] Example 2 This embodiment focuses on the training process of a deep reinforcement learning (DRL) model, and elaborates in detail on the model structure construction, training data preparation, iterative optimization and parameter tuning in Embodiment 1 to ensure that the trained model meets the control requirements of stable wind, turbulent wind and extreme wind conditions.

[0042] Specifically, the training data comes from The multi-modal sensor array of the wind turbine generator (conforming to acquisition rules) includes: a high-frequency lidar (model LR100, sampling frequency...). The wind speed profile data collected by the inertial measurement unit (IMU6050, sampling frequency) The blade vibration data collected by the acoustic sensor (model AS200, sampling frequency) The aerodynamic noise data collected and the vibration sensor (model VS300, sampling frequency) were used. The tower displacement data is collected, and the generator power and speed data (sampling frequency) of the SCADA system are acquired simultaneously. ).

[0043] In the preferred scheme, the data preprocessing module executes according to the following logic: First, it performs an integrity check on the collected data. ,Right now Remove outliers that exceed the physical range; This then triggers data alignment, because the time resolution difference between SCADA data and vibration data is as large as 1000 times. The linear interpolation method using formula (1) Interpolate SCADA data to Resolution, where the target time Let this be the sampling time for the vibration data. , These are adjacent sampling times of SCADA data. , The power value at the corresponding moment, and the data timestamp deviation after interpolation. .

[0044] In this embodiment, the physical constraint module embeds fluid dynamics parameters, such as air density. Pick gravitational acceleration Sensor installation height Bernoulli's equation through formula (2) Computational fluid pressure With wind speed The correlation characteristics, when wind speed hour, This feature is then concatenated with vibration and load data to form a 64-dimensional feature vector. .

[0045] In practice, the attention mechanism module is based on sensor data importance scoring and vibration data scoring. SCADA data scoring The weighting coefficients are calculated using formula (3). When the blade vibration data is scored SCADA power data scoring At that time, vibration data weight Ensure the weighting of key dynamic data.

[0046] To address the issue of insufficient samples under extreme operating conditions, i.e., the initial coverage is only... The GAN generator (using the DCGAN architecture) is triggered according to formula (4). , formula (5) Synthetic samples: The generator takes a 100-dimensional random vector as input. Outputs the wind speed simulating extreme wind conditions. Load data, the discriminator uses real samples from historical extreme operating condition records of wind farms. With synthetic samples The probability comparison was used to update the parameters, and after 500 training rounds, the coverage of extreme samples increased to [percentage missing]. ,Right now .

[0047] In one feasible approach, the DRL model adopts the GNNDQN architecture, the specific structure of which includes: Graph Neural Network (GNN) layer: Models the wind turbine system as a dynamic graph, with nodes including blades ( (units), gearboxes, generators, towers (total) Each node has a corresponding edge representing the aerodynamic structural coupling relationship between nodes (such as the tower shadow effect between the blade and the tower), and outputs 32-dimensional node features. Dual-Critic Network: Input GNN features and operating parameters, Critic 1 outputs... Commentator 2 output According to formula (16) Take the minimum value as the target value ; Strategy network: Output variable pitch PID coefficients ( , , The Lyapunov stability constraint is embedded in formula (6) and the yaw angle command. Initial weights Lyapunov function , For load fluctuation amplitude, This refers to the speed deviation.

[0048] The training parameters are set as follows: the experience pool capacity is 10,000 records, i.e. A single sample consists of 32 samples, i.e. The target network updates every 100 iterations, that is... Learning rate , Discount factor Exploration rate , , Reward function weights , That is, under extreme wind conditions .

[0049] In this embodiment, model training is iterated according to the following steps (total number of iterations). Second-rate): 1. Data Input: Input the preprocessed multimodal features (including synthetic samples) into the GNN layer, and output node features; 2. Value calculation: The two critic networks output values ​​separately. , According to formula (16) get Current network output ; 3. Loss calculation: using formula (6) Calculate total loss ,in for and The mean square error, Dynamically adjust based on system stability (when unstable). ); 4. Strategy Update: If TD error ( , formula (17) ), trigger formula (18) L2 regularization ( The policy network parameters are updated through gradient descent. 5. Reward feedback: After each iteration, according to formula (11) Calculate reward value ,in (Power loss threshold) (Speed ​​setting value), when , Then: ; 6. Convergence criterion: When the loss value of 100 consecutive iterations... And the fluctuation range of the reward value At that time, the model training converged, the parameters were fixed (number of parameters: $800,000), and the training time was reduced compared to traditional DQN. .

[0050] Example 3 This embodiment is based on the GNN-DQN model trained and solidified in Embodiment 2, and is applied to stable wind conditions commonly encountered in onshore wind farms during spring. In this embodiment, after the wind turbine generator completes its startup self-test, the operating condition monitoring cycle is triggered. Data collection logic: High-frequency lidar acquisition of wind speed profile: wind speed Wind speed variation rate ; The inertial measurement unit collects data on blade vibration: blade root vibration acceleration 0.3g ( ); Vibration sensors collect tower displacement data: lateral displacement at the top of the tower. ; SCADA system synchronously acquires: generator power Rotation speed .

[0051] Specifically, the data preprocessing module verifies the collected data: data integrity. ( If no outliers are found, the output is the multi-source data that has passed the verification, which will be used as the input for subsequent feature extraction.

[0052] In the preferred scheme, the data heterogeneity detection module determines the current data heterogeneity: SCADA data ( ) and vibration data ( The difference is 1000 times. ), triggering the feature extraction process: 1. Data alignment: via formula (1) Interpolate SCADA data to Resolution, for example hour, ( ), ( After interpolation ; 2. Physical constraint embedding: according to formula (2) Computational fluid pressure Extract the correlation features between pressure and wind speed; 3. Weight allocation: using formula (3) Calculate weights and score vibration data. (Smooth vibration under stable wind conditions), SCADA data score Weight , ; 4. Operating Condition Output: Output operating condition type "Steady Wind", dynamic characteristic parameters (wind speed change rate) Load fluctuation amplitude ), confidence level ( ), while extreme sample coverage ( ), no need to trigger GAN synthesis.

[0053] 3.3-GNN-DQN Model Inference In practice, the AI ​​co-controller's computing power monitoring module detects CPU utilization. ( When GPU resources are idle, model loading is triggered. Operating condition determination: Read the "Stable Wind" type and wind speed change rate output from the previous steps. ( No need to adjust model parameters (exploration rate) Reward weight , ); Stability constraints: according to formula (6) Calculate loss ,in , Total loss ; Command output: Model inference outputs variable pitch PID coefficients ( , , ), yaw angle command ( ).

[0054] The security assessment module detected that the command did not include a "security verification passed" flag, triggering a self-punishment mechanism. 1. Deviation calculation: according to formula (7) The deviation between the calculation command and the safety threshold, yaw rate command ( ), load command Design load ( ), No deviation; 2. Identifier Generation: Generate a "Security Verification Passed" identifier and output the final command. , , Yaw angle ).

[0055] In this embodiment, the communication module detects the packet loss rate of the OPCUA channel. ( ),Delay ( This triggers the following logic: Priority determination: The instruction is "routine adjustment", allocating a normal control channel (bandwidth percentage) ); Transmission confirmation: After the main control PLC receives the instruction... Feedback "command received" signal ( No timeout retransmission; Action executed: The main control PLC controls the pitch system to adjust the pitch angle according to the PID coefficients. The yaw system adjusts the angle to Delayed response from the implementing agency .

[0056] Specifically, the data acquisition module receives actual data from the actuator: the actual pitch angle. Actual yaw angle Actual power Deviation between data timestamp and instruction execution time ( This triggers the following process: 1. Deviation calculation: according to formula (9) Calculate relative deviation ( ), ( ); 2. Sample Judgment: Bias Samples not marked as "valid training samples" are found when the number of samples in the experience pool is insufficient. This does not trigger model iterative optimization. Example 4 This embodiment addresses turbulent wind conditions and, based on the pre-trained model from Embodiment 2, ensures the dynamic adaptability of the control process. In this embodiment, meteorological sensors continuously collect wind speed data. Turbulence intensity 0.25, wind speed variation rate ( ), triggering the data collection logic: High-frequency lidar: wind speed Wind speed variation rate turbulent fluctuation frequency ; Inertial measurement unit: blade root vibration acceleration 1.2g ( ); Vibration sensor: Lateral displacement at the top of the tower ( ); SCADA system: generator power Rotation speed ( ).

[0057] After data preprocessing, output validated data (integrity). ), which serves as the input for subsequent feature extraction.

[0058] In the preferred solution, the following logic applies: 1. Data alignment: via formula (1) After interpolation, the timestamp deviation between SCADA data and vibration data ; 2. Physical constraints: according to formula (2) calculate Load fluctuation amplitude ( ); 3. Weighting: Vibration data scoring (Significant vibration under turbulent wind), according to formula (3) calculate The weights are tilted towards vibration data; 4. Operating Condition Output: Outputs "Turbulent Wind" type and dynamic characteristic parameters (wind speed change rate). Load fluctuation amplitude ), confidence level .

[0059] In practice, the following model parameters will be adjusted (based on operating condition switching): Exploration rate adjustment: From Down to Reduce strategy exploration and avoid command fluctuations under turbulent winds; Reward weighting adjustment: Increased from 0.4 to Prioritize ensuring stable rotational speed ( (close to the set value) Stability constraints: According to formula (6) Pick ( Total loss ; Command output: Model inference outputs variable pitch PID coefficients ( , , ), yaw angle command yaw rate .

[0060] Among them, the safety assessment module detected the yaw rate. ( ), but load command Design load ( ), trigger formula (7) Calculate the deviation According to formula (8) Generate penalty signal (Under normal wind conditions) ); Policy update: The penalty signal is fed back to the model, and the learning rate changes from... Rise to Re-infer and output correction instructions: Load Design load ( ), yaw rate ; Secondary verification: This generates a "safety verification passed" flag, preventing mechanical shocks caused by load exceeding the threshold and reducing the strategy instability rate. .

[0061] In this embodiment, the communication logic is as follows: Channel status: OPCUA channel packet loss rate ( Trigger backup Channel switching, delay ( ); Bandwidth allocation: The command is "normal adjustment", bandwidth percentage. No emergency order was given to seize the position; Execution feedback: Actual pitch angle Actual power Rotation speed .

[0062] Meanwhile, the wind farm cloud detected Typhoon ( The number of model iterations reached Second, network latency ( ), triggering federated learning: 1. Parameter upload: Each edge node (AI co-controller) extracts the parameters of the feature extraction layer (Formula (13)). encoder weights ), encrypted and uploaded to the cloud; 2. Global training: Cloud-based training according to formula (13) , formula (14) The autoencoder is activated to extract global fault features, and the parameters are aggregated using a federated averaging algorithm, thereby improving training efficiency. ; 3. Model Deployment: The distilled lightweight model (500,000 parameters) is deployed to the edge nodes, improving the model's adaptability to turbulent wind conditions after the update. .

[0063] Specifically, according to formula (9) Calculate the deviation: ( ), No iteration is triggered.

[0064] Example 5 This embodiment addresses extreme wind conditions and, based on the pre-trained model from Embodiment 2, ensures control safety and continuity under extreme scenarios. In this embodiment, the meteorological sensor continuously operates for 10 minutes (…). Wind speed was collected. ( ), turbulence intensity 0.35, triggering high-frequency acquisition ( ): High-frequency lidar: wind speed Wind speed variation rate ; Inertial measurement unit: blade root vibration acceleration 2.5g (close to the equipment limit of 3g); Ultrasonic probe (middle of tower): Tower vibration 0.5g ( ),trigger sampling( ); Oil sensor: Oil contamination level ISO4406 class( class).

[0065] Data preprocessing integrity ( ), and output it to subsequent steps.

[0066] In the preferred scheme, extreme sample coverage was detected. ( Triggering GAN synthesis: 1. Feature extraction: according to formula (2) calculate Load fluctuation amplitude ; 2. GAN synthesis: according to formula (4) , formula (5) Generated wind speed Load The sample of the design payload, after synthesis, has an increased coverage rate. ; 3. Operating Condition Output: Outputs "Extreme Wind" type and dynamic characteristic parameters (wind speed change rate). Load fluctuation amplitude ), confidence level .

[0067] In practice, the following logic is triggered: Model parameters: exploration rate Reward weight (Prioritizing speed stability), Lyapunov weighting (The system is close to being unstable); Command output: PID coefficients for pitch control ( , , ), yaw angle command (Stop windward movement during extreme winds); Mode switching: Wind speed detected According to formula (12) Establish the system state equations, where (Speed) Load, Tower displacement, predicting future Second( Frequency response ( This triggers a switch to "Frequency Support Mode"; Pitch rate adjustment: From Rise to ; Wind accuracy relaxed: from to .

[0068] Among them, the pitch rate was detected. ( ), load Design load ( No deviation; simultaneously, the oil sensor continuously... Second-rate( Pollution level collected Level, sampling interval Triggering a fault warning: Fault code: "GearBox_Oil_Pollution" generated, with additional oil temperature. ; Sample labeling: Samples labeled as "high-value samples" are given priority for inclusion in federated learning training, and anomalies are identified 72 hours in advance.

[0069] In this embodiment, the communication logic is as follows: Emergency Order: Wind Speed ​​Exceeds... The instruction is marked "urgent," and a real-time control channel is allocated. ; Simulation verification: Trigger OpenFAST simulation, load Mark "Simulation verification passed"; Action executed: Pitch system presses To stop rowing The yaw system is locked.

[0070] Subsequent steps provide feedback on actual data: pitch angle Rotation speed Frequency response (Regression safety threshold), deviation No iteration.

[0071] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0072] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A wind turbine end-to-end adaptive control method based on dynamic sensing and co-evolutionary DRL, characterized in that, Includes the following steps: S1. Collect the operating status and environmental wind condition data of the wind turbine generator set through a multi-modal sensor array, preprocess the collected operating status and environmental wind condition data, and output the verified multi-source data. S2. Input the verified multi-source data into a physics-enhanced convolutional encoder, embed fluid dynamics constraints for feature extraction, generate working condition types, dynamic feature parameters and confidence levels, and synthesize extreme working condition samples through a generative adversarial network. S3. Input the dynamic feature parameters and confidence scores into the pre-trained deep reinforcement learning DRL model deployed in the AI ​​co-controller. The deep reinforcement learning DRL model integrates the regularization mechanism driven by dual critics and TD error, embeds Lyapunov stability constraints, adaptively switches parameter configurations according to the working condition type, and outputs variable pitch PID coefficients and yaw angle control commands. S4. Correct the pitch PID coefficient and yaw angle control command through a self-punishment mechanism. S5. Using an event-driven communication framework, the corrected pitch PID coefficients and yaw angle control commands are transmitted to the main control PLC for execution according to priority, triggering the wind turbine to perform pitch and yaw angle adjustments. S6. Collect actual operating data of the actuator and feed it back to the deep reinforcement learning (DRL) model in step S3 to achieve online iterative optimization of the strategy.

2. The wind turbine end-to-end adaptive control method based on dynamic perception and co-evolutionary DRL as described in claim 1, characterized in that, In step S1, the multimodal sensor array includes a high-frequency lidar, an inertial measurement unit, an acoustic sensor, a vibration sensor, an accelerometer, an inclinometer, a gyroscope, and an ultrasonic probe. Operational status and environmental wind condition data include SCADA system data, vibration sensor data, meteorological sensor data, wind speed profile, blade vibration, tower displacement, and aerodynamic load data; When the wind turbine generator completes its startup self-test or the preset operating condition monitoring cycle is reached, the multi-modal sensor array is triggered to start data acquisition.

3. The wind turbine end-to-end adaptive control method based on dynamic perception and co-evolutionary DRL as described in claim 1 or 2, characterized in that, The specific operating logic of the multimodal sensor array in step S1 is as follows: S11. When the wind turbine generator starts its self-test, it triggers sensor fault diagnosis, sends test signals to each sensor, and receives feedback response time < This is normal; S12. After the self-test passes, the accelerometer detects that the gearbox speed is greater than... Start Sampling, the inclinometer is used when the blade pitch angle is greater than... Start Sampling is initiated when the gyroscope is activated during yaw system startup. Sampling, ultrasonic probe under tower vibration > Start sampling; S13. When any sensor detects an abnormality in its own data, it triggers the activation of redundant sensors and sends a "sensor abnormality" warning to the AI ​​co-controller. in, This is expressed as the maximum tolerable response time of the sensor feedback; This represents the rotational speed threshold at which the accelerometer starts sampling; This represents the angle threshold at which the inclinometer starts sampling; This represents the vibration threshold at which the ultrasonic probe initiates sampling. This is expressed as the sampling frequency of the accelerometer; This is expressed as the sampling frequency of the inclinometer; This is expressed as the sampling frequency of the gyroscope; This is expressed as the sampling frequency of the ultrasonic probe.

4. The wind turbine end-to-end adaptive control method based on dynamic perception and co-evolutionary DRL as described in claim 1, characterized in that, The specific steps of step S2 are as follows: S21. Based on the verified multi-source data, it is determined that there is data heterogeneity, triggering the physical knowledge-enhanced convolutional encoder to start the feature extraction process; The data alignment module receives the collected data and uses linear interpolation to unify the low-resolution data to a preset high-resolution data. S22. The attention mechanism module is used to extract physical constraint features by embedding fluid dynamics constraints. At the same time, the attention mechanism module is triggered to calculate the weight coefficients based on the importance scores of each sensor data output by S1 and dynamically allocate data weights. The final output includes the working condition type, dynamic feature parameters and confidence level. The working condition type includes steady wind, turbulent wind and extreme wind. The dynamic feature parameters include wind speed change rate and load fluctuation amplitude. S23. When the working condition classification module collects extreme working condition samples and the generator of the generative adversarial network is in a ready state, the generator is triggered to start sample synthesis, and the synthesized samples are added to the sample library.

5. The wind turbine end-to-end adaptive control method based on dynamic perception and co-evolutionary DRL according to any one of claims 1 to 4, characterized in that, The specific steps of step S3 are as follows: S31. Based on the dynamic feature parameters and confidence level, the AI ​​co-controller triggers the loading of the pre-trained deep reinforcement learning (DRL) model. S32. Read the working condition type based on the loaded deep reinforcement learning (DRL) model. If the working condition type changes, trigger the model parameter configuration switch. S33. Subsequently, the model integrates a dual-criterion and TD error-driven regularization mechanism, embedding Lyapunov stability constraints; whereby, when the deep reinforcement learning DRL model performs Q-value calculation, the dual-criterion architecture is triggered to run in parallel: Critics 1 calculates based on dynamic feature parameters. ,in, Represented as the Q-function of commentator 1; This is represented as the parameter matrix for commentator 1; commentator 2 calculates based on the same parameters. ,in, Represented as the Q function of commentator 2; Represented as the parameter matrix of commentator 2; Pick and The minimum value is taken as the target Q value, and then the TD error is calculated. Let the threshold of TD error be set. ,like This triggers the regularization mechanism, applying an L2 regularization term to the policy network parameters. According to the regularization term Adjust the learning rate from Down to ,in, This is represented as the learning rate after regularization; Set a threshold number of iterations for TD error stability. ,when And maintain In the next iteration, the regularization mechanism is turned off, and the initial learning rate is restored; S34. Input dynamic characteristic parameters for inference, and output variable pitch PID coefficients and yaw angle control commands.

6. The wind turbine end-to-end adaptive control method based on dynamic perception and co-evolutionary DRL as described in claim 1, characterized in that, The specific steps of step S4 are as follows: S41. Safety assessment of triggering the self-penalty mechanism based on the variable pitch PID coefficient and yaw angle control command; S42. Call the preset safety threshold library. Safety thresholds include pitch angle ≤ yaw rate ≤ , load ≤ The deviation between the calculation instruction and the threshold. ; S43, if Then a penalty signal is generated. and will issue a punishment signal Feedback is sent to the deep reinforcement learning DRL model of S3 via the backpropagation channel; S44. When the deep reinforcement learning (DRL) model receives a penalty signal When this happens, the strategy network parameters are updated, and the corrected pitch PID coefficients and yaw angle control commands are re-output. S45, if continuous If the output pitch PID coefficient and yaw angle control command still exceed the safety domain, the model will be loaded with the historical best strategy until the output pitch PID coefficient and yaw angle control command fall into the safety domain, and the final pitch PID coefficient and yaw angle control command with the "safety verification passed" flag will be generated.

7. The wind turbine end-to-end adaptive control method based on dynamic perception and co-evolutionary DRL as described in claim 6, characterized in that, The specific steps of step S5 are as follows: S51. If the final pitch PID coefficient and yaw angle control command with the "safety verification passed" flag is received, and the communication module detects that the OPC UA Pub / Sub channel status is normal, the command priority determination is triggered; if the pitch PID coefficient and yaw angle control command type is "emergency adjustment", a real-time control channel is allocated; if the pitch PID coefficient and yaw angle control command type is "normal adjustment", a normal control channel is allocated. S52, then the communication module is triggered to transmit instructions to the main control PLC according to priority, and at the same time the instruction transmission confirmation mechanism is started. After the main control PLC receives the instruction... The system will send a "receive command" signal within a specified time; if the timeout occurs, a retransmission will be triggered. When the main control PLC receives the command and its self-test is fault-free, the actuator will be triggered to start the pitch and yaw angle adjustment actions.

8. The wind turbine end-to-end adaptive control method based on dynamic perception and co-evolutionary DRL as described in claim 1, characterized in that, The specific steps of step S6 are as follows: When the data acquisition module receives the actual operating data fed back by the main control PLC actuator, the actual operating data includes the actual pitch angle, actual yaw angle, and generator power, and the data timestamp is less than the S5 command execution time deviation. When the deviation calculation module is activated, it is triggered. The time deviation threshold for data correlation; Calculate the relative deviation between the actual data and the S3 output command value. Set the deviation threshold ,like If the deviation data is marked as a "valid training sample", it is input into the experience pool of the deep reinforcement learning DRL model in S3; when the number of samples in the experience pool is ≥ The bar or model iteration counter reaches At this time, the deep reinforcement learning (DRL) model is triggered to start online iterative optimization, updating the policy parameters of the DRL model; among which... This represents the number of samples in the experience pool. This represents the number of iterations.

9. The wind turbine end-to-end adaptive control method based on dynamic perception and co-evolutionary DRL as described in claim 5, characterized in that, In step S32, when the deep reinforcement learning (DRL) model training iteration counter shows that it has reached the target value... The experience pool detected that the number of stored samples was ≥ When a condition is met, the experience replay module is triggered to randomly draw from the experience pool. Samples, sample format is ,in, This represents the operating parameters output by S2 at time t; This represents the control command output by the model at time t; This represents the reward value fed back by S6 at time t; This represents the operating parameters output by S2 at time t+1; The number of iterations; To store the number of samples; The number of randomly selected samples; When sample extraction is complete and the target network detects that its parameters have not been updated ≥ At this time, the target network parameters are updated, and the current network parameters are updated. Copy to target network ,in, This is represented as the parameter matrix of the current network; This is represented as the parameter matrix of the target network; This represents the number of times its own parameters have not been updated. Then, the target Q-value is calculated, and the network Q-value is compared with the target Q-value. Calculate the loss and update the current network parameters using gradient descent; where, This is represented as the Q-function of the current network; Set reward weight coefficient threshold When extreme wind conditions are detected, trigger Adjusted to Prioritize ensuring stable rotational speed.

10. The wind turbine end-to-end adaptive control method based on dynamic perception and co-evolutionary DRL as described in claim 5, characterized in that, In step S33, a threshold number of times the TD error stabilizes is set. TD error decay trigger threshold Threshold for the number of TD error fluctuations The trigger threshold for increased exploration rate ; When the absolute value of TD error is continuous less than At this time, the exploration rate decay mechanism of the deep reinforcement learning (DRL) model is triggered: the exploration rate decreases from... linear decay to The decay step size is / iterations, where This is expressed as the exploration rate after decay; This represents the decay step size of the exploration rate; When the absolute value of TD error is continuous Next > At that time, the exploration rate increase mechanism is triggered: the exploration rate increases from... linearly improved to Increase step size to / iteration, simultaneously triggering empirical pool sample diversity verification; if sample coverage < Uncovered operating condition types > If a sample is found, it will trigger the download of supplementary samples from the cloud; in, This represents the step size for increasing the exploration rate; This is expressed as a threshold for sample coverage. The threshold is represented by the uncovered operating condition type; supplementary samples include extreme wind conditions and fault condition samples.

Images