Control method for variable pitch system of wind power plant

Abstract

The invention relates to the technical field of wind power generation, in particular to a control method for a variable pitch system of a wind power plant. Comprising the steps that a general strategy model based on meta-learning is constructed, initial parameters irrelevant to a fan are obtained through cross-fan meta-training, and rapid fine adjustment adaptation is supported; deploying an intelligent agent taking the initial parameter as a starting point in each fan, and performing local strategy optimization by using deep reinforcement learning; designing a composite reward function integrating the generating capacity, the mechanical load and the power smoothness, and guiding the intelligent agent to realize multi-target dynamic balance; building a federated learning collaborative network, encrypting and uploading model parameters to update by each fan, and aggregating and distributing through a central server to realize full-field collaborative optimization; through closed-loop cooperation of the technology, the system is continuously and adaptively optimized. By fusing meta learning, deep reinforcement learning and federal learning, the overall power generation efficiency, the equipment reliability and the operation intelligence level of the wind power plant are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of technology, and specifically to a control method for a wind farm pitch control system. Background Technology

[0002] Wind energy, as a crucial component of clean energy, relies on the precise control of wind farms for its efficient and stable utilization. The pitch control system, a key actuator in wind turbine generators, is responsible for controlling the power captured and the load borne by the turbine by adjusting the blade pitch angle. Its control performance directly impacts power generation efficiency, equipment lifespan, and grid stability. Traditional wind farm pitch control methods primarily rely on controllers designed based on fixed mathematical models, such as proportional-integral-derivative (PID) controllers and their improved versions. However, in actual operation, wind resources exhibit strong nonlinearity, time-varying characteristics, and uncertainties, including complex turbulence, wind shear effects, and complex wake interference between turbines within the wind farm. This makes it difficult for traditional controllers based on fixed models and parameters to maintain optimal performance across the entire operating range, often resulting in insufficient adaptability and decreased control accuracy in dynamically changing environments. Furthermore, traditional control methods typically focus on a single objective (such as maximum power point tracking), making it difficult to effectively coordinate multiple conflicting control objectives such as maximizing power generation efficiency, minimizing mechanical load, and smoothing power output. In recent years, although some studies have attempted to introduce single-unit intelligent control algorithms, their models have weak generalization capabilities, making it difficult to directly transfer them to wind turbines of different models or under different wind conditions. Furthermore, they lack wind farm-level collaborative mechanisms, failing to optimize the overall output of the wind farm and mitigate the adverse effects of wake at the system level. Therefore, there is an urgent need for a novel pitch control method that can adapt to dynamic environments, intelligently balance multi-objective conflicts, and possess wind farm-level collaborative optimization capabilities to overcome the current technical bottlenecks faced by wind farm operations.

[0003] Therefore, existing technologies still need further development. Summary of the Invention

[0004] The purpose of this invention is to overcome the above-mentioned technical deficiencies and provide a control method for a wind farm pitch system to solve the problems existing in the prior art.

[0005] To achieve the above-mentioned technical objectives, the present invention provides a control method for a wind farm pitch system, comprising: S100. Construct a general strategy model based on meta-learning. By performing cross-wind turbine meta-training on historical data of heterogeneous wind turbines, the model can obtain wind turbine-independent initial parameters. When a new wind turbine is connected or the operating conditions change, only short-term operating data needs to be collected to quickly fine-tune and adapt the model. S200. Deploy an intelligent agent control unit locally on each wind turbine. Starting from the initial parameters of the general strategy model, it runs as an edge computing node and uses deep reinforcement learning algorithms to continuously interact and explore in the simulation environment and real-time system, and optimizes the control strategy through local data. S300 designs a multi-dimensional composite reward function to dynamically evaluate the pitch control effect. The function comprehensively considers multiple objective indicators such as power generation, mechanical load, and power output smoothness, guiding the agent to autonomously balance the optimization priorities under different operating conditions.

[0006] Specifically, the S100 training adopts a model-independent meta-learning framework, which optimizes the initial parameters of the model to enable it to converge quickly on various virtual working conditions and enhances the generalization ability of the dynamic characteristics of the fresh air unit through a gradient update mechanism.

[0007] Specifically, a working condition classifier is introduced during the meta-training process to classify the current working condition based on real-time wind speed, turbulence intensity, and load signals, and dynamically select the closest subset of initial parameters in the meta-model to accelerate the fine-tuning and adaptation process.

[0008] Specifically, the deep reinforcement learning in S200 adopts an actor-critic architecture, in which the actor network generates the paddle-changing action, the critic network evaluates the value of the action, and the interaction data is stored through an experience replay buffer to improve the learning stability.

[0009] Specifically, both the actor network and the critic network adopt a deep neural network structure. The network inputs include wind speed time series data, wind turbine speed, pitch angle and load measurement values, and the output is a continuous pitch angle control command.

[0010] Specifically, the multidimensional composite reward function in S300 is a weighted sum function, in which the power generation reward is positively correlated with the actual power output, the mechanical load penalty is negatively correlated with the tower vibration amplitude and blade stress, the power smoothness reward is negatively correlated with the power change rate, and the weight coefficients can be dynamically adjusted according to the grid dispatching needs.

[0011] Specifically, the method further includes: S400: Construct a wind farm-level federated learning collaborative network. Each wind turbine agent encrypts the updated model parameters after local training and distributes them after global updates are aggregated through the central server, enabling all wind turbines in the farm to implicitly learn the wake interference rules and collaborative strategies. The federated learning collaborative network in the S400 adopts an asynchronous aggregation strategy, which allows some wind turbines to temporarily postpone parameter uploading due to communication delays. The central server merges available updates through a weighted average algorithm to ensure the robustness and consistency of the model update across the entire field.

[0012] Specifically, differential privacy noise is added to parameter updates during the federated learning process to protect the local data privacy of each wind turbine, and a model version management mechanism is used to prevent model degradation caused by malicious attacks.

[0013] Specifically, the intelligent agent control unit integrates a fault-tolerant module. When a sensor failure or communication interruption is detected, it automatically switches to a backup control strategy based on historical data and triggers a model re-initialization process.

[0014] Specifically, the method further includes: S500, through closed-loop collaboration of meta-learning, deep reinforcement learning and federated learning, enables the pitch system to achieve an intelligent transition from sensing environmental changes to adaptive decision-making, continuously optimizing the dynamic balance between power generation efficiency, equipment lifespan and grid demand.

[0015] Beneficial effects: Compared with existing technologies, the control method for wind farm pitch control systems provided by this invention constructs a hierarchical and collaboratively evolving intelligent control system by deeply integrating cutting-edge artificial intelligence technologies such as meta-learning, deep reinforcement learning, and federated learning, achieving significant beneficial effects in many aspects.

[0016] First, this invention endows the control system with powerful rapid transfer and adaptation capabilities through pre-training of a general strategy model based on meta-learning. This model learns meta-knowledge of "how to learn quickly" from massive amounts of heterogeneous wind turbine data, enabling it to perform precise and rapid fine-tuning and adaptation using only short-term operating data when facing newly connected wind turbines or new operating conditions. This fundamentally solves the problems of large data volumes, long cycles, and high costs associated with traditional methods for modeling new wind turbines or retraining controllers, achieving "plug-and-play" control strategies and greatly improving the agility of wind farm expansion, renovation, and operation and maintenance.

[0017] Secondly, by deploying edge agents based on deep reinforcement learning and designing a composite reward function that integrates multiple engineering objectives, each wind turbine is equipped with the ability to autonomously explore and make decisions in complex and dynamic environments. Through continuous interaction with simulation and real-world environments, the agent learns how to dynamically balance multiple mutually constraining objectives such as power generation, mechanical load, and power smoothness, thereby finding the control strategy with the optimal long-term overall benefit. This method overcomes the limitations of traditional single-objective or fixed-weight optimization methods, achieving a better balance between economic efficiency and safety.

[0018] Furthermore, by constructing a field-level collaborative network based on federated learning, a novel approach was adopted to achieve full-field knowledge sharing and strategy collaborative optimization while protecting the local data privacy of each wind turbine. Each wind turbine agent implicitly learns the wake distribution patterns and globally optimal collaborative strategies of the entire wind farm by exchanging encrypted model parameter updates, rather than raw sensitive data. This not only effectively improves the overall power generation efficiency of the entire field and alleviates power generation losses and load unevenness caused by wake effects, but also meets increasingly stringent data security and privacy protection requirements.

[0019] Finally, this invention integrates a complete fault-tolerant module and deeply links with the central monitoring system, ensuring the high reliability and maintainability of the intelligent control system. The system can automatically and seamlessly switch to a safe backup strategy in case of sensor failure, communication interruption, or other abnormal situations, and automatically reinitialize the intelligent model after fault recovery. Simultaneously, a visual interface provides maintenance personnel with transparent status monitoring and flexible manual intervention, achieving an organic combination of human expert experience and machine intelligence, ensuring the long-term stable and reliable operation of the system. In summary, this invention systematically improves the intelligence level and overall performance of wind farm pitch control from multiple levels, including individual adaptation, multi-objective optimization, group collaboration, and system reliability. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the control method of the wind farm pitch system provided in a specific embodiment of the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments in this application, other similar embodiments obtained by those skilled in the art without creative effort should all fall within the scope of protection of this application. Furthermore, directional terms mentioned in the following embodiments, such as "up," "down," "left," and "right," are only for reference to the directions in the accompanying drawings; therefore, the directional terms used are for illustrative purposes and not for limiting the invention.

[0022] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments.

[0023] Please see Figure 1 This invention provides a control method for a wind farm pitch control system, comprising: S100. Construct a general strategy model based on meta-learning. By performing cross-wind turbine meta-training on historical data of heterogeneous wind turbines, the model can obtain wind turbine-independent initial parameters. When a new wind turbine is connected or the operating conditions change, only short-term operating data needs to be collected to quickly fine-tune and adapt the model.

[0024] It should be further explained that the core of this invention is to construct a three-layer progressive intelligent control architecture of "meta-initialization - individual learning - group collaboration". On the cloud or a high-performance server, a massive dataset containing historical SCADA data from hundreds or thousands of wind turbines of dozens of different models and geographical locations is used to perform meta-learning pre-training of the S100. The goal of this step is to produce a "general policy base model" whose parameter initialization points have a universal and excellent initial response to the dynamic characteristics of various wind turbines, laying the foundation for rapid adaptation in the future.

[0025] S200. Deploy an intelligent agent control unit locally on each wind turbine. Starting from the initial parameters of the general strategy model, it runs as an edge computing node and uses deep reinforcement learning algorithms to continuously interact and explore in the simulation environment and real-time system, optimizing the control strategy through local data.

[0026] It should be further explained that the parameters of this "base model" are deployed to the edge agent control unit of each specific wind turbine (an industrial control computer equipped with a GPU or a dedicated AI acceleration chip can be selected) as the initial policy network of its deep reinforcement learning agent.

[0027] S300 designs a multi-dimensional composite reward function to dynamically evaluate the pitch control effect. The function comprehensively considers multiple objective indicators such as power generation, mechanical load, and power output smoothness, guiding the agent to autonomously balance the optimization priorities under different operating conditions.

[0028] It should be further explained that the intelligent agent executes S200 on the digital twin simulation model of the wind turbine and the actual wind turbine. Through real-time interaction with the environment and by utilizing the feedback of the multi-objective reward signal designed by S300, it continuously optimizes its own strategy, making it increasingly adaptable to the specific location, wind conditions and mechanical state of the wind turbine.

[0029] S400 constructs a wind farm-level federated learning collaborative network. Each wind turbine agent encrypts the updated model parameters after local training and distributes them through a central server after aggregating the global updates, enabling all wind turbines in the farm to implicitly learn the wake interference rules and collaborative strategies.

[0030] It should be further explained that at the wind farm level, the S400 initiates a background-running federated learning collaborative network. Each wind turbine's agent periodically (e.g., hourly) encrypts and sends the updated parameters (not the raw data) of its policy network to the central server. The server aggregates all updates to form a new global model, which is then distributed back to each wind turbine, enabling all turbines in the farm to share learning experiences and collaboratively address global issues such as wake ripples.

[0031] S500, through closed-loop collaboration of meta-learning, deep reinforcement learning and federated learning, enables the pitch system to achieve an intelligent transition from sensing environmental changes to adaptive decision-making, continuously optimizing the dynamic balance between power generation efficiency, equipment lifespan and grid demand.

[0032] It should be further explained that S500 describes a closed-loop linkage of these three levels: meta-learning provides "meta-knowledge" for rapid startup, deep reinforcement learning enables individuals to "accumulate experience" and "refine skills", and federated learning promotes "knowledge sharing" and "co-evolution" of the group, ultimately forming an intelligent pitch ecosystem that can continuously improve itself.

[0033] Understandably, this invention systematically solves the problems of generalization, adaptability, and coordination that traditional single controllers cannot simultaneously address. Meta-learning solves the problems of slow "cold start" and large data requirements for controllers under new wind turbines or new operating conditions; deep reinforcement learning enables each wind turbine to autonomously find multi-objective optimal solutions in complex and ever-changing environments; federated learning, while protecting data privacy, achieves coordinated optimization of strategies for all wind turbines in the field, thereby improving power generation and reducing fatigue load at the system level.

[0034] Specifically, the S100 training adopts a model-independent meta-learning framework, which optimizes the initial parameters of the model to enable it to converge quickly on various virtual working conditions and enhances the generalization ability of the dynamic characteristics of the fresh air unit through a gradient update mechanism.

[0035] It should be further noted that the model-independent meta-learning framework employs the MAML algorithm. The specific training process is as follows: Prepare a meta-training set containing historical data from N wind turbines, with data from each wind turbine i constituting an independent task. Each task The data is randomly divided into support sets. and query set The support set is used for task-specific fast adaptation (inner loop), and the query set is used to evaluate the performance of the adapted model and guide meta-optimization (outer loop). Define a parameter as... General strategy model For example, a deep neural network. The goal of the MAML algorithm is to find an optimal set of initialization parameters. This makes it possible for tasks to be distributed For any new task sampled from the dataset, the model can perform well on the task's query set after a few gradient updates using only a small amount of data from the task's support set. A single meta-training iteration contains the following sub-steps: First, sample a batch of tasks (e.g., batchsizeB=8). For each task in the batch... Calculate its inner loop update: using its support set Calculate the loss function Relative to parameters The gradient is calculated, and gradient descent is performed in one or more steps. The parameters updated in the inner loop are: in, This is the learning rate of the inner loop, with an optimal value of 0.01. This value is chosen to ensure effective adaptation within a small number of iterations (usually 1-5 steps), avoiding instability due to excessively large step sizes or insufficient adaptation due to excessively small step sizes.

[0036] Then, use the updated parameters In query set Calculate loss After completing the inner loop update and query set loss calculation for all tasks within the batch, the outer loop update (meta update) is performed: calculating the sum of the query set losses for all tasks with respect to the initial parameters. The gradient, and for Update: in, This is the outer loop learning rate (meta-learning rate), with a preferred value of 0.001. Choose a smaller value. This is to ensure the stability of the meta-optimization process, avoid oscillations when aggregating second-order gradient information from multiple tasks, and thus learn more robust initial parameters. Loss function Defined as a negative reward for control performance, this could be, for example, the mean squared error between the model-predicted pitch action sequence and the "ideal" action (or the action generated by the data-driven controller) in the actual collected operational data, combined with a penalty term for power tracking error. Through thousands of such meta-training iterations, the final result is... It possesses a strong and rapid adaptability. When faced with a brand-new wind turbine, only about two hours of representative operating data from that turbine need to be collected as a support set, based on the aforementioned... Starting from this point, perform 1-3 steps of internal loop updates to obtain the adapted controller parameters for this wind turbine.

[0037] Understandably, this invention fundamentally solves the problem of the large amount of data and time required for traditional methods to model or retrain controllers for new wind turbines, realizing "plug-and-play" control strategies and rapid deployment, and greatly improving the efficiency of intelligent transformation and operation and maintenance of wind farms.

[0038] Specifically, a working condition classifier is introduced during the meta-training process to classify the current working condition based on real-time wind speed, turbulence intensity, and load signals, and dynamically select the closest subset of initial parameters in the meta-model to accelerate the fine-tuning and adaptation process.

[0039] It should be further explained that, in order to further improve the efficiency and adaptation accuracy of meta-learning, we have... (The sentence is incomplete and requires more context to be translated accurately.) In addition, a lightweight condition classifier is trained in parallel. This classifier It is a small neural network whose input feature vector Includes: the average wind speed over the past T seconds (T is preferably 600 seconds, i.e., 10 minutes). Standard deviation of wind speed (characterizing turbulence intensity) and the root mean square value of the vibration acceleration in the front and rear directions of the tower. The output is the probability distribution of K predefined work condition categories. These categories can be predefined based on engineering experience and include: Category 1: "Stable operation at low wind speeds" ); Category 2: "High turbulence near rated wind speed" ); Category 3: "High wind speed constant power operation" ); Category 4: "Extreme Gust Conditions" (Detection of sudden wind speed changes exceeding the threshold).

[0040] Furthermore, during the meta-training phase, we perform training on each task... The data is first processed using a classifier. (its parameters) and (Joint training) classifies each data segment. Then, instead of training a single set of initialization parameters... Instead, it trains a dedicated initialization parameter for each working condition category k. Specifically, in each meta-training iteration, for a batch of tasks, we first determine the main working condition categories of the support set data for each task (as determined by...). The parameters are then determined and assigned to the corresponding category groups. Then, the MAML inner-outer loop update is performed for each category group, but this time each group updates its own specific initial parameters. Operating condition classifier The parameters are also updated through gradients throughout the meta-training process to learn how to better identify discriminative operating conditions from the data. In online applications, when the fresh air system is connected or its operating conditions change, it first utilizes its initial short-term operating data (e.g., 10 minutes of data) and a pre-trained operating condition classifier. Determine its dominant operating condition category k. Then, select the corresponding... As a starting point for rapid fine-tuning, rather than a universal approach. Based on this, fine-tuning is then performed using complete short-term data (e.g., 2 hours) from the wind turbine. Operating condition classifier. The network structure can be a two-layer fully connected network: an input layer with 3 nodes, a hidden layer with 16 nodes (using the ReLU activation function), and an output layer with K nodes (using the Softmax activation function).

[0041] Understandably, this invention improves the "one-size-fits-all" general initialization into a "categorized" refined initialization through meta-learning based on operating conditions. This is equivalent to preparing different "solution templates" for different "problem types," making the starting point of the online fine-tuning process closer to the optimal solution under that operating condition. This further reduces the amount of data and iteration steps required for fine-tuning, significantly accelerates the controller's adaptation speed to new environments, and improves the initial performance of control.

[0042] Specifically, the deep reinforcement learning in S200 adopts an actor-critic architecture, in which the actor network generates the paddle-changing action, the critic network evaluates the value of the action, and the interaction data is stored through an experience replay buffer to improve the learning stability.

[0043] It should be further explained that in the locally deployed agent control unit of each wind turbine, the proximal policy optimization algorithm is used as the specific implementation of deep reinforcement learning. The PPO algorithm is an efficient actor-critic algorithm. The actor network (i.e., the policy network) Responsible for determining the current status Output Action The probability distribution of (pitch angle change). Commentator network (i.e., value network). Responsible for assessing the current status The long-term expected cumulative reward. Each interaction between the agent and the environment (wind turbine simulation model or actual wind turbine) generates a transition tuple. These experiences are stored in a fixed-size FIFO experience replay buffer, preferably containing 100,000 experiences. The PPO algorithm stabilizes policy updates through importance sampling and pruning mechanisms. Its core objective is to maximize the following pruned alternative objective function: in, This represents the probability ratio between the old and new strategies. It is an estimate of the advantage function, representing the state... Next action How much better it is compared to the average case is usually calculated using the generalized dominance estimation algorithm, and its formula is: ,in γ is the discount factor (preferably 0.99), and λ is the GAE parameter (preferably 0.95). The clip function will... Limited to Within the interval, ε is the pruning coefficient, with an optimal value of 0.2. This pruning operation prevents excessively large magnitudes in a single policy update, ensuring training stability. The update objective of the critic network is to minimize the mean squared error loss of the value function. ,in It is the target value, usually obtained through return estimation.

[0044] Furthermore, in actual training, the agent control unit periodically (e.g., every 1024 new experiences) randomly samples a minibatch (preferably 64 bytes) of data from the experience replay buffer to simultaneously update the actor network and the critic network. The learning rates of the actor network and the critic network can be set separately, with the actor network preferably having a learning rate of 3e-4 and the critic network preferably having a learning rate of 1e-3.

[0045] Understandably, the PPO algorithm, combined with the experience replay mechanism, enables the agent to learn efficiently and stably from interactions with the environment. Experience replay breaks the temporal correlation between data, improving data utilization; while PPO's pruning mechanism effectively constrains the step size of policy updates, avoiding performance crashes caused by inappropriate step sizes in traditional policy gradient methods, thus ensuring the robustness and convergence of the learning process in complex, continuously controlled scenarios such as wind turbines.

[0046] Specifically, both the actor network and the critic network adopt a deep neural network structure. The network inputs include wind speed time series data, wind turbine speed, pitch angle and load measurement values, and the output is a continuous pitch angle control command.

[0047] It should be further explained that the actor network and the critic network share a common feature extraction backbone network, which is then split into two independent output heads. The specific network structure is as follows: Input status It is a feature matrix containing L historical time steps (L is preferably 20, corresponding to 20 seconds of data, and the control period is 1 second).

[0048] The feature vector at each time step contains 7 dimensions: current wind speed. Wind turbine generator speed Generator power Current pitch angle Vibration acceleration in the front and rear directions of the tower lateral vibration acceleration of the tower and the dominant frequency amplitude of the torsional vibration of the transmission chain. .

[0049] Therefore, the input state matrix has a dimension of 20×7. First, this matrix is ​​processed through a one-dimensional convolutional layer using 16 filters with a width (kernel size) of 3, a stride of 1, and "same" padding to maintain the temporal dimension length. The convolutional layer is followed by a ReLU activation function. Its output then passes through a max-pooling layer with a pooling window size of 2 and a stride of 2, reducing the temporal dimension length to 10. Subsequently, the pooled output is flattened into a feature vector. This feature vector is then processed through two fully connected layers: The first fully connected layer has 128 neurons and uses the ReLU activation function; The second fully connected layer has 64 neurons and also uses the ReLU activation function.

[0050] Furthermore, this 64-dimensional vector represents the shared feature. Starting from this shared feature, the network is divided into two heads. The actor's network head: The 64-dimensional shared feature is input into a fully connected layer with two neurons. The first neuron outputs the mean of the action. The log-standard deviation of the output action of the second neuron (This is a state-dependent, independent parameter used to control exploration.) The actual action standard deviation is calculated through... Calculated. The final action. From Gaussian distribution obtained from sampling, i.e. .this It is a continuous value between -1 and 1, representing the change (increment) in the pitch angle. To map it to the actual pitch angle command, scaling and limiting are required. ,in It is the maximum pitch rate per step, for example, 2 degrees / second, then It is limited to the physical limits of the fan (such as 0 to 90 degrees).

[0051] Furthermore, the specific design of the commentator network head is as follows: The same 64-dimensional shared features are input into a fully connected layer with one neuron (no activation function), directly outputting the state value estimate for that state. , a scalar. All weights of the network are initialized using a Kaiming normal distribution.

[0052] Understandably, the network structure designed in this invention can effectively handle multi-source, multi-variable time-series signals in wind turbine control. Convolutional and pooling layers can automatically extract local patterns and temporal dependencies from signals such as wind speed and vibration. The shared backbone network design improves computational efficiency and ensures that actors and commentators have a consistent understanding of the state characteristics. Ultimately, this network can map high-dimensional, dynamic environmental states into precise, smooth, continuous pitch angle control commands, which is key to achieving high-performance, high-stability pitch control.

[0053] Specifically, the multidimensional composite reward function in S300 is a weighted sum function, in which the power generation reward is positively correlated with the actual power output, the mechanical load penalty is negatively correlated with the tower vibration amplitude and blade stress, the power smoothness reward is negatively correlated with the power change rate, and the weight coefficients can be dynamically adjusted according to the grid dispatching needs.

[0054] It should be further explained that at each control time step t, the agent executes an action. Afterwards, the environment shifted to a new state. and generate a scalar reward. The reward function of this invention is a weighted sum of three components, specifically designed as follows: in, , , These are the weighting coefficients for power generation bonus, load penalty, and smoothness penalty, respectively. Their default preferred values ​​are: , , .choose As a benchmark, maximizing power generation is set as the core objective. The setting is intended to impose a moderate but continuous penalty on mechanical loads (especially fatigue loads) without excessively suppressing power generation. This value, after simulation and debugging, can achieve a good long-term balance between power generation benefits and extending equipment life. The setting is intended to impose a slight penalty on rapid fluctuations in power output, with an absolute value much smaller than the load penalty, reflecting that power smoothing has a lower priority than load safety. The specific implementations of each component function are as follows: Power Generation Bonus Component The aim is to encourage wind turbines to capture as much wind energy as possible. Below rated wind speed, the reward is positively correlated with the ratio of current power to theoretical maximum power; above rated wind speed, the reward is positively correlated with the accuracy of power tracking of rated power. One specific implementation is as follows: in, It is the actual power at the next moment. Current wind speed The maximum theoretical power that the fan can achieve (determined by the fan's Cp curve). That is the rated wind speed. This is the rated power. Load penalty component. The aim is to reduce fatigue damage to critical components. The main considerations are the bending moments in the tower's forward and backward directions and the flapping moments of the blades. A simplified implementation is based on the square of the measured signal amplitude to approximately reflect the relationship between fatigue damage and the square of the stress amplitude (according to Miner's rule and the SN curve): in, It is the estimated bending moment in the front and rear directions at the bottom of the tower at time t+1 (which can be estimated by vibration acceleration integration and model). It is the estimated value of the leaf root waving bending moment. and This is the corresponding reference bending moment value, used for normalization. Power smoothness penalty component. Designed to suppress drastic fluctuations in power output, it is based on the square of the power change rate: in, It is the control cycle (1 second). In addition, the weighting coefficient... The system can dynamically adjust power levels based on grid dispatch instructions via a central monitoring system. For example, when the grid requires very stable power, power can be temporarily adjusted... Adjust to -0.05; in extreme wind conditions, to protect the wind turbine, it can be... Adjusted to -0.1.

[0055] Understandably, this reward function cleverly transforms a complex, mutually constraining multi-objective engineering optimization problem into a single scalar signal that a reinforcement learning agent can understand and optimize. Through carefully designed weights and component functions, the agent is guided to learn a comprehensive optimal control strategy that can balance power generation revenue, equipment fatigue life, and grid power quality in the long term—something that traditional single-objective controllers struggle to achieve.

[0056] Specifically, the federated learning collaborative network in S400 adopts an asynchronous aggregation strategy, allowing some wind turbines to temporarily postpone parameter uploading due to communication delays. The central server merges available updates through a weighted average algorithm to ensure the robustness and consistency of the entire model update.

[0057] It should be further noted that the wind farm-level federated learning collaborative network operates on a round-by-round basis. Assume there are M wind turbines participating in the federated learning. Each round of federated learning includes the following steps: The first step is parameter distribution: At the beginning of each round, the central server distributes the current global model parameters. The message was sent to all online wind turbine nodes.

[0058] The second step is local training: each wind turbine receives... Then, these parameters are used as the initial parameters for the local agent actor network. Subsequently, wind turbine i independently executes the described PPO local training process using its local runtime data collected in the most recent period (e.g., the past 6 hours) as the training set, performing E local epochs (E is preferably 5). After local training is completed, the updated local model parameters are obtained. Calculate the local model update amount (i.e., gradient or parameter difference). To control the update magnitude and facilitate subsequent weighting, Prune the crop so that its L2 norm does not exceed a threshold C (C is preferably 1.0), i.e. .

[0059] Step 3: Update and upload: Fan i will update the cropped amount of data. and the amount of local data used for training in this round. The experience records (i.e., the number of records) are packaged, timestamped, and then attempted to be uploaded to the central server. Considering the instability of industrial field networks, this invention employs an asynchronous aggregation strategy. The central server sets an aggregation time window. (Preferred timeframe is 300 seconds, or 5 minutes). After the server initiates this round of aggregation, it will wait and receive updates from each wind turbine within this time window.

[0060] Step 4, Asynchronous Aggregation: Within the Time Window After completion, the central server aggregates the updates successfully received from N wind turbines (N≤M). A weighted average algorithm is used, with weights determined by the amount of local data each wind turbine participated in during this training round. The aggregation formula is proportional to the data volume to reflect the difference in contribution from different data volumes. Furthermore, after aggregation is complete, the server uses... Update the global model. For example, in... For wind turbines that fail to upload updates within the specified time window, their updates will be temporarily discarded, and the wind turbine will continue to use the old global model. Run and wait to participate in the next round of federated learning. Aggregation time window The choice of seconds is based on engineering trade-offs: too short a time (e.g., 60 seconds) may cause many wind turbines to be discarded due to local training or communication delays, resulting in insufficient upload time and reduced participation and convergence speed; too long a time (e.g., 1800 seconds) results in too low a global model update frequency, failing to respond promptly to rapid changes in the wind field environment. 300 seconds is a compromise, ensuring that most wind turbines can complete local training and upload once under normal circumstances, while maintaining a relatively high model evolution frequency (twice per hour).

[0061] Understandably, the asynchronous aggregation strategy greatly improves the practicality and robustness of federated learning in complex network environments in industrial settings. It avoids the "tailing" effect caused by waiting for the slowest node in synchronous federated learning, ensuring that the collaborative learning process of the entire wind farm can continue even if communication between some nodes is temporarily interrupted or computation is delayed, thus achieving a good balance between model update efficiency and system reliability.

[0062] Specifically, differential privacy noise is added to parameter updates during the federated learning process to protect the local data privacy of each wind turbine, and a model version management mechanism is used to prevent model degradation caused by malicious attacks.

[0063] It should be further explained that, to enhance privacy and security, differential privacy processing is performed on the model before uploading updates to the wind turbine nodes. Specifically, local training is completed on each wind turbine to obtain the pruned model update vector. After that, do not upload directly. Instead, it first adds random noise that follows a Laplace distribution. For the update vector... Each scalar component in The component after adding noise is: in, Let represent random noise sampled from a Laplace distribution with a mean of 0 and a scale parameter of b. The scale parameter b determines the magnitude of the noise, which is related to the privacy budget ε and the sensitivity Δf of the update vector, as follows: In this invention, since the updated vector has already been... L2 norm pruning was performed (threshold C=1.0), therefore its sensitivity Δf=2C=2.0 (in the worst case, the addition or removal of a data sample can change the L1 norm of the update vector by at most 2C). The privacy budget ε is an important hyperparameter that controls the strength of privacy protection. A smaller ε provides stronger privacy protection but also introduces more noise, potentially harming model accuracy. Empirically, the optimal value for ε can be chosen between 0.5 and 2.0. For example, choosing ε=1.0 would result in a smaller scale parameter... The wind turbine node will add noise to the update vector. and its data volume weight Upload after encryption.

[0064] Furthermore, during server-side aggregation, since the expected value of Laplace noise is zero, the impact of noise on the global model is reduced when a large number of nodes participate in the averaging process. Regarding model security, a model version management mechanism is implemented on the central server. Each successful aggregation generates a new version of the global model. Afterwards, the server immediately evaluates its performance on a high-fidelity wind turbine simulation model test platform built from historical safety data, calculating metrics such as average reward. Simultaneously, the server saves several previous historical versions of the model (e.g., the most recent 10 versions). If the new model... Performance on the test platform compared to the previous version If the decline exceeds a preset threshold (e.g., an average reward drop of more than 10%), a security alert is triggered. At this point, the server does not immediately deploy the problematic server. Instead, it performs a rollback operation, restoring the global model to the previous stable version. At the same time, a malicious node detection program is launched, for example, to calculate the updates uploaded by each node. With the final aggregation direction ( The cosine similarity is used to mark nodes with extremely low similarity (e.g., below -0.5) as suspicious nodes, and their updates are temporarily blocked in subsequent rounds to observe whether the model performance recovers.

[0065] Understandably, the introduction of differential privacy technology theoretically ensures that even if an attacker obtains all encrypted model updates uploaded by a wind turbine, they cannot infer any sensitive information (such as the operating status at a specific time) from the turbine's local training data, thus meeting stringent data privacy protection regulations. Model version management and robust aggregation mechanisms constitute a security defense, effectively detecting and mitigating potential poisoning attacks, ensuring the robustness of the federated learning process and the security of the final global model. This is crucial for deploying artificial intelligence systems in critical infrastructure.

[0066] Specifically, the intelligent agent control unit integrates a fault-tolerant module. When a sensor failure or communication interruption is detected, it automatically switches to a backup control strategy based on historical data and triggers a model re-initialization process.

[0067] It should be further explained that the fault-tolerant module is an independent software monitoring process deployed within the edge intelligent agent control unit of each wind turbine, running in real time with high priority. This module continuously monitors three types of signals: First, check the health status of key input sensors, including anemometers, speed encoders, pitch and angle encoders, accelerometers, etc. This is determined by checking whether the signal values ​​are within a reasonable physical range (e.g., wind speed 0-30 m / s), whether they remain unchanged for a long time (stuck), and whether the signal variance is abnormally low (failed).

[0068] Secondly, the network connection status with the central server of the wind farm is detected by periodically sending and receiving "heartbeat" packets.

[0069] Thirdly, the "health" indicators of the agent model itself, such as the variance of the actor network's output actions (whether it is abnormally high, indicating overexploration or policy instability), and the confidence interval of the critic network's value estimation (whether it is abnormally wide). The fault-tolerant module triggers failover when any of the following conditions are met: 1. The signal from any critical sensor is determined to be in a fault state and continues to exceed a time threshold. (Preferred time: 10 seconds).

[0070] 2. The network connection interruption with the central server lasted for more than [duration missing]. (Preferred time: 60 seconds).

[0071] 3. The model's own health indicators exceed the preset safety boundary for multiple consecutive control cycles (e.g., 20 consecutive cycles). Once triggered, the fault tolerance module immediately performs the following operations: First, in the next control cycle (within 1 second), the control of the pitch controller is seamlessly switched from the deep reinforcement learning agent to a preset, rule-based backup controller. This backup controller can be a gain-scheduled PID controller, whose PID parameter table is obtained by system identification and optimization tuning based on the historical normal operation data of the wind turbine under different wind speed ranges.

[0072] Secondly, detailed fault events are recorded in local logs and the central monitoring system, including the triggering cause, timestamp, and operating status before the switchover. Simultaneously, the agent enters "safe mode," suspending data collection from interactions with the environment and updates to the experience replay buffer, but continuing to record all sensor data (if available) and backup controller instructions in read-only mode.

[0073] Once the fault is resolved (e.g., sensor repair or replacement, network restoration), and the system has been running stably under the backup controller for a period of time (e.g., 300 seconds), the fault-tolerant module will initiate the model re-initialization process. This process includes: a) Use the most recent operational data (e.g., 30 minutes) collected after the fault is recovered to create a new fine-tuning dataset.

[0074] b) Load the most recent known good global model parameter or local model parameter from the central server or local secure storage as the starting point.

[0075] c) Employ a fast fine-tuning method similar to S100, but with stronger regularization (e.g., a smaller learning rate), and retrain the agent model on a new fine-tuning dataset with a small number of iterations (e.g., 10-20 times).

[0076] d) Validate the performance of the retrained model in a simulation environment. After successful validation, in the next low-wind-speed, stable operating window, gradually (e.g., linearly transitioning over 1 minute) return the control weights from the backup controller to the reinitialized agent. Threshold Seconds and The choice of seconds is based on engineering practice: 10 seconds is sufficient to filter out most transient signal interference without allowing the fault state to persist for too long; 60 seconds gives the field network a certain self-recovery time, avoiding frequent switching caused by short-term fluctuations.

[0077] Understandably, this invention provides crucial safety redundancy and robustness guarantees for AI control systems that heavily rely on data and models. It ensures that even if any link in the sensing, communication, or model fails, the wind turbine can still operate safely under a mature and reliable backup strategy, avoiding equipment risks that may arise from AI decision-making errors. Simultaneously, the automated model re-initialization process enables the system to quickly rebuild its intelligent control capabilities after fault recovery, without human intervention, greatly improving system availability and automation.

[0078] Specifically, the fault-tolerant module is linked with the wind farm's central monitoring system to report abnormal events and model performance indicators in real time, and provides an operation and maintenance intervention interface through a visual interface to support manual decision-making optimization.

[0079] It should be further explained that the linkage between the fault-tolerant module and the wind farm's central monitoring system is achieved through a defined standard communication protocol and data interface. The fault-tolerant module periodically (e.g., every 10 seconds) sends a "heartbeat" message to the central monitoring system, containing a summary of the local agent's status, such as: current control mode ("AI active control" / "backup control"), model average reward (moving average of the last 100 steps), model confidence score, local experience replay buffer size, and the time of the last successful participation in federated learning. When a failover event occurs, the fault-tolerant module immediately generates a high-level alarm event and pushes it to the central monitoring system in real time via a message queue. The event message includes the fault type, trigger value, failover time, and snapshots of key parameters before and after the failover. The central monitoring system has a dedicated "Intelligent Pitch System Health Management" visualization interface. This interface is based on a single-line diagram of the wind farm, with each turbine represented by an icon. The icon color reflects its agent status: green (normal AI control), yellow (model performance warning), and red (failover to backup control). Clicking on any turbine icon will bring up a detailed panel displaying: 1. Real-time curves: Displays real-time trend curves for parameters such as wind speed, power, pitch angle, critical loads, real-time reward value, and model value estimation.

[0080] 2. Event Log: Lists all events related to the wind turbine, such as fault switching, model updates, and federated learning participation, in reverse chronological order.

[0081] 3. Model Performance Dashboard: Displays historical statistics such as the model's average reward, load penalty component, and smoothness penalty component in digital and dashboard format (e.g., average values ​​for the most recent 1 hour and 24 hours).

[0082] 4. Manual Intervention Interface: Provides a set of operable controls. For example: ① Forced mode switching: Operation and maintenance personnel can manually select "Forced AI Mode" or "Forced Backup Mode" from the drop-down menu for testing or emergency use.

[0083] ② Reward weight adjustment: Three sliders are provided, each corresponding to a weight coefficient of the reward function. , , Maintenance personnel can dynamically adjust these weights based on special grid dispatch instructions (such as "prioritize power supply" or "prioritize equipment") or on-site observations. After adjustment, the new weights will be issued to the corresponding wind turbine agents, affecting their subsequent learning.

[0084] ③ Trigger Model Retraining: Button. When maintenance personnel believe that the model performance of a certain wind turbine is poor, they can click this button. The system will instruct the wind turbine agent to perform an enhanced local training (increasing the number of training rounds) in the background using data from the most recent period (such as the past 24 hours) and generate a training report.

[0085] ④ Inject Demonstration Data: Operations personnel are allowed to upload a segment of historical operational data that they believe is well-controlled (“demonstration trajectory”). This data will be added to the experience replay buffer of the wind turbine's intelligent agent to guide it in learning better strategies. All manual interventions performed through the interface are recorded in detail, including the operator, time, operation content, and reason. This data is then correlated with subsequent wind turbine performance changes to form a closed-loop feedback loop, used to continuously optimize the strategies and fault-tolerant logic of the entire intelligent control system.

[0086] Understandably, this invention constructs a "human-in-the-loop" hybrid intelligent operation and maintenance framework. Through a deeply integrated monitoring and visualization system, the "black box" decision-making process of the AI ​​controller is presented to operation and maintenance personnel in an understandable way, greatly enhancing the system's transparency and credibility. The flexible human intervention interface empowers operation and maintenance experts to oversee AI decision-making and guide AI learning at critical moments, achieving a complementary advantage between human experience and machine intelligence. This is crucial for ensuring the long-term stable, efficient, and reliable operation of complex industrial AI systems.

[0087] In a preferred embodiment, this application also provides an electronic device, the electronic device comprising: The computer device includes a memory and a processor, wherein the memory stores computer-readable instructions that, when executed by the processor, implement the control method for the wind farm pitch system. The computer device can be broadly categorized as a server, terminal, or any other electronic device with the necessary computing and / or processing capabilities. In one embodiment, the computer device may include a processor, memory, network interface, communication interface, etc., connected via a system bus. The processor of the computer device can be used to provide the necessary computing, processing, and / or control capabilities. The memory of the computer device may include a non-volatile storage medium and internal memory. The non-volatile storage medium may store an operating system, computer programs, etc. The internal memory can provide an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface and communication interface of the computer device can be used to connect and communicate with external devices via a network. When the computer program is executed by the processor, it performs the steps of the method of the present invention.

[0088] This invention can be implemented as a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the steps of the methods of embodiments of the invention to be performed. In one embodiment, the computer program is distributed across multiple network-coupled computer devices or processors, such that the computer program is stored, accessed, and executed in a distributed manner by one or more computer devices or processors. A single method step / operation, or two or more method steps / operations, may be executed by a single computer device or processor or by two or more computer devices or processors. One or more method steps / operations may be executed by one or more computer devices or processors, and one or more other method steps / operations may be executed by one or more other computer devices or processors. One or more computer devices or processors may execute a single method step / operation, or execute two or more method steps / operations.

[0089] Those skilled in the art will understand that the method steps of this invention can be performed by a computer program instructing related hardware, such as a computer device or processor, to perform the steps of this invention when executed. Depending on the context, any references herein to memory, storage, databases, or other media may include non-volatile and / or volatile memory. Examples of non-volatile memory include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state drive, etc. Examples of volatile memory include random access memory (RAM), external cache memory, etc.

[0090] The technical features described above can be combined arbitrarily. Although not all possible combinations of these technical features are described, any combination of these technical features should be considered to be covered by this specification, provided that such combination does not contain contradictions.

[0091] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A control method for a wind farm pitch control system, characterized in that, Includes the following steps: S100. Construct a general strategy model based on meta-learning. By performing cross-wind turbine meta-training on historical data of heterogeneous wind turbines, the model can obtain wind turbine-independent initial parameters. When a new wind turbine is connected or the operating conditions change, only short-term operating data needs to be collected to quickly fine-tune and adapt the model. S200. Deploy an intelligent agent control unit locally on each wind turbine. Starting from the initial parameters of the general strategy model, it runs as an edge computing node and uses deep reinforcement learning algorithms to continuously interact and explore in the simulation environment and real-time system, and optimizes the control strategy through local data. S300 designs a multi-dimensional composite reward function to dynamically evaluate the pitch control effect. The function comprehensively considers multiple objective indicators such as power generation, mechanical load, and power output smoothness, guiding the agent to autonomously balance the optimization priorities under different operating conditions.

2. The control method for a wind farm pitch control system according to claim 1, characterized in that, The S100 training adopts a model-independent meta-learning framework, which optimizes the initial parameters of the model to enable it to converge quickly on various virtual working conditions and enhances the generalization ability of the dynamic characteristics of the fresh air machine through a gradient update mechanism.

3. The control method for a wind farm pitch control system according to claim 2, characterized in that, The meta-training process introduces a working condition classifier, which classifies the current working condition based on real-time wind speed, turbulence intensity, and load signals, and dynamically selects the closest subset of initial parameters in the meta-model to accelerate the fine-tuning and adaptation process.

4. The control method for a wind farm pitch control system according to claim 1, characterized in that, The deep reinforcement learning in S200 adopts an actor-critic architecture, in which the actor network generates the paddle-changing action, the critic network evaluates the value of the action, and the interaction data is stored in the experience replay buffer to improve the learning stability.

5. The control method for a wind farm pitch control system according to claim 4, characterized in that, Both the actor network and the critic network adopt a deep neural network structure. The network inputs include wind speed time series data, wind turbine speed, pitch angle and load measurement values, and the output is a continuous pitch angle control command.

6. The control method for a wind farm pitch control system according to claim 1, characterized in that, The specific form of the multidimensional composite reward function in S300 is a weighted sum function, in which the power generation reward is positively correlated with the actual power output, the mechanical load penalty is negatively correlated with the tower vibration amplitude and blade stress, the power smoothness reward is negatively correlated with the power change rate, and the weight coefficients can be dynamically adjusted according to the grid dispatch requirements.

7. The control method for a wind farm pitch control system according to claim 1, characterized in that, The method further includes: S400: Construct a wind farm-level federated learning collaborative network. Each wind turbine agent encrypts the updated model parameters after local training and distributes them after global updates are aggregated through the central server, enabling all wind turbines in the farm to implicitly learn the wake interference rules and collaborative strategies. The federated learning collaborative network in the S400 adopts an asynchronous aggregation strategy, which allows some wind turbines to temporarily postpone parameter uploading due to communication delays. The central server merges available updates through a weighted average algorithm to ensure the robustness and consistency of the model update across the entire field.

8. The control method for a wind farm pitch control system according to claim 7, characterized in that, Differential privacy noise is added to parameter updates during the federated learning process to protect the local data privacy of each wind turbine and to prevent model degradation caused by malicious attacks through a model version management mechanism.

9. The control method for a wind farm pitch control system according to claim 1, characterized in that, The intelligent agent control unit integrates a fault-tolerant module. When a sensor failure or communication interruption is detected, it automatically switches to a backup control strategy based on historical data and triggers a model re-initialization process.

10. The control method for a wind farm pitch control system according to claim 1, characterized in that, The method further includes: S500, through closed-loop collaboration of meta-learning, deep reinforcement learning and federated learning, enables the pitch system to achieve an intelligent transition from sensing environmental changes to adaptive decision-making, continuously optimizing the dynamic balance between power generation efficiency, equipment lifespan and grid demand.

Images