Fan optimal torque coefficient online optimization method and system fused with air density

Abstract

The invention discloses an air density fused fan optimal torque coefficient online optimization method and system, belongs to the technical field of wind turbine generator control, and solves the technical problems that torque coefficients are fixed, dynamic air density changes cannot be adapted, and unit individual performance differences cannot be compensated in the prior art. The method comprises the following steps: constructing an air density-oriented adaptive Markov decision process model, and obtaining a state vector according to the air density; constructing an intelligent agent architecture based on a double-delay depth deterministic strategy gradient, wherein the intelligent agent architecture comprises a strategy network, two value networks and corresponding target network clusters; training and optimizing the agent architecture based on the double-delay depth deterministic strategy gradient; and performing forward propagation calculation on the state vector by adopting the trained agent architecture to obtain an optimal torque coefficient in the current state. The method is suitable for a wind turbine generator torque control scene.

Description

Technical Field

[0001] This invention belongs to the field of wind turbine control technology, specifically relating to an online optimization method for the optimal torque coefficient of a wind turbine that incorporates air density. Background Technology

[0002] Wind power generation is a key area in the current development of renewable energy. The control system of wind turbine units, especially maximum power point tracking (MPPT) control, is crucial for improving power generation efficiency. One of the core aspects of MPPT control is torque control. In the wind speed range below the rated wind speed, the industry commonly adopts an optimal torque control strategy based on the square of the generator speed. This involves using a lookup table (LUT) or fixed-parameter PI control to assign an optimal torque coefficient K, ensuring that the generator reference torque is proportional to the square of the speed. The theoretical basis of this method is to operate the unit at the optimal tip speed ratio, thereby obtaining the maximum wind energy capture coefficient. To cope with changes in operating conditions, some existing technologies attempt to use a reference air density. and measured air density This allows for simple linear compensation of the torque coefficient K.

[0003] However, existing technical solutions have significant flaws and shortcomings. First, the optimal torque coefficient K, which relies on the traditional lookup table method, is usually calculated offline during the turbine design phase and fixed in the controller. This completely ignores the dynamic changes in air density in the actual operating environment of the wind farm. Air density is a key factor determining wind power, and it fluctuates in real time with changes in temperature, air pressure, and altitude. A fixed K value cannot guarantee that the turbine will operate at its true optimal tip speed ratio most of the time. Second, even with a simple air density compensation scheme, the "theoretical" K value it relies on cannot solve the problem of the "difference between theoretical and actual engineering values." Due to manufacturing tolerances, installation differences, and especially "individual differences" caused by blade erosion and fouling after long-term operation, the optimal aerodynamic characteristics of each wind turbine have deviated from the initial design values. Existing technology lacks an online, adaptive method to optimize for the individual characteristics of each turbine, resulting in the turbine operating in a suboptimal range and causing significant power generation losses.

[0004] Ma Yanfeng et al. disclosed a "Method for Subsynchronous Oscillation Analysis of Doubly Fed Wind Farm Grid-Connected Systems Based on Complex Torque Coefficient Method" in invention patent CN202311084359.7. This method analyzes the stability of subsynchronous oscillations in power systems containing doubly fed wind turbines by establishing a complex frequency domain model and solving for the complex torque coefficient. However, this method is essentially an offline stability analysis tool, its core purpose being to diagnose oscillation risks rather than to optimize power generation performance online. It does not provide a control strategy that can adaptively adjust the optimal torque coefficient to achieve maximum power point tracking (MPPT) based on real-time changes in air density and individual unit performance differences.

[0005] Fu Lei et al. disclosed a "Wind Turbine Torque Control System Based on Actual Aerodynamic Coefficient" in invention patent CN202211738046.4. This method calculates the "actual aerodynamic coefficient" by collecting geometric data of the wind turbine blades and adjusts the torque based on power demand. However, this method relies on prior knowledge and modeling of the blade geometry, making it a model-based control. It cannot effectively address deviations between the "actual engineering value" and the "theoretical calculated value" caused by unmodeled factors such as blade fouling and wear. More importantly, it does not provide an online, data-driven self-learning mechanism to dynamically adapt to real-time air density changes in order to optimize the true maximum power point torque coefficient. Summary of the Invention

[0006] This invention provides an online optimization method and system for wind turbine torque coefficient based on air density. Deployed in the edge computing unit of the wind turbine, it constructs an intelligent agent capable of online self-learning to generate an optimal torque coefficient in real time that matches the current operating conditions (especially air density). Its purpose is to solve the technical problems of fixed torque coefficients, inability to adapt to dynamic changes in air density, and inability to compensate for individual turbine performance differences in existing technologies.

[0007] Firstly, the present invention aims to provide an online optimization method for the optimal torque coefficient of a wind turbine that incorporates air density, comprising the following steps: S1: Construct a Markov decision process model adapted to air density, and obtain the state vector based on air density; wherein, the Markov decision process model includes a state space, an action space, and a reward function; S2: Construct an agent architecture based on dual-delay deep deterministic policy gradient, wherein the agent architecture includes a policy network, two value networks and a corresponding target network cluster; S3: Training and optimizing agent architecture based on dual-delay deep deterministic policy gradient; S4: The trained agent architecture is used to perform forward propagation calculation on the state vector obtained in S1 to obtain the optimal torque coefficient in the current state.

[0008] Furthermore, a preferred solution is provided: the state space is all the environmental information on which the agent makes decisions, including dynamically changing air density, current generator speed, rate of change of generator speed, current output power, and torque coefficient value output by the agent in the previous control cycle. The action space is a set of operations performed by the intelligent agent, used to output the torque coefficient; The reward function is a multi-objective reward function.

[0009] Furthermore, an optimal solution is provided: the reward function is expressed as: , in, Indicates the execution of an action Later Instantaneous power measured at any given time; It is the square of the rate of change of speed, which is used as a penalty term to suppress drastic fluctuations in speed and prevent the optimization process from being too aggressive and causing unit oscillation; This represents the positive weighting coefficient used to balance the generation gain; This represents the positive weighting coefficient used to balance the stability of power generation.

[0010] Furthermore, a preferred solution is provided: the policy network is a deep multilayer perceptron, whose input layer dimension matches the dimension of the state vector, used to map the observed state vector to an optimal torque coefficient.

[0011] Furthermore, a preferred solution is provided: the two value networks have the same structure and independent parameters, and both value networks are deep multilayer perceptrons.

[0012] Furthermore, a preferred embodiment is provided: the target network cluster includes a target policy network and two target value networks, and the parameters of the target network cluster are replicated from the main network using a Polyak averaging method.

[0013] Furthermore, a preferred embodiment is provided: S3 includes: S31: The agent architecture observes the rewards and the state vector of the next moment from the environment and stores them in the experience replay pool; S32: Randomly sample a small batch of experience tuples from the experience replay pool, calculate the target action, and introduce target policy smoothing noise to prevent the policy from overfitting to the Q-value function. Calculate the target Q-value and the final temporal difference target value using two target value networks. Update their parameters by minimizing the mean squared error loss of the two value networks. S33: Update the parameters of the policy network by maximizing the output of any one of the value networks.

[0014] Secondly, the purpose of this invention is to propose an online optimization system for the optimal torque coefficient of a wind turbine that integrates air density. This system is based on an online optimization method for the optimal torque coefficient of a wind turbine that integrates air density, as described in any one or more of the above-mentioned schemes. The system includes: Markov Decision Module: Used to construct a Markov decision process model adapted to air density, and obtain a state vector based on air density; wherein, the Markov decision process model includes a state space, an action space, and a reward function; Agent architecture module: used to construct an agent architecture based on dual-delay deep deterministic policy gradient, the agent architecture including a policy network, two value networks and a corresponding target network cluster; Architecture training module: used to train and optimize agent architecture based on dual-latency deep deterministic policy gradient; Coefficient optimization module: Used to perform forward propagation calculation on the state vector obtained by S1 using the trained agent architecture to obtain the optimal torque coefficient in the current state.

[0015] Thirdly, the present invention aims to provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, and when the processor runs the computer program stored in the memory, the processor executes an online optimization method for the optimal torque coefficient of a fan based on any one or more of the above-described schemes.

[0016] Fourthly, the present invention aims to provide a computer-readable storage medium for storing a computer program that executes the online optimization method for the optimal torque coefficient of a wind turbine based on air density, as described in any one or more of the above-described schemes.

[0017] Compared with the prior art, the advantages of the present invention are: First, this invention achieves dynamic adaptation to changes in air density. By using real-time air density as a key dimension of the reinforcement learning state space, the control strategy no longer relies on fixed parameter tables or simple compensation formulas. The agent can learn the complex nonlinear relationship between air density and the optimal torque coefficient, enabling it to drive the unit to operate close to the actual optimal tip speed ratio under any atmospheric conditions, thereby significantly improving power generation.

[0018] Secondly, this invention addresses the issues of individual unit differences and performance degradation. The algorithm is entirely data-driven; it learns from the actual operating data of each unit and optimizes accordingly. The value is the optimal solution for the "actual engineering value" (such as blade wear and fouling) of the unit, rather than a universal "theoretical value," perfectly solving the performance difference problem between different unit locations. Finally, this invention is a closed-loop online optimization system with high robustness and autonomy. While pursuing maximum power generation, it also considers the stability of unit operation and ensures operational safety through the design of a penalty term in the reward function.

[0019] This invention is applicable to torque control scenarios for wind turbine generators. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0021] Figure 1 This is a framework diagram of an online optimization system for the optimal torque coefficient of a wind turbine that integrates air density, as described in a specific embodiment of the present invention. Figure 2 This is a flowchart illustrating an online optimization method for the optimal torque coefficient of a wind turbine that incorporates air density, as described in a specific embodiment of the present invention. Figure 3 This is a schematic diagram of the TD3 agent model architecture according to a specific embodiment of the present invention. Detailed Implementation

[0022] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application can also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0023] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0025] Implementation Method 1 An online optimization method for the optimal torque coefficient of a wind turbine, incorporating air density, is deployed in the edge computing unit of the wind turbine. This method constructs an intelligent agent capable of online self-learning to generate, in real time, the optimal torque coefficient K that matches the current operating conditions (especially air density). The method includes the following steps: S1: Construct a Markov Decision Process (MDP) model adapted to air density, obtaining the state vector based on air density. The purpose of this step is to mathematically abstract the MPPT torque coefficient optimization problem of wind turbines into a Markov Decision Process (MDP), which is a prerequisite for subsequent reinforcement learning algorithms. This process includes precise definitions of the state space, action space, and reward function.

[0026] The Markov decision process model includes a state space. Action space and reward function ; The state space The environmental information upon which the agent bases its decisions includes dynamically changing air density, current generator speed, rate of change of generator speed, current output power, and torque coefficient value output by the agent in the previous control cycle, forming a specific state vector. exist A moment is defined as .in It is the air density calculated in real time using high-precision meteorological sensors or atmospheric state equations; This is the current generator speed; It is the rate of change of generator speed, i.e. , used to characterize the dynamic processes of a system; It is the output power at the current moment; This is the torque coefficient value output by the agent in the previous control cycle. Its introduction allows the agent to perceive the continuous impact of its previous action.

[0027] The action space This is the set of operations performed by the agent to output the torque coefficient. The innovation of this step lies in the fact that the agent does not directly output the torque value, but instead outputs the core parameter of the optimal torque control law, namely the torque coefficient. .action Defined as a preset security range Continuous values ​​within Therefore, the agent's task is to learn an optimal policy. That is, in a given state At that time, the output is the optimal torque coefficient that maximizes long-term returns. .

[0028] The reward function is a scalar signal used to evaluate the quality of the agent's actions. To maintain stable unit operation while increasing power generation, this step designs the reward function as a multi-objective reward function. A moment is defined as: , in, Indicates the execution of an action Later Instantaneous power measured at any given time; It is the square of the rate of change of speed, which is used as a penalty term to suppress drastic fluctuations in speed and prevent the optimization process from being too aggressive and causing unit oscillation; This represents the positive weighting coefficient used to balance the generation gain; This represents the positive weighting coefficient used to balance the stability of power generation.

[0029] S2: Construct an agent architecture based on Twin Delayed Deep Deterministic Policy Gradient (TD3). The agent architecture includes a policy (actor) network, two value (commentator) networks, and a corresponding target network cluster. To overcome the problems of overestimation of Q-value and training instability in the traditional DDPG algorithm, this implementation preferred to adopt the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm.

[0030] The policy network is a deep multilayer perceptron, whose input layer dimension matches the dimension of the state vector, and is used to map the observed state vector to an optimal torque coefficient.

[0031] Specifically, the function of the policy network is to implement a deterministic policy function. This function directly outputs the observed state. Mapped to an optimal torque coefficient Let the parameters of the actor network be... Its policy function is expressed as In its implementation, this actor network is a deep multilayer perceptron (MLP), whose input layer dimension is the same as the state vector. Dimensions (i.e., 5 dimensions: The network contains multiple hidden layers, such as two layers of 256 neurons each, and uses a Modified Linear Unit (ReLU) as the activation function to introduce non-linearity. Its output layer is a single neuron, designed to ensure the output... Always within the preset safety range Internally, the output layer uses a scaled tanh function as the activation function, mathematically represented as follows: .

[0032] The two value networks have the same structure and independent parameters, and both value networks are deep multilayer perceptrons.

[0033] Specifically, to address the problem of overestimation of Q-values, the TD3 algorithm employs a dual-commenter network structure. This implementation constructs two identical but parameter-independent (...) and ) of critics network and Each commentator network is a deep MLP, and its input layer is state. and actions The concatenation, with input dimension as... The hidden layers (e.g., two layers of 256 neurons each, ReLU activated) process the state-action pair information, while the output layer is a single neuron with a linear activation function used to estimate the Q-value of the state-action pair. and By training two independent Q-value estimates, this implementation will employ a "Clipped Double Q-learning" strategy in the subsequent training step (S3). That is, when calculating the TD target, the smaller Q-value estimate from the two commentator networks will always be selected, thereby effectively suppressing overestimation.

[0034] The target network cluster includes a target policy network and two target value networks. The parameters of the target network cluster are replicated from the main network using a Polyak averaging method.

[0035] Specifically, to ensure the stability of the training process, this implementation creates corresponding target network clusters for the actor network and the two critic networks. This includes a target actor network. (parameter ) and two target critic networks (parameter )and (parameter ). Parameters of these target networks Instead of updating directly through backpropagation, parameters are slowly replicated from the main network using Polyak averaging (soft update). Let the soft update coefficient be... The update rules are as follows: , , , Furthermore, to further improve stability, this implementation introduces a "delayed policy update" mechanism. Specifically, two main commentator networks... and The actor network is updated in each training iteration. The three target network clusters then operate at a lower frequency (each An update is performed only after the commentator network has been updated. This delayed update ensures that the value network (commentator) converges to a more accurate Q-value estimate before the policy network (actor) is updated, thereby improving the quality and stability of the overall policy.

[0036] S3: Training and Optimizing the Agent Architecture Based on Dual-Delay Deep Deterministic Policy Gradient; This step describes the specific process of online closed-loop learning and optimization of the agent on edge computing units. This process strictly follows the core ideas of the TD3 algorithm, namely truncated dual-Q learning, delayed policy update, and target policy smoothing.

[0037] S31, Collection and storage of empirical data: the agent in... Perform actions at all times Subsequently, rewards were observed from the environment. and the state at the next moment The system will use this complete interactive experience tuple. Store in a large-capacity experience replay buffer. middle.

[0038] S32, Truncating Double-Q Learning and Critics Network Updates: In the Experience Replay Pool Randomly sample N empirical tuples in a minibatch. First, calculate the target action. And introduce target policy to smooth noise. (A truncated random noise with a mean of 0) to prevent the policy from overfitting the Q-value function: , Next, the target Q-value is calculated using two target critic networks. And it adopts the "truncation" idea, that is, taking the minimum value of the two: , Then, the final time-difference (TD) target value is calculated. : , in This is a discount factor for future rewards. Finally, by minimizing the networks of the two main commentators... and Their respective mean squared error (MSE) losses To jointly update their parameters: .

[0039] S33, Delayed Policy Gradient and Actor Network Update: According to the delayed update mechanism, only when a specific condition is met (number of training iterations) will the update be applied. Divide The actor network and target network are only updated when the parameter value is 0. The actor network parameters... By maximizing Update using network output (using (This is the standard practice for TD3), its policy gradient The calculation is as follows: , Actor Network Parameters Update along the direction of the gradient.

[0040] The parameters of the policy network are updated by maximizing the output of any one of the value networks.

[0041] S4: The trained agent architecture is used to perform forward propagation calculation on the state vector obtained in S1 to obtain the optimal torque coefficient in the current state.

[0042] Specifically, in each control cycle The edge computing unit of the wind turbine obtains information including real-time air density through the state perception module. The state vector included .Should Immediately entered into the current (already updated online) actor network middle.

[0043] The actor network performs a forward propagation calculation and outputs the optimal torque coefficient for the current state. .Should The value is sent to the wind turbine's main controller, which immediately applies this coefficient to calculate the final generator reference torque command. : , The instruction is then executed by the converter system, thereby completing a closed-loop optimal torque control based on air density adaptation. This process continues in a loop, with the agent continuously optimizing itself under the drive of S3 while executing control.

[0044] Implementation Method 2 This embodiment provides an online optimization system for the optimal torque coefficient of a wind turbine that integrates air density. The system is based on the online optimization method for the optimal torque coefficient of a wind turbine that integrates air density as described in Embodiment 1. The system includes: Markov Decision Module: Used to construct a Markov decision process model adapted to air density, and obtain a state vector based on air density; wherein, the Markov decision process model includes a state space, an action space, and a reward function; Agent architecture module: used to construct an agent architecture based on dual-delay deep deterministic policy gradient, the agent architecture including a policy network, two value networks and a corresponding target network cluster; Architecture training module: used to train and optimize agent architecture based on dual-latency deep deterministic policy gradient; Coefficient optimization module: Used to perform forward propagation calculation on the state vector obtained by S1 using the trained agent architecture to obtain the optimal torque coefficient in the current state.

[0045] Implementation Method 3 This embodiment is a further example illustrating the online optimization system for the optimal torque coefficient of a fan that integrates air density, as described in Embodiment 2.

[0046] See Figure 1 This embodiment provides an online optimization system for the optimal torque coefficient of a wind turbine based on air density. This system is built on the local edge computing unit of the wind turbine. At the physical layer, the system mainly includes a sensor layer, an edge computing unit, and the wind turbine main controller. The sensor layer is the data foundation of this embodiment and must include sensors for measuring atmospheric pressure and temperature to calculate air density in real time. In addition, it also includes a tool for measuring generator speed. and output power Conventional sensors are used. The edge computing unit is the execution core of the algorithm in this embodiment. It is a high-performance embedded computing platform that deploys an intelligent decision-making module (i.e., the actor network trained in S3) and an online learning module. This unit also has a data storage area for implementing the experience replay pool. The wind turbine main controller is the control core of a traditional wind turbine. In this embodiment, it is responsible for receiving the optimal torque coefficient sent by the edge computing unit. Commands, combined with real-time rotational speed Perform the final torque command calculation. .

[0047] Implementation Method 4 This embodiment is a further illustrative example of the online optimization method for the optimal torque coefficient of a fan that incorporates air density, as described in Embodiment 1.

[0048] See Figure 2 , Figure 2 The following is a flowchart of the specific method, which runs on the system architecture described in Embodiment 3.

[0049] After the system starts, initialization is performed first. This step is completed in the edge computing unit, including loading pre-trained network model parameters or randomly initializing the actor network. Dual-Critic Network and the parameters of their corresponding target network clusters. And clear the experience replay pool.

[0050] During the data acquisition and preprocessing stage, the system collects raw data in real time through the sensor layer, such as atmospheric pressure. Ambient temperature Generator speed and active power The preprocessing module first applies the atmospheric state equation... The key real-time air density was calculated. At the same time, for and Low-pass filtering is performed to eliminate high-frequency noise, and the rate of change of rotational speed is calculated. Finally, And the actions of the previous cycle Combine them into a vector, and then normalize it to form the final state vector. Input to the neural network.

[0051] In the model application (decision-making) phase, the state The actor network is fed into the edge computing unit for deployment. The actor network performs a fast forward propagation calculation and outputs an optimal value relative to the current air density. Matching torque coefficient value .Should The value is sent to the wind turbine's main controller via the communication bus. The main controller immediately executes the torque command, completing the closed-loop control of the wind turbine unit.

[0052] The model learning and training phase is the core of this implementation method, enabling adaptation and online optimization. The wind turbine performs actions. Then, the system at the next moment A new state was observed. The immediate reward is calculated based on the reward function defined in S1. This complete empirical tuple Experience replay pool stored in edge computing units The online learning module runs asynchronously in the background, continuously drawing data from various sources. A small batch of empirical data was randomly sampled from the data.

[0053] See Figure 3 , Figure 3 The TD3 agent model architecture used in this invention is illustrated in detail. Sampled data is used to update... Figure 3 The network in question: First, the target TD target value is calculated using the target network cluster and the truncated double-Q learning mechanism. Then, calculate The mean squared loss between the outputs of the two main commentator networks is updated via backpropagation. and The system then determines whether the delayed update condition is met, for example, after every d=2 commentator updates. If met, the actor network is updated according to the policy gradient update logic, and all target networks are soft-updated. This training process continues in a loop, executing in parallel with the control process.

[0054] During the verification and application phase, the model of this invention can undergo sufficient offline pre-training in a high-precision simulation environment to obtain a set of optimal initial parameters before being deployed to edge computing units. In practical applications, due to the continuous operation of the online learning module, this system can automatically track and adapt to the performance curve drift of wind turbines caused by factors such as blade fouling and wear, while dynamically responding to the air density at each moment. The system adapts to changes, thereby continuously and adaptively optimizing the unit's operating point within the maximum power capture range throughout its entire lifecycle.

[0055] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A method for online optimization of the optimal torque coefficient of a fan based on air density, characterized in that, Includes the following steps: S1: Construct a Markov decision process model adapted to air density, and obtain the state vector based on air density; wherein, the Markov decision process model includes a state space, an action space, and a reward function; S2: Construct an agent architecture based on dual-delay deep deterministic policy gradient, wherein the agent architecture includes a policy network, two value networks and a corresponding target network cluster; S3: Training and optimizing agent architecture based on dual-delay deep deterministic policy gradient; S4: The trained agent architecture is used to perform forward propagation calculation on the state vector obtained in S1 to obtain the optimal torque coefficient in the current state.

2. The method for online optimization of the optimal torque coefficient of a fan based on air density according to claim 1, characterized in that, The state space is all the environmental information on which the agent makes decisions, including dynamically changing air density, current generator speed, rate of change of generator speed, current output power, and torque coefficient value output by the agent in the previous control cycle. The action space is a set of operations performed by the intelligent agent, used to output the torque coefficient; The reward function is a multi-objective reward function.

3. The method for online optimization of the optimal torque coefficient of a fan based on air density according to claim 2, characterized in that, The reward function is expressed as follows: ， in, Indicates the execution of an action Later Instantaneous power measured at any given time; It is the square of the rate of change of speed, which is used as a penalty term to suppress drastic fluctuations in speed and prevent the optimization process from being too aggressive and causing unit oscillation; This represents the positive weighting coefficient used to balance the generation gain; This represents the positive weighting coefficient used to balance the stability of power generation.

4. The method for online optimization of the optimal torque coefficient of a fan based on air density according to claim 1, characterized in that, The policy network is a deep multilayer perceptron, whose input layer dimension matches the dimension of the state vector, and is used to map the observed state vector to an optimal torque coefficient.

5. The method for online optimization of the optimal torque coefficient of a fan based on air density according to claim 1, characterized in that, The two value networks have the same structure and independent parameters, and both value networks are deep multilayer perceptrons.

6. The method for online optimization of the optimal torque coefficient of a fan based on air density according to claim 1, characterized in that, The target network cluster includes a target policy network and two target value networks. The parameters of the target network cluster are replicated from the main network using a Polyak averaging method.

7. The method for online optimization of the optimal torque coefficient of a fan based on air density according to claim 6, characterized in that, S3 includes: S31: The agent architecture observes the rewards and the state vector of the next moment from the environment and stores them in the experience replay pool; S32: Randomly sample a small batch of experience tuples from the experience replay pool, calculate the target action, and introduce target policy smoothing noise to prevent the policy from overfitting to the Q-value function. Calculate the target Q-value and the final temporal difference target value using two target value networks. Update their parameters by minimizing the mean squared error loss of the two value networks. S33: Update the parameters of the policy network by maximizing the output of any one of the value networks.

8. An online optimization system for the optimal torque coefficient of a fan, incorporating air density, characterized in that, The system is based on an online optimization method for the optimal torque coefficient of a wind turbine that integrates air density, as described in any one of claims 1-7. The system includes: Markov Decision Module: Used to construct a Markov decision process model adapted to air density, and obtain a state vector based on air density; wherein, the Markov decision process model includes a state space, an action space, and a reward function; Agent architecture module: used to construct an agent architecture based on dual-delay deep deterministic policy gradient, the agent architecture including a policy network, two value networks and a corresponding target network cluster; Architecture training module: used to train and optimize agent architecture based on dual-latency deep deterministic policy gradient; Coefficient optimization module: Used to perform forward propagation calculation on the state vector obtained by S1 using the trained agent architecture to obtain the optimal torque coefficient in the current state.

9. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes an online optimization method for the optimal torque coefficient of a wind turbine based on integrated air density, according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program that executes an online optimization method for the optimal torque coefficient of a wind turbine based on air density, as described in any one of claims 1-7.

Images