Wind power plant primary frequency modulation active power distribution method based on near-end strategy optimization

Abstract

The invention discloses a wind power plant primary frequency modulation active power distribution method based on near-end strategy optimization. The method comprises the following steps: obtaining a pre-trained PPO model; the PPO model is obtained by carrying out iterative training on a near-end strategy optimization algorithm model based on acquired operation state data of each wind turbine generator of the wind power plant and the frequency of a power grid accessed by the wind power plant, and is used for realizing an optimal solution of a power distribution strategy of each wind turbine generator under the target of minimizing the frequency deviation of the power grid and minimizing the frequency adjustment time; the operation state data comprises the rotating speed, the pitch angle and the output active power of the wind turbine generator; and acquiring operation state data and power grid frequency of each wind turbine generator of the wind power plant in real time, preprocessing, inputting into the PPO model, generating a frequency modulation power distribution instruction of each wind turbine generator as a power distribution strategy, and issuing the power distribution strategy to each wind turbine generator for execution. Active power distribution of the wind turbine generator in a complex power grid frequency regulation environment can be realized, and the primary frequency regulation capability of a wind power plant is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system control and optimization, and in particular relates to a method for allocating active power of primary frequency regulation in a wind farm based on proximal strategy optimization. Background Art

[0002] With the continuous expansion of wind power generation, wind farms, as an important part of the new power system, have attracted much attention for their ability to participate in grid frequency regulation. In conventional thermal power and hydropower, the units have large inertia and strong frequency response capabilities, but wind turbines have insufficient inertial response capabilities due to their lightweight rotors and uncontrollable wind speed input. When the load on the grid changes, if the active power output cannot be adjusted in time through primary frequency regulation response, it may cause instability in the grid frequency and even cause large-scale grid failures. For this reason, wind farms are required to have the ability of primary frequency regulation to quickly respond to fluctuations in grid frequency and maintain the dynamic stability of the system.

[0003] However, due to the volatility and intermittency of wind energy, the application of traditional primary frequency regulation methods in wind farms faces many difficulties. Existing control strategies are often based on fixed frequency regulation curves or models, which are difficult to adapt to the rapid changes in external environments such as wind speed and load in real time, resulting in delayed or ineffective frequency regulation response. In addition, insufficient coordination between wind turbines and frequent participation of a single unit in frequency regulation may lead to low frequency response efficiency and unreasonable power distribution, thereby exacerbating system instability. Existing algorithms are difficult to take into account both the efficiency of frequency regulation and the dynamic characteristics of wind turbines at the same time, which puts higher requirements on large-scale scheduling and control of wind farms. Summary of the invention

[0004] In order to solve the above problems existing in the prior art, the present invention provides a method for allocating active power of primary frequency modulation in a wind farm based on proximal strategy optimization. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0005] A wind farm primary frequency modulation active power allocation method based on proximal strategy optimization is applied to a wind farm central control system as an intelligent agent hardware carrier, the method comprising:

[0006] Acquire a pre-trained PPO model; wherein the PPO model is obtained by iteratively training a proximal strategy optimization algorithm model based on the collected operating status data of each wind turbine in the wind farm and the grid frequency of the wind farm, and the PPO model is used to achieve the optimal solution of the power allocation strategy of each wind turbine under the goal of minimizing the grid frequency deviation and minimizing the frequency adjustment time; the operating status data includes the speed, pitch angle and output active power of the wind turbine;

[0007] The operating status data of each wind turbine in the wind farm and the grid frequency of the wind farm are collected in real time, and input into the PPO model after preprocessing to generate frequency modulation power allocation instructions for each wind turbine as a power allocation strategy, and send them to each wind turbine for execution.

[0008] In one embodiment of the present invention, the frequency modulation power allocation instruction of any wind turbine generator set includes:

[0009] Power increase command or power decrease command.

[0010] In one embodiment of the present invention, the rotational speed, pitch angle and output active power of any wind turbine are respectively detected by a rotational speed sensor, a pitch angle measuring instrument and a power sensor deployed on the wind turbine; the grid frequency is detected by a grid frequency monitoring device deployed at the grid connection point.

[0011] In one embodiment of the present invention, the training process of the PPO model includes:

[0012] Carry out data collection, including collecting the operating status data of each wind turbine in the wind farm and the grid frequency of the wind farm, and obtain the status information after preprocessing;

[0013] Input the state information into the proximal strategy optimization algorithm model after parameter initialization, generate power adjustment instructions for each wind turbine group as corresponding action information through its strategy network, and send them to each wind turbine group for execution;

[0014] Calculate the reward value according to the frequency modulation effect after each wind turbine executes the corresponding action information;

[0015] Based on the reward value, performing network optimization on the strategy network and the value network;

[0016] The process from data collection to network optimization is repeated until the model converges to obtain a trained PPO model.

[0017] In one embodiment of the present invention, the preprocessing comprises:

[0018] Data cleaning and data normalization.

[0019] In one embodiment of the present invention, the parameter initialization of the proximal policy optimization algorithm model includes:

[0020] Randomly initialize the parameters θ of the policy network and φ of the value network in the proximal policy optimization algorithm model, set the weights α and β of the reward function, and set the discount factor γ.

[0021] In one embodiment of the present invention, the calculation formula of the reward value is expressed as:

[0022] R t =αR freq +βR Δt ;

[0023] R freq =-|f t -50|;

[0024] Among them, R t is the reward value corresponding to time t; R freq is the frequency deviation; R Δt is the adjustment time of the grid frequency; f t is the grid frequency corresponding to time t.

[0025] In one embodiment of the present invention, after generating a frequency modulation power allocation instruction for each wind turbine as a power allocation strategy and sending it to each wind turbine for execution, the method further includes:

[0026] The operating status data of each wind turbine in the wind farm, the grid frequency of the wind farm, and the frequency modulation effect of the frequency modulation power allocation instruction of each wind turbine are continuously collected, and the data are fed back into the intelligent agent to generate a new state-action-reward data set. The PPO model is incrementally trained using the new state-action-reward data set, thereby realizing online updating and adaptive optimization of the intelligent agent.

[0027] Beneficial effects of the present invention:

[0028] The embodiment of the present invention provides a method for allocating active power for primary frequency modulation of a wind farm based on proximal strategy optimization, and uses the PPO algorithm in deep reinforcement learning to establish a power allocation strategy model for primary frequency modulation of a wind farm. By acquiring the speed, pitch angle, output active power and grid frequency of each wind turbine in the wind farm in real time, the power allocation strategy of each wind turbine in the primary frequency modulation process is optimized, aiming to improve the overall frequency response capability and power regulation efficiency of the wind farm. Compared with traditional methods, the PPO algorithm has stronger adaptability and robustness, can achieve coordinated optimization between wind turbines under different environmental and load conditions, improve the accuracy and speed of primary frequency modulation response, and maximize the frequency modulation efficiency of the wind farm and the stability of the grid. At the same time, the method reduces unnecessary frequency modulation frequency of the unit, reduces equipment wear and maintenance costs. The present invention is particularly suitable for dynamic active power regulation control during primary frequency modulation of a wind farm, and has broad application prospects in frequency stability and grid-connected control of new energy power systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a schematic diagram of the active power-frequency droop characteristic curve of a wind farm;

[0030] Figure 2It is a flow chart of a method for allocating active power of primary frequency regulation in a wind farm based on proximal strategy optimization provided by an embodiment of the present invention;

[0031] Figure 3 is a signal acquisition framework diagram of an embodiment of the present invention;

[0032] Figure 4 is a schematic diagram of the model training process of an embodiment of the present invention;

[0033] Figure 5 It is a schematic diagram of the use process of the wind farm primary frequency modulation active power allocation method based on proximal strategy optimization provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0034] The present invention is further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0035] See also Figure 1 Understand the actual needs of primary frequency regulation in wind farms, Figure 1 A simplified frequency regulation curve of a single wind turbine is given. When the system frequency changes from the rated value f N Reduce (i.e. from the rated value f N When the system frequency moves to f′, the moving direction is B→A), the active power of the wind turbine is increased to maintain the frequency stability; on the contrary, if the system frequency is higher than the rated value (i.e., from the rated value f N Move toward f″, moving direction B→D), then the active power output of the wind turbine is reduced. When there are multiple units in the wind farm participating in frequency regulation, coordinated control of multiple units is required to jointly achieve frequency stability of the system.

[0036] However, existing algorithms are difficult to take into account both the efficiency of frequency regulation and the dynamic characteristics of wind turbines, which places higher demands on the large-scale scheduling and control of wind farms.

[0037] In view of the above problems, an embodiment of the present invention provides a wind farm primary frequency regulation active power allocation method based on proximal policy optimization (PPO) algorithm, which is applied to the wind farm central control system as an intelligent hardware carrier to ensure efficient active power allocation in a complex power grid frequency regulation environment.

[0038] Among them, the PPO algorithm is a widely used policy gradient algorithm in deep reinforcement learning. It combines the flexibility of the policy gradient method and the stability of the trust region optimization method, and is an important breakthrough in the field of policy optimization. This enables the proximal policy optimization algorithm to maintain learning efficiency and high stability when achieving global optimization.

[0039] like Figure 2 As shown, the wind farm primary frequency modulation active power allocation method based on proximal strategy optimization in an embodiment of the present invention may include the following steps:

[0040] S1, obtain the pre-trained PPO model;

[0041] The PPO model is obtained by iteratively training the proximal strategy optimization algorithm model based on the collected operating status data of each wind turbine in the wind farm and the grid frequency of the wind farm. The PPO model is used to achieve the optimal solution of the power allocation strategy of each wind turbine under the goal of minimizing the grid frequency deviation and minimizing the frequency adjustment time; the operating status data includes the speed, pitch angle and output active power of the wind turbine;

[0042] In order to facilitate understanding of the embodiments of the present invention, the training process of the PPO model is first described here.

[0043] Specifically, the training process of the PPO model includes:

[0044] Step 1, performing data collection, including collecting the operating status data of each wind turbine in the wind farm and the grid frequency of the wind farm, and obtaining status information after preprocessing;

[0045] Step 1 is mainly for data collection and preprocessing. The signal collection framework is as follows: Figure 3 As shown. This framework requires high-precision monitoring equipment to be configured on each wind turbine in the wind farm, including a speed sensor, a pitch angle meter, and a power sensor, which are used to collect the operating status data of the wind turbine in real time to characterize the frequency regulation capability of the wind turbine. Specifically, the speed sensor is used to collect the speed of the wind turbine, the pitch angle meter is used to collect the pitch angle of the wind turbine, and the power sensor is used to collect the output active power of the wind turbine.

[0046] At the same time, grid frequency monitoring equipment is deployed at the grid connection point to collect the frequency of wind farms connected to the grid (referred to as grid frequency) to characterize the real-time operation status of the grid. The sampling accuracy of the grid frequency monitoring equipment is required to reach ±0.003Hz, and the sampling period is not more than 20ms to ensure high-precision and low-latency acquisition of grid frequency change information, and provide accurate real-time data support for frequency modulation control.

[0047] The data collected by the above equipment is transmitted to the central control system of the wind farm through a high-speed communication network. The transmission delay is required to be less than 20ms to ensure real-time frequency regulation requirements.

[0048] The central control system is the hardware form of the intelligent body, which can be a wind farm server; the proximal strategy optimization algorithm model is deployed in the intelligent body. The data transmitted to the central control system needs to be preprocessed, which includes data cleaning and data normalization, and the status information is obtained after preprocessing. Among them, data cleaning is to remove noise or outliers to ensure the accuracy of the data; data normalization is to unify the data dimension so that it is suitable for input into the proximal strategy optimization algorithm model, which is convenient for subsequent model training and application.

[0049] Step 2, inputting the state information into the parameter-initialized proximal strategy optimization algorithm model, generating power adjustment instructions for each wind turbine as corresponding action information through its strategy network, and sending them to each wind turbine for execution;

[0050] The state information is a proximal policy optimization algorithm model initialized according to the time step input parameters, and the model selects an action output at a corresponding time step based on the input state.

[0051] For the model training process, please refer to Figure 4 The core of the training process is to generate the frequency regulation power allocation instructions of the wind turbines through the policy network in the proximal policy optimization algorithm model, and optimize the parameters of the policy network and value network of the proximal policy optimization algorithm model through the reward mechanism to achieve dynamic optimization of the primary frequency regulation active power allocation of the wind farm.

[0052] For step 2, it is first necessary to initialize the parameters of the proximal policy optimization algorithm model. Specifically, the parameter initialization of the proximal policy optimization algorithm model includes:

[0053] Randomly initialize the parameters θ of the policy network and φ of the value network in the proximal policy optimization algorithm model, set the weights α and β of the reward function, and set the discount factor γ.

[0054] Then, the state information is input into the parameter-initialized proximal strategy optimization algorithm model, and the corresponding action information is output through the strategy network of the proximal strategy optimization algorithm model, that is, the power adjustment instructions of each wind turbine are generated. The frequency modulation power allocation instruction of any wind turbine includes:

[0055] Power increase command or power decrease command.

[0056] The strategy network uses a multi-layer neural network structure. After inputting the above state information, the strategy network outputs the action information of each wind turbine. In practical applications, the action information can indicate continuous power adjustment or a discrete power adjustment scheme under specific conditions. It can be expressed as:

[0057] a t =π θ(S t );

[0058] Among them, S t Represents the state information corresponding to time t; π θ represents the policy network, whose parameter is θ; a t Represents the action information corresponding to time t.

[0059] The output power adjustment command is transmitted to each wind turbine through the wind farm central control system, ensuring that the power adjustment command can be responded to within 2 seconds and the adjustment can be completed within 15 seconds, so as to ensure that the wind farm can quickly and effectively adjust the grid frequency and avoid large frequency fluctuations.

[0060] Step 3, calculating the reward value according to the frequency modulation effect after each wind turbine executes the corresponding action information;

[0061] After the policy network generates action information and executes it, the proximal policy optimization algorithm model calculates the reward value based on the FM effect. The reward mechanism is designed to optimize the FM performance.

[0062] In the embodiment of the present invention, the factors considered in the reward value include the adjustment time of the grid frequency, that is, the time taken from the start of adjustment to the completion of adjustment, and the frequency deviation of the grid frequency, that is, the frequency regulation accuracy of the wind turbine. Specifically, the smaller the frequency deviation, the higher the reward value; if the grid frequency returns to the target value in a shorter time, that is, the shorter the adjustment time, the system will give a higher reward. The proximal strategy optimization algorithm model performs strategy optimization according to the reward value, strengthens efficient power allocation decisions, and reduces invalid or unreasonable frequency regulation actions.

[0063] Specifically, the calculation formula of the reward value is expressed as:

[0064] R t =αR freq +βR Δt ;

[0065] R freq =-|f t -50|;

[0066] Among them, R t is the reward value corresponding to time t; R freq is the frequency deviation; R Δt is the adjustment time of the grid frequency; f t is the grid frequency corresponding to time t, and 50 is the target value of the grid frequency.

[0067] Step 4: Based on the reward value, perform network optimization on the strategy network and the value network;

[0068] After obtaining the reward value, the proximal policy optimization algorithm model updates the parameters of the policy network and the value network through optimization. In the embodiment of the present invention, the weight of the policy network is continuously adjusted by the PPO algorithm to ensure that the policy optimization after each frequency adjustment is based on the result of real-time feedback from the system. In addition, the clipping mechanism is used to control the amplitude of the policy update to avoid excessive model updates that lead to system instability. Specifically:

[0069] First, the proximal policy optimization algorithm model ensures the stability of training by limiting the policy update range. The optimization objective function of the policy network is set to:

[0070]

[0071] Among them, L CLIP (θ) is the objective function of the policy network; r t (θ) is the probability ratio of the new and old strategies; is the advantage function, which is used to measure the superiority of the current action; S t is the state information corresponding to time t; ∈ is a preset constant, usually 0.1 or 0.2; V θ (·) represents the policy network; It means to find the mathematical expectation.

[0072] Subsequently, the proximal policy optimization algorithm model updates the value network parameters by minimizing the prediction error of the state value. The objective function of the value network is set to:

[0073]

[0074] Among them, L value (φ) represents the objective function of the value network; V φ (·) represents the value network.

[0075] Finally, the policy network and value network are optimized separately using the gradient descent method (such as the Adam optimizer), which can be expressed as:

[0076]

[0077] Among them, η represents the learning rate; Indicates the calculation of the gradient, and ← indicates a change to .

[0078] Step 5: Repeat the process from data collection to network optimization until the model converges to obtain the trained PPO model.

[0079] Specifically, the above steps 1 to 4 are repeated until the model converges to ensure that the policy network can generate the optimal power allocation instruction. At this time, the trained proximal policy optimization algorithm model is recorded as a PPO model.

[0080] The specific processing process of the PPO model can be understood in conjunction with the existing proximal strategy optimization algorithm.

[0081] S2, collecting the operating status data of each wind turbine in the wind farm and the grid frequency of the wind farm in real time, inputting them into the PPO model after preprocessing, generating the frequency modulation power allocation instructions of each wind turbine as the power allocation strategy, and issuing them to each wind turbine for execution.

[0082] Specifically, the trained PPO model (i.e., intelligent agent) is applied to wind farm frequency regulation. The application process is as follows: Figure 5 The intelligent agent receives the operating status data of each wind turbine in the wind farm and the grid frequency of the wind farm in real time. As described in the previous training process, the speed, pitch angle and output active power of any wind turbine are detected by the speed sensor, pitch angle measuring instrument and power sensor deployed on the wind turbine respectively; the grid frequency is detected by the grid frequency monitoring equipment deployed at the grid connection point.

[0083] The collected data is preprocessed and then input into the PPO model as state information. Based on the current state input, the agent generates the optimal frequency modulation power allocation instruction (i.e. Figure 5 The power deviation command in the power grid is used to ensure that each wind turbine responds to the grid frequency fluctuation in a coordinated manner. The frequency modulation power allocation command output by the intelligent agent is sent to each wind turbine through the central control system of the wind farm, and responds within 1 second and completes power adjustment within 5 to 10 seconds to avoid the impact of large fluctuations on the grid and ensure the rapidity and stability of frequency modulation.

[0084] Furthermore, after generating a frequency modulation power allocation instruction for each wind turbine as a power allocation strategy and issuing it to each wind turbine for execution, the method further includes:

[0085] The operating status data of each wind turbine in the wind farm, the grid frequency of the wind farm, and the frequency modulation effect of the frequency modulation power allocation instruction of each wind turbine are continuously collected, and the data are fed back into the intelligent agent to generate a new state-action-reward data set. The PPO model is incrementally trained using the new state-action-reward data set, thereby realizing online updating and adaptive optimization of the intelligent agent.

[0086] Specifically, in the actual operation of a wind farm, environmental factors such as wind speed and grid load may change rapidly, which requires the agent to dynamically adjust its frequency regulation strategy to adapt to the new environment. Therefore, the agent has an online update function, which can adaptively optimize the strategy network and value network by continuously acquiring status information.

[0087] During the online update process, the newly collected wind farm status data (such as wind turbine speed, pitch angle, output power, grid frequency, etc.) and frequency modulation effect feedback are regularly input into the agent to generate a new state-action-reward data set. The PPO model uses these new data to incrementally train the policy network and value network (for details, please refer to the training process above), while ensuring that the update range is limited to avoid model instability.

[0088] Online updates introduce a real-time feedback mechanism that continuously monitors changes in grid frequency and the power output of wind turbines. If it is detected that the frequency regulation effect does not meet expectations, the power allocation plan will be adjusted immediately and the frequency regulation strategy will be re-optimized. Through long-term accumulated feedback data, the frequency regulation effect is gradually optimized, so that the wind farm can continuously optimize the frequency regulation strategy in the case of drastic changes in wind speed or grid load fluctuations, ensuring that the frequency regulation capability of the wind farm always remains efficient and stable.

[0089] As online updates proceed, the agent can accumulate more status data and continuously improve its adaptability and robustness. In long-term operation, the agent can further improve frequency regulation efficiency and reduce equipment losses through deep optimization of accumulated data, providing a more accurate and reliable frequency regulation solution for wind farms.

[0090] It can be seen that this method has strong adaptive capabilities, can adapt to different environments and load conditions, and can maintain real-time adjustments during the actual operation of the wind farm, so that the system can still maintain high frequency regulation efficiency and stability under different working conditions.

[0091] In summary, in order to solve the problems of slow frequency modulation response speed, unreasonable unit power allocation, low frequency modulation efficiency, etc. in the prior art, the embodiment of the present invention provides a method for allocating active power for primary frequency modulation of a wind farm based on proximal strategy optimization, and uses the PPO algorithm in deep reinforcement learning to establish a power allocation strategy model for primary frequency modulation of a wind farm. By acquiring the speed, pitch angle, output active power and grid frequency of each wind turbine in the wind farm in real time, the power allocation strategy of each wind turbine in the primary frequency modulation process is optimized, aiming to improve the overall frequency response capability and power regulation efficiency of the wind farm. Compared with traditional methods, the PPO algorithm has stronger adaptability and robustness, can achieve coordinated optimization between wind turbines under different environments and load conditions, improve the accuracy and speed of primary frequency modulation response, and maximize the frequency modulation efficiency of the wind farm and the stability of the power grid. At the same time, the method reduces unnecessary unit frequency modulation frequency, reduces equipment wear and maintenance costs. The present invention is particularly suitable for dynamic active power regulation control during primary frequency modulation of wind farms, and has broad application prospects in frequency stability and grid-connected control of new energy power systems.

[0092] Specifically, the present invention has the following beneficial effects:

[0093] 1. By using the proximal policy optimization (PPO) algorithm, the present invention can adapt to changes in wind speed and grid load in real time, and quickly distribute active power instructions to each wind turbine, significantly improving the response speed of primary frequency regulation of the wind farm and reducing frequency deviation;

[0094] 2. The present invention reasonably distributes power output based on the real-time operating status and frequency regulation capability of each wind turbine, avoiding the problem of uneven power distribution or excessive participation of a single unit in traditional frequency regulation methods, and improving the overall frequency regulation efficiency;

[0095] 3. By rationally allocating frequency modulation tasks, the present invention reduces the frequent participation of wind turbines in frequency modulation, effectively reduces mechanical wear of equipment, prolongs the service life of wind turbines, and reduces maintenance costs;

[0096] 4. The PPO algorithm has strong adaptability and robustness in dealing with variable wind speed and grid load environments. The present invention can work stably in complex actual operating environments, ensuring that the system can still maintain efficient frequency control when facing external disturbances;

[0097] 5. While improving the frequency regulation response capability of the wind farm, the present invention also enhances the grid connection capability of the wind farm, ensuring that the frequency stability of the power grid can be effectively guaranteed under the condition of high wind power penetration rate.

[0098] It should be noted that, in the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification.

[0099] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A method for allocating active power of primary frequency modulation in a wind farm based on proximal strategy optimization, characterized in that: Applied to a wind farm central control system as a hardware carrier of an intelligent agent, the method comprises: Acquire a pre-trained PPO model; wherein the PPO model is obtained by iteratively training a proximal strategy optimization algorithm model based on the collected operating status data of each wind turbine in the wind farm and the grid frequency of the wind farm, and the PPO model is used to achieve the optimal solution of the power allocation strategy of each wind turbine under the goal of minimizing the grid frequency deviation and minimizing the frequency adjustment time; the operating status data includes the speed, pitch angle and output active power of the wind turbine; The operating status data of each wind turbine in the wind farm and the grid frequency of the wind farm are collected in real time, and input into the PPO model after preprocessing to generate frequency modulation power allocation instructions for each wind turbine as a power allocation strategy, and send them to each wind turbine for execution.

2. The method for allocating active power of primary frequency modulation in a wind farm based on proximal strategy optimization according to claim 1 is characterized in that: The frequency regulation power allocation instructions of any wind turbine include: Power increase command or power decrease command.

3. The method for allocating active power of primary frequency modulation in a wind farm based on proximal strategy optimization according to claim 1 is characterized in that: The rotational speed, pitch angle and output active power of any wind turbine are detected by the rotational speed sensor, pitch angle measuring instrument and power sensor deployed on the wind turbine respectively; the grid frequency is detected by the grid frequency monitoring equipment deployed at the grid connection point.

4. The method for allocating active power of primary frequency modulation in a wind farm based on proximal strategy optimization according to claim 2 is characterized in that: The training process of the PPO model includes: Carry out data collection, including collecting the operating status data of each wind turbine in the wind farm and the grid frequency of the wind farm, and obtain the status information after preprocessing; Input the state information into the proximal strategy optimization algorithm model after parameter initialization, generate power adjustment instructions for each wind turbine group as corresponding action information through its strategy network, and send them to each wind turbine group for execution; Calculate the reward value according to the frequency modulation effect after each wind turbine executes the corresponding action information; Based on the reward value, performing network optimization on the strategy network and the value network; The process from data collection to network optimization is repeated until the model converges to obtain a trained PPO model.

5. The method for allocating active power of primary frequency modulation in a wind farm based on proximal strategy optimization according to claim 1 or 4, characterized in that: The preprocessing comprises: Data cleaning and data normalization.

6. The method for allocating active power of primary frequency modulation in a wind farm based on proximal strategy optimization according to claim 5 is characterized in that: The parameter initialization of the proximal strategy optimization algorithm model includes: Randomly initialize the parameters θ of the policy network and φ of the value network in the proximal policy optimization algorithm model, set the weights α and β of the reward function, and set the discount factor γ.

7. The method for allocating active power of primary frequency modulation in a wind farm based on proximal strategy optimization according to claim 4 is characterized in that: The calculation formula of the reward value is expressed as: R t =αR freq +βR Δt ; R freq =- | f t -50 | ; Among them, R t is the reward value corresponding to time t; R freq is the frequency deviation; R Δt is the adjustment time of the grid frequency; f t is the grid frequency corresponding to time t.

8. The method for allocating active power of primary frequency modulation in a wind farm based on proximal strategy optimization according to claim 1, characterized in that: After generating the frequency modulation power allocation instruction of each wind turbine set as the power allocation strategy and issuing it to each wind turbine set for execution, the method further includes: The operating status data of each wind turbine in the wind farm, the grid frequency of the wind farm, and the frequency modulation effect of the frequency modulation power allocation instruction of each wind turbine are continuously collected, and the data are fed back into the intelligent agent to generate a new state-action-reward data set. The PPO model is incrementally trained using the new state-action-reward data set, thereby realizing online updating and adaptive optimization of the intelligent agent.

Images