Wind power plant yaw control method based on wind process

Abstract

The invention relates to a wind power plant yaw control method based on a wind process. The method comprises the following steps: 1, establishing a wind power plant simulation model; 2, acquiring operation state data of the wind power plant as an initial state of the wind power plant simulation model; 3, acquiring a wind speed fluctuation sequence and a wind direction fluctuation sequence of the wind power plant as the state of the wind power plant yaw control model; 4, performing off-line learning to obtain optimal yaw angle set values under different wind regime sequences; enabling the wind power plant yaw control model to observe the environment and guide a wind turbine generator to act, and completing online learning; the wind power plant yaw control model is established, the optimal yaw angle of each wind turbine generator in the wind power plant is obtained through the wind speed sequence and the wind direction sequence, and the total generated power of the wind power plant is greatly improved; and the timeliness requirement of solving the wind power plant yaw control model in practical engineering is met.

Description

Technical Field

[0001] The present invention relates to a yaw control method for a wind farm based on a wind process, belonging to the technical field of wind power generation. Background Art

[0002] With the continuous growth of the global demand for clean energy, wind power generation, as an important form of renewable energy utilization, has developed rapidly. The scale of wind farms is expanding day by day, and numerous wind turbines are densely distributed within a limited area; during the operation of wind farms, the wake interaction between wind turbines has become a prominent problem.

[0003] In the current research on yaw control of wind farms, most consider controlling the yaw angle of wind turbines under specific wind speeds and wind directions, and fail to take into account the influence of the time series of wind speed and wind direction, reducing the optimization effect of the yaw control wake of the wind farm; and the optimal yaw angle setting values of wind turbines under different wind speeds and wind directions vary greatly. If only the average wind speed and average wind direction are relied on to set the yaw angle, the optimization result is very limited; if the yaw angle is adjusted according to the real-time fluctuating wind speed and wind direction, it is difficult to track the optimal yaw angle in time due to the limitation of the yaw speed.

[0004] Chinese invention application with publication number CN116658359A discloses a real-time collaborative yaw control method for a wind farm, including the following steps: 1) obtaining the measured wind direction and wind speed data of the wind farm and the wind turbine layout data; 2) establishing a wind farm model based on two-dimensional coordinates according to the wind turbine layout data to simulate the spatial distribution of the wind farm; 3) establishing a wind turbine model and a wind turbine wake model; 4) combining the wind farm model with the wind turbine model and the wind turbine wake model to establish a wind farm collaborative yaw control model; 5) according to the wind direction and wind speed data collected in real time by the wind farm, using the Bayesian optimization algorithm to optimize the yaw angle parameters of each wind turbine at each moment to obtain the yaw angle of each wind turbine when the wind farm generates the maximum power, that is, the optimal wind turbine collaborative yaw control strategy; 6) simulating the working conditions of a real wind farm through the yaw control model to verify the power effect of the wind farm under the optimal wind turbine collaborative yaw control strategy; the above scheme has the following problems or defects: 1. Slow response to environmental changes: The environment of the wind farm is dynamically changing, such as the real-time changes of wind speed and wind direction, etc.; the Bayesian optimization constructs a model based on historical data and responds relatively slowly to the dynamic changes of the environment. 2. Lack of real-time interaction mechanism: The interaction between the Bayesian optimization and the environment during the optimization process is relatively weak, mainly relying on the pre-set objective function and sampling strategy. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a wind farm yaw control method based on a wind process that obtains the optimal yaw angle of each wind turbine in the wind farm through a wind speed sequence and a wind direction sequence.

[0006] The present invention adopts the following technical solutions:

[0007] The yaw control method for a wind farm based on the wind process of the present invention adopts the following steps:

[0008] S1. Collect information on the wind farm environment, terrain, fan coordinates, and wind turbine types, and establish a wind farm simulation model;

[0009] S2. Obtain the operating state data of the wind farm as the initial state of the wind farm simulation model;

[0010] S3. Obtain the wind speed fluctuation sequence and wind direction fluctuation sequence of the wind farm as the state of the wind farm yaw control model;

[0011] S4. The wind farm yaw control model based on the reinforcement learning algorithm (PPO) interacts with the wind farm simulation model for offline learning to obtain the optimal yaw angle setting values under different wind condition sequences; further, the wind farm yaw control model observes the environment and guides the actions of the wind turbines to complete online learning.

[0012] The establishment of the wind farm simulation model in S1 of the present invention includes a wind turbine model, a wake model, and an optimization model, which jointly simulate the main process of the wind farm absorbing wind energy;

[0013] The wind turbine model simulates the wind turbine absorbing wind energy and adopts the C p - wind speed, C T - wind speed curve as the model;

[0014] During operation, the wind energy capture coefficient C p and the wind turbine thrust coefficient C T of the wind turbine are determined according to the wind speed;

[0015] According to the wind turbine thrust coefficient C T calculate the blocking effect of the wind turbine on the oncoming flow and transmit it to the wake model to calculate the maximum wind speed loss in the wake area; the wind energy capture coefficient C p represents the ability of the wind turbine to convert wind energy into mechanical energy and can be used to calculate the power of the wind turbine. Its calculation expression is as follows:

[0016]

[0017] Among them, A is the swept area of the wind wheel, ρ is the air density, u is the wind wheel wind speed, γ is the fan yaw angle, and C p is the wind energy capture coefficient.

[0018] The wake model of the present invention includes an unbiased wake model, a wake deflection model, and a wake superposition model; there are multiple wind turbines in a wind farm, and the downstream wind turbines are simultaneously affected by the wakes of multiple wind turbines. It is necessary to consider the superposition effect of the wake areas of multiple wind turbines and calculate the incoming flow velocity of the downstream wind turbines; the wake superposition model is used to superpose the unbiased wakes or yawed wakes of each upstream wind turbine to obtain the incoming flow velocity of the wind turbine; the unbiased wake model uses the BPA Gaussian wake model; the wake deflection model uses the Bastankhah Gaussian wake deflection model.

[0019] The wake superposition model uses the sum of squares superposition to calculate the influence of the wake superposition of multiple upstream units, and the formula is as follows:

[0020]

[0021] where, u i is the incoming flow velocity of the target unit i, N is the number of upstream N units in the wake superposition area of this unit, u ∞ is the incoming flow velocity, u j is the incoming flow velocity in front of a certain upstream unit j, and u ji is the wind speed at the position of unit i in the wake area of unit j.

[0022] The expression of the optimization model of the overall power P of the wind farm simulation model in S1 of the present invention is as follows:

[0023]

[0024] where, P i is the wind turbine power of the i-th unit, C pi is the wind energy capture coefficient of the i-th unit, u i is the wind turbine speed of the i-th unit, and γ i is the yaw angle of the i-th unit.

[0025] In S2 of the present invention, the initial state of the wind farm simulation model needs to obtain the parameters of each wind turbine, including power, yaw angle, rotational speed, and pitch angle for setting.

[0026] In S3 of the present invention, the wind speed fluctuation sequence and wind direction fluctuation sequence of the wind farm are divided into the current wind condition sequence and the wind condition sequence in the future time period; the current wind condition sequence uses the lidar data, nacelle anemometer data, and nacelle lidar data in the wind farm, and the measured data is used after cleaning and filtering; the wind condition sequence in the future time period needs to build a new anemometer tower around the wind farm, and the wind conditions in a wider range are used as the wind conditions in the future time period of the wind farm; the wind speed fluctuation sequence and wind direction fluctuation sequence are defined as (ws 0 ,…,ws t-1 ,wd 0,…,wd t-1 ), where ws 0 ,…,ws t-1 represents the wind speed fluctuation sequence in t minutes, and wd 0 ,…,wd t-1 represents the wind direction fluctuation sequence in t minutes.

[0027] In the offline learning stage of S4 of the present invention, with the historical wind speed, wind direction sequence and the initial yaw angle as the input, the agent interacts with the wind farm simulation model, aiming to learn to control the yaw angle of the wind turbine to maximize the total power generation under different wind speed and wind direction conditions; initially, experience is accumulated by randomly selecting yaw angle actions, the agent observes the environmental state and outputs actions; after sufficient experience, the policy network (Actor) and value network (Critic) based on the reinforcement learning algorithm (PPO) are trained according to the sampled data to optimize the policy, and at the same time, the experience is stored in the data buffer for subsequent training and policy improvement;

[0028] In the online learning stage, the agent is placed in the actual wind farm environment, monitors the environmental conditions in real time, and commands the adjustment of the yaw action of the wind turbine according to the policy learned based on the reinforcement learning algorithm (PPO); during this period, the agent continuously interacts with the environment, collects new data and stores it in the experience set, and slightly updates the policy network and value network in combination with the real-time data to adapt to environmental changes; by continuously updating the policy and value function, the online learning can dynamically optimize the yaw angle setting.

[0029] In step S4 of the present invention, the construction of the wind farm yaw control model based on the reinforcement learning algorithm (PPO) is carried out by the following steps;

[0030] S401. Determine the structures of the policy network (Actor) and value network (Critic). The policy network (Actor) uses a multi-layer perceptron. The input layer is determined according to the state space, multiple hidden layers and activation functions are set, and the output layer outputs the action probability through softmax according to the action space; the value network is similar, and the state value of a single neuron is output.

[0031] S402. Collect the data of the interaction between the agent and the environment. Set an experience replay buffer during offline learning, calculate the advantage function for the sampled data, update the policy network and value network respectively using the objective function and mean square error loss function based on the reinforcement learning algorithm (PPO) and perform gradient descent with the optimizer; in online learning, deploy the wind farm yaw control model to the actual environment, the agent selects actions according to the policy and executes them, collects new data into the experience pool, and periodically extracts and slightly updates the two networks in combination with the real-time data to adapt to environmental changes.

[0032] The initial state of the wind turbine of the present invention is the shutdown state, the power is 0 kW, the yaw angle is 0°, the rotational speed is 0 rpm, and the pitch angle is 90°.

[0033] The positive effects of the present invention are as follows: The present invention establishes a yaw control model for a wind farm, and obtains the optimal yaw angles of each wind turbine in the wind farm through the wind speed sequence and the wind direction sequence. Compared with the traditional control method for average wind speed and average wind direction, the total power generation of the wind farm is greatly improved; at the same time, the online learning of the yaw control of the wind farm can utilize the data feedback in actual operation to reduce the influence of the uncertainty of wind conditions, wake calculation, integrated model, etc. on the improvement effect and improve the power generation efficiency. The present invention meets the timeliness requirements for solving the yaw control model of the wind farm in actual engineering, and avoids the problems that it is difficult to meet the timeliness requirements when using online optimization for a long time, and the optimization accuracy is low for wind conditions that have not been pre-calculated when using offline optimization. Through the method of the present invention, the power generation of the wind farm can be better than that of the traditional maximum power point tracking (MPPT) control, effectively providing the optimal yaw angles of each wind turbine during the operation of the wind farm, and being able to perform real-time optimization according to the changes in the wind sequence, improving the power generation efficiency of the wind farm and making the overall power generation of the wind farm reach the maximum value. Description of the Drawings

[0034] Appendix Figure 1 is the flow chart of the method of the present invention;

[0035] Appendix Figure 2 is C p - wind speed curve;

[0036] Appendix Figure 3 is C T - wind speed curve;

[0037] Appendix Figure 4 is the schematic diagram of the yaw control system for a wind farm based on PPO;

[0038] Appendix Figure 5 is the example diagram of power improvement relative to MPPT after offline training. Detailed Embodiments

[0039] The present invention will be described in detail below in conjunction with the appendix Figures 1-5 :

[0040] As shown in the appendix Figure 1 , the yaw control method for a wind farm based on the wind process of the present invention comprises the following steps:

[0041] Step S1. Regarding the setting of the wind farm, the wind turbine selects the 5MW unit of the National Renewable Energy Laboratory (NREL) of the United States as the research object, the hub height is 90m, the wind wheel diameter D is 126m, the generator efficiency is 1.0, C p - wind speed, C T - wind speed curve as shown in the appendix Figure 2 , 3As shown, the current C of the wind turbine can be determined according to the real-time wind speed p and C T value. The layout of the wind farm adopts a 3*3 layout, with a row spacing of 8D and a column spacing of 5D; the default settings of the model are used as the initial state of the wind farm simulation model;

[0042] The establishment of the wind farm simulation model includes a wind turbine model, a wake model, and an optimization model, which jointly simulate the main process of the wind farm absorbing wind energy; the wind turbine model simulates the wind turbine absorbing wind energy, using the C p - wind speed and C T - wind speed curve as the model; during operation, the wind energy capture coefficient C of the wind turbine is determined according to the wind speed p and the thrust coefficient C of the wind turbine T ;

[0043] According to the thrust coefficient C of the wind turbine T calculate the obstruction effect of the wind turbine on the incoming flow and transfer it to the wake model to calculate the maximum wind speed loss in the wake area; the wind energy capture coefficient C p represents the ability of the wind turbine to convert wind energy into mechanical energy and can be used to calculate the power of the wind turbine. Its calculation expression is as follows

[0044]

[0045] In the formula, A is the swept area of the wind turbine rotor, ρ is the air density, u is the wind speed of the wind turbine rotor, γ is the yaw angle of the wind turbine, and C p is the wind energy capture coefficient;

[0046] The wake model includes a non-yawed wake model, a wake deflection model, and a wake superposition model; there are 9 wind turbines in the wind farm, and the downstream wind turbines are simultaneously affected by the wakes of multiple upstream wind turbines. The superposition effect of the wake areas of multiple wind turbines needs to be considered to calculate the incoming flow wind speed of the downstream wind turbines; in the yaw control of the wind turbines in the wind farm, the wind turbines may be non-yawed or yawed. Therefore, a wake superposition model is needed to superpose the non-yawed wakes or yawed wakes of each upstream wind turbine to obtain the incoming flow wind speed of the downstream wind turbines; the non-yawed wake model uses the BPA Gaussian wake model, and the wake deflection model uses the Bastankhah Gaussian wake deflection model; the wake superposition model uses the sum of squares superposition to calculate the influence of the wake superposition of multiple upstream units; the formula is as follows:

[0047]

[0048] where u i is the incoming flow wind speed of the target unit i, N is the number of units in the wake superposition area of N upstream units of this unit, u ∞ is the incoming flow wind speed, and u jis the incoming wind speed in front of the upstream unit j, u ji is the wind speed at the location of unit i in the wake area of unit j.

[0049] The expression of the optimization model of the whole-field power P under the yaw control of the wind farm in step S1 is as follows:

[0050]

[0051] where P i is the wind turbine power of the i-th unit, C pi is the wind energy capture coefficient of the i-th unit, u i is the wind turbine speed of the i-th unit, γ i is the yaw angle of the i-th unit.

[0052] In step S3, the wind speed fluctuation sequence and the wind direction fluctuation sequence of the wind farm are divided into the current wind condition sequence and the wind condition sequence in the future time period; the current wind condition sequence can use data such as lidar data, nacelle anemometer data, and nacelle lidar data in the wind farm, and use the measured data after cleaning and filtering; the wind condition sequence in the future time period requires a new anemometer tower to be built around the wind farm, and the wind conditions in a wider range are used as the wind conditions in the future time period of the wind farm; the wind speed fluctuation sequence, the wind direction fluctuation sequence, and the set value of the yaw angle of the wind turbine are used as the state of the wind farm yaw control model; the wind speed fluctuation sequence and the wind direction fluctuation sequence are defined as (ws 0 , …, ws t-1 , wd 0 , …, wd t-1 ), where ws 0 , …, ws t-1 represents the wind speed fluctuation sequence at t minutes, and wd 0 , …, wd t-1 represents the wind direction fluctuation sequence at t minutes; in this embodiment, considering the five-minute wind condition, the wind condition sequence is defined as (ws 0 , …, ws 4 , wd 0 , …, wd 4 ); where ws 0 , …, ws 4 represents the wind speed fluctuation sequence at 5 minutes, and wd 0 , …, wd 4 represents the wind direction fluctuation sequence at 5 minutes.

[0053] Step S4 starts the offline learning and online learning of the wind farm yaw control model based on PPO; the schematic diagram of the wind farm yaw control model based on PPO is as shown in the appendix Figure 4As shown in the figure, during the offline learning stage, with historical wind speed, wind direction sequences, and the initial yaw angle as inputs, the intelligent agent (composed of a policy network, a value network, and an interaction and learning mechanism; the intelligent agent perceives environmental changes, analyzes data, makes decisions, and executes actions to improve the operating efficiency and safety of the wind farm) interacts with the wind farm simulation model, aiming to learn to adjust the yaw angle of the wind turbines to maximize the total power generation under different wind speed and wind direction conditions; initially, experience is accumulated by randomly selecting yaw angle actions. The intelligent agent observes the environmental state and outputs actions, introducing random disturbances to increase policy diversity; after sufficient experience, the policy network (Actor) and value network (Critic) of the PPO algorithm are trained based on sampled data to optimize the policy. At the same time, the experience is stored in the data buffer for subsequent training and policy improvement.

[0054] During the online learning stage, the intelligent agent is placed in the actual wind farm environment, monitors the environmental conditions in real time, and commands the adjustment of the yaw actions of the wind turbines according to the policy learned by PPO; during this period, the intelligent agent continuously interacts with the environment, collects new data and stores it in the experience set, and slightly updates the policy network and value network in combination with real-time data to adapt to environmental changes such as wind speed fluctuations and wind direction changes; by continuously updating the policy and value function, online learning can dynamically optimize the yaw angle setting and improve the power generation efficiency of the wind farm.

[0055] For the state setting of the wind farm yaw control model, in order to better represent the state of the wind turbines, the yaw angle setting value of the wind turbines is used as one of the states; the state is s(ws 0 ,…,ws 4 ,wd 0 ,…,wd 4 ,y 0 ), where ws 0 ,…,ws 4 represents the 5-minute wind speed fluctuation sequence, wd 0 ,…,wd 4 represents the 5-minute wind direction fluctuation sequence, and the time resolution of the wind condition fluctuation data is in minutes; y 0 is the yaw angle setting value of the wind turbine, and y PPO is the yaw angle setting value of the wind turbine after the action. The action setting is to use the action a as the action value of the yaw angle setting value of the wind turbine, with the clockwise direction being positive; regarding the reward setting of the model, the power generation is normalized, and the improvement rate of the power generation after the wind farm yaw control relative to the power generation under the maximum power point tracking (MPPT) control strategy is used as the reward; the total power generation of the wind farm within 5 minutes is calculated as W PPO ; the expression of the reward r optimized by PPO is as follows:

[0056]

[0057] An example table of the training process parameters is shown as follows:

[0058]

[0059] During the training process, 200 random wind conditions were used for training. After experiencing the training of the first wind process, the agent has certain decision-making ability. In subsequent training, the decision-making ability of the agent can be further improved; after the training is completed, the power improvement example of the optimized performance under 10 types of wind conditions compared with the MPPT strategy is shown in the appendix Figure 5 As shown, it can be seen that compared with the maximum power point tracking (MPPT) control, the yaw control model of the wind farm has a relatively obvious increase in the total power generation of the wind farm under most wind conditions; the actual environment required for the online learning of the yaw control model of the wind farm can be set as a wind farm simulation model with different wake parameters as the actual wind farm environment. Wake expansion coefficient k w The empirical formula is as follows

[0060] k w = 0.3837I u + 0.003678

[0061] where I u represents the turbulence intensity; 0.3837 is set to 0.5 as the actual wind farm environment; the training interaction is carried out 400 times with 200 random wind conditions; after the online learning of the yaw control model of the wind farm based on the reinforcement learning algorithm (PPO) interacts with the actual environment, the optimization effect of power improvement is further enhanced.

[0062] In summary, through the offline training and online learning of the yaw control model of the wind farm and the wind farm simulation model, the optimal yaw angles of each wind turbine in the wind farm under different wind condition sequences are obtained; the yaw angles are set according to the measured wind condition sequence, effectively improving the total power generation of the wind farm.

[0063] Compared with the traditional control method for average wind speed and average wind direction, the method of the present invention greatly improves the total power generation of the wind farm; at the same time, the online learning of the yaw control of the wind farm can utilize the data feedback in actual operation to reduce the influence of the uncertainty of wind conditions, wake calculation, integrated model, etc. on the improvement effect and improve the power generation efficiency. The present invention meets the timeliness requirements for solving the yaw control model of the wind farm in actual engineering, avoiding the problems that the online optimization takes a long time and is difficult to meet the timeliness requirements, and the optimization accuracy for wind conditions not pre-calculated is low when using offline optimization. Through the method of the present invention, a better wind farm power generation than the traditional maximum power point tracking (MPPT) control can be obtained, effectively providing the optimal yaw angles of each wind turbine during the operation of the wind farm, and being able to perform real-time optimization according to the change of the wind sequence, improving the power generation efficiency of the wind farm and making the overall power generation of the wind farm reach the maximum value.

[0064] Obviously, the above embodiments are merely certain embodiments of the present invention and are not used to limit the present invention; at the same time, for those skilled in the art, the present invention can have various changes and modifications; any modifications, equivalent substitutions, and improvements made within the spirit and scope of the present invention are all within the protection scope of the present invention.

Claims

1. A wind farm yaw control method based on wind process, characterized in that: Use the following steps: S1. Collect information on wind farm environment, terrain, wind turbine coordinates and wind turbine type, and establish a wind farm simulation model; S2, obtaining wind farm operation status data as the initial state of the wind farm simulation model; S3, obtaining a wind speed fluctuation sequence and a wind direction fluctuation sequence of the wind farm as the state of a yaw control model of the wind farm; S4. The wind farm yaw control model based on the reinforcement learning algorithm (PPO) interacts with the wind farm simulation model to conduct offline learning and obtain the optimal yaw angle setting value under different wind condition sequences; further, the wind farm yaw control model observes the environment and guides the action of the wind turbine to complete online learning.

2. A wind farm yaw control method based on wind process according to claim 1, characterized in that: The establishment of the wind farm simulation model in S1 includes the wind turbine model, wake model and optimization model, which together simulate the main process of wind farm absorbing wind energy; The wind turbine model simulates the wind turbine absorbing wind energy, using C p - Wind speed, C T - Wind speed curve as a model; The wind energy capture coefficient C of the wind turbine is determined according to the wind speed during operation. p and wind turbine thrust coefficient C T ; According to the thrust coefficient C of the wind turbine T Calculate the wind turbine's obstruction to the incoming flow and pass it to the wake model to calculate the maximum wind speed loss in the wake area; wind energy capture coefficient C p It indicates the ability of wind turbines to convert wind energy into mechanical energy. It can be used to calculate the power of wind turbines. The calculation expression is as follows: Among them, A is the swept area of ​​the wind rotor, ρ is the air density, u is the wind speed of the wind rotor, γ is the yaw angle of the wind turbine, and C p is the wind energy capture coefficient.

3. A wind farm yaw control method based on wind process according to claim 2, characterized in that: The wake model includes the non-yaw wake model, the wake deflection model and the wake superposition model. There are multiple wind turbines in the wind farm, and the downstream wind turbines are affected by the wakes of multiple wind turbines at the same time. It is necessary to consider the superposition effect of the wake areas of multiple wind turbines to calculate the incoming wind speed of the downstream wind turbines. The wake superposition model is used to superimpose the non-yaw wakes or yaw wakes of each upstream wind turbine to obtain the incoming wind speed of the wind turbine. The non-yaw wake model adopts the BPA Gaussian wake model; The wake deflection model adopts the Bastankhah Gaussian wake deflection model; The wake superposition model uses square sum superposition to calculate the influence of wake superposition of multiple upstream units. The formula is as follows: Among them, u i is the inflow wind speed of target unit i, N is the wake superposition area of ​​the upstream N units, u ∞ is the incoming wind speed, u j is the inflow wind speed before a certain upstream unit j, u ji is the wind speed in the wake area of ​​unit j at the location of unit i.

4. The method for yaw control of a wind farm based on wind process according to claim 2, characterized in that: The expression of the optimization model of the entire wind farm power P of the wind farm simulation model in S1 is as follows: Among them, P i is the rotor power of the i-th unit, C pi is the wind energy capture coefficient of the i-th unit, u i is the wind speed of the rotor of the i-th unit, γ i is the yaw angle of the i-th unit.

5. The method for yaw control of a wind farm based on wind process according to claim 1, characterized in that: The initial state of the wind farm simulation model in S2 needs to obtain the parameters of each wind turbine, including power, yaw angle, speed and pitch angle for setting.

6. The method for yaw control of a wind farm based on wind process according to claim 1, characterized in that: The wind speed fluctuation sequence and wind direction fluctuation sequence of the wind farm in S3 are divided into the current wind condition sequence and the wind condition sequence in the future time period; The current wind condition sequence uses the lidar data, cabin wind speed and direction instrument data, and cabin lidar data in the wind farm, and uses the measured data after cleaning and filtering. The wind condition sequence for the future time period requires the construction of new wind towers around the wind farm, and uses a wider range of wind conditions as the wind conditions for the wind farm in the future time period. The wind speed fluctuation sequence and wind direction fluctuation sequence are defined as (ws0,…,ws t-1 ,wd0,…,wd t-1 ), where ws0,…,ws t-1 represents the wind speed fluctuation sequence of t minutes, wd0,…,wd t-1 Represents the wind direction fluctuation sequence of t minutes.

7. The method for yaw control of a wind farm based on wind process according to claim 1, characterized in that: In the offline learning stage of S4, the historical wind speed, wind direction sequence and initial yaw angle are used as inputs, and the intelligent agent interacts with the wind farm simulation model, aiming to learn to control the yaw angle of the wind turbine to maximize the total power generation under different wind speed and wind direction conditions; initially, experience is accumulated by randomly selecting yaw angle actions, and the intelligent agent observes the environmental state and outputs actions; after sufficient experience, the policy network (Actor) and value network (Critic) based on the reinforcement learning algorithm (PPO) are used to train based on sampled data to optimize the strategy, and the experience is stored in the data temporary storage area for subsequent training and strategy improvement; In the online learning phase, the agent is placed in an actual wind farm environment, monitors the environmental conditions in real time, and commands the wind turbine yaw action adjustment based on the strategy learned by the reinforcement learning algorithm (PPO). During this period, the agent continuously interacts with the environment, collects new data and stores it in the experience set, and combines real-time data to slightly update the strategy network and value network to adapt to environmental changes. By continuously updating the policy and value function, online learning can dynamically optimize the yaw angle setting.

8. The method for yaw control of a wind farm based on wind process according to claim 7, characterized in that: The construction of the wind farm yaw control model based on the reinforcement learning algorithm (PPO) in step S4 adopts the following steps; S401, determine the structure of the policy network (Actor) and the value network (Critic). The policy network (Actor) uses a multi-layer perceptron, determines the input layer according to the state space, sets multiple hidden layers and activation functions, and the output layer outputs the action probability according to the action space through softmax. The value network is similar, outputting the state value of a single neuron; S402, collect the data of interaction between the agent and the environment, set the experience playback buffer in offline learning, sample the data to calculate the advantage function, use the reinforcement learning algorithm (PPO) objective function and the mean square error loss function to update the policy network and the value network respectively, and use the optimizer gradient descent; In online learning, the wind farm yaw control model is deployed to the actual environment. The intelligent agent selects actions according to the strategy, collects new data into the experience pool, and regularly extracts and combines real-time data to update the two networks slightly to adapt to environmental changes.

9. The method for controlling the yaw of a wind farm based on wind process according to claim 5, characterized in that: The initial state of the wind turbine is shutdown, with a power of 0 kW, a yaw angle of 0°, a speed of 0 rpm, and a pitch angle of 90°.

Images