An offshore wind farm wake cooperative control method based on a DD-MAPPO-MPC architecture

Abstract

The application discloses a kind of based on DD-MAPPO-MPC architecture's offshore wind farm wake synergic control method, belongs to wind farm wind energy capture efficiency optimization and unit synergic control technology;For the problem that upstream wind turbine wake of offshore wind farm influences downstream wind turbine power output, yaw control is carried out to wind turbine to reduce wind farm power loss.The specific method includes: using Date-Driven to create high-precision digital wind farm simulator, through actual environment data, training is carried out using graph neural network, and future wind field environment data is predicted;Using RL algorithm, the optimal control strategy of wind farm is obtained by using MAPPO model learning;Using MPC, according to the future wind field data calculated by DD and the control strategy provided by MAPPO, the wind turbine is controlled.The application is based on actual environment at sea, modifies the running state of wind turbine, reduces the influence of airflow between wind turbines, and improves the efficiency of wind farm.

Description

Technical Field

[0001] This invention relates to wind farm wind energy capture efficiency optimization and unit coordinated control technology, and particularly to a method for wake coordinated control of offshore wind farms. Background Technology

[0002] Offshore wind power has become a key area for large-scale expansion due to its unique advantages such as stable wind energy resources, high wind speeds, low turbulence, and no occupation of land resources. Installed capacity and grid-connected scale have continued to climb, and its share of energy supply in the power system has been steadily increasing. To maximize the utilization of offshore wind energy resources, wind farms generally adopt high-density array-style turbine layouts, with increasingly smaller turbine spacing and a continuously increasing number of turbines per farm. However, this layout also brings significant engineering challenges, namely the turbine wake effect. When upstream wind turbines capture wind energy, they create a low-speed, highly turbulent wake region downstream. This causes a significant decrease in the inflow wind speed to downstream turbines and a sharp increase in turbulence intensity, directly leading to two core problems: First, power generation efficiency loss. Wake interference can reduce the overall output power of a wind farm by 10% to 30%, and the power loss can be even higher in complex wind conditions and large-scale wind farms, severely restricting wind energy capture efficiency and economic benefits. Second, equipment safety risks. The unsteady aerodynamic loads brought by the wake can exacerbate fatigue damage to key components such as turbine blades, transmission chains, and towers, shortening equipment service life and increasing the difficulty and cost of offshore operation and maintenance.

[0003] Existing technologies for suppressing wake effects and coordinating wind farm control have several limitations. Traditional analytical wake models lack accuracy. Mainstream classic wake models such as Jensen and FLORIS are based on idealized physical assumptions and cannot accurately adapt to complex real-world environments such as offshore gusts, sudden wind direction changes, and ocean-atmospheric boundary layer turbulence. They suffer from large errors in wake distribution and power prediction, making it difficult to support high-precision control. Single control methods have performance shortcomings. Single-agent reinforcement learning optimizes only a single wind turbine, easily getting trapped in local optima and failing to achieve coordinated decision-making across the entire wind farm, thus failing to maximize total power generation. While multi-agent reinforcement learning can achieve global strategy optimization, it lacks real-time wind farm prediction and equipment safety constraint verification, resulting in overly aggressive control actions that easily exceed wind turbine operating boundaries and have poor engineering feasibility. Model predictive control has rolling optimization and constraint handling capabilities, but it relies on accurate prediction models, limiting the accuracy of traditional physical models. Furthermore, without a globally optimal strategy to guide it, optimization efficiency is low and control lag is significant. Furthermore, the integration of data-driven and control algorithms is low. Pure data-driven modeling can fit the characteristics of the actual wind field, but it lacks collaborative decision-making and real-time optimization closed loop, and cannot transform accurate predictions into executable and efficient control commands. Various control algorithms also lack the support of high-precision and fast-response virtual wind fields, resulting in insufficient matching between control strategies and actual wind fields.

[0004] In summary, existing wake control technologies for offshore wind farms cannot simultaneously meet the core requirements of accurate wake prediction, global collaborative optimization, engineering operation safety, and real-time control responsiveness, making it difficult to fundamentally solve the wake interference problem in high-density offshore wind farms. Therefore, there is an urgent need to develop a novel wake collaborative control method that integrates data-driven modeling, multi-agent collaborative decision-making, and model predictive control to break through traditional technical bottlenecks and improve the overall power generation efficiency and operational stability of offshore wind farms. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the prior art by providing an optimized method for wind farm control to solve the wake interference problem between wind farms and thus help improve the power generation efficiency of wind farms.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] A method for coordinated wake control of offshore wind farms based on the DD-MAPPO-MPC architecture includes the following steps:

[0008] Step S1: Data-driven virtual wind farm modeling: Collect actual environmental data and wind turbine operation data of offshore wind farms. Using individual wind turbines in the wind farm as graph nodes and the wake influence relationship between wind turbines as graph edges, construct a digital twin virtual wind farm model based on graph neural network (GNN). Output wind farm wake distribution data and wind turbine power prediction data in the prediction time domain through the virtual wind farm model, providing an interactive environment for reinforcement learning training and providing wind farm state input in the prediction time domain for model predictive control.

[0009] Step S2, MAPPO multi-agent cooperative control strategy optimization: Each wind turbine in the wind farm is defined as an independent agent. The virtual wind farm model constructed in step S1 is used as the training interaction environment. The optimization objective is to maximize the total power generation of the wind farm. The MAPPO algorithm is trained through multi-agent near-end strategy optimization to obtain the global cooperative optimal control strategy of the wind farm.

[0010] Step S3, MPC rolling optimization control solution: Based on the predicted time-domain wind field state output by the virtual wind field model described in step S1, and the global collaborative optimal control strategy obtained in step S2 as the optimization reference target, the optimal control sequence of the wind turbine in the control time domain is obtained through rolling time-domain optimization solution under the preset wind turbine operation safety constraints.

[0011] Step S4, Control Command Execution and Power Acquisition: The optimal control sequence obtained in step S3 is converted into real-time control commands for the yaw angle and pitch angle of the corresponding wind turbines, and sent to each wind turbine in the wind farm for execution, thereby achieving wake suppression and acquiring the actual output power of the wind farm.

[0012] Furthermore, in step S1, constructing the virtual wind field model specifically includes the following sub-steps:

[0013] Step S11: Construct the wind turbine state and motion dynamics equations: Define the wind turbine's state variables as a multi-dimensional vector containing incoming wind speed, incoming wind direction, turbulence intensity, wind turbine yaw angle, and wind turbine output power; define the wind turbine's motion variables as a control vector containing yaw angle adjustment and pitch angle adjustment; and construct the wind turbine state transition equations:

[0014]

[0015] in, This represents the wind field state at time k. This indicates the fan control action at time k. This is the state transition function; Indicates the incoming wind speed. Indicates the angle of the incoming wind direction. Indicates turbulence intensity. This represents the yaw angle of the i-th wind turbine. This represents the output power of the i-th fan. This indicates the adjustment angle of the yaw angle for the i-th wind turbine. This represents the change in the pitch angle of the i-th wind turbine;

[0016]

[0017]

[0018]

[0019] Step S12: Construct a wake loss prediction model: Based on the spatial relationship between the fans, the incoming airflow conditions, and the fan operating parameters, construct an effective inflow velocity model for a single fan.

[0020]

[0021] in, Let represent the effective inflow wind speed of the i-th wind turbine at time k. This represents the free wind speed at time k that is unaffected by the upstream wake. This represents the wake loss coefficient of the i-th wind turbine at time k;

[0022] Step S13: Construct a wind farm power output model: Based on the effective inflow wind speed, yaw angle and pitch angle of the wind turbine, construct a single wind turbine output power model, and superimpose them to obtain a total power prediction model for the entire wind farm.

[0023] Step S14: Construct the loss function for the data-driven model:

[0024]

[0025] in, This indicates the actual output power of the fan. This represents the predicted output power of the wind turbine. L represents the error between the predicted and actual data.

[0026] Step S14: Constructing the data-driven model loss function and GNN message passing mechanism: Using the mean square error between the actual output power of the wind turbine and the predicted power of the model as the loss function, train and optimize the parameters of the GNN model; the graph convolution message passing function of the GNN model is:

[0027]

[0028] in For the first In the layer network Feature vectors of each wind turbine node , For trainable weight matrix, For activation function, For the first The upstream of the typhoon affects the wind turbine assembly.

[0029] Furthermore, the wake loss coefficient in step S12 The calculations are based on the wind turbine rotor radius, wake radius, thrust coefficient, and axial induction coefficient, specifically:

[0030]

[0031] in The radius of the wake; Where is the radius of the wind turbine. For the first The thrust coefficient of the typhoon generator unit For the first Axial induction coefficient of typhoon generator.

[0032] Furthermore, in step S13, the output power model of a single wind turbine is as follows:

[0033]

[0034]

[0035] in, Indicates the first The output power of the typhoon generator This represents the total power of the wind farm. It is the first The pitch angle of a typhoon turbine. Indicates leaf tip speed; Represents the viscous stress tensor; This indicates air density, typically 1.225 kg / m³. 3 , For the first The power coefficient of typhoon generator units; The swept area of ​​the wind turbine rotor. for Time of the first The yaw angle of a typhoon turbine; the total power of a wind farm is the sum of the output power of all individual wind turbines.

[0036] Furthermore, in step S2, the MAPPO algorithm training uses a globally shared reward function, the expression of which is:

[0037]

[0038]

[0039] Where N is the total number of wind turbines in the wind farm. Indicates the first Changes in the yaw angle of the typhoon turbine. Indicates the first Changes in the pitch angle of the typhoon turbine These are the power weighting coefficient, yaw penalty coefficient, and pitch penalty coefficient, respectively. R is the globally shared reward function, and A is the dominance function.

[0040] Furthermore, in step S2, the MAPPO algorithm updates its strategy using a pruning objective function, which is:

[0041]

[0042]

[0043] in, These are the trainable parameters of the policy network. for A time-limiting function restricts the update magnitude of the PPO strategy to prevent training breakdown. for The advantage function at time, The cropping threshold, The pruning function is used; the advantage function is calculated based on the global shared reward function and the state value of the value network.

[0044] Furthermore, in step S2, the final total loss function of the MAPPO algorithm is:

[0045]

[0046]

[0047] in, Let be the mean squared error loss function of the value network. This is the value loss weighting coefficient. The strategy entropy weight coefficient, The policy entropy of the policy network is given; by minimizing the total loss function, the policy network and value network are iteratively updated to obtain the globally cooperative optimal control policy.

[0048] Furthermore, in step S3, the objective function for MPC rolling optimization is:

[0049]

[0050] in, It is the target total power of the wind farm output by the MAPPO algorithm. The future provided by DD Time based on control quantity The predicted total power of the wind farm, It is a control variable composed of yaw angle and pitch angle. It is the change in the amount of movement. This is the penalty coefficient for changes in the control quantity.

[0051] Furthermore, in step S3, the preset wind turbine operation safety constraints include:

[0052]

[0053]

[0054]

[0055]

[0056]

[0057]

[0058] in, for The yaw angle of the wind turbine at all times, This is the minimum allowable yaw angle for the wind turbine. This represents the maximum permissible yaw of the wind turbine. for At any given moment, the wind turbine's blade pitch angle This is the minimum permissible blade pitch angle for the wind turbine. This represents the maximum permissible pitch angle of the wind turbine. for Fan power at all times This represents the minimum output power of the fan. This represents the maximum output power of the fan. for The mechanical load of the fan at all times. This is the maximum mechanical load that the wind turbine can withstand.

[0059] Furthermore, in step S4, the actual output power of the wind farm and the operating data of the wind turbines collected will be fed back to the virtual wind farm model in step S1, which will be used to update and optimize the parameters of the GNN model online, thereby realizing the closed-loop iterative optimization of the control strategy.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] (1) Using DD to construct a virtual wind field model, the wake loss and power mapping relationship can be directly learned from wind conditions, turbine location and control actions, replacing the traditional analytical model, and providing high-precision and ultra-fast wind field prediction for reinforcement learning MAPPO and model predictive control MPC.

[0062] (2) MAPPO uses multi-agent collaboration to obtain the global optimal strategy, which is closer to the actual application of wind farms and maximizes the power generation of the entire field.

[0063] (3) MAC uses the MAPPO global optimal strategy to predict future wind field environment information using a data-driven model, performs online rolling optimization under constraints, and outputs safe, smooth, and executable yaw and pitch angle commands. Attached Figure Description

[0064] Figure 1 This is a simplified flowchart of a method for coordinated wake control of offshore wind farms based on the DD-MAPPO-MPC architecture proposed in this invention.

[0065] Figure 2 This is a detailed flowchart of a method for coordinated wake control of offshore wind farms based on the DD-MAPPO-MPC architecture proposed in this invention;

[0066] Figure 3 , Figure 4 These are the reward value change graph and the initial environment graph of the wind farm during the training process of this invention. Detailed Implementation

[0067] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0068] like Figure 1 As shown, a method for coordinated wake control of offshore wind farms based on the DD-MAPPO-MPC architecture includes the following steps:

[0069] Step S1: Data-driven virtual wind farm modeling: Collect actual environmental data and wind turbine operation data of offshore wind farms. Using individual wind turbines in the wind farm as graph nodes and the wake influence relationship between wind turbines as graph edges, construct a digital twin virtual wind farm model based on graph neural network (GNN). Output wind farm wake distribution data and wind turbine power prediction data in the prediction time domain through the virtual wind farm model, providing an interactive environment for reinforcement learning training and providing wind farm state input in the prediction time domain for model predictive control.

[0070] This method constructs a GNN virtual wind farm model through data-driven methods, which can accurately predict wake distribution and wind turbine power. It provides a safe and efficient interactive training environment for MAPPO reinforcement learning and provides MPC with the input for predicting the wind farm state in the time domain. It overcomes the problems of insufficient accuracy of traditional physical models, high cost of trial and error on real machines, and strong control lag, and effectively improves the overall power generation efficiency and operational safety of wind farms.

[0071] Step S2, MAPPO Multi-Agent Cooperative Control Strategy Optimization: Each wind turbine in the wind farm is defined as an independent agent. Using the virtual wind farm model constructed in Step S1 as the training environment, and maximizing the total power generation of the entire wind farm as the optimization objective, the MAPPO algorithm is trained through multi-agent near-end strategy optimization to obtain the globally optimal cooperative control strategy for the wind farm. The reward value change graph and the initial environment graph of the wind farm during the training process are shown below. Figure 3 , Figure 4 As shown;

[0072] By using a single wind turbine as an intelligent agent and combining it with a global shared reward mechanism, the system breaks through the dilemma of local optima, guides wind turbines to coordinate and avoid wakes, and balances power generation revenue, equipment losses and load safety to achieve a dynamic balance among the three, significantly improving the overall power generation efficiency and meeting the long-term operation needs of offshore wind farms.

[0073] Step S3, MPC rolling optimization control solution: Based on the predicted time-domain wind field state output by the virtual wind field model described in step S1, and the global collaborative optimal control strategy obtained in step S2 as the optimization reference target, the optimal control sequence of the wind turbine in the control time domain is obtained through rolling time-domain optimization solution under the preset wind turbine operation safety constraints.

[0074] Based on virtual wind farm prediction data and MAPPO optimal strategy, the optimal control sequence is solved through MPC rolling optimization. This ensures the maximization of wind farm power generation efficiency, strictly adheres to equipment safety constraints, avoids trial and error losses on real machines, and can respond to wind condition changes in real time, improving control accuracy and stability. It is suitable for complex offshore operating scenarios and provides a scientific and feasible optimal control solution for subsequent control execution.

[0075] Step S4, Control Command Execution and Power Acquisition: The optimal control sequence obtained in step S3 is converted into real-time control commands for the yaw angle and pitch angle of the corresponding wind turbines, and sent to each wind turbine in the wind farm for execution, thereby achieving wake suppression and acquiring the actual output power of the wind farm.

[0076] Furthermore, in step S1, constructing the virtual wind field model specifically includes the following sub-steps:

[0077] Step S11: Construct the wind turbine state and motion dynamics equations: Define the wind turbine's state variables as a multi-dimensional vector containing incoming wind speed, incoming wind direction, turbulence intensity, wind turbine yaw angle, and wind turbine output power; define the wind turbine's motion variables as a control vector containing yaw angle adjustment and pitch angle adjustment; and construct the wind turbine state transition equations:

[0078] in, This represents the wind field state at time k. This indicates the fan control action at time k. This is the state transition function; Indicates the incoming wind speed. Indicates the angle of the incoming wind direction. Indicates turbulence intensity. This represents the yaw angle of the i-th wind turbine. This represents the output power of the i-th fan. This indicates the adjustment angle of the yaw angle for the i-th wind turbine. This represents the change in the pitch angle of the i-th wind turbine;

[0079]

[0080]

[0081]

[0082] Step S12: Construct a wake loss prediction model: Based on the spatial relationship between the fans, the incoming airflow conditions, and the fan operating parameters, construct an effective inflow velocity model for a single fan.

[0083]

[0084] in, Let represent the effective inflow wind speed of the i-th wind turbine at time k. This represents the free wind speed at time k that is unaffected by the upstream wake. This represents the wake loss coefficient of the i-th wind turbine at time k;

[0085] Step S13: Construct a wind farm power output model: Based on the effective inflow wind speed, yaw angle and pitch angle of the wind turbine, construct a single wind turbine output power model, and superimpose them to obtain a total power prediction model for the entire wind farm.

[0086] Step S14: Construct the loss function for the data-driven model:

[0087]

[0088] in, This indicates the actual output power of the fan. This represents the predicted output power of the wind turbine. L represents the error between the predicted and actual data.

[0089] Step S14: Constructing the data-driven model loss function and GNN message passing mechanism: Using the mean square error between the actual output power of the wind turbine and the predicted power of the model as the loss function, train and optimize the parameters of the GNN model; the graph convolution message passing function of the GNN model is:

[0090]

[0091] in For the first In the layer network Feature vectors of each wind turbine node , For trainable weight matrix, For activation function, For the first The upstream of the typhoon affects the wind turbine assembly.

[0092] Furthermore, the wake loss coefficient in step S12 The calculations are based on the wind turbine rotor radius, wake radius, thrust coefficient, and axial induction coefficient, specifically:

[0093]

[0094] in The radius of the wake; Where is the radius of the wind turbine. For the first The thrust coefficient of the typhoon generator unit For the first Axial induction coefficient of typhoon generator.

[0095] Furthermore, in step S13, the output power model of a single wind turbine is as follows:

[0096]

[0097]

[0098] in, Indicates the first The output power of the typhoon generator This represents the total power of the wind farm. It is the first The pitch angle of a typhoon turbine. Indicates leaf tip speed; Represents the viscous stress tensor; This indicates air density, typically 1.225 kg / m³. 3 , For the first The power coefficient of typhoon generator units; The swept area of ​​the wind turbine rotor. for Time of the first The yaw angle of a typhoon turbine; the total power of a wind farm is the sum of the output power of all individual wind turbines.

[0099] Furthermore, in step S2, the MAPPO algorithm training uses a globally shared reward function, the expression of which is:

[0100]

[0101]

[0102] Where N is the total number of wind turbines in the wind farm. Indicates the first Changes in the yaw angle of the typhoon turbine. Indicates the first Changes in the pitch angle of the typhoon turbine These are the power weighting coefficient, yaw penalty coefficient, and pitch penalty coefficient, respectively. R is the globally shared reward function, and A is the dominance function.

[0103] Furthermore, in step S2, the MAPPO algorithm updates its strategy using a pruning objective function, which is:

[0104]

[0105]

[0106] in, These are the trainable parameters of the policy network. for A time-limiting function restricts the update magnitude of the PPO strategy to prevent training breakdown. for The advantage function at time, The cropping threshold, The pruning function is used; the advantage function is calculated based on the global shared reward function and the state value of the value network.

[0107] Furthermore, in step S2, the final total loss function of the MAPPO algorithm is:

[0108]

[0109]

[0110] in, Let be the mean squared error loss function of the value network. This is the value loss weighting coefficient. The strategy entropy weight coefficient, The policy entropy of the policy network is given; by minimizing the total loss function, the policy network and value network are iteratively updated to obtain the globally cooperative optimal control policy.

[0111] Furthermore, in step S3, the objective function for MPC rolling optimization is:

[0112]

[0113] in, It is the target total power of the wind farm output by the MAPPO algorithm. The future provided by DD Time based on control quantity The predicted total power of the wind farm, It is a control variable composed of yaw angle and pitch angle. It is the change in the amount of movement. This is the penalty coefficient for changes in the control quantity.

[0114] Furthermore, in step S3, the preset wind turbine operation safety constraints include:

[0115]

[0116]

[0117]

[0118]

[0119]

[0120]

[0121] in, for The yaw angle of the wind turbine at all times, This is the minimum allowable yaw angle for the wind turbine. This represents the maximum permissible yaw of the wind turbine. for At any given moment, the wind turbine's blade pitch angle This is the minimum permissible blade pitch angle for the wind turbine. This represents the maximum permissible pitch angle of the wind turbine. for Fan power at all times This represents the minimum output power of the fan. This represents the maximum output power of the fan. for The mechanical load of the fan at all times. This is the maximum mechanical load that the wind turbine can withstand.

[0122] Furthermore, in step S4, the actual output power of the wind farm and the operating data of the wind turbines collected will be fed back to the virtual wind farm model in step S1, which will be used to update and optimize the parameters of the GNN model online, thereby realizing the closed-loop iterative optimization of the control strategy.

Claims

1. A method for coordinated wake control of offshore wind farms based on the DD-MAPPO-MPC architecture, characterized in that, Includes the following steps: Step S1: Data-driven virtual wind farm modeling: Collect actual environmental data and wind turbine operation data of offshore wind farms. Using individual wind turbines in the wind farm as graph nodes and the wake influence relationship between wind turbines as graph edges, construct a digital twin virtual wind farm model based on graph neural network (GNN). Output wind farm wake distribution data and wind turbine power prediction data in the prediction time domain through the virtual wind farm model, providing an interactive environment for reinforcement learning training and providing wind farm state input in the prediction time domain for model predictive control. Step S2, MAPPO multi-agent cooperative control strategy optimization: Each wind turbine in the wind farm is defined as an independent agent. The virtual wind farm model constructed in step S1 is used as the training interaction environment. The optimization objective is to maximize the total power generation of the wind farm. The MAPPO algorithm is trained through multi-agent near-end strategy optimization to obtain the global cooperative optimal control strategy of the wind farm. Step S3, MPC rolling optimization control solution: Based on the predicted time-domain wind field state output by the virtual wind field model described in step S1, and the global collaborative optimal control strategy obtained in step S2 as the optimization reference target, the optimal control sequence of the wind turbine in the control time domain is obtained through rolling time-domain optimization solution under the preset wind turbine operation safety constraints. Step S4, Control Command Execution and Power Acquisition: The optimal control sequence obtained in step S3 is converted into real-time control commands for the yaw angle and pitch angle of the corresponding wind turbines, and sent to each wind turbine in the wind farm for execution, thereby achieving wake suppression and acquiring the actual output power of the wind farm.

2. The offshore wind farm wake collaborative control method based on the DD-MAPPO-MPC architecture according to claim 1, characterized in that, In step S1, constructing the virtual wind field model specifically includes the following sub-steps: Step S11: Construct the wind turbine state and motion dynamics equations: Define the wind turbine's state variables as a multi-dimensional vector containing incoming wind speed, incoming wind direction, turbulence intensity, wind turbine yaw angle, and wind turbine output power; define the wind turbine's motion variables as a control vector containing yaw angle adjustment and pitch angle adjustment; and construct the wind turbine state transition equations: in, This represents the wind field state at time k. This indicates the fan control action at time k. This is the state transition function; Indicates the incoming wind speed. Indicates the angle of the incoming wind direction. Indicates turbulence intensity. This represents the yaw angle of the i-th wind turbine. This represents the output power of the i-th fan. This indicates the adjustment angle of the yaw angle for the i-th wind turbine. This represents the change in the pitch angle of the i-th wind turbine; Step S12: Construct a wake loss prediction model: Based on the spatial relationship between the fans, the incoming airflow conditions, and the fan operating parameters, construct an effective inflow velocity model for a single fan. in, Let represent the effective inflow wind speed of the i-th wind turbine at time k. This represents the free wind speed at time k that is unaffected by the upstream wake. This represents the wake loss coefficient of the i-th wind turbine at time k; Step S13: Construct a wind farm power output model: Based on the effective inflow wind speed, yaw angle and pitch angle of the wind turbine, construct a single wind turbine output power model, and superimpose them to obtain a total power prediction model for the entire wind farm. Step S14: Construct the loss function for the data-driven model: in, This indicates the actual output power of the fan. This represents the predicted output power of the wind turbine. L represents the error between the predicted and actual data. Step S14: Constructing the data-driven model loss function and GNN message passing mechanism: Using the mean square error between the actual output power of the wind turbine and the predicted power of the model as the loss function, train and optimize the parameters of the GNN model; the graph convolution message passing function of the GNN model is: in, For the first In the layer network Feature vectors of each wind turbine node , For trainable weight matrix, For activation function, For the first The upstream of the typhoon affects the wind turbine assembly.

3. The offshore wind farm wake collaborative control method based on the DD-MAPPO-MPC architecture according to claim 2, characterized in that, The wake loss coefficient in step S12 The calculations are based on the wind turbine rotor radius, wake radius, thrust coefficient, and axial induction coefficient, specifically: in, The radius of the wake; Where is the radius of the wind turbine. For the first The thrust coefficient of the typhoon generator unit For the first Axial induction coefficient of typhoon generator.

4. The method for coordinated wake control of offshore wind farms based on the DD-MAPPO-MPC architecture according to claim 2, characterized in that, In step S13, the output power model of a single wind turbine is as follows: in, Indicates the first The output power of the typhoon generator This represents the total power of the wind farm. It is the first The pitch angle of a typhoon turbine. Indicates leaf tip speed; Represents the viscous stress tensor; This indicates air density, typically 1.225 kg / m³. 3 , For the first The power coefficient of typhoon generator units; The swept area of ​​the wind turbine rotor. for Time of the first The yaw angle of a typhoon turbine; the total power of a wind farm is the sum of the output power of all individual wind turbines.

5. The offshore wind farm wake collaborative control method based on the DD-MAPPO-MPC architecture according to claim 1, characterized in that, In step S2, the MAPPO algorithm training uses a globally shared reward function, the expression of which is: Where N is the total number of wind turbines in the wind farm. Indicates the first Changes in the yaw angle of the typhoon turbine. Indicates the first Changes in the pitch angle of the typhoon turbine These are the power weighting coefficient, yaw penalty coefficient, and pitch penalty coefficient, respectively. R is the globally shared reward function, and A is the dominance function.

6. The offshore wind farm wake collaborative control method based on the DD-MAPPO-MPC architecture according to claim 5, characterized in that, In step S2, the MAPPO algorithm updates its strategy using a pruning objective function, which is: in, These are the trainable parameters of the policy network. for A time-limiting function restricts the update magnitude of the PPO strategy to prevent training breakdown. for The advantage function at time, The cropping threshold, The pruning function is used; the advantage function is calculated based on the global shared reward function and the state value of the value network.

7. The offshore wind farm wake collaborative control method based on the DD-MAPPO-MPC architecture according to claim 6, characterized in that, In step S2, the final total loss function of the MAPPO algorithm is: in, Let be the mean squared error loss function of the value network. This is the value loss weighting coefficient. The strategy entropy weight coefficient, The policy entropy of the policy network is given; by minimizing the total loss function, the policy network and value network are iteratively updated to obtain the globally cooperative optimal control policy.

8. The method for coordinated wake control of offshore wind farms based on the DD-MAPPO-MPC architecture according to claim 1, characterized in that, In step S3, the objective function for MPC rolling optimization is: in, It is the target total power of the wind farm output by the MAPPO algorithm. The future provided by DD Time based on control quantity The predicted total power of the wind farm, It is a control variable composed of yaw angle and pitch angle. It is the change in the amount of movement. This is the penalty coefficient for changes in the control quantity.

9. The offshore wind farm wake collaborative control method based on the DD-MAPPO-MPC architecture according to claim 8, characterized in that, In step S3, the preset safety constraints for wind turbine operation include: in, for The yaw angle of the wind turbine at all times, This is the minimum allowable yaw angle for the wind turbine. This represents the maximum permissible yaw of the wind turbine. for At any given moment, the wind turbine's blade pitch angle This is the minimum permissible blade pitch angle for the wind turbine. This represents the maximum permissible pitch angle of the wind turbine. for Fan power at all times This represents the minimum output power of the fan. This represents the maximum output power of the fan. for The mechanical load of the fan at all times. This is the maximum mechanical load that the wind turbine can withstand.

10. The offshore wind farm wake collaborative control method based on the DD-MAPPO-MPC architecture according to claim 1, characterized in that, In step S4, the actual output power of the wind farm and the operating data of the wind turbines are collected and fed back to the virtual wind farm model in step S1. This data is used to update and optimize the parameters of the GNN model online, thereby achieving closed-loop iterative optimization of the control strategy.

Images