A wind farm coordinated yaw control method based on a multi-layer artificial intelligence system

Abstract

The present invention belongs to the field of wind power generation, and discloses a collaborative yaw control method for a wind farm based on a multi-layer artificial intelligence system. A numerical simulation wake data set is used to train a single wind turbine yaw wake machine learning model, and combined with a superposition model, an accurate and rapid prediction of the power of the entire field under the yaw control state is achieved. Intelligent control optimization consists of two stages. The first stage intelligently partitions the wind farm based on the wake interference pattern between wind turbines to reduce the dimension of the problem to speed up the optimization efficiency. The second stage uses a Bayesian machine learning network to call intelligent power prediction limited data based on the partition constraint to establish an approximate probabilistic relationship between total power generation and yaw combination, and simultaneously learns and optimizes the power objective function to further accelerate the positioning of the optimal yaw state. The collaborative yaw control method based on the above system can accurately and efficiently determine the optimal yaw control strategy for the wind farm based on real-time incoming flow information, and significantly weaken the wake influence between wind turbines.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind power generation, and in particular relates to a wind farm collaborative yaw control method based on a multi-layer artificial intelligence system. Background Art

[0002] For wind turbines arranged according to a certain layout, the operation of upstream turbines will produce a wake effect, resulting in a loss of wind speed and an increase in turbulence in the downstream flow field. This will cause power loss to the downstream turbines and increase fatigue in the turbine structure, ultimately reducing their service life. Therefore, adopting a reasonable coordinated yaw control strategy to weaken the wake effect between wind turbines can not only significantly increase the total power generation of offshore wind farms, but also effectively reduce the fatigue load on the turbine structure, thereby extending the overall life of the wind turbines.

[0003] Current research typically combines wind farm power prediction models with related intelligent algorithms to determine the optimal yaw layout of wind turbines. Existing wind farm power predictions primarily rely on analytical wake models, such as the Jensen wake model and the Frandsen wake model. These two analytical wake models are widely used in the industry and are incorporated into several international industry-standard software packages, such as WindPro, WAsP, and Floris. However, these models have significant flaws. They not only fail to effectively capture the turbulent characteristics of the near- and far-wake regions, but also fail to accurately depict the non-centrosymmetric morphology of the wind turbine's skewed wake under yaw conditions. Furthermore, they rely heavily on empirical parameters, requiring individual experiments or numerical simulations to correct them for each situation. While power prediction based on high-precision CFD simulations can overcome these shortcomings, it consumes significant computing resources, especially for large wind farms, and is therefore unsuitable for real-time power control in actual wind farms.

[0004] At present, there are many studies on the whole-field wake and power prediction technology of wind farms. For example, patent application CN-115859812A discloses a wind turbine wake model construction and wind farm layout optimization method based on machine learning. It uses a numerical simulation data set to train a machine learning wake model and applies it to wind farm layout optimization. However, this wake model fails to consider the impact of the yaw state on the wake and cannot be applied to real-time coordinated yaw control of wind farms.

[0005] Existing offshore wind farms typically employ a greedy yaw control strategy, where each turbine faces the wind head-on to maximize its own power, without considering the interference effects between turbines. This can lead to a certain loss in the wind farm's total power generation. In comparison, an optimized collaborative control strategy modulates the upstream turbine's wake trajectory by adjusting its yaw angle, reducing the impact of its wake on downstream turbines at the expense of its own power generation, thereby significantly improving power generation efficiency. However, traditional optimization algorithms, such as genetic algorithms (GAs), particle swarm optimization (PSOs), and covariance matrix adaptive evolutionary strategies (CMAESs), require a large amount of sampled data to reach the optimal state for complex wind farm systems with multiple input parameters. This can lead to excessive iterations, increasing optimization computation time and making them unsuitable for real-time optimization control of wind farms. Furthermore, sensor noise in the inflow's average wind speed and turbulence intensity, as well as control errors in the yaw eccentricity angle, can cause uncertainty in the monitoring data, leading to errors in wind farm power generation estimation, which in turn hinders the implementation of accurate real-time yaw control.

[0006] Currently, there are many studies on collaborative yaw control technology for wind farms. For example, patent application CN 111615589A discloses a method and device for collaboratively controlling wind turbines in a wind farm. The method uses measured data to train a machine learning model to establish a correlation between a pair of wind turbine time series data (including environmental conditions, wind turbine internal status, wind farm information) and the power ratio of upstream and downstream wind turbines for application in wind farm collaborative control. However, the method is only suitable for short-term adjustments and when wind speed fluctuations are small. For any pair of wind turbines in an irregularly arranged wind farm, a machine learning model needs to be established and trained independently. At the same time, changes in the wind farm layout require retraining of the machine learning model.

[0007] Existing wind farm collaborative yaw control technology lacks a universal and efficient control method based on accurate power prediction. Especially as the scale of wind farms continues to expand, the optimization dimension increases significantly, which brings great challenges to fast and accurate real-time yaw control. Summary of the Invention

[0008] To solve the above technical problems, the present invention provides a wind farm collaborative yaw control method based on a multi-layer artificial intelligence system. The multi-layer artificial intelligence system, which is composed of intelligent power prediction and intelligent control optimization, accurately and efficiently determines the optimal yaw control strategy of the wind farm based on real-time incoming flow information, thereby realizing real-time collaborative yaw control of the wind farm.

[0009] The present invention provides a wind farm coordinated yaw control method based on a multi-layer artificial intelligence system, the method comprising an intelligent power prediction stage and an intelligent control optimization stage;

[0010] In the intelligent power prediction stage, a numerical simulation wake dataset is used to train a single wind turbine yaw wake machine learning model. This model is then combined with a wake superposition model to predict the entire wind farm's wake and output power under yaw control.

[0011] In the intelligent control optimization stage, the wind farm is intelligently partitioned based on the wake interference pattern between wind turbines, and a Bayesian machine learning network is constructed for yaw optimization, achieving real-time collaborative yaw control of the wind farm.

[0012] Furthermore, the intelligent power prediction steps are:

[0013] Step 1: Determine the inflow and control conditions for generating a wake database based on the wind turbine operating parameters, and perform a series of fluid dynamics numerical simulations on the wake field of a single wind turbine based on the inflow and control conditions.

[0014] Step 2: Divide the turbine hub height plane into N subdomains evenly along the direction perpendicular to the wind speed and number them according to their spatial positions (1, 2, …, N). Output the velocity and turbulence intensity of the numerically simulated wake field corresponding to each subdomain to form a wake field subdataset, including N wake velocity subdatasets and N wake turbulence intensity subdatasets.

[0015] Step 3: Build an artificial neural network model consisting of a three-variable input layer, a hidden layer, and an output layer with the same number of output nodes as the subdomain wake field; the activation function of the hidden layer is sigmoid, and the activation function of the output layer is ReLU; the optimization algorithm is Adam;

[0016] Step 4: Use the artificial neural network built in step 3 to independently train the wake field sub-datasets to obtain the ANN (artificial neural network) yaw wake sub-model. The training of different sub-models is performed through parallel computing.

[0017] Step 5: Aggregate the N ANN yaw wake sub-models in the order of the corresponding sub-domain numbers to form a single wind turbine yaw wake model (including velocity model and turbulence model);

[0018] Step 6: Based on the wind farm inflow conditions, determine the upstream and downstream relationship of the wind turbines in the wind farm, and determine the inflow conditions for each wind turbine in order from upstream to downstream. The inflow conditions for each wind turbine are obtained by superimposing the wake fields of all upstream wind turbines. The wake field of the upstream wind turbine is solved using the single-turbine yaw wake model based on its inflow and yaw control conditions.

[0019] Step 7: Based on the inflow and yaw angle of each wind turbine in the wind farm obtained in step 6, combined with the power-wind speed curve, determine its power generation. Add them together to obtain the predicted total power generation of the wind farm, completing the intelligent power prediction.

[0020] Furthermore, in step 1, the Reynolds average method or the actuation line coupling numerical simulation method is used for the numerical simulation of the fluid dynamics of the wake of a single wind turbine, and the inflow and control conditions of the wake database generated are uniformly selected within the operating range of the wind turbine, including the wind speed at the hub height, the turbulence intensity and the yaw angle of the wind turbine.

[0021] Furthermore, for different wind turbine types, the calculation domain size of the numerical simulation of the fluid dynamics of the single wind turbine wake is determined according to the maximum impact area of ​​the wind turbine wake within the operating range, and is judged by the set wake velocity loss threshold.

[0022] Furthermore, in step 2, the output nodes are arranged at equal distances in the numerically simulated wake field, and each wake velocity sub-dataset (wake turbulence sub-dataset) contains the velocity (turbulence) at all output nodes in the wake field corresponding to its subdomain.

[0023] Furthermore, in step 4, the input layer of the ANN yaw wake sub-model contains three variables related to inflow and control, namely, the hub height wind speed u hub , turbulence intensity I and yaw angle γ; the output layer is the velocity field or turbulence intensity field of the wake, which is represented by velocity loss Δu and additional turbulence intensity ΔI respectively.

[0024] Furthermore, in step 6, the inflow condition of the most upstream wind turbine is determined by the undisturbed inflow condition of the wind farm, and the inflow condition of the downstream wind turbine is obtained by superimposing the single wind turbine wake fields of all upstream wind turbines using an empirical wake superposition model.

[0025] Furthermore, different superposition models are selected according to the differences in the superposition of velocity and turbulence.

[0026] Furthermore, the intelligent control optimization steps are:

[0027] Step 8: Intelligently partition the target wind farm based on the wake interference pattern between wind turbines. Divide the wind farm into m parallel partitions along the incoming flow direction. The partition rows are numbered in order (1, 2, ..., m). The yaw angles of wind turbines in the same partition are kept consistent, and the last wind turbine is in a zero-yaw state.

[0028] Step 9: Based on the partition constraints in step 8, randomly generate yaw combinations. Given the wind farm inflow conditions, repeat steps 6 and 7 to generate a series of initial training data sets consisting of yaw combinations and corresponding total powers.

[0029] Step 10: Build a Bayesian machine learning network with the total wind farm power as the objective function. In the first iteration, a Gaussian process is used to initially establish an approximate probability distribution between the yaw combination and the total power using the initial training dataset to complete the learning task.

[0030] Step 11: Subsequent iterations simultaneously learn and optimize the objective function. After completing the same learning task as in the initial iteration, the optimization phase uses the learned approximate probability distribution to establish a corresponding acquisition function. The acquisition function is based on the expectation and variance of the power prediction probability distribution, and its extreme value indicates the area where the optimal power is likely to be high. By maximizing the acquisition function, the optimal yaw combination is searched for.

[0031] Step 12: Given the wind farm inflow conditions, repeat steps 6 to 7 to determine the total power corresponding to the possible optimal yaw combination. The two form a new data set and are incorporated into the training dataset.

[0032] Step 13: Repeat steps 11 and 12 until the power generation obtained from multiple adjacent iterations remains stable and meets the optimization convergence conditions, the optimal coordinated yaw control strategy is determined, and the intelligent control optimization is completed.

[0033] Furthermore, in step 8, in the intelligent partitioning of the wind farm, the specific steps for determining the partition row number i to which any wind turbine in the wind farm belongs are as follows:

[0034] (1) Draw a series of parallel lines passing through the fan position perpendicular to the incoming flow direction. Take the first row of baselines and calculate the distance between the first and last rows, denoted as l. The row spacing s is:

[0035]

[0036] (2) Calculate the distance between the fan and the baseline, denoted as d k ;

[0037] (3) The row number i of the fan is:

[0038] i=ceil(d k / s)(i=1,2,…,m).

[0039] The beneficial effects described in the present invention are as follows: the wind farm power prediction system based on the ANN yaw wake model adopted by the present invention can not only ensure the accuracy comparable to CFD numerical simulation, but also achieve efficiency similar to analytical simulation, which well meets the accuracy and efficiency requirements of the subsequent control optimization process for power data acquisition. The control optimization system combines the Bayesian machine learning network with the intelligent partitioning method. The Bayesian machine learning network establishes an approximate probabilistic relationship between total power generation and yaw combination through Gaussian process (GP), and uses the limited data set provided by the intelligent power prediction to simultaneously learn and optimize the objective function, efficiently determine the optimal yaw control strategy, and on this basis, cooperate with the intelligent partitioning method to further reduce the optimization dimension and significantly improve the control efficiency. The method described in the present invention is suitable for real-time collaborative yaw control considering different incoming flow conditions under any wind farm layout, and can accurately and efficiently provide the optimal yaw control strategy based on the incoming flow information measured in real time, effectively solving the complex optimization problem of highly nonlinear and high variable dimension in the collaborative yaw control of large offshore wind farms. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a flow chart of the wind farm coordinated yaw control method;

[0041] Figure 2 It is a schematic diagram of the structure of the ANN yaw wake submodel;

[0042] Figure 3 It is a schematic diagram of intelligent zoning of wind farms;

[0043] Figure 4 It is a demonstration of intelligent zoning of an example wind farm;

[0044] Figure 5 This is a schematic diagram of the multi-layer artificial intelligence architecture of the wind farm collaborative yaw control method;

[0045] Figure 6 This is an example of the yaw layout and wind speed cloud map before and after wind farm control optimization;

[0046] Figure 7 This is a comparison of the optimization efficiency of the example wind farm using intelligent zoning and traditional independent control;

[0047] Figure 8 This is an example of a wind farm that compares the power improvement effect of using intelligent zoning and traditional independent control considering different wind direction distributions. DETAILED DESCRIPTION

[0048] In order to make the contents of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments in conjunction with the accompanying drawings.

[0049] like Figure 1As shown in FIG, a wind farm collaborative yaw control method based on a multi-layer artificial intelligence system according to the present invention takes a 4×4 regular wind farm with 16 wind turbines as an example. Figure 4 As shown, the fan interval is 7D. The specific steps are:

[0050] Step 1: The Vestas V80 2MW wind turbine model is selected, with a rotor diameter of 80m. Based on the operating parameters of this wind turbine, the inflow and control conditions for generating the wake database are determined, including the wind speed and turbulence at the hub height and the turbine yaw angle. A series of fluid dynamics numerical simulations of the single wind turbine wake field are performed using the Reynolds Average Method (RANS) / Actuation Line Method (ALM) coupled numerical simulation method. The computational domain is 28D × 12D × 5D:

[0051] The wind speed distribution range is 5m / s–15m / s, with an interval of 1m / s; the turbulence distribution range is 2%–26%, with an interval of 2%; the yaw angle distribution range is 0–30°, with an interval of 3°, for a total of 1443 operating conditions;

[0052] Step 2: Divide the horizontal plane at 70m at the turbine hub height into 120 subdomains at intervals of 8m (0.1D) along the vertical wind speed direction. Output the velocity and turbulence in the wake field corresponding to the numerical simulation in each subdomain according to the spatial position number (1, 2, ..., 120) to form a wake subdataset, including 120 wake velocity subdatasets and 120 wake turbulence subdatasets. Each subdataset consists of the wake field data (velocity or turbulence) at 280 output nodes uniformly distributed at intervals of 8m (0.1D) along the flow direction in the corresponding subdomain.

[0053] Step 3: Build an artificial neural network model consisting of a three-variable input layer, a hidden layer, and an output layer with the same number of output nodes as the subdomain wake field; the activation function of the hidden layer is sigmoid, and the activation function of the output layer is ReLU; the optimization algorithm is Adam;

[0054] Step 4: Use the artificial neural network built in step 3 to train the wake field sub-datasets in step 2 independently to obtain the ANN yaw wake sub-model. The sub-model structure is as follows Figure 2 As shown, an input layer of size 3 contains three variables related to inflow and control, namely the hub height wind speed u hub, turbulence intensity I and yaw angle γ, one hidden layer with a size of 10 and an activation function of "sigmoid", one output layer with a size of 280, containing the velocity or turbulence data of all output nodes on the corresponding subdomain, represented by the velocity loss Δu or the additional turbulence intensity ΔI; the training of different sub-models is carried out in parallel computing, with a total of 240 sub-models for the velocity field and turbulence field (120 ANN yaw wake velocity sub-models and 120 ANN yaw wake turbulence sub-models);

[0055] Step 5: Summarize the 120 ANN yaw wake sub-models in the order of the corresponding sub-domain numbers to form a single-turbine yaw wake model of the VestasV80 wind turbine (including the velocity model and turbulence model);

[0056] Step 6: Figure 4 In the regular wind farm shown, the upstream and downstream relationships between wind turbines are determined according to the incoming flow direction; the inflow of the upstream turbine is determined by the undisturbed incoming flow of the wind farm. For downstream turbines, based on the inflow and control conditions of the upstream turbines, the tail flow fields of individual turbines are superimposed in order from upstream to downstream to obtain the tail flow data for the entire wind farm, thereby obtaining the corresponding inflow for each turbine. The velocity superposition model uses the square sum loss ratio model: Turbulence intensity superposition adopts the turbulent kinetic energy superposition model: Subscript i represents the target downstream wind turbine, subscript j represents the upstream wind turbine of wind turbine i, and subscript inflow represents the undisturbed inflow of the wind farm;

[0057] Step 7: Based on the inflow and yaw angle of each wind turbine in the wind farm obtained in step 6, combined with the power-wind speed curve, determine its power generation, and then obtain the total power generation prediction result of the wind farm;

[0058] Step 8: Intelligently partition the target wind farm according to the wake interference pattern between wind turbines, such as Figure 3 As shown in the figure, the wind farm is divided into 4 rows with equal spacing along the incoming flow direction. The yaw angles of wind turbines in the same row are consistent, and the last wind turbine is in a zero yaw state. The specific steps to determine the row number i of any wind turbine in the wind farm are as follows:

[0059] (1) Draw a series of parallel lines perpendicular to the incoming flow direction passing through the fan position. Take the first line as the reference line. Calculate the distance between the first and last lines as 26D. Then the line spacing s = 26 / 4D:

[0060] (2) Calculate the distance between the fan and the baseline, denoted as d k ;

[0061] (3) The row number of the fan is determined as follows:

[0062] i=ceil(d k / s)(i=1,2,…,m)

[0063] The results of the example wind farm intelligent zoning are as follows: Figure 4 As shown;

[0064] Step 9: Based on the partition constraints in step 8, randomly generate yaw combinations. Select the wind farm inflow conditions as u = 12 m / s and I = 2%. Repeat steps 6 and 7 to generate a series of initial training data sets consisting of yaw combinations and corresponding total power.

[0065] Step 10: Build Figure 5 The Bayesian machine learning network shown in the figure takes the total power of the wind farm as the objective function. The Gaussian process (GP) is used in the first iteration to initially establish the approximate probability distribution relationship between the yaw combination and the total power generation using the initial training data set: P total (γ|D 1:n ,θ)~N(μ,σ 2 ); where γ represents the yaw combination, P is the total power, and D is the training data set of the Bayesian learning network;

[0066] Step 11: The objective function is then learned and optimized simultaneously in subsequent iterations. After completing the same learning task as in the initial iteration, the optimization phase uses the learned approximate probability distribution to establish a corresponding acquisition function. The optimal yaw combination is searched for by maximizing the acquisition function. The acquisition function is established based on the expected μ and variance σ of the power prediction probability distribution. The expected improvement (EI) acquisition function is usually used:

[0067]

[0068] Step 12: Given the wind farm inflow conditions u = 12 m / s, I = 2%, repeat steps 6 to 7 to determine the total power corresponding to the possible optimal yaw combination. The two form a new data set and are incorporated into the training dataset D.

[0069] Step 13: Repeat steps 11 and 12 until the generated power obtained from multiple adjacent iterations remains stable and meets the optimization convergence conditions, and the optimal coordinated yaw control strategy is determined.

[0070] like Figure 6 As shown, its small Figure 6-1 To show the example wind farm control optimization before and after yaw layout diagram, Figure 6-2To illustrate wind speed contours before and after optimization of an example wind farm control, traditional independent control refers to omitting the intelligent zoning in step 8 of the present invention. Each wind turbine's yaw angle is optimized independently, while all other steps remain unchanged. As can be seen from the figure, control optimization minimizes the impact of strong wake turbulence on downstream wind turbines. Furthermore, in terms of wake attenuation, intelligent zoning effectively reduces the dimensionality of the optimization problem while still achieving similar results to traditional independent control. Figure 7 The difference between the power improvement effect and optimization efficiency between the two control strategies is further quantitatively compared. Figure 7 The intelligent partitioning strategy achieves an optimized power output of 22.26 MW, a 30% increase, slightly less than the 22.32 MW achieved by the traditional independent control strategy. However, its optimization efficiency is significantly improved, reaching optimality after approximately 50 iterations, nearly four times the 200 stable iterations required by traditional independent control. Figure 8 A comparison of the power improvement effects between the two control strategies considering different wind direction distributions is given. Although the decentralized wind direction distribution will increase the difference in optimized power between intelligent partitioning and traditional independent control to a certain extent, overall, compared with the significant improvement in optimization efficiency, the small difference in the optimal solution is negligible, thus verifying that the wind farm collaborative yaw control method integrating intelligent partitioning still has great application value under different wind direction distributions.

[0071] The above description is only a preferred embodiment of the present invention and is not intended to further limit the present invention. All equivalent changes made using the contents of the present invention description and drawings are within the scope of protection of the present invention.

Claims

1. A wind farm coordinated yaw control method based on a multi-layer artificial intelligence system, characterized in that: The method includes an intelligent power prediction stage and an intelligent control optimization stage; In the intelligent power prediction stage, a numerical simulation wake dataset is used to train a single wind turbine yaw wake machine learning model. This model is then combined with a wake superposition model to predict the entire wind farm's wake and output power under yaw control. In the intelligent control optimization phase, the wind farm is intelligently partitioned based on the wake interference pattern between wind turbines, and a Bayesian machine learning network is constructed for yaw optimization. Realize real-time coordinated yaw control of wind farms; The intelligent power prediction includes the following steps: Step 1: Determine the inflow and control conditions for generating a wake database based on the wind turbine operating parameters, and perform a series of fluid dynamics numerical simulations on the wake field of a single wind turbine based on the inflow and control conditions. Step 2: Divide the turbine hub height plane into N subdomains evenly along the direction perpendicular to the wind speed. Number them according to their spatial positions (1, 2, …, N). Output the velocity and turbulence intensity of the numerically simulated wake field corresponding to each subdomain to form a wake field subdataset, including N wake velocity subdatasets and N wake turbulence intensity subdatasets. Step 3: Build an artificial neural network model consisting of a three-variable input layer, a hidden layer, and an output layer with the same number of output nodes as the subdomain wake field; the activation function of the hidden layer is sigmoid, and the activation function of the output layer is ReLU; the optimization algorithm is Adam; Step 4: Use the artificial neural network built in step 3 to independently train the wake field sub-datasets to obtain the ANN yaw wake sub-model. The training of different sub-models is performed through parallel computing. Step 5: Aggregate the N ANN yaw wake sub-models in the order of the corresponding sub-domain numbers to form a single wind turbine yaw wake model; Step 6: Based on the wind farm inflow conditions, determine the upstream and downstream relationship of the wind turbines in the wind farm, and determine the inflow conditions for each wind turbine in order from upstream to downstream. The inflow conditions for each wind turbine are obtained by superimposing the wake fields of all upstream wind turbines. The wake field of the upstream wind turbine is solved using the single-turbine yaw wake model based on its inflow and yaw control conditions. Step 7: Based on the inflow and yaw angle of each wind turbine in the wind farm obtained in step 6, combined with the power-wind speed curve, determine its power generation. Add them together to obtain the predicted total power generation of the wind farm, completing the intelligent power prediction.

2. The wind farm coordinated yaw control method based on a multi-layer artificial intelligence system according to claim 1, characterized in that: In step 1, the Reynolds average method or the actuation line coupling numerical simulation method is used for the numerical simulation of the fluid dynamics of the single wind turbine wake. The inflow and control conditions of the wake database generated are uniformly selected within the wind turbine operating range, including the wind speed at the hub height, turbulence intensity and wind turbine yaw angle.

3. The wind farm coordinated yaw control method based on a multi-layer artificial intelligence system according to claim 2, characterized in that: For different wind turbine types, the calculation domain size of the numerical simulation of the fluid dynamics of the single wind turbine wake is determined according to the maximum impact area of ​​the wind turbine wake within the operating range, and is judged by the set wake velocity loss threshold.

4. The wind farm coordinated yaw control method based on a multi-layer artificial intelligence system according to claim 1, characterized in that: In step 2, the output nodes are arranged at equal distances in the numerical simulation wake field, and each wake turbulence sub-dataset contains the velocities at all output nodes in the wake field corresponding to its subdomain.

5. The wind farm coordinated yaw control method based on a multi-layer artificial intelligence system according to claim 1, characterized in that: In step 4, the input layer of the ANN yaw wake sub-model contains three variables related to inflow and control, namely, the hub height wind speed u hub , turbulence intensity I and yaw angle The output layer is the velocity field or turbulence intensity field of the wake, which is represented by the velocity loss ∆u and the additional turbulence intensity ∆I respectively.

6. The wind farm coordinated yaw control method based on a multi-layer artificial intelligence system according to claim 1, characterized in that: In step 6, the inflow condition of the most upstream wind turbine is determined by the undisturbed inflow condition of the wind farm, and the inflow condition of the downstream wind turbine is obtained by superimposing the single wind turbine wake fields of all upstream wind turbines using the empirical wake superposition model.

7. The wind farm coordinated yaw control method based on a multi-layer artificial intelligence system according to claim 6, characterized in that: According to the difference of velocity and turbulence superposition, different superposition models are selected.

8. The wind farm coordinated yaw control method based on a multi-layer artificial intelligence system according to claim 1, characterized in that: Intelligent control optimization includes the following steps: Step 8: Intelligently partition the target wind farm based on the wake interference pattern between wind turbines. Divide the wind farm into m parallel partitions along the incoming flow direction. The partition rows are numbered sequentially (1, 2, …, m). The yaw angles of wind turbines within the same partition are kept consistent, and the last wind turbine is in a zero-yaw state. Step 9: Based on the partition constraints in step 8, randomly generate yaw combinations. Given the wind farm inflow conditions, repeat steps 6 and 7 to generate a series of initial training data sets consisting of yaw combinations and corresponding total powers. Step 10: Build a Bayesian machine learning network with the total wind farm power as the objective function. In the first iteration, a Gaussian process is used to initially establish an approximate probability distribution between the yaw combination and the total power using the initial training dataset to complete the learning task. Step 11: The objective function is then learned and optimized simultaneously in subsequent iterations. After completing the same learning task as in the initial iteration, the learned approximate probability distribution is used to establish a corresponding acquisition function. The acquisition function is established based on the expectation and variance of the power prediction probability distribution. Its extreme value represents the area where the optimal power occurs. The optimal yaw combination is searched by maximizing the acquisition function. Step 12: Given the wind farm inflow conditions, repeat steps 6 to 7 to determine the total power corresponding to the optimal yaw combination. The two form a new data set and are incorporated into the training dataset. Step 13: Repeat steps 11 and 12 until the power generation obtained from multiple adjacent iterations remains stable and meets the optimization convergence conditions, the optimal coordinated yaw control strategy is determined, and the intelligent control optimization is completed.

9. The wind farm coordinated yaw control method based on a multi-layer artificial intelligence system according to claim 8, characterized in that: In step 8, in the intelligent zoning of the wind farm, the specific steps for determining the zone row number i to which any wind turbine in the wind farm belongs are as follows: (1) Draw a series of parallel lines passing through the fan position perpendicular to the incoming flow direction. Take the first row of baselines and calculate the distance between the first and last rows, denoted as l. The row spacing s is: ， (2) Calculate the distance between the fan and the baseline, denoted as d k ; (3) The row number i of the fan is: i = ceil (d k / s) (i = 1, 2, …, m)。

Images