Wind power plant yaw angle optimization control model construction method and device and control strategy and device
Through the Gaussian wake model and sparse directed graph combined with the multi-agent graph attention MASAC algorithm, a wind farm yaw angle optimization control model is constructed, which solves the problems of high calculation costs and long optimization time in the existing technology, and realizes the wind farm Effective improvement and robustness of total power.
Patent Information
- Application Number
- CN202410831990.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-26
- Publication Date
- 2024-11-19
AI Technical Summary
When dealing with complex wake and variable wind direction environments, the existing wind farm yaw optimization strategy has high calculation costs and long optimization time, and fails to fully consider the coordinated optimization of multiple variables, resulting in limited room for total power improvement in the wind farm.
The Gaussian wake model and sparse directed graph are used to construct the wind farm yaw angle optimization control model, and the multi-agent graph attention MASAC algorithm (GAT-SGGMASAC) is used to perform layered optimization control, reducing the computational complexity and improving optimization efficiency, learning Optimal yaw angle strategy to increase the total wind farm power.
The calculation cost and optimization time of the wind farm yaw angle optimization strategy have been reduced, adapted to complex and variable wind direction environments, effectively improved the total output power of the wind farm, and had better optimization effects and robustness.
Smart Images

Figure CN118981935A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a method and device for constructing a yaw angle optimization control model of a wind farm, as well as a control strategy and device, and belongs to the field of power optimization control of wind farm power generation. Background Art
[0002] In the actual application of wind turbines, the distance between wind turbines in a wind farm cannot be too far due to factors such as construction cost and terrain restrictions. The rear-row wind turbines are inevitably affected by the wake effect, resulting in a certain degree of power loss in the wind farm, which can cause the wind farm to lose up to 10% to 40% of the rated power. In order to reduce the total power loss of the wind farm caused by the wake effect, it is necessary to optimize the power control, and the current mainstream method to solve the total power loss is to perform yaw optimization control.
[0003] Among the existing wind farm yaw optimization strategies, two strategies are often considered. One is the wind farm yaw control technology based on the axial induction factor: the document "Active output optimization of offshore wind farms based on wake effect" proposes a wind farm total power optimization strategy based on particle swarm algorithm to maximize the total power of the wind farm, taking the axial induction factor as the optimization variable. The results show that the total power of the whole field is increased by 9.96% after optimization. The document "From the Cover: Wind farm power optimization through wake steering" proposes a yaw optimization control strategy combining game theory and greedy algorithm, taking the axial induction factor as the optimization variable. Experiments have shown that the use of this strategy can effectively increase the power of the wind farm. The other is the control technology based on the wind farm wake yaw: the document "Control-oriented model for secondary effects of wake steering" proposes an optimization strategy based on the wake deflection model, which achieves a power increase of nearly 40% for the whole field under full wake interference conditions, and an average power increase rate of 7.1%. The paper "Optimization of wind turbine yawangles in a wind farm using a three-dimensional yawed wake model" proposed a wind farm yaw strategy based on particle swarm optimization, using the yaw angle as a control variable to maximize the total power of the wind farm. However, the above studies did not fully consider the coordinated optimization of multiple variables, and could not adapt to the various wind farm layouts in practice. The accuracy of modeling the wind farm optimization problem is very high. Therefore, the strategies proposed in the above studies have many restrictions when used, and there is still room for improvement in the current wind farm yaw strategy.
[0004] As a model-free artificial intelligence control strategy, deep reinforcement learning has strong learning ability and is expected to solve the defects of the above strategies. At present, some research has been carried out in the field of wind farm control. The paper "Artificial Neural Network Based Reinforcement Learning for Wind Turbine Yaw Control" proposes a wind farm yaw control strategy based on q learning, which improves the total output power of the wind farm through yaw control. The paper "Multi-Agent Reinforcement Learning Control of a Hydrostatic Wind Turbine-Based Farm" proposes a multi-agent deep reinforcement learning strategy called multi-agent strategy optimization (MAPO), which learns the optimal operation curve of the wind farm through repeated interaction between wind turbines and the environment, thereby gradually improving the control strategy of the wind farm. The above studies show that deep reinforcement learning has certain advantages in wind farm yaw control, but it only optimizes the control from the overall perspective of the wind farm, and has not yet taken into account the implicit correlation between complex wakes. In addition, the strategy calculation cost is high and time-consuming, and it cannot be applied to complex and changeable wind direction environments. Summary of the invention
[0005] In order to solve the above technical problems, the technical problem to be solved by the present invention is to provide a method and device for constructing a wind farm yaw angle optimization control model, as well as a control strategy and device, which have low yaw control strategy operation cost, short optimization time and increased total power generation.
[0006] The present invention adopts the following technical solution:
[0007] To achieve the above objectives, in a first aspect, the present invention provides a method for constructing a yaw angle optimization control model, which comprises the following steps:
[0008] S1. Obtaining the location information, environmental information and wind turbine parameters of all wind turbines in the wind farm;
[0009] S2, transform the coordinates of the wind turbine group, and calculate the wind farm wake environment parameters based on the Gaussian wake model;
[0010] S3. construct an initial wake directed graph based on environmental parameters, perform sparse processing on the initial wake directed graph, disconnect the edges with lower weights in the initial wake directed graph to obtain a sparse wake directed graph, and construct a subset of wind turbines based on the sparse wake directed graph;
[0011] S4. Based on the subset of wind turbines, a wind farm yaw angle optimization control model is constructed, and the yaw angle optimization control model is trained based on the wind farm environment.
[0012] Furthermore, the wind farm wake environment parameters in S2 include wake overlap ratio, deficit wind speed and distance between wind turbines;
[0013] According to the aerodynamic theory of wind turbines, the Gaussian wake model formula is (1):
[0014]
[0015] In the formula, V xmin V represents the wind speed at the center of the wake at position x behind the wind rotor, that is, the minimum wind speed of the wake cross section at that position; x is the wind speed at each location on the cross section at position x behind the wind rotor; C T represents the thrust coefficient of the current wind turbine; σ is the standard deviation of the Gaussian at the position x behind the wind rotor; h is the center height of the wind rotor; R is the radius of the wind rotor; Δr is a constant. When x approaches 0, Δr can make V xmin It is equal to the wind speed value immediately behind the wind rotor; z represents the coordinate along the blade; y represents the coordinate in the vertical direction;
[0016] Since the wake only affects some downstream wind turbines, the wind speed loss of these turbines is less than that of being completely covered by the wake. Therefore, the wind speed loss in the single wake model is multiplied by the ratio of the area where the wake overlaps with the rotating disk of the wind rotor, so as to calculate the actual wind speed at the downstream turbine. The loss wind speed at any position can be calculated according to the Gaussian wake model formula as shown in formula (2). The area of the wake overlap can be calculated according to the cosine theorem. The influence between wind turbines is shown in formula (3), where is the wake overlap ratio;
[0017]
[0018]
[0019] In the formula, V ij represents the effective wind speed at the location of wind turbine j; V0 is the inflow wind speed of upstream wind turbine i; S represents the area of the overlap between the cross section of the wake of the upstream wind turbine i at the downstream wind turbine j and the plane swept by the wind rotor of wind turbine j; j represents the rotor area of the downstream wind turbine j; θ1 and θ2 are the arcs of the overlapping wake and rotor parts, respectively, which are calculated by the cosine theorem; R1 represents the rotor radius of wind turbine j; R2 represents the rotor radius of wind turbine i; C T,i represents the thrust coefficient of wind turbine i;
[0020] The unit coordinates are transformed to obtain the cross-flow coordinates of the downstream wind turbine. Since the wake deflection in the simulation experiment is tested under a single average wind speed, the speed change caused by wind resource turbulence is limited. Therefore, a linear function is used to parameterize the lateral deviation of the wake caused by the wake rotation. The following is a detailed description.
[0021] Detailed calculation steps:
[0022] 1. Linear function representation of wake deviation caused by rotation;
[0023] y w,rotation (x) = a d +b d x (4)
[0024] Among them, a d and b d It is a parameter that needs to be determined by empirical methods; w,rotation (x) is a linear function used to parameterize the lateral deviation of the wake caused by rotation;
[0025] 2. Linear function representation of wake deviation caused by yaw;
[0026] By dividing the tangent of the wake centerline angle by x, the lateral offset of the wake center caused by the yaw of the wind rotor is obtained, as shown in the following formula:
[0027]
[0028] In the formula, y w,yaw,j (x) represents the lateral offset of the wake center relative to the hub of the wind turbine j rotor caused by yaw; x represents the axial distance from the center of the wind rotor; D is the diameter of the wind rotor; k d is the coefficient of wake deflection degree caused by yaw; a j is the rotor axial induction factor; γ j is the yaw angle; Indicates the thrust coefficient of the fan when it is yawed;
[0029] 3. Combining the rotation-induced component and the yaw-induced component to obtain the cross-flow coordinates, the following formula:
[0030] y w,j (x) = y w,rotation (x)+y w,yaw,j (x) (6)
[0031] In the formula, y w,j (x) represents the lateral position coordinate of the wake center relative to the hub of the wind turbine j; y w,rotation (x) is used to parameterize the lateral deviation of the wake caused by rotation; y w,yaw,j (x) represents the lateral displacement of the wake center relative to the hub of the wind turbine j caused by yaw;
[0032] Based on the Gaussian wake model formula, the environmental parameters of wake overlap ratio, deficit wind speed and distance between wind turbines are calculated.
[0033] Furthermore, in step S3, the method for constructing the initial wake directed graph is as follows:
[0034] Define a directed graph G = (V, E, A), where the vertex set V = {v i |i=1,2,K,N} represents each wind turbine in the wind farm; each edge of the graph represents the wake distribution between wind turbines; the non-negative adjacency matrix A = (w ij ) N×N Indicates the wake influence strength between wind turbines; defines a pair of elements e ij =(v i ,v j ), which is the vertex v of the directed graph i To the vertex v j The edge weight of the connection, e ij It represents the wake effect of upstream wind turbine i on downstream wind turbine j, that is, the influence of wake effect is unidirectional; the definition of wake intensity weighted non-negative matrix A is as follows:
[0035]
[0036] Among them, w ij represents the weight value of the wake effect of the upstream wind turbine i on the downstream wind turbine j; for the element w in the wake intensity weighting matrix A ij If the upstream wind turbine i has a wake effect on the downstream wind turbine j, the wake weight coefficient is expressed as w ij ; Otherwise the element remains 0;
[0037] The method to construct a sparse wake directed graph is:
[0038] Determine the pruning threshold, the global pruning threshold is ε k = k*ε, where the coefficient k is a hyperparameter that controls the sparsity of the wake directed graph; the threshold ε is the average value of the entire wake weight coefficient, generating a pruned directed graph G p =(V,E p ,A p ), where V represents the node set of the directed graph, E p represents the edge set after pruning, A p represents the adjacency matrix after pruning; according to the pruning threshold, the edges corresponding to the weight matrix A are removed to obtain a sparse tail flow directed graph;
[0039] The method to generate a subset of wind turbines is:
[0040] The free stream wind speed V ∞ The front exhaust fan is used as the leading node of each subset, and the BFS algorithm is used to traverse the connected vertices to generate different directed graph subsets n. i , determine the directed graph subset n i Is there a shared set S between i ; If the shared wind turbines belong to multiple subsets N i , then calculate the shared wind turbine group T in each subset i The weighted degree centrality C indeg The sum of all C indeg Sort and share wind turbines T i Assign it to the subset with the largest degree centrality weight, and construct a decoupled sparse tail directed graph G s =(V,E s ,A s ), based on the sparse wake directed graph, the on-site units are divided into multiple wind turbine subsets N i ;
[0041] Weighted degree centrality is a comprehensive indicator to measure the importance of shared vertices, which takes into account the number of adjacent vertices and the weight of the connection. indeg The calculation method is as follows:
[0042]
[0043] Where v represents the vertex in the subgraph; w1(v) represents the in-degree weight of the v node; and L represents the in-degree value.
[0044] Furthermore, the following steps are used to construct the yaw angle optimization control model:
[0045] S41, construct a hierarchical optimization control network and establish several fan subsets N i To achieve hierarchical optimization control, the upper layer uses each wind turbine subset processed by the sparse directed graph as an intelligent agent to optimize between subsets, and the lower layer uses each wind turbine subset N as an intelligent agent to optimize between subsets. i Each wind turbine in the wind turbine subset N is an intelligent agent in i The optimization is performed within the wind farm, and the upper optimization target is set to maximize the total power of the wind farm, and the lower ... i The total power of the inner wind farm is the largest; the upper and lower optimization processes are simultaneously optimized under different time granularity conditions. The lower layer will be optimized in each time step, and the upper layer will be optimized once every 10 time steps. At this time, information such as wind turbine power and control variables are exchanged between the upper and lower layers;
[0046] S42. Construct a reinforcement learning optimization control network based on the sparse wake multi-agent graph attention MASAC algorithm (GAT-Sparse Graph Grouping Multi-Agent Soft Actor-Critic, GAT-SGGMASAC); GAT-SGGMASAC is based on the Actor-Critic framework. The Actor module uses a fully connected layer for feature transformation, and then uses two layers of graph attention layers (GAT) for superposition, and then uses three fully connected layers to realize the nonlinear mapping of system state to action strategy. Similarly, the Critic module structure also adopts the form of superposition of two layers of GAT and three layers of full connection; the hierarchical control network is integrated with the reinforcement learning optimization control network to obtain the GAT-SGGMASAC yaw angle optimization control model;
[0047] S43, train the GAT-SGGMASAC network, and the thrust and output power calculation methods of the wind turbine are shown in equations (9) and (10);
[0048]
[0049] Where i = 1, 2, ..., N, N is the total number of wind turbines; A d is the rotor disk area of wind turbine i; U i is the wind speed on the rotor plane at wind turbine i; ρ is the air density; and are the thrust coefficient and power coefficient of the i-th wind turbine, respectively, which are related to the axial induction factor a i and yaw angle φ i The calculation methods are shown in formula (11) and formula (12):
[0050]
[0051]
[0052] In the formula, μ is the calculation hyperparameter;
[0053] The total power calculation formula for a wind farm with N wind turbines is shown in (13);
[0054]
[0055] Control the yaw angle φ of each wind turbine in the wind farm i and thrust coefficient The problem of increasing the total power of a wind farm is considered as an optimal control problem with the reward function (14):
[0056]
[0057] Where t represents the current time step; γ is the power conversion coefficient; x(t) is the control variable, which includes the thrust coefficient and yaw angle variable [C′ T1 (t),...,C′ TN (t),φ1(t),...,φ N (t)] T ;
[0058] The optimization goal is to find a set of optimal control strategies x * (t) maximizes the reward J(t), and the optimization problem is formulated as:
[0059]
[0060] The optimization problem needs to meet the following conditions:
[0061]
[0062] In the formula, C′ T,min , C′ T,max ,φ min ,φ max are the lower and upper bounds of the thrust coefficient and yaw angle, respectively.
[0063] In a second aspect, the present invention provides a wind farm yaw angle optimization control strategy, which adopts the following steps:
[0064] S11, obtaining location information, environmental information and wind turbine parameters of all wind turbines in the wind farm;
[0065] S12, transforming the coordinates of the wind turbine group, and calculating the wind farm wake distribution in the simulation environment based on the Gaussian wake model;
[0066] S13. Inputting the wind farm wake environment information into the yaw angle optimization control model as described in any one of claims 1 to 4 to obtain an optimal yaw angle action strategy.
[0067] Furthermore, the process of calculating the wind farm wake environment information is as follows:
[0068] The coordinates of the unit are transformed to obtain the cross-flow coordinates of the downstream wind turbine. Based on the Gaussian wake model formula, the environmental parameters of the wake overlap ratio, the deficit wind speed, and the distance between wind turbines are calculated.
[0069] Furthermore, the process of obtaining the optimal yaw angle action includes:
[0070] Based on the wind farm wake environment information, the yaw angle optimization control model is used and the maximum reward valuation is output according to the forward calculation to obtain the optimal action strategy for the real-time yaw angle of the wind farm;
[0071] Specifically, by controlling the yaw angle and thrust coefficient of each wind turbine in the wind farm, the optimal control strategy is quickly solved to maximize the reward J(t), so that the wind turbine can obtain the optimal action strategy for the real-time yaw angle and achieve an increase in the total power of the wind farm.
[0072] Furthermore, the verification process of the optimal yaw angle action strategy is also included, and the specific operations are as follows:
[0073] The wind farm wake environment information is respectively input into the hierarchical reinforcement learning optimization control strategy and the greedy strategy to obtain a first total power output value and a second total power output value respectively, and the first total power output value and the second total power output value are compared to obtain an evaluation result of the wind farm yaw angle optimization control strategy.
[0074] In a third aspect, the present invention provides a device for constructing a yaw angle optimization control model for a wind farm, which is used to implement the yaw angle optimization control model construction method, and the device comprises:
[0075] The modeling data acquisition module is used to obtain the coordinates of all wind turbines in the wind farm, wind direction, wind speed, distance between wind turbines, and environmental information of the wind rotor radius;
[0076] The parameter acquisition module is used to transform the unit coordinates to obtain the cross-flow coordinates of the downstream wind turbine, and calculate the environmental parameters of the wake overlap ratio, the loss wind speed, and the distance between wind turbines based on the Gaussian wake model formula;
[0077] A relationship building module is used to calculate the wake weight coefficient under the environmental parameters, and to build a sparse directed graph based on the pruning threshold and the wake weight coefficient; based on the sparse directed graph, multiple wind turbine subsets N are obtained. i ;
[0078] The model building module builds a GAT-SGGMASAC yaw angle optimization control model based on the wake coupling relationship information of the wind turbine subset.
[0079] In a fourth aspect, the present invention provides a wind farm yaw angle optimization control strategy device, which is used to implement the yaw angle optimization control strategy, and the device comprises:
[0080] The control data acquisition module is used to obtain the location information, environmental information and wind turbine parameters of all wind turbines in the wind farm;
[0081] The parameter acquisition module is used to transform the unit coordinates to obtain the cross-flow coordinates of the downstream wind turbine, and calculate the environmental parameters of the wake overlap ratio, the loss wind speed, and the distance between wind turbines based on the Gaussian wake model formula;
[0082] A model output module is used to input the environmental parameters into the yaw angle optimization control model as described in any one of claims 1 to 4, and obtain a real-time wind farm yaw angle optimal action strategy based on the maximum reward valuation output by the model forward calculation.
[0083] The positive effects of the present invention are as follows: the present invention obtains the coordinates of all wind turbines in the wind farm, determines the parameter values of wind direction, wind speed, distance between wind turbines, and radius of wind rotors; obtains key environmental parameters of the wind farm, and ensures that the model can accurately reflect the actual wind farm conditions; transforms the coordinates of the wind turbines, and calculates the wake distribution of the wind farm based on the Gaussian wake model; utilizes coordinate transformation and Gaussian wake model calculation to accurately describe the influence of the wake effect between the front and rear wind turbines mathematically, which is convenient for optimizing the yaw angle; constructs an original wake directed graph according to environmental parameters, performs sparse processing on the directed graph, disconnects the edges with lower weights in the directed graph, obtains a sparse wake directed graph, and constructs a subset of wind turbines based on the sparse wake directed graph; uses The directed graph is used to represent the wake influence between wind turbines in a wind farm. The sparse processing is to simplify the directed graph, remove the wake interference with less influence (lower weight), reduce the computational complexity, make the model more efficient, and retain the main wake influence path to ensure the accuracy of the optimization process. The subset division method is helpful for local optimization and improves the efficiency and effect of the optimization calculation. Based on the subset of wind turbines, a wind farm yaw angle optimization control model is constructed, and the yaw angle optimization control model is trained based on the wind farm environment. Through the training of the model, the model can learn a set of strategies for optimal yaw angle settings under different wind conditions, so as to maximize the total power output of the wind farm. Then the real-time wind farm wake environment information is input into the yaw angle optimization control model to obtain the yaw angle optimal action strategy.
[0084] The present invention reduces the computing resources of the yaw control strategy, shortens the optimization time, can effectively improve the yaw angle control effect of the wind farm under complex and changeable wind direction environment, and further improves the total output power of the wind farm, and has better optimization effect and robustness. BRIEF DESCRIPTION OF THE DRAWINGS
[0085] Figure 1 A flow chart for constructing a yaw angle optimization control model according to an embodiment of the present invention;
[0086] Figure 2 This is a network structure diagram of the GAT-SGGMASAC model according to an embodiment of the present invention;
[0087] Figure 3 It is a reward curve diagram of different reinforcement learning yaw control model training processes in an embodiment of the present invention;
[0088] Figure 4A diagram showing total power output of a wind farm with different yaw control strategies according to an embodiment of the present invention;
[0089] Figure 5 This is a flow chart of a wind farm yaw angle optimization control strategy according to an embodiment of the present invention. DETAILED DESCRIPTION
[0090] The following will be described in detail with reference to the accompanying drawings and in combination with embodiments.
[0091] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.
[0092] Embodiment 1:
[0093] like Figure 1 As shown, this embodiment provides a method for constructing a yaw angle optimization control model, which adopts the following steps:
[0094] S1. Obtaining the location information, environmental information and wind turbine parameters of all wind turbines in the wind farm;
[0095] The location information refers to the coordinates of the wind turbine group, the environmental information refers to the corresponding wind direction and wind speed, and the wind turbine parameters refer to the radius of the wind rotor;
[0096] S2, transform the coordinates of the wind turbine group, and calculate the wind farm wake environment parameters based on the Gaussian wake model;
[0097] The calculation process of the wind farm wake environmental parameters is as follows: the unit coordinates are transformed to obtain the cross-flow coordinates of the downstream wind turbines, and based on the Gaussian wake model formula, the environmental parameters of the wake overlap ratio, the loss wind speed and the distance between wind turbines are calculated; the specific process is as follows: the Gaussian wake model formula (1) is obtained according to the aerodynamic theory of wind turbines:
[0098]
[0099] In the formula, V xmin V represents the wind speed at the center of the wake at position x behind the wind rotor, that is, the minimum wind speed of the wake cross section at that position; x is the wind speed at each location on the cross section at position x behind the wind rotor; C T represents the thrust coefficient of the current wind turbine; σ is the standard deviation of the Gaussian at the position x behind the wind rotor; h is the center height of the wind rotor; R is the radius of the wind rotor; when x approaches 0, Δr is a constant, which can make V xminIt is equal to the wind speed value immediately behind the wind rotor; z represents the coordinate along the blade direction, and y represents the coordinate in the vertical direction;
[0100] Since the wake only affects some downstream wind turbines, the wind speed loss of these turbines is less than that of being completely covered by the wake. Therefore, the wind speed loss in the single wake model is multiplied by the area ratio of the wake and the wind rotor rotating disk overlap to calculate the actual wind speed at the downstream turbine. The loss wind speed at any position can be calculated according to the Gaussian wake model formula as shown in formula (2). The influence of the wake overlap area between wind turbines can be calculated according to the cosine theorem as shown in formula (3). is the wake overlap ratio;
[0101]
[0102] In the formula, V ij represents the effective wind speed at the location of wind turbine j; V0 is the inflow wind speed of upstream wind turbine i; S represents the area of the overlap between the cross section of the wake of the upstream wind turbine i at the downstream wind turbine j and the plane swept by the wind rotor of wind turbine j; j represents the rotor area of the downstream wind turbine j; θ1 and θ2 are the arcs of the overlapping wake and rotor parts, respectively, which are calculated by the cosine theorem; R1 represents the rotor radius of wind turbine j; R2 represents the rotor radius of wind turbine i; C T,i represents the thrust coefficient of wind turbine i;
[0103] Transform the unit coordinates to obtain the cross-flow coordinates of the downstream fan (cross-flow coordinates are a commonly used concept in fluid dynamics and fluid mechanics, describing the flow characteristics perpendicular to the main flow direction;
[0104] Similar to the xy-axis decomposition projection), since the wake deflection in the simulation experiment is tested at a single average wind speed, the speed change caused by wind resource turbulence is limited, so a simple linear function is used to parameterize the lateral deviation of the wake caused by the wake rotation. The following are the detailed calculation steps:
[0105] 1. Linear function representation of wake deviation caused by rotation;
[0106] y w,rotation (x) = a d +b d x (4)
[0107] Among them, a d and b d It is a parameter that needs to be determined by empirical methods; w,rotation (x) is a linear function used to parameterize the lateral deviation of the wake caused by rotation;
[0108] 2. Linear function representation of wake deviation caused by yaw;
[0109] By integrating the tangent of the wake centerline angle divided by x, the lateral offset of the wake center caused by the yaw of the wind rotor is obtained;
[0110]
[0111] In the formula, y w,yaw,j (x) represents the lateral offset of the wake center relative to the hub of the wind turbine j rotor caused by yaw; x represents the axial distance from the center of the wind rotor; D is the diameter of the wind rotor; k d is the coefficient of wake deflection caused by yaw; a j is the rotor axial induction factor; γ j is the yaw angle; Indicates the thrust coefficient of the fan when it is yawed;
[0112] 3. Combining the rotation-induced component and the yaw-induced component to obtain the cross-flow coordinates, the following formula:
[0113] y w,j (x) = y w,rotation (x)+y w,yaw,j (x) (6)
[0114] In the formula, y w,j (x) represents the lateral position coordinate of the wake center relative to the hub of the wind turbine j; y w,rotation (x) is used to parameterize the lateral deviation of the wake caused by rotation; y w,yaw,j (x) represents the lateral displacement of the wake center relative to the hub of the wind turbine j caused by yaw;
[0115] Based on the Gaussian wake model formula, the environmental parameters of wake overlap ratio, deficit wind speed and distance between wind turbines are calculated.
[0116] S3. construct an initial wake directed graph based on environmental parameters, perform sparse processing on the initial wake directed graph, disconnect the edges with lower weights in the initial wake directed graph to obtain a sparse wake directed graph, and construct a subset of wind turbines based on the sparse wake directed graph;
[0117] The process of constructing a subset of wind turbines includes: constructing an initial wake directed graph based on the environmental parameters, obtaining a sparse directed graph through pruning based on a pruning threshold and a wake weight coefficient, and dividing the on-site wind turbines into multiple subsets of wind turbines based on the sparse directed graph; the specific process is as follows:
[0118] The method for constructing the initial wake directed graph is as follows:
[0119] Define a directed graph G = (V, E, A), where the vertex set V = {v i |i=1,2,K,N} represents each wind turbine in the wind farm; each edge of the graph represents the wake distribution between wind turbines; the non-negative adjacency matrix A = (w ij ) N×N Indicates the wake influence strength between wind turbines; defines a pair of elements e ij =(v i ,v j ), which is the vertex v of the directed graph i To the vertex v j The edge weight of the connection, e ij It represents the wake effect of upstream wind turbine i on downstream wind turbine j, that is, the influence of wake effect is unidirectional; the definition of wake intensity weighted non-negative matrix A is as follows:
[0120]
[0121] Among them, w ij represents the weight value of the wake effect of the upstream wind turbine i on the downstream wind turbine j; for the element w in the wake intensity weighting matrix A ij If the upstream wind turbine i has a wake effect on the downstream wind turbine j, the wake weight coefficient is expressed as w ij ; Otherwise the element remains 0;
[0122] The method to construct a sparse wake directed graph is:
[0123] Determining the pruning threshold is an important parameter in performing pruning operations. It determines which parameters or weight values will be removed to simplify the model. The global pruning threshold is ε k = k*ε, where the coefficient k is a hyperparameter that controls the sparsity of the wake directed graph; the threshold ε is the average value of the entire wake weight coefficient, generating a pruned directed graph G p =(V,E p ,A p ), where V represents the node set of the directed graph, E p represents the edge set after pruning, A p represents the adjacency matrix after pruning; according to the pruning threshold, the edges corresponding to the weight matrix A are removed to obtain a sparse tail flow directed graph;
[0124] The method to generate a subset of wind turbines is:
[0125] The free stream wind speed V ∞ The front exhaust fan is used as the leading node of each subset, and the BFS algorithm is used to traverse the connected vertices to generate different directed graph subsets n. i , determine the directed graph subset ni Is there a shared set S between i ; If the shared wind turbines belong to multiple subsets N i (The subsets are N1, N2, N3, etc. formed after sparse processing, and the shared subsets such as S1 are formed by wind turbines in N1 that belong to multiple subsets at the same time, which is called a shared subset). Then calculate the shared wind turbines T in each subset. i The weighted degree centrality C indeg The sum of all C indeg Sort and share wind turbines T i Assign it to the subset with the largest degree centrality weight, and construct a decoupled sparse tail directed graph G s =(V,E s ,A s ), based on the sparse wake directed graph, the on-site units are divided into multiple wind turbine subsets N i ;
[0126] Weighted degree centrality is a comprehensive indicator to measure the importance of shared vertices, which takes into account the number of adjacent vertices and the weight of the connection. indeg The calculation method is as follows:
[0127]
[0128] Where v represents the vertex in the subgraph; w1(v) represents the in-degree weight of the v node; and L represents the in-degree value.
[0129] S4. constructing a wind farm yaw angle optimization control model based on the subset of wind turbines, and training the yaw angle optimization control model based on the wind farm environment;
[0130] The process of constructing a wind farm yaw angle optimization control model includes: constructing a hierarchical reinforcement learning optimization control network based on a subset of wind turbines after sparse processing, wherein the upper layer optimizes between wind turbine subsets, and the lower layer optimizes within wind turbine subsets to achieve hierarchical optimization control; the reinforcement learning optimization control network is based on the Actor-Critic framework, and adopts a one-layer full connection, two-layer graph attention and three-layer full connection sequentially superimposed to obtain a reinforcement learning optimization control network; the hierarchical control network is integrated with the reinforcement learning optimization control network to obtain a yaw angle optimization control model; the specific steps are as follows:
[0131] S41, build a hierarchical optimization control network, and establish multiple wind turbine subsets N i To achieve hierarchical optimization control, the upper layer uses each wind turbine subset processed by the sparse directed graph as an intelligent agent to optimize between subsets, and the lower layer uses each wind turbine subset N as an intelligent agent to optimize between subsets. i Each wind turbine in the wind turbine subset N is an intelligent agent in iThe optimization is performed within the wind farm, and the upper optimization target is set to maximize the total power of the wind farm, and the lower ... i The total power of the inner wind farm is the largest; the upper and lower optimization processes are simultaneously optimized under different time granularity conditions. The lower layer will be optimized in each time step, and the upper layer will be optimized once every 10 time steps. At this time, information such as wind turbine power and control variables are exchanged between the upper and lower layers;
[0132] S42. Construct a reinforcement learning optimization control network based on the sparse wake multi-agent graph attention MASAC algorithm (GAT-Sparse Graph Grouping Multi-Agent Soft Actor-Critic, GAT-SGGMASAC); GAT-SGGMASAC is based on the Actor-Critic framework. The Actor module uses a fully connected layer for feature transformation, and then uses two layers of graph attention layers (GAT) for superposition, and then uses three fully connected layers to realize the nonlinear mapping of system state to action strategy. Similarly, the Critic module structure also adopts the form of superposition of two layers of GAT and three layers of full connection; the hierarchical control network is integrated with the reinforcement learning optimization control network to obtain the GAT-SGGMASAC yaw angle optimization control model;
[0133] S43, train the GAT-SGGMASAC network, and the thrust and output power calculation methods of the wind turbine are shown in equations (9) and (10);
[0134]
[0135] Where i = 1, 2, ..., N, N is the total number of wind turbines; A d is the rotor disk area of wind turbine i; U i is the wind speed on the rotor plane at wind turbine i; ρ is the air density; and are the thrust coefficient and power coefficient of the i-th wind turbine, respectively, which are related to the axial induction factor a i and yaw angle φ i The calculation methods are shown in formula (11) and formula (12):
[0136]
[0137]
[0138] In the formula, μ is the calculation hyperparameter;
[0139] The total power calculation formula for a wind farm with N wind turbines is shown in (13);
[0140]
[0141] Control the yaw angle φ of each wind turbine in the wind farm i and thrust coefficient The problem of increasing the total power of a wind farm is considered as an optimal control problem with the reward function (14):
[0142]
[0143] Where t represents the current time step; γ is the power conversion factor; x(t) is the control variable, which includes the thrust coefficient and yaw angle variables.
[0144] The optimization goal is to find a set of optimal control strategies x * (t) maximizes the reward J(t), and the optimization problem is formulated as:
[0145]
[0146] The optimization problem needs to meet the following conditions:
[0147]
[0148] In the formula, C′ T,min , C′ T,max ,φ min ,φ max are the lower and upper bounds of the thrust coefficient and yaw angle, respectively.
[0149] In this embodiment, in order to verify the effect of the yaw angle optimization control model, the yaw angle optimization control model construction method of this embodiment is compared with the three reinforcement learning yaw control models of MATD3, DDPG and DQN. The GAT-SGGMASAC model has better optimization effect and requires less time for optimization.
[0150] Specifically, in order to more intuitively reflect the performance of the yaw optimization model of the method of this embodiment, the training process of GAT-SGGMASAC is compared with other reinforcement learning yaw control models. The reward curve of the learning process is as follows: Figure 3 As shown, the ordinate is the reward for wind farm training, and the abscissa is the training round. Each round includes multiple steps of interaction between the yaw control method and the wind farm environment. After each round of training, the cumulative reward record given by the wind farm environment is plotted. The higher the reward, the better the performance of the control strategy.
[0151] In order to more intuitively reflect the rapidity of the yaw optimization model of the method of this embodiment, the optimization time required for the yaw angle optimization control model construction method of this embodiment and the three reinforcement learning yaw control models of MATD3, DDPG and DQN are compared under four wind angle conditions of 0°, 10°, 30° and 60°, as shown in Table 1:
[0152] Table 1
[0153]
[0154] Embodiment 2:
[0155] like Figure 4 As shown, this embodiment provides a wind farm yaw angle optimization control strategy, including:
[0156] S11, obtaining the location information, environmental information and wind turbine parameters of all wind turbines in the wind farm; the location information refers to the coordinates of the wind turbines, the environmental information refers to the corresponding wind direction and wind speed, and the wind turbine parameters refers to the radius of the wind rotor;
[0157] S12, transform the coordinates of the wind turbine generator set, and calculate the wind farm wake distribution in the simulation environment based on the Gaussian wake model. The specific calculation process is as follows:
[0158] According to the basic assumption of the Gaussian wake model, that is, the wake cross section is Gaussian, the wind speed V at the position x behind the wind rotor at a distance x′ from the wake center is obtained. x , as shown in formula (17).
[0159] V x =V xmin f(x′ / x) (17)
[0160] Where V xmin V is the wind speed at the center of the wake at position x behind the wind rotor, that is, the minimum wind speed of the wake cross section at that position; x is the wind speed at the point x′ from the center of the wake on the cross section at position x behind the wind rotor; x′ is the vertical distance from the axis of the wind rotor;
[0161] For any incoming wind speed, the wake region is Gaussian, as shown in formula (18);
[0162]
[0163] Where σ is the standard deviation of the Gaussian at the position x behind the wind rotor; V x is the wind speed distribution on the wake cross section;
[0164] The wake boundary of the Gaussian wake model is the same as that of the traditional Jensen wake model. The wake velocity at the x position behind the wind rotor can be finally derived, as shown in formula (19);
[0165]
[0166] In the formula, V xmin V represents the wind speed at the center of the wake at position x behind the wind rotor, that is, the minimum wind speed of the wake cross section at that position; x is the wind speed at each location on the cross section at position x behind the wind rotor; C T represents the thrust coefficient of the current wind turbine; σ is the standard deviation of the Gaussian at the position x behind the wind rotor; h is the center height of the wind rotor; R is the radius of the wind rotor; when x approaches 0, Δr is a constant, which can make V xmin It is equal to the wind speed value immediately behind the wind rotor; z represents the coordinate along the blade direction, and y represents the coordinate in the vertical direction;
[0167] The calculation process of the wind farm wake environmental parameters is as follows: transforming the unit coordinates to obtain the cross-flow coordinates of the downstream wind turbines, and calculating the environmental parameters of the wake overlap ratio, the loss wind speed, and the distance between wind turbine units based on the Gaussian wake model formula;
[0168] Specifically, the specific process of obtaining the wind farm wake environmental parameters in this embodiment is the same as the process of obtaining the environmental parameters in the first embodiment;
[0169] S13, inputting the wind farm wake environment information into a yaw angle optimization control model to obtain an optimal yaw angle action strategy;
[0170] The process of obtaining the optimal yaw angle action includes: based on the wind farm wake environment information, using the yaw angle optimization control model and outputting the maximum reward valuation according to the forward calculation, obtaining the real-time wind farm yaw angle optimal action strategy; specifically:
[0171] By controlling the yaw angle and thrust coefficient of each wind turbine in the wind farm, the optimal control strategy is quickly solved to maximize the reward J(t). The reward J(t) in this embodiment is the same as the reward J(t) solution process in the first embodiment, so that the wind turbine obtains the optimal action strategy for the real-time yaw angle, and the total power of the wind farm is improved.
[0172] The verification process of the optimal yaw angle action strategy is as follows:
[0173] The wind farm wake environment information is respectively input into the hierarchical reinforcement learning optimization control strategy and the greedy strategy to obtain the first total power output value and the second total power output value respectively, and the first total power output value and the second total power output value are compared to obtain the evaluation result of the wind farm yaw angle optimization control strategy. Specifically, in order to verify that the wind farm yaw angle optimization control strategy of this embodiment can improve the total power generation of the wind farm, different wind speed conditions of 6m / s to 14m / s are set, and the yaw control of the wind farm is performed using the yaw angle optimization control model construction method of this embodiment, four reinforcement learning yaw control strategies of MATD3, DDPG and DQN, and the greedy strategy. The total power output of the wind farm is as follows: Figure 4 As shown, when four active yaw control strategies and the greedy strategy are used for optimization under different wind speed conditions, the yaw angle optimization control model construction method of this embodiment has the best effect on increasing the total power of the wind farm. Compared with the greedy strategy, the power increase level of the wind farm in the wind speed range of 6m / s to 14m / s is 11.1% to 17.2%.
[0174] Embodiment three:
[0175] This embodiment provides a device for constructing a yaw angle optimization control model for a wind farm, which is used to implement the method for constructing a yaw angle optimization control model described in Embodiment 1. The device includes:
[0176] A data acquisition module is used to obtain the location information, environmental information and wind turbine parameters of all wind turbines in the wind farm;
[0177] The modeling parameter acquisition module is used to transform the unit coordinates to obtain the cross-flow coordinates of the downstream wind turbine. Based on the Gaussian wake model formula, the environmental parameters of the wake overlap ratio, the loss wind speed, and the distance between wind turbines are calculated.
[0178] A relationship building module is used to calculate the wake weight coefficient under the environmental parameters, and to build a sparse directed graph based on the pruning threshold and the wake weight coefficient; and to build multiple wind turbine subsets N based on the sparse directed graph. i ;
[0179] The model building module is used to build a GAT-SGGMASAC yaw angle optimization control model based on the wind turbine subset relationship information.
[0180] This embodiment has all the advantages of the method for constructing the yaw angle optimization control model described in the first embodiment.
[0181] Embodiment 4:
[0182] This embodiment provides a wind farm yaw angle optimization control strategy device, which is used to implement the yaw angle optimization control strategy described in the second embodiment, and the device includes:
[0183] The control data acquisition module is used to obtain the location information, environmental information and wind turbine parameters of all wind turbines in the wind farm;
[0184] The parameter acquisition module is used to transform the unit coordinates to obtain the cross-flow coordinates of the downstream wind turbine, and calculate the environmental parameters of the wake overlap ratio, the loss wind speed, and the distance between wind turbines based on the Gaussian wake model formula;
[0185] The model output module is used to input the environmental parameters into the yaw angle optimization control model constructed by the method of embodiment 1, and obtain the real-time wind farm yaw angle optimal action strategy according to the maximum reward valuation output by the model forward calculation.
[0186] This embodiment has all the advantages of the wind farm yaw angle optimization control strategy described in the second embodiment.
[0187] The present invention obtains the coordinates of all wind turbines in a wind farm, determines the parameter values of wind direction, wind speed, distance between wind turbines, and radius of wind rotors; obtains key environmental parameters of the wind farm, and ensures that the model can accurately reflect the actual wind farm conditions; transforms the coordinates of the wind turbines, and calculates the wake distribution of the wind farm based on the Gaussian wake model; utilizes the coordinate transformation and the Gaussian wake model calculation to accurately describe the influence of the wake effect between the front and rear wind turbines mathematically, which is convenient for optimizing the yaw angle; constructs an original wake directed graph according to the environmental parameters, performs sparse processing on the directed graph, disconnects the edges with lower weights in the directed graph, obtains a sparse wake directed graph, and constructs a subset of wind turbines based on the sparse wake directed graph; uses the directed graph to represent The invention shows the wake influence between wind turbines in the wind farm. The sparse processing is to simplify the directed graph, remove the wake interference with less influence (lower weight), reduce the computational complexity, make the model more efficient, and retain the main wake influence path to ensure the accuracy of the optimization process. The subset division method is helpful for local optimization and improves the efficiency and effect of the optimization calculation. Based on the subset of wind turbines, a wind farm yaw angle optimization control model is constructed, and the yaw angle optimization control model is trained based on the wind farm environment. Through the training of the model, the model can learn a set of strategies for optimal yaw angle setting under different wind conditions, so as to maximize the total power output of the wind farm. Then the real-time wind farm wake environment information is input into the yaw angle optimization control model to obtain the optimal action strategy for the yaw angle.
[0188] The present invention reduces the computing resources of the yaw control strategy, shortens the optimization time, can effectively improve the yaw angle control effect of the wind farm under complex and changeable wind direction environment, and further improves the total output power of the wind farm, and has better optimization effect and robustness.
[0189] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.
Claims
1. A method for constructing a wind farm yaw angle optimization control model, characterized in that: Use the following steps: S1. Obtaining the location information, environmental information and wind turbine parameters of all wind turbines in the wind farm; S2, transform the coordinates of the wind turbine group, and calculate the wind farm wake environment parameters based on the Gaussian wake model; S3. construct an initial wake directed graph based on environmental parameters, perform sparse processing on the initial wake directed graph, disconnect the edges with lower weights in the initial wake directed graph to obtain a sparse wake directed graph, and construct a subset of wind turbines based on the sparse wake directed graph; S4. Based on the subset of wind turbines, a wind farm yaw angle optimization control model is constructed, and the yaw angle optimization control model is trained based on the wind farm environment.
2. The method for constructing a yaw angle optimization control model according to claim 1, characterized in that: The wind farm wake environmental parameters described in S2 include wake overlap ratio, deficit wind speed and distance between wind turbines; According to the aerodynamic theory of wind turbines, the Gaussian wake model formula is (1): In the formula, V xmin V represents the wind speed at the center of the wake at position x behind the wind rotor, that is, the minimum wind speed of the wake cross section at that position; x is the wind speed at each location on the cross section at position x behind the wind rotor; C T represents the thrust coefficient of the current wind turbine; σ is the standard deviation of the Gaussian at the position x behind the wind rotor; h is the center height of the wind rotor; R is the radius of the wind rotor; Δr is a constant. When x approaches 0, Δr can make V xmin It is equal to the wind speed value immediately behind the wind rotor; z represents the coordinate along the blade; y represents the coordinate in the vertical direction; Since the wake only affects some downstream wind turbines, the wind speed loss of these turbines is less than that of being completely covered by the wake. Therefore, the wind speed loss in the single wake model is multiplied by the ratio of the area where the wake overlaps with the rotating disk of the wind rotor, so as to calculate the actual wind speed at the downstream turbine. The loss wind speed at any position can be calculated according to the Gaussian wake model formula as shown in formula (2). The area of the wake overlap can be calculated according to the cosine theorem. The influence between wind turbines is shown in formula (3), where is the wake overlap ratio; In the formula, V ij represents the effective wind speed at the location of wind turbine j; V0 is the inflow wind speed of upstream wind turbine i; S represents the area of the overlap between the cross section of the wake of the upstream wind turbine i at the downstream wind turbine j and the plane swept by the wind rotor of wind turbine j; j represents the rotor area of the downstream wind turbine j; θ1 and θ2 are the arcs of the overlapping wake and rotor parts, respectively, which are calculated by the cosine theorem; R1 represents the rotor radius of wind turbine j; R2 represents the rotor radius of wind turbine i; C T,i represents the thrust coefficient of wind turbine i; The unit coordinates are transformed to obtain the cross-flow coordinates of the downstream wind turbine. Since the wake deflection in the simulation experiment is tested at a single average wind speed, the speed change caused by wind resource turbulence is limited. Therefore, a linear function is used to parameterize the lateral deviation of the wake caused by the wake rotation. The following are the detailed calculation steps:
1. Linear function representation of wake deviation caused by rotation; y w,rotation (x)=a d +b d x (4) Among them, a d and b d It is a parameter that needs to be determined by empirical methods; w,rotation (x) is a linear function used to parameterize the lateral deviation of the wake caused by rotation; 2. Linear function representation of wake deviation caused by yaw; By dividing the tangent of the wake centerline angle by x, the lateral offset of the wake center caused by the yaw of the wind rotor is obtained, as shown in the following formula: In the formula, y w,yaw,j (x) represents the lateral offset of the wake center relative to the hub of the wind turbine j rotor caused by yaw; x represents the axial distance from the center of the wind rotor; D is the diameter of the wind rotor; k d is the coefficient of wake deflection caused by yaw; a j is the rotor axial induction factor; γ j is the yaw angle; Indicates the thrust coefficient of the fan when it is yawed; 3. Combining the rotation-induced component and the yaw-induced component to obtain the cross-flow coordinates, the following formula: and w,j (x)=y w,rotation (x)+y w,yaw,j (x) (6) In the formula, y w,j (x) represents the lateral position coordinate of the wake center relative to the hub of the wind turbine j; y w,rotation (x) is used to parameterize the lateral deviation of the wake caused by rotation; y w,yaw,j (x) represents the lateral displacement of the wake center relative to the hub of the wind turbine j caused by yaw; Based on the Gaussian wake model formula, the environmental parameters of wake overlap ratio, deficit wind speed and distance between wind turbines are calculated.
3. The method for constructing a yaw angle optimization control model according to claim 1, characterized in that: In step S3, the method for constructing the initial wake directed graph is as follows: Define a directed graph G = (V, E, A), where the vertex set V = {v i |i=1,2,K,N} represents each wind turbine in the wind farm; each edge of the graph represents the wake distribution between wind turbines; the non-negative adjacency matrix A = (w ij ) N×N Indicates the wake influence strength between wind turbines; defines a pair of elements e ij =(v i ,v j ), which is the vertex v of the directed graph i To the vertex v j The edge weight of the connection, e ij It represents the wake effect of upstream wind turbine i on downstream wind turbine j, that is, the influence of wake effect is unidirectional; The wake intensity weighted non-negative matrix A is defined as follows: Among them, w ij represents the weight value of the wake effect of the upstream wind turbine i on the downstream wind turbine j; for the element w in the wake intensity weighting matrix A ij If the upstream wind turbine i has a wake effect on the downstream wind turbine j, the wake weight coefficient is expressed as w ij ; Otherwise the element remains 0; The method to construct a sparse wake directed graph is: Determine the pruning threshold, the global pruning threshold is ε k = k*ε, where the coefficient k is a hyperparameter that controls the sparsity of the wake directed graph; the threshold ε is the average value of the entire wake weight coefficient, generating a pruned directed graph G p =(V,E p ,A p ), where V represents the node set of the directed graph, E p represents the edge set after pruning, A p represents the adjacency matrix after pruning; according to the pruning threshold, the edges corresponding to the weight matrix A are removed to obtain a sparse tail flow directed graph; The method to generate a subset of wind turbines is: The free stream wind speed V ∞ The front exhaust fan is used as the leading node of each subset, and the BFS algorithm is used to traverse the connected vertices to generate different directed graph subsets n. i , determine the directed graph subset n i Is there a shared set S between i ; If the shared wind turbines belong to multiple subsets N i , then calculate the shared wind turbine group T in each subset i The weighted degree centrality C indeg The sum of all C indeg Sort and share wind turbines T i Assign it to the subset with the largest degree centrality weight, and construct a decoupled sparse tail directed graph G s =(V,E s ,A s ), based on the sparse wake directed graph, the on-site units are divided into multiple wind turbine subsets N i ; Weighted degree centrality is a comprehensive indicator to measure the importance of shared vertices, which takes into account the number of adjacent vertices and the weight of the connection. indeg The calculation method is as follows: Where v represents the vertex in the subgraph; w1(v) represents the in-degree weight of the v node; and L represents the in-degree value.
4. The method for constructing a yaw angle optimization control model according to claim 3, characterized in that: The following steps are used to construct the yaw angle optimization control model: S41, construct a hierarchical optimization control network and establish several fan subsets N i To achieve hierarchical optimization control, the upper layer uses each wind turbine subset processed by the sparse directed graph as an intelligent agent to optimize between subsets, and the lower layer uses each wind turbine subset N as an intelligent agent to optimize between subsets. i Each wind turbine in the wind turbine subset N is an intelligent agent in i The optimization is performed within the wind farm, and the upper optimization target is set to maximize the total power of the wind farm, and the lower ... i The total power of the inner wind farm is the largest; the upper and lower optimization processes are simultaneously optimized under different time granularity conditions. The lower layer will be optimized in each time step, and the upper layer will be optimized once every 10 time steps. At this time, information such as wind turbine power and control variables are exchanged between the upper and lower layers; S42. Construct a reinforcement learning optimization control network based on the sparse wake multi-agent graph attention MASAC algorithm (GAT-Sparse Graph Grouping Multi-Agent Soft Actor-Critic, GAT-SGGMASAC); GAT-SGGMASAC is based on the Actor-Critic framework. The Actor module uses a fully connected layer for feature transformation, and then uses two layers of graph attention layers (GAT) for superposition, and then uses three fully connected layers to realize the nonlinear mapping of system state to action strategy. Similarly, the Critic module structure also adopts the form of superposition of two layers of GAT and three layers of full connection; the hierarchical control network is integrated with the reinforcement learning optimization control network to obtain the GAT-SGGMASAC yaw angle optimization control model; S43, train the GAT-SGGMASAC network, and the thrust and output power calculation methods of the wind turbine are shown in equations (9) and (10); Where i = 1, 2, ..., N, N is the total number of wind turbines; A d is the rotor disk area of wind turbine i; U i is the wind speed on the rotor plane at wind turbine i; ρ is the air density; and are the thrust coefficient and power coefficient of the i-th wind turbine, respectively, which are related to the axial induction factor a i and yaw angle φ i The calculation methods are shown in formula (11) and formula (12): C Pi =4a i / (cos(μφ i )-a i ) 2 (12) In the formula, μ is the calculation hyperparameter; The total power calculation formula for a wind farm with N wind turbines is shown in (13); Control the yaw angle φ of each wind turbine in the wind farm i and thrust coefficient The problem of increasing the total power of a wind farm is considered as an optimal control problem with the reward function (14): Where t represents the current time step; γ is the power conversion coefficient; x(t) is the control variable, which includes the thrust coefficient and yaw angle variable [C′ T1 (t),...,C′ TN (t),φ1(t),...,φ N (t)] T ; The optimization goal is to find a set of optimal control strategies x * (t) maximizes the reward J(t), and the optimization problem is formulated as: The optimization problem needs to meet the following conditions: In the formula, C′ T,min , C′ T,max ,φ min ,φ max are the lower and upper bounds of the thrust coefficient and yaw angle, respectively.
5. A wind farm yaw angle optimization control strategy, characterized in that: Use the following steps: S11, obtaining location information, environmental information and wind turbine parameters of all wind turbines in the wind farm; S12, transforming the coordinates of the wind turbine group, and calculating the wind farm wake distribution in the simulation environment based on the Gaussian wake model; S13. Inputting the wind farm wake environment information into the yaw angle optimization control model as described in any one of claims 1 to 4 to obtain an optimal yaw angle action strategy.
6. The wind farm yaw angle optimization control strategy according to claim 5, characterized in that: The process of calculating the wind farm wake environment information: The coordinates of the unit are transformed to obtain the cross-flow coordinates of the downstream wind turbine. Based on the Gaussian wake model formula, the environmental parameters of the wake overlap ratio, the deficit wind speed, and the distance between wind turbines are calculated.
7. The wind farm yaw angle optimization control strategy according to claim 6, characterized in that: The process of obtaining the optimal yaw angle action includes: Based on the wind farm wake environment information, the yaw angle optimization control model is used and the maximum reward valuation is output according to the forward calculation to obtain the optimal action strategy for the real-time yaw angle of the wind farm; Specifically, by controlling the yaw angle and thrust coefficient of each wind turbine in the wind farm, the optimal control strategy is quickly solved to maximize the reward J(t), so that the wind turbine can obtain the optimal action strategy for the real-time yaw angle and achieve an increase in the total power of the wind farm.
8. The wind farm yaw angle optimization control strategy according to claim 5, characterized in that: It also includes the verification process of the optimal yaw angle action strategy, and the specific operations are as follows: The wind farm wake environment information is respectively input into the hierarchical reinforcement learning optimization control strategy and the greedy strategy to obtain a first total power output value and a second total power output value respectively, and the first total power output value and the second total power output value are compared to obtain an evaluation result of the wind farm yaw angle optimization control strategy.
9. A device for constructing a wind farm yaw angle optimization control model, characterized in that: The device is used to implement the method for constructing a yaw angle optimization control model according to any one of claims 1 to 4, comprising: The modeling data acquisition module is used to obtain the coordinates, wind direction, wind speed, distance between wind turbines and environmental information of the wind rotor radius of all wind turbines in the wind farm; The parameter acquisition module is used to transform the unit coordinates to obtain the cross-flow coordinates of the downstream wind turbine, and calculate the environmental parameters of the wake overlap ratio, the loss wind speed, and the distance between wind turbines based on the Gaussian wake model formula; A relationship building module is used to calculate the wake weight coefficient under the environmental parameters, and to build a sparse directed graph based on the pruning threshold and the wake weight coefficient; based on the sparse directed graph, multiple wind turbine subsets N are obtained. i ; The model building module builds a GAT-SGGMASAC yaw angle optimization control model based on the wake coupling relationship information of the wind turbine subset.
10. A wind farm yaw angle optimization control strategy device, characterized in that: For implementing the yaw angle optimization control strategy according to any one of claims 5 to 8, the device comprises: The control data acquisition module is used to obtain the location information, environmental information and wind turbine parameters of all wind turbines in the wind farm; The parameter acquisition module is used to transform the unit coordinates to obtain the cross-flow coordinates of the downstream wind turbine, and calculate the environmental parameters of the wake overlap ratio, the loss wind speed, and the distance between wind turbines based on the Gaussian wake model formula; A model output module is used to input the environmental parameters into the yaw angle optimization control model as described in any one of claims 1 to 4, and obtain a real-time wind farm yaw angle optimal action strategy based on the maximum reward valuation output by the model forward calculation.
Citation Information
Cited By
Wind power plant yaw control method, system and device based on layered optimization and medium
CN119755011A
Wind farm yaw control method, system, device and medium based on hierarchical optimization
CN119755011B
Wind turbine generator operation control optimization method and system based on deep learning
CN119982338A
Wind-storage-load cooperative control method and device considering containment control of deep and far sea wind power plant
CN120262529A
Wind power plant power dynamic optimization method and system based on spectral clustering partition
CN120300942A
