Wind power plant power optimization method and system based on BFGS and knowledge generalization mechanism data driving

Abstract

The invention belongs to the technical field of generation power optimization of a wind power plant, and provides a wind power plant power optimization method and system based on BFGS and knowledge generalization mechanism data driving. According to the method, an optimization framework of a quasi-Newton method is adopted, and according to wind power plant power generation data information measured in real time, the gradient estimated by synchronous perturbation stochastic approximation and Hessian matrix inversion estimated by BFGS are utilized to complete updating of all fan control actions; a knowledge generalization mechanism is designed, and real-time generalization and full utilization are carried out on learned action knowledge; the method does not depend on a wind power plant generation power model, and solves the problems that an existing optimization scheme is slow in convergence speed and difficult to effectively improve the limitation that the modeling wind power plant generation power is difficult to analyze.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind farm power generation optimization, and in particular relates to a wind farm power optimization method and system based on BFGS and knowledge generalization mechanism data drive. Background Art

[0002] Wakes are time-, space-, and parameter-dependent, and are also affected by topography. This makes wake characteristics extremely complex for mountainous wind farms, making it difficult to establish analytical wake models and wind farm power models, thus limiting the use of model-driven power optimization methods.

[0003] Data-driven power optimization methods do not rely on a wind farm's power generation model. Instead, they obtain the optimal wind turbine control inputs solely through control inputs and measurement data. These methods are suitable for wind farm power generation optimization problems that are difficult to analytically model. However, wind conditions in real wind farms are complex and random, and the wake coupling between wind turbines is strongly correlated with wind conditions, causing the optimal solution to wind farm power optimization problems to rapidly vary. Currently, existing data-driven power optimization methods converge slowly and can only adapt to fixed or slowly changing wind conditions, making it difficult to effectively improve wind farm power generation. Summary of the Invention

[0004] In order to solve the above problems, the present invention proposes a data-driven wind farm power optimization method and system based on BFGS and knowledge generalization mechanism. The present invention adopts the optimization framework of the quasi-Newton method, and according to the measured wind farm power generation data information, uses the gradient estimated by synchronous perturbation random approximation and the inverse Hessian matrix estimated by BFGS to complete the update of all wind turbine control actions. At the same time, the present invention designs a knowledge generalization mechanism to generalize and fully utilize the learned action knowledge in real time, further accelerating the convergence speed of the proposed method. The present invention does not rely on the power generation model of the wind farm, but only uses measured data to achieve the purpose of quickly optimizing the power generation of the entire wind farm, solving the problems of the slow convergence speed of existing optimization schemes and the difficulty in effectively improving the power generation of wind farms.

[0005] In order to achieve the above object, the present invention is implemented through the following technical solutions: In a first aspect, the present invention provides a data-driven wind farm power optimization method based on BFGS and a knowledge generalization mechanism, comprising: According to the power generation efficiency of the wind farm under different wind directions, the entire wind direction interval is divided into multiple sub-intervals; For each divided subinterval, define the wind farm power generation efficiency optimization subproblem; According to the defined power generation efficiency optimization sub-problem, based on BFGS and knowledge generalization mechanism, a data-driven wind farm power optimization scheme is obtained; wherein, when the optimization scheme is determined, in the action decision stage, for the current wind direction, the optimization algorithm corresponding to the sub-interval decides the action: if the logical variable is true, then the action update is completed before the action is generated. When the action is updated, the estimated Newton step path is selected as the search direction, and the projection operator is used to ensure that the decision action meets the control constraints of all wind turbines. After the action update is completed, the perturbation action is generated with the help of the perturbation parameter and perturbation vector; if the logical variable is not true, the action update is not performed, and the perturbation action is directly generated. action; in the action evaluation stage, the wind farm executes the action and collects the power generation data and wind condition data of the wind farm to obtain the power generation efficiency of the wind farm; in the strategy update stage, if the wind direction data at two consecutive sampling moments belong to the same wind direction sub-interval, the implemented action is effectively evaluated in the actual wind farm: if the logical variable is true, the power generation efficiency is retained and the logical variable is set to false; if the logical variable is not true, the gradient information of the current iteration point is estimated according to the synchronous perturbation stochastic approximation, and the inverse Hessian matrix is ​​estimated with the help of the quasi-Newton method BFGS. Based on the knowledge generalization mechanism, the learned action knowledge is generalized and fully utilized in real time, and the logical variable is set to true.

[0006] Furthermore, within each subinterval, the wake coupling strength between wind turbines at different wind directions is considered to be the same. It is assumed that only one wake coupling mode exists, and the subproblem is regarded as a static optimization problem, and the corresponding optimal solution is regarded as a constant vector.

[0007] Furthermore, the estimated Newton step size is selected as the search direction: ; in, is the estimated Newton step, is the number of iterations; At the iteration point The inverse Hessian matrix estimated at ; is the estimated gradient.

[0008] Furthermore, the projection operator is used to ensure that the decision action satisfies the control constraints of all wind turbines: ; in, It's about The projection operator, is the step length.

[0009] Furthermore, if the wind direction data at two consecutive moments belong to the same subinterval, the implemented action is effectively evaluated in the actual wind farm: if the logical variable is true, the power generation efficiency is retained and the logical variable is set to false; otherwise, the gradient information of the current iteration point is estimated based on the synchronous perturbation stochastic approximation, and the inverse Hessian matrix is ​​estimated with the help of the quasi-Newton method BFGS.

[0010] Furthermore, if the wind direction data at two consecutive moments belong to the same subinterval, based on the knowledge generalization mechanism, the learned action knowledge is generalized and fully utilized in real time, and the logical variable is set to true.

[0011] In a second aspect, the present invention further provides a wind farm power optimization system driven by data based on BFGS and knowledge generalization mechanism, comprising: The partitioning module is configured to: divide the entire wind direction interval into multiple sub-intervals according to the power generation efficiency of the wind farm under each wind direction; The optimization sub-problem definition module is configured to: define a wind farm power generation efficiency optimization sub-problem for each divided wind direction sub-interval; The power optimization module is configured to obtain a data-driven wind farm power optimization solution based on the defined power generation efficiency optimization sub-problem, BFGS and knowledge generalization mechanism; when the optimization solution is determined, in the action decision stage, the optimization algorithm corresponding to the sub-interval decides the action for the current wind direction: if the logical variable is true, the action update is completed before the action is generated. When the action is updated, the estimated Newton step path is selected as the search direction, and the projection operator is used to ensure that the decided action meets the control constraints of all wind turbines. After the action update is completed, the perturbation parameter and perturbation vector are used to generate the perturbation action; if the logical variable is not true, the action update is not performed. Perturbation actions are directly generated; in the action evaluation stage, the wind farm executes the action and collects the wind farm's power generation data and wind condition data to obtain the wind farm's power generation efficiency; in the strategy update stage, if the wind direction data at two consecutive sampling moments belong to the same subinterval, the implemented action is effectively evaluated in the actual wind farm: if the logical variable is true, the power generation efficiency is retained and the logical variable is set to false; if the logical variable is not true, the gradient information of the current iteration point is estimated based on the synchronous perturbation random approximation, and the inverse Hessian matrix is ​​estimated with the help of the quasi-Newton method BFGS. Based on the knowledge generalization mechanism, the learned action knowledge is generalized and fully utilized in real time, and the logical variable is set to true.

[0012] In a third aspect, the present invention further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the wind farm power optimization method based on BFGS and knowledge generalization mechanism data-driven according to the first aspect.

[0013] In a fourth aspect, the present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the processor executes the program, the steps of the wind farm power optimization method based on BFGS and knowledge generalization mechanism data-driven described in the first aspect are implemented.

[0014] In a fifth aspect, the present invention further provides a computer program product, comprising a computer program. When the computer program is executed by a processor, the steps of the wind farm power optimization method based on BFGS and knowledge generalization mechanism data-driven described in the first aspect are implemented.

[0015] Compared with the prior art, the present invention has the following beneficial effects: The present invention first divides the entire wind direction interval into multiple subintervals according to the power generation efficiency of the wind farm under each wind direction; then, for each divided wind direction subinterval, a power generation efficiency optimization subproblem of the wind farm is defined; finally, based on the defined power generation efficiency optimization subproblem, a data-driven wind farm power optimization scheme is obtained based on BFGS and a knowledge generalization mechanism; wherein, the optimization framework of the quasi-Newton method is adopted, the estimated Newton step path is selected as the search direction, and the decision action is guaranteed to meet the control constraints of all wind turbines through a projection operator; after the action update is completed, the perturbation action is generated with the help of perturbation parameters and perturbation vectors, and is executed by the wind farm; according to the measured wind farm power generation data information, the gradient is estimated by synchronous perturbation random approximation and the inverse Hessian matrix is ​​estimated by BFGS, and based on the knowledge generalization mechanism, the learned action knowledge is generalized and fully utilized in real time. The present invention does not rely on the wind farm power generation model, but only uses measured data to achieve the purpose of quickly optimizing the power generation of the entire wind farm, solving the problems of the slow convergence speed of the existing optimization scheme and the difficulty in effectively improving the power generation of the wind farm. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The drawings constituting a part of the specification of this embodiment are used to provide a further understanding of this embodiment. The schematic embodiments and descriptions of this embodiment are used to explain this embodiment and do not constitute an improper limitation on this embodiment.

[0017] Figure 1 This is a schematic diagram of the method of Example 1 of the present invention; Figure 2 Flowchart of the D_BFGS strategy of Example 1 of the present invention; Figure 3 This is the wind farm layout of Example 1 of the present invention; Figure 4 This is the simple wind condition of Example 1 of the present invention; Figure 5 The power generation efficiency trajectory of the wind farm in Example 1 of the present invention; Figure 6 is the average power generation efficiency trajectory of the wind farm in Example 1 of the present invention; Figure 7 is the wind direction in Example 1 of the present invention; Figure 8 The wind farm of embodiment 1 of the present invention is The power generation efficiency trajectory when Figure 9 The wind farm of embodiment 1 of the present invention is The power generation efficiency trajectory when Figure 10 The wind farm of embodiment 1 of the present invention is The power generation efficiency trajectory. DETAILED DESCRIPTION

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0019] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present application belongs.

[0020] Indicates the joint control action of all fans, Indicates fan The control action of the axial induction factor is called Axial Induction Factor (AIF).

[0021] yes The feasible domain of and Respectively The lower and upper bounds of .

[0022] express feasible domain.

[0023] Indicates the incoming wind direction of the wind farm; Represents the actual power generation efficiency model of the wind farm.

[0024] Example 1: Wind power essentially converts wind energy into electrical energy, and this energy conversion results in a reduction in wind energy. As a result, while wind turbines extract energy from the wind, they also create areas of reduced wind speed downstream. This phenomenon is known as the wake effect, with the reduced wind speed being called a wake, and the area of ​​reduced wind speed being called a wake zone. If the wake zone formed by an upstream wind turbine overlaps with the rotor rotation plane of a downstream wind turbine, the wind speed reaching the downstream wind turbine will be lower than that reaching the upstream turbine, resulting in a reduction in the downstream wind turbine's output power compared to the upstream turbine. Therefore, the wake effect can lead to aerodynamic coupling (wake coupling) between wind turbines. Furthermore, to reduce installation, operation, and maintenance costs, wind turbines are densely spaced within wind farms, typically with spacing of several hundred meters. This can lead to wake coupling between a single wind turbine and multiple other turbines within a wind farm, ultimately reducing the overall wind farm's power generation. However, practical wind farms generally employ a greedy policy (i.e., maximum power point tracking or single-turbine maximum wind energy capture). Under this strategy, each wind turbine only considers maximizing its own power generation, ignoring the wake coupling between wind turbines, resulting in suboptimal power generation for the entire wind farm. Taking the Lillgrund wind farm as an example, when wind speeds are below rated, the power generation efficiency loss due to wake coupling can exceed 30%. Furthermore, according to statistics, wake coupling causes an annual power generation loss of 20% at the Horns Rev offshore wind farm. This shows that the wake effect severely reduces the power generation efficiency of wind farms. This not only severely damages the economic benefits of wind power, but also prevents the social and environmental benefits of wind power as a clean energy source from being fully realized. For this reason, the problem of optimizing wind farm power generation considering the wake effect has received widespread attention.

[0025] Wakes are time-, space-, and parameter-dependent, and are also affected by topography. This makes wake characteristics extremely complex for mountainous wind farms, making it difficult to establish analytical wake models and wind farm power models, thus limiting the use of model-driven power optimization methods.

[0026] Data-driven power optimization methods do not rely on a wind farm power model. Instead, they determine the optimal wind turbine control inputs solely through control inputs and measurement data. These methods are suitable for wind farm power optimization problems that are difficult to model analytically. These methods primarily include those based on safe experimental dynamics (SED), random search, extreme value search, and zero-order feedback optimization. These methods converge slowly and are either only adaptable to fixed wind conditions or to slowly varying wind conditions. However, wind conditions in real wind farms are highly complex and random, and the wake coupling between wind turbines is strongly dependent on wind conditions. This results in rapidly time-varying optimal solutions to wind farm power optimization problems. Therefore, it is necessary to research data-driven power optimization algorithms that converge quickly to effectively improve wind farm performance.

[0027] To address the problem of optimizing wind farm power generation that is difficult to analytically model, this embodiment provides a data-driven wind farm power optimization method based on BFGS and a knowledge generalization mechanism. This method does not rely on a wind farm power generation model and only uses measured data to achieve the goal of rapidly optimizing the power generation of the entire wind farm. This addresses the limitation of existing optimization schemes, which have slow convergence speeds and are unable to effectively improve the power generation of wind farms (which are difficult to analytically model).

[0028] The method in this embodiment employs a quasi-Newton optimization framework. Based on measured wind farm power generation data, it uses gradients estimated using synchronous perturbation stochastic approximation and the inverse Hessian matrix estimated using BFGS to update all wind turbine control actions. Furthermore, the proposed scheme generalizes the learned actions in real time, further accelerating its convergence. The details are detailed below: The power generation efficiency optimization problem of a wind farm is defined as follows: ; Due to the complex terrain (such as mountains), the characteristics of the wind turbine wake are highly complex, and even a simplified wake model is difficult to establish, and thus the wind farm power generation efficiency analytical model is difficult to establish. Therefore, the power optimization problem can only be solved by data-driven optimization methods to obtain the optimal joint control action. , maximizing the power output of the wind farm.

[0029] In actual wind farms, the widespread application of the greedy strategy makes the historical power generation data under this strategy easily available. Therefore, the power generation efficiency of the real wind farm under all wind directions is According to the obtained power generation efficiency data, the entire wind direction interval is divided into Subinterval, that is . Indicates the subintervals, where and Respectively The lower and upper bounds of , is the set of all subinterval subscripts, that is For any ,Require: ; in, is a small positive constant. The power generation efficiency of wind farms should be guaranteed According to wind direction In the sub-interval Only negligible changes occur within Therefore, for each divided wind direction sub-interval, the wake coupling strength between wind turbines at different wind directions can be approximately considered to be the same, and it can be assumed that only one wake coupling mode exists.

[0030] Therefore, for the wind direction sub-interval , define the following wind farm power generation efficiency optimization sub-problem: ; in, Indicates wind direction The power generation efficiency model of the wind farm is .because When , the wake coupling pattern between wind turbines remains almost unchanged, so the subproblem can be regarded as a static optimization problem, and the corresponding optimal solution can be regarded as a constant vector.

[0031] Different sub-problems have different wake coupling patterns, resulting in different optimal solutions, and each sub-problem can be regarded as a power generation efficiency optimization problem of a wind farm under fixed wind conditions.

[0032] In summary, the optimization problem of wind farm power generation efficiency under time-varying wind direction can be regarded as the continuous switching of wind farm power generation efficiency optimization sub-problems defined in each wind direction sub-interval.

[0033] In order to solve the problem of optimizing the power generation efficiency of wind farms under time-varying wind conditions, such as Figure 2 As shown in the figure, the following wind farm fast power optimization scheme based on BFGS and knowledge generalization mechanism data driven is proposed, referred to as Strategy: S1. Initialization: definition, , For true, , , , , , yes dimensional unit matrix, , , , , , , , , , , , .

[0034] for ,have: S2, action decision: if , ,and If true, the estimated Newton step size is selected. As search direction:

[0035] in, At the iteration point The inverse Hessian matrix estimated at The projection operator is used to ensure that the decision action satisfies the control constraints of all wind turbines: ; in, It's about The projection operator, is the step size. The generated perturbation action is: ; ; in, is the perturbation parameter, is a random perturbation vector.

[0036] otherwise( , ,and Not true), the generated perturbation action is: ; .

[0037] S3. Action evaluation: Wind farm execution action , collect wind farm power generation data and wind condition data (including wind direction ), and calculate the wind farm Power generation efficiency .

[0038] S4. Policy update: if , , ,and is true, then , is false.

[0039] otherwise( , , ,and is false), , the gradient is estimated as: , , in, is the random perturbation vector No. elements, ; The estimate of the inverse of the Hessian matrix is: let , , ,if ,but , otherwise, ; Knowledge generalization is: for any ,if , then determine the neighbor class: , Determine the iteration phase: , if ( No. OK column elements), then , in, yes No. OK Column elements; ; is true.

[0040] In the initialization step, is the initial measured wind direction; It is a logical variable and has only two values, true and false; Representation Algorithm The number of iterations; It's an algorithm The initial iteration action of The estimated wind direction falls within the subinterval When the wind farm power generation efficiency model is in the initial iteration action The gradient at is the inverse of the initialized Hessian matrix; is a constant greater than 0; 、 and The algorithms are Learning rate (or step size), perturbation parameters, random perturbation vector; is a constant less than 0; Indicates the maximum number of iterations allowed when the algorithm shares knowledge (current solution); It represents the maximum number of iterations allowed when the algorithm accepts shared knowledge. The entire iterative process of accepting knowledge is divided into three stages. express; and and wind direction subintervals respectively The numbered set of wind direction subintervals that are very close and relatively close; that is, for wind direction subintervals There are two types of neighbors defined: ,but (near neighbor); for ,but (Distant Neighbor). is a matrix for judging whether knowledge generalization is sufficient, where The interval for measuring neighbor wind direction is located in neighbor classes, the corresponding algorithm is in the At the iteration stage, the minimum number of iterations required to accept knowledge is, is the weight matrix, where When generalized knowledge is expressed, the neighbor wind direction interval is located at neighbor classes, the corresponding algorithm is in the The proportional coefficient at the iteration stage.

[0041] Proposed The strategy consists of three steps: action decision, action evaluation, and strategy update. In the action decision part, there are mainly two steps: action update and action generation. , then for the current wind direction , by the subinterval The corresponding optimization algorithm is used to decide the action. If is true, the action update must be completed before generating the action. When updating the action, the estimated Newton step path is selected as the search direction, and the projection operator is used to ensure that the decision action meets the control constraints of all wind turbines. After the action update is completed, the perturbation parameter , the perturbation vector , generates a perturbation action, which is executed by the wind farm. If it is not true, no action update is required and the perturbation action is generated directly and evaluated by the wind farm. In the action evaluation phase, the power generation data and wind condition data of the wind farm are collected and the wind farm’s Power generation efficiency .

[0042] In the strategy update phase, if the wind direction data at two consecutive sampling moments and Belong to the same wind direction sub-interval, the action to be implemented It has been effectively evaluated in actual wind farms. If true, Retain and order is false; otherwise, the gradient information of the current iteration point is estimated based on the synchronous perturbation random approximation, and the inverse Hessian matrix is ​​estimated with the help of the quasi-Newton method BFGS, laying the foundation for the next action update.

[0043] This embodiment uses the solution obtained by the algorithm corresponding to the current wind direction interval to guide the action update of the optimization algorithm corresponding to the adjacent wind direction interval, so as to further utilize the learned knowledge and accelerate the convergence speed of the algorithm. Although the wind turbine wake coupling modes corresponding to adjacent wind direction intervals are inconsistent, the front and rear positions between the wind turbines in the two modes are similar, which makes the wake coupling modes also similar. This provides an important basis for knowledge generalization. In the knowledge generalization link, the optimization algorithms corresponding to the two types of neighbors in the wind direction sub-interval have different degrees of utilization of generalized knowledge, among which the algorithms corresponding to the close neighbors will pay more attention to the utilization of generalized knowledge; at the same time, the iterative process of the neighbor corresponding optimization algorithm accepting knowledge is divided into three stages, and the earlier the stage, the more the optimization algorithm utilizes the generalized knowledge. The above two ideas of knowledge utilization are expressed through the weight matrix In the knowledge generalization stage, first ensure that the number of iterations of the algorithm corresponding to the current wind direction interval is less than This is because when the number of iterations is large, the solution obtained gets closer and closer to the optimal solution, and its reference value for solving the optimization problem corresponding to the adjacent interval also decreases. Next, for the neighbors of the current wind direction interval, it is necessary to determine whether the number of iterations of the neighbor's corresponding optimization algorithm is sufficient; if it is sufficient, generalized knowledge will no longer be accepted. Otherwise, the neighbor class is determined and the iteration stage of the neighbor's corresponding algorithm is determined. Finally, if the number of iterations of the algorithm corresponding to the current wind direction interval is greater than the number of iterations of the algorithm corresponding to its neighbor by a certain constant value, it means that the knowledge is sufficient and generalization of knowledge can be completed.

[0044] This embodiment adopts a quasi-Newton optimization framework. Based on measured wind farm power generation data, it uses the gradient estimated by synchronous perturbation stochastic approximation and the inverse Hessian matrix estimated by BFGS to update the control actions of all wind turbines. Furthermore, the proposed scheme generalizes the learned actions in real time, further accelerating its convergence. Simulation tests show that compared to the baseline strategy, the proposed method can rapidly optimize wind farm power generation based on measured data in a model-free scenario. To verify the effectiveness and feasibility of the D_BFGS strategy proposed in this embodiment, two simulation test cases were set up.

[0045] Assuming the fan diameter is 126 meters and the air density is for , entrance wind speed for .like Figure 3 As shown in Figure 2, the performance of the proposed scheme is tested on a wind farm with 25 wind turbines. The spacing between adjacent wind turbine pairs is 560 meters. By simulating the exact wind farm power model (actually unknown) using the FLORIS model and applying the greedy strategy to the model, the power generation data of the wind farm under all wind directions can be obtained. It is assumed that this data is historical power generation data from a real wind farm. Based on this data, the power generation efficiency of the wind farm under all wind directions can be calculated. Set the constant is 0.02, the entire wind direction interval can be divided into 203 sub-intervals, namely .

[0046] propose The parameters of the strategy are set to (Greedy Strategy), , , , Randomly generated by symmetric Bernoulli distribution. For any wind direction subinterval , define the close neighbors as the two wind direction subintervals adjacent to it, and the far neighbors as the ones adjacent to its close neighbors, except Wind direction sub-intervals other than itself, For example, for the wind direction subinterval ,but , ;for ,but , , , , , , .

[0047] To illustrate the advantages of the proposed strategy, the following optimization schemes were selected for simulation comparison: ① a greedy strategy; ② an optimal strategy; the strategy obtained using a gradient method based on the FLORIS model is referred to as the optimal strategy. For real wind farms, the optimal strategy is unknown due to the difficulty in obtaining an accurate power generation model. ③ a randomized projected simplex (SPS) strategy. This data-driven centralized optimization scheme demonstrates superior wind farm performance compared to the data-driven safety experimental dynamics (SED) strategy.

[0048] Comprehensive performance test: The simulated wind conditions are as follows Figure 4 As shown. Wind direction in angle set Table 1 shows the percentage of wind farm power generation efficiency improved by the proposed D_BFGS strategy. Figure 5 and 6 The trajectories of wind farm power generation efficiency and average power generation efficiency under different control strategies are shown respectively.

[0049]

[0050] like Figure 5 As shown in the figure, under a given number of iterations, the proposed D_BFGS strategy makes the wind farm power generation efficiency and The wind directions all converged to the optimal value. However, the data-driven SPS strategy failed to optimize the power generation efficiency of the wind farm. This shows that the proposed D_BFGS strategy has a faster convergence speed than the SPS strategy. In addition, Figure 5 It shows that the proposed strategy can adapt to time-varying wind direction. As shown in Table 1, compared with the greedy strategy and SPS strategy, the proposed D_BFGS strategy significantly improves the power generation efficiency of the wind farm under both wind directions. Figure 6 It shows that the proposed strategy exhibits higher average power generation efficiency and thus generates more power generation. In summary, the proposed D_BFGS strategy has superior power generation performance.

[0051] Knowledge generalization ability test: In this case, in order to test the knowledge generalization ability of the proposed strategy, assume that the wind direction starts from 0 degrees. For every 400 increase, the wind direction increases by one degree, e.g. Figure 7The simulation test results are shown in Figures 8 to 10 As shown in Figure 2, the D_BFGS_N strategy indicates that the proposed D_BFGS strategy does not use the knowledge generalization function.

[0052] like Figure 8 、 Figure 9 and Figure 10 As shown, the proposed D_BFGS strategy converges faster than the SPS strategy and exhibits more stable power generation performance, avoiding oscillations in wind farm power generation. Furthermore, compared to the greedy strategy, the proposed strategy significantly improves wind farm power generation under many wind directions. Furthermore, it can be noted that the proposed D_BFGS strategy exhibits faster convergence (compared to the D_BFGS_N strategy) due to the addition of the knowledge generalization function. This confirms the effectiveness of the knowledge generalization function.

[0053] Example 2: This embodiment provides a wind farm power optimization system driven by data based on BFGS and a knowledge generalization mechanism, including: The partitioning module is configured to: divide the entire wind direction interval into multiple sub-intervals according to the power generation efficiency of the wind farm under each wind direction; The optimization sub-problem definition module is configured to: define a wind farm power generation efficiency optimization sub-problem for each divided wind direction sub-interval; The power optimization module is configured to obtain a data-driven wind farm power optimization solution based on the defined power generation efficiency optimization sub-problem, BFGS and knowledge generalization mechanism; when the optimization solution is determined, in the action decision stage, the optimization algorithm corresponding to the sub-interval decides the action for the current wind direction: if the logical variable is true, the action update is completed before the action is generated. When the action is updated, the estimated Newton step path is selected as the search direction, and the projection operator is used to ensure that the decided action meets the control constraints of all wind turbines. After the action update is completed, the perturbation parameter and perturbation vector are used to generate the perturbation action; if the logical variable is not true, the action update is not performed. Perturbation actions are directly generated; in the action evaluation stage, the wind farm executes the action and collects the wind farm's power generation data and wind condition data to obtain the wind farm's power generation efficiency; in the strategy update stage, if the wind direction data at two consecutive sampling moments belong to the same subinterval, the implemented action is effectively evaluated in the actual wind farm: if the logical variable is true, the power generation efficiency is retained and the logical variable is set to false; if the logical variable is not true, the gradient information of the current iteration point is estimated based on the synchronous perturbation random approximation, and the inverse Hessian matrix is ​​estimated with the help of the quasi-Newton method BFGS. Based on the knowledge generalization mechanism, the learned action knowledge is generalized and fully utilized in real time, and the logical variable is set to true.

[0054] The working method of the system is the same as the wind farm power optimization method based on BFGS and knowledge generalization mechanism data driven in Example 1, and will not be described in detail here.

[0055] Example 3: This embodiment provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the steps of the wind farm power optimization method based on BFGS and knowledge generalization mechanism data-driven described in Example 1 are implemented.

[0056] Example 4: This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the processor executes the program, the steps of the wind farm power optimization method based on BFGS and knowledge generalization mechanism data-driven described in Example 1 are implemented.

[0057] Example 5: This embodiment provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the steps of the wind farm power optimization method based on BFGS and knowledge generalization mechanism data drive described in Example 1 are implemented.

[0058] The above description is merely a preferred embodiment of this embodiment and is not intended to limit this embodiment. Those skilled in the art will readily appreciate that this embodiment may be modified and varied in various ways. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this embodiment shall be within the scope of protection of this embodiment.

Claims

1. A data-driven wind farm power optimization method based on BFGS and knowledge generalization mechanism, characterized by: include: According to the power generation efficiency of the wind farm under different wind directions, the entire wind direction interval is divided into multiple sub-intervals; For each divided subinterval, define the wind farm power generation efficiency optimization subproblem; According to the defined power generation efficiency optimization sub-problem, based on BFGS and knowledge generalization mechanism, a data-driven wind farm power optimization scheme is obtained; wherein, when the optimization scheme is determined, in the action decision stage, the optimization algorithm corresponding to the sub-interval decides the action for the current wind direction: if the logical variable is true, then the action update is completed before the action is generated. When the action is updated, the estimated Newton step path is selected as the search direction, and the projection operator is used to ensure that the decision action meets the control constraints of all wind turbines. After the action update is completed, the perturbation action is generated with the help of the perturbation parameter and perturbation vector; if the logical variable is not true, the action update is not performed, and the perturbation is generated directly. action; in the action evaluation phase, the wind farm executes the action and collects the power generation data and wind condition data of the wind farm to obtain the power generation efficiency of the wind farm; in the strategy update phase, if the wind direction data at two consecutive sampling moments belong to the same subinterval, the implemented action is effectively evaluated in the actual wind farm: if the logical variable is true, the power generation efficiency is retained and the logical variable is set to false; if the logical variable is not true, the gradient information of the current iteration point is estimated based on the synchronous perturbation random approximation, and the inverse Hessian matrix is ​​estimated with the help of the quasi-Newton method BFGS. Based on the knowledge generalization mechanism, the learned action knowledge is generalized and fully utilized in real time, and the logical variable is set to true.

2. The wind farm power optimization method based on BFGS and knowledge generalization mechanism data driven according to claim 1, characterized in that: In each subinterval, the wake coupling strength between wind turbines under different wind directions is considered to be the same. It is assumed that only one wake coupling mode exists, and the subproblem is regarded as a static optimization problem, and the corresponding optimal solution is regarded as a constant vector.

3. The wind farm power optimization method based on BFGS and knowledge generalization mechanism data-driven according to claim 1, characterized in that: Select the estimated Newton step size as the search direction: ； in, is the estimated Newton step, is the number of iterations; At the iteration point The inverse Hessian matrix estimated at ; is the estimated gradient.

4. The wind farm power optimization method based on BFGS and knowledge generalization mechanism data drive according to claim 1, characterized in that: The projection operator is used to ensure that the decision action satisfies the control constraints of all wind turbines: ； in, It's about The projection operator, is the step length; Indicates the number of iterations.

5. The wind farm power optimization method based on BFGS and knowledge generalization mechanism data drive according to claim 1, characterized in that: If the wind direction data at two consecutive moments belong to the same subinterval, the implemented action is effectively evaluated in the actual wind farm: if the logical variable is true, the power generation efficiency is retained and the logical variable is set to false; On the contrary, the gradient information of the current iteration point is estimated based on the synchronous perturbation random approximation, and the inverse Hessian matrix is ​​estimated with the help of the quasi-Newton method BFGS.

6. The wind farm power optimization method based on BFGS and knowledge generalization mechanism data drive according to claim 1, characterized in that: If the wind direction data at two consecutive moments belong to the same subinterval, based on the knowledge generalization mechanism, the learned action knowledge is generalized and fully utilized in real time, and the logical variable is set to true.

7. A data-driven wind farm power optimization system based on BFGS and knowledge generalization mechanism is characterized by: include: The partitioning module is configured to: divide the entire wind direction interval into multiple sub-intervals according to the power generation efficiency of the wind farm under each wind direction; The optimization sub-problem definition module is configured to: define a wind farm power generation efficiency optimization sub-problem for each divided wind direction sub-interval; The power optimization module is configured to obtain a data-driven wind farm power optimization solution based on the defined power generation efficiency optimization sub-problem, BFGS and knowledge generalization mechanism; when the optimization solution is determined, in the action decision stage, the optimization algorithm corresponding to the sub-interval decides the action for the current wind direction: if the logical variable is true, the action update is completed before the action is generated. When the action is updated, the estimated Newton step path is selected as the search direction, and the projection operator is used to ensure that the decided action meets the control constraints of all wind turbines. After the action update is completed, the perturbation parameter and perturbation vector are used to generate the perturbation action; if the logical variable is not true, the action update is not performed. Perturbation actions are directly generated; in the action evaluation stage, the wind farm executes the action and collects the wind farm's power generation data and wind condition data to obtain the wind farm's power generation efficiency; in the strategy update stage, if the wind direction data at two consecutive sampling moments belong to the same subinterval, the implemented action is effectively evaluated in the actual wind farm: if the logical variable is true, the power generation efficiency is retained and the logical variable is set to false; if the logical variable is not true, the gradient information of the current iteration point is estimated based on the synchronous perturbation random approximation, and the inverse Hessian matrix is ​​estimated with the help of the quasi-Newton method BFGS. Based on the knowledge generalization mechanism, the learned action knowledge is generalized and fully utilized in real time, and the logical variable is set to true.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the wind farm power optimization method based on BFGS and knowledge generalization mechanism data-driven are implemented as described in any one of claims 1 to 6.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that: When the processor executes the program, the steps of the wind farm power optimization method based on BFGS and knowledge generalization mechanism are implemented as described in any one of claims 1 to 6.

10. A computer program product, characterized in that The computer program product includes a computer program, and when the computer program is executed by a processor, the steps of the wind farm power optimization method based on BFGS and knowledge generalization mechanism data drive are implemented as described in any one of claims 1 to 6.

Images