Reversible neural network model assisted high-dimensional Bayesian wind field layout optimization method

Abstract

The invention belongs to the field of wind power, and discloses a reversible neural network model-assisted high-dimensional Bayesian wind field layout optimization method, which comprises the following steps of: 1, modeling a scene to be laid out and generating a feasible layout scene; 2, designing a reversible neural network structure and a loss function of high-dimensional layout information compression; and step 3, optimizing the high-dimensional Bayesian wind field layout assisted by the reversible neural network. According to the method, the problem that the traditional method either adopts an analytical wake flow model to sacrifice precision or adopts a heuristic algorithm to sacrifice efficiency is solved, an efficient framework based on Bayesian optimization is provided, a high-fidelity numerical wake flow model is adopted, and both optimization efficiency and result precision are improved.

Description

Technical Field

[0001] This invention relates to the field of wind power, specifically to a high-dimensional Bayesian wind farm layout optimization method assisted by a reversible neural network model. Background Technology

[0002] Wind energy plays a vital role in the global energy transition as a clean and renewable energy source. With the continuous expansion of wind farms, wake effects inevitably occur between wind turbines during actual operation. Wake effects refer to the reduction in airflow velocity and enhancement of turbulence caused by upstream turbines during operation. This effect significantly reduces the effective incoming wind speed of downstream turbines, increases load fluctuations, and consequently affects the power output and equipment lifespan of the entire wind farm. Therefore, optimizing wind farm layout has become an important research and engineering practice topic in order to mitigate the adverse effects of wake effects. By rationally designing the spatial distribution of wind turbines within the farm area, power generation or revenue can be maximized under given wind resources and terrain conditions, while also taking into account various constraints.

[0003] Existing methods and research can be divided into two categories based on the wake models used. One category is based on analytical wake models, such as the Jensen model and the Gaussian model. These models simplify the description of wake velocity decay and turbulent diffusion through physical approximations and empirical formulas, and have advantages such as high computational efficiency, few parameter requirements, and ease of integration into optimization frameworks. Many studies have used these wake models to explicitly or implicitly model the objective function of wind field design, and then use gradient-based algorithms or intelligent heuristic algorithms to solve the optimization problem, obtaining the optimal wind field layout scheme. However, because these wake models are based on strong simplification assumptions, they are difficult to accurately describe the flow field characteristics under non-uniform inflow conditions and multiple wake superpositions, resulting in limited prediction accuracy and potentially leading to significant discrepancies between the final results and actual conditions.

[0004] Another approach uses numerical wake models based on computational fluid dynamics, such as the Reynolds-averaged Navier-Stokes equations and large eddy simulations, to perform high-fidelity modeling of wake interactions. These models can accurately and comprehensively describe the velocity field, turbulence characteristics, and evolution of the wind turbine wake, exhibiting high prediction accuracy. However, numerical models are described by partial differential equations governing fluid behavior, making it extremely difficult to explicitly construct the objective function for layout optimization. Furthermore, gradient-based algorithms cannot be directly used to solve these problems. In addition, while some studies have utilized heuristic algorithms for convergence, the numerical models themselves are computationally expensive, requiring significant computational resources and time for a single simulation. Combining this with computationally intensive heuristic algorithms can lead to extremely long convergence times.

[0005] In summary, analytical models are suitable for large-scale, rapid optimization, but their accuracy is limited; numerical models offer high accuracy, but their computational burden is too heavy. Finding a balance between high-precision modeling and efficient optimization computation has become a significant challenge in wind field layout optimization research and applications. Summary of the Invention

[0006] The purpose of this invention is to provide a high-dimensional Bayesian wind field layout optimization method assisted by a reversible neural network model, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A high-dimensional Bayesian wind field layout optimization method assisted by a reversible neural network model includes:

[0009] Step 1: Model the scene to be laid out and generate feasible layout scenes:

[0010] Step 2: Perform high-dimensional layout information preprocessing and design the structure and loss function of the invertible neural network model;

[0011] Step 3: Perform high-dimensional Bayesian wind field layout optimization assisted by a reversible neural network model, including:

[0012] In the main optimization phase, the reversible neural network model is embedded into the initialization phase in the form of forward propagation. In the initialization phase, several layout schemes are uniformly sampled on the feasible layout dataset using the Latin hypercube sampling technique. These layout scheme representations are simultaneously subjected to forward propagation of the reversible neural network model and high-fidelity computational fluid dynamics simulation to obtain the corresponding low-dimensional representations and corresponding objective function values, thereby constructing an initial low-dimensional representation dataset of input-output pairs. The probabilistic surrogate model in Bayesian optimization is trained on the initial low-dimensional representation dataset of input-output pairs to complete the initialization.

[0013] During the iterative optimization process in the main optimization phase, the reversible neural network model is embedded in the form of backpropagation; the acquisition function is optimized on the probabilistic surrogate model constructed in the low-dimensional space, and the next evaluation point is selected by balancing exploration and utilization; then, this evaluation point is paired with its corresponding objective function value and filled into the low-dimensional representation dataset, the probabilistic surrogate model is updated on this basis, and then the next iteration of optimization is entered until the stopping condition is met and the optimal solution is output.

[0014] Furthermore, the scene modeling in step 1 includes:

[0015] Determine how to represent the location of wind turbines in the wind farm, and use a grid-type coordinate system to characterize the location of the wind turbines;

[0016] In the grid-based representation, the wind field to be deployed is divided into several grids. The side length of each grid depends on the rotor diameter D of the wind turbines to be deployed in the wind field. The set of all grids with a total of M vertices is denoted as . , which represents the available locations for the N wind turbines to be deployed, x i and y i These represent the positions of the vertices along the x-axis and y-axis, respectively, and i is the corresponding fan number. Let N represent the integer space, and each layout is represented as a set of N vertices that satisfy the constraints:

[0017]

[0018] In the formula, Symbols representing layout schemes;

[0019] Each feasible layout must satisfy the minimum distance constraint between wind turbines:

[0020]

[0021] Where, d ij d represents the distance between fan i and fan j. min This refers to the minimum distance between wind turbines determined based on the geographical location of the wind farm to be located and the selected wind turbine model.

[0022] Furthermore, the scene modeling in step 1 also includes:

[0023] Define the objective function for wind farm layout optimization as annual power generation (AEP):

[0024]

[0025] Among them, D year W represents the number of days in a year, and f represents the number of wind conditions considered in the wind field. j This indicates the frequency of the j-th wind condition. This represents the power generation capacity of the i-th wind turbine under the j-th wind condition;

[0026] The wind turbine is modeled using a brake disc model, and the wake effect is calculated using a high-fidelity numerical model RANS. The brake disc model is incorporated into the standard RANS equations as a source term, as shown in the following formula:

[0027]

[0028] Where t is time, x i With x j These are Cartesian coordinates, where the subscripts i and j represent the horizontal and vertical directions, respectively. and Let f represent the time-averaged velocity and pressure, ρ and μ be the air density and dynamic viscosity, respectively. i The source terms are represented by the brake disc; finally, the Reynolds stress terms are closed using the k-ε model. .

[0029] Furthermore, the feasible layout scenario generation in step 1 includes:

[0030] Based on the modeled layout scenario, a backtracking method is used to automatically generate all feasible layouts for that scenario, and all feasible layouts are stored in a dataset. middle.

[0031] Furthermore, the high-dimensional layout information preprocessing in step 2 includes:

[0032] For the layout representation consisting of the positions of N wind turbines to be placed, a unique integer index is assigned to each vertex in the wind field divided by grid. Each layout is represented by recording the index corresponding to the vertex where the wind turbine is placed, thereby reducing the dimension of the layout representation from 2N to N.

[0033] For the feasible layout dataset generated in step 1 Conduct an internal dimension analysis;

[0034] feasible layout dataset Normalization is performed to obtain the normalized feasible layout dataset. .

[0035] Furthermore, the reversible neural network model in step 2 includes several reversible coupling layers;

[0036] A reversible neural network model learns a deterministic nonlinear transformation f θ , will input Mapped to a specific latent space of the same dimension And has an easily obtainable inverse mapping ,get:

[0037]

[0038] in, Indicates backpropagation The output obtained, It represents the space of real numbers.

[0039] Furthermore, in step 2, the reversible coupling layer of the reversible neural network model is an affine coupling layer of RNVP. Each layer of this affine coupling layer contains four sub-networks for learning different nonlinear transformations, and the sub-networks adopt a multilayer perceptron.

[0040] The output vector of the invertible neural network model Perform dimensional partitioning:

[0041]

[0042] in, This indicates a low-dimensional representation of latent variables. It is the intrinsic dimension of the dataset. Represents additional information hidden variables; during training, After concatenating with the all-zero vector, reverse propagation is performed to reconstruct the data.

[0043]

[0044] in, This indicates a reconstruction without additional information. Representative and Vectors of the same shape consisting entirely of zeros.

[0045] Furthermore, the loss function design for the reversible neural network model in step 2 includes:

[0046] The comprehensive loss function L consists of three different loss terms:

[0047]

[0048] In the formula, For standard reconstruction loss, This represents the reconstruction loss using only the partitioned low-dimensional representation. The additional information retention loss indicates how much information is stored in the additional dimensions that are divided, where ω1, ω2, and ω3 are weighting coefficients;

[0049] in, The calculation method is as follows:

[0050]

[0051] In the formula, To find the 2-norm, Represents the desire for expectation;

[0052] in, The calculation method is as follows:

[0053]

[0054] in, The calculation method is as follows:

[0055] .

[0056] Furthermore, step 2 also includes:

[0057] In the preprocessing stage, the reversible neural network model is pre-trained on all feasible layout schemes to achieve compression and accurate reconstruction of high-dimensional layout information. After the training error converges, the reversible neural network model learns a deterministic bijective mapping, which includes injectivity and surjectivity. At this point, the reversible neural network model is saved.

[0058] Compared with existing technologies, the advantages are as follows: This invention solves the problems of traditional methods either using analytical wake models to evaluate wind field wakes, resulting in insufficient accuracy, or using heuristic algorithms to perform numerous objective function evaluations, sacrificing efficiency. It proposes an efficient framework based on Bayesian optimization, fully utilizing computationally expensive high-fidelity numerical wake model simulation results, thus balancing optimization efficiency with improved accuracy. Furthermore, it introduces a pre-trained invertible neural network model for the first time to address the dimensionality curse problem of Bayesian optimization in high-dimensional spaces, providing a new approach for extending the algorithm's application to other fields. Attached Figure Description

[0059] Figure 1 A complete flowchart of a high-dimensional Bayesian wind field layout optimization method assisted by a reversible neural network model.

[0060] Figure 2 This is a schematic diagram of a reversible neural network model. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] Please see Figure 1 A high-dimensional Bayesian wind field layout optimization method assisted by a reversible neural network model includes:

[0063] Step 1, perform scene modeling and feasible layout scene generation, including:

[0064] Step 1.1: Model the scene to be laid out.

[0065] In this step, we mainly model the wind farm scenario to be laid out. First, we determine how to represent the positions of the wind turbines in the wind farm. The existing representation methods are mainly Cartesian coordinates (continuous) and grid coordinates (discrete). Since existing research has shown that the grid representation method does not hinder the acquisition of the optimal solution for layout optimization, and it is naturally convenient for handling the distance constraints between wind turbines, this example uses grid coordinates to represent the positions of the wind turbines.

[0066] In a grid-based representation, the wind field to be deployed is first divided into several grids. The side length of each grid typically depends on the rotor diameter D of the wind turbines to be deployed in the wind field. Therefore, the set of all grids with a total of M vertices can be denoted as... , which represents the available locations for the N wind turbines to be deployed. i and y i Both are integers, representing the order of the vertex along the x-axis and y-axis, respectively. Representing the integer space. The representation of each layout, that is, the decision variables of the wind farm layout optimization problem, can be written as a set of N vertices satisfying the constraints:

[0067]

[0068] In the formula, Symbols representing layout schemes.

[0069] Specifically, each feasible layout must satisfy the minimum distance constraint between wind turbines:

[0070]

[0071] Where, d ij d represents the distance between fan i and fan j. min The minimum distance between wind turbines is determined based on the geographical location of the wind farm to be laid out and the selected wind turbine model, and is usually between 2D and 7D.

[0072] Finally, an objective function for wind farm layout optimization needs to be defined. This example uses Annual Energy Production (AEP):

[0073]

[0074] Among them, D year This represents the number of days in a year, either 365 or 366. W and N represent the number of wind conditions and the number of wind turbines considered in the wind farm, respectively. j This indicates the frequency of the j-th wind condition. This represents the power generation capacity of the i-th wind turbine under the j-th wind condition.

[0075] The power generation P of the wind turbine is determined by the layout scheme. The objective function is a function of wind turbines, and different layouts will result in drastically different power generation for each turbine in the wind farm. To accurately evaluate the fluid behavior in the wind farm under different layout schemes and obtain a more realistic objective function, this invention uses a brake disc model to model the wind turbine and combines it with a high-fidelity numerical model—Reynolds-averaged Navier-Stokes (RANS)—to calculate the wake effect. The brake disc model is added to the standard RANS equations as source terms, as shown in the following formula:

[0076]

[0077] Where t is time, x i With x j These are Cartesian coordinates, where the subscripts i and j represent the horizontal and vertical directions, respectively. and Let f represent the time-averaged velocity and pressure, ρ and μ be the air density and dynamic viscosity, respectively. i This represents the source term from the brake disc. Finally, the present invention uses a k-ε model to close the Reynolds stress term. The above equations are discretized in space and time using the second-order numerical form of the finite volume method, and can be solved using open-source or commercial computational fluid dynamics toolkits. The wind turbine blades are modeled based on blade element momentum theory. Since this invention does not involve any theoretical innovation, it will not be elaborated further here.

[0078] Step 1.2: Generate feasible layout scenarios

[0079] Subsequently, based on the layout scenario modeled in step 1.1, a backtracking method is used to automatically generate all feasible layouts for the scenario, and all feasible layouts are stored in a dataset. In particular, when generating feasible layouts, this invention proposes and uses a novel layout representation method suitable for grid-type position representations, which is described in detail in step 2.

[0080] Step 2 involves preprocessing the high-dimensional layout information and designing the structure and loss function of the invertible neural network model, including:

[0081] Step 2.1: Perform high-dimensional layout information preprocessing.

[0082] In this step, we propose and use a novel wind farm layout representation method suitable for grid-based location representation. For the layout representation of N wind turbines to be placed, traditional methods require a 2N-dimensional vector (each turbine requires an x ​​and y coordinate) for representation. This invention assigns a unique integer index to each vertex in the grid-divided wind farm, and each layout is represented by recording the index corresponding to the vertex where the turbine is placed, thereby reducing the layout representation dimension from 2N to N.

[0083] Furthermore, to confirm the target dimension for information compression in the subsequent reversible neural network, a manifold learning algorithm, such as local linear embedding, needs to be used on the feasible layout dataset generated in step 1. Conduct an internal dimension analysis. Finally, Normalization is performed to obtain the normalized feasible layout dataset. Prepare to input a specially designed reversible neural network for dimensionality reduction.

[0084] Step 2.2: Design the reversible neural network structure and loss function.

[0085] Reversible neural networks are a special type of neural network, initially proposed for generative models and probability density estimation. A reversible neural network model consists of several reversibly coupled layers that learn a deterministic nonlinear transformation f. θ , will input Mapped to a specific latent space of the same dimension And due to its structural characteristics, it has an easily obtainable inverse mapping (i.e. Therefore:

[0086]

[0087] in, Indicates backpropagation The output obtained, It represents the space of real numbers.

[0088] This invention proposes a specially designed reversible neural network, characterized by latent space dimension segmentation and loss function design, achieving compression and accurate reconstruction of layout information. Reversible neural network models are typically implemented by stacking reversibly coupled layers, such as... Figure 2 As shown on the far left. Reversible coupling layers come in various forms. This invention selects an improved version of an affine coupling layer called RNVP, which possesses strong nonlinear characterization capabilities and appropriate flexibility. Its structure is as follows... Figure 1 As shown in the middle, each affine coupling layer contains four sub-networks for learning different nonlinear transformations. The structure of the sub-networks can be arbitrarily complex without compromising the reversibility of the reversible neural network. The sub-networks in this invention employ a multilayer perceptron, such as... Figure 2 The rightmost one in the middle.

[0089] Specifically, in this invention, we consider the output vector of the invertible neural network. Perform dimensional partitioning:

[0090]

[0091] in, These represent low-dimensional latent variables, used to store useful information in the layout representation, i.e., the low-dimensional representation of the layout. It is the number of intrinsic dimensions of the dataset analyzed by the manifold learning algorithm in step 2.1. Latent variables represent additional information, used to store redundant information (or not any useful information). This dimensionality splitting operation facilitates the design of the penalty loss term during training and subsequent optimization in the low-dimensional representation space. Furthermore, it allows for the testing of information in the latent space... Does it encompass all the valid information in the layout representation? During training, we concatenate it with an all-zero vector and then reconstruct it through backpropagation.

[0092]

[0093] in, This indicates a reconstruction without additional information. Representative and Vectors of all zeros with the same shape. Based on this, for the layout information compression and accurate reconstruction tasks required by this invention, we propose and design a comprehensive loss function consisting of three different loss terms. The first loss term is the standard reconstruction loss. :

[0094]

[0095] In the formula, To find the 2-norm, It represents the desire for something.

[0096] It is worth noting that, although the reversibility of reversible neural networks theoretically guarantees... However, due to the limitations of computer floating-point precision during model training, this loss term will never be zero. Adding this loss term can control the model to remain stable during gradient descent, and is used to monitor the numerical stability of model parameters during training and detect floating-point errors. The increased loss value can serve as an indicator of anomalies such as gradient explosion.

[0097] The second loss term is designed as a reconstruction loss that uses only the partitioned low-dimensional representation for reconstruction. The calculation method is as follows:

[0098]

[0099] As the main source of loss and the objective of this model, the smaller the value of this loss term, the more accurate and complete the information compression of the layout representation. When the loss term is 0, it means that the model has learned a definite invertible mapping, making the latent space... In It can fully represent the original high-dimensional layout information.

[0100] The last loss term is designed to penalize the latent space. The amplitude of these dimensions is intended to gradually approach 0 (i.e., contain no information) through training, and is defined as:

[0101]

[0102] In the formula, The residual loss for additional information indicates how much information is stored in the additional (redundant) dimensions; this design aims to assist... Together, they guide the model parameters to update in the desired direction, compressing all effective information in the layout representation into a specified dimension in the latent space. Ablation experiments in real-world cases also demonstrate the necessity of this loss term. In summary, the total loss function L of this invertible neural network model can be expressed as a weighted sum of the various loss terms:

[0103]

[0104] Wherein ω1, ω2 and ω3 are the weight coefficients of each loss term, which are set to 0.1, 1 and 0.5 respectively in this invention according to empirical method.

[0105] In the preprocessing stage, an invertible neural network with a specially designed model structure and loss function is first pre-trained on all feasible layout schemes to achieve compression and accurate reconstruction of high-dimensional layout information. After the training error converges, the model learns a deterministic bijective (satisfying both injectivity and surjectivity) mapping, at which point the invertible neural network model is saved.

[0106] Step 3: Perform reversible neural network-assisted high-dimensional Bayesian wind field layout optimization, including:

[0107] As described in step 1, this invention employs a high-fidelity numerical model based on fluid dynamics to solve for the annual power generation of wind farms under different layouts. However, due to the high computational cost of the numerical model and the difficulty in explicitly modeling the objective function caused by its high fidelity, this problem becomes an expensive black-box optimization problem. Bayesian optimization is a powerful sequential optimization method, well-suited for solving black-box optimization problems, but its performance significantly decreases with increasing dimensionality of decision variables (also known as the "curse of dimensionality"). Therefore, to address this issue, this invention proposes for the first time and utilizes a theoretically bijective invertible neural network embedded in Bayesian optimization to assist in solving the curse of dimensionality problem caused by high-dimensional layout spaces. The overall flowchart of this framework is shown below. Figure 2 As shown.

[0108] In the main optimization phase, the reversible neural network model is first embedded into the initialization phase via forward propagation. In the initialization phase, Latin hypercube sampling is first applied to the feasible layout dataset. Several layout schemes are uniformly sampled. These layout schemes are simultaneously processed through 1. forward propagation of a reversible neural network to obtain corresponding low-dimensional representations, and 2. high-fidelity computational fluid dynamics simulation to obtain corresponding objective function values, thereby constructing an initial input-output pair low-dimensional representation dataset. Subsequently, The probabilistic surrogate model in Bayesian optimization—the Gaussian process—is trained and initialized. In this example, the mean function and kernel function of the Gaussian process are a 0-mean function and a Matérn 5 / 2 kernel, respectively.

[0109] During the iterative optimization process in the main optimization phase, the reversible neural network is embedded in the form of backpropagation. Specifically, the acquisition function is optimized on a probabilistic surrogate model constructed in a low-dimensional space. The next evaluation point is selected by balancing exploration and exploitation. At this point, the evaluation point is a point in the low-dimensional representation space and has no physical meaning. However, through the backpropagation of the reversible neural network, its corresponding layout scheme can be reconstructed, and then its corresponding objective function value can be calculated using high-fidelity simulation. Subsequently, this evaluation point is paired with its corresponding objective function value and populated into the low-dimensional representation dataset. The probabilistic proxy model is then updated based on this, and then enters the next iteration for optimization until the stopping condition is met and the optimal solution is output.

[0110] In summary, the high-dimensional Bayesian wind field layout optimization method assisted by the reversible neural network model proposed in this invention has been successfully implemented. This method introduces a theoretically bijective reversible neural network, pre-trained on feasible layouts using a specially designed model structure and loss function, achieving compression and accurate reconstruction of high-dimensional layout information. Furthermore, by cleverly embedding it into the Bayesian optimization process, it effectively solves the "curse of dimensionality" problem suffered by Bayesian optimization in high-dimensional layout spaces, guiding it to optimize within a low-dimensional representation space containing all feasible layouts. This effectively utilizes the efficient optimization capability of Bayesian optimization in solving expensive black-box optimization problems, achieving a balance between result accuracy and optimization efficiency while using high-fidelity computational fluid dynamics simulations to evaluate the objective functions of different layout schemes.

[0111] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A high-dimensional Bayesian wind field layout optimization method assisted by a reversible neural network model, characterized in that, include: Step 1: Model the scene to be laid out and generate feasible layout scenes: Step 2: Perform high-dimensional layout information preprocessing and design the structure and loss function of the invertible neural network model; Step 3: Perform high-dimensional Bayesian wind field layout optimization assisted by a reversible neural network model, including: In the main optimization phase, the reversible neural network model is embedded into the initialization phase in the form of forward propagation. In the initialization phase, several layout schemes are uniformly sampled on the feasible layout dataset using the Latin hypercube sampling technique. These layout scheme representations are simultaneously subjected to forward propagation of the reversible neural network model and high-fidelity computational fluid dynamics simulation to obtain the corresponding low-dimensional representations and corresponding objective function values, thereby constructing an initial low-dimensional representation dataset of input-output pairs. The probabilistic surrogate model in Bayesian optimization is trained on the initial low-dimensional representation dataset of input-output pairs to complete the initialization. During the iterative optimization process in the main optimization phase, the invertible neural network model is embedded in the form of backpropagation; the acquisition function is optimized on the probabilistic surrogate model constructed in the low-dimensional space, and the next evaluation point is selected by balancing exploration and utilization. Subsequently, this evaluation point is paired with its corresponding objective function value and populated into a low-dimensional representation dataset. The probabilistic proxy model is then updated based on this, and the process proceeds to the next iteration until the stopping condition is met and the optimal solution is output.

2. The high-dimensional Bayesian wind field layout optimization method assisted by a reversible neural network model according to claim 1, characterized in that, The scene modeling in step 1 includes: Determine how to represent the location of wind turbines in the wind farm, and use a grid-type coordinate system to characterize the location of the wind turbines; In a grid-based representation, the wind field to be deployed is divided into several grids. The side length of each grid depends on the rotor diameter D of the wind turbines to be deployed in the wind field. The set of all grids with a total of M vertices is denoted as […]. , which represents the available locations for the N wind turbines to be deployed, x i and y i These represent the positions of the vertices along the x-axis and y-axis, respectively, and i is the corresponding fan number. Let N represent the integer space, and each layout is represented as a set of N vertices that satisfy the constraints: ， In the formula, Symbols representing layout schemes; Each feasible layout must satisfy the minimum distance constraint between wind turbines: ， Where, d ij d represents the distance between fan i and fan j. min This refers to the minimum distance between wind turbines determined based on the geographical location of the wind farm to be located and the selected wind turbine model.

3. The high-dimensional Bayesian wind field layout optimization method assisted by a reversible neural network model according to claim 2, characterized in that, The scene modeling in step 1 also includes: Define the objective function for wind farm layout optimization as annual power generation (AEP): ， Among them, D year W represents the number of days in a year, and f represents the number of wind conditions considered in the wind field. j This indicates the frequency of the j-th wind condition. This represents the power generation capacity of the i-th wind turbine under the j-th wind condition; The wind turbine is modeled using a brake disc model, and the wake effect is calculated using a high-fidelity numerical model RANS. The brake disc model is incorporated into the standard RANS equations as a source term, as shown in the following formula: ， Where t is time, x i With x j These are Cartesian coordinates, where the subscripts i and j represent the horizontal and vertical directions, respectively. and Let f represent the time-averaged velocity and pressure, ρ and μ be the air density and dynamic viscosity, respectively. i The source terms are represented by the brake disc; finally, the Reynolds stress terms are closed using the k-ε model. .

4. The high-dimensional Bayesian wind field layout optimization method assisted by a reversible neural network model according to claim 3, characterized in that, The generation of feasible layout scenarios in step 1 includes: Based on the modeled layout scenario, a backtracking method is used to automatically generate all feasible layouts for that scenario, and all feasible layouts are stored in a dataset. middle.

5. The high-dimensional Bayesian wind field layout optimization method assisted by a reversible neural network model according to claim 1, characterized in that, Step 2, which involves preprocessing high-dimensional layout information, includes: For the layout representation consisting of the positions of N wind turbines to be placed, a unique integer index is assigned to each vertex in the wind field divided by grid. Each layout is represented by recording the index corresponding to the vertex where the wind turbine is placed, thereby reducing the dimension of the layout representation from 2N to N. For the feasible layout dataset generated in step 1 Conduct an internal dimension analysis; feasible layout dataset Normalization is performed to obtain the normalized feasible layout dataset. .

6. The method for optimizing high-dimensional Bayesian wind field layout assisted by a reversible neural network model according to claim 5, characterized in that, The reversible neural network model in step 2 includes several reversible coupling layers; A reversible neural network model learns a deterministic nonlinear transformation f θ , will input Mapped to a specific latent space of the same dimension And has an easily obtainable inverse mapping ,get: ， in, Indicates backpropagation The output obtained, It represents the space of real numbers.

7. The method for optimizing high-dimensional Bayesian wind field layout assisted by a reversible neural network model according to claim 6, characterized in that, In step 2, the reversible coupling layer of the reversible neural network model is the affine coupling layer of RNVP. Each layer of the affine coupling layer contains 4 sub-networks for learning different nonlinear transformations. The sub-networks adopt multilayer perceptrons. The output vector of the invertible neural network model Perform dimensional partitioning: ， in, This indicates a low-dimensional representation of latent variables. It is the intrinsic dimension of the dataset. Represents additional information hidden variables; during training, After concatenating with the all-zero vector, reverse propagation is performed to reconstruct the data. ， in, This indicates a reconstruction without additional information. Representative and Vectors of the same shape containing all zeros.

8. The high-dimensional Bayesian wind field layout optimization method assisted by a reversible neural network model according to claim 6, characterized in that, The loss function design for the reversible neural network model in step 2 includes: The comprehensive loss function L consists of three different loss terms: ， In the formula, For standard reconstruction loss, This represents the reconstruction loss using only the partitioned low-dimensional representation. The additional information retention loss indicates how much information is stored in the additional dimensions that are divided, where ω1, ω2, and ω3 are weighting coefficients; in, The calculation method is as follows: ， In the formula, To find the 2-norm, Represents the desire for expectation; in, The calculation method is as follows: ， in, The calculation method is as follows: 。 9. The high-dimensional Bayesian wind field layout optimization method assisted by a reversible neural network model according to claim 1, characterized in that, Step 2 also includes: In the preprocessing stage, the reversible neural network model is pre-trained on all feasible layout schemes to achieve compression and accurate reconstruction of high-dimensional layout information. After the training error converges, the reversible neural network model learns a deterministic bijective mapping, which includes injectivity and surjectivity. At this point, the reversible neural network model is saved.

Images