A generative optimization design method for wind turbine blades based on generative adversarial networks

Abstract

A generative optimization design method for wind turbine blades based on a generative adversarial network comprises the following steps: using a simulation algorithm to obtain detailed data of existing wind turbine blades to form a training data set; generating a new three-dimensional structure of a wind turbine blade through an encoder and a generative adversarial network; evaluating the power generation efficiency of the newly designed wind turbine blade using a generative adversarial network; optimizing the wind turbine blade design through adversarial training of a generative neural network and a predictive neural network, and ultimately outputting an optimized wind turbine blade design. The present invention solves the problem that it is difficult to simultaneously take into account multiple design objectives in wind turbine blade optimization, such as aerodynamic performance, structural strength, and cost, and that it is easy to fall into a local optimal solution; through multi-objective training, an optimized design that simultaneously meets multiple objectives is generated, and through feedback from a discriminator, the generated design is gradually improved to find the blade shape with the best overall performance.

Description

Technical Field

[0001] The present invention belongs to the field of wind power generation, and in particular relates to a generative optimization design method for wind turbine blades based on a generative adversarial network. Background Art

[0002] The capacity of modern wind turbines has grown steadily, from a few hundred kilowatts in the early days to several megawatts (MW) and even over ten MW today. The development of large-scale turbines has significantly increased the power generation capacity of wind farms. Wind turbine blades, which are increasingly made of new materials and optimized design techniques, are becoming lighter and more efficient. Blade optimization is a critical component of wind power generation technology, directly impacting turbine efficiency and reliability. Advances in computing power and algorithmic technology are also driving the development of blade optimization methods.

[0003] Currently, wind turbine blade optimization mostly relies on CFD for high-precision aerodynamic simulations. This multi-scale coupled simulation, combined with large-scale climate models and small-scale local flow field models, allows for precise predictions of wind resources and blade performance, and allows for evaluation of blade design performance under varying wind conditions. Simulations can optimize blade shape, angles, and structure to improve wind capture efficiency. However, CFD simulations and optimization algorithms require significant computing resources, especially when multi-objective optimization and high-resolution simulations are involved. The computational time and cost are very high. Traditional optimization algorithms, such as genetic algorithms (GAs) and particle swarm optimization (PSOs), can mimic natural selection and genetic mechanisms, finding the optimal blade design through iterative search. However, these algorithms are computationally expensive and time-consuming, and each fitness evaluation may require complex simulations or calculations. In particular, fitness evaluations in wind blade optimization often involve CFD simulations, further increasing computational overhead. With the advancement of artificial intelligence (AI), wind blade optimization methods will continue to evolve towards intelligent, automated, and integrated approaches. AI and machine learning technologies will be more deeply integrated into the optimization process, improving design efficiency and quality. Furthermore, multidisciplinary optimization methods will continue to advance, enabling the global optimal design of wind turbine blades. In terms of computing resources and experimental verification, more advanced computing technologies and experimental methods will be used to ensure the practical feasibility and reliability of the optimization results.

[0004] Based on the above analysis, the technical problem to be solved by the present invention is: it is difficult to take into account multiple design objectives at the same time in the optimization of wind turbine blades, such as aerodynamic performance, structural strength and cost, and it is easy to fall into a local optimal solution; it is difficult to generate an optimized design that meets multiple objectives at the same time through multi-objective training, and to gradually improve the generated design through the feedback of the discriminator to find the blade shape with the best comprehensive performance. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a generative optimization design method for wind turbine blades based on a generative adversarial network, which solves the problem that it is difficult to simultaneously take into account multiple design objectives in wind turbine blade optimization, such as aerodynamic performance, structural strength and cost, and is easy to fall into local optimal solutions; through multi-objective training, an optimized design that meets multiple objectives at the same time is generated, and through the feedback of the discriminator, the generated design is gradually improved to find the blade shape with the best comprehensive performance.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0007] A generative optimization design method for wind turbine blades based on a generative adversarial network, the steps are as follows:

[0008] Step 1: Use simulation algorithms to perform large-scale numerical simulations on existing wind turbine blades to obtain blade geometry data, mesh unit division results, lift and other characteristic evaluations, as well as stress, strain and deformation data. Analyze the vibration characteristics of different wind turbine blades under different environments and the influence mechanism of various structural parameters on the dynamic response of wind turbine blades to form a training data set. This generates a data set that can be used in subsequent steps.

[0009] Step 2: Obtain the front view, left side view, and top view of the existing wind turbine blade structure to form the original data set; obtain the original data sets of three different views; use the encoder to reduce the dimensionality of the original data sets of each view to eliminate the influence of the connection structure between the buoys on the subsequent generative optimization design, and only retain the key information of the buoys in the image; obtain the key information data set after dimensionality reduction; input the encoded original data set into the generative adversarial network to generate a new three-view image of the wind turbine blade structure;

[0010] The generated coded image is decoded and restored to the same dimension as the original dataset through dimensionality increase, and the three-dimensional structure of the wind turbine blade is reconstructed based on the three-view image. A new three-dimensional structure of the wind turbine blade is obtained, i.e., a newly designed wind turbine blade structure.

[0011] Step 3: Use the generative adversarial network to evaluate the power generation of the new wind turbine blades and obtain the power generation efficiency of the new blades;

[0012] Step 4: Based on the original data set, the generative neural network is used to analyze the original data set and identify data features and attributes. The data attributes are modified by adding noise to the attributes to obtain modified data. The modified data is passed to the predictive neural network. The predictive neural network calculates the probability that the generated output belongs to the original data set, providing a reference for the generative neural network and reducing the randomization of noise in the next cycle to obtain a probability result.

[0013] Step 5: Conduct adversarial training on the generative neural network and the predictive neural network. The generative neural network maximizes the possibility of error in the predictive neural network, while the predictive neural network minimizes its own error possibility. During the training iteration process, the two neural networks will continue to evolve and compete with each other until a state of equilibrium is reached. Finally, the optimized and trained wind turbine blade design is output.

[0014] Preferably, the sub-steps of step 1 are:

[0015] Step 1.1: First, determine the appropriate simulation algorithm and set relevant parameters, including simulation accuracy and calculation range;

[0016] Then, the model of the existing wind turbine blade is imported into the simulation software and the simulation process is started. Through the simulation, the geometric data of the blade is obtained, including information such as the shape and size of the blade. The grid unit division result is obtained.

[0017] Finally, the blade model is divided into fine grids to more accurately analyze the blade's characteristics; evaluate the blade's lift and drag characteristics, as well as stress, strain, and deformation data under various loading conditions;

[0018] Step 1.2: Analyze the vibration characteristics of different wind turbine blades under different environments and the influence mechanism of various structural parameters on the dynamic response of wind turbine blades; collect operating data of different wind turbine blades under different environmental conditions, including wind speed and direction;

[0019] These data are analyzed to study the vibration characteristics of the blades and determine which structural parameters have a significant impact on the dynamic response of the blades. Through analysis, a training data set is formed, which contains various characteristic data and structural parameter information of the blades, providing data support for subsequent steps.

[0020] Preferably, the sub-steps of step 2 are:

[0021] Step 2.1: Obtain the front, left, and top views of the existing wind turbine blade structure. Use image acquisition equipment or software tools to photograph or model the existing wind turbine blade from multiple angles to obtain the front, left, and top views of the blade structure. Combine these three views to form an original dataset, which contains structural information of the blade at different angles.

[0022] Step 2.2: Perform dimensionality reduction on the original datasets for each view. Select an appropriate encoder and input the original datasets for each view into the encoder. The encoder learns the data features and performs dimensionality reduction on the data, eliminating the influence of the connection structure between the buoys on the subsequent generative optimization design, and retaining only the key information of the buoys in the image. After dimensionality reduction, a reduced dataset of key information is obtained.

[0023] Step 2.3: Input the encoded original dataset into the GAN; set the parameters of the GAN, such as the structure and learning rate of the generator and discriminator; input the encoded original dataset into the GAN and start the network training process; the GAN generates a new three-view image of the wind turbine blade structure through continuous learning and optimization;

[0024] Step 2.4: Decode the generated coded image and reconstruct the three-dimensional structure; input the generated coded image into the decoder and restore it to the same dimension as the original data set through a dimensionality increase operation; then, based on the information of the three-view image, use three-dimensional modeling software or an algorithm to reconstruct the three-dimensional structure of the wind blade; finally, obtain a new three-dimensional structure of the wind blade, that is, a newly designed wind blade structure.

[0025] Preferably, the sub-steps of step 3 are:

[0026] Step 3.1: Input the newly designed wind turbine blade structure into the generative adversarial network. The network analyzes the structure and characteristics of the blade and predicts the power generation of the blade under different operating conditions.

[0027] Step 3.2: Determine whether the newly designed blade has higher power generation efficiency by evaluating the power generation efficiency of the new blade.

[0028] Preferably, the sub-steps of step 4 are:

[0029] Step 4.1. Analyze the original dataset using a generative neural network. Select an appropriate generative neural network structure and parameters and input the original dataset into the generative neural network. The generative neural network analyzes the original dataset and identifies the features and attributes of the data by learning the features and attributes of the data.

[0030] Step 4.2: Modify data attributes by adding noise or random changes to them. In the generative neural network, identify the attributes that need to be modified and add noise or random changes to them. By modifying data attributes, you increase the diversity of the data and provide more possibilities for subsequent training processes.

[0031] Step 4.3: Obtain the modified data and pass it to the prediction neural network; the modified data is used as the output of the generating neural network to obtain the modified data; the modified data is passed to the prediction neural network, and the prediction neural network will analyze and evaluate the data;

[0032] Step 4.4: The prediction neural network calculates the probability that the output generated belongs to the original dataset. The prediction neural network learns the characteristics and distribution of the original dataset and calculates the probability that the generated output data belongs to the original dataset. This probability value will provide a reference for the generation neural network to help the generation neural network adjust its parameters and reduce the randomization of noise points in the next cycle.

[0033] Step 4.5: Obtain the probability result; the prediction neural network outputs the calculated probability result, which will be fed back to the generation neural network to guide the optimization process of the generation neural network.

[0034] Preferably, the sub-steps of step 5 are:

[0035] Step 5.1. Conduct adversarial training. Combine the generative neural network and the predictive neural network and conduct adversarial training. During the training process, the generative neural network maximizes the probability of the predictive neural network making mistakes, while the predictive neural network minimizes the probability of its own mistakes.

[0036] Step 5.2: The two neural networks continue to evolve and compete with each other. During the training iteration process, the generative neural network and the predictive neural network will continuously learn and optimize their own parameters to improve performance. The two neural networks will continue to evolve and compete with each other, and through continuous interaction and adjustment, the performance of the entire system will gradually improve.

[0037] Step 5.3: Until a balance state is reached and the optimized trained wind blade design is output; as the training progresses, the generative neural network and the predictive neural network will gradually reach a balance state; in the balance state, the generative neural network can generate more realistic and optimized wind blade design data, and the predictive neural network can also accurately evaluate the authenticity of this data; finally, the optimized trained wind blade design is output to provide a reference for the actual wind blade production and application.

[0038] Preferably, in step 2, the encoder solves the problem that the connection structure between the buoys affects the generation of the optimized design structure through encoding and decoding operations; the operation method is:

[0039] Select 3 on the original image 3-sized regions are convolved with convolution kernels of the same size:

[0040]

[0041] Where, Indicates encoding, represents the pixel value of the i-th row and j-th column of the image, represents the convolution operation, p Represents the original image data; h represents the convolution kernel, and h The values ​​of each parameter should be determined through a large number of numerical simulation studies to form an existing wind turbine blade database and training;

[0042] The decoding operation is achieved by restoring the high-level features of the image and increasing the dimension. A feature pixel is selected on the encoded image and the size is 3 The deconvolution kernel of 3 is used for convolution calculation:

[0043]

[0044] Where, Indicates decoding, p Represents any pixel on the image; C represents the deconvolution kernel, and C The values ​​of each parameter should be determined through a large number of numerical simulation studies to form an existing wind turbine blade database and training.

[0045] A generative optimization design system for wind turbine blades based on a generative adversarial network adopts the generative optimization design method for wind turbine blades based on a generative adversarial network.

[0046] The present invention can achieve the following beneficial effects:

[0047] 1. Using generative adversarial networks to optimize the structure of wind turbine blades. By combining the advantages of generators and discriminators, efficient design generation, multi-objective optimization, data enhancement and simulation cost reduction, improved design innovation, automation and efficiency, as well as improved adaptability and generalization capabilities are achieved.

[0048] 2. By leveraging deep learning models such as autoencoders and variational autoencoders, complex 3D structures can be reconstructed and generated. This improves the automation of 3D modeling and reduces the time and cost of manual modeling. The encoder compresses high-dimensional 3D blade structure data into a low-dimensional latent space, effectively reducing data storage and processing. The decoder accurately reconstructs the original 3D structure from the low-dimensional latent representation, ensuring the integrity of design details. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The present invention will be further described below with reference to the accompanying drawings and examples:

[0050] Figure 1 This is a simulation model of the fan blade of the present invention;

[0051] Figure 2 Schematic diagram of the three-dimensional structure reconstruction method based on encoding-decoding of the present invention;

[0052] Figure 3 Generate a neural network schematic for the present invention;

[0053] Figure 4 Schematic diagram of the generative adversarial network for the present invention. DETAILED DESCRIPTION

[0054] The preferred solution is Figures 1 to 4 As shown in Figure 1, a generative optimization design method for wind turbine blades based on a generative adversarial network is proposed. The method is as follows:

[0055] Step 1: Data Preparation:

[0056] Step 1.1: Determine the simulation algorithm and parameters and perform simulation:

[0057] First, when determining the appropriate simulation algorithm, factors such as the physical characteristics of the wind turbine blades, the operating environment, and the required accuracy must be considered. For example, a finite element analysis algorithm can be selected, which enables precise mechanical analysis of complex structures. Furthermore, relevant parameters should be set, such as setting the simulation accuracy to high, medium, or low, and selecting the appropriate accuracy level based on specific needs. The calculation scope can be determined based on the size of the wind turbine blades and the operating environment, ensuring that the blades' behavior under various operating conditions is covered.

[0058] Next, the existing wind turbine blade model is imported into the simulation software. This process ensures the accuracy and completeness of the model, including information such as the blade's geometry and material properties. Once the simulation is started, the software analyzes the blade according to the predefined algorithms and parameters.

[0059] Through simulation, the geometric data of the blade is obtained, including shape information such as the length, width, and thickness of the blade, as well as the curved features of the blade surface. At the same time, the grid unit division results are obtained. Dividing the blade model into fine grids can improve the accuracy of the analysis. For example, the blade can be divided into thousands or even more small grid cells. This allows for more accurate analysis of the stress, strain, and deformation of the blade under different loads. In addition, the lift, drag, and other characteristics of the blade can be evaluated. These characteristics are important performance indicators of the blade during operation. Through simulation, the lift and drag values ​​of the blade under different wind speeds and directions can be obtained.

[0060] Step 1.2: Analyze vibration characteristics and influencing mechanisms and form a training data set:

[0061] Collect operational data, such as wind speed and direction, from different wind turbine blades under different environmental conditions. This data can be obtained through field measurements, sensor data collection, or simulation experiments. For wind speed, an anemometer can be used to record wind speed values ​​at different locations and times; for wind direction, a wind vane can be used.

[0062] The collected data is analyzed to study the blade's vibration characteristics. By monitoring and analyzing blade vibration at different wind speeds and directions, the blade's vibration frequency, amplitude, and other characteristics are determined. Furthermore, the impact of various structural parameters on the wind turbine blade's dynamic response is analyzed. For example, the impact of blade thickness, length, and width on performance indicators such as lift, drag, and stress is analyzed.

[0063] This analysis generates a training dataset. This dataset contains various blade characteristic data and structural parameter information, such as blade geometry, lift and drag characteristics, stress, strain, and deformation data, as well as environmental data such as wind speed and direction. This data will provide data support for subsequent steps in training generative adversarial networks and neural networks.

[0064] Step 2: New design generation:

[0065] Step 2.1: Get multiple views and form the original dataset:

[0066] Use an image acquisition device or software tool to photograph or model the existing wind turbine blade from multiple angles. If using an image acquisition device, use a high-resolution camera to photograph the blade from different angles, ensuring you can capture front, left, and top views. If using software tools, model the blade using 3D modeling software and capture views from different angles.

[0067] After obtaining the front, left, and top views of the blade structure, these three views are combined to form the original dataset. This dataset contains structural information of the blade at different angles, providing input data for subsequent dimensionality reduction and generative adversarial networks.

[0068] Step 2.2: Perform dimensionality reduction:

[0069] Select an appropriate encoder. The selection of an encoder should take into account the characteristics of the data and the goal of dimensionality reduction. For example, you can choose an autoencoder or principal component analysis (PCA). Input the original dataset of each view into the encoder, which learns the data's characteristics and performs dimensionality reduction on the data.

[0070] During the dimensionality reduction process, the encoder eliminates the influence of the connection structure between the buoys on the subsequent generative optimization design. This is because the connection structure may interfere with the generative adversarial network's learning and generation of key information about the blades. Through dimensionality reduction, only the key information about the buoys in the image, such as their shape, size, and position, is retained.

[0071] After dimensionality reduction, the reduced key information dataset is obtained, which will be used as the input of the generative adversarial network.

[0072] Step 2.3: Input the generative adversarial network to generate three images:

[0073] Set the parameters of the generative adversarial network, such as the generator and discriminator structures and learning rate. The generator structure can be designed based on the characteristics of the three-dimensional images of the wind turbine blade structure to be generated, for example, a convolutional neural network (CNN) architecture can be used. The discriminator structure must be able to accurately determine whether the generated images are realistic. The learning rate setting needs to be adjusted based on the size and complexity of the training data to ensure rapid network convergence.

[0074] The encoded raw dataset is fed into the generative adversarial network, initiating the network's training process. During training, the network continuously learns and optimizes, generating new three-dimensional views of the wind turbine blade structure. The generator attempts to produce realistic three-dimensional views, while the discriminator determines whether the generated images are realistic and provides feedback to the generator, prompting continuous improvement.

[0075] Step 2.4: Decode and reconstruct the 3D structure:

[0076] The resulting encoded image is fed into a decoder, whose architecture mirrors the encoder and can be restored to the same dimensions as the original dataset through a dimensionality-upgrading operation. During this dimensionality-upgrading process, the decoder restores the image's details and features, making the resulting image clearer and more realistic.

[0077] Then, based on the information from the three views, 3D modeling software or algorithms are used to reconstruct the 3D structure of the wind turbine blade. By fusing the information from the three views, geometric modeling techniques can be used to construct a 3D model of the blade. For example, triangular patch stitching can be used to convert the lines and contours in the three views into the surface of the 3D model.

[0078] The final result is a new three-dimensional structure of the wind turbine blade, that is, a completely new design of the wind turbine blade structure. This structure will serve as input for subsequent steps to conduct power generation evaluation and optimization training.

[0079] Step 3: Power generation evaluation:

[0080] Step 3.1: Input the new blade structure for power generation prediction:

[0081] The newly designed wind turbine blade structure is fed into a generative adversarial network. The network analyzes the blade's structure and characteristics to predict the blade's power generation under different operating conditions. For example, based on information such as the blade's shape, size, and material, as well as environmental factors like wind speed and direction, the network can calculate the blade's lift and drag under different operating conditions, thereby predicting power generation.

[0082] Step 3.2: Evaluate power generation efficiency to determine performance:

[0083] The new blade's power generation efficiency is evaluated. This efficiency is determined by calculating the ratio of the power generated by the blade to the wind energy input. The newly designed blade's power generation efficiency is compared with the existing blade to determine whether it offers higher power generation efficiency. If the new blade's power generation efficiency is higher, the design is successful. If the power generation efficiency does not improve or even decreases, further design optimization is necessary.

[0084] Step 4: Optimize training:

[0085] Step 4.1: Analyze the original dataset using the generated neural network:

[0086] Select the appropriate neural network structure and parameters. The neural network structure can be selected based on the characteristics of the original dataset and the optimization goal. For example, a deep neural network (DNN) or a recurrent neural network (RNN) can be used. Parameter settings include the number of network layers, the number of neurons, and the activation function, which need to be adjusted based on the scale and complexity of the data.

[0087] The original dataset is fed into a generative neural network. The generative neural network analyzes the original dataset and identifies its features and attributes by learning the data's characteristics and attributes. For example, the network can learn features such as a blade's geometry, material properties, lift and drag characteristics, and the relationships between these features.

[0088] Step 4.2: Add noise to modify data attributes:

[0089] In the generative neural network, the attributes that need to be modified are determined. These attributes can be selected based on the optimization goal and the characteristics of the data. For example, geometric parameters such as blade thickness, length, and width, or performance indicators such as lift and drag can be selected.

[0090] Add noise or random variations to these attributes. You can use methods like random number generators or normal distributions to add noise. By modifying data attributes, you increase data diversity and provide more possibilities for subsequent training. For example, you can generate data for blades of varying thicknesses by varying the thickness of the blade to explore their impact on performance.

[0091] Step 4.3: Pass the modified data to the prediction neural network:

[0092] The modified data is used as the output of the generative neural network to obtain the modified data, which will contain new blade structure and performance information.

[0093] The modified data is passed to the predictive neural network. The predictive neural network will analyze and evaluate the data. The structure and parameters of the predictive neural network also need to be selected and adjusted according to the characteristics of the data and the optimization goal.

[0094] Step 4.4: Calculate the probability that the output belongs to the original dataset:

[0095] Predictive neural networks learn the characteristics and distribution of the original dataset and calculate the probability that the generated output data belongs to the original dataset. This probability value reflects the degree of similarity between the generated data and the original data. A high probability value indicates that the generated data is relatively similar to the original data; a low probability value indicates that the generated data is significantly different from the original data.

[0096] This probability value will provide a reference for the generative neural network, helping it adjust parameters to reduce the randomness of noise in the next cycle. For example, if the probability value is low, the generative neural network can adjust the amount of noise added or select different attributes to modify to improve the quality of the generated data.

[0097] Step 4.5: Get probability results and guide optimization:

[0098] The predictive neural network outputs the calculated probability result. This result is fed back to the generative neural network to guide its optimization process. The generative neural network can adjust its parameters and operations based on the probability result to improve the quality and authenticity of the generated data. This process also helps the predictive neural network continuously learn and optimize its judgment ability, improving the accuracy of its evaluation of generated data.

[0099] Step 5: Adversarial Training:

[0100] Step 5.1: Perform adversarial training:

[0101] The generative neural network and the predictive neural network are combined for adversarial training. During training, the goal of the generative neural network is to maximize the probability of error in the predictive neural network, that is, to generate data that can fool the predictive neural network into thinking it is real data. The goal of the predictive neural network, on the other hand, is to minimize its own error probability, that is, to accurately determine whether the generated data is real.

[0102] Step 5.2: Neural Network Evolution and Adversarial:

[0103] During the training iterations, the generative and predictive neural networks continuously learn and optimize their parameters to improve performance. The generative neural network continuously adjusts its noise addition strategy and data attribute modification methods to generate more realistic data. The predictive neural network continuously learns the characteristics and distribution of the original dataset to improve the accuracy of its judgment of the generated data.

[0104] The two neural networks will continuously evolve and compete with each other. Through constant interaction and adjustment, the performance of the entire system will gradually improve. For example, the data generated by the generative neural network will become more and more realistic, and the judgment ability of the predictive neural network will also become stronger.

[0105] Step 5.3: Reach equilibrium and output the optimized design:

[0106] As training progresses, the generative and predictive neural networks gradually reach a state of equilibrium, where the generative neural network can generate more realistic and optimized wind blade design data, while the predictive neural network can accurately assess the authenticity of this data.

[0107] Ultimately, the optimized wind turbine blade design is output. This design will provide a reference for actual wind turbine blade production and application, thereby improving the power generation efficiency, reliability, and performance of wind turbines.

[0108] In step 2, the encoder uses encoding-decoding operations to solve the problem of the connection structure between the buoys affecting the generation of the optimized design structure. The specific operation method is as follows:

[0109] We select 3×3 regions of the original image and perform convolution with convolution kernels of the same size. This process can be understood as extracting local features from the image. The convolution kernel performs a weighted summation of the local regions of the image, thereby extracting the image's features.

[0110] The formula is:

[0111]

[0112] Where, p Represents the original image data; h represents the convolution kernel, and h The values ​​of each parameter should be determined through a large number of numerical simulation studies to form an existing wind turbine blade database and training.

[0113] The decoding operation is achieved by restoring the high-level features of the image and increasing the dimension. Specifically, a feature pixel is selected from the encoded image and the size is 3 The deconvolution kernel of 3 is used for convolution calculation:

[0114]

[0115] Where, p Represents any pixel on the image; C represents the deconvolution kernel, and C The values ​​of each parameter should be determined through a large number of numerical simulation studies to form an existing wind turbine blade database and training.

[0116] Finally, a new wind turbine blade structure was proposed using the above-mentioned generative optimization design method and the encoding-decoding based three-dimensional structure reconstruction method. The vibration reduction effectiveness of the new wind turbine blade was verified by combining numerical simulation and large-scale water tank tests.

[0117] The Generative Adversarial Network (GAN) of the present invention is a deep learning architecture that trains two neural networks to compete against each other to generate more realistic new data from a given training dataset, such as new images from an existing image database. An adversarial network is a network that trains two different networks and pits them against each other. One network generates new data by taking input data samples and modifying them as much as possible, while the other network attempts to predict whether the generated data output belongs to the original dataset. In other words, the prediction network determines whether the generated data is real or fake. The generation network generates newer, improved versions of the fake data values ​​until the prediction network can no longer distinguish the fake data values ​​from the original data values.

[0118] A generative adversarial network-based wind turbine blade optimization design system adopts the above design method. The system includes the following main modules:

[0119] Data acquisition module: This module is responsible for collecting various data of existing wind turbine blades, including geometric data, vibration characteristics data, operational data, etc. This data can be obtained through sensors, measuring equipment, or databases.

[0120] Simulation module: Use simulation algorithms to perform large-scale numerical simulations on existing wind turbine blades, obtain various characteristic data of the blades, and analyze the influence mechanism of vibration characteristics and structural parameters on dynamic response under different environments.

[0121] View acquisition and processing module: obtains the front view, left view, and top view of the existing wind turbine blade structure, and performs dimensionality reduction and encoding and decoding operations to provide input data for the generative adversarial network.

[0122] Generative adversarial network module: Generates a new three-view image of the wind blade structure, and obtains a new three-dimensional structure of the wind blade through decoding and three-dimensional reconstruction.

[0123] Power Generation Evaluation Module: Evaluate the power generation of new wind turbine blades to determine their power generation efficiency.

[0124] Neural network module: includes generating neural networks and predicting neural networks, and performs optimization training by adding noise, modifying data attributes, calculating probabilities, and other operations.

[0125] Adversarial training module: performs adversarial training on the generative neural network and the predictive neural network to improve the performance of the system and the quality of the output results.

[0126] Output module: Outputs optimized and trained wind turbine blade designs to provide a reference for actual production and application.

[0127] Through the collaborative work of these modules, the system can achieve generative optimization design of wind turbine blades based on generative adversarial networks, thereby improving the performance and power generation efficiency of wind turbine blades.

[0128] The above embodiments are merely preferred technical solutions of the present invention and should not be construed as limiting the present invention. The scope of protection of the present invention shall be the technical solutions set forth in the claims, including equivalent alternatives to the technical features of the technical solutions set forth in the claims. In other words, equivalent alternatives and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A generative optimization design method for wind turbine blades based on generative adversarial networks, characterized by The following steps are involved: Step 1: Use simulation algorithms to perform large-scale numerical simulations on existing wind turbine blades to obtain blade geometry data, mesh unit division results, lift and other characteristic evaluations, as well as stress, strain, and deformation data. Analyze the vibration characteristics of different wind turbine blades in different environments and the influence mechanism of various structural parameters on the dynamic response of wind turbine blades to form a training data set; Obtain a data set that can be used in subsequent steps; Step 2: Obtain the front view, left view, and top view of the existing wind turbine blade structure to form an original data set; Obtain original datasets from three different views. Use an encoder to reduce the dimensionality of each view's original dataset, eliminating the influence of the connection structure between buoys on the subsequent generative optimization design, and retaining only the key information of the buoys in the image. Obtain the key information dataset after dimensionality reduction; input the encoded original dataset into the generative adversarial network to generate a new three-view diagram of the wind turbine blade structure; The generated coded image is decoded and restored to the same dimension as the original dataset through dimensionality increase, and the three-dimensional structure of the wind turbine blade is reconstructed based on the three-view image. A new three-dimensional structure of the fan blade is obtained, that is, a completely new designed fan blade structure; Step 3: Use the generative adversarial network to evaluate the power generation of the new wind turbine blade; Get the power generation efficiency of the new blade; Step 4: Based on the original data set, the generative neural network is used to analyze the original data set and identify data features and attributes. The data attributes are modified by adding noise to the attributes to obtain modified data. The modified data is passed to the predictive neural network. The predictive neural network calculates the probability that the generated output belongs to the original data set, providing a reference for the generative neural network and reducing the randomization of noise in the next cycle. Used to obtain probability results; Step 5: Conduct adversarial training on the generative neural network and the predictive neural network. The generative neural network maximizes the possibility of error in the predictive neural network, while the predictive neural network minimizes its own error possibility. During the training iteration process, the two neural networks will continue to evolve and compete with each other until a state of equilibrium is reached. Finally, the optimized and trained wind turbine blade design is output.

2. The method for generative optimization design of wind turbine blades based on a generative adversarial network according to claim 1, characterized in that: The sub-steps of step 1 are: Step 1.1: First, determine the appropriate simulation algorithm and set relevant parameters, including simulation accuracy and calculation range; Then, the model of the existing wind turbine blade is imported into the simulation software and the simulation process is started; Acquiring geometric data of the blade through simulation, wherein the geometric data includes information including the shape and size of the blade; Get the grid unit division result, Finally, the blade model is divided into fine grids to more accurately analyze the blade's characteristics; evaluate the blade's lift and drag characteristics, as well as stress, strain, and deformation data under various loading conditions; Step 1.2: Analyze the vibration characteristics of different wind turbine blades under different environments and the influence mechanism of various structural parameters on the dynamic response of wind turbine blades; Collect operating data of different wind turbine blades under different environmental conditions, including wind speed and direction; The operating data is analyzed to study the vibration characteristics of the blades and determine which structural parameters have a significant impact on the dynamic response of the blades. Through the analysis, a training data set is formed, which contains various characteristic data and structural parameter information of the blades.

3. The method for generative optimization design of wind turbine blades based on a generative adversarial network according to claim 1, characterized in that: The sub-steps of step 2 are: Step 2.1: Obtain the front view, left side view, and top view of the existing wind turbine blade structure; Use image acquisition equipment or software tools to photograph or model existing wind turbine blades from multiple angles to obtain the front, left, and top views of the blade structure. Combine these three views to form an original dataset containing structural information from different blade angles. Step 2.2: Perform dimensionality reduction on the original dataset of each view; Select an appropriate encoder and input the original dataset of each view into the encoder. The encoder learns the data features and performs dimensionality reduction on the data, eliminating the influence of the connection structure between the buoys on the subsequent generative optimization design, and retaining only the key information of the buoys in the image. After the dimensionality reduction process, the key information dataset is obtained. Step 2.3: Input the encoded original dataset into the generative adversarial network; Set the parameters of the generative adversarial network, such as the structure of the generator and discriminator, and the learning rate; input the encoded original dataset into the generative adversarial network to start the network training process; The generative adversarial network generates new three-dimensional views of the wind turbine blade structure through continuous learning and optimization; Step 2.4: Decode the generated coded image and reconstruct the 3D structure; input the generated coded image into the decoder and restore it to the same dimension as the original dataset through dimensionality increase operation; Then, based on the information of the three-view drawings, the three-dimensional structure of the wind blade is reconstructed using three-dimensional modeling software or an algorithm; ultimately, a new three-dimensional structure of the wind blade is obtained, that is, a newly designed wind blade structure.

4. The method for generative optimization design of wind turbine blades based on a generative adversarial network according to claim 1, characterized in that: The sub-steps of step 3 are: Step 3.1: Input the newly designed wind turbine blade structure into the generative adversarial network. The network analyzes the structure and characteristics of the blade and predicts the power generation of the blade under different operating conditions. Step 3.2: Determine whether the newly designed blade has higher power generation efficiency by evaluating the power generation efficiency of the new blade.

5. The method for generative optimization design of wind turbine blades based on a generative adversarial network according to claim 1, characterized in that: The sub-steps of step 4 are: Step 4.

1. Analyze the original dataset using a generative neural network. Select an appropriate generative neural network structure and parameters and input the original dataset into the generative neural network. The generative neural network analyzes the original dataset and identifies the features and attributes of the data by learning the features and attributes of the data. Step 4.2: Modify data attributes by adding noise or random changes to them. In the generative neural network, identify the attributes that need to be modified and add noise or random changes to them. By modifying data attributes, you increase the diversity of the data and provide more possibilities for subsequent training processes. Step 4.3: Obtain the modified data and pass it to the prediction neural network; the modified data is used as the output of the generating neural network to obtain the modified data; the modified data is passed to the prediction neural network, and the prediction neural network will analyze and evaluate the data; Step 4.4: The prediction neural network calculates the probability that the output generated belongs to the original dataset. The prediction neural network learns the characteristics and distribution of the original dataset and calculates the probability that the generated output data belongs to the original dataset. This probability value will provide a reference for the generation neural network to help the generation neural network adjust its parameters and reduce the randomization of noise points in the next cycle. Step 4.5: Obtain the probability result; the prediction neural network outputs the calculated probability result, which will be fed back to the generation neural network to guide the optimization process of the generation neural network.

6. The method for generative optimization design of wind turbine blades based on a generative adversarial network according to claim 1, characterized in that: The sub-steps of step 5 are: Step 5.

1. Conduct adversarial training. Combine the generative neural network and the predictive neural network and conduct adversarial training. During the training process, the generative neural network maximizes the probability of the predictive neural network making mistakes, while the predictive neural network minimizes the probability of its own mistakes. Step 5.2: The two neural networks continue to evolve and compete with each other. During the training iteration process, the generative neural network and the predictive neural network will continuously learn and optimize their own parameters to improve performance. The two neural networks will continue to evolve and compete with each other, and through continuous interaction and adjustment, the performance of the entire system will gradually improve. Step 5.3: until a balance state is reached and the optimized wind turbine blade design is output; as the training progresses, the generative neural network and the predictive neural network will gradually reach a balance state; In a balanced state, the generative neural network can generate more realistic and optimized wind blade design data, and the predictive neural network can also accurately evaluate the authenticity of this data; Finally, the optimized and trained wind blade design is output to provide a reference for the actual production and application of wind blades.

7. The method for generative optimization design of wind turbine blades based on a generative adversarial network according to claim 1, characterized in that: In step 2, the encoder solves the problem of the connection structure between the buoys affecting the generation of the optimized design structure through encoding and decoding operations; the operation method is: Select 3 on the original image 3-sized regions are convolved with convolution kernels of the same size: Where, Indicates encoding, represents the pixel value of the i-th row and j-th column of the image, represents the convolution operation, p Represents the original image data; h represents the convolution kernel, and h The values ​​of each parameter should be determined through a large number of numerical simulation studies to form an existing wind turbine blade database and training; The decoding operation is achieved by restoring the high-level features of the image and increasing the dimension. A feature pixel is selected on the encoded image and the size is 3 The deconvolution kernel of 3 is used for convolution calculation: Where, Indicates decoding, p Represents any pixel on the image; C represents the deconvolution kernel, and C The values ​​of each parameter should be determined through a large number of numerical simulation studies to form an existing wind turbine blade database and training.

8. A generative optimization design system for wind turbine blades based on a generative adversarial network, characterized by: A generative optimization design method for wind turbine blades based on a generative adversarial network according to any one of claims 1 to 7 is adopted.

Images