Optimization design method of tidal current energy water turbine based on generator and predictor

Abstract

The invention provides a tidal current energy water turbine intelligent design method based on a generator and a predictor and a water turbine structure optimization design framework. According to the method, a compact generator is obtained by training a variational self-encoding generative adversarial network and is used for parametric modeling of tidal current energy water turbine rotor blades, and an efficient performance predictor based on the combination of an artificial neural network and transfer learning is established to realize intelligent prediction of blade performance. The generator and the predictor are coupled with the multi-target genetic algorithm, efficient design of the tidal current energy water turbine blade is achieved, the performance of the tidal current energy water turbine blade can be rapidly predicted, and then efficient optimization design of the tidal current energy water turbine is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of optimal design of tidal current turbines, and specifically relates to an optimal design method of tidal current turbines based on a generator and a predictor. Background Art

[0002] The efficient development and utilization of marine energy is one of the important challenges in the 21st century. Compared with traditional energy sources, tidal current energy has the advantages of being clean, pollution-free, and highly predictable. Tidal current turbines are important mechanical devices for obtaining marine tidal current energy at present, and the efficient optimal design of their hydrodynamic performance is the main research issue at present.

[0003] The rotor blade is the main component for energy capture of tidal current turbines. At present, the optimization method for rotor blades first requires parametric modeling of the blades including chord length, twist angle, and percentage thickness distribution, then combines with water tunnel experiments or computational fluid dynamics to obtain hydrodynamic performance, and finally uses optimization algorithms to explore the turbine design with the best hydrodynamic performance.

[0004] However, in the parametric modeling during the optimization process currently, it is necessary to pre-define parameters for the design space in advance, and then find the global optimum within the pre-defined parameter range. As the scale of the design space increases, the computational amount of performance evaluation also increases exponentially, which brings huge design costs.

[0005] To solve the above problems, it is a common method to perform feature dimensionality reduction on the turbine model, and then obtain a compact parametric representation to improve the optimization efficiency. In addition, with the development of deep learning, a series of intelligent algorithms can greatly improve the speed and accuracy of performance evaluation. Summary of the Invention

[0006] Aiming at the above existing problems, the present invention aims to provide an optimal design method of tidal current turbines based on a generator and a predictor, including establishing a dataset of the geometric structure and hydrodynamic performance of tidal current turbines based on Latin hypercube sampling and blade element momentum theory; training a generator model capable of representing turbine designs based on a variational autoencoder generative adversarial network, improving the compactness of turbine structure representation while retaining the optimal representation ability, thereby accelerating the optimization convergence speed; establishing a predictor model based on transfer learning and artificial neural networks to achieve rapid and intelligent prediction of turbine geometric features to hydrodynamic performance parameters; in the final optimal design, coupling the generator and the predictor, using a filter to improve the generation quality of the generator, using the blade element momentum theory to judge the prediction accuracy of the predictor model and dynamically fine-tuning the network weights of the predictor.

[0007] To achieve the above object, the technical solutions adopted by the present invention are as follows:

[0008] A method for optimizing the design of a tidal current energy turbine based on a generator and a predictor, comprising the following steps:

[0009] Step S1: Based on Bessel parametric modeling and Latin hypercube theory, generate a dataset of geometric parameters of the tidal current energy turbine blades, and then divide the sample dataset into a training dataset and a test dataset according to a data ratio of 8:1:1;

[0010] Step S2: Select the NACA4415 airfoil as the reference airfoil, combine it with the dataset of blade geometric parameters obtained in step one to perform airfoil transformation, generate a three-dimensional tidal current energy turbine model, and based on the blade element momentum theory, calculate the performance of the three-dimensional tidal current energy turbine model to generate the hydrodynamic performance of the turbine;

[0011] Step S3: Use a variational autoencoder generative adversarial network to construct a turbine generator model that maps from a latent vector to the geometric parameters of the tidal current energy turbine blades, and couple a Savitzky-Golay filter to the last layer of the generator to smooth some outliers;

[0012] Step S4: Use an artificial neural network to construct a predictor model that maps from the geometric parameters of the tidal current energy turbine blades to the hydrodynamic performance, pre-train it on the original dataset generated in step one, and perform transfer learning on the target domain dataset generated by the generator model obtained in step three;

[0013] Step S5: Based on the multi-objective non-dominated genetic algorithm, optimize the geometric parameters of the tidal current energy turbine with the maximization of the power coefficient and the minimization of the thrust coefficient as the optimization objectives.

[0014] Further, the specific process of training the generator described in step S3 is as follows:

[0015] Step S3.1: First, normalize the chord length, twist angle, and percentage thickness distribution in the blade geometric parameter dataset respectively;

[0016] Step S3.2: Input the normalized data into the variational autoencoder generative adversarial network, reconstruct new blade geometric parameter data through three steps: a feature encoder, variational inference, and a feature generator, and use a feature discriminator to discriminate the features of the newly generated blade geometric parameter data, and then perform iterative weight optimization with the generator;

[0017] Step S3.3: Use a Savitzky-Golay filter in the last layer of the generator to smooth some outliers in the newly generated turbine.

[0018] Further, the specific process of optimizing the generator weights described in step S3.2 is as follows:

[0019] Step S3.2.1: Use an encoder to extract features from the input normalized data to obtain a low-dimensional latent vector, and through variational inference, make the latent vector conform to the standard normal distribution. The variational inference formula is: Q(z|x) = N(z; μ(x), diag(σ(x))), where x represents the input data, z represents the latent vector, μ(x) represents the mean, σ(x) represents the standard deviation, and the diagonal matrix in the formula indicates that all elements in the latent vector are independent;

[0020] Step S3.2.2: Introduce the method of adversarial training, and use the cross-entropy loss function to iteratively optimize the weights of the generator and the discriminator to improve the generation quality of the generator. The cross-entropy loss function is: where D is the discriminator and G is the generator. Through adversarial training, the generation ability of the generator can be iteratively enhanced.

[0021] Furthermore, the specific process of training the predictor described in step S4 is as follows:

[0022] Step S4.1: Normalize the blade geometric parameter dataset obtained in step one and input it into the artificial neural network model, and use the hydrodynamic performance obtained in step S2 as the output for neural network pre-training;

[0023] Step S4.2: Use the generator model trained in step S3, use Gaussian noise as the input, reconstruct a large number of high-quality hydroturbine datasets, and use the blade element momentum theory to obtain the corresponding hydrodynamic performance;

[0024] Step S4.3: Freeze the first two layers of the artificial neural network model pre-trained in step S4.1, use the dataset generated in step S4.2 as the target domain, and perform transfer learning and retraining to improve the generalization and accuracy of the neural network. The code for freezing the neural network is: for layer in ANN.layers[:2]: layer.trainable = False.

[0025] Furthermore, in step S5, in the multi-objective optimization process, the generator generates a large number of tidal current turbines, and the predictor is used to evaluate the performance of the generated turbines. In addition, to improve the accuracy of performance evaluation, the blade element momentum theory is used as the judgment criterion, and the predictor parameters are dynamically fine-tuned.

[0026] The beneficial effects of the present invention are: The present invention discloses an optimized design method for tidal current turbines based on a generator and a predictor. Compared with the prior art, the improvements of the present invention are as follows:

[0027] For the optimal design of tidal current turbines, the present invention designs an optimal design method for tidal current turbines based on a generator and a predictor, which can couple the generator and the predictor, without relying on the predefined parameter intervals in traditional parametric modeling and the performance evaluation based on computational fluid dynamics. Compared with the traditional optimal design of tidal current turbines, the optimization time is shortened by more than 96%. Description of the Drawings Figure 1 It is a flow chart of the optimal design method for tidal current turbines based on a generator and a predictor; Figure 2 It is the structure of the variational autoencoder generative adversarial network model; Figure 3 It is the comparison between the optimal design result of the tidal current turbine based on a generator and a predictor and the optimal result based on the blade element momentum theory; Detailed Embodiments Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention. The optimal design method for tidal current turbines based on a generator and a predictor proposed by the present invention takes the horizontal axis tidal current turbine as an example, and includes four parts: the construction of the data set, the construction of the turbine generator, the construction of the hydrodynamic performance predictor, and the optimization of the turbine geometric structure based on the multi-objective non-dominated optimization algorithm. Embodiment 1: Refer to Figure 1 As shown, an intelligent design method for tidal current turbines based on a generator and a predictor includes the following steps: I. Construction of the data set 1. Construction of the data set of turbine geometric parameters In the present invention, based on the Latin hypercube theory, sampling of the blade geometric parameters is carried out, and constraints on the decreasing chord length, twist angle, and percentage thickness distribution are set during the sampling process to obtain 208 groups of blade geometric data that meet the requirements of actual production and manufacturing. 2. Construction of the data set of turbine hydrodynamic performance In the present invention, the hydrodynamic performance is evaluated based on the blade element momentum theory. The blade element momentum theory is a theoretical method for analyzing and evaluating the performance of rotating machinery. Its core idea is to divide the mechanical rotor or blade into several small units, and then analyze the hydrodynamic force of each unit to finally obtain the overall performance. Furthermore, the present invention uses a Python script to call the Xfoil software to evaluate the hydrodynamic performance of the blade unit, and uses the lift and drag coefficients at different angles of attack calculated by Xfoil as the input of the blade element momentum theory to obtain the power coefficient and thrust coefficient of the water turbine at different tip speed ratios. II. Construction of the water turbine generator The present invention obtains a generator by training a variational autoencoder generative adversarial network to generate high-dimensional water turbine geometric parameters from a latent vector. Figure 2 Fig. shows the model structure of the variational autoencoder generative adversarial network. Furthermore, the present invention selects 80% of the data in the dataset as the training set and 10% as the validation set. In addition, a dynamic learning rate Adam optimizer is used, with an initial learning rate of 0.0001 and a decay rate of 0.001, and 100,000 training iterations are performed. Furthermore, the present invention couples the Savitzky-Golay filter with the last layer of the generator to smooth some outliers in the newly generated water turbine. III. Construction of the hydrodynamic performance predictor An artificial neural network is used to construct a predictor model from the blade geometric parameters of the tidal energy water turbine to the hydrodynamic performance. The independently normalized blade geometric parameters are used as the input, and the corresponding hydrodynamic performance is used as the output to quickly obtain the hydrodynamic performance of the water turbine rotor. In addition, transfer learning is used to improve the generalization ability and accuracy of the model. Furthermore, in the pre-training, 208 groups of data generated by Bezier curves are used as the input, and the corresponding hydrodynamic performance is used as the output. 80% of the data in the dataset is selected as the training set and 10% as the validation set. In addition, a dynamic learning rate Adam optimizer is used, with an initial learning rate of 0.01 and a decay rate of 0.001, and 5,000 training iterations are performed. Furthermore, in the optimization of the predictor performance using transfer learning, 208 groups of data generated by the generator are used as the input, and the hydrodynamic performance calculated by the blade element momentum theory is also used as the output. With other training parameters unchanged, the first two fully connected layers of the predictor are frozen for parameter fine-tuning. IV. Optimization of the water turbine geometric structure The present invention couples the trained generator, predictor with a multi-objective optimization algorithm, uses the generator to generate a water turbine rotor from a low-dimensional latent vector, and uses the predictor to predict its hydrodynamic performance. In addition, the blade element momentum theory is used to evaluate the accuracy of the predictor and fine-tune the predictor weights. The present invention uses the multi-objective genetic algorithm (NSGA-II) for multi-objective optimization. First, a population is randomly generated, and then the processes of selection, crossover, and mutation are simulated to eliminate individuals with poor fitness and retain individuals with strong fitness. After several generations of evolution, the optimal individuals in the population are obtained. To further verify the optimization method described in the present invention, under the same optimization conditions, this method is compared and verified with the traditional blade element momentum theory. Table 1 gives the comparison of the optimization parameters and time of the two methods. Figure 3 It is the Pareto front of the optimization result of the water turbine. Iteration settings and time comparison of two optimization frameworks In the embodiment of the present invention, under the same conditions, the optimization design method of the tidal energy turbine based on the generator and predictor proposed by the present invention only uses 4% of the time and obtains optimization results similar to those of the method based on the blade element momentum theory. Furthermore, it shows that this optimization method effectively reduces the optimization cost of the tidal energy turbine. The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for optimizing the design of a tidal current turbine based on a generator and a predictor, characterized in that, It includes the following steps: Step S1: Based on Bessel parametric modeling and Latin hypercube theory, generate a dataset of geometric parameters of tidal current turbine blades, and then divide the sample dataset into a training dataset and a test dataset according to a data ratio of 8:1:1; Step S2: Select the NACA4415 airfoil as the reference airfoil, combine it with the blade geometric parameter dataset obtained in Step 1 for airfoil transformation to generate a three-dimensional tidal current turbine model, and based on the blade element momentum theory, calculate the performance of the three-dimensional tidal current turbine model to generate the hydrodynamic performance of the turbine; Step S3: Use a variational autoencoder generative adversarial network to construct a turbine generator model that maps from a latent vector to the geometric parameters of tidal current turbine blades, and couple a Savitzky-Golay filter at the last layer of the generator to smooth some outliers; Step S4: Use an artificial neural network to construct a predictor model that maps from the geometric parameters of tidal current turbine blades to hydrodynamic performance, pre-train it on the original dataset generated in Step 1, and perform transfer learning on the target domain dataset generated by the generator model obtained in Step 3; Step S5: Based on the multi-objective non-dominated genetic algorithm, optimize the geometric parameters of the tidal current turbine with the maximization of the power coefficient and the minimization of the thrust coefficient as the optimization objectives.

2. The optimization design method of a tidal current turbine based on a generator and a predictor according to claim 1, wherein: The specific training process of the generator described in Step S3 is as follows: Step S3.1: First, normalize the chord length, twist angle, and percentage thickness distribution in the blade geometric parameter dataset respectively; Step S3.2: Input the normalized data into the variational autoencoder generative adversarial network, reconstruct new blade geometric parameter data through three steps: feature encoder, variational inference, and feature generator, and use the feature discriminator to discriminate the features of the newly generated blade geometric parameter data, and then perform iterative weight optimization with the generator; Step S3.3: Use a Savitzky-Golay filter at the last layer of the generator to smooth some outliers in the newly generated turbine.

3. The optimal design method of the tidal current turbine based on the generator and the predictor according to claim 2, characterized in that: The specific process of generator weight optimization described in Step S3.2 is as follows: Step S3.2.1: Use the encoder to extract features from the input normalized data to obtain a low-dimensional latent vector, and through variational inference, make the latent vector conform to the standard normal distribution. The variational inference formula is: Q(z|x) = N(z; μ(x), diag(σ(x))), where x represents the input data, z represents the latent vector, μ(x) represents the mean, σ(x) represents the standard deviation, and the diagonal matrix in the formula indicates that all elements in the latent vector are independent; Step S3.2.2: Introduce an adversarial training method, and use the cross-entropy loss function to iteratively optimize the weights of the generator and the discriminator to improve the generation quality of the generator. The cross-entropy loss function is: where D is the discriminator and G is the generator. Through adversarial training, the generation ability of the generator can be iteratively enhanced.

4. The optimal design method of the tidal current turbine based on the generator and the predictor according to claim 1, wherein: The specific training process of the predictor described in Step S4 is as follows: Step S4.1: Normalize the blade geometric parameter dataset obtained in Step 1 and input it into the artificial neural network model, using the hydrodynamic performance of the water turbine obtained in Step S2 as the output for neural network pre-training; Step S4.2: Use the generator model trained in Step S3, with Gaussian noise as the input, to reconstruct a large number of high-quality water turbine datasets, and use the blade element momentum theory to obtain the corresponding hydrodynamic performance; Step S4.3: Freeze the first two layers of the artificial neural network model pre-trained in Step S4.1, and use the dataset generated in Step S4.2 as the target domain for transfer learning retraining to improve the generalization and accuracy of the neural network.

5. The optimized design method of a tidal current turbine based on a generator and a predictor according to claim 1, characterized in that: In Step S5, the multi-objective optimization process generates a large number of tidal current turbines through the generator, and uses the predictor to evaluate the performance of the generated turbines. In addition, to improve the accuracy of performance evaluation, the blade element momentum theory is used as the judgment criterion, and the parameters of the predictor are dynamically fine-tuned.

Images