叶继红, 杨振宇. 基于生成式对抗网络的风场生成研究[J]. 工程力学, 2021, 38(10): 1-11. DOI: 10.6052/j.issn.1000-4750.2020.10.0721
引用本文: 叶继红, 杨振宇. 基于生成式对抗网络的风场生成研究[J]. 工程力学, 2021, 38(10): 1-11. DOI: 10.6052/j.issn.1000-4750.2020.10.0721
YE Ji-hong, YANG Zhen-yu. RESEARCH ON GENERATION OF WIND FIELDS BASED ON GAN[J]. Engineering Mechanics, 2021, 38(10): 1-11. DOI: 10.6052/j.issn.1000-4750.2020.10.0721
Citation: YE Ji-hong, YANG Zhen-yu. RESEARCH ON GENERATION OF WIND FIELDS BASED ON GAN[J]. Engineering Mechanics, 2021, 38(10): 1-11. DOI: 10.6052/j.issn.1000-4750.2020.10.0721

基于生成式对抗网络的风场生成研究

RESEARCH ON GENERATION OF WIND FIELDS BASED ON GAN

  • 摘要: 生成式对抗网络(Generative Adversarial Network,GAN)是人工智能领域较为重要的思想与方法。该文提出基于GAN生成风场,其中,由于GAN需要训练数据,考虑到实测风场数据的困难与匮乏,该文通过改进循环预前模拟法生成训练数据,并利用GANθ与GANΔθ生成风场,前者用于生成单点相位谱,后者用于生成单位距离相位谱差值,以克服传统模拟方法湍流度略低、频谱特性失真等不足;进而基于中心递进法利用GANθ、GANΔθ的结果生成相位谱;利用相位谱、幅值谱生成风场。从数据分布角度定性评估了GAN结果质量;利用1-NN算法定量评估了GAN结果质量;从风场特性角度将GAN生成的风场与目标风场进行了对比验证。通过定性、定量及对比验证可得:基于GAN生成的数据分布与目标分布相接近,生成的风场特性与目标风场相接近,说明基于GAN生成风场方法通过学习数据分布特性生成数据,具有良好的适应能力与生成数据能力。

     

    Abstract: The Generative Adversarial Network (GAN) is an important thought and method in the field of artificial intelligence. A generation method of wind fields is proposed in this paper based on GAN. To overcome the lacking of on-site measurements of wind field data, the modified precursor method is firstly used to generate the training data in GAN (i.e. GANθ and GANΔθ). The former model (GANθ) is used to generate single-point phase spectrum while the latter (GANΔθ) is to create phase spectrum difference per unit distance. This can overcome the disadvantages of slightly lower turbulence and distortion of spectral characteristics in traditional simulation methods. Based on the Center Progressive method, the phase spectrum is obtained from the results of GANθ and GANΔθ models. Wind fields are generated by the phase spectrum and amplitude spectrum. The quality of GAN-based results is evaluated qualitatively and quantitatively through data distribution and 1-NN algorithm, respectively. Furthermore, the characteristics of wind fields predicted by GAN are compared with those of the target wind fields. The evaluation and comparison results show that the data distribution generated by GAN agrees well with the target distribution, and the characteristics of generated wind fields are in a good agreement with those of target wind fields. This indicates that the GAN-based generation method of wind fields, by generating data through learning data distribution, has good adaptability and capability of data generation.based generation method of wind fields, by generating data through learning data distribution, has good adaptability and capability of data generation.

     

/

返回文章
返回