Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Enhancing wind field resolution in complex terrain through a knowledge-driven machine learning approach

Jacob Wulff Wold, Florian Stadtmann, Adil Rasheed, Mandar Tabib, Omer San, Jan-Tore Horn

TL;DR

The paper tackles the challenge of real-time, high-resolution wind-field prediction in complex terrain by coupling mesoscale and microscale simulations and delivering a 3D wind-field super-resolution model based on a physics-informed ESRGAN variant. It replaces perceptual loss with gradient- and divergence-based losses rooted in atmospheric physics, and finds that a non-adversarial, fully convolutional generator can outperform trilinear interpolation while maintaining physical consistency. The method leverages open data from HARMONIE-SIMRA, achieves a PSNR of 47.14 dB and a mean wind-vector error of 0.24 m/s, and supports upscaling factors up to 16×, enabling faster digital-twin-style analyses. The results demonstrate that physics-based losses can obviate the need for adversarial training, yielding stable, accurate reconstructions and enabling practical deployment for wind-energy applications in complex terrain. An open data release and discussion of digital-twin integration highlight the study's practical impact for real-time wind-field forecasting and wind-farm optimization.

Abstract

Atmospheric flows are governed by a broad variety of spatio-temporal scales, thus making real-time numerical modeling of such turbulent flows in complex terrain at high resolution computationally intractable. In this study, we demonstrate a neural network approach motivated by Enhanced Super-Resolution Generative Adversarial Networks to upscale low-resolution wind fields to generate high-resolution wind fields in an actual wind farm in Bessaker, Norway. The neural network-based model is shown to successfully reconstruct fully resolved 3D velocity fields from a coarser scale while respecting the local terrain and that it easily outperforms trilinear interpolation. We also demonstrate that by using appropriate cost function based on domain knowledge, we can alleviate the use of adversarial training.

Enhancing wind field resolution in complex terrain through a knowledge-driven machine learning approach

TL;DR

The paper tackles the challenge of real-time, high-resolution wind-field prediction in complex terrain by coupling mesoscale and microscale simulations and delivering a 3D wind-field super-resolution model based on a physics-informed ESRGAN variant. It replaces perceptual loss with gradient- and divergence-based losses rooted in atmospheric physics, and finds that a non-adversarial, fully convolutional generator can outperform trilinear interpolation while maintaining physical consistency. The method leverages open data from HARMONIE-SIMRA, achieves a PSNR of 47.14 dB and a mean wind-vector error of 0.24 m/s, and supports upscaling factors up to 16×, enabling faster digital-twin-style analyses. The results demonstrate that physics-based losses can obviate the need for adversarial training, yielding stable, accurate reconstructions and enabling practical deployment for wind-energy applications in complex terrain. An open data release and discussion of digital-twin integration highlight the study's practical impact for real-time wind-field forecasting and wind-farm optimization.

Abstract

Atmospheric flows are governed by a broad variety of spatio-temporal scales, thus making real-time numerical modeling of such turbulent flows in complex terrain at high resolution computationally intractable. In this study, we demonstrate a neural network approach motivated by Enhanced Super-Resolution Generative Adversarial Networks to upscale low-resolution wind fields to generate high-resolution wind fields in an actual wind farm in Bessaker, Norway. The neural network-based model is shown to successfully reconstruct fully resolved 3D velocity fields from a coarser scale while respecting the local terrain and that it easily outperforms trilinear interpolation. We also demonstrate that by using appropriate cost function based on domain knowledge, we can alleviate the use of adversarial training.
Paper Structure (31 sections, 30 equations, 18 figures, 10 tables)

This paper contains 31 sections, 30 equations, 18 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Theory
  3. Meso- and microscale simulation model
  4. Generative adversarial networks
  5. Relativistic Average GANs
  6. Single Image Super-Resolution GANs
  7. Loss modification in this work
  8. Quantitative evaluation metrics
  9. Method and set-up
  10. Data generation and investigation
  11. Data preprocessing
  12. Model structure
  13. Generator
  14. Discriminator
  15. Hyperparameters
  16. ...and 16 more sections

Figures (18)

  • Figure 1: Area covered by the HARMONIE-SIMRA model, HARMONIE to the left, and SIMRA to the right, as it has been previously presented by Rasheed et. al. Rasheed2014amw
  • Figure 2: Shape of the dataset. Showing every fifth layer of data in the vertical coordinate. The $z$-coordinate has been multiplied by five to highlight the shape. Yellow marks high wind speed, and blue marks low wind speed.
  • Figure 3: Zoomed in unscaled wind field. Showing the 10 bottommost layers in a 32x32 slice of the data area. Yellow marks high wind speed, and blue marks low wind speed.
  • Figure 4: Turbulent Terrain effects. Showing the 16 bottommost layers in a chaotic region of the wind field.
  • Figure 5: Probability for pixels being included in a subsample using uniform and beta (0.25,0.25) distribution for subsample selection. The beta distribution shifts the probability towards the edges and is used here.
  • ...and 13 more figures