Super-resolution of near-surface wind speed in mountainous areas with machine learning

Near-surface wind speed in mountainous areas is estimated using a super-resolution (SR) model based on the hybrid downsampled skip-connection/multi-scale model (DSMC/MS) proposed by Fukami et al (JFM, Vol.870, 2019, pp. 106-120). The Reynolds-averaged Navier-Stokes (RANS) is implemented for flow fields in two mountainous areas (10 km x 10 km) to obtain wind speeds at 10 m above the ground as the training, validation and test data for the SR model. The low-resolution and high-resolution wind speeds are obtained by RANS at spatial resolutions of 160 m and 20 m, respectively. The training data are 128 x 128 grid points (2.56 km x 2.56 km) craped from the high-resolution wind speeds, and the input data of the low-resolution wind speeds are extracted at the corresponding areas. The SR model using the DSC/MS accurately estimates the high-resolution wind speeds when both the low-resolution wind speeds and the terrain elevation (the horizontal resolution of 20 m) are used as the model inputs. By comparison, traditional linear interpolation of the low-resolution wind speeds fails to reconstruct the high-resolution wind speed with high accuracy. For the test data at another site which has not been used in training, the SR model using the DSC/MS roughly captures the high-resolution wind speeds.