Estimation of near-surface wind speed over complex terrain with machine learning

High-resolution wind speeds near the ground surface over complex terrain are reconstructed by the low-resolution wind speeds using a machine-learned super-resolution (SR) model. The hybrid downsampled skip-connection/multi-scale model proposed by Fukami et al (JFM, Vol.870, 2019, pp. 106-120) are applied in the machine-learned SR model. The Reynolds-averaged Navier-Stokes is implemented for flow fields over complex terrains to obtain the high-resolution wind speeds (20 m resolution) at 10 m above the ground. The low-resolution wind speeds (160 m or 320 m resolution) were obtained by applying average pooling to high-resolution wind speeds. The SR model using only the low-resolution wind speeds as its input can not dramatically improve the reproducibility of the high-resolution wind speeds. However, the SR model using both the low-resolution wind speeds and the terrain elevation (the horizontal resolution of 20 m) as its inputs accurately estimates the high-resolution wind speeds. For the dataset at another site which has not been used in training, the SR model using the two inputs approximately reconstructs the high-resolution scalar velocities.