Adversarial image-to-image model to obtain highly detailed wind fields from mesoscale simulations in urban environments

We propose a conditional Generative Adversarial Network (cGAN) that can produce detailed local wind fields in urban areas, comparable in level of detail to those from Computational Fluid Dynamics (CFD) simulations, that are generated from coarser Numerical Weather Prediction (NWP) data. In our approach, the cGAN is trained using NWP data as input and CFD as targets. Both CFD and NWP data are presented to the network as images, using an image-to-image model based on Pix2Pix to transform coarse meteorological conditions into detailed local wind fields. The methodology is tested in a residential district in a large Spanish city, Zaragoza. The model predictions show significant agreement with the actual CFD results, while reducing the computational time from eight hours to seconds. Feature engineering of image channels effectively reduces the model error, especially in the wind direction, achieving a mean absolute error in the wind speed of 0.35m/s and a wind direction error of 27.0°. •Image-to-image cGAN model translates NWP to CFD-level wind data.•Model reduces simulation time from hours to seconds.•Feature engineering on image channels improves accuracy.