Downscaling from Mesoscale to Microscale in Complex Terrain using a Compound Generative Adversarial Network

High resolution simulations of turbulent flow over complex terrain provide accurate wind energy resource assessments and inform turbine design requirements. Large Eddy Simulations (LES) can capture the fine scale features of turbulent flow; however, they are difficult to configure and prohibitively computationally expensive for use outside of the research community. Hence a compound Generative Adversarial Network (GAN) was developed to leverage the skill of an LES and produce a high resolution (30 m) flow field with plausible fine-scale turbulence features and characteristics when conditioned on a coarse, computationally-cheap mesoscale grid. Training data included flow and temperature fields from an LES in the vicinity of Mt. Hood in the Pacific Northwest with high-resolution topographical information. The trained compound GAN generates high-resolution simulations over a sizable domain (90 km X 180 km) in seconds using a single general purpose graphical processing unit as compared to the large high performance computing resources needed to produce the similar LES. GAN-generated winds from the test dataset compared well with the LES in terms of energy spectra, flow statistics, and visual inspection. Finally, our models were applied in a region of complex terrain significantly different from the training region. Statistical analysis of GAN output and LES in this independent secondary testing region established that our models were able to leverage the skill of LES in complex terrain for independent cases and regions. This result demonstrates that a deep learning approach to downscaling has the potential to greatly reduce cost and increase availability of valuable high-resolution data products.