A comparative study of convolutional neural network models for wind field downscaling

Kevin Höhlein, Michael Kern, Timothy Hewson, Rüdiger Westermann

Meteorol Appl. 2020; 27:e1961 · 18 August 2020 · ECMWF, Technical University of Munich (TUM)

In one sentence. Compares CNN architectures and proposes DeepRU, a U-Net variant, to downscale ERA5-driven 100 m wind forecasts from 31 km to 9 km over the Alps.

Abstract

We analyze the applicability of convolutional neural network (CNN) architectures for downscaling of short-range forecasts of near-surface winds on extended spatial domains. Short-range wind field forecasts (at the 100 m level) from ECMWF ERA5 reanalysis initial conditions at 31 km horizontal resolution are downscaled to mimic HRES (deterministic) short-range forecasts at 9 km resolution. We evaluate the downscaling quality of four exemplary model architectures and compare these against a multi-linear regression model. We conduct a qualitative and quantitative comparison of model predictions and examine whether the predictive skill of CNNs can be enhanced by incorporating additional atmospheric variables, such as geopotential height and forecast surface roughness, or static high-resolution fields, like land-sea mask and topography. We further propose DeepRU, a novel U-Net-based CNN architecture, which is able to infer situation-dependent wind structures that cannot be reconstructed by other models. Inferring a target 9 km resolution wind field from the low-resolution input fields over the Alpine area takes less than 10 milliseconds on our GPU target architecture, which compares favorably to an overhead in simulation time of minutes or hours between low- and high-resolution forecast simulations.

Labels

Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.