Towards Operational Downscaling of Low Resolution Wind Fields using Neural Networks

Numerical weather prediction models with high resolution (of order kms or less) can deliver very accurate low-level winds. The problem is that one cannot afford to run simulations at very high resolution over global or other large domains for long periods because the computational power needed is prohibitive.

Instead, we propose using neural networks to downscale low-resolution wind-field simulations (input) to high-resolution fields (targets) to try to match a high-resolution simulation. Based on short-range forecasts of wind fields (at the 100m level) from the ECMWF ERA5 reanalysis, at 31km resolution, and the HRES (deterministic) model version, at 9km resolution, we explore two complementary approaches, in an initial “proof-of-concept” study.

In a first step, we evaluate the ability of U-Net-type convolutional neural networks to learn a one-to-one mapping of low-resolution input data to high-resolution simulation results. By creating a compressed feature-space representation of the data, networks of this kind manage to encode important flow characteristics of the input fields and assimilate information from additional data sources. Next to wind vector fields, we use topographical information to inform the network, at low and high resolution, and include additional parameters that strongly influence wind-field prediction in simulations, such as vertical stability (via the simple, compact metric of boundary layer height) and the land-sea mask. We thus infer weather-situation and location-dependent wind structures that could not be retrieved otherwise.

In some situations, however, it will be inappropriate to deliver only a single estimate for the high-resolution wind field. Especially in regions where topographic complexity fosters the emergence of complex wind patterns, a variety of different high-resolution estimates may be equally compatible with the low-resolution input, and with physical reasoning. In a second step, we therefore extend the learning task from optimizing deterministic one-to-one mappings to modelling the distribution of physically reasonable high-resolution wind-vector fields, conditioned on the given low-resolution input. Using the framework of conditional variational autoencoders, we realize a generative model, based on convolutional neural networks, which is able to learn the conditional distributions from data. Sampling multiple estimates of the high-resolution wind vector fields from the model enables us to explore multimodalities in the data and to infer uncertainties in the predictand.

In a future customer-oriented extension of this proof-of-concept work, we would envisage using a target resolution higher than 9km - say in the 1-4km range - to deliver much better representivity for users. Ensembles of low resolution input data could also be used, to deliver as output an “ensemble of ensembles”, to condense into a meaningful probabilistic format for users. The many exciting applications of this work (e.g. for wind power management) will be highlighted.