Publication
Wind modelling in complex topography remains a challenging task because operational Numerical Weather Prediction (NWP) models have a spatial resolution that is too coarse to capture all the meaningful terrain-induced phenomena like channeling, blocking, speed-ups, or thermally driven flows. In addition, higher resolution NWP models are so computationally expensive that it is not possible to generate long time series, which are often necessary for the study of wind-induced phenomena like snow deposition and transport or wind energy potential assessment. Downscaling schemes are often used to generate higher resolution wind fields from coarser resolution models. They can approximately be categorized as dynamic, which use nested NWP models or flow solvers, or as statistical, for which relevant predictors are used to interpolate the wind fields to a higher resolution grid. On the one hand, the computational cost of dynamic downscaling schemes does not allow to generate long time series at high spatial resolution. On the other hand, statistical downscaling suffers from the same inaccuracies as their driving coarse-resolution datasets. We aimed at developing a downscaling scheme with the following capabilities: I. Fast enough to generate long time series of high-resolution (50 m) wind fields for large domains (whole mountain range, 10000 sq. km); II. Utilize the accuracy of measurement station networks. III. Generate wind fields that exhibit terrain dependencies, and not a simple interpolation of the coarser dataset. Our novel downscaling model uses several convolutional neural networks that combine, at multiple spatial resolutions, 2D information from a NWP model and various topographic descriptors. This combination of information is performed at the location of hundreds of measurement stations, which data are used as predictands. When trained, this model can predict at any location the wind vector given the local atmospheric state from the NWP model and the local topography. For our test case, the Swiss Alps, we use data from the NWP model COSMO-1 and from 330 automatic measurement stations. The generated high-resolution wind fields show good agreement against observations and against high-resolution simulations from the Weather Research and Forecasting (WRF) model.