Downscaling of Wind Predictions in Complex Terrain Using Bayesian Inference

Fine-resolution, short-range real-time forecasting of atmospheric boundary layer flow over complex terrain demands substantial computational resources and skilled operators. Statistical downscaling from available coarse-resolution meteorological forecasts to a relevant finer scale simplifies the procedure and shortens the computation time by orders of magnitude.

A statistical downscaling experiment was performed for winds throughout the mixed layer depth above a complex terrain in Israel's northern Negev desert region. Dynamical downscaling was performed by the WRF model to create an historical database comprised of pairs of coarse (4.5 km) and fine (0.5 km) resolution forecasts, each pair referring to a specific time. The coarser-resolution data were defined as predictors while the finer-resolution data served as predictands for statistical downscaling.

The database was used to test two statistical downscaling algorithms: the commonly used minimal distance Analog and a novel Bayesian inference aided analog (hereafter Bayesian algorithm). Both algorithms are based on searching in the database for coarse resolution analogs of a current coarse resolution forecast, and suggesting the equivalent finer-resolution past forecast as the current best forecast. The Bayesian algorithm suggested here is unique in the way it defines the analogy. instead of referring to minimal differences between predictors, it calculates the Bayesian conditional probability of a given past forecast to be the optimal analog. The comparison of the two algorithms shows that the Bayesian approach outperforms the traditional approach.

The Bayesian algorithm reproduces the fine resolution dynamically-downscaled surface winds with an average absolute direction difference of 43 degrees and 20 degrees for calm winds and moderate/strong winds respectively. for higher altitudes inside the mixed layer the error grows only by 10-20 degrees. The whole procedure is extremely fast (a few seconds) and easy to operate, which makes it suitable for real-time non-expert fast-response applications.