Bernstein Quantile Networks for probabilistic downscaling of gridded extreme precipitation forecasts at hectometric scales

Probabilistic post-processing methods have over the last decades successfully been applied to enhance forecasts from numerical weather prediction (NWP) models. In this work, the Bernstein Quantile Networks (BQN) method is applied to gridded precipitation data for probabilistic downscaling of deterministic NWP model forecasts to hectometric-scale resolutions. In BQN, the predictive distribution is specified by a Bernstein polynomial whose coefficients are linked to input features by a neural network, enabling flexible distributional shapes to adequately represent the underlying forecast uncertainty. Models are trained using a quantile loss function that is extended to handle the point mass at zero through a censoring approach. Few restrictions on the predictive distribution combined with quantile loss make BQN applicable for more or less any target variable without modifications.

Forecast data from the Destination Earth initiative are here used to train BQN models. High-resolution, on-demand NWP runs targeting extreme events in Europe at hectometric resolutions (500m-750m) serve as the target, while coarse-resolution global IFS forecasts at 4.4 km resolution are used as inputs. As the forecast domains vary dynamically in response to anticipated extreme events across Europe, BQN models must generalise across diverse spatial and meteorological contexts. An additional complexity is that the amount of training data is limited. To address this, a few network architectures are explored and not least quantile loss variants, including adaptive weighting schemes and tail-focused penalties, to better capture extreme precipitation events with limited training data. The method provides fully specified marginal predictive distributions on grids. Possible extensions to generating scenarios by learning dependency structures are discussed.