Can neural networks outperform quantile mapping for post-processing seasonal weather forecast variables over the Alpine region?

The possibility to use seasonal weather forecasts is of paramount importance in hydrological and socio-economical applications. However, current seasonal weather forecasts from global numerical weather prediction (NWP) models inherit systematic biases resulting from inaccurate representation and parameterization of local to global scale environmental processes. Therefore, the hydrological community frequently uses the quantile mapping (QM) statistical postprocessing for bias correction and downscaling of the meteorological inputs (i.e., daily precipitation and temperature) to hydrological models. The QM often assumes a linear and static relationship between quantiles of observed and simulated data over time. These limitations can be relaxed by employing a Neural Network (NN) based postprocessing method. In this context, the objective of this study is to compare the accuracy of QM and NN statistical postprocessing of ensemble seasonal weather forecasts over the Trentino-South Tyrol region (north-eastern Italian Alps), characterised by complex topography. 

The study uses the latest fifth-generation seasonal weather forecast system (SEAS5) total precipitation and 2m-temperature dataset produced by European Centre for Medium-Range Weather Forecast (ECMWF), available at a horizontal grid resolution of 0.125° x 0.125° with 25 ensemble members in a re-forecast period from 1981 to 2016. The respective reference dataset is a high-resolution gridded observation (250 m x 250 m) over the region of interest. The QM method derives a functional relationship between the variable of interest and the corresponding predictor, whereas the NN based methods can be used with a set of predictors to learn the linear and non-linear relationships in a data-driven way.

The analysis is divided into training (1981 – 2010, 30 years) and testing (2011 – 2016, 6 years) period to compare the cumulative ranked probability scores (CRPS) of both the statistical postprocessing methods. The statistical postprocessing is implemented univariately on the daily dataset (2m temperature and precipitation) over a month for each lead time. The raw forecasts and postprocessed forecasts are compared with particular focus on the effects of the forecast lead time and location, as well as diurnal and seasonal cycles in forecast performance. The postprocessed forecasts revealed a significant improvements compared to the raw forecasts.