A neural network ensemble downscaling system (SIBILLA) for seasonal forecasts over Italy: winter case studies

Abstract

A novel statistical downscaling system for seasonal predictions is presented, based on an ensemble of neural networks with Bayesian regularization. The system SIBILLA (Statistical Integrated Bayesian Information system for Large to Local area Analysis) is able to take multiple predictor fields and/or time series as inputs. Gridded fields are compressed using empirical orthogonal functions, and a canonical correlation analysis is performed between predictors and each predictand. The first canonical variates are used as effective predictors in a neural network ensemble system. Final outputs for each parameter are expressed as a probability distribution for each station/grid point in the space of observations, as a result of the convolution of Gaussian mixtures. A first example of application in the Italian area is presented. An overall increase in skill score performances with respect to European Centre for Medium-Range Weather Forecasts (ECMWF) System 4 direct model output for the period 1981–2010 was found, even if probably not as high as desirable in a fully operational system.