Simulation of seasonal precipitation and raindays over Greece: a statistical downscaling technique based on artificial neural networks (ANNs)

A statistical downscaling technique based on artificial neural network (ANN) was employed for the estimation of local changes on seasonal (winter, spring) precipitation and raindays for selected stations over Greece. Empirical transfer functions were derived between large-scale predictors from the NCEP/NCAR reanalysis and local rainfall parameters. Two sets of predictors were used: (1) the circulation-based 500 hPa and (2) its combination along with surface specific humidity and raw precipitation data (nonconventional predictor). The simulated time series were evaluated against observational data and the downscaling model was found efficient in generating winter and spring precipitation and raindays. The temporal evolution of the estimated variables was well captured, for both seasons. Generally, the use of the nonconventional predictors are attributed to the improvement of the simulated results. Subsequently, the present day and future changes on precipitation conditions were examined using large-scale data from the atmospheric general circulation model HadAM3P to the statistical model. The downscaled climate change signal for both precipitation and raindays, partly for winter and especially for spring, is similar to the signal from the HadAM3P direct output: a decrease of the parameters is predicted over the study area. However, the amplitude of the changes was different. Copyright (c) 2006 Royal Meteorological Society