A hybrid generalised linear and Levenberg–Marquardt artificial neural network approach for downscaling future rainfall in North Western England

Abstract

This paper describes a novel technique for downscaling daily rainfall which uses a combination of a generalised linear model (GLM) and artificial neural network (ANN) to downscale rainfall. A two-stage process is applied, an occurrence process which uses the GLM model and an amount process which uses an ANN model trained with a Levenberg–Marquardt approach. The GLM-ANN was compared with other three downscaling models, the traditional neural network (ANN), multiple linear regression (MLR) and Poisson regression (PR). The models are applied for downscaling daily rainfall at three locations in the North West of England during the winter and summer. Model performances with respect to reproduction of various statistics such as correlation coefficient, autocorrelation, root mean square errors (RMSE), standard deviation and the mean rainfall are examined. It is found that the GLM-ANN model performs better than the other three models in reproducing most daily rainfall statistics, with slight difficulties in predicting extremes rainfall event in summer. The GLM-ANN model is then used to project future rainfall at the three locations employing three different general circulation models (GCMs) for SRES scenarios A2 and B2. The study projects significant increases in mean daily rainfall at most locations for winter and decreases in summer.