Statistical Downscaling with Gamma Distribution and Elastic Net Regularization

Case Study : Monthly Rainfall 1981-2013 at Indramayu

Abstract

Rainfall data are more than or equal zero and can be represented using Gamma distribution. In statistical downscaling the local scale rainfall data are used as the response variable to develop functional relation with global scale precipitation data of Global Circulation Model (GCM) output as the predictor variables. Generally, GCM output are multicollinear and regularization method can solve the problem. This paper develops a statistical downscaling model with the response of Gamma distribution using ridge regularizations and elastic net regularizations. Data are monthly rainfall in Indramayu at 1981-2013 and monthly precipitation data of GCM output in 1981-2013. The result shows that the elastic net (standard deviation of RMSEP value is 22.7 mm/month and standard deviation of correlation between actual and predicted value is 0.20 in four years) is more consistent than ridge regularization (standard deviation of RMSEP value is 35.7 mm/month and standard deviation of correlation between actual dan predict rainfall is 0.22 in four years) in predicting a next year rainfall.

Keywords: Gamma Distribution, Statistical Downscaling, Global Circulation Model, Ridge, Elastic Net