Extreme Rainfall Prediction using Bayesian Quantile Regression in Statistical Downscaling Modeling

Abstract

Statistical Downscaling (SD) is a model that uses satellite data from General Circulation Models (GCM), which in climatology are very useful in predicting climate for the next few decades. GCM data is generally ill-conditioned, which is high dimension and multicollinearity, so a special technique is needed to handle this poorly conditioned. One of the variable selection techniques and handling of multicollinearity which is currently highly developed is regularization techniques including Adaptive Lasso, where selective parameters are adaptive, which can differ for each regression coefficient. Until now, predictions of extreme rainfall in Indonesia have not used Adaptive Lasso in SD modeling. This paper aims to predict the amount of rainfall (in millimeters) at moderate extreme (quantile 0.75) and high extreme rainfall (quantile 0.9 da 0.95) and handling poorly conditioned GCM data with Adaptive Lasso techniques and building predictive models of local rainfall by utilizing GCM data using the Bayes quantile regression model. Response in the form of monthly rainfall at Indramayu district, West Java Indonesia, and 49 explanatory variables in the form of GCM precipitation data in the period January 1981 - December 2013, which handled multicollinearity and variable selection using the Adaptive Lasso. The results are very satisfying with correlation between predicted and real data is above 0.91 for and the RMSEP is less than 50.