Statistical downscaling technique using partial least square multi response for monthly rainfall forecasting

Abstract

Rainfall is one of the climate components that is often used as a reference, especially in agriculture. This thing cannot be separated from the fact that cropping pattern planning will need to pay attention to the amount of rainfall in the future. This research aims to build a rainfall forecasting model at 77 station points in Jember Regency using Statistical Downscaling (SD) technique. SD technique is an effort to connect the circulation of global scale variables (explanatory variables) and local scale variables. ln this research, the global scale variable used is Global Circular Model (GCM) with 3 variables which are precipitation, air temperature, and sea level pressure. The local scale variable used is monthly rainfall data in Jember Regency at 77 station points. The models used are Partial least Square (PLS) multi response, PLS single response, and Principal Component Regression (PCR) models. PLS single response gives a more accurate result compared to other models. This is indicated by the RMSEP value in PLS single response, PLS multi response, and PCR are 64.75, 66.41; and 67.11. But overall, the PLS multi response is more effective and efficient as a functional model for Statistical Downscaling to predict rainfall in Jember Regency. Because the running time on the PLS multi response is shorter than the other two models. PLS multi response provides forecast results at 77 stations in one process, while PLS single response and PCR require 77 processes to provide forecasts at 77 stations.