FORECASTING MONTHLY RAINFALL IN PANGKEP REGENCY USING STATISTICAL DOWNSCALING MODEL WITH ROBUST PRINCIPAL COMPONENT REGRESSION TECHNIQUE

Abstrak

A General Circulation Model (GCM) is a global climate model commonly used to predict local-scale climate patterns. However, the spatial resolution of GCMs is typically on a global scale, which is inadequate for predicting local climate. Statistical downscaling (SD) is used to transform climate information from a global scale to a smaller scale for local-scale climate predictions. GCM data have large dimensions and high correlations between grids, so principal component regression (PCR) is used in SD. The minimum covariance determinant (MCD) and minimum vector variance (MVV) methods are used in principal component analysis to obtain robust principal components (PCs). The data used in this study were the monthly rainfall data in Pangkep Regency for the period from January 1999 to December 2022 as the response variable, which were obtained from the Meteorology, Climatology, and Geophysics Agency (BMKG) Region IV Makassar. The predictor variable data were GCM precipitation data (64 variables) for the same period and three dummy variables. This study aimed to obtain rainfall forecasts in Pangkep Regency for the year 2023 based on a robust PCR model using results from MCD and MVV. The modeling results indicated that both the MCD and MVV methods provided similar model accuracy, with a coefficient of determination of approximately 91%. The PCR model with two PCs from the MVV method and dummy variables was identified as the best model for explaining the variability in rainfall data in Pangkep Regency. Additionally, the 2023 rainfall forecast results showed that both methods yielded relatively similar accuracy. The addition of dummy variables in the PCR model improved both the model accuracy and rainfall forecasts. The PCR model with three PCs from MVV and dummy principal component variables produced accurate rainfall forecasts based on a high correlation value (0.974) and the smallest mean absolute percentage error (7.290).