A comparison of the principal component regression methods and the robust principal component regression with minimum vector variance in statistical downscaling models

Abstract

Principal component regression (PCR) is commonly used in statistical downscaling (SD) models to forecast local-scale rainfall data based on global-scale rainfall. The global circulation model (GCM) is one of the global scale climate models available in grid form and can improve the accuracy of rainfall forecasting. Furthermore, minimum vector variance (MVV) is used in PCR models to produce a variance-covariance matrix that is robust against outliers. This study aims to compare the results of the estimated local rainfall data in Pangkep Regency obtained from the classic PCR with the PCR with MVV. Dummy variables of the K-means cluster technique are added to overcome the heterogeneous residual variance. By using the Pangkep district rainfall data for the 1997-2017 period, it was found that the addition of a dummy variable produces a better accuracy than the model without dummy variables. The classic PCR model with dummy variables has an R2 value of 92.44%-92.86% and the root mean square error (RMSE) is 81.80-83.75. While the robust PCR model with MVV yield an R2 value of 92.42%-92.76% and the RMSE value is 82.41-83.79. Based on the results of model validation (data period 2018), the robust PCR model with MVV is a better than the Classical PCR model. Robust PCR model with MVV and involving 5 main components is able to generate more accurate predictions of rainfall data.