| dc.description.abstract | One of the newest forecasting techniques today is the Statistical Downscaling (SDs) technique. The SDs technique is a
procedure for inferring high-resolution information from low-resolution variables. Forecasting rainfall using the SDs
technique is to build a function that can predict the value of a response variable using predictor variables, for example,
the variables in the Global Circular Model (GCM). In this study, forecasting will be carried out using the Partial Least
Square (PLS) model and compared with the PLS model that has been time segmented namely the REBUS-PLS model.
We use four latent variables consisting of three exogenous latent variables and one endogenous latent variable. The
exogenous variable ξ1
is precipitation, ξ2
is air pressure, and ξ3
is temperature, while the endogenous variable is
monthly rainfall. The measurement model is a functional rule that describes the mathematical relationship between
exogenous latent variables ξ1
, ξ2
, and ξ3 with their corresponding manifests. After obtaining the structural model and
measurement model, then parameter estimation is carried out. The PLS model obtained was then tested for the goodness
of the model with several indicators, namely R2
, mean redundancy, and Goodness of Fit. The values obtained are
70.05%, 49.098%, and 76.11%. There are 4 segmentations which are segment 1 (33 months), segment 2 (29 months),
segment 3 (50 months), and segment 4 (32 months). The validity and reliability tests were carried out again in each
segment. Furthermore, the goodness of the model is also tested on each local model. The R-square values generated in
segment 1, segment 2, segment 3, and segment 4 are 97.13%, 97.52%, 85.05%, and 91.38%. Overall, the PLS model
has a smaller RMSE than the REBUS-PLS model at 25 observation stations. Meanwhile, at the other 52 observation
stations, the accuracy of the REBUS-PLS model is better than the PLS model. | en_US |
