Abstract
Climate parameter projections obtained by global and regional models (GCM and RCM, respectively) offer a challenge to many researchers in terms of controlling the quality of the outcome data using several scales. In the literature, the proposed models, namely statistical downscaled and regression-based models, are mostly used to adjust the RCM data series. Contrariwise, in practice, these conceptual models perform poorly in certain cases and at certain scales. In this regard, a new downscaling model is proposed herein for annual rainfall projection, based on fusion models, namely polynomial regression (Poly_R), classification and regression tree (CRT), and principal component regression (PCR). The proposed model downscales the rainfall data projected by the coupled model intercomparison phase five (CMIP5) under different representative concentration pathway (RCP) scenarios (2.6, 4.5, 6.0, and 8.5) using overlapping data between the observation and the CMIP5 historical data. This process aims to define the framework for how to use the output equations and algorithm to correct data forecasting by RCM. Generally, the model can be summarized into three levels of analysis, starting with an iterative downscaling using a trendline model that is obtained by Poly_R fitting. Then, the CRT is used to classify and predict the data in subsets. Finally, multiple regression is given by a PCR model using principal components and standardized variables. The final model is also used to downscale the predicted data obtained by both previous models. The results provide the best performance of the fusion model in all RCP cases, compared to the delta change correction and linear scale models. This performance is proved by R2 scores which range between 0.87 and 0.95.