Data-driven ensemble model to statistically downscale rainfall using nonlinear predictor screening approach

The Artificial Intelligence (AI) models i.e., Artificial Neural Network (ANN) and Least Square Support Vector Machine (LSSVM) were used to statistically downscale and project rainfall data from CMIP5 General Circulation Models (GCMs) for Tabriz and Ardabil synoptic stations in north west Iran. Since one of the important issues in statistical downscaling of GCMs is to select dominant variables among plenty of large-scale climate data (predictors), a predictors screening framework, which integrates Wavelet-Entropy (WE) and Self-Organizing Map (SOM) was developed in this study to statistically downscale mean monthly rainfall values. The advantage of proposed method is to reduce the noise and dimensionality of data as well as selecting reliable inputs of downscaling model for future rainfall projection. To this end, five GCMs (i.e., Can-ESM2, BNU-ESM, CSIRO-ACCESS1, GFDL-ESM2G and INM-CM4) were employed. ANN, LSSVM and Multiple Linear Regression (MLR) models were trained to capture relationship between the predictors and the stations' observed rainfall values (predictand). Then, ensemble techniques were applied on the outputs of the downscaling models. The calibration, validation and projection of the proposed downscaling models were performed over the periods; Jan. 1951 to Dec. 1991, Jan. 1992 to Dec. 2005 and Jan. 2017 to Dec. 2100, respectively. The projection of rainfall for near and distant future (i.e., 2017-2050 and 2050-2100) by the proposed multi-GCM ensemble framework yielded to rainfall alteration pattern; 40-41% and 35-42% decrease at Tabriz station and 6-12% and 5-13% increase at Ardabil station under RCPs 4.5 and 8.5, respectively.