A Neural Network Architecture for Statistical Downscaling Technique : A Case Study in Indramayu District

Abstract

This research is focused on the development of statistical downscaling model using neural network technique to predict SOND rainfall in Indramayu. SST and rainfall data from multimodel ensemble outputs (derived from 18 ensemble members of ECHAM5 model under SRES A1B scenario) is used as predictor to predict SOND rainfall in each station. SST domains were selected by using cluster and correlation analyses, which were divided into three sets, namely SST lag 1 (August), lag 2 (July), and lag 3 (June). The Artificial Neural Network (ANN) employed in this study was multilayer perceptron with hidden layer as many as 5, 10, 20, and 40, and was trained with back propagation. The results show that the observed value lies between the maximum and minimum values of the predicted data. It is shown that the lagged SST provides better relationship with the observed data, and the optimum number of hidden neurons in neural networks is 5. Maximum correlation resulted from the models is 0.796 with an average of about 0.6. It is found that the prediction results tend to overestimate low rainfall and underestimate high rainfall found in the observed data.