Comparison of MLP-ANN Scheme and SDSM as Tools for Providing Downscaled Precipitation for Impact Studies at Daily Time Scale

Abstract

Statistical downscaling has become an important part in most of the watershed scale climate change investigations. It is usually performed using multiple regression-based models. Basic working principle of such models is to develop a suitable relationship between the large scale (predictors) and the local climatic parameters called predictands. The development of such relationships using linear regression becomes very challenging when the local parameter to be downscaled is complex in nature such as precipitation. For this reason, use of nonlinear data driven techniques including Artificial Neural Networks (ANNs) is becoming more and more popular. Therefore, an attempt has been made in the study presented here to introduce a new Multi-Layer Perceptron (MLP) ANN-based scheme to develop a robust predictors-predictand relationship to be used as a downscaling model at daily time scale. The efficiency of this model has been compared with a popularly used model called Statistical Down Scaling Model (SDSM), for daily precipitation at the Clutha watershed in New Zealand. The results show that the model developed based on ANN scheme exhibits better performance than the SDSM. Hence, it is concluded that the use of artificial intelligence techniques such as ANN can greatly help in developing more efficient predictor-predictand models for even for precipitation being the toughest climate variable to model