Indian summer monsoon rainfall prediction: A comparison of iterative and non-iterative approaches

Abstract

Accurate forecasting of Indian summer monsoon rainfall (ISMR) is highly desirable, however it is one of the toughest challenges the scientific community is facing today. Among the various scientific approaches, statistical, empirical, and numerical modeling are popular and widely used in the recent time for monsoon forecast, whereas for some reason soft computing has not yet gained any such attention, which although is a potential approach and may be able to solve this problem. Thus, it is highly desirable to examine the performance of existing soft computing techniques and based on the outcomes accordingly make improvements and/or design and develop new ones for the issue in question. With this motivation, in this study we examine the performance of single layer feed-forward neural network (SLFN) in ISMR forecast and for the first time apply random vector functional link network (RVFL) and regularized online sequential network − RVFL (ROS-RVFL) for the forecast purpose and subsequently examine their relative performance with respect to SLFN. Seasonal forecast of ISMR is made a year in advance using the previous year’s data. Six sets of forecasts are made, by varying the length of the past data (training period), ranging from 5 to 10 years. It is found that ROS-RVFL is more accurate and computationally more efficient than SLFN and RVFL. Secondly, irrespective of the techniques, when we consider the last 8 years of past data for the training purpose the result is found to be relatively more accurate, and the ensemble mean of forecast with 8 and 9 years of training period gives the best result among all the trials. Based on statistical and detail analyses finally it is concluded that ROS-RVFL with ensemble mean of 8–9 years training period is a promising approach for ISMR forecast.