Application of machine learning ensemble models for rainfall prediction

Practical information can be drawn from rainfall for making long-term water resources management plans, taking flood mitigation measures, and even establishing proper irrigation systems. Given that a large amount of data with high resolution is required for physical modeling, this study proposes a new standalone sequential minimal optimization (SMO) regression model and develops its ensembles using Dagging (DA), random committee (RC), and additive regression (AR) models (i.e., DA-SMO, RC-SMO, and AR-SMO) for rainfall prediction. First, 30-year monthly data derived from the year 1988 to 2018 including evaporation, maximum and minimum temperatures, maximum and minimum relative humidity rates, sunshine hours, and wind speed as input and rainfall as the output were acquired from a synoptic station in Kermanshah, Iran. Next, based on the Pearson correlation coefficient (r-value) between input and output variables, different input scenarios were formed. Then, the dataset was separated into three subsets: development (1988–2008), calibration (2009–2010), and validation (2011–2018). Finally, the performance of the developed algorithms was validated using different visual (scatterplot and boxplot) and quantitative (percentage of BIAS, root mean square error, Nash–Sutcliffe efficiency, and mean absolute error) metrics. The results revealed that minimum relative humidity had the greatest effect on rainfall prediction. The most effective input scenario featured all the input variables except for wind speed. Our findings indicated that the DA-SMO ensemble algorithm outperformed other algorithms.