Abstract

State-of-the-art random forest (RF) models have been documented as versatile tools to solve regression and classification problems in hydrology. They can model stochastic time series by bagging different decision trees. This article introduces a new hybrid RF model that increases the forecasting accuracy of RF-based models. The new model, called GARF, is attained by integrating genetic algorithm (GA) and hybrid random forest (RF), in which different decision trees are bagged. We applied GARF to model and forecast a multitemporal drought index (SPEI-3 and SPEI-6) at two meteorology stations (Beypazari and Nallihan) in Ankara, Turkey. We compared the associated results with classic RF, standalone extreme learning machine (ELM), and a hybrid ELM model optimized by Bat algorithm (Bat-ELM) to verify the new model accuracy. The performance assessment was performed using graphical and statistical analysis. The forecasting results demonstrated that the GARF outperformed the benchmark models. GARF achieved the least error in a quantitative assessment for the prediction of both SPEI-3 and SPEI-6, particularly in the testing period. The results of this study showed that the new model can improve the forecasting accuracy of the classic RF technique up to 30% and 40% at Beypazari and Nallihan stations, respectively.