Improving drought modeling based on new heuristic machine learning methods
Rana Muhammad Adnan, Hongliang Dai, Alban Kuriqi, Özgür Kişi, Mohammad Zounemat‐Kermani
Abstract
Drought modeling is vital for designing and managing water resource systems due to its significant effects on agriculture and other components of the environment. This study evaluates the prediction accuracy of two newly developed heuristic methods, optimally pruned extreme learning machine (OP-ELM) and dynamic evolving neural-fuzzy inference system (DENFIS) in drought modeling based on four standard precipitation indexes (SPI), i.e., SPI3, SPI6, SPI9 and SPI12 calculated for three meteorological stations of Pakistan. The prediction capability of the two methods is tested using the different time lag input combinations of each drought index and compared with multivariate adaptive regression spline (MARS). For the SPI3, SPI6, and SPI12 drought indexes, it was found that the DENFIS model provided better accuracy than the OP-ELM and MARS models. However, in the case of SPI6, the OP-ELM model performed better than the other models. The effect of periodicity on the prediction accuracy of the models is also evaluated. It was found that adding periodicity as inputs for models generally gave good forecasting results; however, importing this component negatively affects models’ accuracy in some cases. The best DENFIS model improved the accuracy of the OP-ELM by 8.35%, 6.06% and 2.51% predicting SPI3, SPI6, and SPI12 of Drosh Station with respect to root mean square error criterion, respectively. This study also demonstrated that the DENFIS heuristic models which have not been previously used for this issue generally provided more accurate predictions than the other models in drought prediction.