Comparison of wavelet and empirical mode decomposition hybrid models in drought prediction

Drought is a major area of interest within the field of water resources management, agriculture, energy resources and community health. Recently researchers have examined not only the mathematical expression of drought indices but also statistical predictions. Accordingly, high accuracy results were obtained using stand-alone machine learning techniques such as artificial neural networks (ANN) and support vector machine (SVM). However, lately, hybrid models have been introduced, which are created by integrating different time series decomposition techniques into standalone models, since the accuracy of stand-alone models used in the drought prediction being low particularly for mid-term and long-term drought predictions. In this study, self-calibrated Palmer Drought Severity Index (sc-PDSI) values were predicted by using three different standalone models and six hybrid models which are performed by two different decomposition techniques, such as Empirical mode decomposition (EMD) and Wavelet decomposition (WD). The main purpose of this study is to evaluate the effect of using EMD and WD for decomposing time series into their sub-bands on drought prediction. sc-PDSI time series were used to achieve 1, 3 and 6-month lead time predictions for Adana and Antalya cities located in the southern part of Turkey. Model performance indicators such as mean square error (MSE), Nash-Sutcliffe efficiency coefficient (NSE) and determination coefficient (R-2) were employed to compare the proposed models. The results revealed that the accuracy of the stand-alone models, particularly in mid-term sc-PDSI predictions, was unsatisfactory. However, the prediction accuracy has been increased significantly with the introduction of EMD and WD techniques. Considering the Adana region, the hybrid wavelet models outperformed the models hybridized by EMD for not only 1-month lead time (NSEWD-ANFIS = 0.981 and NSEEMD-M5 = 0.890), but also for 3 month (NSEWD-SVM = 0.878 and NSEEMD-ANFIS = 0.811) and 6-month (NSEWD-ANFIS = 0.857 and NSEEMD-ANFIS = 0.783) lead times. According to the obtained results for Antalya region, similar findings were also observed among hybrid models. Thus, it is concluded that the predictions made using WD have higher accuracy than EMD, and the correct wavelet type selection has a significant effect on the results.