Akademik Veri Yönetim Sistemi
Journal of Atmospheric and Solar-Terrestrial Physics, cilt.283, 2026 (SCI-Expanded, Scopus)
Accurate and reliable forecasting of hydrological drought is of critical importance for the sustainable management of water resources and the reduction of drought risks. In this study, the Streamflow Drought Index (SDI) was computed using monthly total streamflow records from the Aşağı Esence, Mavi, and Taşağıl streamflow observation stations located within the Konya Closed Basin, covering the period from January 1986 to September 2023. For the three stations, drought forecasting at 1-, 3-, 6-, 9-, and 12-month time scales was performed using Residual Empirical Mode Decomposition (REMD) and Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) as preprocessing techniques, in conjunction with Ridge Regression, Lasso Regression (LSS), Partial Least Squares Regression (PLS), and Quantile Regression Neural Network (QRNN) machine learning (ML) models. Accordingly, four standalone ML models and eight hybrid ML models were developed. The Partial Autocorrelation Function method was used to determine the input variables for these ML models, and lags at the 5% significance level were included. The performance of both standalone and hybrid ML models was evaluated using metrics such as Root Mean Square Error (RMSE), Overall Index, Nash–Sutcliffe Efficiency, Wilmott's refined index, Combined Accuracy, Relative Squared Error, Theil's Inequality Coefficient, and Ratio of RMSE, as well as scatter plots using REC curves, Taylor diagrams, and seasonal error boxplot analyses. The results showed that the hybrid ML models demonstrated clear superiority over the standalone ML models across all stations and time scales. The REMD–PLS hybrid model exhibited near-perfect prediction performance across all time scales, particularly at the Mavi and Taşağıl stations. The CEEMDAN–QRNN and REMD–QRNN models performed particularly well at short time scales, whereas the REMD–PLS, REMD–LSS, and REMD–QRNN models yielded more successful results in medium and long-time scales at the Aşağı Esence station. In general, as the time scale increased, prediction accuracy increased, while error levels systematically decreased. These findings demonstrate that REMD-based hybrid ML models are a powerful and reliable tool in hydrological drought monitoring, early warning systems, and long-term water resource planning.