Shahnazi et al. (2026) A novel implementation of a decomposition-enhanced hybrid GWO–KELM model with LUBE for constructing prediction intervals of groundwater drought

Identification

• Journal: Earth Science Informatics
• Year: 2026
• Date: 2026-03-01
• Authors: Saman Shahnazi, Kiyoumars Roushangar, Armin Farshbaf, Hossein Hashemi
• DOI: 10.1007/s12145-026-02093-y

Research Groups

• Department of Water Resources Engineering, Faculty of Civil Engineering, University of Tabriz, Iran
• Water Sciences and Hydroinformatics Research Center, Khazar University, Azerbaijan
• Division of Water Resources Engineering, Faculty of Engineering, Lund University, Sweden
• Centre for Advanced Middle Eastern Studies, Lund University, Sweden
• United Nations University Hub on Water in a Changing Environment (WICE) at Lund University, Environment and Health (UNUINWEH), United Nations University Institute for Water, Sweden

Short Summary

This study developed a novel decomposition-enhanced hybrid Grey Wolf Optimizer (GWO)–Kernel Extreme Learning Machine (KELM) model with Lower–Upper Bound Estimation (LUBE) for multi-horizon point and interval forecasting of groundwater drought (Standardized Groundwater Index, SGI). The Variational Mode Decomposition (VMD)–GWO–KELM model consistently outperformed other approaches, especially for short-term forecasts, providing reliable and sharp prediction intervals.

Objective

• To develop and evaluate a novel decomposition-enhanced hybrid Grey Wolf Optimizer (GWO)–Kernel Extreme Learning Machine (KELM) model, integrated with the Lower–Upper Bound Estimation (LUBE) framework, for multi-horizon (3, 9, and 12 months) point and interval forecasting of the Standardized Groundwater Index (SGI) in a semi-arid aquifer.
• To account for hydrogeological heterogeneity by clustering piezometric wells based on their drought-response behaviors using Self-Organizing Maps (SOM).
• To systematically compare the effectiveness of various signal decomposition techniques (CEEMDAN, VMD, EWT, EFD) in enhancing forecasting accuracy and uncertainty quantification.

Study Configuration

• Spatial Scale: Marand Plain, East Azerbaijan Province, Iran. Analysis of 20 piezometric wells in an unconfined aquifer.
• Temporal Scale: Monthly groundwater level records spanning 25 years (April 1996 to March 2021). Forecasting horizons included 3-month (short-term), 9-month (mid-term), and 12-month (long-term) lead times. Data split: 70% for training (1996–2012) and 30% for testing (2013–2021).

Methodology and Data

• Models used:
  • Hybrid Forecasting Model: Grey Wolf Optimizer (GWO) optimized Kernel Extreme Learning Machine (KELM).
  • Signal Decomposition Techniques: Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), Variational Mode Decomposition (VMD), Empirical Wavelet Transform (EWT), Empirical Fourier Decomposition (EFD).
  • Clustering Algorithm: Self-Organizing Maps (SOM).
  • Uncertainty Quantification: Lower–Upper Bound Estimation (LUBE) framework.
  • Drought Index Calculation: Standardized Groundwater Index (SGI) derived using nonparametric Kernel Density Estimation (KDE).
• Data sources: Monthly groundwater level records from 20 piezometric wells, provided by the Water Resources Management Organization (WRMO) of Iran.

Main Results

• Self-Organizing Maps (SOM) clustering successfully classified the 20 piezometric wells into 5 distinct drought-response zones, with an optimal Silhouette Coefficient (SC) of 0.68, revealing spatial heterogeneity in groundwater drought patterns.
• Decomposition-enhanced models significantly improved short-term (3-month) point forecasting accuracy compared to the standalone GWO–KELM model, with paired t-test results confirming statistical significance (p < 0.05).
• The VMD–GWO–KELM model consistently demonstrated superior point forecasting performance across most wells and horizons, particularly for short-term predictions (e.g., Well P8: Determination Coefficient (DC) = 0.943, Root Mean Square Error (RMSE) = 0.248 for 3-month lead-time).
• For wells exhibiting strong monotonic negative drought trends, preprocessing methods were less effective, sometimes reducing forecasting accuracy due to over-decomposition of noise and under-extraction of meaningful modes.
• The LUBE framework, integrated with hybrid models, successfully constructed reliable prediction intervals (PIs). The VMD–GWO–KELM model achieved an optimal trade-off between Prediction Interval Coverage Probability (PICP) and Normalized Mean Prediction Interval Width (NMPIW) for short- and mid-term forecasts.
• The EWT–GWO–KELM model demonstrated superior long-term reliability in constructing PIs.
• The study highlighted the critical importance of using complementary performance metrics like Kling–Gupta Efficiency (KGE) and Bias Factor (BF) alongside DC and RMSE, as conventional metrics alone could mask significant model failures in reproducing observed data characteristics.

Contributions

• Novel integration of the Grey Wolf Optimizer (GWO) and Kernel Extreme Learning Machine (KELM) with multiple advanced signal decomposition techniques (CEEMDAN, VMD, EWT, EFD) for multi-horizon point forecasting of the Standardized Groundwater Index (SGI).
• First-time application of Self-Organizing Maps (SOM) for clustering piezometric wells based on their drought-response behaviors, effectively addressing hydrogeological heterogeneity in groundwater drought forecasting.
• Innovative integration of the Lower–Upper Bound Estimation (LUBE) framework with the GWO-KELM model for direct estimation and optimization of prediction intervals, providing robust uncertainty quantification for groundwater drought forecasts.
• Comprehensive comparative analysis of four distinct signal decomposition techniques, identifying VMD as the most effective for enhancing point forecast accuracy and EWT for long-term prediction interval reliability in groundwater drought modeling.
• Emphasized the necessity of employing a diverse set of performance metrics (including KGE and BF) for a more reliable and holistic evaluation of groundwater drought forecasting models, beyond traditional correlation and error measures.

Funding

The authors declare that no funds, grants, or other support were received during the preparation of this manuscript.

Citation

@article{Shahnazi2026novel,
  author = {Shahnazi, Saman and Roushangar, Kiyoumars and Farshbaf, Armin and Hashemi, Hossein},
  title = {A novel implementation of a decomposition-enhanced hybrid GWO–KELM model with LUBE for constructing prediction intervals of groundwater drought},
  journal = {Earth Science Informatics},
  year = {2026},
  doi = {10.1007/s12145-026-02093-y},
  url = {https://doi.org/10.1007/s12145-026-02093-y}
}