Khalil et al. (2025) A novel optimized machine learning ensemble approach for future drought assessment

Identification

• Journal: Journal of Hydrology
• Year: 2025
• Date: 2025-12-12
• Authors: Rashida Khalil, Zulfiqar Ali
• DOI: 10.1016/j.jhydrol.2025.134782

Research Groups

• College of Statistical Sciences, University of the Punjab, Lahore, Pakistan
• Department of Statistics, Lahore College for Women University, Lahore, Pakistan

Short Summary

This study proposes a novel multi-model ensemble (MME) framework integrating Learning Vector Quantization (LVQ) and Optimized Learning Vector Quantization (OLVQ) to enhance the reliability of precipitation forecasts from General Circulation Models (GCMs). The findings demonstrate that both LVQ and OLVQ significantly outperform traditional ensemble methods, with OLVQ providing further substantial improvements in reducing prediction errors and increasing correlation.

Objective

• To improve the reliability of precipitation forecasts from General Circulation Models (GCMs) by developing a novel multi-model ensemble (MME) framework that integrates Learning Vector Quantization (LVQ) and Optimized Learning Vector Quantization (OLVQ) to dynamically assign weights to individual models based on their performance.

Study Configuration

• Spatial Scale: Tibetan Plateau (TP) region, across 32 meteorological stations.
• Temporal Scale: Historical monthly precipitation simulations from 1961–2014.

Methodology and Data

• Models used: Learning Vector Quantization (LVQ), Optimized Learning Vector Quantization (OLVQ), Multi-model ensemble (MME) framework. Performance was compared against traditional ensemble methods (e.g., average (AVE) and coefficient of variation (COE) based weighting).
• Data sources: Historical monthly precipitation simulations from 18 General Circulation Models (GCMs) under Coupled Model Intercomparison Project Phase 6 (CMIP6). Meteorological station observations were used for performance evaluation.

Main Results

• Both LVQ and OLVQ approaches substantially outperformed traditional ensemble methods for precipitation forecasting.
• The LVQ approach reduced Normalized Root Mean Square Error (NRMSE) by 72 % (0.1383 ± 0.0522) and Normalized Relative Absolute Error (NRAE) by 80 % (0.5967 ± 0.2080) when compared with the average (AVE) method.
• The OLVQ approach provided additional improvements, achieving 13.6 % lower NRMSE than LVQ (0.1196 ± 0.0352) and 54 % lower than COE (0.2576 ± 0.3327).
• OLVQ also demonstrated a 22 % higher correlation than the AVE approach.

Contributions

• Proposes a novel multi-model ensemble (MME) framework that integrates Learning Vector Quantization (LVQ) for dynamically assigning weights to individual GCMs based on their performance.
• Introduces an advanced variant, Optimized Learning Vector Quantization (OLVQ), which incorporates a metaheuristic optimization algorithm to further refine the weighting mechanism, minimize prediction errors, and enhance model performance discrimination.
• Demonstrates significant improvements in the reliability of precipitation forecasts compared to conventional ensemble methods, offering a promising approach for more informed decision-making in future climate projections.

Funding

• Not explicitly mentioned in the provided paper text.

Citation

@article{Khalil2025novel,
  author = {Khalil, Rashida and Ali, Zulfiqar},
  title = {A novel optimized machine learning ensemble approach for future drought assessment},
  journal = {Journal of Hydrology},
  year = {2025},
  doi = {10.1016/j.jhydrol.2025.134782},
  url = {https://doi.org/10.1016/j.jhydrol.2025.134782}
}