Alsumaiei (2026) Complexity-efficiency dynamics of metaheuristic-optimized recurrent neural network models for drought forecasting in hyper-arid Kuwait

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2026
• Date: 2026-03-03
• Authors: Abdullah A. Alsumaiei
• DOI: 10.1016/j.ejrh.2026.103300

Research Groups

Department of Civil Engineering, College of Engineering and Petroleum (COEP), Kuwait University, Kuwait.

Short Summary

This study develops and benchmarks metaheuristic-optimized recurrent neural network models (LSTM, GRU) for drought forecasting in hyper-arid Kuwait using the distribution-free Precipitation Index (PI12, PI24), finding that longer aggregation windows enhance stability and that compact architectures often achieve comparable accuracy to more complex, optimized models with greater efficiency.

Objective

• To develop an accurate and operationally practical drought forecasting tool for hyper-arid environments.
• To clarify the complexity-efficiency trade-off in metaheuristic-optimized recurrent neural network models for drought forecasting using the Precipitation Index (PI) in hyper-arid Kuwait.

Study Configuration

• Spatial Scale: Kuwait (hyper-arid region in the Arabian Peninsula, Western Asia), focusing on data from a single representative station (Kuwait International Airport).
• Temporal Scale: Monthly precipitation data from 1958 to 2017 (60 years) for drought index calculation. Drought forecasting at 12-month (PI12) and 24-month (PI24) accumulation scales, with a one-step-ahead (one month) forecasting horizon.

Methodology and Data

• Models used:
  • Long Short-Term Memory (LSTM) neural networks (baseline and metaheuristic-optimized variants).
  • Gated Recurrent Unit (GRU) neural network (compact benchmark).
  • Metaheuristic optimization algorithms: Bat Algorithm (BA), Ant Colony Optimization (ACO), and Grey Wolf Optimization (GWO) for LSTM hyperparameter tuning (number of hidden units, learning rate, dropout rate).
  • Adam optimization algorithm for network training.
• Data sources:
  • Long-term monthly precipitation data from Kuwait International Airport (KIA).
  • Precipitation Index (PI) at 12-month (PI12) and 24-month (PI24) accumulation scales, derived from precipitation data.

Main Results

• All recurrent neural network configurations successfully captured the dominant temporal structure of the Precipitation Index (PI) series.
• Forecasts based on the 24-month PI (PI24) consistently exhibited enhanced phase alignment and reduced high-frequency variability compared to PI12, demonstrating the stabilizing effect of longer temporal aggregation.
• For PI12, the ACO-optimized LSTM achieved the lowest Root Mean Square Error (RMSE) of 0.1887 and the highest coefficient of determination (R²) of 0.796, closely followed by the GRU (RMSE 0.1896, R² 0.794).
• For PI24, performance converged across most models, with the baseline LSTM and BA-optimized LSTM both achieving an RMSE of 0.0939 and R² of 0.853. ACO and GRU models showed very similar performance (RMSE 0.0941–0.0944, R² 0.851–0.852).
• The Grey Wolf Optimizer (GWO) often resulted in models with substantially higher complexity (e.g., 41,200 trainable parameters for PI12) without delivering commensurate accuracy gains, and sometimes showed lower accuracy (e.g., PI24 GWO RMSE 0.1020, R² 0.826).
• While some pairwise performance differences were statistically significant, the numerical magnitude of these differences was small, especially for PI24, indicating marginal practical gains from increased model complexity.
• Compact recurrent architectures, such as the GRU (3162–5934 trainable parameters) and smaller LSTM configurations (e.g., ACO-optimized LSTMs with 4216–7912 parameters), achieved near-optimal predictive skill with significantly lower computational cost and parameter footprint.
• The average annual precipitation in Kuwait is approximately 0.110–0.120 meters, with a long-term mean of 0.1075 meters at Kuwait International Airport.

Contributions

• Introduces and benchmarks metaheuristic-optimized recurrent neural network models for drought forecasting in hyper-arid Kuwait using the Precipitation Index (PI), a deterministic and distribution-free drought indicator well-suited for zero-inflated rainfall conditions.
• Provides a systematic evaluation of the complexity–efficiency trade-off in drought forecasting models by quantifying both accuracy and computational cost (trainable parameters, training time), offering practical insights for operational deployment.
• Compares multiple metaheuristic optimization algorithms (ACO, BA, GWO) for LSTM hyperparameter tuning under identical data partitions and search budgets, alongside non-optimized baseline LSTM and compact GRU models.
• Demonstrates that temporal aggregation (e.g., PI24) is the dominant factor for forecast stability and fidelity in hyper-arid regions, and that parsimonious recurrent architectures can achieve comparable performance to more complex, heavily optimized models.

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Citation

@article{Alsumaiei2026Complexityefficiency,
  author = {Alsumaiei, Abdullah A.},
  title = {Complexity-efficiency dynamics of metaheuristic-optimized recurrent neural network models for drought forecasting in hyper-arid Kuwait},
  journal = {Journal of Hydrology Regional Studies},
  year = {2026},
  doi = {10.1016/j.ejrh.2026.103300},
  url = {https://doi.org/10.1016/j.ejrh.2026.103300}
}