Alsumaiei (2026) Physics-constrained neural network for daily pan evaporation forecasting in hyper-arid climates optimized by the Bat Algorithm

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-01-13
• Authors: Abdullah A. Alsumaiei
• DOI: 10.1016/j.jhydrol.2026.134936

Research Groups

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

Short Summary

This paper proposes a hybrid Physics-Constrained Neural Network (PCNN) optimized by the Bat Algorithm (BA) to accurately forecast daily pan evaporation in hyper-arid climates, demonstrating high predictive accuracy and physical consistency.

Objective

• To develop and validate a hybrid Physics-Constrained Neural Network (PCNN) framework, optimized by the Bat Algorithm (BA), for accurate and physically plausible daily pan evaporation forecasting in hyper-arid regions.

Study Configuration

• Spatial Scale: Two stations in Kuwait: Kuwait International Airport (KIA) and Abdaly.
• Temporal Scale: Daily.

Methodology and Data

• Models used: Physics-Constrained Neural Network (PCNN), Bat Algorithm (BA). The PCNN incorporates physical constraints based on surface energy balance theory (vapor pressure deficit, net radiation, aerodynamic resistance) into its loss function.
• Data sources: Daily meteorological data and Class A pan evaporation data from two observation stations (Kuwait International Airport and Abdaly).

Main Results

• The PCNN-BA model achieved high accuracy and good generalizability in forecasting daily pan evaporation.
• Root Mean Square Error (RMSE) values were approximately 1.046 × 10⁻⁸ m/s at Kuwait International Airport (KIA) and 1.373 × 10⁻⁸ m/s at Abdaly.
• Coefficient of Determination (R²) values were 0.953 at KIA and 0.884 at Abdaly.
• The model's consistency with thermodynamic principles was confirmed by a new metric, physics residual RMSE (PRMSE).
• The model demonstrated robustness against synthetic noise in inputs, indicating insensitivity to uncertainty.
• The proposed framework showed competitive predictive accuracy while explicitly enforcing physical consistency compared to established data-driven models.

Contributions

• Introduction of a novel hybrid framework (PCNN-BA) that integrates physical constraints into a neural network's loss function for pan evaporation forecasting, ensuring both accuracy and physical plausibility.
• Development of a new metric, physics residual RMSE (PRMSE), to confirm the model's consistency with thermodynamic principles.
• Demonstration of the model's robustness to input uncertainties, making it suitable for data-limited and water-scarce conditions.
• Provision of a computationally efficient and scalable tool directly applicable to water resources management, desert agriculture, and irrigation scheduling in hyper-arid environments.

Funding

• Not specified in the provided text.

Citation

@article{Alsumaiei2026Physicsconstrained,
  author = {Alsumaiei, Abdullah A.},
  title = {Physics-constrained neural network for daily pan evaporation forecasting in hyper-arid climates optimized by the Bat Algorithm},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.134936},
  url = {https://doi.org/10.1016/j.jhydrol.2026.134936}
}