Suresh et al. (2026) Physics-Guided Deep Learning with Bayesian Optimization for Enhanced River Streamflow Prediction

Identification

• Journal: Water Resources Management
• Year: 2026
• Date: 2026-02-16
• Authors: Y. Suresh, M. Sahaya Sheela, Pamarthi Sunitha, Saisubramaniam Gopalakrishnan
• DOI: 10.1007/s11269-026-04520-7

Research Groups

• Department of Computer Science and Engineering, PSNA College of Engineering and Technology, Dindigul, Tamil Nadu, India
• Department of Electronics and Communication Engineering, Vel Tech Rangarajan Dr.Sagunthala R&D Institute of Science and Technology, Chennai, Tamil Nadu, India
• Department of Electronics and Communication Engineering, Aditya University, Surampalem, Andhra Pradesh, India
• Department of Information Technology, Hindustan Institute of Technology and Science, Chennai, Tamil Nadu, India

Short Summary

This study introduces PIDeepONet, a novel hybrid deep learning model integrating Physics-Guided Loss (PGL) and Bayesian Optimization (BO), to enhance the accuracy and physical plausibility of river streamflow predictions using only observational data. The model effectively bridges the gap between traditional physics-based and purely data-driven approaches, demonstrating superior performance in both random and temporal data splits for two Indian river basins.

Objective

• To develop and validate PIDeepONet, a hybrid physics-guided deep learning framework that combines a Physics-Guided Loss (PGL) function and Bayesian Optimization (BO), for accurate, physically plausible, and scalable river streamflow prediction using only streamflow observations.

Study Configuration

• Spatial Scale: Two major tributaries of the Cauvery River basin in Southern India: Hemavathi River (Akkihebbal station, 12.5986°N, 76.4006°E) and Kabini River (Hommaragalli station, 12.1142°N, 76.4631°E). The Cauvery basin covers approximately 81,155 square kilometers.
• Temporal Scale:
  • Hemavathi River: January 1, 2004, to December 31, 2024 (6,365 data points).
  • Kabini River: June 7, 2019, to December 31, 2024 (1,608 data points).
  • Data was split into 70% for training, 15% for validation, and 15% for testing, using both random shuffling and strict temporal splitting strategies.

Methodology and Data

• Models used: Physics-Informed Deep Operator Network (PIDeepONet) enhanced with a Physics-Guided Loss (PGL) framework and Bayesian Optimization (BO) for hyperparameter tuning. The model was compared against BO-only and PGL-only optimization methods.
• Data sources: River discharge and water level data acquired from the National Water Informatics Centre (NWIC) of the Central Water Commission (CWC), India.

Main Results

• The hybrid BO-PGL PIDeepONet model consistently outperformed single optimization methods (BO-only or PGL-only).
• For Hemavathi River, random data splitting yielded an R² of 0.99976 and a Root Mean Square Error (RMSE) of 0.711 cubic meters per second (m³/s).
• For Kabini River, random data splitting achieved an R² of 0.99988 and an RMSE of 1.926 m³/s.
• Under strict temporal data splitting, the model demonstrated strong generalization with R² values of 0.94715 for Hemavathi and 0.95914 for Kabini.
• The model produced physically plausible and accurate streamflow predictions even with limited or extreme datasets, suitable for forecasting flash floods and fires.

Contributions

• Development of a novel hybrid Physics-Informed Deep Operator Network (PIDeepONet) architecture that integrates Bayesian Optimization (BO) for efficient hyperparameter tuning and a Physics-Guided Loss (PGL) function for enforcing hydrological consistency.
• Introduction of a PGL function that explicitly incorporates hydrological principles, significantly improving model reliability and reducing unphysical predictions, especially under extreme flow conditions or data scarcity.
• Implementation of automated hyperparameter optimization using BO, which reduces manual intervention, accelerates convergence, and enhances computational efficiency in hydrological modeling.
• Creation of a scalable and adaptable framework for real-time streamflow forecasting, capable of predicting extreme events and applicable in both data-scarce and dynamically evolving hydrological environments, effectively bridging the gap between physics-based and purely data-driven models.

Funding

No funding was received to assist with the preparation of this manuscript.

Citation

@article{Suresh2026PhysicsGuided,
  author = {Suresh, Y. and Sheela, M. Sahaya and Sunitha, Pamarthi and Gopalakrishnan, Saisubramaniam},
  title = {Physics-Guided Deep Learning with Bayesian Optimization for Enhanced River Streamflow Prediction},
  journal = {Water Resources Management},
  year = {2026},
  doi = {10.1007/s11269-026-04520-7},
  url = {https://doi.org/10.1007/s11269-026-04520-7}
}