Data-Driven Modeling

Deep Learning-based Large-sample Modelling
Overcoming Limitations of Data Gaps
Addressing Dam Regulations on Streamflow
Physical Interpretability and Laws

Deep Learning-based Large-sample Modelling

We test the first instance of a Long Short-Term Memory (LSTM) model for simultaneous simulation of multiple Indian watersheds. The study demonstrates that DL models can significantly improve streamflow predictions over conventional methods, even under data-scarce conditions.

Read the Full paper here: Nikunj K. Mangukiya, Ashutosh Sharma, Chaopeng Shen (2023) “How to enhance hydrological predictions in hydrologically distinct watersheds of the Indian subcontinent?”, Hydrological Processes, 37(7), e14936. https://doi.org/10.1002/hyp.14936

Overcoming Limitations of Data Gaps

A deep learning approach was developed to simulate daily flows from aggregated and intermittent observations, addressing the challenge of temporal gaps in the data. The study shows that reliable daily predictions can be achieved even with monthly or weekly observations, offering a practical solution for hydrological modeling in data-limited and sparsely sampled catchments.

Read the Full paper here: Nikunj K. Mangukiya, Ashutosh Sharma (2025). Deep learning‐based approach for enhancing streamflow prediction in watersheds with aggregated and intermittent observations, Water Resources Research, 61, e2024WR037331. https://doi.org/10.1029/2024WR037331

Addressing Dam Regulations on Streamflow

Develops a hybrid differentiable process-based hydrological model (dPLHBVRes) that integrates reservoir dynamics within a neural-parameterized HBV framework. The model enables joint learning of streamflow and unobserved internal hydrological states, bridging the gap between process understanding and predictive performance. It demonstrates that combining process knowledge with data-driven flexibility improves model reliability and interpretability in human-influenced catchments.

Read the Full paper here: Mangukiya & Sharma (2025) Integrating Reservoir Dynamics into Differentiable Process-Based Hydrological Model for Enhanced Streamflow Estimation. Water Resources Research, 61(7), e2025WR040268. https://doi.org/10.1029/2025WR040268

Physical Interpretability and Laws

We investigate physics-informed deep learning using the Mass-Conserving Long Short-Term Memory (MC-LSTM) framework, which explicitly enforces hydrological laws such as mass conservation to improve the reliability, stability, and interpretability of streamflow predictions. MC-LSTM demonstrates state-of-the-art performance in capturing high-flow dynamics and shows enhanced robustness in highly human-influenced, data-scarce, and semi-arid watersheds compared to conventional LSTM models.

Read the Full paper here: Sahu, G., Mangukiya, N. K., & Sharma, A. (2025). Does MC-LSTM model improve the reliability of streamflow prediction in human-influenced watersheds? Journal of Hydrology, 665, 134711. https://doi.org/10.1016/j.jhydrol.2025.134711

Figures on this page are adapted from research publications of our lab and related studies. All rights to the original figures remain with the respective copyright holders.

Page updated

Google Sites

Report abuse