Author(s): Ashutosh Sharma; Nikunj Mangukiya; Sweta Dash
Linked Author(s): Ashutosh Sharma
Keywords: Deep learning differentiable modeling process-based hydrology reservoir inflow prediction
Abstract: Accurate dam inflow prediction is critical for ensuring dam safety, optimizing operations, supporting downstream applications, and improving reservoir management. Recently, data-driven models have outperformed traditional hydrological models in streamflow prediction and other geoscientific variables. However, these models often lack interpretability, which is essential for building trust and ensuring reliability, especially in contexts where dam safety and operations are at stake. In this study, we demonstrate that a differentiable process-based model can provide more accurate inflow predictions compared to purely data-driven models, such as Long Short-Term Memory (LSTM) networks, and traditional models, such as the Hydrologic Engineering Center - Hydrologic Modeling System (HEC-HMS). Specifically, we utilized a process-based hydrological model in a simplified form as the backbone and integrated it with LSTM neural networks for parameterization, creating a differentiable parameter learning (dPL) framework. To evaluate the reliability and efficacy of the proposed differentiable framework, we conducted a case study to predict inflow for the Rengali Dam located in Odisha, India, across the Brahmani River basin. The results demonstrated the superior performance of the proposed dPL framework compared to the LSTM model for inflow prediction. At the Rengali Dam, the dPL framework achieved a significantly higher Nash–Sutcliffe efficiency (0.80) than the LSTM model (0.52), indicating improved predictive accuracy. The dPL model also reduced high flow bias from 19.10% to 8.84%, lowered RMSE from 515.01 m3/s to 331.78 m3/s, and achieved a higher correlation coefficient (0.92 vs. 0.85). Similar improvements were observed at the Gomlai and Panposh gauge stations, where the dPL framework consistently outperformed the LSTM model in terms of Nash–Sutcliffe efficiency, high flow bias, RMSE, and overall bias. These results highlight the robustness and reliability of the proposed framework. Overall, the proposed differentiable framework seamlessly integrates the accuracy of data-driven models with the interpretability and reliability of process-based models, paving the way for digital transformation in hydrological modeling. This approach not only accommodates models with varying process complexity but also opens new avenues for learning and incorporating physics from big data.
Year: 2025