Author(s): Sungjin Kim; Sewoong Chung

Linked Author(s):

Keywords: Process-guided deep learning; Physics-informed model; Water temperature prediction; Turbidity modeling; Stratified reservoir; CE-QUAL-W2

Abstract: This study developed a Process-Guided Long Short-Term Memory (PG-LSTM) model to predict the vertical distributions of water temperature and turbidity in a deep stratified reservoir. The PG-LSTM model was trained by incorporating physical constraints derived from a hydrodynamic model (CE-QUAL-W2) into its cost function. Specifically, penalty terms were added when the model predictions violated the principles of energy and mass conservation. The proposed model was applied to the Soyang-gang Reservoir in South Korea to evaluate its performance. Results demonstrated that the PG-LSTM model not only improved predictive accuracy for temperature and turbidity profiles but also maintained better compliance with energy and mass conservation laws compared with the conventional LSTM model. These findings indicate that integrating physical process knowledge into data-driven deep learning frameworks can enhance both physical consistency and predictive capability. The PG-LSTM approach highlights the potential of physics-informed artificial intelligence models for improving the reliability of water quality prediction in complex stratified lake systems.

DOI:

Year: 2026