Author(s): Sewoong Chung; Sungjin Kim

Linked Author(s): Sungjin Kim, Se-Woong Chung

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

Abstract: This study aimed to develop and evaluate a process-guided deep learning (PGDL) model for predicting the vertical distributions of water temperature and turbidity in a deep stratified reservoir, with the goal of improving both predictive accuracy and physical consistency for reservoir operation support. To achieve this, we developed a Process-Guided Long Short-Term Memory (PG-LSTM) model that incorporates physical constraints derived from the two-dimensional hydrodynamic and water quality model CE-QUAL-W2 into the training process. Specifically, penalty terms were added to the loss function to discourage violations of energy conservation for temperature and mass conservation for suspended sediments. Unlike previous PGDL studies, which have largely focused on temperature prediction alone and have generally relied on simplified one-dimensional process representations, this study extends the framework to the simultaneous prediction of temperature and turbidity profiles in a real deep reservoir using dual conservation constraints linked to a 2D mechanistic model. The proposed model was applied to the Soyang-gang Reservoir in the Republic of Korea. Results showed that PG-LSTM reduced temperature RMSE by 48% and 65% and turbidity RMSE by 51% and 49% relative to the standard LSTM and calibrated W2 model, respectively, while also maintaining better physical consistency than the conventional LSTM. These results demonstrate that embedding process-based knowledge into data-driven deep learning can improve both the reliability and predictive skill of vertical water quality simulations. The developed PG-LSTM model therefore provides a robust decision-support tool for the operation of selective withdrawal facilities and for managing complex vertical water quality dynamics in stratified reservoirs.

DOI: https://doi.org/10.64697/iahr.proc.hic2026.15

Year: 2026