GIST Scholar
College of Engineering
Department of Environment and Energy Engineering
1. Journal Articles
Hierarchical prediction of streamflow and organic carbon at event scale using a physics-guided AI hybrid framework
- Type
- Article
- Citation
- JOURNAL OF HYDROLOGY, v.679
- Issued Date
- 2026-10
- Abstract
- Agricultural watersheds are major contributors to storm-driven export of streamflow (Q) and organic carbon, including particulate organic carbon (POC) and dissolved organic carbon (DOC), which strongly influence downstream water quality and aquatic biogeochemical processes. However, accurate prediction of event-scale hydro-carbon dynamics remains challenging because process-based models require extensive calibration, whereas purely data-driven models often lack physical consistency. This study presents a physics-guided spatial-temporal hybrid framework (PhyST-HC) that integrates uncalibrated Soil and Water Assessment Tool-Carbon (SWAT-C) simulations with Graph Neural Networks (GNNs) and Transformer-based attention mechanisms for hierarchical prediction of Q, POC, and DOC. Distributed SWAT-C-derived hydrological and carbon-state variables serve as physics-guided watershed descriptors, while GNNs capture upstream-downstream connectivity and the Transformer learns temporal watershed memory. The framework was evaluated in the agriculturally dominated Tuckahoe Creek Watershed (TCW), USA. Results showed that PhyST-HC achieved Q performance comparable to a SUFI-2-calibrated SWAT-C model (KGE = 0.82) without watershed-specific calibration, while substantially outperforming both SUFI-2 and SWAT-LSTM for organic carbon prediction (POC: KGE = 0.68 vs. 0.04 and 0.34; DOC: KGE = 0.74 vs. 0.43 and -0.05, respectively). Ablation analysis confirmed the importance of physics-guided SWAT-derived watershed-state variables beyond standalone GNN and Transformer models. The framework also improved representation of event-scale POC flushing, delayed DOC transport, and watershed connectivity under contrasting storm conditions. These findings demonstrate the potential of physics-guided hybrid learning for accurate and transferable event-scale hydro-carbon prediction in agricultural watersheds.
- Publisher
- ELSEVIER
- ISSN
- 0022-1694
- Appears in Collections:
- Department of Environment and Energy Engineering > 1. Journal Articles
- 파일 목록
-
-
관련 파일이 존재하지 않습니다.
-
Items in Repository are protected by copyright, with all rights reserved, unless otherwise indicated.