Physics-Informed Deep Learning for Karst Spring Prediction: Integrating Variational Mode Decomposition and Long Short-Term Memory with Attention
Water · 08.07.2025 · 17 (14) · 2043
DOI: 10.3390/w17142043 ↗Kurz gefasst
Fachartikel in der Zeitschrift Water, online erschienen am 08.07.2025. Verfasst von Liangjie Zhao und 6 weiteren Autor:innen. Die Arbeit wurde bisher 6-mal zitiert (Crossref) und nennt 37 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Key Research and Development Program of China, Karst Water Resources and Environment Academician Workstation of Guizhou Province, Basic Research Fund.
Abstract (Originalsprache)
Accurately forecasting karst spring discharge remains a significant challenge due to the inherent nonstationarity and multi-scale hydrological dynamics of karst hydrological systems. This study presents a physics-informed variational mode decomposition long short-term memory (VMD-LSTM) model, enhanced with an attention mechanism and Monte Carlo dropout for uncertainty quantification. Hourly discharge data (2013–2018) from the Zhaidi karst spring in southern China were decomposed using VMD to extract physically interpretable temporal modes. These decomposed modes, alongside precipitation data, were input into an attention-augmented LSTM incorporating physics-informed constraints. The model was rigorously evaluated against a baseline standalone LSTM using an 80% training, 15% validation, and 5% testing data partitioning strategy. The results demonstrate substantial improvements in prediction accuracy for the proposed framework compared to the standard LSTM model. Compared to the baseline LSTM, the RMSE during testing decreased dramatically from 0.726 to 0.220, and the NSE improved from 0.867 to 0.988. The performance gains were most significant during periods of rapid conduit flow (the peak RMSE decreased by 67%) and prolonged recession phases. Additionally, Monte Carlo dropout, using 100 stochastic realizations, effectively quantified predictive uncertainty, achieving over 96% coverage in the 95% confidence interval (CI). The developed framework provides robust, accurate, and reliable predictions under complex hydrological conditions, highlighting substantial potential for supporting karst groundwater resource management and enhancing flood early-warning capabilities.
Förderung
- National Key Research and Development Program of China
- Karst Water Resources and Environment Academician Workstation of Guizhou Province
- Basic Research Fund
Zitieren
APA
Zhao, L., Fazi, S., Luan, S., Wang, Z., Li, C., Fan, Y., & Yang, Y. (2025). Physics-Informed Deep Learning for Karst Spring Prediction: Integrating Variational Mode Decomposition and Long Short-Term Memory with Attention. Water, 17(14), 2043. https://doi.org/10.3390/w17142043
Vancouver
Zhao L, Fazi S, Luan S, Wang Z, Li C, Fan Y, et al. Physics-Informed Deep Learning for Karst Spring Prediction: Integrating Variational Mode Decomposition and Long Short-Term Memory with Attention. Water. 2025;17(14):2043. doi:10.3390/w17142043
BibTeX
@article{Zhao2025,
author = {Zhao, Liangjie and Fazi, Stefano and Luan, Song and Wang, Zhe and Li, Cheng and Fan, Yu and Yang, Yang},
title = {Physics-Informed Deep Learning for Karst Spring Prediction: Integrating Variational Mode Decomposition and Long Short-Term Memory with Attention},
journal = {Water},
year = {2025},
volume = {17},
number = {14},
pages = {2043},
doi = {10.3390/w17142043}
}
EUVIMED