.png){kind=icon}
.png){kind=icon}
Return
Future-informed FOCAL MODE: a multimodal deep learning framework for forecasting dissolved oxygen in reservoirs
.png){kind=icon}
PRE
AI
F
J
N
C
A
DOI：10.1007/s10661-026-15096-5.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Dissolved oxygen (DO) is essential for reservoir ecosystem health but is often reduced by dam operations that alter water quality. Agencies frequently mitigate these low levels through liquid oxygen injection. Since this costly intervention requires advance planning and logistical coordination, accurate long-range DO forecasts are critical for timely, cost-effective oxygenation. To address this need, we design Future-informed FOurier, Cross-modal Attention LSTM using Multi-modal Observational Data Engine, referred to as Future-informed FOCAL MODE—an advanced predictive modeling framework that unifies data-driven learning with physics-based insight. The architecture features a dual-channel design: the first channel learns from rich, multi-dimensional historical datasets collected from diverse environmental sources, capturing long-term patterns and interdependencies across modalities. Each modality contributes a distinct aspect of the system under study, offering a comprehensive perspective despite differences in resolution and temporal alignment. The second channel integrates forward-looking inputs from physics-informed models that reflect the governing laws of environmental dynamics. This future-informed stream accounts for evolving climate change, not captured in historical data. While the former channel forecasts DO from learned past patterns, the latter predicts DO from anticipated environmental conditions. The integration of these complementary perspectives enables more accurate, adaptive, and robust DO predictions under both current and future scenarios. By combining retrospective learning with anticipatory modeling, Future-informed FOCAL MODE adapts historical learning to future trends, enhancing climate responsiveness. In a Watts Bar Dam case study (Tennessee, USA), the model showed strong accuracy RMSE $$\sim $$ 1, demonstrating resilient, climate-aware DO forecasts.
Keywords:
Time series analysis
Deep learning
Dissolved oxygen prediction
Physics-informed modeling
Multi-modal data integration
Journal
IF:
3
Papers:
2.1K
Citations:
3.5W
Organization
Cited Papers