Deep learning of flood forecasting by considering interpretability and physical constraints
Problem
Deep learning flood forecasting models often lack interpretability and physical realism, limiting trust in predictions.
Why it matters
Accurate short-term flood forecasts support flood warning and risk management.
What the AI did
An LSTM was enhanced with a feature-time attention mechanism and physics-guided constraints enforcing monotonic input-output relationships, tested for 1-6 hour flood prediction in the Luan River Basin, China.
Results
NSE 0.988 at t+1 h and 0.908 at t+6 h; outperforms standard LSTM and attention-only variants.
Impact
Outperforms baselines in one basin; broader adoption not stated.
Record
| Venue | Hydrology and earth system sciences |
|---|---|
| Published | 2025 |
| Work type | article |
| Authors | Ting Zhang, Ran Zhang, Jianzhu Li, Ping Feng |
| Citations | 16 (OpenAlex, at import time) |
| References | 37 |
| Model or method | PHY-FTMA-LSTM (physics-guided feature-time multi-head attention LSTM) |
| Organization | not stated in source |
| Primary source | https://doi.org/10.5194/hess-29-5955-2025 |
| Open access | gold |
| Keywords | flood forecasting, LSTM, physics-guided, interpretability, Nash-Sutcliffe efficiency |
Importance rating
significant — thousands to millions affected
Abstract gives no flood impact figures; short-term flood forecasting broadly supports warning and risk management, implying significant benefit to flood-prone communities.
More in Climate & Environment
| # | Paper | Year | Citations | Importance |
|---|---|---|---|---|
| 1 | Learning skillful medium-range global weather forecasting | 2023 | 1,416 |
8
|
| 2 | Accurate medium-range global weather forecasting with 3D neural networks | 2023 | 1,631 |
8
|
| 3 | FuXi: a cascade machine learning forecasting system for 15-day global weather forecast | 2023 | 355 |
8
|
| 4 | Skilful nowcasting of extreme precipitation with NowcastNet | 2023 | 451 |
8
|
| 5 | Forecasting Renewable Energy Generation with Machine Learning and Deep Learning: Current Advances and Future Prospects | 2023 | 378 |
8
|