ClimateOpenAlex
Heuristic editor, no API keyVerdict: NotableMass-Conserving LSTM With Dual States for Streamflow Prediction: Separating Quickflow and Slow Storage
Abstract We introduce a Mass-Conserving Long Short-Term Memory with Dual States (MC-LSTM-DS) with the intention of separating short- and long-term memory for predicting streamflow.
VerdictWorth a reader's time today.
Abstract
Abstract We introduce a Mass‐Conserving Long Short‐Term Memory with Dual States (MC‐LSTM‐DS) with the intention of separating short‐ and long‐term memory for predicting streamflow. It enforces water balance through a new partition gate on precipitation. We benchmark MC‐LSTM‐DS against a standard Long Short‐Term Memory (LSTM) and the original Mass‐Conserving Long Short‐Term Memory (MC‐LSTM) on CAMELS‐IND for 158 basins. Skill score analysis of the three models reveals that enforcing mass conservation degrades performance in the semi‐arid and tropical monsoon regions. However, the addition of another cell state in MC‐LSTM‐DS improves performance over MC‐LSTM in these regions. The decomposition of model predictions suggests that the added long cell‐state captures slow storage and releases in the model, while the original state tracks quickflow. Their relative contributions vary systematically with climate, providing a hydrological representation of the cell states. The newly introduced partition mechanism captures signatures of different runoff‐generating processes. A basin‐scale water‐balance check suggests additional effective inflows in some monsoon‐dominated regions. This highlights the catchments where data or missing process representations (e.g., groundwater) may limit strict closure. We also evaluated our model's cross‐regional robustness on the CAMELS‐US data set across 531 basins. The performance of MC‐LSTM‐DS on the CAMELS‐US matches that of LSTM and MC‐LSTM in Nash‐Sutcliffe Efficiency and attains state‐of‐the‐art Kling‐Gupta Efficiency and FHV (High Flow Bias). This indicates its potential generalization across regions. This study establishes a binational benchmark and provides an interpretable, mass‐conserving deep learning framework for operational streamflow prediction across diverse hydroclimates.
The editor's rubric
| Dimension | Level | Weight | What that level means |
|---|---|---|---|
| Leverage | ███░░ 3 | 16% | A method or resource many groups across the field will adopt within a year. |
| Magnitude | ███░░ 3 | 20% | Large gain: roughly 2x, or a clear new state of the art on a hard, unsaturated problem. |
| Evidence | ███░░ 3 | 20% | Solid: multiple benchmarks or cohorts, ablations, fair baselines, released code or data. |
| Novelty | ██░░░ 2 | 10% | A new combination of known ideas. |
| Trajectory | ███░░ 3 | 14% | A clear path to scale. |
| Stakes | ██░░░ 2 | 20% | Benefits a professional community (practitioners, clinicians, engineers). |
Editor’s rationale
Heuristic triage from title and abstract text only, not a reading of the paper. Cues found: method (we propose); breadth (generalizes); gains (state of the art); verification (multiple benchmarks); scale (scalable, efficient); stakes (climate).
How the score was computed
- Merit
- 5.4 / 10
- Adjusted merit
- 4.6 / 10
- Attention
- 27%
- Freshness
- 85%
- Citations1 (reference 15, via openalex, Oct 2, 2026, 07:30 UTC)
- Field-weighted citation impact2.5 (reference 3, via openalex, Oct 2, 2026, 07:30 UTC)