EVALUATING TRANSFORMER AND LSTM FRAMEWORKS FOR PREDICTION IN UNGAUGED BASINS
Taye Akinrele 1 , James Halgren 2 , Noorbakhsh Amiri Golilarz 3 Sudip Mittal 4 , Shahram Rahimi 5 · 2026
Read EVALUATING TRANSFORMER AND LSTM FRAMEWORKS FOR PREDICTION IN UNGAUGED BASINS by Taye Akinrele 1 , James Halgren 2 , Noorbakhsh Amiri Golilarz 3 Sudip Mittal 4 , Shahram Rahimi 5 (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-02791.txt.
Description
Akinrele and co-authors compare two sequence models for streamflow prediction where no gauges exist. Ungauged basins lack direct observations. Uncertainty rises and extreme events get harder to anticipate. Watershed networks merge many tributaries into downstream channels. The study tests an encoder-only Transformer against an LSTM. Retrospective simulations from the NOAA National Water Model supply the data. Upstream-only and combined configurations are both run.
The LSTM wins overall in both configurations. Adding downstream information boosts every model. Median NNSE rises by more than 60 percent. The authors read the result as a test of architectural inductive bias. Recurrent memory fits this upstream reconstruction task better than an encoder-only Transformer. Downstream context acts as a strong auxiliary constraint across architectures. The work guides model choice for hydrologic inference with sparse data.
Catalog record
| CREATOR | Taye Akinrele 1 , James Halgren 2 , Noorbakhsh Amiri Golilarz 3 Sudip Mittal 4 , Shahram Rahimi 5 |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 19 KB |
| LICENSE | http://creativecommons.org/licenses/by-sa/4.0/ |
| RIGHTS | CC BY SA |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.02791