LSTM beats Transformer for streamflow prediction in ungauged basins
Adding downstream data boosts prediction accuracy by over 60% for both models.
A new study published on arXiv (arXiv:2606.02791) pits a Transformer encoder against a classic LSTM for a critical hydrological task: inferring upstream streamflow in ungauged basins. The research team—Taye Akinrele, James Halgren, Noorbakhsh Amiri Golilarz, Sudip Mittal, and Shahram Rahimi from Mississippi State University—used retrospective simulations from NOAA's National Water Model as their testbed. They ran two configurations: one using only upstream data and another combining upstream with downstream observations.
The results are a clear win for the older architecture. The LSTM showed stronger overall performance in both settings, suggesting that recurrent memory's inductive bias is better suited for reconstructing upstream flows from limited hydrologic information. More strikingly, adding downstream context boosted median Nash-Sutcliffe efficiency (NNSE)—a key accuracy metric—by over 60% for all models. The authors emphasize this is not a leaderboard exercise but a deeper probe into architectural inductive biases, with practical implications for flood forecasting and water resource management in data-sparse regions.
- Encoder-only Transformer underperformed LSTM for upstream streamflow inference across all test configurations.
- Adding downstream hydrologic context improved median NNSE by over 60% for both architectures.
- Study used NOAA National Water Model retrospective simulations to evaluate models in ungauged basins.
Why It Matters
Proves simpler recurrent models still outperform Transformers for certain spatiotemporal physics tasks, guiding AI deployment in hydrology.