Title: Assessing the Efficacy of Transformer and LSTM Architectures for Streamflow Prediction in Data-Scarce Regions

Abstract: Watershed systems are characterized by convergent structures where various tributaries join to form downstream channels, thereby consolidating a range of hydrological processes originating from upstream areas. In regions lacking direct observational data, this scarcity of information heightens uncertainty and hampers the capacity to forecast extreme weather events. This research investigates whether an encoder-only Transformer architecture offers superior performance compared to Long Short-Term Memory (LSTM) networks for inferring upstream streamflow when hydrological data is limited. The analysis utilizes retrospective simulations generated by the NOAA National Water Model (NWM).

The findings reveal that, across both upstream-only and combined setup configurations, the LSTM model demonstrated more robust overall performance than the Transformer. The addition of downstream data significantly enhanced predictive accuracy for all models, raising the median Nash-Sutcliffe Efficiency (NNSE) by over 60%. Rather than framing this work as a competitive benchmark, the authors view the experiments as an examination of architectural inductive bias within the context of hydrologic sequence inference. The data suggests that recurrent memory mechanisms are better suited to the task of upstream reconstruction than encoder-only Transformers. Furthermore, the inclusion of downstream hydrologic context serves as a potent auxiliary constraint, markedly boosting prediction accuracy across different architectural frameworks.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC