Answers are grounded in this paper and cite their sources.
Hourly rainfall–runoff modelling with deep learning involves long input sequences and rapid short-term streamflow variability. Conventional LSTMs may encounter architectural bottlenecks: information from distant time steps must be carried through a compact recurrent state, and sigmoid-bounded gates can limit how quickly stored information is revised. This study examines whether extended long short-term memory (xLSTM), which introduces exponential gating and expanded memory structures, can improve large-sample hourly streamflow prediction. An xLSTM-based model and LSTM-based benchmarks were evaluated under the same input setting and general training hyperparameters. Experiments used hourly streamflow and meteorological records from 430 Canadian basins over 2011–2023. The xLSTM-based model achieved comparable overall accuracy for temporal extrapolation and improved overall accuracy for spatial extrapolation. These results suggest that xLSTM is a plausible alternative to standard LSTM, although further work is needed to refine xLSTM configurations and evaluate their robustness across larger and more diverse datasets.
No takes yet. Share an insight, caveat, or question.
Yu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: