Abstract
Accurate hourly streamflow prediction is crucial for managing water resources, particularly in smaller basins with short response times. This study evaluates six deep learning (DL) models, including Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Convolutional Neural Network (CNN), and their hybrids (CNN-LSTM, CNN-GRU, CNN-Recurrent Neural Network (RNN)), across two basins in Northwest Spain over a ten-year period. Findings reveal that GRU models excel, achieving Nash-Sutcliffe Efficiency (NSE) scores of approximately 0.96 and 0.98 for the Groba and Anllóns catchments, respectively, at 1-hour lead times. Hybrid models did not enhance performance, which declines at longer lead times due to basin-specific characteristics such as area and slope, particularly in smaller basins where NSE dropped from 0.969 to 0.24. The inclusion of future rainfall data in the input sequences has improved the results, especially for longer lead times from 0.24 to 0.70 in the Groba basin and from 0.81 to 0.92 in the Anllóns basin for a 12-hour lead time. This research provides a foundation for future exploration of DL in streamflow forecasting, in which other data sources and model structures can be utilized.