Enhancing LSTM streamflow modeling in data-scarce rain-dominated basins: the impact of using multiple precipitation products as model inputs

Abstract

Uruguay relies heavily on hydropower, with three dams along the Río Negro providing nearly half of the country’s electricity. Accurate streamflow modeling is therefore critical for dam operations and energy reliability. Recently, long short-term memory (LSTM) networks have emerged as a leading approach for streamflow modeling.To train LSTMs, daily meteorological forcings (e.g., precipitation, temperature, radiation) alongside static catchment attributes (e.g., geophysical and climatological) are typically required. Large-sample datasets such as CARAVAN have enabled the global application of LSTMs. However, CARAVAN’s reanalysis-based precipitation has been shown to reduce model performance in the US compared to other local products, specifically in the Midwest and the Southeast, regions characterized by rain-dominated basins, the same hydrological regime found across Uruguay. This study presents the evaluation of LSTM-based streamflow modeling in Uruguay, training and evaluating models across eleven sub-basins in the Río Negro basin under four precipitation forcings: CARAVAN, MSWEP, CHIRPS, and local rain gauges. To do so, precipitation products are first compared across a subset of rain-dominated US basins, and the resulting insights are used to interpret model performance in Uruguay. It assesses model performance and hydrological signatures under both individual and combined precipitation product inputs. Results show that MSWEP provides the best performance among gridded precipitation products across both the US and Uruguayan basins, with precipitation peak representation identified as the primary driver of performance differences between products. In Uruguay, combining MSWEP with local rain gauge data yielded the best overall results across most hydrological signatures, while CARAVAN reanalysis precipitation consistently underperformed. These findings highlight the importance of precipitation product selection for LSTM-based streamflow modeling in rain-dominated and data-scarce regions.