Previsão de vazão em cenário de mudanças climáticas: uma análise comparativa de modelos LSTM e TFT na sub-bacia do Rio Canoas (Bacia do Rio Uruguai)

Abstract

Hydrological modeling in the Uruguay River Basin faces challenges due to climate non-stationarity, which weakens the validity of historical patterns. This study evaluates the Temporal Fusion Transformer (TFT) and Long ShortTerm Memory (LSTM) architectures for daily streamflow forecasting and drought detection, comparing a historical regime (1980–1990) and a recent regime (2011–2018) while using the Oceanic Niño Index (ONI) as a covariate. The LSTM achieved competitive performance for nowcasting (t + 1), whereas the TFT outperformed it at subseasonal horizons (7, 15, and 30 days), with LogNSE > 0.86 and greater stability under low-flow conditions (RMSEdry). The TFT’s interpretability supports its ability to capture relevant physical drivers. Overall, model choice should be guided by the forecasting horizon, with the TFT offering greater resilience for water planning under climate change.