Data-Driven Detection of Climate–Streamflow Dependencies and Multi-Year Hydrological Persistence in Brazilian Reservoir Systems

Understanding how climate variability is reflected in streamflow is essential for reservoir management and hydropower planning. This study investigated how temporal scale influences climate–streamflow relationships, persistence characteristics, and predictability in two Brazilian reservoirs: Três Marias (São Francisco Basin) and Serra da Mesa (Tocantins Basin). Monthly streamflow and climate-index records (Pacific Decadal Oscillation (PDO), El Niño–Southern Oscillation (ENSO), and Antarctic Oscillation (AAO)) from 1979–2020 were analyzed using a 12-month moving average (MA12) filter to emphasize low-frequency variability. Temporal filtering strengthened climate–streamflow relationships, particularly for PDO and AAO, revealing signals that were less apparent in the original monthly series. Lagged-correlation analyses identified contrasting persistence structures between the reservoirs. Três Marias exhibited multi-year persistence timescales (22–27 months), whereas Serra da Mesa showed shorter and more heterogeneous response timescales, ranging from an immediate PDO response to approximately 14–19 months for ENSO and AAO. Forecasting experiments using benchmark models (Persistence and Linear Regression) and deep learning architectures (LSTM and TCN) showed limited predictive skill on the raw monthly series but substantially improved performance after temporal filtering. For the MA12-filtered series, the benchmark models achieved the highest performance in both reservoirs (R2≈0.95 in Três Marias and R2≈0.93 in Serra da Mesa). Overall, the results indicate that temporal scale strongly influences the detectability of climate signals, the persistence of streamflow variability, and the predictability of reservoir inflows.