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

Understanding climate-streamflow dependencies is crucial for evaluating reservoir impacts and adaptive water management. This study analyzed streamflow in two key Brazilian reservoirs, Três Marias (São Francisco Basin) and Serra da Mesa (Tocantins Basin), using monthly records from 1979 to 2020. A 12-month moving average temporal filter enhanced low-frequency climate signals to assess hydrological variability and memory. Temporal smoothing substantially clarified climate–streamflow dependencies, with correlation gains reaching 106% for PDO, 204% for ENSO, and more than 4,200% for the Antarctic Oscillation (AAO) in Três Marias. The filtered analysis revealed contrasting hydrological memory structures: Três Marias exhibited multi-year memory with maximum correlations at approximately 22–27 months, while Serra da Mesa showed faster response times of 4–12 months. To evaluate predictive implications, streamflow forecasting was performed using two deep learning architectures: LSTM (recurrent neural network baseline) and TCN (temporal convolutional network). TCN substantially outperformed LSTM in Três Marias (R2 = 0.95 vs. 0.05), demonstrating that convolutional architectures effectively exploit low-frequency persistence when scale-aware preprocessing reveals it. These findings show that temporal filtering provides an effective framework for detecting climate–streamflow dependencies and hydrological memory, with direct implications for seasonal-to-decadal forecasting and climate-informed reservoir management under changing conditions.