Previsão mensal de vazão para o rio Jequitinhonha

Streamflow forecasting is pivotal for mitigating socio‑economic risks and optimizing water‑resources operations, whose inherent variability affects energy production, public supply, and extreme‑event management. Aiming to bridge the national research gap on machine‑learning applications to the Jequitinhonha River, this study developed a CatBoost model capable of delivering reliable monthly forecasts. Daily discharge observations served to train the model, and 31‑day (one‑month) ahead predictions were evaluated with MAPE, RMSE, and KGE, together with 5th–95th‑percentile prediction intervals. The baseline configuration yielded a MAPE of 15.7 %, an RMSE of 26.7 m³/ s, and a KGE of 0.36; expanding the lag window subsequently improved the KGE to 0.54, while the 90 % prediction interval satisfactorily encompassed the observed values. CatBoost thus proved adept at reproducing the Jequitinhonha’s flow regime, yet further advances will depend on longer time series and refined hyperparameter tuning to enhance correlation and diminish bias, consolidating the approach as a promising decision‑support tool for regional water‑resources management.