MLP-based hydrological forecasting in the Madeira River Basin, Amazonia: a prelude toward robust modeling of hydrological extremes

Recent decades have been associated with an apparent intensification of hydrological extremes across the Madeira River Basin (Amazonia), reinforcing the need for forecasting frameworks that are reproducible, leakage-safe, and operationally defensible. An integrated machine-learning workflow is implemented to forecast the downstream Standardized Streamflow Index (SSI-12) at gauging station 15700000, using a strictly time-ordered monthly dataset and an explicitly controlled validation protocol. The supervised learning design is defined via forward target shifting (h = 1) and explicit representation of hydrological memory through antecedent lag terms (1-12 months), consistent with the persistence embedded in accumulated standardized indices. Data preparation comprises temporal harmonization, conversion to consistent numeric formats, and reconstruction of residual gaps through KNN imputation to better preserve multivariate covariability in predictor space. A parsimonious modeling pipeline is adopted, combining standardization (training statistics only) with mutual-information-based feature screening to enforce predictor compactness and reduce redundancy. Hyperparameters and feature subset size are optimized via RandomizedSearchCV under TimeSeriesSplit cross-validation, with NSE used as the primary refit criterion. Final fitting is refined through external early stopping on a held-out validation segment, monitoring a robust Huber loss to stabilize training under heteroscedastic conditions. Out-of-sample skill assessed through RMSE, MAE, R2, NSE, and KGE indicates strong predictability and close phase agreement between forecasts and observations. Nevertheless, a persistence-type baseline remains superior on validation and test partitions, underscoring the pronounced short-term autocorrelation intrinsic to SSI-12 and setting a stringent benchmark for incremental gains. Residual behavior under extremes further indicates heteroscedasticity and systematic peak attenuation, motivating extreme-aware refinements centered on residual learning relative to persistence, event-centric feature engineering incorporating exogenous hydroclimatic drivers, and tail-sensitive optimization to improve fidelity during high-impact episodes.