Machine Learning for Daily Streamflow Prediction in Coastal-Andean Basins: Evaluating Climate-Only and Hydrological Persistence Predictors

Morales-Davila J.; Argumedo K.O.; Arteaga A.C.; Altez S.P. · International Conference on Civil, Structural and Transportation Engineering (2026) · DOI: 10.11159/iccste26.225

Resumen

– Daily streamflow prediction in coastal-Andean basins with intermittent flow regimes represents a critical challenge for water resources management in regions characterized by high seasonal variability and limited hydrological monitoring. Although machine learning (ML) models have demonstrated effectiveness under data-scarce conditions, most existing studies do not quantify the differential impact of hydrological persistence predictors across seasons — a gap that is particularly relevant in semi-arid basins of the Peruvian Pacific. This study evaluates three ML models — Multiple Linear Regression (MLR), Random Forest (RF), and XGBoost — under two predictor scenarios: climate-only inputs (E1, 9 variables) and climate inputs combined with observed streamflow lags (E2, 12 variables) — applied to the Fortaleza, Pativilca, and Supe basins over the period 2000–2015, with performance metrics disaggregated by wet and dry season. During the test period, E2 achieved Nash–Sutcliffe Efficiency (NSE) ≥ 0.895 across all basin–model combinations, classified as "Very Good" according to Moriasi et al., compared to values between 0.697 and 0.857 under E1. The most critical improvements were observed during the dry season: E1 yielded negative NSE values and Percent Bias (PBIAS) of up to +157.2% in the Supe basin, reduced to below 8% under E2, with ΔNSE_Dry reaching +1.029. Variable importance analysis revealed a structural regime shift between scenarios: Pcp_acc7 dominated under E1, whereas Q_lag1 concentrated the maximum normalized importance (1.00) under E2 across all models and basins. Predictor set selection exerted a greater influence on model performance than algorithm choice. These results quantify the operational value of streamflow telemetry in Peruvian Pacific basins and establish hydrological persistence as the key determinant of dry-season predictive accuracy.

Instituciones: Universidad Privada del Norte