Projection of river discharge trends using machine learning and statistical uncertainty quantification

ABSTRACT Reliable river discharge forecasting is essential for flood mitigation and water management in climate-vulnerable areas like Taiwan. This study combines machine learning (ML) models – Random Forest, Support Vector Regression, Multilayer Perceptron, Gradient Boosting Machine, and Long Short-Term Memory – with statistical uncertainty quantification to predict peak discharges. Using historical annual maximum data (1975–2023) from gauging stations, it projects 2025–2040 trends. Empirical distributions (Log-Pearson Type III, Gumbel, Generalized Extreme Value) estimate 100-year return periods. Mann-Kendall analysis shows rising peaks in central Taiwan (e.g., Zhuoshui River) and declines in south/east (e.g., Gaoping, Hualien Rivers), indicating increased flood risks and water shortages. Bootstrap resampling reveals wider uncertainty in variable regions, while Radial Basis Function interpolation maps spatial extremes. Log-Pearson Type III fits best for most stations. Findings underscore adaptive strategies and resilient planning, enhancing accuracy for policymakers and engineers in hydrological risk management.