Enhancing streamflow prediction in a dam-regulated river by integrating mechanism and machine learning models

Dongjiang River Basin, China Daily streamflow prediction in dam-regulated rivers remains a critical challenge in contemporary hydrology, particularly given the growing global prevalence of regulated river systems. In order to reduce high-flow prediction errors while mitigating RF overfitting through HSPF constraints and maintaining robust validation performance, this work develops a hybrid streamflow prediction framework combining Hydrological Simulation Program-FORTRAN (HSPF) and Random Forest (RF) to improve daily streamflow prediction in dam-regulated rivers. The proposed hybrid methodology strategically integrates the Hydrological Simulation Program—FORTRAN (HSPF), which provides physics-based simulations of watershed-scale rainfall-runoff processes, with the Random Forest (RF) algorithm, which effectively captures nonlinear dam operation patterns. This integration addresses key limitations associated with standalone modeling approaches. Validation through multiple metrics demonstrates the integrated framework's statistically superior performance compared to individual HSPF and RF (NSE=0.83) models across all flow regimes. Notably, the ensemble approach reduces extreme flow prediction errors by 4–25 % while mitigating RF's overfitting tendency (validation NSE decline: 0.83→0.49) through mechanistic constraints. This nested simulation paradigm establishes a novel pathway for reconciling data-driven flexibility with physical consistency in regulated basin modeling. • Hybrid model integrates HSPF and RF for dam-regulated streamflow simulation. • RF addresses dam operation non-linearity, enhancing prediction accuracy. • Hybrid model reduces overfitting observed in standalone RF model. • Ensemble approach improves extreme event simulation. • Streamflow are transformed to extreme flows by dam operation.