Towards minimizing prediction uncertainty of hydrological models through Physics-aware machine learning models

Reliable streamflow simulations are vital for effective management of water resources. However, hydrological model simulations are often accompanied with uncertainty arising from input, model structure, and parameters. Quantifying and reducing all sources of uncertainty is often challenging, however crucial for improving the reliability of the model simulations. The study therefore aims to develop two process-aware machine learning models to reduce prediction uncertainty in the hydrological model simulations. For this endeavor, the potential of state-of-the-art machine learning models are exploited to reduce the total prediction uncertainty of the process-based hydrological models. This is achieved by coupling lumped hydrological models (HBV, and HyMOD) with a machine learning model (Random Forest, RF) through a Bayesian-based data assimilation technique (Particle filter, PF). The developed modeling framework is capable of improving the process understanding of hydrological models when integrated with Bayesian-based data assimilation technique that allows robust uncertainty quantification and machine learning model with relatively outstanding predictive ability. The resulting models (HBV-PF-RF and HyMOD-PF-RF) examined on the Sunkoshi River basin in Nepal exemplified a significant improvement in the model accuracy and reduction in prediction uncertainty. For example, the Nash-Sutcliffe Efficiency improved from 0.77 to 0.95 in calibration and 0.70 to 0.82 in validation for the HBV-PF-RF. Equivalently, an improvement of 0.81 to 0.97 and 0.83 to 0.84 was observed in calibration and validation respectively of the HyMOD-PF-RF.