Quantifying Streamflow Prediction Uncertainty Through Process‐Aware Data‐Driven Models

ABSTRACT The hydrological model simulation accompanied with uncertainty quantification helps enhance their overall reliability. Since uncertainty quantification including all the sources (input, model structure and parameter) is challenging, it is often limited to only addressing model parametric uncertainty, neglecting other uncertainty sources. This paper focuses on exploiting the potential of state‐of‐the‐art data‐driven models (or DDMs) in quantifying the prediction uncertainty of process‐based hydrological models. This is achieved by integrating the robust predictive ability of the DDMs with the process understanding ability of the hydrological models. The Bayesian‐based data assimilation (DA) technique is used to quantify uncertainty in process‐based hydrological models. This is accomplished by choosing two DDMs, random forest algorithm (RF) and support vector machine (SVM), which are distinctly integrated with two process‐based hydrological models: HBV and HyMOD. Particle filter algorithm (PF) is chosen for uncertainty quantification. All these combinations led to four different process‐aware DDMs: HBV‐PF‐RF, HBV‐PF‐SVM, HyMOD‐PF‐RF and HyMOD‐PF‐SVM. The assessment of these models on the Baitarani, Beas and Sunkoshi river basins exemplified an improvement in the accuracy of the daily streamflow simulations with a reduction in the prediction uncertainty across all the models for all the basins. For example, the nash‐sutcliffe efficiency improved by 54.69% and 10.61% in calibration and validation of the Baitarani river basin, respectively. Equivalently, average bandwidth improved by 79.37% and 71.59%, respectively. This signified the (a) potential of the DDMs in quantifying and reducing the prediction uncertainty of the hydrological model simulations, (b) transferability of the model with an appreciable performance irrespective of the choice of basins having varying topography and climatology and (c) ability to perform significantly irrespective of different process‐based and DDMs being involved, thereby ensuring generalizability. Thus, the framework is expected to assist in effective decision‐making, including various environmental management and disaster preparedness.