New approaches toward efficient and robust uncertainty quantification in real-time flood forecasting

In the past decade, with three main problems in a rainfall-runoff model of real-time flood prediction including accuracy, predictability, and computational efficiency, many approaches have been developed. However, most of the prior studies concentrated on the objective of improving predictive accuracy without evaluating the effectiveness. Therefore, in this study, a new coupling unified framework is proposed, which can improve the predictive performance with three issues above. First, to increase the forecast accuracy and predictability, the sequential data assimilation is implemented to address the uncertainty sources such as model state and model parameter. Ensemble Kalman filter (EnKF) (Evensen 1994) is selected instead of the particle filter (PF) because this method is computationally cheaper than PF as it generally requires less ensemble members based on the sequential Monte Carlo method. In addition, in initialization of EnKF in prior studies the initial ensemble of model state and parameter are generated according to the uniform distributions in the corresponding ranges and the initial state values of zero are set (so-called Random method) while in this study, Generalized Likelihood Uncertainty Estimation (GLUE) (Beven and Binley 1992) method is suggested as an approach to initialize the ensemble member of states and parameters to robust increase the convergence in assimilation EnKF. The use of GLUE also improves the accuracy and predictability of the forecast. Finally, the EnKF method required a large of a number of models runs to obtain the ensemble results lead to very PCU intensive while time is important for timely flood warning and emergency responses. Therefore, the surrogate model based on Polynomial chaos expansion (PCE) (Ghanem and Spanos 1991; Wiener 1938) is considered to alternate the conceptual rainfall runoff (CRR) model, which can significantly increase the computational efficiency of the forecast. In this work, we investigate and tackle and essential estimation problem that is the accuracy, predictability, and computational efficiency of the hydrological model on real-time flood forecasting depending on different approaches of coupling framework, the optimal unified framework to robust quantify and reduce uncertainties in hydrologic prediction is proposed. The results of 18 approaches will show that: (i) the efficiency of computation of proposed unified framework on real-time forecasting (ii) the effect of GLUE on initialization of real-time forecasting; (iii) the capable of PCE model on real-time forecasting; and (iv) the effect of EnKF on real-time forecasting.