Toward Computationally Efficient and Accurate Real-time Ensemble Flood Forecasting using Physics-Informed Surrogate Modeling with Extrapolating Capability

Abstract

Extreme floods occur more frequently than in the past due to climate warming, and they have more profound socio-economic impacts. Flood forecasting is one of the important components of flood risk management and mitigation but is subject to multiple uncertainties caused by meteorological inputs, initial states, model structures, and model parameters. Numerous research efforts investigated the uncertainties in the tasks of flood prediction. However, at present we entirely lack comprehensive studies that can handle long-lasting challenges of computational burden, inaccuracy, and unreliable predictability in real-time ensemble flood forecasting with uncertainty quantification. This dissertation aims to gain comprehensive knowledge of building novel modeling frameworks for computationally efficient and accurate real-time ensemble flood forecasting with uncertainty quantification. In this dissertation, a series of innovative methodologies have been developed for accurate, robust, and efficient uncertainty quantification of hydrological models in predicting floods. These methods include: (i) a unified modeling framework based on generalized likelihood uncertainty estimation (GLUE) framework coupled with polynomial chaos expansion (PCE) for fast and robust quantifying and understanding the parameter uncertainty of hydrological model in flood predictions; (ii) a novel modeling framework, for computationally efficient and accurate real-time ensemble flood forecasting with uncertainty quantification, which combines three modeling techniques together for the first time: surrogate modeling, parameter inference, and data assimilation; (iii) a novel, robust and efficient surrogate data assimilation approach for real-time flood forecasting using PCE to replace internal processes of Ensemble Kalman filters (EnKFs); and (iv) a new surrogate model, named polynomial chaos-kriging (PCK), that can provide reliable ensemble results, even for extreme events that deviate significantly from the training data space. Corresponding major accomplishments of this dissertation are abridged as follows. (i) The PCE surrogate model is firstly integrated into the GLUE framework to offset the computational demands of an uncertainty quantification task. It provides the benefits of an interpretable, probabilistic framework on which to make inferences about the drivers of model behavior, as well as the sensitivities of the model’s output to the uncertain inputs. (ii) The novel framework of real-time ensemble flood forecasting embraces the benefits of three modeling techniques together for the first time: (1) PCE surrogates can significantly decrease computational time; (2) Parameter inference (GLUE) allows for model faster convergence, reduced uncertainty, and superior accuracy of simulated results; and (3) EnKFs assimilate errors that occur during forecasting. This framework provides a holistic, robust approach to accounting and understanding the uncertainties of hydrological parameters and vastly reducing the computational burden of ensemble simulations in real-time flood prediction. This modeling framework contributes to a shift in modeling paradigm arguing that complex, high‐fidelity hydrologic and hydraulic models should be increasingly adopted for real‐time and ensemble flood forecasting. (iii) The power of surrogate approaches is further exploited to develop new surrogate filters by replacing the internal processes of the EnKFs with PCE. A comprehensive investigation into how to configure a surrogate filter indicates that the new partial (replacing part of original filters) and invariant (valid for entire time periods) approaches are preferred in terms of accuracy and efficiency, which helps directly reduce the number of dimensions and bridge the gap between hindcasting and real-time forecasting. This proposed surrogate filter will be a promising alternative tool for performing computationally-intensive data assimilation in high-dimensional problems. And (iv) a new surrogate model named polynomial chaos-kriging (PCK) is developed by combining the advantages of two well-known surrogate models, PCE and kriging. This combination enabled streamflow prediction for extreme events that deviated significantly from the trained data space, and allowed for quantifying predictive uncertainty robustly and efficiently. This finding will ultimately inspire novel designs toward a potentially more comprehensive surrogate model.