Document Type
Presentation
Publication Date
8-1-2014
Abstract
Floods are the most common and widespread disasters in the world and are responsible for a greater number of damaging events than any other type of natural event. However, due to various uncertainties that originate from simulation models, observations, and forcing data, it is still insufficient to obtain accurate flood forecasting results with the required lead times. Recently, ensemble forecasting techniques based on data assimilation (DA) have become increasingly popular, due to their potential ability to explicitly handle the various sources of uncertainty in operational hydrological models. Difficulty lies in DA for flood forecasting because nonlinearity increases sharply during flood events and the probability of streamflow cannot be characterized by the Gaussian assumption. Particle filtering (PF), also known as sequential Monte Carlo (SMC) methods, is a Bayesian learning process in which the propagation of all uncertainties is conducted by a suitable selection of randomly generated particles without any assumptions about the nature of the distributions. In this paper, the performance of ensemble Kalman filtering (EnKF) and PF is assessed for short-term streamflow forecasting with a distributed hydrologic model, namely, the water and energy transfer processes (WEP) model. To mitigate the drawbacks of conventional filters, the ensemble square root filter (EnSRF) and the regularized particle filter (RPF) are implemented. For both the EnSRF and the RPF, sequential data assimilation is performed within a lag-time window to consider the response times of internal hydrologic processes. The proposed methods are applied to several catchments in Korea and Japan to assess their performance. The discussion will be focused on how non-Gaussian and non-linear property of floods affects updating results by two DA methods.
Comments
Session R32, Hydrologic Modeling: Forecasting