University of Twente Student Theses


Tracking the uncertainty in streamflow prediction through a hydrological forecasting system

Pham, Trang Van (2011) Tracking the uncertainty in streamflow prediction through a hydrological forecasting system.

[img] PDF
Abstract:A flood forecasting system is a complex system which consists of many different components and each of these components can contain, to some extent, an uncertainty. Studying the uncertainties in flood forecasting, quantifying and propagating them through the system can help to gain more information about the different sources of uncertainty that may affect the forecasts. This information can later be added to the forecasts to improve their quality. These issues bring several challenges to the study of flow forecasting uncertainty: firstly, what is the impact of different sources of uncertainty on the quality of flood? Secondly, forecasts among all sources of uncertainty that stem from different components of the system, which sources significantly affect flood forecasts? Thirdly, which methods can be used to efficiently quantify and propagate those uncertainties through a forecasting model? Finally, which measures should be used to evaluate the uncertainty quantification and their impact on the quality of the forecasts? The aims of this research is to quantify and propagate the different kinds of uncertainty sources which play a role in flood forecasting; and to investigate methods to assess the quantified uncertainties and proper measures to evaluate the uncertainty quantification. In this research, the GRPE forecasting system, an ensemble prediction system based on the lumped GRP hydrologic model, is applied to three catchments in France. The uncertainties from precipitation data (input precipitation which is used for flow simulation and forecast precipitation used for flow forecasting), hydrological initial conditions (discharge data) and model parameters, which are acknowledged as important sources of uncertainty in hydrological modelling and forecasting, are studied. They are individually quantified and then propagated together through the forecasting system with an experimental approach by multiplying the simulations. The model structure uncertainty is not considered in the scope of this research. Methods for uncertainty quantification are defined and applied to each source of uncertainty. Two ensemble prediction systems from ECMWF and Météo-France are used to account for the forecast precipitation uncertainty for lead times from 1 to 9 days. For the uncertainty of input precipitation data, geo-statistical simulations of spatially averaged rainfall, conditioned on point data, and available for one study catchment, are chosen to provide the multiple statistical realizations of daily spatial rainfall fields over the study area. The hydrologic initial condition uncertainty is quantified by using an ensemble of discharges to update the state of the routing reservoir of the forecasting model. These discharges are retrieved from the analysis of uncertainties affecting the rating curves of each study catchment. Ten different periods of data, with the length of 5 years each, are selected to calibrate the model and thus to account for calibration period uncertainty. Finally, the Generalized Likelihood Uncertainty Estimator (GLUE) method is alternatively used to quantify the parameterization uncertainty. This is done by taking a large number of 125.000 sets of parameters to find the confidence intervals. To assess the results of uncertainty quantification, two probabilistic evaluation measures, the Brier (Skill) Score and the reliability diagram, are employed. In addition, confidence intervals of the forecasts are used to visualize the outcome of the research. Uncertainty in flood forecasting 3 The results show that input precipitation uncertainty does not have noticeable impact on the forecast output. This may be due to the method used to quantify the uncertainties from this source, which may be inappropriate to correctly capture them. For the catchments studied, this source of uncertainty can, therefore, be neglected when propagating different sources of uncertainty through the system. The other sources of uncertainty show large impacts on flow forecasts. Initial condition uncertainty shows large impacts for small lead times (up to 2 days). After that, forecast precipitation uncertainty has the largest impact; this impact is more significantly pronounced at high lead times. Depending on the catchment, parameter uncertainty can have more impact if it is evaluated from the variation of the calibration period or from the GLUE method. Based on the results of this research, and on the catchments and methods investigated, it is recommended to take into consideration the uncertainty of forecast precipitation, initial condition and model parameters in flood forecasting. There are different ways to account for parameter uncertainty, but the proposed approach of using different calibration periods proved to be a simple method but able to improve the quality of the forecast outputs.
Item Type:Essay (Master)
Cemagref Antony-Research Unit HBAN
Faculty:ET: Engineering Technology
Subject:56 civil engineering
Programme:Civil Engineering and Management MSc (60026)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page