Views
The EFFS flood forecasting cascade
Introduction
The GLUE methodology was used in one of the first studies of making use of the ECMWF ensemble weather prediction system in flood forecasting. This work was carried out under the EU funded European Flood Forecasting System (EFFS) led by the EU Joint Research Centre, Ispra, Italy (De Roo et al., 2003; Pappenberger et al., 2005). EFFS later evolved in the operational European Flood Alert Service that routinely makes up to 10 year forecasts for all the major river basins of Europe.
Case study: details
The EFFS system involved a cascade of models from the ECMWF 50 member ensemble weather forecasts, through the LISFLOOD-FF rainfall-runoff model running at a 1 km grid scale and hourly time step, to the LISFLOOD-FP 2 dimensional flood routing model running at a grid scale of 100m. Initial conditions are set up prior to event running a daily water balance version, LISFLOOD-WB, with measured rainfall inputs. Prior assessment of the LISFLOOD and LISFLOOD-FP to obtain a set of behavioural models for use within GLUE was carried out using historical data (Pappenberger et al., 2004a)
The cascade therefore involved a propagation of uncertainty from the atmospheric inputs, through the rainfall-runoff model to the hydraulic model. At each level in the cascade there were uncertain model inputs and uncertain model parameters, leading to uncertain model predictions that could be compared against any observed data to allow the GLUE weights associated with each model to be updated as if in real time.
To make the cascade of predictions computationally feasible, a way of classifying the rainfall-runoff model responses into 6 functionally similar types was implemented so that only representative model runs needed to be made (Pappenberger et al., 2004b, 2005), each weighted within the GLUE prediction framework to be equivalent to a set of equivalent models.
Results
The methodology for cascading uncertainty through the flood forecasting system was demonstrated by an application to the Meuse catchment, upstream of Maaseik (~21,000 km2) which flows through France, Luxembourg, Belgium and the Netherlands. Discharge observations are also available at the Borgharen gauging station, at the upstream boundary of the 35 km reach simulated by the LisFlood?-FP flood inundation model.
Two different 10 day ahead forecast scenarios were available for the model evaluation, one starting on the 21 of January 1995 at 12:00 and the other one on the 22 of January 1995 also at 12:00. The second set has been chosen to evaluate possible updating techniques for the runoff model, which could be also extended to the inundation evaluation.
The rainfall-runoff model has been calibrated on hourly flow data starting from 23rd of December 1994 right up to the starting time of the forecasts. The likelihood measure used is the Multicomponent Mapping method as described in Pappenberger and Beven (2004a). This methodology allows the definition of errors around evaluation data as a measure designed the form of a pyramidal frustrum in time and magnitude around observations. The performance of hydrograhs is computed according to combinations of the measure over all time steps. The result is a form of fuzzy measure that can be used in weighting those simulation models that are retained as behavioural. The behavioural parameter sets are then clustered into functional classes, which show a similar flow behaviour.
The flood plain model was pre-conditioned on observational data available for the December 1993 flood event which consisted of a mosaic of air photo images of maximum inundation extent and two stage hydrographs internal to the model domain.
During the event the weights associated with each model component can be updated in real time. This does not, however, preclude the possibility that the models available will not be able to reproduce the observed discharges (as seen in Figure 1),
Figure 1: 10 Day ahead discharge predictions at Borgharen gauging station on the River Meuse. A. Weighted combination of ECMWF ensemble inputs with 6 representative rainfall runoff models and hydraulic routing models. B. Control run with 6 representative rainfall runoff models. C. Deterministic forecast with 6 representative rainfall-runoff models (from Pappenberger et al., HESS, 2005).
Figure 2: Quantile predictions for inundation model at time of the SAR overpass. Solid line is estimate of actual inundation at that time from the satellite image (from Pappenberger et al., HESS, 2005).
Comment
This case study demonstrates the use of GLUE in a cascade of model components with more than one source of conditioning data, as well as the possibility of updating the likelihoods weights as if in real-time. An additional study that used updating of the weights associated with behavioural inundation models in a real-time situation is provided by Romanowicz and Beven (2003).
Further information on using GLUE in the calibration of flood inundation models may be found in GLUE Case Study 2 <add link>
References
-
Beven, K. (2006). "A manifesto for the equifinality thesis." Journal of Hydrology 320(1-2): 18-36.
-
De Roo, A P J and 20 others, 2003, Development of a European flood forecasting system, Int. J. River Basin Management, 1(1), 49-59
-
Pappenberger, F, Beven, K J, de Roo, A., Thielen, J., and Gouweleeuw, G, 2004a, Uncertainty analysis of the rainfall runoff model LisFlood? within the Generalized Likelihood Uncertainty Estimation (GLUE), J. River Basin Management, 2, 123-133.
-
Pappenberger, F and Beven, K J, 2004b, Functional Classification and Evaluation of Hydrographs based on Multicomponent Mapping (Mx), J. River Basin Management, 2, 89-100.
-
Pappenberger, F., Beven, K.J., Hunter N., Gouweleeuw, B., Bates, P., de Roo, A., Thielen, J., 2005, Cascading model uncertainty from medium range weather forecasts (10 days) through a rainfall-runoff model to flood inundation predictions within the European Flood Forecasting System (EFFS). Hydrology and Earth System Science, 9(4),381-393.