In flood forecasting and assessing flood-induced risks the responsible and timely computer-based modelling is of vital importance. Hydraulic and hydrologic (physically-based) models traditionally play here an important role and proved their usefulness. In recent years it has been shown that the so-called data-driven models (including various statistical approaches, neural networks, machine learning, fuzzy systems and chaos theory) successfully complement the physically-based ones and allow for quantification of predictions uncertainty. Data-driven modelling and its practical use in flood-related problems is the major topic of this work package.
Combination of physically-based simulation models and data-driven methods (hybrid modelling) leads to additional possibilities. Various types could be used in conjunction being combined or cross-checking each other. This approach will be considered as well.
The problem of quantification of uncertainty and risk will receive special attention. Comparison of various methods will be made, their applicability for flood warning systems tested, and recommendations provided.
Research in this work package will be organised along a number of subpackages (topics) associated both with domain areas and the techniques used.

Subpackage 1: Data-driven methods in river flood forecasting. In river flood forecasting the prediction of flows plays a crucial role. For example, the models used by RIZA, RIKZ and WL provide such forecasts based on the measurements at Lobith and upstream and involve hydrodynamic modelling. Novel data-driven methods allow for improvement. This was demonstrated under the RIKZ/RWS NAUTILUS project where neural networks, fuzzy logic and information theory models were used to generate forecasts that served as boundary conditions for the ZeeDelta model. In a number of other studies (on Huai River in China, Bagmati catchment in Nepal, Sieve catchment in Italy) data-driven models were used to establish rainfall-runoff relationships and make forecasts in flood management context.
The subpackage will focus on a number of open issues allowing for the following: improving accuracy of data-driven models, testing various types of them and their combinations (mixtures of specialised models), development of models with the mixture of nominal and real-valued inputs, for problems with noisy data and with data of high variability (especially important for flood forecasting), combining such models with physically-based hydrodynamic and hydrological models (thus building hybrid models), increasing the usability of models and incorporating them into decision frameworks.
The issue of complementing the flood forecasts with the uncertainty estimates will be addressed as well. Research efforts will be focussed on advancing these approaches to a level of practical tools. This topic is closely linked to the EU-funded FLOODsite project (2004-2008). PhD study at IHE and 1-2 MSc students are foreseen.
Participating researchers: D.P. Solomatine, R.K. Price (both IHE), P. van Gelder (TUD), A. Mynett, (WL), M. Verlaan (RIKZ).

Subpackage 2: Predicting sea water levels and surges in the coastal zone. Safe operation of the coastal defence structures depends on the accurate forecast of high sea water levels. In the framework of the NAUTILUS project (RWS) and the Delft Cluster 1 ‘Data mining’ project (in cooperation with Directie Noordzee) it was convincingly demonstrated that the use of data-driven methods like neural networks and non-linear dynamics (chaos theory) could considerably improve such predictions by complementing the Dutch Continental Shelf Model. Apart from that, these methods make it possible to generate uncertainty estimates of water level predictions (confidence intervals and linguistic descriptions). These approaches will be further extended and thoroughly tested with the new data. This topic has links to the EU-funded FLOODsite project (2004-2008). PhD study at IHE and 1-2 MSc students are foreseen.
Participating researchers: D.P. Solomatine, R.K. Price (both IHE), H. Keyser (RWS-DNZ).

Subpackage 3. Methods of multi-variate statistics, extreme values analysis and Bayesian statistics. These methods traditionally play an important role in analysing uncertainty and making risk assessment. They closely relate to a number of methods used in data-driven modelling and, in fact, form the basis for many of them. The novel methods from this category and their combinations will be addressed, and compared to other methods of quantifying uncertainty.
Extreme quantile estimates of environmental variables, such as wind, waves, discharge and sea level, corresponding to return periods of the order of 50 to 10,000 years are utilized in the design and assessment of civil engineering infrastructures. The peaks-over-threshold (POT) method is a widely used approach for extreme value estimation. Although POT approach is conceptually simple and has received considerable attention, its practical applications are confounded by the problem of inhomogeneous data and of selection of a suitable threshold, which is not known a priori. Research is necessary on the variation of quantile uncertainty (bias and variance) as a function of threshold with the purpose of developing empirical criteria for optimal threshold selection. The bootstrap method can be used only in case of a single random sample. Standard non-parametric version of Efron’s bootstrap method is rarely applicable to extreme quantile estimation, because the available sample may not contain any observations in the region of tail extrapolation. To overcome these difficulties, other algorithms should be developed. PhD study on this topic at TUD is foreseen.
Participating researchers: P. van Gelder, J.K. Vrijling (both TUD), D.P. Solomatine (IHE), A. Vrouwenvelder, M. de Wit (both TNO-Bouw).

Subpackage 4: Improving probability analyses of 1:10,000-yr storm-surge levels and waves characteristics using novel methods of dating deposits. For obtaining historical extreme storm-surge levels traditional practice is to use the one-century based extrapolation methods. In the dunes of North-Holland, extreme storm-surge deposits occur about 7 m above sea level. These deposits can be dated (using optically stimulated luminescence) and added to the existing data series used in extrapolation. A similar approach can be applied for reconstruction of waveheight/-period combinations during extreme storms. These data can be obtained from the Oyster Grounds, north of the Wadden Islands, at water depths of 35m and more.
Participating researchers: Ad van der Spek, Sytze van Heteren (both TNO-NITG), Marcel Stive (TUD-CiTG), Jacob Wallinga (TUD-IRI).

Subpackage 5. Management and cross-links with other WPs of the project. This topic is established to stress and organisationally ensure the tight cooperation with the other WPs of the project. Especially close links are foreseen with the following work packages:
WP A1 (“Genese van rivierafvoergolven”) where data-driven methods will be combined with the physically-based watershed models. Participating researchers: D.P. Solomatine (IHE).
WP B (“Loads on and strength of flood defence systems”), where the use of data-driven modeling in analysing the strength of flood defence systems is foreseen, in connection with the GeoBrain-initiative at GeoDelft. Apart from linking pure data like (possible) flood levels, also knowledge exchange on the techniques used will take place. Participating researchers: A.R. Koelewijn (GD).

This WP is based on the results achieved in the project “Data mining, knowledge discovery and data-driven modelling” of the phase 1 of Delft Cluster and will extensively use its knowledge and algorithmic base.
The work package is not only oriented at research into the effective methods of flood forecasting and uncertainty analysis, but also at making them practical tools in operational flood forecasting and risk management.

Safety against flooding HOME