Activity 1: Propagation of integral uncertainty through composite (hybrid) models

Methods for propagating integral uncertainty (that is including all sources) through complex models (employing various modeling paradigms, thus referred to as composite, or hybrid models) will be reviewed at the start of the project. In the first 18 months, attention will be focussed on the applicability of (1) efficient strategies of sampling in resampling uncertainty methods; (2) fuzzy logic methods (partly based on the foundation laid in the OSIRIS EU project). FLOODsite is novel in extending this work to include evaluation and development of statistical and data-driven methods for analysing the propagation of uncertainty through models. The following strands are foreseen.

Firstly, the development of non-parametric methods to analyse the integral (total) model uncertainty in the situations when model errors are not stationary across the space of possible hydrologic conditions. Such methods will be based on machine learning, and, more generally, computational intelligence techniques.

Secondly, efforts to enhance the resampling (Monte-Carlo) methods to make them more applicable in analysing the complex models’ parametric uncertainty. An approach to be tested here is combining the mentioned methods with the global optimization techniques.

Thirdly, the complex character of models may lead to the necessity of surrogate modelling, i.e. the use of data-driven (machine learning) techniques in models emulation. This is essential because the run-time of these models usually prohibits full Monte Carlo analysis.

In terms of computational intelligence technologies to be considered, attention will be given to fuzzy logic, artificial neural networks, decision and model trees and committee machines. These methods are to be evaluated in the first 18 months. In subsequent phases of the research these methods will be further refined and used for the localized non-parametric assessment of integral (total) model uncertainty based on model errors. The methodologies developed in this sub-task deal with model uncertainty, and are aimed at introducing uncertainty estimates into model forecasts of various types, and hence, through that, to decision practice at various levels of risk management, both for long-term planning, and emergency management. The methods developed will be implemented in software (research quality) and tested and compared in the context of pilot studies in (Activity 3).

Within this activity, Partner 20: UNESCO-IHE Institute for Water Education, Delft has developed uncertainty method so-called UNEEC (UNcertainty Estimation based on Local Errors and Clustering).