Foundational methods for model verification and uncertainty analysis (Invited)
Abstract
Before embarking on formal methods of uncertainty analysis that may entail unnecessarily restrictive assumptions and sophisticated treatment, prudence dictates exploring one's data, model candidates and applicable objective functions with a mixture of methods as a first step. It seems that there are several foundational methods that warrant more attention in practice and that there is scope for the development of new ones. Ensuing results from a selection of foundational methods may well inform the choice of formal methods and assumptions, or suffice in themselves as an effective appreciation of uncertainty. Through the case of four lumped rainfall-runoff models of varying complexity from several watersheds we illustrate that there are valuable methods, many of them already in open source software, others we have recently developed, which can be invoked to yield valuable insights into model veracity and uncertainty. We show results of using methods of global sensitivity analysis that help: determine whether insensitive parameters impact on predictions and therefore cannot be fixed; and identify which combinations of objective function, dataset and model structure allow insensitive parameters to be estimated. We apply response surface and polynomial chaos methods to yield knowledge of the models' response surfaces and parameter interactions, thereby informing model redesign. A new approach to model structure discrimination is presented based on Pareto methods and cross-validation. It reveals which model structures are acceptable in the sense that they are non-dominated by other structures across calibration and validation periods and across catchments according to specified performance criteria. Finally we present and demonstrate a falsification approach that shows the value of examining scenarios of model structures and parameters to identify any change that might have a specified effect on a prediction.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2013
- Bibcode:
- 2013AGUFM.H31I..02J
- Keywords:
-
- 1804 HYDROLOGY Catchment;
- 0550 COMPUTATIONAL GEOPHYSICS Model verification and validation;
- 1847 HYDROLOGY Modeling;
- 3275 MATHEMATICAL GEOPHYSICS Uncertainty quantification