Hydrosystem models are subject to uncertainty in model structure, parameters, forcing terms and data. To specifically address conceptual uncertainty (the uncertainty in model choice), Bayesian model averaging (BMA) is chosen as an integrated modeling framework. BMA is a formal statistical approach that rests on Bayesian probability theory. Posterior model weights are obtained for a set of alternative conceptual models based on their individual skill in reproducing the observed data and the principle of parsimony. These weights can be used for model ranking, model selection or model averaging. The conceptual uncertainty within the set of considered models can be quantified as so-called between- model variance.
Despite the rigorous procedure, a number of obstacles exist(ed) to the wide-spread use of BMA in scientific research and practice. Some of these have been addressed in the course of my PhD project. Several options for evaluating the BMA equations efficiently and accurately have been compared (numerical methods vs. information criteria). Further, the shift in justified model complexity with an increasing amount of available data has been investigated, and a statistical concept to assess the robustness of model ranking against noisy input or output data has been proposed.
The focus of this post-doctoral project is on further extending the applicability of BMA by handling combinations of physically-based models with statistical error models. Error models are introduced to reduce the bias in predictions. The optimal complexity of such error models can be identified within the BMA framework. The goal is to infer improved model structures with realistic parameters that yield most accurate and reliable prognoses.
