HR: 1340h
AN: H43A-0492    [Abstracts]
TI: Combining the Strengths of Physically Based Models with Statistical Modelling Tools Using a Hierarchical Mixture of Experts Framework
AU: * Marshall, L A
EM: lucy@civeng.unsw.edu.au
AF: School of Civil and Environmental Engineering, The University of New South Wales, UNSW , Sydney, NSW 2052 Australia
AU: Sharma, A
EM: a.sharma@unsw.edu.au
AF: School of Civil and Environmental Engineering, The University of New South Wales, UNSW , Sydney, NSW 2052 Australia
AU: Nott, D
EM: d.nott@unsw.edu.au
AF: School of Mathematics, The University of New South Wales, UNSW, Sydney, NSW 2052 Australia
AB: Rigidity in a modelling framework has been known to result in considerable bias in cases where the system behaviour is closely linked to the catchment antecedent conditions. An alternative to accommodate such variations in the system makeup is to enable the model to be flexible enough to evolve as antecedent conditions change. We present a framework that incorporates such flexibility by expressing the model through the combination of a number of different model structures. Each structure is adopted at a given time with a probability that depends on the current hydrologic state of the catchment. This framework is known as a Hierarchical Mixture of Experts (HME). When applied in a hydrological context, the HME approach has two major functions. It can act as a powerful predictive tool where simulation is extended beyond the calibration period. It also offers a basis for model development and building based on interpretation of the final model architecture in calibration. The probabilistic nature of HME means that it is ideally specified using Bayesian inference. The Bayesian approach also formalises the incorporation of uncertainty in the model specification. The interpretability of the overall HME framework is largely influenced by the individual model structures. One model which can be applied in the HME context is the popular Topmodel. Topmodel is a modelling tool that allows the simulation of distributed catchment response to rainfall. Many different versions of the basic model structure exist as the underlying concepts are challenged by different catchment studies. One modification often made is to the description of the baseflow recession. This study will investigate the predictive capability of Topmodel when the model is specified using both a Bayesian and HME approach. The specification of the distribution of model errors is investigated by definition of several different probability distributions. The HME approach is applied in a framework that compares two subtle modifications of the Topmodel structure, to see if one version is favoured over another at different times. The study will critically assess possible limitations of the HME approach, including computational issues, the effect of data availability and model complexity. The HME approach is largely a statistical modelling tool, whilst Topmodel is based on physical principles. If we apply the HME approach to a distributed model, are the results physically interpretable? Do the component models truly represent different catchment `states' and are these states predictable?
DE: 1805 Computational hydrology
DE: 1847 Modeling
DE: 1873 Uncertainty assessment (3275)
SC: Hydrology [H]
MN: Fall Meeting 2005