Download files
Abstract
Understanding the uncertainties associated with streamflow prediction in hydrological modelling has gained much interest over the past years, mainly because these uncertainties remain ambiguous and reduce the reliability of predicted streamflow hydrographs. Consequently, numerous approaches have been developed to address these uncertainties. This thesis provides an insight into the development of reducing two important sources of modelling uncertainty, model parameters and the underlying structure of the model prescribed and as well as improve the streamflow prediction.
Hierarchical Mixture of Experts (HME) cleverly provides a framework that addresses both parameter and structural uncertainties. The HME rationale consists of multiple components (hydrological models) or ensembles, which are then combined together in a dynamic fashion, providing opportunity to mix and share information. The responses from these models are assigned to different events of the catchment data, which vary from highflows, lowflows to periods following highflows or lowflows. The gating function within the HME calculates the model weights probabilistically, prioritising the preferred model according to the catchment predictors. To address the parameter uncertainty, the HME components incorporate Sequential Monte Carlo sampler with Bayesian inference as the preferred parameter sampler.
The HME with predictors has proven to reduce the structural uncertainty by achieving a better Bayesian Information Criterion (BIC). The BIC has been demonstrated to significantly improve across 50 catchments in Australia when multiple predictors are combined together. In this thesis, the mechanism of selecting the best combination of predictors, allowing for a better model mixing, according to different catchments characteristics is specified. Through this HME rationale, it has also been demonstrated that the coefficients can be regionalised to another catchments with similar important characteristics although located in different regions. This concept will be useful when implemented for prediction at ungauged catchments.