Multi-model approach for combined outputs of rainfall-runoff models

Abstract

The multi-model approach is the integration of results obtained from competing models for improving modelling results that will be better than the result of the best individual model in the combination. It is an alternative framework for improving the accuracy and reliability of river flow simulations. To date, many applications in hydrology have applied this approach for improving the modelling results rather than developing new rainfall-runoff models. The proposed research aims to develop the multi-model approach for river flow simulations through case studies of two contrasting catchments located in Thailand and New Zealand, respectively. The research involves the use of three types of rainfall-runoff models, two empirical black-box models, two conceptual models and a semi-distributed physically based model. The two empirical black-box models selected for this research are the linear perturbation model (LPM) and the linear varying gain factor model (LVGFM). The two conceptual models selected in this study are the soil moisture accounting and routing (SMAR) model and the Nedbør-Afrstrømnings Model (NAM). The semi-distributed physically based model selected is the soil and water assessment tool (SWAT). These models were used to provide daily discharge estimates for both catchments. This research addresses important issues for developing the multi-model approach, namely, the optimal number of rainfall-runoff models applied in the multi-model combination systems, the comparison of non-linear combination methods (i.e. gene expression programming (GEP) and artificial neural networks (ANNs)), and the uncertainty analysis of the developed multi-model combination systems. The results showed that this research provides an improved multi-model approach for river flow simulation through case studies of Thailand and New Zealand catchments. The comparative results of the combination methods found that the most suitable method of combining methods in a multi-model combination system is based on the catchment type. The optimal number of rainfall-runoff models which best perform in multi-model combination systems depends on the selection of the number of rainfall-runoff models and the catchment type. The uncertainty analysis of multi-model results demonstrates that use of the bootstrap technique is very effective for quantifying the uncertainty associated with the combined results of the ANN multi-model combination systems. The research work detailed in this thesis has been reported in four journal papers and presented at three international conferences.