Hierarchical Mixture of Experts and Diagnostic Modeling Approach to Reduce Hydrologic Model Structural Uncertainty

Abstract:

The choice of hydrologic model structures is one of the sources of uncertainty in representing hydrological process. In most applications, a single comprehensive hydrologic model structure might not be able to capture the entire complex and multi-scale interactions among the different components of the hydrologic process adequately. Calibrating such model can result in displacement of errors from structure to parameters, which in turn leads to over-correction and biased parameter values. An alternative to a single model structure is to develop local expert structures that are well suited in representing specific components of the hydrologic system and adaptively integrate them based on an indicator state variable. In this study, the Hierarchical Mixture of Experts (HME) architecture with a modified gating network function is applied to integrate two runoff module structures of the HBV model. The runoff module structures (i.e., buckets number and orientation) are proposed based on their expertise in representing recession flow and flow duration curve. This process based diagnostic framework of local experts provides a skilled platform for HME to effectively capture each distinct characteristic of the hydrograph and stochastically adapt to catchment response through soil moisture as an indicator variable. The approach is tested using two previously studied catchments, the Guadalupe River (Texas) and the French Broad River (North Carolina) from the Model Parameter Estimation Experiment (MOPEX). The results show that the HME approach has a better performance over a single model for both catchments in terms of the Nash Sutcliffe and correlation coefficient. Furthermore, we have developed and applied a comprehensive performance assessment matrix based on information theory to evaluate the differences between model and observation in terms of different characteristics of the hydrograph.