Simultaneous Calibration of Hydrologic Model Structure and Parameters Using a Blended Model
Abstract
The identification of the optimal model structure within a hydrologic model has been a challenge noted in the literature and practice for decades. The use of a one size fits all fixed model structure approach is common practice, but is limiting in both predictive performance and hydrologic insight. A number of flexible model frameworks have emerged in literature, and provide the tools for enumeration and comparison of related model structures. However, no direct calibration of model structure was studied until recently, where an integer calibration approach allowed the model to switch between pre-defined hydrologic process algorithms during the calibration. This study builds on these concepts and presents a novel approach for directly calibrating the structure of a hydrologic model using only continuous variables in the optimization scheme; this configuration is referred to as the blended model. This experiment is carried out using the Raven Hydrologic Modelling Framework, an open source project focused on a flexible and modular approach to hydrologic modelling. This study takes the form of a model calibration experiment, in which the blended model structure is optimized and compared with 108 discrete permutations of optimized model structures. Each of these models are deployed for twelve of the MOPEX catchments, and individually calibrated with multiple parallel iterations using the Dynamically Dimensioned Search algorithm. The results indicate that the simultaneous calibration of model structure and parameters produces a robust performance across the twelve selected catchments, with an overall average calibration NSE value of 0.73, and the top rank across all structures when averaged over all catchments and iterations. The results also provide insight on the structural identifiability within each hydrologic process group, and generates similar insights to the exhaustive enumeration of optimized model structures with a fraction of the computational cost. This approach has the potential to both improve overall predictive model performance and reduce the effort required in calibrating multiple model structures in exploratory studies.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.U11D..08C