A Framework for Algorithmic Improvement to Mitigate the Effects of Equifinality in the Calibration of High-dimensional Parameters for Hydrological Models

• Published in: Water resources management Vol. 38; no. 1; pp. 251 - 267
• Main Authors: Tang, Xiaoyang, Tang, Deshan, Zhang, Fulin
• Format: Journal Article
• Language: English
• Published: Dordrecht Springer Netherlands 01.01.2024; Springer Nature B.V
• Subjects: Algorithms; Atmospheric Sciences; Calibration; Civil Engineering; Earth and Environmental Science; Earth Sciences; Environment; Evolutionary algorithms; Geotechnical Engineering & Applied Earth Sciences; Hydrogeology; Hydrologic models; Hydrology; Hydrology/Water Resources; Objective function; Parameter sensitivity; Parameters; prediction; Searching; shrinkage; water; Hydrological models; Equifinality; Parameter calibration; Algorithm improvement; SCE_UA algorithm
• ISSN: 0920-4741; 1573-1650
• DOI: 10.1007/s11269-023-03667-x

When performing hydrological model parameter calibration, equifinality inevitably reduces the simulation and prediction ability of hydrological models. To lessen the impact of equifinality, a novel algorithmic improvement framework is proposed in this paper. This framework allows the parameters to be searched hierarchically in order of sensitivity size and shrinks the original ranges of the parameters before the final search. The shuffled complex evolution (SCE_UA) algorithm, which is the most popular method for addressing hydrological model calibration issues, is improved using this new framework yielding HSRS_SCE algorithm, which stands for the SCE_UA algorithm with hierarchical search (HS) and range shrinkage (RS). A 26-dimensional parametric calibration problem is constructed and solved in this study utilizing 12 schemes based on the HSRS_SCE algorithm with various parameters and a control scheme based on the SCE_UA algorithm. The results show that the HSRS_SCE algorithm can not only produce calibrated parameter results significantly superior to those of the traditional scheme (p < 0.05) but also produce objective function values 26.1% better than those of the traditional scheme and reduce the search time through parallel computing.