Sensitivity and Interaction Analysis Based on Sobol’ Method and Its Application in a Distributed Flood Forecasting Model

Water

Hui Wan¹  Jun Xia¹  Liping Zhang¹  Dunxian She¹  Yang Xiao¹  Lei Zou¹ 

[1] State Key Laboratory of Water Resources and Hydropower Engineering Science, Wuhan University, No.8 Donghu South Road, Wuhan 430072, China; E-Mails:

关键词: sensitivity analysis;    interaction analysis;    correlated parameters;    objective functions;    flood magnitudes;    Sobol’ method;   

DOI  :  10.3390/w7062924

来源: mdpi

【 摘 要 】

Sensitivity analysis is a fundamental approach to identify the most significant and sensitive parameters, helping us to understand complex hydrological models, particularly for time-consuming distributed flood forecasting models based on complicated theory with numerous parameters. Based on Sobol’ method, this study compared the sensitivity and interactions of distributed flood forecasting model parameters with and without accounting for correlation. Four objective functions: (1) Nash–Sutcliffe efficiency (E_NS); (2) water balance coefficient (WB); (3) peak discharge efficiency (E_P); and (4) time to peak efficiency (E_TP) were implemented to the Liuxihe model with hourly rainfall-runoff data collected in the Nanhua Creek catchment, Pearl River, China. Contrastive results for the sensitivity and interaction analysis were also illustrated among small, medium, and large flood magnitudes. Results demonstrated that the choice of objective functions had no effect on the sensitivity classification, while it had great influence on the sensitivity ranking for both uncorrelated and correlated cases. The Liuxihe model behaved and responded uniquely to various flood conditions. The results also indicated that the pairwise parameters interactions revealed a non-ignorable contribution to the model output variance. Parameters with high first or total order sensitivity indices presented a corresponding high second order sensitivity indices and correlation coefficients with other parameters. Without considering parameter correlations, the variance contributions of highly sensitive parameters might be underestimated and those of normally sensitive parameters might be overestimated. This research laid a basic foundation to improve the understanding of complex model behavior.

【 授权许可】

CC BY   
© 2015 by the authors; licensee MDPI, Basel, Switzerland.