Streamflow simulation gives the major information on water systems to water resources planning and management. The monthly river flows in dry season often exhibit high autocorrelation.The headwater catchment of the Yellow River basin monthly flow series in dry season exhibits this clearly. However, existing models usually fail to capture the high-dimensional, nonlinear dependence. To address this issue, a stochastic model is developed using canonical vine copulas in combination with nonlinear correlation coefficients. Kendall's tau values of different pairs of river flows are calculated to measure the mutual correlations so as to select correlated streamflows for every month. Canonical vine copula is used to capture the temporal dependence of every month with its correlated streamflows. Finally, monthly river flow by the conditional joint distribution functions conditioned upon the corresponding river flow records was generated. The model was applied to the simulation of monthly river flows in dry season at Tangnaihai station, which controls the streamflow of headwater catchment of Yellow River basin in the north of China. The results of the proposed method possess a smaller mean absolute error (MAE) than the widely-used seasonal autoregressive integrated moving average model. The performance test on seasonal distribution further verifies the great capacity of the stochastic-statistical method.
IntroductionWith the global population continuing to increase, water resources are becoming ever more vital by more demand for urbanization and agricultural intensification [1,2]. In water resources planning, streamflow simulation in dry season is a paramount process in water and drought management, determination of river water flow potentials, environmental flow analysis, agricultural practices, and hydro-power generation [3,4].Compared with models which consider relatively steady physiographic, geological, soil, land use, and plant cover attributes in a site or watershed, the statistical models are simpler and more reliable for their principle of identifying relations between output variables with their predictors without any explicit knowledge of the physical processes [5,6]. The traditional statistical model consists of the parametric and nonparametric models. The most famous parametric models are autoregressive moving average models and autoregressive integrated moving average models proposed by Box and Jenkins [7]. They are established based on the linear regression method with auto-correlation function and partial autocorrelation function. The models and their variants are widely used for their practical nature, but also show limits such as normal assumption and linearity or inaccuracy coursed by transformation [8]. Lall and Sharma [9] proposed the nonparametric model of nearest neighbor resampling method to fit the skewed marginal distribution and nonlinear structure of river flow. This kind of nonparametric model performs well in inheriting statistic features of historical record in case of the high-class dataset...