Dependence structure analysis of multisite river inflow data using vine copula-CEEMDAN based hybrid model

• Published in: PeerJ (San Francisco, CA) CA), 2020-11, Vol.8, p.e10285-e10285, Article e10285
• Main Authors: Nazir, Hafiza Mamona, Hussain, Ijaz, Faisal, Muhammad, Mohamd Shoukry, Alaa, Abdel Wahab Sharkawy, Mohammed, Fawzi Al-Deek, Fares, Ismail, Muhammad
• Format: Article
• Language: English
• Subjects: Canonical-vine; Case studies; Complete ensembe empirical mode decomposition with adaptive noises; Computational Science; Creeks & streams; Decomposition; Ecohydrology; Hydraulic flow; Natural Resource Management; Neural networks; Pair copula construction; Random variables; River basins; Rivers; Spatial and Geographic Information Science; Statistics; Stochastic models; Stream flow; Time series
• ISSN: 2167-8359; 2167-8359
• DOI: 10.7717/peerj.10285

Several data-driven and hybrid models are univariate and not considered the dependance structure of multivariate random variables, especially the multi-site river inflow data, which requires the joint distribution of the same river basin system. In this paper, we proposed a Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) Vine copula-based approach to address this issue. The proposed hybrid model comprised on two stages: In the first stage, the CEEMDAN is used to extract the high dimensional multi-scale features. Further, the multiple models are used to predict multi-scale components and residuals. In the second stage, the residuals obtained from the first stage are used to model the joint uncertainty of multi-site river inflow data by using Canonical Vine. For the application of the proposed two-step architecture, daily river inflow data of the Indus River Basin is used. The proposed two-stage methodology is compared with only the first stage proposed model, Vector Autoregressive and copula-based Autoregressive Integrated Moving Average models. The four evaluation measures, that is, Mean Absolute Relative Error (MARE), Mean Absolute Deviation (MAD), Nash-Sutcliffe Efficiency (NSE) and Mean Square Error (MSE), are used to observe the prediction performance. The results demonstrated that the proposed model outperforms significantly with minimum MARE, MAD, NSE, and MSE for two case studies having significant joint dependance. Therefore, it is concluded that the prediction can be improved by appropriately modeling the dependance structure of the multi-site river inflow data.