Forward Prediction of Runoff Data in Data-Scarce Basins with an Improved Ensemble Empirical Mode Decomposition (EEMD) Model

Data scarcity is a common problem in hydrological calculations that often makes water resources planning and engineering design challenging. Combining ensemble empirical mode decomposition (EEMD), a radial basis function (RBF) neural network, and an autoregression (AR) model, an improved EEMD prediction model is proposed for runoff series forward prediction, i.e., runoff series extension. In the improved model, considering the decomposition-prediction-reconstruction principle, EEMD was employed for decomposition and reconstruction and the RBF and AR model were used for component prediction. Also, the method of tracking energy differences (MTED) was used as stopping criteria for EEMD in order to solve the problem of mode mixing that occurs frequently in EEMD. The orthogonality index (Ort) and the relative average deviation (RAD) were introduced to verify the mode mixing and prediction performance. A case study showed that the MTED-based decomposition was significantly better than decomposition methods using the standard deviation (SD) criteria and the G. Rilling (GR) criteria. After MTED-based decomposition, mode mixing in EEMD was suppressed effectively (|Ort| < 0.23) and stable orthogonal components were obtained. For this, annual runoff series forward predictions using the improved EEMD-based prediction model were significantly better (RAD < 11.1%) than predictions by the rainfall-runoff method and the AR model method. Thus, this forward prediction model can be regarded as an approach for hydrological series extension, and shows promise for practical applications.