Hybrid Daily Streamflow Forecasting Based on Variational Mode Decomposition Random Vector Functional Link Network-Based Ensemble Forecasting

Abstract

Streamflow forecasting using advance machine learning models have received great importance during the last few years regarding its importance for water resources management, especially for facing climate change. Several approaches based on the exploitation of a wide variety of models have been proposed and successfully applied for accurately daily and monthly streamflow forecasting. However, since streamflow and rainfall are closely interconnected, they were always combined for building more robust forecasting models. While, other climatic variables, i.e., temperature and evapotranspiration, were rarely, if ever, combined for streamflow forecasting, an important part of the developed models used only the value of streamflow measured at previous time lag as input variables. Recently, the use of signal processing decomposition algorithms, i.e., wavelet decomposition (WD) and more recently the variational mode decomposition (VMD), has attracted considerable attention and its success was highlighted up to this date without serious criticism. In the present chapter, we introduce a new scheme for daily streamflow forecasting using the random vector functional link network (RVFL) combined with the VMD. The VMD was used for decomposing the streamflow signal, and then the different intrinsic mode functions (IMF) were used as input variables. For more in depth conclusions, obtained results using the RVFL were compared to those using the extreme learning machine (ELM). Models accuracies were evaluated using several performances metrics and, overall, our best estimation resulted in an overall low RMSE and MAE, and high correlation between measured and predicted streamflow. Furthermore, the best forecasting accuracies were obtained using the RVFL combined with the VMD, for which the R and NSE values were ranged from 0.922 to 0.995 and from 0.850 to 0.991 using the RVFL_VMD compared to the values of 0.836–0.947 and 0.691–0.898 obtained using the ELM_VMD. It was found that the gained improvement in terms of model performances was more significant using the RVFL models compared to the ELM models.