Complementary data-intelligence model for river flow simulation

Abstract

Despite of diverse progressions in hydrological modeling techniques, the necessity of a robust, accurate, reliable, and trusted expert system for real-time stream flow prediction still exists. The intention of the present study is to establish a new complementary data-intelligence (DI) model called wavelet extreme learning machine (WA-ELM) for forecasting river flow in a semi-arid environment. The monthly river flow data for the period 1991-2010 is used to calibrate and validate the applied predictive model, developed using antecedent flow data as predictor. The prediction efficiency of the developed WA-ELM model is validated against stand-alone ELM model. The performance of the models is diagnosed using multiple statistical metrics and graphical analysis visualization. The results reveal that incorporation of data pre-processing wavelet approach with ELM model enhances the river flow predictability. In quantitative term, the root-mean-square error (RMSE) and mean absolute error (MAE) measurements are reduced by 65% and 67% using WA-ELM over ELM model, respectively. The Taylor diagram reveals much closer proximity and the Violin plot shows similar distribution of WA-ELM simulated river flow to the observed river flow compared to stand-alone ELM simulated river flow. The hybridization of wavelet decomposition method with ELM model improves the ability of ELM model to extract the required information for modeling the non-stationary and high stochastic river flow pattern. Overall, the study reveals that WA-ELM can be a reliable methodology for modeling river flow in semi-arid environment and for different regimes (i.e., low-, medium- and high-flow).