Assessing Combinations of Artificial Neural Networks Input Meteorological Parameters to Improve Daily Runoff Simulation

Abstract

Hydrological models are one of the tools used for reconstruction or simulation, forecasting for the anticipation of future changes in the flow of a river, which allows better management of water resources during low flow periods and anticipation of flood risks during high water periods. Using artificial neural network (ANN) in the fields of hydrology and water resources, became advocated due to its capability of tackling, modeling and forecasting the problems that are nonlinear or stochastic within the Rainfall-Runoff (R-R) system. This research aims to test the practicability of using ANNs with two input configurations; to model the R-R relationship in two stations Mirebek and Ain Berda located in the Seybouse basin in Algeria. The 1st (ANN IP) considers only precipitation as an input variable for the daily flow simulation. The 2nd (ANN IIP,T,Hu) considers a combination of the Temperature and Humidity with precipitation at the model input. The results of the two models were compared through performance metrics, viz., Root Mean Square Error (RMSE),mean absolute error (MAE), Pearson's correlation coefficient (R), Nash Sutcliffe Efficiency coefficient (NSE), and through graphical interpretation (scatter plots and time series). Better flow simulations were provided by the three-input model for the two stations; where R=0.90, NSE=80.5% for Mirebek station and R=0.85, NSE=72% for Ain Berda. This result has confirmed that as much input variables are numerous, as more the model of ANN is efficient. The finding of this study indicates that the developed ANN models could be considered as a powerful tool for predicting Runoff.