Evaluating Different Machine Learning Models for Runoff and Suspended Sediment Simulation

Abstract

In the present study, prediction of runoff and sediment at Polavaram and Pathagudem sites of the Godavari basin was carried out using machine learning models such as artificial neural network (ANN) and adaptive neuro-fuzzy inference system (ANFIS). Different combinations of antecedent stage, current day stage and antecedent runoff for current day runoff prediction and antecedent runoff, current day runoff and antecedent sediment for current day sediment prediction were explored using Gamma test (GT) to select the effective input variables for runoff and sediment prediction. The performance during training and testing periods of the ANN and ANFIS models were evaluated quantitatively through various statistical indices and qualitative by visual observation. After comparing the qualified results of different ANN and ANFIS models it was found that ANN model with double hidden layers and ANFIS model with membership function (Triangular, 3) performed well for runoff and sediment predictions, respectively for Pathagudem site. ANFIS model with membership function (Triangular, 3) and ANFIS model with membership function (Gaussian, 3) shown the best results for runoff and sediment prediction, respectively, for Polavaram site. The effect of input variables on the selected models was also validated by the way of sensitivity analysis. The results of sensitivity analysis was found that the current day runoff mostly depends on present day stage and present day sediment depends on current day runoff.