MARS for Prediction of Shear Force and Discharge in Two-Stage Meandering Channel

Abstract

Accurate prediction of shear stress distribution along the boundary in an open channel is the key to the solution of numerous critical engineering problems. This paper investigated the distribution of boundary shear force at the cross section of meandering compound channels. The research focused on developing a model for predicting shear force using a machine learning (ML) method, the multivariate adaptive regression spline (MARS). A nonparametric regression methodology was adopted in MARS for developing a model of shear force percentage in the floodplain of two-stage meandering channels. The width ratio, relative depth, sinuosity, bed slope, and meander belt width ratio of the channel were input variables to the model. The influence of each parameter on predicting the percentage of shear force in the floodplain was also analyzed by adopting a sensitivity analysis. Performance of the MARS model was evaluated by three different machine learning techniques—the group method of data handling (GMDH), support vector machine (SVM) and k-nearest neighbor (KNN)—through different statistical measures. The results indicated that the proposed MARS model predicted the shear force percentage in the floodplain satisfactorily, with a coefficient of determination (R2) of 0.94 and 0.93 and a scatter index (SI) of 0.053 and 0.044 for the training and testing phases, respectively. Moreover, the model was successfully applied for validating the two available overbank discharge values for the Baitarani River at Anandapur (drainage area of 8,570 km2), giving the minimum errors of the evaluated methods in terms of mean absolute scaled error (MASE) of 0.014 and 0.066 for flow depths of 7.5 and 8.63 m, respectively.