Optimizing bed shear stress prediction in open flow channels: an investigation of heuristic machine learning techniques

The prediction of hydraulic parameter such as bed shear stress (τb) is a challenging task in context of flash floods. In order to predict bed stresses amidst rapidly fluctuating discharge, this study aims to utilize machine learning (ML) models to adequately predict τb in open channels. In this regard, four ML algorithms such as K-nearest neighbors (KNN), artificial neural networks (ANN), multilayer perceptron (MLP) and random forest (RF) have been incorporated to predict τb using seven input features (IFs) from IF1 to IF7 in 577 data points. The coefficient of determination (R2), Taylors diagram, Sensitivity analysis, SHapley Additive exPlanations, Regression error characteristics curves, error metrics, and box plots have been inspected. Also, K-fold cross validation is conducted for comprehensive assessment of predictive achievement of ML models. The results revealed that RF and KNN showed better results in comparison to other models having R2 values equal to 0.99 in IF4. The identified best-fit model, RF serves as a valuable tool for engineers, enabling to make accurate and reliable predictions of τb. The predictive capability empowers proactive measures to prevent potential damage from flash floods to hydraulic infrastructures ensuring community safety and protection of vital water resources in future flood-related challenges.