Evaluation of Data-Driven and Empirical Models for the Estimation of Sediment Removal Efficiency in Settling Basins

Recent studies have demonstrated the ability of data-driven models to predict sediment removal efficiency for settling basins, hydropower schemes, canals, and water treatment units. Empirical models, such as those cited herein, have also been employed for the estimation of sediment removal efficiency. This study evaluated the effectiveness of these empirical models as well as of data-driven models, such as the adaptive boost regressor (ADBR), extreme gradient boost regressor (XGBR), random forest (RF), and random tree (RT). A set of 328 observations on removal efficiency was used to determine the effectiveness of these models. Data analysis found that there was a high degree of variability in the performance of empirical models when measured by statistical indicators, such as the coefficient of determination (R2), root-mean-square error (RMSE), Nash-Sutcliffe efficiency (NSE), mean absolute error (MAE), and index of agreement (IA). The data-driven models outperformed empirical models in the prediction of sediment removal efficiency. The values of R2, RMSE, NSE, MAE, and IA ranged from 0.320 to 0.528, 0.319 to 0.251, 0.121 to 0.457, 0.249 to 0.195, and 0.662 to 0.827, respectively, for empirical models, and from 0.931 to 0.974, 0.092 to 0.055, 0.927 to 0.974, 0.064 to 0.038, and 0.981 to 0.993 for data-driven models. Among the empirical models, the Salmasi model exhibited a higher degree of accuracy, followed by the USBR, Garde, Raju, Camp-Dobbins, and Sumer models. The data-driven models were ranked from best to worst as ADBR, XGBR, RF, and RT, respectively. Additionally, the most suitable range of input parameters was assessed using the best-performing data-driven model ADBR. Data-driven models effectively captured the relationships between variables and the dynamic nature of the system under consideration, resulting in significantly improved predictive accuracy.