Predicting characteristics of dune bedforms using PSO-LSSVM

Dunes have a large influence on hydraulic roughness, and, thereby, on water levels which could affect the navigability of rivers and performance of hydraulic structures. The present study investigated the variation of geometric and topographic characteristics of dune bedforms and flow features as measured in laboratory studies (data sets from laboratory experiments) to estimate the roughness coefficient and characteristics of dune height. The Least Squares Support Vector Machine (LSSVM), which was optimized using Particle Swarm Optimization (PSO), was used as the Meta model approach to predict the values of interest. Developed models were separated into three categories: modeling using flow characteristics, modeling of flow and bedform characteristics, and modeling by using flow and sediment characteristics. It was found that for estimation of the roughness coefficient in open channels with dune bedforms, models developed based on flow and sediment characteristics performed more successfully. The model with input parameters of flow and grain Reynolds numbers (Re and Rb, respectively) and the ratio of the hydraulic radius (R) to the median grain diameter (D50) yields a squared correlation coefficient (R2) of 0.8609, a coefficient of determination (DC) of 0.7361, and a root mean square error (RMSE) of 0.0034 for a test series of Manning's roughness coefficient which was the most accurate model. Results proved the key role of flow Reynolds number (Re) values as an input feature for all models predicting the roughness coefficient. Accordingly, classic approaches led to poor results in comparison. On the other hand, results obtained for estimated values of relative dune height led to moderate prediction quality, which albeit, outperformed classic approaches.