Quantitative simulation of diverged flow using machine learning techniques and FLOW3D numerical modeling

Intouduction:

Groynes at water intake locations significantly increase the flow diverted from rivers by optimizing incoming water control. Vaghefi and Ghodsian (2017) experimentally studied flow patterns around a T-shaped Groyne within a 90-degree arc using a moving bed. Shaker and Kashfipour (2013) compared flow velocity and shear stress distribution with and without Groynes. Behnam-Talab et al. (2018) simulated porous Groynes using FLOW-3D software. Shahinejad et al. (2022) applied a multi-objective algorithm to optimize T-shaped Groyne dimensions, achieving superior results, compared to previous designs. Zare and Honer (2016) investigated how simple Groynes reduce lateral erosion in river arches under laboratory conditions, emphasizing the influence of Groynes on erosion patterns. These studies collectively highlight the importance of Groyne design in enhancing water extraction and mitigating erosion.The review of literatures confirms that both laboratory and numerical studies have been conducted to examine the characteristics of various types of Groynes and the impact of flow patterns on them. However, there is a lack of studies addressing the simultaneous application and comparison of numerical and data-driven models in the investigation of geometric and hydraulic characteristics, particularly concerning the effect on the amount of diverted discharge from a canal into intake featuring T-shaped and L-shaped Groynes. Consequently, this research aims to evaluate the performance of two MLMs, specifically SVM and GEP in comparison with the Computational Fluid Dynamics (CFD)-based FLOW-3D model, on a laboratory scale.

Groynes play a crucial role in river engineering by regulating river flow. This study assesses the efficacy of two machine learning algorithms—support vector machine (SVM) and gene expression programming (GEP)—in comparison with FLOW-3D software for simulating diverted flow in a laboratory setting. The experimental model was tested in a laboratory flume with T-shaped and L-shaped Groynes positioned at 90 and 135-degree angles to channel the discharge into the intake system. The machine learning models incorporated three independent variables: the flow Froude number, the angle of water intake, and the relative length of the Groynes. Out of 96 laboratory data points, 70% were allocated for model training and 30% for model testing. Model performance was assessed using the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) indices.

 The results indicated that the GEP model surpassed the SVM model. For the L-shaped Groyne, the values for (R², MAE, RMSE) during both the training and testing phases were (0.9325, 0.9878, 1.2536) and (0.9836, 0.4102, 0.6325), respectively. For the T-shaped Groyne, the corresponding values were (0.9025, 1.2534, 1.8502) during training and (0.9873, 0.3337, 0.4972) during testing. In the FLOW-3D model, after calibration and validation, a Manning's roughness coefficient of 0.035 and the Prandtl's mixing length model were chosen for turbulence simulation. The performance indices during the testing phase for the L-shaped and T-shaped Groynes were (0.9607, 0.9363, 1.2070) and (0.9513, 1.1256, 1.3759), respectively. The GEP model showed a relative advantage over the FLOW-3D model.

Concutions:

 This study compares the performance of MLMs (SVM, GEP) with FLOW-3D in simulating diverted flow using T-shaped and L-shaped Groynes. Results from laboratory flume tests showed GEP outperformed SVM and FLOW-3D, particularly in simulating flow diversion, evaluated by RMSE, MAE, and R² performance indices.