Hydraulic conveying characteristics of particles in bend based on numerical simulation and explainable stacking machine learning model

Abstract

As a hydraulic lifting pipeline structure widely used in deep-sea oil, gas transportation, and sediment dredging projects, the pipeline configuration is related to the improvement of transportation efficiency and pipeline safety. Particularly, the bending section consumes the most energy and withstands severe erosion. Understanding and predicting the conveying characteristics of two-phase flow in bends is therefore crucial. In this study, CFD-DEM (computational fluid dynamics-discrete element method) simulation method is employed to calculate various cases, considering five parameters: pipeline bending radius and angle, conveying velocity, particle diameter, and concentration, to explore the influence of these parameters on pressure drop and erosion rate of pipeline and result in a data set of hundreds of cases. Based on this data set, seven machine learning models are trained to predict pressure drop and erosion rate, respectively. To enhance model accuracy, the stacking method in ensemble learning is employed to combine multiple models with good performance. Additionally, the Optuna and SHAP (SHapley Additive exPlanation) methods are utilized to optimize hyperparameters and explain the degree to which parameters impact the predictions. The result demonstrates that pressure drop is almost unaffected by bending radius, while erosion rate initially decreases and then increases with bending angle, and both increase with other parameters. Among the evaluated models, artificial neural network, XGBoost, and random forest all demonstrate high prediction accuracy. The stacking model further improves the accuracy, with mean absolute error improving by 21.7% and 32.2%, and the SHAP method demonstrated good interpretability, which is basically consistent with CFD-DEM results.