Prediction of train aerodynamic coefficients under diverse shape parameters and yaw angles

Acquiring aerodynamic coefficients of a high-speed train considering its shape parameters and environmental yaw angles typically requires resource-intensive model tests or numerical simulations. To address this issue, this paper proposes an innovative surrogate model approach to cost-efficiently predict the aerodynamic coefficients. Six critical shape variables are chosen to construct a parametric train model, concurrently integrating the yaw angle (0-90 degrees) to generate a sample space using optimal Latin hypercube design. Then, four original regression algorithms [polynomial regression, support vector regression (SVR), least square support vector regression (LSSVR), and Kriging] and three improved regression algorithms (IPSO-SVR, IPSO-LSSVR, and IPSO-Kriging), incorporating an improved particle swarm optimization (IPSO) algorithm with SVR, LSSVR, and Kriging, are introduced to construct surrogate models. Finally, the prediction accuracy, prediction uncertainty and generalization potential of each surrogate model are compared in terms of the side force coefficient (Cs), lift force coefficient (Cl) and rolling moment coefficient (Cm). The results show that the IPSO-Kriging model outperforms the other surrogate models by exhibiting higher prediction accuracy and generalization performance, although the IPSO-LSSVR model provides a better assessment of the prediction uncertainty in the Cl. The absolute percentage error of IPSO-Kriging is within 5% for all test samples, which implies that this model can provide an effective and economical alternative for model tests or computational fluid dynamic simulations to acquire aerodynamic coefficients.