1. Introduction
The development of high-speed trains necessitates a thorough investigation of their aerodynamic performance. In addition to drag optimizations to reduce energy requirements, the stability in crosswind conditions, the induced flow in the vicinity of the train (slipstream velocities) and other aerodynamic aspects have to be considered.
Three tools are commonly used to assess the aerodynamic performance of high-speed trains: full-scale tests, tests at model scale and numerical simulations. Especially early on in the development process, model scale testing and numerical simulations are invaluable as cost-effective development environments. For model scale experiments, there are two common approaches, namely moving model testing and wind tunnel testing. With moving model testing, longer train models can be investigated, and the relative motion between train and ground is described correctly. Drawbacks of moving model testing are the requirement for unique testing facilities, and equipping the model for accurate force measurements is difficult. They are best suited for trackside measurements including measuring slipstream velocities, nose pressures and trackside structure loading [
1]. Drag evaluations are usually restricted to cases where wind tunnel measurements are not possible, e.g., for long train models [
2].
On the contrary, wind tunnel tests allow fast, reliable and repeatable measurements of aerodynamic forces. Their application is mainly for drag and crosswind sensitivity studies and limited to short train models. During wind tunnel tests, the model is normally mounted above a static ground plane, and thus the relative motion between train and ground is not modeled. As a consequence, a boundary layer grows along the ground plane, resulting in unrealistic flow conditions underneath the train. Several studies using computational fluid dynamics (CFD) have analyzed the importance of a moving ground for drag and lift prediction.
In [
3], Xia et al. used CFD to investigate a three car train model with different ground simulations. While the drag of the first car decreased by only four drag counts (one drag count:
) when switching from a full moving ground to a static ground, the total drag of the three car train dropped by 6.4%. Additionally, the lift increased significantly, especially on the head car. They concluded that a moving ground is important for underbody investigations. Xia et al. also studied different moving ground conditions, where only part of the ground is moving akin to belts in automotive wind tunnels. They showed that using only a narrow belt between the wheels results in significant deviations from the full moving ground, whereas with a sufficiently wide belt these deviations are negligible.
Zhang et al. [
4] also used a three car train model. They simulated using a stationary ground and a moving ground and assessed how rotating wheels affect the moving ground results. The difference in drag between moving and static ground for the first car was only two drag counts, but the overall drag was 6% higher using the moving ground. The effect of rotating wheels was small (three drag counts total). Large parts of the drag increases using the moving ground were caused on the bogies, again highlighting the need for a moving ground in this region.
In contrast to the previous two studies, Niu et al. [
5] simulated the actual train movement instead of applying a moving ground boundary condition, thus replicating a moving model test. The moving head car’s drag was 10% smaller than the stationary head car’s drag. Significant pressure and velocity differences were observed along the underbody.
While multiple studies using CFD are available, experimental results using a moving ground are scarce. Baker and Brockie [
6] analyzed wind tunnel results using moving and stationary grounds with the aim of extrapolating the full scale drag for journey time calculations, where accurate absolute drag values are necessary. They concluded that the ground simulation produces drag differences of around 10%, while the uncertainty of extrapolating to full-scale led to differences of up to 30% of the measured full scale drag. Thus, the extrapolation error is larger than the error caused by using a static ground, and a moving ground may not be necessary for such applications.
Kwon et al. [
7] and Kwon et al. [
8] published results of wind tunnel tests using different ground simulations. In [
7], a high-speed train model was tested in two different wind tunnels, one using a moving belt and the other the usual static ground plane. They concluded that for evaluating drag reduction measures, the static ground plane is sufficient, while for absolute drag measurements, a moving ground is necessary. The drag of the three car train model was decreased by about 12% in the static ground tests. In [
8], Kwon et al. tested two train models with different lengths and nose shapes in the same two wind tunnels with additional tests using a tangential blowing system in a third wind tunnel. The conclusions remained the same, adding that a tangential blowing system can be an alternative to a moving belt system.
The aim of the following study is to provide additional experimental results, using the same wind tunnel with both a moving and stationary ground and a standardized train model. The focus is not only on absolute drag values, but also on the importance for geometry changes. With increasing train speeds and energy costs, even small drag optimizations of high-speed trains are important, and thus need to be predicted accurately during development. At a speed of 300 km/h the aerodynamic drag accounts for about 75% of the driving resistance of a modern high-speed train on open, flat ground (using the “High Speed 300” reference data from [
9]). This proportion is increasing rapidly with even higher train velocities. Bogie fairings as a drag reduction measure are slowly adopted in new high-speed trains across the world and the flow around them is affected by the underbody flow, which makes them ideal for this study. Bogie fairings offer a significant potential for drag reduction. For instance, Wang et al. [
10] simulated the drag reductions due to bogie fairings on a three car train. Using bogie fairings, the drag of the head car decreased by about 22% compared to the original model. The total drag of the three car model shrank by about 13%. The extensive experimental results are used to validate a CFD simulation setup. Lastly, the bogie fairings are simulated and compared to the wind tunnel tests.
2. Materials and Methods
Prior to this study, a 1:20 scale model of a high-speed train was built. It is based on the simplified ICE 3 reference geometry of the crosswind assessment standard EN 14067-6 [
11]. The model consists of the first car and a downstream body, separated by a 5 mm gap to ensure that no force is transferred between the bodies. The model features a modular design to allow for geometry changes and easy access to any measurement equipment located inside the model. The segments were 3D printed, sanded, filled and finally painted to ensure a smooth surface finish. Then, 125 pressure taps were drilled into the model, with most of them located on the train nose (
Figure 1). The model scale and length was chosen due to the size of the wind tunnel’s test section and length of the central belt of its five-belt system (called centerbelt). Train models with additional train cars would need to be at a significantly smaller scale, resulting in difficulties to reach sufficiently high Reynolds numbers.
It is expected that the moving ground has a more significant impact on geometry changes made on a train’s underbody than changes made higher up. As the underbody of the reference geometry is already smooth, bogie fairings were chosen as a design change to evaluate the influence of the ground simulation. The height of the bogie fairings is so that about two thirds of the front and rear bogie cutouts are covered (
Figure 2). For the wind tunnel tests, the bogie fairings were again 3D printed and taped to the model.
All wind tunnel measurements were carried out in the model scale wind tunnel of the Institute of Automotive Engineering (IFS) of the University of Stuttgart. It is a closed circuit wind tunnel with an open test section and its nozzle is 1.575 m wide and 1.05 m tall. The ground simulation system consists of a five-belt system, a boundary layer pre-suction and tangential blowing in front of the 1.70 m long and 0.25 m wide center belt. For comparison, the first car’s width is 0.148 m (59% of the centerbelt width) and its length is 1.301 m (77% of the centerbelt length). As the four wheel rotating units are connected to the wind tunnel balance, they could affect lift measurements and were therefore removed for this study. Wind speeds of up to 80 m/s are possible, however the wind speed in this study was 50 m/s. The resulting Reynolds number is about
(using a characteristic length L = 0.15 m) and thus in the range of constant pressure drag as recommended in [
12]. Prior measurements with wind speeds ranging between 80 km/h and 270 km/h supported this recommendation. Forces and moments were measured using the wind tunnel’s dynamic underfloor balance. Multiple runs were repeated once and on all occasions the drag coefficients of both runs were within one drag count. Further information about the flow quality, measurement equipment and control systems of the wind tunnel are published in [
13].
As the centerbelt is used as the moving ground, the model is positioned according to the true flat ground configuration, as defined in [
11]. Thus, the top of rail plane is 11.75 mm (235 mm full-scale) above the moving ground (
Figure 3).
During the moving ground (MG) measurements, the full ground simulation was used, i.e., the centerbelt, boundary layer pre-suction and tangential blowing were switched on. For the static ground (SG) measurements, all ground simulation systems were turned off. In this case, the boundary layer thickness at the turntable center in the empty test section is about 40 mm [
13]. This is still below the 30% of the model height threshold set out in EN 14067-6. Consequently, measurements using the static ground would be valid for crosswind assessments without any need for additional measures to reduce the boundary layer thickness of the incoming flow.
To measure the forces acting on the model, it is connected by two steel crossbeams to four vertical struts. The struts are connected to the wind tunnel’s underfloor balance (
Figure 4a). Similar struts are commonly used in the wind tunnel with 1:4 scale road vehicles. The aerodynamic forces on the struts and crossbeams were quantified by additional measurements where the model was held with two rods by the wind tunnel’s model manipulator and thus, only the crossbeams and struts were connected to the underfloor balance (
Figure 4b). Measurements with wind speed between 80 km/h and 270 km/h confirmed that the drag of the crossbeams and struts is Reynolds number independent in the considered range (
. Especially for measurements at higher yaw angles, the interference effects between model support and train model are not captured using this method, but they were assumed small compared to the forces acting on the train model.
Four Pressure Systems ESP-32HD pressure scanners were used for the surface pressure measurements. Their pressure range is about ±7000 Pa, and the static accuracy is within ±0.10% FS. The pressure scanners were positioned inside the first car and connected to an external Pressure Systems DTC Initium data acquisition system. Pressures were sampled at 10 Hz for 30 s and subsequently time-averaged. All 125 pressures were measured simultaneously. To enable comparisons between measurements using different ground configurations as well as between experiments and simulations, the time-averaged pressures are corrected using the corresponding static pressure gradient of the empty test section. The pressure gradients with all ground simulation systems enabled as well as with all systems disabled are published in [
13].
For the fluid simulations PowerFLOW 5.5b, a CFD package using the Lattice-Boltzmann-Method (LBM), was employed. The LBM is based on the mesoscopic kinetic theory. In contrast to the microscopic kinetic theory where particle motions are tracked, the mesoscopic kinetic theory involves tracking particle distributions [
14]. The Boltzmann equation (Equation (1)) describes the evolution of the particle distribution function
, with position
, particle velocity
, time
and the collision operator
. Macroscopic variables like density and velocity are moments of the particle distribution function. The left side of Equation (1) is the advection of the particle distribution function and the right side represents the redistribution of the particle distribution function due to collisions [
14]. The collision operator in its simplest form is the Bhatnagar–Gross–Krook (BGK) collision operator (Equation (2)) with the relaxation time constant
and the equilibrium distribution
.
PowerFLOW uses the D3Q19 model (three spatial dimensions, 19 discrete velocities) for discretizing the Boltzmann equation on a lattice. The discretized form using the BGK collision operator is given as Equation (3) with discrete velocities
and time step
. Interactions between the fluid flow and surfaces are calculated using a volumetric boundary scheme [
15]. At each facet particle bounce-back (no-slip), reflection (slip) or a combination of bounce-back and reflection (wall functions) processes occur [
15].
Simulations are inherently transient and turbulence is modeled with the very large eddy simulation (VLES) approach. This approach is similar to hybrid Reynolds Averaged Navier–Stokes (RANS)—large eddy simulation (LES) turbulence modeling of finite volume Navier–Stokes based solvers. In PowerFLOW, large scale turbulence is resolved by the voxel grid (cells of the lattice are called voxels), whereas small scale turbulence is modeled by a modified Renormalization Group (RNG)
k-ε turbulence model. The turbulence is represented in the BGK collision operator by adjusting the relaxation time scale
to
as given in Equation (4) [
15,
16,
17]. In Equation (4),
is the temperature,
the local strain parameter with the magnitude of the strain tensor
and
is a model coefficient (
). The turbulent kinetic energy
is given in Equation (5) and the turbulent dissipation
is calculated using Equation (6) [
15,
16,
17].
is the density,
is the molecular viscosity,
is the eddy viscosity and
is the stress tensor.
,
,
,
and
are model coefficients [
17]. For high Reynolds number flows, e.g., flows around vehicles, wall functions are used [
15].
The simulation was built using the same train geometry at 1:20 scale, also maintaining the gap between first car and downstream body. Instead of replicating the wind tunnel geometry, an idealized setup was employed: The model support was omitted and the simulation volume was significantly larger than the wind tunnel test section, resulting in a blockage of below 0.1% at 0° yaw. The idealized setup approach was chosen for its simplicity and to avoid introducing additional modeling errors that might be difficult to quantify (e.g., modeling of the boundary layer pre-suction). As the wind tunnel pressure measurements are corrected using the wind tunnel’s pressure gradient, the comparison of the pressure measurements and the simulations are valid.
The inlet velocity is set to 50 m/s (same as in the experiment) and the outlet is a zero pressure outlet. The walls of the simulation domain are frictionless walls, whereas two different approaches were used for the ground (
Figure 5): In the first case, the complete ground surface is modeled as a moving wall, with the velocity set to the wind speed (
Figure 5a). The second approach segments the ground in multiple section (
Figure 5b): Firstly, a moving wall section sized and positioned like the centerbelt in the wind tunnel. The surface roughness of this centerbelt section is matched to the roughness of the wind tunnel’s centerbelt. Secondly, segments either side and downstream of the centerbelt are sized according to the test section dimensions. These segments are treated as no-slip walls. The remaining ground surfaces are slip walls, thus avoiding a boundary layer upstream of the centerbelt. The simulations are run for 0.5 s of physical time and averaged over the last 0.3 s.
The simulation domain is refined using multiple refinement regions. A general refinement box (voxel size 8 mm) extends about 15 train heights upstream of the model and 25 train heights downstream. It is 768 mm (96 local voxels) wide and 512 mm (64 local voxels) tall. The term local voxels refers to the number of voxels with the voxel size of the respective refinement region, i.e., 96 voxels of voxel size 8 mm. The far field is then constructed using subsequently coarser boxes (
Figure 6a).
Near the surface of the train model additional refinement regions are employed, of which three are shown in
Figure 6b. A 15 mm offset (local voxel size: 1 mm) encloses all train surfaces. To resolve the flow around the train two additional volume refinements are used: Firstly, a 40 mm (20 local voxels) offset that is extended 200 mm in front of the train model and 250 mm behind the downstream body. Secondly, an 80 mm (20 local voxels) offset that extends 350 mm in front of the model and 600 mm behind the downstream body.
Select regions are further refined. The train nose up to the constant cross section of the car body, the bogies and the sharp edges of the bogie cutouts and inter-car gap are refined with a 7.5 mm offset of local voxel size 0.5 mm (
Figure 6b (1–3)). The front lip (cowcatcher) is refined with a 2.5 mm offset of local voxel size 0.25 mm (
Figure 6b (1)). These refinements result in a wall y
+ of around 30 on the train nose and around 60 along the remaining train surfaces for a wind speed of 50 m/s.
A grid sensitivity study was performed by scaling the voxel sizes. Three different grid sizes were investigated. The edge length of the smallest voxels varied between 0.25 mm (fine), 0.3125 mm (medium) and 0.375 mm (coarse). All other voxel sizes are scaled accordingly, but the simulation domain dimensions were unchanged. The surface pressure profiles for three sections are shown in
Figure 7. For the first section, differences in pressure between the three grids can be observed primarily on the underbody. The lowest pressures near the 0° angle indicator occur on the finest grid. Analysis of the force development profiles along the head car showed some differences near the front bogie, and thus in a region strongly influenced by the upstream underbody flow. For the sections downstream of the front bogie cutout and in front of the rear bogie cutout, deviations between the fine and medium grids are small. The underbody surface pressure (−30° to 30°) is higher on the coarse grid. All subsequent results were obtained using the fine grid.