1. Introduction
As wind farms become larger, the interactions between wind turbines increase. Wake effects can reduce turbine performance by up to
$40\,\%$
when the wind direction aligns with the wind farm layout (Stevens & Meneveau Reference Stevens and Meneveau2017; Porté-Agel et al. Reference Porté-Agel, Bastankhah and Shamsoddin2020). Traditional wind farm design strategies aim to control turbines individually and ignore their interactions (Howland et al. Reference Howland, Bossuyt, Martínez-Tossas, Meyers and Meneveau2016; Boersma et al. Reference Boersma, Doekemeijer, Gebraad, Fleming, Annoni, Scholbrock, Frederik and van Wingerden2017), but these strategies become less effective for larger wind farms. Active yaw control is a popular strategy to optimise wind farm power production (Meyers et al. Reference Meyers, Bottasso, Dykes, Fleming, Gebraad, Giebel, Göcmen and van Wingerden2022; Shapiro, Starke & Gayme Reference Shapiro, Starke and Gayme2022). Adjusting the turbine yaw angle can steer wakes away from downstream turbines, effectively reducing wake effects and improving overall wind farm performance. Simulations (Gebraad et al. Reference Gebraad, Teeuwisse, van Wingerden, Fleming, Ruben, Marden and Pao2016; Li, Liu & Lu Reference Li, Liu and Lu2024), wind tunnel experiments (Bastankhah & Porté-Agel Reference Bastankhah and Porté-Agel2019; Zong & Porté-Agel Reference Zong and Porté-Agel2021) and field measurements (Howland et al. Reference Howland, Lele and Dabiri2019, Reference Howland, Quesada, Martínez, Larrañaga, Yadav, Chawla, Sivaram and Dabiri2022) have demonstrated that yaw control can improve wind farm performance. For better evaluating and improving this control strategy, a deep understanding of the complex dynamics of yawed wakes is important and necessary (Wang et al. Reference Wang, Cai, Wang, Wang, Chen, Hou, Zhou, Xu, Zhang and Li2023). This highlights the need for more fundamental insights into tip vortex behaviour in yawed wakes.
The most prominent flow features in the near wake of a turbine are the helical vortex structures shed from the tips and roots of the rotor blades (Porté-Agel et al. Reference Porté-Agel, Bastankhah and Shamsoddin2020). Tip vortices obstruct momentum exchange between the free stream and the near-wake region by separating this region from the outer flow and thus significantly slow down wake recovery (Medici Reference Medici2005; Lignarolo et al. Reference Lignarolo, Ragni, Krishnaswami, Chen, Simão Ferreira and van Bussel2014, Reference Lignarolo, Ragni, Scarano, Ferreira and van Bussel2015; Porté-Agel et al. Reference Porté-Agel, Bastankhah and Shamsoddin2020). The evolution and stability of tip and root vortices behind a non-yawed turbine have been extensively studied in previous theoretical (Widnall Reference Widnall1972; Okulov Reference Okulov2004; Okulov & Sørensen Reference Okulov and Sørensen2007), numerical (Ivanell et al. Reference Ivanell, Mikkelsen, Sørensen and Henningson2010; Sarmast et al. Reference Sarmast, Dadfar, Mikkelsen, Schlatter, Ivanell, Sørensen and Henningson2014; Yang et al. Reference Yang, Hong, Barone and Sotiropoulos2016) and experimental studies (Hong et al. Reference Hong, Toloui, Chamorro, Guala, Howard, Riley, Tucker and Sotiropoulos2014; Biswas & Buxton Reference Biswas and Buxton2024a ). A single helical vortex exhibits three primary unstable modes: the short wave mode, long wave mode and mutual mode (Widnall Reference Widnall1972). The instability of the helical vortex depends on the wave number of the disturbance and its geometric parameters. Additionally, the distance between successive helical vortices is an important factor influencing their stability (Gupta & Loewy Reference Gupta and Loewy1974; Okulov Reference Okulov2004).
The vortex structures around the tip vortices play an important role in the instability of tip vortices (Sørensen Reference Sørensen2011). Therefore, models often simplify the rotor wake by
$N$
tip vortices with the circulation
$\Gamma$
and a root vortex with the circulation
$-N \Gamma$
. However, this vortex system is unconditionally unstable (Okulov & Sørensen Reference Okulov and Sørensen2007). In reality, the circulation on the blade changes in the radial direction, so a trailing vortex sheet is generated along the blade and advected downstream. Due to the mutual induction between tip vortices, hub vortices and the trailing vortex sheet, a complex roll-up process occurs. This process eventually leads to the formation of discrete and strong tip vortices surrounded by a compact vortex in the centre. These are stable and can be observed in experiments (Landgrebe Reference Landgrebe1972; Okulov & Sørensen Reference Okulov and Sørensen2007). Furthermore, in utility-scale field experiments, centrifugal instability was identified as an important mechanism for tip vortex dynamics (Yang et al. Reference Yang, Hong, Barone and Sotiropoulos2016). Additionally, blade design (Dong et al. Reference Dong, Qin, Li and Yang2023) and tip speed ratio (Biswas & Buxton Reference Biswas and Buxton2024a
) strongly affect tip vortex dynamics. Furthermore, the free stream turbulence enhances the entrainment rate into the wake, leading to a faster tip vortex breakdown (Gambuzza & Ganapathisubramani Reference Gambuzza and Ganapathisubramani2023).
For non-yawed turbines, the stability of tip vortices can be analysed using different methods, such as Fourier analysis (Ivanell et al. Reference Ivanell, Mikkelsen, Sørensen and Henningson2010), proper orthogonal decomposition (POD, Lumley Reference Lumley1967; Mula & Tinney Reference Mula and Tinney2015) and dynamic mode decomposition (DMD, Schmid Reference Schmid2010; Sun et al. Reference Sun, Tian, Zhu, Hua and Du2021). These mode decomposition methods can capture the primary modes induced by the tip vortices. In particular, Sarmast et al. (Reference Sarmast, Dadfar, Mikkelsen, Schlatter, Ivanell, Sørensen and Henningson2014) employed both POD and DMD to analyse the instability of wind turbine wakes. They discovered that the most unstable mode is characterised by an out-of-phase displacement of successive helices, resulting in local vortex pairing. This implies that the main cause of wake destabilisation is mutual induction instability. Sun et al. (Reference Sun, Tian, Zhu, Hua and Du2021) used DMD to show that the leading energetic modes represent the downstream, swinging and rotating motion of the tip vortices. A rotating reference frame reveals that disturbances originating from small-scale turbulence around the tip vortices lead to tip vortex instability.
Mode decomposition methods can also analyse the interaction of coherent structures across different scales, providing insights into the energy transfer process in the turbine wake. For instance, Lignarolo et al. (Reference Lignarolo, Ragni, Scarano, Ferreira and van Bussel2015) analysed the tip vortex interaction using triple decomposition. They emphasised that pairing instability is a precursor to the onset of three-dimensional vortex distortion and breakdown, leading to enhanced turbulent mixing and kinetic energy transport in the far-wake region. Kinjangi & Foti (Reference Kinjangi and Foti2023) used sparsity-promoting dynamic mode decomposition (spDMD, Jovanović et al. Reference Jovanović, Schmid and Nichols2014) to extract coherent structures. They analysed the scale-specific transfer of coherent kinetic energy among mean, coherent and random scales, identifying interactions between tip vortices and root vortices through this energy transfer. They found that the flow becomes highly turbulent after tip vortices break up, generating numerous coherent structures with different spatial–temporal scales. Biswas & Buxton (Reference Biswas and Buxton2024b ) studied the coherent energy transfer in the near wake of a turbine at various tip-speed ratios using a multi-scale triple decomposition. They analysed the intricate network of triadic energy transfer among modes with different frequencies, identifying a variety of energy sources such as energy production from the mean flow and nonlinear triadic energy production contributing to the stability of the tip vortices.
For yawed turbines, the axial symmetry of the vortex structure is disrupted, creating a curled wake profile. As a result, the evolution and instability of tip vortices behind a yawed rotor are significantly different from those behind a non-yawed rotor. For example, the evolution and instability characteristics of the tip vortices are asymmetric, with the vortices on the windward side being stronger and the vortex core size larger (Wang et al. Reference Wang, Cai, Wang, Wang, Chen, Hou, Zhou, Xu, Zhang and Li2023). Also the counter-rotating vortex pair (CVP) is a crucial feature of a yawed wind turbine wake. Note that CVP exists in any steady flow around a lifting body (Liu et al. Reference Liu, Wu, Su and Kang2017). Thus, it also exists in the far wake of both a yawed drag disk and a yawed wind turbine. In particular, the CVP is created by the accompanying trailing vortices from the top and bottom of the disk (Howland et al. Reference Howland, Bossuyt, Martínez-Tossas, Meyers and Meneveau2016; Shapiro, Gayme & Meneveau Reference Shapiro, Gayme and Meneveau2018), similar to a wind turbine blade which creates a lift force and has a trailing tip and root vortex on each end (Zong & Porté-Agel Reference Zong and Porté-Agel2020). The difference between a rotating turbine and a static disk under yawed conditions is actually related to the fact that the location and strength distribution of vortices are different, although both systems create an integral lift force deflecting the wake. In particular, for a rotating wind turbine the tip vortices are located around the outer edge of the wake and the root vortex is located in the center of the wake, while for a static disk the two counter rotating vortex cores are located at the top and bottom. Meanwhile, for a rotating wind turbine the tip and root vortex strength changes with angular position of the blade due to the change in relative velocity, while for a static disk the two counter rotating vortex strength is the same due to the conservation of total circulation.
Howland et al. (Reference Howland, Bossuyt, Martínez-Tossas, Meyers and Meneveau2016) experimentally found that a CVP forms behind a porous disk under yawed conditions, even when the disk is not rotating. Bastankhah & Porté-Agel (Reference Bastankhah and Porté-Agel2016) observed that a CVP forms behind a model wind turbine with high yaw angles, resulting in a kidney-shaped cross-section of the wake in the far-wake region. They hypothesised that the formation of the CVP can be explained by analysing the continuity equation. Shapiro et al. (Reference Shapiro, Gayme and Meneveau2018) viewed a yawed turbine as a lifting surface with an elliptical distribution of transverse lift and used Prandtl’s lifting line theory to predict the transverse velocity and the circulation of the shed CVP. Later, Shapiro, Gayme & Meneveau (Reference Shapiro, Gayme and Meneveau2020) demonstrated that the decay of the CVP in the wake of a yawed turbine operating within an atmospheric boundary layer is primarily driven by the gradual cancellation of vorticity along the wake’s line of symmetry due to cross-diffusion. Martínez-Tossas et al. (Reference Martínez-Tossas, Annoni, Fleming and Churchfield2019) and Martínez-Tossas & Branlard (Reference Martínez-Tossas and Branlard2020) developed curled wake models using a collection of vortices distributed along the straight line going from the bottom to the top of the rotor and along the circumference of the rotor, respectively. Both models predict an elliptic distribution of vortex strength, but neither can accurately predict wake displacement in the far-wake region. Finally, Bastankhah et al. (Reference Bastankhah, Shapiro, Shamsoddin, Gayme and Meneveau2022) developed an analytical model based on a vortex sheet to predict the shape of curled wakes. This model was subsequently integrated into a wake model to analytically describe the mean velocity deficit distribution in the wake of a yawed turbine.
Zong & Porté-Agel (Reference Zong and Porté-Agel2020) argued that the CVP formed behind a yawed wind turbine originates from the complex interactions between the hub vortex and the streamwise components of the blade tip vortices, which is fundamentally different from the case of a yawed drag disk where the hub vortex is absent. They proposed a point vortex model (PVM) in combination with the vortex cylinder theory to simulate the formation process of the CVP and predict the development of yawed wakes. As the vortex cylinder theory for a non-yawed turbine (Branlard & Gaunaa Reference Branlard and Gaunaa2015) was used to predict the total circulation, the model results must be multiplied by a factor of 0.45 to achieve good agreement with the data.
A good fundamental understanding of the evolution and instability of tip vortices behind a yawed wind turbine is essential to improve yaw control strategies in utility-scale wind farms (Meneveau Reference Meneveau2024). However, research on the evolution and instability of tip vortices in yawed wake conditions remains limited, emphasising the need to study their deformation and instability characteristics. This study consists of two main components: (i) we introduce a model to predict the evolution of tip vortices from their generation to destabilisation, which agrees well with simulation results; and (ii) we use spDMD to identify the critical mode types that influence the development and breakdown of tip vortices. Specifically, the first part of the study focuses on time-averaged wake structures, which are critical for understanding the long-term velocity deficits that directly influence the power output of downstream turbines. In contrast, the second part addresses the temporal dynamics of the wake, which are more closely associated with unsteady loading and fatigue on downstream turbine components. These two aspects are closely interconnected. As will be shown in § 5, the averaged mode identified in the spectral analysis corresponds to the time-averaged wake structure described in the first part, providing a dynamic validation of the steady-state observations. Moreover, the yaw-induced asymmetries observed in the mean vorticity field are visible in the modal decomposition as asymmetric interactions – particularly between tip and root vortex modes – demonstrating how time-averaged features are represented in the unsteady vortex dynamics. Additionally, the merging modes identified in the dynamic analysis help explain the vortex pairing behaviour hinted at in the time-averaged results, further reinforcing the link between both perspectives. By combining these two complementary approaches – time-averaged and spectral – we offer a more complete physical picture of yawed turbine wake behaviour, with relevance for both power optimisation and structural load assessment.
The organisation of the paper is as follows. In § 2, the simulation method and numerical set-up are explained. In § 3, the differences in the tip vortex evolution and instability characteristics between yawed and non-yawed wakes are analysed. In § 4, a physical model is proposed that is able to predict the evolution of tip vortices from their generation to destabilisation. The theoretical prediction is further compared with the simulation results for different yaw angles and tip speed ratios. In § 5, spDMD is employed to examine the spectral properties of the vortex structures in the far-wake region and to identify the critical mode types that are relevant to the development and breakdown of tip vortices. We conclude with a summary of the main findings in § 6. Some derivation details of the physical model are provided in Appendices A and B.
2. Numerical method
2.1. Large eddy simulation method
Large eddy simulation (LES) is used to integrate the spatially filtered Navier–Stokes equations (Albertson Reference Albertson1996; Liu et al. Reference Liu, Franceschini, Oliveira, Galeazzo, Carmo and Stevens2022):
where the tilde denotes spatial filtering,
$\widetilde {\boldsymbol{{u}}}$
is the velocity,
$\widetilde {\boldsymbol{\omega }} = \nabla \times \widetilde {\boldsymbol{{u}}}$
is the vorticity,
$\boldsymbol{{f}}_{{wt}}$
is the force due to the wind turbine, implemented using an actuator line model (ALM),
$\widetilde {p}$
is the modified pressure, and
$\boldsymbol{\tau }$
denotes the sum of the deviatoric part of the sub-grid scale shear stress and the viscous stress,
Here,
$\nu$
is the kinematic viscosity of air,
$C_s$
is the Smagorinsky coefficient,
$\unicode{x1D6E5} _{{grid}} =(\Delta x \Delta y \Delta z)^{1/3}$
is the grid scale with
$\Delta x, \Delta y, \Delta z$
the grid spacings in the streamwise (
$x$
), spanwise (
$y$
) and vertical (
$z$
) directions, and the superscript
$T$
denotes a matrix transpose. Following Sarmast et al. (Reference Sarmast, Dadfar, Mikkelsen, Schlatter, Ivanell, Sørensen and Henningson2014), the simulations are conducted at Reynolds number
$Re=U_0R/\nu =10^5$
, where
$U_0$
is the wind speed at infinity and
$R$
is the turbine radius. Note that the Reynolds number above
$10^5$
has only minor effects on the overall wake behaviour (Sørensen et al. Reference Sørensen, Shen and Munduate1998; Chamorro, Arndt & Sotiropoulos Reference Chamorro, Arndt and Sotiropoulos2012).
The ALM was initially developed for horizontal axis wind turbines by Sørensen & Shen (Reference Sørensen and Shen2002), where the blades were represented by a force distribution along a line extending from the hub to each blade tip. The advantage of ALM is that it allows detailed unsteady simulations to be carried out without resolving the entire flow field near the blades. As a result, ALM has become a widely used approach for studying the fundamental characteristics of turbine wakes (e.g. Gupta & Wan Reference Gupta and Wan2019; Abraham, Martínez-Tossas & Hong Reference Abraham, Martínez-Tossas and Hong2021; Korb, Asmuth & Ivanell Reference Korb, Asmuth and Ivanell2023). Importantly, the previous studies have shown that ALM can successfully capture the primary flow features in the near wake (see e.g. Sarmast et al. Reference Sarmast, Dadfar, Mikkelsen, Schlatter, Ivanell, Sørensen and Henningson2014; Nilsson et al. Reference Nilsson, Shen, Sørensen, Breton and Ivanell2015; Sørensen et al. Reference Sørensen, Mikkelsen, Henningson, Ivanell, Sarmast and Andersen2015) as observed in experiments. Therefore, this method has been employed to examine the stability of tip vortices of the non-yawed turbine in uniform inflow conditions (Ivanell et al. Reference Ivanell, Mikkelsen, Sørensen and Henningson2010; Sarmast et al. Reference Sarmast, Dadfar, Mikkelsen, Schlatter, Ivanell, Sørensen and Henningson2014). Note that there are also some limitations in ALM. For example, Melani et al. (Reference Melani, Mohamed, Cioni, Balduzzi and Bianchini2024) showed that ALM produces a more diffused vortex core than the blade-resolved simulation. However, they also showed that the circulation of the tip vortex can be well predicted by ALM. Thus, the trajectories of the vortex centres are largely unaffected. Building on this established capability, we employ ALM in the present study to analyse the near- and far-wake dynamics of yawed turbines.
Specifically, the turbine blades in ALM are represented by distributed body forces on the quarter-chord lines of the blades. The flow velocities
$(u_x, u_y, u_z)$
are interpolated to the actuator points in the global coordinate system
$(x, y, z)$
using trilinear interpolation. Subsequently, the local relative azimuthal velocity
$u_\theta$
is defined in the local coordinate system
$(r, \theta , x)$
of the blade as follows:
where
$\Omega$
is the rotor rotation speed,
$r$
is the radial location and
$\theta$
is the azimuthal angle. The angle of attack
$\alpha$
for each actuator line point is
where
$\psi$
is the twist angle of the blade and
$\phi$
is the relative angle deviation. The lift and drag forces per unit span are
where
$\rho$
is the air density,
$u_{{rel}} = (u_\theta ^2 + u_x^2)^{1/2}$
is the velocity relative to the rotating blade,
$c$
is the chord length, and
$C_l$
and
$C_d$
are the lift and drag coefficients that can be obtained from tabulated aerofoil data. These forces are subsequently transferred to the global coordinate system using the standard force projection method.
The pseudo-spectral LES code is the same as that used by Liu & Stevens (Reference Liu and Stevens2021), which originates from the work by Albertson (Reference Albertson1996), and later contributions by Bou-Zeid, Meneveau & Parlange (Reference Bou-Zeid, Meneveau and Parlange2005), Calaf, Meneveau & Meyers (Reference Calaf, Meneveau and Meyers2010), Gadde, Stieren & Stevens (Reference Gadde, Stieren and Stevens2021), Liu et al. (Reference Liu, Franceschini, Oliveira, Galeazzo, Carmo and Stevens2022) and many others. The computational grid is uniformly distributed in the horizontal directions and staggered in the vertical direction. The code is pseudo-spectral, using periodic boundary conditions in the streamwise and spanwise directions. The vertical direction uses the second-order centred finite difference method, with zero shear stress and no penetration at the top and bottom. Time integration is performed using a second-order Adams–Bashforth method and all simulations used the standard Smagorinsky model with
$C_{s} = 0.16$
. We note that the sub-grid model has limited influence on the near-wake flow structures (Martínez-Tossas et al. Reference Martínez-Tossas, Churchfield, Yilmaz, Sarlak, Johnson, Sørensen, Meyers and Meneveau2018).
2.2. Simulation set-up
We consider the National Renewable Energy Laboratory (NREL) 5 MW wind turbine (Jonkman et al. Reference Jonkman, Butterfield, Musial and Scott2009), which has a rotor diameter of
$D=126$
m. The optimal tip-speed ratio for maximum energy extraction is
$\lambda _{{opt}}=\Omega R/U_0=7.55$
, with
$\Omega$
the angular rotation speed of the blades. We simulated uniform inflow at
$U_0=8$
m s–1 without an applied pressure gradient. We consider the tip-speed ratio’s
$\lambda = 6$
and
$\lambda = 8$
and non-yawed (
$\beta =0^\circ$
) and yawed (
$\beta = 20^\circ , 30^\circ$
) conditions.
Grid convergence verification of the turbine with the yaw angle
$\beta =30^\circ$
and the tip speed ratio
$\lambda =8$
. Here,
$\Gamma _0$
is the total circulation of tip vortices,
$C_T=8T/(\rho U_0^2 \pi D^2)$
is the turbine thrust coefficient with
$\rho$
denoting the density of air and
$C_P=8P/(\rho U_0^3 \pi D^2)$
is the turbine power coefficient.

All simulations are performed in a domain of size
$L_x \times L_y \times L_z = 12D \times 4D \times 4D$
along the streamwise, spanwise and vertical directions, using a uniform grid with
$1680 \times 560 \times 560$
nodes. The turbine is located at
$3D$
downstream of the inlet in the middle of the
$y$
–
$z$
plane, and the turbine centre is set as the origin of the coordinate system. A fringe region of
$7.5\,\%$
of the streamwise domain length is used to force the inflow condition (Stevens, Graham & Meneveau Reference Stevens, Graham and Meneveau2014). The time step for the numerical simulation is
${\textrm{d}}t=0.001 T$
, where
$T=2\pi /\Omega$
is the rotation period of the turbine. The simulations are run for
$70T$
and statistics are collected after the simulations reach a stationary state (
$t\gt 35 T$
).
Table 1 compares the dimensionless total circulation
$\Gamma _0/(U_0 D)$
(to be discussed in detail in § 4), the turbine thrust coefficient
$C_T$
, and the turbine power coefficient
$C_P$
of the wind turbine with the yaw angle
$\beta =30^\circ$
and the tip speed ratio
$\lambda =8$
obtained from the grid resolution of
$840\times 280\times 280$
,
$1260\times 420\times 420$
and
$1680\times 560\times 560$
. The table shows clearly that the values of
$\Gamma _0, C_T, C_P$
obtained from different grid resolution agree very well with each other, where the maximum magnitude of the relative error is less than
$0.2\,\%$
. This good agreement confirms the grid convergence of the present study.
3. Wake structure and instability characteristics
Figure 1 shows the evolution and instability of tip vortices for
$\lambda =8$
under yawed and non-yawed conditions. For convenience, the turbine centre is set at the location
$(x,y,z)=(0,0,0)$
, where
$x,y,z$
are the streamwise, spanwise and vertical coordinates that form a right-handed coordinate system. The figure shows that under non-yawed conditions, the wake remains circular (figure 1
c), while the wake becomes kidney-shaped due to the formation of the CVP (figure 1
d) under yawed conditions. Following Sørensen et al. (Reference Sørensen, Mikkelsen, Henningson, Ivanell, Sarmast and Andersen2015), we divide the wake into three regions: the near wake, transition region and far wake. The quantitative criteria will be detailed later. It is worth noting that this division differs from Neunaber et al. (Reference Neunaber, Hölling, Stevens, Schepers and Peinke2020), where a decay region is also included.
Contour surfaces of the instantaneous vortices (
$Q=0.02$
) in a wind turbine wake for
$\lambda =8$
and (a)
$\beta =0^{\circ }$
and (b)
$\beta =30^{\circ }$
. Here,
$Q$
is the second invariant of the velocity gradient tensor. The iso-surfaces are coloured based on their streamwise position. Blue dashed line, the onset of the transition region; red dashed line, the onset of the far wake region; black dashed line, the streamwise location where the
$(y,z)$
-plane locates. The corresponding distributions of the instantaneous vortices at the plane
$x=3D$
for (c)
$\beta =0^{\circ }$
and (d)
$\beta =30^{\circ }$
. The dashed circle and ellipse denote the trajectory of the blade tip projected on the
$(y,z)$
-plane.

In the near-wake region, the tip vortices are smooth. Under non-yawed conditions, the trajectories of tip vortices projected on the cross-stream plane are circular, whereas they become elliptical under yawed conditions. In the transition region, adjacent tip vortices interact, pair and merge, triggering their instability. The destabilisation of the vortices is very different in yawed and non-yawed conditions. Under non-yawed conditions, the pitch of tip vortices is nearly uniform across azimuthal angles, leading to uniform vortex pairing and a narrow transition region (figure 1
a). Under yawed conditions, blades at
$z\lt 0$
(pointing into the plane) experience a velocity component in the negative streamwise direction, which reduces the advection velocity of the tip vortices. Blades at
$z\gt 0$
(pointing out of the plane) experience a velocity component in the positive streamwise direction, increasing the advection velocity of the vortices. This variation affects the pitch of the shed vortices at different locations, creating a relatively wide transition region (figure 1
b). In the far wake, the helical vortex structures break down into small-scale vortices, causing the wake to become turbulent.
When modelling wind turbine wakes, it is crucial to differentiate between the near- and far-wake regions because of their unique characteristics. In the near-wake region, the mean velocity deficit follows a super-Gaussian distribution (Blondel & Cathelain Reference Blondel and Cathelain2020; Blondel Reference Blondel2023), whereas in the far-wake region, it transitions to a Gaussian distribution (Vahidi & Porté-Agel Reference Vahidi and Porté-Agel2022). To accurately define the boundaries between the near- and far-wake regions, Biswas & Buxton (Reference Biswas and Buxton2024a ) suggested using the strength of the spanwise velocity spectra, which we now use for yawed turbines. The strength of the spanwise velocity spectra is defined as follows:
where
$v$
is the spanwise velocity. The relative strength and its frequency-averaged value are defined compared with the blade passing frequency
$3f_r$
(
$f_r= 1/T$
is the rotor frequency):
Variation of the relative strength of the spanwise velocity spectra
$\hat S_v(x,f)$
for
$\lambda =8$
and (a)
$\beta =0^\circ$
and (b)
$\beta =30^\circ$
. Black solid line, the frequency-averaged relative strength
$\langle \hat S_v(x,f) \rangle _f$
; blue dashed line, onset of the transition region where
$\langle \hat S_v(x,f) \rangle _f =0.1$
; red dashed line, onset of the far-wake region where
$\hat {S}_v=1$
.

Figure 2 shows the variation in the relative strength of the spanwise velocity spectra,
$\hat {S}_v(x,f)$
, for
$f\lt f_r$
for yawed and non-yawed conditions. In the near-wake region, the relative strength
$\hat S_v$
is very small. The reason is that the flow in this region is laminar with a dominant frequency of
$3f_r$
. As the tip vortices evolve downstream, they interact with each other, leading to the rapid growth of
$\hat S_v(x,f)$
. We define the onset of the transition region as the point where
$\langle \hat S_v(x,f) \rangle _f = 0.1$
. This occurs at
$x=1.14D$
for non-yawed conditions and
$x=1.01D$
for yawed conditions, indicating that yaw accelerates the instability onset of tip vortices.
In the far-wake region, the helical vortices break down into smaller-scale vortex structures. As a result, low frequencies replace the blade rotation frequency
$3f_r$
as the dominant frequencies. We define the onset of the far-wake region as the point where the strength of low frequencies (
$f \lt f_r$
) exceeds that of the blade passing frequency (
$3f_r$
), i.e.
$\langle \hat S_v(x,f) \rangle _f=1$
(Biswas & Buxton Reference Biswas and Buxton2024a
). We find the onset of the far-wake region at
$x=1.70D$
for the non-yawed conditions and at
$x=2.14D$
for the yawed conditions. This indicates that the tip vortices of the yawed turbine travel further downstream before they transition into small-scale turbulence. Note that results of
$\hat S_v(x,f)$
for
$f\gt f_r$
usually do not exhibit the behaviour as those for
$f\lt f_r$
that can be conveniently employed to identify the onset of the far-wake region, and are therefore not shown in the figure.
4. A physical model of vortex structure evolution of yawed turbine
4.1. Physical model
Active yaw control in wind farms critically depends on the wake velocity deficit of yawed turbines. Therefore, a good fundamental understanding of the CVP dynamics is essential as it controls the deflection and recovery of yawed turbine wakes. The formation mechanism of the CVP has been widely discussed in previous studies (Howland et al. Reference Howland, Bossuyt, Martínez-Tossas, Meyers and Meneveau2016; Bastankhah & Porté-Agel Reference Bastankhah and Porté-Agel2016; Shapiro et al. Reference Shapiro, Gayme and Meneveau2018, Reference Shapiro, Gayme and Meneveau2020; Martínez-Tossas et al. Reference Martínez-Tossas, Annoni, Fleming and Churchfield2019; Martínez-Tossas & Branlard Reference Martínez-Tossas and Branlard2020; Zong & Porté-Agel Reference Zong and Porté-Agel2020; Bossuyt et al. Reference Bossuyt, Scott, Ali and Cal2021; Bastankhah et al. Reference Bastankhah, Shapiro, Shamsoddin, Gayme and Meneveau2022). In particular, Zong & Porté-Agel (Reference Zong and Porté-Agel2020) proposed that the spanwise force generated by the yawed turbine induces the formation of the CVP, with both tip and root vortices playing essential roles. As the CVP evolves downstream, it generates a non-uniform spanwise velocity field, leading to the deflection of the wake and the development of a kidney-shaped velocity deficit.
Zong & Porté-Agel (Reference Zong and Porté-Agel2020) proposed a point vortex model (PVM) to simulate the CVP formation process and predict the time-averaged velocity deficit in yawed wakes. The PVM assumes the strength of tip vortices is independent of the azimuthal angle
$\theta$
. When the entire rotor disk is divided into
$N$
arc segments, the local circulation
$\Gamma _i$
within the
$i$
th segment depends only on the projection of the local velocity vector onto the streamwise direction, as follows:
where
$\boldsymbol{{e}}_x$
denotes the unit vector along the
$x$
-direction,
$\Gamma _0$
is the total circulation and
$\boldsymbol{{C}}_i$
is the local velocity vector of the tips,
\begin{equation} \boldsymbol{C}_i =\left [ \begin{matrix} U_0 (1-a_n \cos ^2 \beta )-\Omega R \sin \theta _i \sin \beta \\[3pt] -\Omega R \sin \theta _i \cos \beta \\[3pt] \Omega R \cos \theta _i \end{matrix} \right ], \end{equation}
where
$a_n$
is the rotor-averaged axial induction factor. Note that (4.2) is the same as that of Zong & Porté-Agel (Reference Zong and Porté-Agel2020) since
$a_n \cos ^2 \beta = a$
, where
$a$
is the induction factor based on the ratio of the averaged streamwise velocity at the rotor disk and far upstream.
To calculate the distribution of local circulation
$\Gamma _i$
, the induction factor
$a_n$
and the total circulation
$\Gamma _0$
must first be determined. Zong & Porté-Agel (Reference Zong and Porté-Agel2020) used a fitting formula to predict
$a_n$
and vortex cylinder theory for non-yawed turbines (Branlard & Gaunaa Reference Branlard and Gaunaa2015) to estimate
$\Gamma _0$
. However, they assumed that
$a_n$
only depends on the yaw angle
$\beta$
, while
$a_n$
can also depend on the tip-speed ratio
$\lambda$
. Moreover, applying the vortex cylinder theory for a non-yawed turbine to yawed conditions introduces inconsistencies, leading to an overestimation of
$\Gamma _0$
. Consequently, the model of Zong & Porté-Agel (Reference Zong and Porté-Agel2020) cannot reliably predict the circulation distribution
$\Gamma _i$
for yawed turbines. In fact, their model requires a correction factor of
$0.45$
to agree with observational data.
To overcome these limitations, we propose to use (i) an advanced version of blade element momentum (BEM) theory (Lu et al. Reference Lu, Li, Li, Liu, Zhang, Liu and Liu2023) with an updated expression of the spanwise velocity far downstream of the yawed turbine without any empirical constant to determine
$a_n$
and (ii) the vortex cylinder theory for a yawed turbine of Branlard & Gaunaa (Reference Branlard and Gaunaa2016) to determine
$\Gamma _0$
by the equivalency of the rotor-averaged axial induced velocity in the rotor plane calculated from the Biot–Savart law and the momentum theory. The detailed description of the determination of the total circulation
$\Gamma _0$
and the axial induction factor
$a_n$
are provided in Appendices A and B, respectively. Here we only list the main results.
The induction factor
$a_n$
can be expressed as (Heck, Johlas & Howland Reference Heck, Johlas and Howland2023)
\begin{equation} a_n =1-\frac {\sqrt {U_0^2 -U_4^2- V_4^2}}{\sqrt {C_T'} U_0 \cos \beta }, \end{equation}
where
$U_4, V_4$
are the streamwise and spanwise velocities far downstream of the turbine (Heck et al. Reference Heck, Johlas and Howland2023),
and (see Appendix B)
We remark that (4.3) and (4.4) are derived from the Bernoulli equations, and the conservation of mass and streamwise momentum balance, respectively, and (4.5) is derived from the definition of axial induction factor
$a_n$
. Thus, their validity is generally ensured and no empirical constant is needed, while Heck et al. (Reference Heck, Johlas and Howland2023) and Lu et al. (Reference Lu, Li, Li, Liu, Zhang, Liu and Liu2023) determined
$V_4$
using lifting line theory. In addition, Heck et al. (Reference Heck, Johlas and Howland2023) regarded the thrust coefficient
${C^{\prime}_T}$
in (4.3) and (4.4) as an input and thus
$a_n=a_n({C^{\prime}_T}, \beta )$
. In contrast, Lu et al. (Reference Lu, Li, Li, Liu, Zhang, Liu and Liu2023) determined theoretically
${C^{\prime}_T}$
by equating its value to the rotor-averaged thrust coefficient calculated from the BEM theory (see also Appendix B),
where
$N_b$
is the number of blades,
$r_0$
and
$R$
are the radii of the blade root and tip,
$\phi =\phi (r,\theta )$
is the inflow angle,
$\psi =\psi (r)$
is the blade twist angle,
$C_l$
and
$C_d$
are the corrected lift and drag coefficients, and
$c=c(r)$
is the blade chord length.
Thus, in this study, the thrust coefficient
${C^{\prime}_T}$
is treated as a function of multiple parameters, including the aerodynamic properties of the blades and the inflow conditions. By combining (4.3)–(4.6), the induction factor
$a_n$
can be solved analytically using an iterative method (for details, see Lu et al. Reference Lu, Li, Li, Liu, Zhang, Liu and Liu2023). Finally, by equating the rotor-averaged axial-induced velocity in the rotor plane calculated from the vortex system of a skewed vortex cylinder (Branlard & Gaunaa Reference Branlard and Gaunaa2016) and that calculated from the momentum theory, the total circulation
$\Gamma _0$
can be expressed as (for details see Appendix A)
By substituting (4.7) into (4.1), the local circulation
$\Gamma _i$
can be determined. The PVM proposed by Zong & Porté-Agel (Reference Zong and Porté-Agel2020) can then be used to simulate the evolution of the yawed turbine wake. The PVM consists of a root vortex
$\Gamma _{N+1}=-\Gamma _0$
located in the centre of the rotor disk and
$N$
point vortices distributed evenly along the edge of the rotor disk with strength
$\Gamma _i \, (i=1,2,\ldots , N)$
. In this vortex configuration, the zero-net total circulation criterion is automatically satisfied. For simplicity, these vortices are approximated as Lamb–Oseen vortices (Wu, Ma & Zhou Reference Wu, Ma and Zhou2006, p. 262). The velocity field
$(V_i,W_i)$
induced by the vortex at
$(y_i,z_i), \, i=1,2,\ldots , N+1$
can be determined as follows (Zong & Porté-Agel Reference Zong and Porté-Agel2020):
\begin{eqnarray} \def\lumina\hspace*{-1.8pc} V_i(y,z,t) \!\!\!\!&=&\!\!\!\! \frac {-\Gamma _i(z-z_i)}{2\pi \left [ (y-y_i)^2+(z-z_i)^2\right ]}\!\left \{ 1- \exp { \left [ -\frac {(y-y_i)^2+(z-z_i)^2}{4\nu _E t + R_i^2} \right ] } \right \}\!,\,\, \end{eqnarray}
\begin{eqnarray} \def\lumina\hspace*{-1.8pc} W_i(y,z,t) \!\!\!\!&=&\!\!\!\! \frac {-\Gamma _i(y-y_i)}{2\pi \left [ (y-y_i)^2 +(z-z_c^i)^2\right ]}\!\left \{ 1- \exp { \left [ -\frac {(y-y_i)^2+(z-z_i)^2}{4\nu _E t + R_i^2} \right ] } \right \}\!,\,\, \end{eqnarray}
where
$R_i$
is the initial core radius of the
$i$
th vortex,
$\nu _E$
is the eddy viscosity in the cross-stream plane and
$t$
is time. The behaviour of (4.8) and (4.9) for
$[(y-y_i)^2+(z-z_i)^2] \gg (4\nu _E t + R_i^2)$
approaches that of a line vortex, while for
$[(y-y_i)^2+(z-z_i)^2] \ll (4\nu _E t + R_i^2)$
, it approaches that of solid-body rotation. These two regions merge around
$[(y-y_i)^2+(z-z_i)^2] \sim (4\nu _E t + R_i^2)$
, which represents an expanding core radius.
The eddy viscosity
$\nu _E$
is determined by the Reynolds stress in the wake, which scales with
$\nu _E \propto a_n U_0 D$
. Although the prefactor
$\nu _E/(a_n U_0 D)$
shows negligible variance for different yaw angles, it does increase as the downstream location increases, with an average value of
$0.002$
in the region considered. Numerical tests indicate that the value of the prefactor has no significant effect on the predicted wake deflection. This is indeed the case since the trajectories of the Lamb–Oseen vortices are mainly governed by their circulations, which are fixed in our model and independent of the grid resolution (see table 1), and the eddy viscosity primarily influences the subsequent evolution of the vortex core sizes. Thus, in this study, we choose the eddy viscosity as
$\nu _E=0.002 a_n U_0 D$
. According to simulation results, we set the initial core radius as
$R_i=0.15D$
for the root vortex,
$R_i=0.02D$
for all tip vortices and the number of point vortices along the rotor disk edge as
$N=91$
. This setting ensures the initial vorticity contours closely resemble those observed in the simulation results. In the far wake, the convective velocity of the tip vortices should be very close to the free stream velocity, and for simplicity, we adopt the free stream velocity
$U_0$
as the convective velocity throughout the model such that the time
$t=x/U_0$
. Because the velocity induced by different vortices can be superimposed linearly, the time evolution of the location
$(y_j, z_j)$
of the
$j$
th vortex can be written as (Zong & Porté-Agel Reference Zong and Porté-Agel2020)
\begin{align} \frac {{\textrm{d}}y_j}{{\textrm{d}}t} = \sum _{i=1}^{N+1} V_i(y_j,z_j,t)(1-\delta _{ij}), \quad \frac {{\textrm{d}}z_j}{{\textrm{d}}t} = \sum _{i=1}^{N+1} W_i(y_j,z_j,t)(1-\delta _{ij}), \end{align}
where
$\delta _{ij}$
is the Kronecker delta. Equations (4.8)–(4.10) constitute a closed set of equations that can be used to simulate the evolution of the cross-stream velocity field under any operation condition, which is solved using an explicit time-marching scheme.
4.2. Validation against simulation data
Figure 3 shows the azimuthal distribution of the non-dimensional local circulation
$\Gamma _i/(U_0D)$
for the yawed wind turbine with the yaw angle
$\beta =20^\circ , 30^\circ$
and tip-speed ratio
$\lambda =6, 8$
. In LES, the local circulation
$\Gamma _i$
is calculated following the method of Zong & Porté-Agel (Reference Zong and Porté-Agel2020). Specifically, the annular area
$0.4D \leq \sqrt {y^2/\cos ^2{\beta } + z^2} \leq 0.6D$
of the cross-section at
$x/D=0.25$
is divided into 36 arc segments and the streamwise vorticity within each segment is integrated to determine
$\Gamma _i$
. The figure shows that the circulation
$\Gamma _i$
varies in a nearly sinusoidal way with the azimuthal angle, and its maximum and minimum values occur at the negative and positive axis of
$z$
(
$z=R\sin \theta$
), respectively.
Figure 3 shows that the Zong & Porté-Agel (Reference Zong and Porté-Agel2020) model deviates significantly from LES. In contrast, the prediction of the present model given by (4.1) with
$a_n$
determined from the updated BEM of Lu et al. (Reference Lu, Li, Li, Liu, Zhang, Liu and Liu2023) and
$\Gamma _0$
determined from (4.7) agrees excellently with LES. With the updated BEM accurately capturing the induction factor and the vortex cylinder model precisely predicting the total circulation, our model reliably predicts both the initial tip vortex distribution and the wake evolution.
Azimuthal distribution of the non-dimensional local circulation
$\Gamma _i/(U_0 D)$
for the yawed turbine with (a)
$\lambda =6, \beta =20^\circ$
, (b)
$\lambda =8, \beta =20^\circ$
, (c)
$\lambda =6, \beta =30^\circ$
and (d)
$\lambda =8, \beta =30^\circ$
at the streamwise location
$x/D=0.25$
. Circles, LES; dashed line, Zong & Porté-Agel (Reference Zong and Porté-Agel2020) model; solid line, present model (4.1).

Figure 4 compares the evolution of streamwise vorticity in the wake of the yawed turbine with
$\beta =30^\circ$
and
$\lambda =8$
calculated by LES and our model. Figure 4(a) illustrates the streamwise vorticity distribution in the near-wake region, revealing three distinct types of vortex structures. The negative root vortex is concentrated near the rotor centre. The negative and positive tip vortices are positioned along a circle with the rotor radius. Finally, the positive and negative trailing edge vortices are located between the root vortices and the tip vortices, and originate from the vortices shedding along the span of the blades (Bastankhah et al. Reference Bastankhah, Shapiro, Shamsoddin, Gayme and Meneveau2022; Biswas & Buxton Reference Biswas and Buxton2024a
).
As the trailing edge vortices are much weaker than the root vortex at the centre, their impact on tip vortex evolution is modelled in the PVM with a point vortex at the rotor centre. The strength of the point vortex is set to
$-\Gamma _0$
, consistent with Kelvin’s circulation theorem (Wu et al. Reference Wu, Ma and Zhou2006), which states that the total circulation of the rotor flow is zero under uniform inflow conditions. Figure 4(a,e) shows the PVM reliably predicts the initial distribution of streamwise vorticity directly behind the rotor.
Contour of the non-dimensional streamwise vorticity
$\omega _x D /U_0$
of the yawed turbine wake with
$\beta =30^\circ$
and
$\lambda =8$
. (a–d, i–l) LES and (e–h, m–p) model.

When the inflow is uniform and the wake’s velocity deficit is small, the Reynolds-averaged momentum equations in the spanwise and vertical directions can be approximated by two-dimensional momentum equations (Zong & Porté-Agel Reference Zong and Porté-Agel2020). This is indeed the case for the streamwise velocity deficits of tip vortices, which are usually less than 20 % of the incoming flow. Therefore, we assume that the initial vorticity distribution in the PVM evolves in the two-dimensional cross-stream plane with convection time
$t=x/U_0$
. Figure 4(b,c) shows the evolution of the vortex system in the transition region. As the flow remains laminar and the molecular viscosity is very small, the interfaces of the tip vortices are still very sharp. Vortex interactions cause the upper (
$z\gt 0$
) and lower (
$z\lt 0$
) halves of the tip vortices to bend clockwise and counterclockwise (that is, the direction pointing from the root vortex to the farthest tip vortices at
$z\gt 0$
and
$z\lt 0$
rotates clockwise and counterclockwise), respectively, gradually transforming the wake shape from elliptical to kidney-shaped. Clearly, the gradual formation of the kidney-shaped wake has been reproduced well by PVM (figure 4
f,g).
Figure 4(d, i–l) shows that the vortex structures in the far wake spread out more rapidly, primarily due to turbulent mixing. Due to vortex interactions, the large-scale vortex structures evolve and remain visible in the organisation of smaller scales. In particular, vortices originally in the upper half gradually merge with those in the central region, forming the upper leg of the CVP; while vortices in the lower half curl into an approximately circular shape, creating the lower leg of the CVP. The formation process of the CVP has also been captured well by our PVM (figure 4 h, m–p). Given that the PVM effectively simulates the evolution of the streamwise vorticity field, it should also be able to accurately capture the spanwise deflection of the turbine wake.
Figure 5 demonstrates that our PVM predicts the deflection of tip vortices at hub height well for the yawed wind turbine with the yaw angle
$\beta =20^\circ , 30^\circ$
and tip-speed ratio
$\lambda =6, 8$
. In LES, the tip vortex locations are identified as local extrema from a Gaussian fit to the spanwise distribution of time-averaged vorticity magnitude. The position of the strongest vortices creates the appearance of a narrower wake, corresponding to the relatively small core of the kidney-shaped wake. Figure 1, showing the three-dimensional evolution of tip vortices, highlights the overall wake expansion. Figure 5 shows that the model accurately predicts the spanwise wake deflection in the majority of the wake, and only overestimates very slightly the spanwise wake deflection in the far wake. The reason is mainly because the model does not fully capture the strong turbulent mixing in the far wake. Another potential reason is the fact that the identification of tip vortices in the far-wake region is very difficult as they break down therein. Clearly, the accuracy of the PVM predictions is more than sufficient to inform further far-wake modelling (Shapiro et al. Reference Shapiro, Gayme and Meneveau2018) and for predicting circulation in tip correction models (Kleine, Hanifi & Henningson Reference Kleine, Hanifi and Henningson2023).
Downstream evolution of the time-averaged positions of tip vortices at hub height for the yawed turbine with (a)
$\lambda =6, \beta =20^\circ$
, (b)
$\lambda =8, \beta =20^\circ$
, (c)
$\lambda =6, \beta =30^\circ$
and (d)
$\lambda =8, \beta =30^\circ$
. Circles, LES; solid lines, model.

5. Modal analysis of vorticity magnitude field of wind turbine
So far, we investigated the time-averaged wake structures. Here, we also study the temporal dynamics of the vortices, which impact the power output and fatigue loads of downstream turbines (Sarmast et al. Reference Sarmast, Dadfar, Mikkelsen, Schlatter, Ivanell, Sørensen and Henningson2014). We employed sparsity-promoting dynamic mode decomposition (spDMD; Jovanović et al. Reference Jovanović, Schmid and Nichols2014) to examine the spectral properties of the vortex structures in the far-wake region for yawed and non-yawed conditions. The goal is to decompose the complex flow into several simpler modes and understand the physical significance of dominant flow modes.
5.1. Sparsity-prompting dynamic mode decomposition
DMD is a powerful method for analysing flow instabilities, breaking down time-resolved flow data into modes. Each mode represents a flow structure with a specific frequency and an associated rate of temporal growth or decay (Schmid Reference Schmid2010; Tu et al. Reference Tu, Rowley, Luchtenburg, Brunton and Kutz2014; Jovanović et al. Reference Jovanović, Schmid and Nichols2014). The standard DMD approach faces some challenges when identifying the most relevant modes. For example, relying on a naive amplitude-based criterion can be inadequate, as it may prioritise modes with large initial amplitudes but significant decay rates. This issue often arises due to high-amplitude noise or statistical outliers, resulting in a suboptimal approximation of the original flow snapshots (Schmid Reference Schmid2022). To overcome these limitations, Jovanović et al. (Reference Jovanović, Schmid and Nichols2014) developed spDMD, which accurately reconstructs data using a weighted combination of mode-dynamics pairs while minimising the number of modes. This method is used here, and its key concepts are explained later. For more details, see Jovanović et al. (Reference Jovanović, Schmid and Nichols2014) and Kutz et al. (Reference Kutz, Brunton, Brunton and Proctor2016).
Consider the nonlinear dynamical system in the time-discrete formulation,
where
$\boldsymbol{{x}}_i$
are the state variables for the
$i$
th snapshot and
$\boldsymbol{{F}}$
represents the nonlinear governing equations. The dynamics of the nonlinear system of (5.1) can be projected onto the dynamics of the observables of
$\boldsymbol{{x}}$
, denoted as
$\phi (\boldsymbol{{x}})$
, via the linear Koopman operator
${\mathcal{K}}$
(Koopman Reference Koopman1931),
An approximate representation of the Koopman operator
${\mathcal{K}}$
can be obtained through a data-driven approach,
where
$\boldsymbol{{A}} \in {\mathbb{C}}^{m \times m}$
is the approximated Koopman operator, and
$\boldsymbol{\Psi }_2=[\boldsymbol{\psi }_1, \boldsymbol{\psi }_2, \ldots , \boldsymbol{\psi }_{n}] \in {\mathbb{R}}^{m \times n}$
and
$\boldsymbol{\Psi }_1=[\boldsymbol{\psi }_0, \boldsymbol{\psi }_1, \ldots , \boldsymbol{\psi }_{n-1}] \in {\mathbb{R}}^{m \times n}$
are the data matrices. Here,
$\boldsymbol{\psi }_i$
represents the
$i$
th snapshot of the flow field,
$m$
is the number of components in each snapshot and
$n$
is the number of snapshots in each data matrix. Each column in
$\boldsymbol{\Psi }_2$
corresponds to the corresponding column in
$\boldsymbol{\Psi }_1$
advanced by one sampling time step.
Due to the linear nature of
$\boldsymbol{{A}}$
, the observable space
$\phi (\boldsymbol{{x}})$
can be chosen as its invariant subspace. Thus, the temporal sequence of observables can be expressed as a superposition of the eigenvalues and eigenvectors of
$\boldsymbol{{A}}$
. However, the number of components in each snapshot usually far exceeds the number of snapshots
$(m \gg n)$
, so extracting eigenvalues and eigenvectors of
$\boldsymbol{{A}}$
directly is impractical. Therefore, the data matrix
$\boldsymbol{\Psi }_1$
is first preprocessed using the singular value decomposition (SVD, see e.g. Tu et al. Reference Tu, Rowley, Luchtenburg, Brunton and Kutz2014),
where
$\boldsymbol{\Sigma } \in {\mathbb{R}}^{n \times n}$
is a diagonal matrix,
$\boldsymbol{{U}} \in {\mathbb{C}}^{m \times n}$
is a semi-unitary matrix with
$\boldsymbol{{U}}^*\boldsymbol{{U}} = \boldsymbol{{I}}$
,
$\boldsymbol{{V}} \in {\mathbb{C}}^{n \times n}$
is a unitary matrix with
$\boldsymbol{{V}}^*\boldsymbol{{V}} = \boldsymbol{{V}} \boldsymbol{{V}}^*=\boldsymbol{{I}}$
,
$\boldsymbol{{I}} \in {\mathbb{R}}^{n \times n}$
is the identity matrix and the superscript
$^*$
denotes the conjugate transpose. Then, a low-dimensional matrix
$\widetilde {\boldsymbol{{S}}}$
can be constructed from a similarity transformation of
$\boldsymbol{{A}}$
,
where (5.3) and (5.4) have been used and
$\boldsymbol{\Sigma }^{+}$
is the Moore–Penrose pseudo-inverse of
$\boldsymbol{\Sigma }$
, which is formed by replacing every non-zero diagonal entry by its reciprocal. Denote by
$\mu _j \in {\mathbb{C}}$
and
$\widetilde {\boldsymbol{\phi }}_j \in {\mathbb{C}}^{n \times 1}$
the
$j$
th eigenvalue and eigenvector of
$\widetilde {\boldsymbol{{S}}}$
,
Then, from (5.5) and (5.6), one obtains
Here,
$\mu _j$
and
$\boldsymbol{\phi }_j$
are the
$j$
th eigenvalue and eigenvector of
$\boldsymbol{{P}} \boldsymbol{{A}}$
, and therefore they are called the DMD eigenvalue and the projected DMD mode (Tu et al. Reference Tu, Rowley, Luchtenburg, Brunton and Kutz2014). Note that
$\boldsymbol{{P}}$
represents the orthogonal projection onto the range of
$\boldsymbol{\Psi }_1$
,
Denote the spatial mode matrix as
$\boldsymbol{\Phi }$
and the initial amplitude matrix as
$\boldsymbol{\alpha }$
,
where
$\alpha _j (j=1,2,\ldots , n)$
is the initial complex amplitude of the projected DMD mode
$\boldsymbol{\phi }_j$
and the superscript
$T$
denotes the transpose. Assuming that the initial flow field can be expressed by the linear combination of the DMD modes in the least-squares sense (Kutz et al. Reference Kutz, Brunton, Brunton and Proctor2016),
Then, the temporal sequence of the original flow field can be reconstructed as follows:
where
$\boldsymbol{{D}}_\alpha = {\textrm{diag}} \, (\alpha _1, \alpha _2, \ldots , \alpha _n)$
is the initial complex amplitude matrix and
$\boldsymbol{{V}}_{{and}}=\boldsymbol{{V}}_{{and}}(\mu _1, \mu _2, \ldots , \mu _n)$
is the Vandermonde matrix.
In spDMD, the balance between accurate data reconstruction and sparsity in mode selection is achieved by regularising the least-squares deviation between the snapshot matrix and the linear combination of DMD modes, adding a term that penalises the
$L_1$
-norm of the DMD amplitude vector (Jovanović et al. Reference Jovanović, Schmid and Nichols2014; Schmid Reference Schmid2022),
where
$\Vert \cdot \Vert _F$
denotes the Frobenius norm,
$\Vert \cdot \Vert _1$
denotes the
$L_1$
-norm and
$\eta \geq 0$
is the regularisation parameter that controls the sparsity of the vector
$\boldsymbol{\alpha }$
. The regularisation parameter
$\eta$
should be selected to minimise the reconstruction error while maximising the elimination of unnecessary or unphysical components from
$\boldsymbol{\alpha }$
. Since there is no clear correlation between
$\eta$
and the number of retained components in
$\boldsymbol{\alpha }$
, a computational sweep of
$\eta$
is required to identify an appropriate value. The performance loss due to the sparsity-promoting method proposed by Jovanović et al. (Reference Jovanović, Schmid and Nichols2014),
\begin{equation} \Pi _{{loss}}=100 \sqrt {\frac { \Vert \boldsymbol{\Psi }_1 - \boldsymbol{\Phi } \boldsymbol{{D}}_\alpha \boldsymbol{{V}}_{{and}} \Vert _F }{ \Vert \boldsymbol{\Psi }_1 \Vert _F }}, \end{equation}
helps us monitor the reconstruction error of the spDMD when scanning
$\eta$
.
5.2. Eigenvalues and spectra of the spDMD modes
To investigate the intricate vortex structures that emerge after the destabilisation of large-scale helical vortices, we analyse the non-dimensional vorticity magnitude fields
$|\omega |D/U_{0}$
in the transition and far-wake region (see figure 6). Specifically, data are collected within the spatial domain:
$x/D \in [2, 4]$
,
$y/D \in [-1, 1]$
and
$z/D \in [-0.8, 0.8]$
. Note that the selected spatial range in the
$y$
-direction is slightly larger than that in
$z$
-direction since the yawed turbine wake deflects mainly in
$y$
-direction (see figures 4 and 5) and the magnitude of the vorticity is much larger than that of the streamwise vorticity (see figures 4 and 6). For the DMD analysis, 613 samples are extracted at a sampling frequency of
$61.2 f_r$
when the simulation is statistically stationary (
$t/T \in [60, 70]$
).
Iso-surface of the non-dimensional (a,b) instantaneous and (c,d) time-averaged vorticity magnitude
$|\boldsymbol{\omega }| D/U_0=20$
for
$\lambda =8$
with (a,c)
$\beta =0^\circ$
and (b,d)
$\beta =30^\circ$
. The iso-surfaces are coloured based on their non-dimensional (a,b) instantaneous and (c,d) time-averaged streamwise velocity.

Figure 7(a,b) shows the DMD eigenvalues on the complex plane for the wake under yawed and non-yawed conditions. All eigenvalues lie on or within the unit circle since the flow has reached the statistical stationary state and are symmetric around
${{Im}} \, (\mu _i)=0$
since they appear in conjugate pairs. This shows that most modes are neutrally stable, with a few exhibiting decay, indicating convective instability of the tip vortices (see figure 2). Figure 7(c,d) illustrates the initial amplitude distribution of the DMD modes as a function of the frequency
$f_i$
(Sarmast et al. Reference Sarmast, Dadfar, Mikkelsen, Schlatter, Ivanell, Sørensen and Henningson2014),
where
$\Delta t_s=1/(61.2 f_r)$
is the time interval between the snapshots in our data matrix. These results are obtained by solving the optimisation problem in (5.12) with the regularisation parameter
$\eta =0$
. Since oscillatory modes form conjugate pairs, the initial amplitude distribution is symmetric around
$f_i/f_r=0$
. Therefore, only half of the spectrum is displayed. The DMD analysis focuses on the beginning of the far-wake region where tip vortices begin to destabilise, capturing numerous complex, small-scale vortex structures with a wide range of characteristic frequencies. Some peaks in the DMD spectra do not correspond to the frequencies of
$3f_r$
or its harmonics and may result from the standard DMD algorithm’s sensitivity to numerical noise (Strom, Polagye & Brunton Reference Strom, Polagye and Brunton2022). This complicates data dimensionality reduction, necessitating further modal filtering using spDMD.
Eigenvalue distribution of the DMD modes for
$\lambda =8$
with (a)
$\beta =0^\circ$
and (b)
$\beta =30^\circ$
. The corresponding initial amplitude distribution of the DMD modes as a function of dimensionless frequency for the turbine at (c)
$\beta =0^\circ$
and (d)
$\beta =30^\circ$
. The mode colours correspond to those used in figure 8.

By solving the optimisation problem in (5.12), the number of retained modes,
$n_r$
, can be obtained as a function of the regularisation parameter
$\eta$
, and the results are shown in figure 8(a,b). Here, we only keep modes with an amplitude larger than
$10^{-6}$
. Figure 8(c,d) shows the optimal performance loss of spDMD as a function of the number of retained modes. For very large value of
$\eta \gtrsim 80\,000$
, only a single mode (i.e. the averaged mode to be discussed later) is extracted. As
$\eta$
decreases, the number of extracted modes gradually increases. For
$\eta \lesssim 600$
, all modes (i.e.
$n_r = 612$
) are extracted and thus the spDMD becomes equivalent to the standard DMD. Note that the energy distribution of the modes in the far-wake region is highly scattered; for example, the first 19 modes reconstruct only approximately
$20\,\%$
of the energy of the original flow field.
(a,b) Retained modes number
$n_r$
as a function of regularisation parameter
$\eta$
for
$\lambda =8$
with (a)
$\beta =0^\circ$
and (b)
$\beta =30^\circ$
. (c,d) Corresponding optimal performance loss of spDMD as a function of
$n_r$
for (c)
$\beta =0^\circ$
and (d)
$\beta =30^\circ$
. Modes extracted earlier when reducing the regularisation parameter
$\eta$
from a very large value are shown in dark colour, while those extracted later are shown in light colour.

Figure 9 presents the amplitude distribution for different numbers of retained modes in the spDMD analysis for non-yawed in panels (a,c,e) and yawed in panels (b,d,f) conditions. This is achieved by solving the optimisation problem in (5.12) with the regularisation parameter
$\eta$
determined from its relation with the number of retained modes
$n_r$
shown in figure 8(a,b). For
$n_r=1$
, the average mode with frequency
$f_i/f_r=0$
is selected for both cases (figure not shown). When
$n_r=7$
for the non-yawed case (figure 9
a) and
$n_r=5$
for the yawed case (figure 9
b), the spDMD identifies not only the averaged mode, but also
$(n_r-1)/2$
pairs of modes with frequencies in the range
$f_i/f_r \in (0,1)$
. As
$n_r$
increases, the number of modes in the frequency range
$f_i/f_r=(0,1)$
gradually increases. These modes are characterised with low frequencies
$f/f_r \in (0,1)$
. Note that these low frequencies are the dominant frequencies in the far wake (see figure 2) and are closely associated with wake meandering induced by the shear layer instability (Li, Dong & Yang Reference Li, Dong and Yang2022). Thus, we refer to these modes as shear modes for brevity. When
$7\lt n_r\leq 21$
for the non-yawed case (figure 9
c) and
$5\lt n_r\leq 15$
for the yawed case (figure 9
d), the spDMD begins to identify modes with frequencies of
$f_i/f_r=3$
and its harmonics or multiples, which we call the harmonic modes as their frequencies are related to the blade passing frequency of the tip/root vortices and its harmonics. Further increasing
$n_r$
increases the number of the identified shear and harmonic modes. When
$21\lt n_r\leq 41$
for the non-yawed case (figure 9
e) and
$15\lt n_r\leq 37$
for the yawed case (figure 9
f), the spDMD begins to identify modes with frequencies
$f_i/f_r \in [1,3)$
. These are the typical frequencies of merging (also known as pairing) dynamics between tip vortices (Felli, Camussi & Di Felice Reference Felli, Camussi and Di Felice2011; Biswas & Buxton Reference Biswas and Buxton2024a
). Note that the amplitude of the average mode is lower under yawed conditions due to reduced thrust force, which leads to a smaller vorticity magnitude than in non-yawed conditions. Furthermore, the shear, harmonic and merging modes are observed at lower mode numbers
$n_r$
when the turbine is yawed. The reason is that the tip vortices break down earlier in yawed conditions (figure 1).
Amplitude distribution of the retained
$n_r$
modes from the spDMD for
$\lambda =8$
with (a,c,e)
$\beta =0^\circ$
and (b,d,f)
$\beta =30^\circ$
. Due to symmetry, we only show the
$(n_r+1)/2$
modes with
$f_i/f_r\geq 0$
. The mode colours correspond to those used in figure 8.

Figure 9 shows that spDMD prioritises low-frequency modes with
$f_i/f_r \leq 3$
. These modes cannot be inferred from the initial amplitude distribution of the DMD modes in figure 7(c,d), as the selection does not only depend on the mode amplitude. These low-frequency dynamics are linked to the growth of low-frequency disturbances in the wake region, which are the most significant features in the far wake (Okulov et al. Reference Okulov, Naumov, Mikkelsen, Kabardin and Sørensen2014; Li et al. Reference Li, Liu and Lu2024). Therefore, in the following, we focus on the modes with
$n_r=41$
for non-yawed conditions (figure 9
e) and
$n_r=37$
for yawed conditions (figure 9
f).
5.2.1. Averaged mode with
$f_i/f_r=0$
The first spDMD mode (
$f_i/f_r=0$
) is shown in figure 10, where the isosurface is reconstructed from
${{Re}} (\boldsymbol{\phi }_i \alpha _i) =0.2$
(without summation). The figure distinctly illustrates the spatial characteristics of the tip vortices (outer edge) and root vortices (centre). Analysis shows that the eigenvalue
$\mu _i \approx 1$
, indicating that this mode does not exhibit temporal oscillations, and its amplitude remains constant over time. This suggests that the mode represents the time-averaged motion of the tip and root vortices. This is indeed the case since the flow structures displayed in figure 10 are very similar to those in figure 6(c,d). In terms of spatial features, the outer edge of the averaged mode for non-yawed condition retains a circular shape (figure 10
a), while that for the yawed condition evolves into a kidney-shaped structure (figure 10
b). Additionally, longitudinal streaks are observed along the tip vortex region for both cases, which may be related to streamwise coherent structures in the frequency range
$f_i/f_r \in (0,1)$
to be discussed later. As these coherent structures include very low frequencies, a dataset spanning only 10 rotations may be too limited to produce a clean, averaged wake shape. Nevertheless, the main findings obtained here should be unaffected.
Averaged mode with
$f_i/f_r=0$
for
$\lambda =8$
with (a)
$\beta =0^\circ$
and (b)
$\beta =30^\circ$
. The iso-surfaces
${{Re}} (\boldsymbol{\phi }_i \alpha _i) =0.2$
are coloured based on their non-dimensional time-averaged streamwise velocity
$U/U_0$
.

5.2.2. Shear modes with
$f_i/f_r \in (0,1)$
Figure 11 shows the shear modes with
$f_i/f_r = 0.32$
, representing the second type of modes identified by spDMD. These modes are predominantly located at the positions of the tip vortices and exhibit longitudinal streak-like structures, i.e. their length scales in the longitudinal direction are significantly larger than those in the transverse direction. This is a common feature of the modes influenced by shear effects (Di Mascio, Muscari & Dubbioso Reference Di Mascio, Muscari and Dubbioso2014; Magionesi et al. Reference Magionesi, Dubbioso, Muscari and Di Mascio2018). As the frequency increases, the longitudinal length scales of these streak-like structures decrease. For the non-yawed condition (figure 11
a), the shear modes are uniformly distributed along the outer edge of the wake. Under yawed conditions (figure 11
b), the shear modes are more concentrated below the hub height
$(z\lt 0)$
than above
$(z\gt 0)$
. This asymmetry arises from yaw misalignment, where the blades encounter a forward velocity component as they rotate into the lower region, thereby intensifying the shear in that area. Consequently, this leads to a top–down asymmetry in the modes. We note that the shear modes with
$f_i/f_r = 0.11$
(figure 9
e, f) correspond to a Strouhal number
$St = f_i D/U_0 = 0.28$
. This corresponds to the value found by Chamorro et al. (Reference Chamorro, Hill, Morton, Ellis, Arndt and Sotiropoulos2013) for far-wake meandering, and is very close to that observed for bluff bodies and circular cylinders where
$St \approx 0.2$
(Okulov et al. Reference Okulov, Naumov, Mikkelsen, Kabardin and Sørensen2014). This suggests that shear modes are associated with the early development of wake meandering.
Shear modes with
$f_i/f_r=0.32$
for
$\lambda =8$
with (a)
$\beta =0^\circ$
and (b)
$\beta =30^\circ$
. The iso-surfaces
${{Re}} (\boldsymbol{\phi }_i \alpha _i) =0.2$
are coloured based on their non-dimensional time-averaged streamwise velocity
$U/U_0$
.

5.2.3. Harmonic modes
Figure 12 displays the harmonic modes with
$f_i/f_r = 3$
identified by spDMD. These harmonic modes are linked to the periodic passage of tip and root vortices shed from the blades (Magionesi et al. Reference Magionesi, Dubbioso, Muscari and Di Mascio2018). In contrast to the shear modes, which primarily occur at the tip vortex positions (figure 11), the harmonic modes are present at both the tip and root vortex positions, suggesting the downstream propagation of tip and root vortices. The harmonic mode with the fundamental frequency
$3f_r$
is dominant in the near-wake region. In the far-wake region, vortex tearing, rolling up and pairing become evident. These nonlinear interactions between tip vortices are largely responsible for the excitation and amplification of higher harmonic modes. While similar to the averaged mode (figure 10), the harmonic modes do not directly interact with the mean flow under non-yawed conditions (figure 12
a). However, under yawed conditions, the harmonic modes directly interact with the mean flow due to the intrinsic flow asymmetry (figure 12
b). This interaction plays a critical role in the formation process of a CVP behind the yawed turbine (Zong & Porté-Agel Reference Zong and Porté-Agel2020). Note that CVP exists also in cases without root vortices, as demonstrated in the lifting line approach of Shapiro et al. (Reference Shapiro, Gayme and Meneveau2018) and the actuator disk wake study of Howland et al. (Reference Howland, Bossuyt, Martínez-Tossas, Meyers and Meneveau2016).
Harmonic modes with
$f_i/f_r=3$
for
$\lambda =8$
with (a)
$\beta =0^\circ$
and (b)
$\beta =30^\circ$
. The iso-surfaces
${{Re}} (\boldsymbol{\phi }_i \alpha _i) =0.06$
are coloured based on their non-dimensional time-averaged streamwise velocity
$U/U_0$
.

5.2.4. Merging modes with
$f_i/f_r \in [1,3)$
Figure 13 shows the merging modes with
$f_i/f_r = 2.1$
for yawed and non-yawed conditions. Similar to the shear modes (figure 11), the merging modes are primarily concentrated at the locations of the tip vortices and are nearly absent in the root vortex region. According to previous studies (Felli et al. Reference Felli, Camussi and Di Felice2011; Biswas & Buxton Reference Biswas and Buxton2024a
), the merging process is mainly associated with frequencies in the range
$f_i/f_r \in [1, 3)$
. The merging of tip vortices is mainly driven by mutual induction (Widnall Reference Widnall1972). This interaction between adjacent tip vortices amplifies small-scale disturbances in the wake over time. Yawing the turbine causes significant variations in tip vortex strength with azimuthal position (figure 3). These strength variations further accelerate the onset of mutual instability, which explains why the merging modes are identified earlier under yawed conditions (see figure 9
e, f).
Merging modes with
$f_i/f_r=2.1$
for
$\lambda =8$
with (a)
$\beta =0^\circ$
and (b)
$\beta =30^\circ$
. The iso-surfaces
${{Re}} (\boldsymbol{\phi }_i \alpha _i) =0.06$
are coloured based on their non-dimensional time-averaged streamwise velocity
$U/U_0$
.

6. Conclusion
We theoretically and numerically investigated the evolution and instability of tip vortices behind yawed and non-yawed wind turbines. First, we analyse the instability growth of tip vortices using high-fidelity large eddy simulations. Following Biswas & Buxton (Reference Biswas and Buxton2024a
), we employ the relative strength of the spanwise velocity spectra
$\hat {S}_v(x, f)$
to study the wake, applying this method to yawed wakes for the first time. In the near-wake region, the flow remains laminar, with smooth tip vortices and
$\hat {S}_v(x, f)$
values close to zero. In the transition region, vortex interactions intensify, leading to a sharp increase in
$\hat {S}_v(x, f)$
. Further downstream, in the far-wake region, large-scale helical vortices disintegrate into smaller turbulent structures and
$\hat {S}_v(x, f)$
stabilises at a high asymptotic value. Our results demonstrate that the transition region occurs earlier under yawed conditions than non-yawed conditions.
Second, we propose a self-consistent physical model that can accurately predict the time-averaged evolution of tip vortices from their generation to destabilisation. The physical model is inspired by the point vortex model proposed by Zong & Porté-Agel (Reference Zong and Porté-Agel2020), but with two essential improvements. In particular, we propose to use (i) an advanced version of blade element momentum theory (Lu et al. Reference Lu, Li, Li, Liu, Zhang, Liu and Liu2023) with an updated expression of the spanwise velocity far downstream of the yawed turbine without any empirical constant to determine
$a_n$
and (ii) the vortex cylinder theory for a yawed turbine of Branlard & Gaunaa (Reference Branlard and Gaunaa2016) to determine
$\Gamma _0$
by the equivalency of the rotor-averaged axial induced velocity in the rotor plane calculated from the Biot–Savart law and the momentum theory. This leads to a more accurate prediction of the azimuthal distribution of the local circulation
$\Gamma _i$
. Excellent agreement with simulations shows that the present model can accurately and efficiently predict the time-averaged evolution of yawed and non-yawed wind turbine wakes.
Finally, we apply sparsity-promoting dynamic mode decomposition (spDMD, Jovanović et al. Reference Jovanović, Schmid and Nichols2014) to analyse the spatial–temporal structures in the initial part of the far-wake region. In contrast to standard DMD, spDMD highlights the significance of low-frequency modes with
$f_i/f_r \leq 3$
(
$f_i$
is the mode frequency and
$f_r$
is the rotor frequency). We classify these modes into four categories based on their associated flow processes: (i) the averaged mode (
$f_i/f_r=0$
), representing the time-averaged flow; (ii) shear modes (
$f_i/f_r=(0,1)$
), associated with the early stage of wake meandering; (iii) harmonic modes (
$f_i/f_r=3n, n=1,2,\ldots$
), linked to blade passage; and (iv) merging modes (
$f_i/f_r \in [1,3)$
), corresponding to the pairing interactions between tip vortices. These modes are observed in both yawed and non-yawed cases, but they emerge earlier in the yawed configurations due to stronger interactions between tip and root vortices.
These low-frequency modes are linked to the growth of disturbances in the wake, significantly influencing wake recovery and the fatigue loads on downstream turbines. The key observation is that the modes are symmetric under non-yawed conditions, but asymmetric under yawed conditions. As a result, in yawed conditions, the modes associated with the tip vortices can interact directly with those of the root vortices. This direct interaction plays a crucial role during the formation process of the counter-rotating vortex pair observed in yawed wakes (Zong & Porté-Agel Reference Zong and Porté-Agel2020).
Funding
This work was supported by the National Natural Science Foundation of China (Grant Nos. 12272371, 12388101), the National Natural Science Fund for Excellent Young Scientists Fund Program (Overseas), and the Supercomputing Center of the University of Science and Technology of China. This project has received funding from the European Research Council Horizon Europe program (Grant No. 101124815). Numerical computations were performed at Hefei advanced computing centre.
Declaration of interests
The authors report no conflict of interest.
Data availability statement
All study data are included in this article.
Appendix A. Derivation of the total circulation
$\Gamma _0$
According to Branlard & Gaunaa (Reference Branlard and Gaunaa2016), the vortex system behind a yawed turbine can be modelled with a vortex cylinder with the circulation
$\Gamma _0$
, a bound vortex disk with the strength
$\gamma _b$
and a root vortex with the circulation
$\Gamma _r$
. The vortex cylinder is formed by the helical tip vorticity, which can be further decomposed into the longitudinal
$(\gamma _l)$
and tangential
$(\gamma _t)$
components. The bound vortex disk
$\gamma _b$
locates at the rotor plane and the root vortex
$\Gamma _r$
is parallel to the longitudinal tip vorticity
$\gamma _l$
. These vorticity components, as well as the pitch of tip vortex
$h$
, the yaw angle
$\beta$
and the wake skew angle
$\chi$
, are shown in figure 14. Since no expansion is assumed, the wake is contained in a cylinder of elliptical cross-section.
Sketch of the vortex system behind a yawed turbine.

The root vortex has the same magnitude as the total rotor circulation,
The bound vortex disk is assumed to have a constant strength for simplicity,
where
$r$
is the radial distance from the rotor centre. For each component of the vortex cylinder, the strength is equal to the total rotor circulation divided by the distance run by the helix in the orthogonal direction of this component,
where
$R$
is the rotor radius,
$h$
is the pitch of the tip vortex and
$\chi$
is the wake skew angle. The wake skew angle
$\chi$
is usually higher than the yaw angle
$\beta$
and their relations can be found for instance from Burton et al. (Reference Burton, Sharpe, Jenkins and Bossanyi2001, p. 105) and Bastankhah & Porté-Agel (Reference Bastankhah and Porté-Agel2016),
where
$a_n$
is the axial induction factor and
$\beta$
is the yaw angle. Note that (A4) can be derived using the Biot–Savart law with the vortex system approximated as that sketched in figure 14 (Burton et al. Reference Burton, Sharpe, Jenkins and Bossanyi2001). Meanwhile, the pitch
$h$
can be approximated by the streamwise velocity in the rotor area and the blade rotation period
$2\pi /\Omega$
,
Using the Biot–Savart law (e.g. Wu et al. Reference Wu, Ma and Zhou2006, § 3.2.2), Branlard & Gaunaa (Reference Branlard and Gaunaa2016) determined the velocities induced by the different vortex components. For our purpose, we only focus on the axial induced velocity in the rotor plane. In particular, the axial induced velocity in the rotor plane from the tangential vorticity is
where
$\theta \in [0, 2\pi ]$
is the azimuthal angle. The axial induced velocity in the rotor plane from the longitudinal vorticity is
The axial induced velocity in the rotor plane from the root vortex is
In addition, the cross-product in the Biot–Savart law (Wu et al. Reference Wu, Ma and Zhou2006) implies that the axial induced velocity in the rotor plane from the bound vortex disk is zero (Branlard & Gaunaa Reference Branlard and Gaunaa2015). Thus, the axial induced velocity in the rotor plane from the whole vortex system is
From (A6)–(A9), the rotor-averaged axial-induced velocity in the rotor plane can be determined as follows:
The rotor-averaged axial induced velocity (
$\langle \Delta U \rangle _A$
) should be the same as that determined by the momentum theory (
$-a_n U_0 \cos \beta$
), and thus,
Finally, substituting (A3)–(A5) into (A11), one obtains
This expression relates the total circulation
$\Gamma _0$
to the rotor-averaged axial induction factor
$a_n$
, which can be modelled by the BEM theory introduced later.
We note that the total circulation
$\Gamma _0$
used by Zong & Porté-Agel (Reference Zong and Porté-Agel2020) is (neglecting the tangential induction factor, see also Bossuyt et al. Reference Bossuyt, Scott, Ali and Cal2021)
where
$k=0.45$
is a factor added artificially to account for the overprediction of
$\Gamma _0$
and
$C_T$
is the thrust coefficient determined by empirical fitting,
Clearly, the total circulation
$\Gamma _0$
determined by (A13) requires empirical fitting of both the factor
$k$
and the thrust coefficient
$C_T$
. In contrast, the total circulation
$\Gamma _0$
determined by (A12) requires no empirical fitting at all.
Appendix B. Advanced blade element momentum theory
This section introduces the advanced version of the BEM theory proposed by Lu et al. (Reference Lu, Li, Li, Liu, Zhang, Liu and Liu2023) with an updated expression of the spanwise velocity far downstream of the yawed turbine. In particular, Appendix B.1 determines the induction factor
$a_n$
from the momentum theory and Appendix B.2 determines the thrust coefficient
${C^{\prime}_T}$
from the blade element theory.
B.1. Momentum theory: determination of
$a_n$
Denote
$U_4$
and
$V_4$
as the streamwise and spanwise velocities far downstream of the turbine and assume that the pressure recovers to the free stream far downstream and the velocity across the turbine disk is continuous. From the Bernoulli equations upstream and downstream of the rotor disk, the axial induction factor
$a_n$
can be determined as (Heck et al. Reference Heck, Johlas and Howland2023)
\begin{equation} a_n = 1-\frac { \sqrt {U^2_0-U_4^2-V_4^2} }{ \sqrt {{C^{\prime}_T}} U_0 \cos \beta }, \end{equation}
where
${C^{\prime}_T}$
is the turbine thrust coefficient to be determined later. Using the conservation of mass and the streamwise momentum balance, the streamwise velocity
$U_4$
can be determined as (Heck et al. Reference Heck, Johlas and Howland2023)
The spanwise velocity
$V_4$
needs special attention. Shapiro et al. (Reference Shapiro, Gayme and Meneveau2018) viewed a yawed turbine as a lifting surface with an elliptic distribution of spanwise lift and thus determined the spanwise velocity on the rotor disk, which is simply taken as
$V_4$
, from the Prandtl lifting line theory (e.g. Anderson Reference Anderson2001, p. 368),
Equation (B3) usually underpredicts the spanwise velocity disturbance at high yaw angles (Shapiro et al. Reference Shapiro, Gayme and Meneveau2018; Heck et al. Reference Heck, Johlas and Howland2023) and, therefore, Lu et al. (Reference Lu, Li, Li, Liu, Zhang, Liu and Liu2023) proposed an empirical constant
$3/2$
as a multiplier of (B3) to improve its performance. Although this empirical correction works well, the determination of
$V_4$
in the momentum theory using the lifting line theory is not consistent with our viewpoint on the vortex system of the yawed turbine (see figure 14). According to Branlard & Gaunaa (Reference Branlard and Gaunaa2016), the spanwise velocity induced by the vortex system can be determined in principle. However, this could be too complicated and goes beyond the momentum theory framework. Actually, the spanwise velocity
$V_4$
can simply be determined within the momentum theory framework when the tangential induction factor is neglected,
Thus, (B4) is adopted in this study.
Now, consider the limiting case
$|V_4| \ll |U_4|$
, the rotor-normal induction can be simplified as (Shapiro et al. Reference Shapiro, Gayme and Meneveau2018; Heck et al. Reference Heck, Johlas and Howland2023)
Thus, the ratio of (B4) to (B3) is
$1/[(1-a_n) \cos \beta ]$
, which approaches to
$3/2$
when
$a_n \rightarrow 1/3$
and
$\beta \rightarrow 0$
. This explains the rationality of the empirical constant
$3/2$
proposed by Lu et al. (Reference Lu, Li, Li, Liu, Zhang, Liu and Liu2023). However, if the updated expression of (B4) is employed, the empirical constant is unnecessary and the theory is more self-consistent.
B.2. Blade element theory: determination of
${C^{\prime}_T}$
The momentum theory alone is not sufficient to determine the value of
$a_n$
theoretically as the thrust coefficient
${C^{\prime}_T}$
is unknown yet. This can be derived from the blade element theory in yawed conditions. Consequently, the thrust coefficient
${C^{\prime}_T}$
is not a constant, but a function of many parameters (e.g. aerodynamic properties of the blades and the inflow conditions).
Denote the normal and tangential components of the wind velocity far upstream relative to the rotor disk as
Then, the normal component of the wind velocity at the turbine location is
and the tangential component of the wind relative to the blade is
where
$\Omega$
is the angular velocity of the blade,
$r$
is the radial coordinate and
$\theta$
is the azimuthal angle. Note that we have neglected the tangential induction factor and taken the axial induction factor as its rotor-averaged value. The squared relative wind speed to the blade is
and the inflow angle
$\phi =\phi (r,\theta )$
is
The average thrust for the rotor area
${\textrm{d}} S=r \,{\textrm{d}} \theta \,{\textrm{d}} r$
is
where
$\rho$
is the air density,
$N_b$
is the number of blades,
$c=c(r)$
is the blade chord length,
$\psi =\psi (r)$
is the blade twist angle, and
$C_l$
and
$C_d$
are the corrected lift and drag coefficients that depend on the local aerofoil type and angle of attack. Thus, the rotor-averaged thrust coefficient
${C^{\prime}_T}$
can be calculated as
\begin{eqnarray} \def\lumina\hspace*{-2.8pc} {C^{\prime}_T} \!\!\!\!&=&\!\!\!\!\frac {1}{\pi R^2} \int _{r_0}^R \!\! \int _0^{2\pi } \frac { 2 }{\rho W_n^2} \frac {\textrm{d}T}{\textrm{d}S} r \,{\textrm{d}} \theta \,{\textrm{d}}r \nonumber\\[10pt] \def\lumina\hspace*{-2.8pc} \!\!\!\!&=&\!\!\!\! \frac {N_b}{2 \pi ^2 R^2} \int _{r_0}^R \!\! \int _0^{2\pi } \frac {1}{\sin \phi } [ C_l (\phi -\psi ) \cot \phi + C_d(\phi -\psi ) ] c(r) \,{\textrm{d}} \theta\, {\textrm{d}}r, \end{eqnarray}
where
$r_0$
and
$R$
are the radii of the blade root and tip, respectively.























































































