This article was downloaded by: [Selcuk Universitesi]

On: 16 January 2015, At: 04:51

Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,
37-41 Mortimer Street, London W1T 3JH, UK

International Journal of Control

Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/tcon20

Optimal control of Formula One car energy recovery

systems
a a b
D.J.N. Limebeer , G. Perantoni & A.V. Rao
a
Department of Engineering Science, University of Oxford, Parks Road, Oxford, OX1 3PJ, UK
b
Department of Mechanical and Aerospace Engineering, University of Florida, Gainesville,
FL 32611–6250, USA
Accepted author version posted online: 14 Apr 2014.Published online: 23 Apr 2014.

Click for updates

To cite this article: D.J.N. Limebeer, G. Perantoni & A.V. Rao (2014) Optimal control of Formula One car energy recovery
systems, International Journal of Control, 87:10, 2065-2080, DOI: 10.1080/00207179.2014.900705

To link to this article: http://dx.doi.org/10.1080/00207179.2014.900705

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) contained
in the publications on our platform. However, Taylor & Francis, our agents, and our licensors make no
representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of the
Content. Any opinions and views expressed in this publication are the opinions and views of the authors, and
are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied upon and
should be independently verified with primary sources of information. Taylor and Francis shall not be liable for
any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoever
or howsoever caused arising directly or indirectly in connection with, in relation to or arising out of the use of
the Content.

This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://
www.tandfonline.com/page/terms-and-conditions
 International Journal of Control, 2014
Vol. 87, No. 10, 2065–2080, http://dx.doi.org/10.1080/00207179.2014.900705

Optimal control of Formula One car energy recovery systems

D.J.N. Limebeera,∗ , G. Perantonia and A.V. Raob
a
Department of Engineering Science, University of Oxford, Parks Road, Oxford, OX1 3PJ, UK; b Department of Mechanical and
Aerospace Engineering, University of Florida, Gainesville, FL 32611–6250, USA
(Received 8 November 2013; accepted 1 March 2014)

The utility of orthogonal collocation methods in the solution of optimal control problems relating to Formula One racing
is demonstrated. These methods can be used to optimise driver controls such as the steering, braking and throttle usage,
and to optimise vehicle parameters such as the aerodynamic down force and mass distributions. Of particular interest is the
optimal usage of energy recovery systems (ERSs). Contemporary kinetic energy recovery systems are studied and compared
with future hybrid kinetic and thermal/heat ERSs known as ERS-K and ERS-H, respectively. It is demonstrated that these
systems, when properly controlled, can produce contemporary lap time using approximately two-thirds of the fuel required
Downloaded by [Selcuk Universitesi] at 04:51 16 January 2015

by earlier generation (2013 and prior) vehicles.

Keywords: numerical optimal control; orthogonal collocation; pseudo-spectral methods; lap-time simulation; Formula One
car modelling; energy recovery systems

1. Introduction tion environment (Åkesson, Årzén, Gäfvert, Bergdahl, &

Minimum-lap-time optimal control problems for Formula Tummescheit, 2010) has also been used to solve minimum-
One cars are not easy to solve. A discussion of some of the lap-time problems (Gustafsson, 2008).
technical difficulties related to the solution of these prob- Much of the more recent work in this area is directed
lems can be found in the existing literature, for example, to improving the solubility of these problems and focuses
Casanova (2000) and Kelly (2008). A historical overview of on simplifying the vehicle and tyre models, and widen-
some of the optimal control literature relating to vehicular ing the use of linear optimal control techniques. These in-
applications can be found in Sharp and Peng (2011). clude the use of quasi-steady-state models (Brayshaw &
A reader familiar with the classical theory of optimal Harrison, 2005) and the use of linear quadratic (LQ) pre-
control and the calculus of variations might think that the view and model predictive (MP) control. One approach
most natural approach to the solution of these problems is to to minimise the burden associated with minimum-lap-time
use first-order necessary conditions to formulate and solve studies is the use of LQ preview to follow a prescribed
a two-point boundary-value problem. This is the approach driving line at a fixed speed (Thommyppillai, Evangelou,
taken in Cossalter, Lio, Lot, and Fabbri (1999), where two & Sharp, 2009). The method proposed makes use of mul-
motorcycle manoeuvring problems are solved. In later de- tiple linear models and a gain scheduling scheme, with its
velopments of this line of research, an indirect method is operation demonstrated on three-track segments. The ideas
used to examine minimum-time manoeuvres for a single- based on linearised models have been developed in Tim-
track car model through a U-turn (Tavernini, Massaro, ings and Cole (2012, 2013), where (linear time-varying)
Velenis, Katzourakis, & Lot, 2013). MP control rather than linear preview control is exploited.
In some recent work, a direct transcription method This approximation allows one to define the problem of
is used to solve various minimum-lap-time optimal con- finding a suboptimal racing line and a speed profile as a
trol problems for a complete lap (Perantoni & Limebeer, convex optimisation problem.
2014). This work employed a car model that reproduces In this paper, we study the use of a direct orthogo-
the longitudinal, lateral and yaw dynamics of the vehicle, a nal collocation method for solving racing-related optimal
realistic tyre model and a curvilinear (differential geomet- control problems. Apart from the use of a different opti-
ric) description of the track. This approach was success- mal control methodology (as compared with Perantoni &
ful in solving several minimum-lap-time optimal control Limebeer, 2014), the car model has been upgraded to use a
problems with a number of control and parameter optimi- more realistic aerodynamic model. A key focus of this paper
sation variations. The Modelica modelling and optimisa- is the optimal control of the ERS. Under the 2013 Formula

∗
Corresponding author. Email: [email protected]


C 2014 Taylor & Francis
 2066 D.J.N. Limebeer et al.

One technical regulations, a kinetic energy recovery system

(KERS) is allowed to recover kinetic energy from the car
during braking, store that energy (usually in a battery that
we will refer to as the energy store (ES)), and then make it
available to propel the car at some later point in the race.
The energy drawn from the ES to propel the car is limited
to 400 KJ on a per-lap basis (Fédération Internationale de
l’Automobile, 2012).
In the lead up to the 2008–2017 Formula One engine
development freeze, it was estimated that an average of four
milliseconds per lap were gained for every million dollars
Figure 1. Plan view of a Formula One car with its basic geomet-
spent on engine refinements (Trabesinger, 2007). This in- ric parameters. The car mass centre is at Mc , while the centre of
dicates that further engine development is bound not only pressure is located at Cp . The body-fixed axes xb and yb are in the
to be prohibitively expensive, but also incremental in terms ground plane.
of future technical progress. In order to improve Formula
One’s ‘green’ credentials, and accelerate road–car relevant solution of a subsidiary optimal control problem. The car
power train developments, it was decided that from 2014 model is standard and is based on a rigid-body represen-
Downloaded by [Selcuk Universitesi] at 04:51 16 January 2015

Formula One cars will have to recover thermal as well as tation of a chassis with longitudinal, lateral and yaw free-
kinetic energy that would otherwise be lost (Fédération In- doms. We will use the tyre description given in Kelly (2008)
ternationale de l’Automobile, 2013). The hope is that the in combination with an upgraded aerodynamic model. The
introduction of hybrid thermal and kinetic energy recovery important geometric modelling quantities are shown in plan
systems will produce gains that are far more significant and view in Figure 1.
cost-effective than further engine development alone. The
results of this paper show that contemporary lap times can
be maintained using less powerful turbo-charged engines, 2.1 Track model
which require only two-thirds of the fuel required by con-
We will model the track using a curvilinear coordinate sys-
temporary cars, when thermal and kinetic energy recovery
tem that follows the vehicle using the track centre-line po-
systems are employed.
sition as the curvilinear abscissa (Cossalter et al., 1999).
The vehicle and track models are covered in Section 2.
Referring to Figure 2, we describe the location of the
Most of this material is standard, or has appeared else-
mass centre of the vehicle in terms of the curvilinear ab-
where, and so this treatment is correspondingly brief. The
scissa s(t) and the vector n(s(t)). The former quantity de-
tyre modelling is covered in Appendix 1 with the vehicle
fines the distance travelled along the track centre line, while
parameters given in Appendix 2. Section 3 briefly reviews
the latter gives the position of the vehicle’s mass centre in
the mathematical background underpinning orthogonal col-
a direction perpendicular to the track centre-line tangent
location transcription techniques. The key is the classical
theory of orthogonal polynomials, which underpins most
numerical integration schemes – particularly integration
quadratures. The results appear in Section 4. The perfor-
mance of the nominal car is studied first, with this prelimi-
nary investigation used as a vehicle for illustrating the use
and features of adaptive pseudo-spectral optimisation tech-
niques. The main results of the paper relate to the optimal
control of energy storage systems in Formula One racing.
We begin with the optimal control of contemporary KERS,
which is then followed by an investigation of the hybrid
kinetic and thermal energy recovery systems to be intro-
duced in the 2014 racing season. The conclusions appear in
Section 5, followed by the Acknowledgements section.

2. Car and track model

The track and car kinematics are modelled using ideas from Figure 2. Curvilinear-coordinate-based description of a track
segment Z. The independent variable s represents the elapsed
classical differential geometry. As was explained in Peran- centre-line distance travelled, with R(s) the radius of curvature
toni and Limebeer (2014), the track description is based on and N (s) the track half-width; nx and ny represent an inertial
measured data with the curvature of the track found as the reference frame.
 International Journal of Control 2067

vector t(s(t)). It is assumed that the travelled distance s(t) is

an increasing function of time, and that ‘time’ and ‘distance’
can be thought of as alternative independent variables. The
standard dot notation will be used to signify derivatives
with respect to time. At any point s, the track’s curvature
is given by C (the s-dependence is implied) and its radius
of curvature is given by R. The track centre-line tangent
vector t will be described in terms of the track orientation
angle θ , with the track’s half-width given by N . The yaw
angle of the vehicle is given by ψ and the angle between
the vehicle and the track by ξ : ψ = θ + ξ . In this coor-
dinate, the system constraints on the track width are easily
expressed in terms of constraints on the magnitude of n.
It follows by routine calculation (Perantoni & Limebeer,
2014) that

u cos ξ − v sin ξ Figure 3. Tyre force system. The inertial reference frame is
ṡ = , (1) shown as nx and ny .
1 − nC
Downloaded by [Selcuk Universitesi] at 04:51 16 January 2015

in which u and v are the longitudinal and lateral components from Equation (3); ω = ψ̇ is the vehicle yaw rate.
of the car’s velocity, respectively. The rate of change of n is
given by
2.2 Car model
ṅ = u sin ξ + v cos ξ. (2) Each tyre produces longitudinal and lateral forces that are
responsive to the tyres’ slip (see Appendix 1). These forces
Differentiating ψ = ξ + θ with respect to time results in together with the steer and yaw angle definitions are given
in Figure 3.
ξ̇ = ψ̇ − C ṡ. (3) Balancing forces in the longitudinal and lateral direc-
tions, while also balancing the yaw moments, gives

2.1.1 Change of independent variable d

The ‘distance travelled’ will be used as the independent M u(t) = Mωv + Fx
dt
variable. This has the advantage of maintaining an explicit d
connection with the track position, as well as reducing (by M v(t) = −Mωu + Fy
dt
one) the number of problem state variables. Since d  
Iz ω(t) = a cos δ(Ff r y + Ff l y ) + sin δ(Ff r x + Ff l x )
dt  
dt
dt = ds = Sf (s)ds, + wf sin δFf r y − cos δFf r x − wr Frr x
ds  
+ wf cos δFf l x − sin δFf l y + wr Frl x
 
where Sf comes from Equation (1), there follows − b Frr y + Frl y , (7)
 −1
ds 1 − nC in which Fx and Fy are the longitudinal and lateral forces,
Sf = = . (4)
dt u cos ξ − v sin ξ respectively, acting on the car. These forces are given by

The quantity Sf is the reciprocal of the component of the

vehicle velocity in the track-tangent direction (on the centre Fx = cos δ(Ff rx + Ff lx ) − sin δ(Ff ry + Ff ly )
line at any point s). There follows + (Frrx + Frlx ) + Fax , (8)

dn Fy = cos δ(Ff ry + Ff ly ) + sin δ(Ff rx + Ff lx )

= Sf (u sin ξ + v cos ξ ) (5)
ds + (Frry + Frly ), (9)
from Equation (2), and
in which Fax is the aerodynamic drag force. These equations
dξ can be expressed in terms of the independent variable s as
= Sf ω − C (6) follows:
ds
 2068 D.J.N. Limebeer et al.

where Droll ∈ [0, 1]. This can either be chosen as represent-

du ing the suspension set-up, or it can be optimised so as to
= Sf (s)u̇ (10)
ds represent an optimised roll balance.
dv
= Sf (s)v̇ (11)
ds
dω 2.3.2 Nonnegative tyre loads
= Sf (s)ω̇. (12) The forces satisfying Equations (13), (14), (15) and (16) are
ds
potentially both positive and negative. Negative forces are
indicative of vertical reaction forces, while positive forces
2.3 Tyre forces are fictitious ‘forces of attraction’. Since the model being
The tyre forces have normal, longitudinal and lateral com- used here has no pitch, roll or heave freedoms, none of the
ponents that act on the vehicle’s chassis at the tyre–ground wheels are free to leave the road, while also keeping faith
contact points and react on the inertial frame. The rear- with Equations (13)–(16).
wheel tyre forces are expressed in the vehicle’s body-fixed To cater for the possible ‘positive force’ situation within
reference frame, while the front tyre forces are expressed in a nonlinear programming environment, we introduce the
a steered reference frame (refer again to Figure 3). In each vector,
case, these forces are a function of the normal load and the ⎡ ⎤
Downloaded by [Selcuk Universitesi] at 04:51 16 January 2015

tyre’s longitudinal and lateral slip. F̄f rz

⎢ F̄ ⎥
⎢ f lz ⎥
F̄z = ⎢ ⎥, (17)
⎣ F̄rrz ⎦
2.3.1 Load transfer
F̄rlz
In order to compute the time-varying tyre loads normal to
the ground plane, we balance the forces acting on the car
and define a vector of nonpositive loads,
in the nz direction and balance moments around the body-
fixed xb - and yb -axes (see Figure 1). Balancing the vertical
forces gives Fz = min(F̄z , 0); (18)

the minimum function min( ·, ·) is interpreted elementwise.

0 = Frrz + Frlz + Ff rz + Ff lz + Mg + Faz , (13)
It is clear that F̄z and Fz will be equal unless at least one
entry of F̄z is positive (i.e. nonphysical). We now argue that
in which the F· ·z ’s are the vertical tyre forces for each of its the model must respect the laws of mechanics at all times
four wheels, g is the acceleration due to gravity and Faz is and so Equations (13), (14) and (15) must be enforced
the aerodynamic down force acting on the car. Balancing unconditionally. In contrast, we assume that the solution to
moments around the car’s body-fixed xb -axis gives Equation (16), which is only an approximate representation
of the suspension system, can be ‘relaxed’ in the event of a
0 = wr (Frlz − Frrz ) + wf (Ff lz − Ff rz ) + hFy , (14) wheel load sign reversal.
Equations (13), (14) and (15) can be written in matrix
in which Fy is the lateral inertial force acting on the car’s form as
mass centre (see Equation (9)). Balancing moments around
the car’s body-fixed yb -axis gives A1 Fz = c, (19)

0 = b(Frrz + Frlz ) − a(Ff rz + Ff lz ) + hFx while Equation (16) is given by

+ (aA − a)Faz , (15)
A2 Fz = 0. (20)
where Fx is the longitudinal inertial force acting on the car’s
In order to deal with the ‘light wheel’ situation, we
mass centre (see Equation (8)), while Faz is the aerodynamic
combine Equations (19) and (20),
down force.
Equations (13), (14) and (15) are a set of linear equa-
tions in four unknowns. A unique solution for the tyre loads A1 0 Fz c
= , (21)
can be obtained by adding a suspension-related roll balance 0 A2 F̄z 0
relationship, in which the lateral load difference across the
front axle is some fraction of the whole: in which Fz in Equation (20) has been replaced by F̄z . In the
situation where all the wheels are normally loaded, F̄z = Fz ,
Ff rz − Ff lz = Droll (Ff rz + Frrz − Ff lz − Frlz ), (16) and Equation (21) reduces to Equations (13), (14), (15) and
 International Journal of Control 2069

2.5 Wheel torque distribution

In order to optimise the vehicle’s performance, one needs
to control the torques applied to the individual road wheels.
The braking system applies equal pressure to the brake cal-
lipers on each axle, with the braking pressures between
the front and rear axles satisfying a design ratio. The drive
torques applied to the rear wheels are controlled by a dif-
ferential mechanism.

2.5.1 Brakes
We approximate equal brake calliper pressures with equal
braking torques when neither wheel on a particular axle is
locked. If a wheel ‘locks up’, the braking torque applied
to the locked wheel may be lower than that applied to the
rolling wheel. For the front wheels this constraint is mod-
Figure 4. Car aerodynamic maps. The drag coefficient CD is
the solid (blue) curve, which is a function of speed (in m/s). The elled as follows:
Downloaded by [Selcuk Universitesi] at 04:51 16 January 2015

down-force coefficient CL , also a function of speed, is given by the

dot–dash (red) curve. The speed-dependent aerodynamic centre 0 = max(ωf r , 0) max(ωf l , 0)(Ff rx − Ff lx ), (24)
of pressure is given by the dashed (magenta) curve in metres from
the front axle. The ‘+’ symbol represents measured data points.
in which ωfr and ωfl are the angular velocities of the front
right and front left wheel, respectively. If either road wheel
‘locks up’, the corresponding angular velocity will be non-
(16). On the other hand, if there is a ‘light wheel’, the
positive and the braking torque constraint (24) will be in-
mechanic equations (13), (14) and (15) will be satisfied by
active.
the nonpositive forces Fz , while the roll balance equation
is satisfied by the now fictitious forces F̄z that contain a
force of attraction. The nonpositive forces Fz are used to
2.5.2 Differential
represent the normal tyre loads in the rest of the model.
It is clear that the four components of F̄z have to satisfy The drive torque is delivered to the rear wheels through a
the nonlinear circularly dependent relationship (21), which limited-slip differential, which is modelled by
will be solved by an NLP algorithm.
R(Flrx − Frrx ) = −kd (ωlr − ωrr ), (25)

in which ωlr and ωrr are the rear-wheel angular velocities,

2.4 Aerodynamic loads R is the wheel radius and kd is a torsional damping coeffi-
The external forces acting on the car come from the tyres cient. The special cases of an open- and a locked-differential
and from aerodynamic influences. The aerodynamic force correspond to kd = 0 and kd arbitrarily large, respectively.
is applied at the centre of pressure, which is located in the Limited slipping occurs between these extremes.
vehicle’s plane of symmetry. The drag and lift forces are
given by
3. Orthogonal collocation methods
1 The use of orthogonal collocation (pseudo-spectral) meth-
Fax = − CD (u) ρ A u2 , (22) ods in trajectory optimisation problems has been gathering
2
pace since the 1980s. An early example was the solution of
the classical brachistochrone problem using a Chebyshev
and polynomial expansion description of the state and control
(Dooren & Vlassenbroeck, 1980). In these techniques, a
1 series of node points, called collocation points, are defined
Faz = CL (u) ρ A u2 , (23) at which the state and control vectors are represented in
2
discrete form. In collocation methods, implicit integration
schemes such as Gauss–Legendre quadrature rules (Hilde-
respectively. The speed-dependent drag and down-force co- brand, 1974; Lanczos, 1956) are used to solve the system
efficients, and the speed-dependent location of the aerody- equations and integrate the stage cost. In Williams (2004),
namic centre of pressure are given in Figure 4. the pseudo-spectral method is generalised to consider
 2070 D.J.N. Limebeer et al.

collocation based on the roots of the derivatives of Jacobi 3.2 Grid refinement
polynomials. The Legendre nodes can be obtained as a par- The numerical solution of an optimal control problem in-
ticular case of this more general formulation. We do not volves three steps: (1) the transcription of the optimal con-
propose to review this literature, but the interested reader trol problem into a nonlinear programming problem (NLP);
will find much of this work described in Patterson and Rao (2) the solution of the (sparse) NLP; and (3) a review of the
(2012) and Garg, Hager, and Rao (2011). The work pre- accuracy of the solution and, if necessary, a refinement of
sented in this paper makes use of a Radau pseudo-spectral the mesh followed by a repeat of steps (1)–(3).
method (Garg et al., 2010, 2011) that has been implemented The accuracy and efficiency of this process can be in-
in the software package GPOPS-II (Patterson & Rao, 2013). fluenced by many things including the transcription pro-
The reader may also like to consult Betts (1998, 2001). cess itself. Factors favouring the use of simple low-order
integration schemes such as the trapezoidal rule include
sparse Jacobian matrices, sparse Hessians and fewer right-
3.1 Optimal control
hand side evaluations (when finite differencing is used to
The purpose of an optimal control computation is to de- compute derivatives) as compared with more complex inte-
termine the state and control associated with a system in gration schemes. Set against these advantages are the fact
order to minimise a performance index. When expressed in that these simple integration schemes have relatively poor
Bolza form, the performance index is given by accuracy and so a finer grid is needed, thereby inflating the
Downloaded by [Selcuk Universitesi] at 04:51 16 January 2015

 number of NLP decision variables required.

  tf
Global orthogonal collocation methods lie at the other
J =  t0 , x(t0 ), tf , x(tf ), p + l(t, x(t), u(t), p)dt,
t0 end of the spectrum, since a single integration segment is
(26) used. In this case, increased accuracy comes with increasing
the degree of the interpolating polynomials used. For prob-
while the system and operating constraints are given by lems whose solutions are smooth pseudo-spectral meth-
⎧ dx ods converge rapidly. The disadvantages associated with
⎪
⎪ − f (t, x(t), u(t), p) = 0 (global) pseudo-spectral methods derive from the fact that
⎨ dt
g (t, x(t), u(t), p) = 0 even smooth problems may require high-order polynomi-
(27)
⎪
⎪ h (t, x(t), u(t), p) ≤ 0 als. In the case of problems with nonsmooth solutions, the
⎩  
gb x(t0 ), x(tf ), u(t0 ), u(tf ), p = 0, convergence rate of a pseudo-spectral method may be slow
(as the degree of the approximating polynomial increases)
where t0 ≤ t ≤ tf is the optimisation interval with tf either and a poor approximation results even if high-degree poly-
fixed, or a free optimisation parameter. The vector p ∈ Rnp nomials are used.
contains fixed parameters to be optimised,1 and x(t) ∈ Rn A second limitation of global orthogonal collocation
and u(t) ∈ Rm are the state and control vectors, respectively. methods is that the use of a high-degree polynomial results
The vector-valued function f (·) ∈ Rn describes the system in a dense NLP. An alternative is to segment the optimal con-
dynamics. The vector functions g(·) ∈ Rng and h(·) ∈ Rnh trol problem (as with conventional integration algorithms),
define the equality and inequality constraints, respectively, and then employ orthogonal collocation techniques within
for the system. The subscript b refers to the boundary con- each segment. GPOPS-II uses a two-tiered grid refinement
straints with gb (·) ∈ Rngb . The scalar function l(·) is the strategy that refines both the problem segmentation and the
stage cost that is a function of the state, the controls and the orthogonal polynomial orders (Darby, Hager, & Rao, 2011;
parameters. Patterson, Hager, & Rao, 2014). If the integration error
Certain numerical techniques (like orthogonal colloca- across a particular segment is uniform, the number of col-
tion methods) were developed with a fixed optimisation location points may be increased. If the error at an isolated
interval in mind. In order to normalise the optimisation in- point within the segment is significantly larger than those
terval in the general case, one replaces t ∈ [t0 , tf ] with τ ∈ at other points within the segment it may be subdivided
[ − 1, 1] by using the change of variable, (at these large-error points).

tf − t 0 tf + t0
t= τ+ . (28)
2 2
4. Results
This mapping can be used with any free finite initial and We begin with a brief study of the nominal vehicle per-
terminal times. formance with a view to introducing some of the technical
Direct methods of the type employed here transcribe, features of orthogonal collocation methods with adaptive
or convert, infinite-dimensional optimal control problems grid refinement (Darby et al., 2011; Patterson et al., 2014).
into finite-dimensional optimisation problems with alge- These preliminary results are also used to reconcile the re-
braic constraints. sults obtained with orthogonal collocation methods with
 International Journal of Control 2071

Figure 6. The final optimisation mesh showing the segment

boundaries and collocation points. The track centre line (green)
with the segment boundaries shown as (red) circles. The black
Downloaded by [Selcuk Universitesi] at 04:51 16 January 2015

dots are the collocation points.

Figure 5. Plan view of the ‘Circuit de Cataluya’ (Barcelona)

with the start–finish line shown as . Shown also are the track
boundaries, the track centre line (green), the initial mesh segment given in Perantoni and Limebeer (2014), with exact agree-
boundaries (red circles) and the collocation points (black dots); ment achieved when the simple constant-coefficient aero-
a magnified version of part of the track is given in Figure 6. All dynamic model given in Perantoni and Limebeer (2014) is
distances are in metres. used.

the simpler techniques employed in Perantoni and Lime-

beer (2014). The main purpose of this section is to study 4.2 Energy recovery systems
racing-related kinetic and thermal energy recovery systems. The prohibitively high cost of engine development, in com-
bination with a drive towards a ‘greener’ sport, resulted
in the introduction of kinetic and thermal energy recovery
4.1 Nominal car performance devices into Formula One. The underlying concept is that
Our study of the use of orthogonal collocation methods the energy recovered during braking, and/or from the en-
in the solution of minimum-lap-time problems will make gine exhaust gases, should be redeployed to improve both
use of the Barcelona Formula One circuit. We begin this the fuel efficiency and the car’s lap times. Hopefully, the
section by briefly reviewing the key features of that circuit. continued development of ERSs of this type will produce
Figure 5 shows a plan view of the circuit together with cost-effective gains that are also relevant to road cars. We
its numbered corners. In preparation for the solution of will study the optimal control of these systems by recognis-
the optimal control problem, one can also see an initial ing the power transfer and energy storage restrictions im-
mesh that comprises 20 segments with four mesh points posed by the technical regulations relating to Formula One
per segment – the mesh points lie along the track’s centre (Fédération Internationale de l’Automobile, 2012, 2013).
line.
Approximate distances from the start–finish line to each
of the corners are given in Table 1. The car and tyre data used Table 1. Approximate distances to mid corner on the ‘Circuit de
in this study are given in Tables A1 and A2 in Appendix 2. Catalunya’ track (in metres from the start–finish line).
A magnified view of turns  10 −
16 is shown in Figure 6.
Corner Distance (m) Corner Distance
This figure shows the centre line and a refined grid that
follows 15 grid refinement steps. The final segmentation is 1 725 9 2750
uneven as are the orders of the approximating polynomials. 2 800 10 3350
The solution of the minimum-lap-time problem for the 3 1050 11 3500
nominal car results in the vehicular speed profile given in 4 1650 12 3650
5 2000 13 3900
Figure 7; the locations of each of the 16 corners are also 6 2250 14 4000
shown. The predicted lap time in this case is 80.50 s, 7 2420 15 4050
with the racing line for part of the circuit shown in Fig- 8 2500 16 4200
ure 8. These results are in broad agreement with those
 2072 D.J.N. Limebeer et al.

Figure 8. Racing line (red) through turns ➉– ; the track bound-

aries are shown in (black) with the centre line (blue).
Figure 7. Car speed (in m/s) as a function of position on the
Downloaded by [Selcuk Universitesi] at 04:51 16 January 2015

Barcelona Formula One track (measured in metres from the start–

finish line). The corner numbers correspond to those given in following path constraints are introduced:
Figure 5.

max
PIC + Pkers − Pm ≥ 0
, (29)
Pkers − Pm H (−Pm ) ≥ 0
4.2.1 Kinetic energy recovery
Contemporary (2013 and prior) KERS facilitates the cap- in which H(·) is the Heaviside step function.2 It is also nec-
 s limit −Pkers ≤ Pkers ≤
max
ture of kinetic energy that derives from braking the moving essary to enforce the KERS power
vehicle. The recovered energy can be stored for later use Pkers , the ES capacity limit 0 ≤ 0 Pkers ds ≤ 400 kJ for all
max

in propelling the car. Referring to Figure 9, the reader will values of s to race distance, and the per-lap ES discharge
appreciate that a vehicle without KERS can operate only limit 0 ≤ lap H (Pkers )Pkers ds ≤ 400 kJ required in the reg-
max
along the horizontal axis as far to the right as PIC ; the en- ulations
 (Fédération Internationale de l’Automobile, 2012);
gine’s peak power limit. The left-hand limit will be dictated lap (·)ds are line integrals carried out around a lap of the
by the braking capacity of the tyres. With the introduction circuit.
of KERS, the vehicle can operate within the cross-hatched The minimum-lap-time optimal control problem was
region; the KERS power limits are Pkersmax
= ±60 kW, while solved using GPOPS-II with an optimal lap time of 80.23
the energy released from the ES may not exceed 400 kJ s achieved. This is 0.27 s faster than the lap time achieved
in any one lap. In order to enforce these restrictions, the without the KERS. The KERS power usage is shown in

Figure 9. Operating regime of the 2013 KERS. The ordinate represents the KERS power flow Pkers , while the abscissa represents the
power Pm delivered to the rear wheels.
 International Journal of Control 2073

Figure 10. Power usage with conventional KERS. The solid

Figure 11. Energy usage with conventional KERS. The (blue)
(black) curve shows the vehicle’s speed (left-hand axis in m/s),
Downloaded by [Selcuk Universitesi] at 04:51 16 January 2015

curve is the vehicle speed divided by 100 m/s, the (red) curve
while the solid (red) curve is the power delivered to/from the
is the ES’s state of charge, while the (green) curve is the energy
KERS ES (right-hand axis in W).
transferred from the ES to the KERS divided by 400 KJ for a
racing lap.
Figure 10, while the energy usage is illustrated in Figure 11.
It is clear from Figure 10 that the KERS operates against
the allowable power limits using a bang-bang (switching) (3) Turbo-compounded engines (Mamat, Romagnoli,
strategy,3 with maximum power drawn from the KERS on & Martinez-Botas, 2012) have been introduced
fast portions of the track. Figure 10 shows that the KERS that include a turbine, a compressor and a
is being deployed between turns and ➀, through turn motor-generator on a common drive shaft; this
➅ and between turns ➈ and ➉. During periods of rapid motor-generator is referred to as the motor-
deceleration, the KERS is recharged, which is evident on the generator-unit-heat (MGU-H). The turbine is by-
entry to most corners. Bang-bang behaviour is not always passed by a controllable waste gate; closing the
possible during recharging, because the KERS regeneration waste gate will increase the power generated by the
power cannot exceed the braking power available. MGU-H, but decrease the power generated by the
Figure 11 shows the vehicle speed, the state of charge engine. There is no limit on the amount of energy
of the KERS4 and the energy transferred from the ES to than can be recovered from the exhaust gases. The
the KERS drive motor as a function of track position. The operation of the thermal energy recovery system
optimal strategy is to maximise the stored energy on the is summarised in Figure 12, which shows a plot
entry to turn in preparation for a period of prolonged of power from the internal combustion (IC) engine
KERS usage out of turn . The KERS is heavily utilised, PIC versus power supplied by the MGU-H; denoted
although its full-energy storage capacity is not required. Ph . When the waste gate is open, the MGU-H re-
The KERS per-lap discharge limit is an active constraint quires 60 kW to operate the compressor and the IC
which is met at the start–finish line; this constraint restricts engine output power is boosted by 20 kW. When
the KERS power usage to the three full-power pulses shown the waste gate is closed, 40 kW is recovered from
in Figure 10. the MGU-H and the IC engine output power drops
to approximately 440 kW. All the figures given here
are representative values only.
4.2.2 Kinetic and thermal energy recoveries
(4) A motor-generator-unit-kinetic (MGU-K) is cou-
The energy recovery rules for 2014 are more complex than pled directly to the engine crankshaft. As before,
those used previously, with the optimal control problem the MGU-K can be used to recover vehicular ki-
correspondingly more intricate. The key differences are the netic energy during braking, or boost the power to
following: the rear wheels during periods of firm acceleration.
(5) The car’s ES can store up to 4 MJ.
(1) The car is restricted to a maximum 100 kg of fuel (6) The ES can accept up to 2 MJ per lap from the
per race. MGU-K.
(2) The fuel mass flow rate must not exceed 0.028 kg/s (7) The MGU-K can draw up to 4 MJ per lap from the
(or 100 kg/hour); ES.
 2074 D.J.N. Limebeer et al.

Figure 12. Operating regime of the 2014 engine and thermal energy recovery system (ERS) at full power; power absorbed by the MGU-H
is deemed positive.
Downloaded by [Selcuk Universitesi] at 04:51 16 January 2015

(8) The power flows to and from the MGU-K is re- In order to monitor the resource constraints associated
stricted to ±120 kW. with the fuel usage, and the ES and MGU-K usage, four
(9) The power flow to and from the MGU-H is un- auxiliary state variables are introduced as follows:
restricted; see Appendix 3 in Fédération Interna-
tionale de l’Automobile (2013). 
F = Ḟ ds. (32)
lap
It is clear that operating the engine with an open waste
gate is very inefficient from an energy conservation point
of view. The engine controls are the fuel-flow rate and the This state is used to monitor the fuel consumption and is
waste gate opening. It is assumed that the engine full-speed constrained by 0 ≤ F ≤ 100/nl kg, in which nl is the number
rotational losses are 40 kW. of laps in the race. As pointed out in point (2) , Ḟ , when
We will now explain how these various constraints were normalised, is constrained to the interval [0, 1]. The second
set up in the optimal control calculation. The power avail- auxiliary state is given by
able to the rear wheels, Pm , is constrained by the following
inequality: Ės = − (Ph + Pk ) . (33)

(Pmax + PWg Wg )Ḟ − Ploss + Pk − Pm ≥ 0. (30)

This state is used to monitor the stored energy and is con-
strained by 4 MJ ≥ Es ≥ 0 over the whole circuit. The third
The Pmax term in Equation (30) represents the power gen- auxiliary state is described by
erated by the IC engine under full fuelling when the waste
gate is closed; the normalised fuel mass flow rate is given ⎧
by Ḟ ∈ [0, 1]. The second term in Equation (30) represents ⎨ Pk Pk > 0, Ph > 0
the power boost resulting from opening the waste gate; Wg ĖES2K = Pk + Ph Pk > −Ph > 0 (34)
⎩
∈ [0, 1] is the waste gate control with PWg = 20 kW. The 0 otherwise.
third term represents the engine’s rotational losses and is set
at a constant 40 kW for illustrative purposes, Pk is the power This state is used to monitor the energy supplied to
delivered to the MGU-K (Pk is positive when the MGU-K the MGU-K from the ES and is constrained by 4 MJ ≥
is motoring) and Pm represents the mechanical power deliv- EES2K ≥ 0. The first alternative corresponds to the case
ered to the rear wheels (positive values accelerate the car). when both electrical machines are running as motors. In
The power generated/absorbed by the MGU-H is given by this case, it is only the MGU-K energy usage that is ‘taxed’.
 max   The second alternative corresponds to the case in which the
Ph = Ph − Phmin Wg + Phmin Ḟ (31) MGU-K is motoring, but the MGU-H is generating. In this
case, the power generated by the MGU-H is off set against
and is a function of the waste gate opening and the fuel flow the MGU-K power requirements, with the difference mon-
rate; Phmin = −40 kW and Phmax = 60 kW. itored.
 International Journal of Control 2075

The fourth auxiliary state is used to tax energy flows

from the MGU-K into the ES and is described by
⎧
⎨ Pk Pk < 0, Ph < 0
ĖK2ES = Pk + Ph 0 > −Ph > Pk (35)
⎩
0 otherwise;

this state is constrained by 0 ≥ EES2K ≥ −2 MJ . The first

alternative corresponds to the case when both electrical ma-
chines are operating as generators. In this case, it is only the
MGU-K energy generation that is penalised. The second al-
ternative corresponds to the case in which the MGU-K is
generating, while the MGU-H is motoring. In this case, the
power required by the MGU-H is off set against the power
generated by the MGU-K, with the difference monitored.
As before, differentiable approximations to the Heaviside
step function are used to enforce constraints (34) and (35). Figure 13. Energy usage in the 2014 ERS. The (cyan) plot is the
vehicle speed divided by 100 m/s; the (red) line is the ES state of
Downloaded by [Selcuk Universitesi] at 04:51 16 January 2015

Inequality constraint (30), in combination with the four aux-

iliary states (32)–(35) and their associated box constraints charge; the (green) plot is the fuel used divided by the race fuel
limit/66 – Barcelona is a 66 lap race; the (blue) line is the energy
are used to model the power and energy constraints required transferred from MGUK to ES divided by 2 MJ; the (magenta)
by the 2014 Formula One rules (Fédération Internationale line is the energy transferred from ES to MGUK divided by 4 MJ.
de l’Automobile, 2013).
The minimum-lap-time optimal control problem for the
and by the fact that the state of charge at the beginning
2014 car was solved using GPOPS-II. The resulting op-
and end of the lap are constrained to be equal. It is likely
timal lap time was 82.43 s, which is 1.93 s slower than
that the greater use of the ES will occur on suboptimal laps
the lap time achieved with the nominal car studied in Sec-
involving the overtaking of other vehicles.
tion 4.1. Contributing to this downgraded performance is
The vehicle’s fuel usage, MGU-K and waste gate con-
the fact that the 2014 car is 50 kg heavier than its prede-
trols, and the power delivered to the rear wheels are illus-
cessor and its IC engine is 120 W less powerful. The com-
trated in Figure 14, where one observes the liberal use of the
monly used 0.03 s/kg ‘rule of thumb’ would cut the car’s
MGU-K to boost the total drive power available at the rear
lap time to 80.94 s if it had a mass of 660 kg rather than
wheels. Again, a bang-bang-type strategy is again being
710 kg.
employed by the MGU-K during motoring. The waste gate
In order to analyse the performance of the 2014 car,
we will begin with the full-lap energy usage as illustrated
in Figure 13. The first thing to observe from the red, blue
and magenta traces is that an optimal lap is utilising fully
the fuel and the MGU-K generation allowances, but not the
MGU-K motoring allowance. Not surprisingly, there are
braking periods when no fuel is being used – this occurs, for
example, on the entries to turns ➀, ➃, ➄ and ➉; see Table 1.
Also as expected, the MGU-K is being used to increase the
vehicle speed on the fast straights between, for example,
turns and ➀, turns ➂ and ➃, and between turns ➈ and
➉. The MGU-K is being used to recharge the ES during
the braking sections that correspond approximately to the
periods of zero fuel usage; see, for example, the sections
entering turns ➀, ➃ and ➄. There is also intermittent energy
recover occurring on the slower section between turns ➉ and
. The gradients of the magenta and blue curves, which
represent the MGU-K usage, differ by a factor of two due
to the way in which they have been normalised. Notable Figure 14. Normalised power usage for the 2014 ERS. The
is the fact that the ES is under utilised on an optimal lap (blue) plot is the mechanical power delivered to the rear wheels
divided by 440 kW; the (green) plot is normalised fuel usage rate
of the Barcelona circuit; indeed only 1.12 MJ of its 4 MJ (‘1’ corresponds to the maximum allowed rate); the (red) plot is
capacity is being used. This is a result of restrictions on the KERS power divided by 120 kW; the (magenta) plot is the waste
total amount of MGU-K generation and motoring allowed, gate position (0-closed, 1-open).
 2076 D.J.N. Limebeer et al.

Figure 15. Energy usage for a 2014 ERS qualifying lap. The Figure 16. Normalised power usage for the 2014 ERS in qual-
(cyan) plot is the vehicle speed divided by 100 m/s; the (red) line ifying. The (blue) plot is the mechanical power delivered to be
Downloaded by [Selcuk Universitesi] at 04:51 16 January 2015

is the ES state of charge; the (blue) line is the energy transferred rear wheels divided by 440 kW; the (green) plot is normalised
from MGUK to ES divided by 2 MJ; the (magenta) line is the fuel usage rate (‘1’ corresponds to the maximum allowed rate);
energy transferred from ES to MGUK divided by 4 MJ. the (red) plot is KERS power divided by 120 kW; the (magenta)
plot is the waste gate position (0-closed, 1-open).

is essentially unused due to its poor energy efficiency. The

MGU-K is again being used to recharge the ES on the entry ➉. This extravagant use of the waste gate derives from the
to corners; several particular instances of this are discussed ability to simply ‘run down’ the ES on a qualifying lap. In
above. In sum, these results are encouraging and suggest contrast to the racing lap, the waste gate is typically open
that contemporary car performances can be achieved, on when the engine is fully fuelled. On the entry to turns ➀,
two-thirds of the current fuel usage, especially if the vehi- ➃, ➆ and ➉, the waste gate is being closed a little before si-
cle’s aerodynamic drag can be improved. multaneously cutting the fuel and the MGU-K. We believe
Figure 15 considers the case of a qualifying lap for that this strategy is necessary in order to save power as the
which there is a fuel mass flow rate limit, but no total fuel ES is fully drained at the end of the lap. Also evident is the
usage restriction. Another fundamental difference between fact that the MGU-K is used regeneratively on the entry to
this case and the one considered in Figure 13 is the free- corners when possible, although its generation quota is not
dom to start the lap with a fully charged ES, and to then fully utilised.
end it with the ES fully discharged. The qualifying lap is
completed in 82.06 s, which is 0.37 s faster than an optimal
racing lap with 100 kg of fuel. In this case, the MGU-K
motoring energy quota is fully utilised, with the MGU-K
regeneration quota only partially used. In its qualifying con-
figuration, the engine is run with the waste gate open for
sustained periods of time when maximum engine power is
needed. During these periods of time, the ES will be sup-
plying both the MGU-K and the MGU-H, with the latter
used to drive the engine boost compressor. The main benefit
of using the MGU-K, in combination with an open waste
gate during qualifying, is the higher top speeds achieved
between turns and turn ➀, between turns ➂ and ➃, and
between turns ➈ and ➉.
The vehicle’s fuel usage, MGU-K and waste gate con-
trols, and the power delivered to the rear wheels during a
qualifying lap are illustrated in Figure 16. Unlike the racing
lap, the car is repeatedly brought to its top speed by the si-
Figure 17. Speed comparison for the three cases studied. The
multaneous use of the MGU-K and a fully open waste gate; (red) curve is a qualifying lap, the (cyan) curve is the racing lap
this can be observed on the straight between turns and with 100 kg of fuel, and the (blue) curve is the racing lap with
➀, and intermittently, most of the way between turns ➂ and 75 kg of fuel.
 International Journal of Control 2077

Figure 17 shows a speed comparison between a qual- 3. Optimal control problems with costs and dynamics that are
ifying lap, a racing lap with 100 kg of fuel and a racing linear in the control usually involve switching strategies. In
lap with only 75 kg of fuel (with a lap time of 85.23 s). these problems, the control takes on limit values dictated
by a switching function that is derivable from the minimum
The energy management strategy used in the 75 kg case is principle; this is the well-known bang-bang principle. In the
broadly the same as that employed in the 100 kg case. All case that the switching function is zero for a finite time, the
the fuel is used, the 2-MJ MGU-K generation quota is fully problem has a singular arc and a bang-singular-bang strategy
utilised, while the MGU-K motoring quota is only partially results (see Kirk, 1970, p. 246).
employed; only 0.8 MJ of the energy storage capacity is re- 4. This is defined as a number between 0 and 1; with 0 repre-
senting a completely discharged and 1 a fully charged ES.
quired. The higher straight line speed for the qualifying lap
over the racing lap is clearly evident. The need to ‘lift off ’
the throttle into turns ➀, ➃ and ➉, thus conserving fuel, is
References
evident from the speed profile in the 75 kg of fuel case.
Åkesson, J., Årzén, K.E., Gäfvert, M., Bergdahl, T., &
Tummescheit, H. (2010). Modeling and optimization with
optimica and jmodelica.org-languages and tools for solving
5. Conclusion large-scale dynamic optimization problems. Computers and
The use of pseudo-spectral methods for vehicular optimal Chemical Engineering, 34, 1737–1749.
control problems has been increasing since the early 1980s. Betts, J.T. (1998). Survey of numerical methods for trajectory
While the majority of these applications relate to aerospace optimization. Journal of Guidance, Control and Dynamics,
Downloaded by [Selcuk Universitesi] at 04:51 16 January 2015

21(2), 193–207.
problems, we have demonstrated that they are also well Betts, J.T. (2001). Practical methods for optimal control and esti-
suited to ground-vehicle-related optimisation problems. mation using nonlinear programming (2nd ed. ). Philadelphia,
The results in this paper focus on the control of ERSs in PA: SIAM.
Formula One racing cars, where the emphasis is on minimis- Brayshaw, D.L., & Harrison, M.F. (2005). A quasi steady state ap-
ing the lap time. This minimum-time type of performance proach to race car lap simulation in order to understand the ef-
fects of racing line and centre of gravity location. Proceedings
index leads to the use of a bang-bang strategy during KERS- of the IMechE Part D: Journal of Automobile Engineering,
assisted and ERS-assisted driving. A strategy reminiscent 219, 725–739.
of bang-bang control is also employed during braking, al- Casanova, D. (2000). On minimum time vehicle manoeuvring: The
though regenerative braking is limited by the braking power theoretical optimal lap (PhD thesis). Bedfordshire: School of
available making the limits for the KERS bang-bang control Engineering, Cranfield University.
Cossalter, V., Lio, M.D., Lot, R., & Fabbri, L. (1999). A general
variable. Following the analysis of contemporary KERS, we method for the evaluation of vehicle manoeuvrability with
analysed the control issues relating to the 2014 ERS that has special emphasis on motorcycles. Vehicle System Dynamics,
both kinetic and thermal energy recovery features. Unlike 31, 113–135.
the conventional KERS, the upcoming ERS has an explicit Darby, C.L., Hager, W.W., & Rao, A.V. (2011). An hp-adaptive
fuel-saving role. It is demonstrated that the ERS in com- pseudospectral method for solving optimal control problems.
Optimal Control Applications and Methods, 32, 476–502.
bination with a lower powered turbo-boosted engines can Dooren, R.V., & Vlassenbroeck, J. (1980). A new look at the
provide contemporary levels of lap-time performance with brachistochrone problem. Journal of Applied Mathematics
approximately 100 kg (rather than the current 150 kg) of and Physics, 31, 785–790.
fuel per race. It is also shown that the IC engine and the Fédération Internationale de l’Automobile. (2012, September
ERS have to be operated in a very different manner during 25). 2013 formula one technical regulations (Tech. Report).
Retrieved from FIA: http://argent.fia.com/web/fia-public.nsf/
qualifying, when the short-term, but profligate use of stored 7EE7D5392A94A532C1257A850051A2C9/$FILE/2013%
electrical energy, is allowed. The potential for redeploying 20TECHNICAL%20REGULATIONS.pdf
these technologies in everyday road cars is self-evident. Fu- Fédération Internationale de l’Automobile. (2013, January 23).
ture work is expected to include a three-dimensional track, 2014 Formula One technical regulations (Tech. Report).
more complicated (and realistic) aerodynamic maps and a Retrieved from FIA: http://www.fia.com/sites/default/files/
regulation/file/1-2014%20TECHNICAL%20REGULATIONS
suspension system. %202014-01-23 0.pdf
Garg, D., Hager, W.W., & Rao, A.V. (2011). Pseudospectral meth-
ods for solving infinite-horizon optimal control problems. Au-
Acknowledgements tomatica, 47(4), 829–837.
This work was supported by the UK Engineering and Physical Garg, D., Patterson, M.A., Darby, C.L., Françolin, C., Huntington,
Sciences Research Council. G.T., Hager, W.W., & Rao, A.V. (2011). Direct trajectory opti-
mization and costate estimation of finite-horizon and infinite-
horizon optimal control problems using a Radau pseudospec-
Notes tral method. Computational Optimization and Applications,
1. Rn denotes the set of n-dimensional real vectors. 49(2), 335–358.
2. H( · ) = 0 when the argument is nonpositive, with H( · ) = 1, Garg, D., Patterson, M.A., Hager, W.W., Rao, A.V., Benson, A.V.,
otherwise. The Heaviside step function is not differentiable at & Huntington, G.T. (2010). A unified framework for the
the origin and so is√
approximated by a smooth function such numerical solution of optimal control problems using pseu-
as H (x) ≈ (1 + x/ x 2 + )/2 in which is ‘small’. dospectral methods. Automatica, 46(11), 1843–1851.
 2078 D.J.N. Limebeer et al.

Gustafsson, T. (2008). Computing the ideal racing line using opti-

vw
mal control (Master’s thesis). Linköping: Linköping Univer- tan α = − , (37)
sity. uw
Hildebrand, F.B. (1974). Introduction to numerical analysis
(2nd ed.). New York, NY: Dover Publications. where R is the wheel radius and ωw the wheel’s spin angular veloc-
Kelly, D.P. (2008). Lap time simulation with transient vehicle ity. The quantities uw and v w are the absolute speed components of
and tyre dynamics (PhD Thesis). Bedfordshire: School of the wheel centre in a wheel-fixed coordinate system. The follow-
Engineering, Cranfield University. ing formulae give the peak values and locations of the lateral and
Kirk, D.E. (1970). Optimal control theory: An introduction. En- longitudinal friction coefficients using linear interpolation (Kelly,
glewood Cliffs, NJ: Prentice-Hall. 2008):
Lanczos, C. (1956). Applied analysis. Englewood Cliffs, NJ: Pren-
tice Hall. μx max2 − μx max1
Mamat, A.M., Romagnoli, A., & Martinez-Botas, R.F. (2012). μx max = (Fz − Fz1 ) + μx max1 , (38)
Design and development of a low-pressure turbine for tur- Fz2 − Fz1
bocompounding applications. International Journal of Gas μy max2 − μy max1
Turbine, Propulsion and Power Systems, 4(3), 1–8. μy max = (Fz − Fz1 ) + μy max1 , (39)
Fz2 − Fz1
Pacejka, H.B. (2008). Tyre and vehicle dynamics (2nd ed.).
κmax2 − κmax1
Oxford: Butterworth-Heinemann. κmax = (Fz − Fz1 ) + κmax1 , (40)
Patterson, M.A., Hager, W.W., & Rao, A.V. (2014). A ph mesh Fz2 − Fz1
refinement method for optimal control. Optimal Control Ap- αmax2 − αmax1
plications and Methods. doi: 10.1002/oca2114 αmax = (Fz − Fz1 ) + αmax1 , (41)
Fz2 − Fz1
Downloaded by [Selcuk Universitesi] at 04:51 16 January 2015

Patterson, M.A., & Rao, A.V. (2012). Exploiting sparsity in di-

rect collocation pseudospectral methods for solving optimal
control problems. Journal of Spacecraft and Rockets, 39(2), where the quantities containing a ‘1’ or a ‘2’ in the subscript are
364–377. measured tyre parameters. Next, the tyre slip is normalised with
Patterson, M.A., & Rao, A.V. (2013). GPOPS-II: A MATLAB respect to the peak slip:
software for solving multiple-phase optimal control problems
using hp-adaptive Gaussian quadrature collocation methods
and sparse nonlinear programming. ACM Transactions on κn = κ/κmax , (42)
Mathematical Software, 39(3), 1–41.
Perantoni, G., & Limebeer, D.J.N. (2014). Optimal control for αn = α/αmax . (43)
a formula one car with variable parameters. Vehicle System
Dynamics. doi:10.1080/00423114.2014.889315 Following normalisation, the slip is characterised by a combined-
Sharp, R.S., & Peng, H. (2011). Vehicle dynamics applications slip coefficient
of optimal control theory. Vehicle System Dynamics, 49(7),
1073–1111. 
Tavernini, D., Massaro, M., Velenis, E., Katzourakis, D.I., & Lot, ρ = αn2 + κn2 . (44)
R. (2013). Minimum time cornering: The effect of road sur-
face and car transmission layout. Vehicle System Dynamics, The friction coefficients in the longitudinal and lateral directions
51(10), 1533–1547. doi:10.1080/00423114.2013.813557 are described by
Thommyppillai, M., Evangelou, S.A., & Sharp, R.S. (2009). Car
driving at the limit by adaptive linear optimal preview control.
Vehicle System Dynamics, 47(12), 1535–1550. μx = μx max sin (Qx arctan(Sx ρ)), (45)
Timings, J.P., & Cole, D.J. (2012). Vehicle trajectory linearisation  
to enable efficient optimisation of the constant speed racing μy = μy max sin Qy arctan(Sy ρ) , (46)
line. Vehicle System Dynamics, 50(6), 883–901.
Timings, J.P., & Cole, D.J. (2013). Minimum maneuver time cal- with
culation using convex optimization. Journal of Dynamic Sys-
tems, Measurement, and Control, 135, 031015-1–031015-9.
Trabesinger, A. (2007). Power games. Nature, 447, 900–903. π
Sx = , (47)
Williams, P. (2004). Jacobi pseudospectral method for solving 2 arctan(Qx )
optimal control problems. Journal of Guidance, Control and π
Dynamics, 27(2), 293–297. Sy = . (48)
2 arctan(Qy )

Finally, the longitudinal and lateral components of the tyre forces

Appendix 1. Tyre friction are given by
The tyre frictional forces are modelled using empirical formulae
that are responsive to the tyre normal loads and the combined κn
slip. The tyre’s longitudinal slip is described by a longitudinal slip Fx = μx Fz , (49)
coefficient κ, while the lateral slip is described by a slip angle α
ρ
αn
(Pacejka, 2008). Following the standard conventions, we use Fy = μy Fz . (50)
ρ
 
Rωw In the car model used here, the normal loads are determined by
κ = − 1+ , (36)
uw solving the load balance equations given in Sections 2.3.1 and
 International Journal of Control 2079
Downloaded by [Selcuk Universitesi] at 04:51 16 January 2015

Figure A1. Longitudinal tyre force versus longitudinal slip; Fx is given in kN. The left-hand diagram considers five normal loads
(at α = 0◦ ); the lightest curve corresponds to 1000 N with the normal load then increased in steps of 1000 N. The right-hand diagram
considers combined slip for five values of side-slip angle at a normal load of 2000 N; the darkest curve corresponds to a slip angle of 0◦
with the slip angle increased in steps of 5◦ .
 
2.3.2. The four-wheel slip angles are given by sin δ(ψ̇wf − u) + cos δ(ψ̇a + v)
αf r = arctan ,
cos δ(u − ψ̇wf ) + sin δ(ψ̇a + v)
   
v − ψ̇b cos δ(ψ̇a + v) − sin δ(ψ̇wf + u)
αrr = arctan , αf l = arctan ,
u − ψ̇wr cos δ(ψ̇wf + u) + sin δ(ψ̇a + v)
  (51)
v − ψ̇b
αrl = arctan ,
u + ψ̇wr

Figure A2. Modulus of the tyre force as function of the longitudinal slip κ and the slip angle α (in degrees) for a normal load of 2000
N. The right-hand diagram shows contours of equal tyre force.
 2080 D.J.N. Limebeer et al.

with the longitudinal slip coefficients given by load is increased. The main reason for introducing aerodynamic
down-force generating systems in Formula One cars is to exploit
  this effect. The right-hand part of Figure A1 shows how the lon-
Rωrr
κrr = − 1 + , gitudinal force is compromised when the tyre is side slipping. As
u − ψ̇wr the side-slip angle increases, the longitudinal peak force reduces
 
Rωrl and moves towards higher slip values.
κrl = − 1 + , The normal-load-dependent inverted witch’s hat in Figure A2
u + ψ̇wr
  gives a three-dimensional representation of the combined slip
Rωf r characteristics of the tyre model at a fixed normal load of 2000 N.
κf r = − 1 + ,
cos δ(u − ψ̇wf ) + sin δ(ψ̇a + v) The tyre’s pure slip characteristics are obtained by taking vertical
  cuts through the lines α = 0 and κ = 0.
Rωf l
κf l = − 1 + . (52)
cos δ(u + ψ̇wf ) + sin δ(ψ̇a + v)
Appendix 2. Vehicle and tyre data
The left-hand part of Figure A1 shows the load dependence This appendix contains nominal values for the tyre and vehicle
of the longitudinal tyre force under pure longitudinal slip con- parameters used in this study. The air density is assumed to be
ditions; larger longitudinal forces are available when the normal ρ = 1.2 kg/m3 , while the car’s frontal area is A = 1.5 m2 .

Table A1. Vehicle parameters.

Downloaded by [Selcuk Universitesi] at 04:51 16 January 2015

Symbol Description Value

Pmax Peak engine power 560/440∗ kW

max
Pkers Maximum KERS power 60/120∗ kW
M Vehicle mass 660/710∗ kg
Iz Moment of inertia about the z-axis 450 kg m2
w Wheelbase 3.4 m
a Distance of the mass centre from the front axle 1.8 m
b Distance of the mass centre from the rear axle w−a
h Centre of mass height 0.3 m
Droll Roll moment distribution (fraction at the front axle) 0.5
wf Front wheel to car centre-line distance 0.73 m
wr Rear wheel to car centre-line distance 0.73 m
R Wheel radius 0.33 m
kd Differential friction coefficient 100 N/rpm

Note: The values marked with an asterisk relate to the 2014 car.

Table A2. Tyre friction parameters.

Symbol Description Value

Fz1 Reference load 1 2000 N

Fz2 Reference load 2 6000 N
μx1 Peak longitudinal friction coefficient at load 1 1.75
μx2 Peak longitudinal friction coefficient at load 2 1.40
κ1 Slip coefficient for the friction peak at load 1 0.11
κ2 Slip coefficient for the friction peak at load 2 0.10
μy1 Peak lateral friction coefficient at load 1 1.80
μy2 Peak lateral friction coefficient at load 2 1.45
α1 Slip angle for the friction peak at load 1 9◦
α2 Slip angle for the friction peak at load 2 8◦
Qx Longitudinal shape factor 1.9
Qy Lateral shape factor 1.9