This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY 1

Explicit Nonlinear Model Predictive Control

for Electric Vehicle Traction Control
Davide Tavernini , Mathias Metzler , Graduate Student Member, IEEE,
Patrick Gruber , and Aldo Sorniotti , Member, IEEE

Abstract— This paper presents a traction control (TC) system performance and the seamless blending of the regenerative and
for electric vehicles with in-wheel motors, based on explicit dissipative braking contributions.
nonlinear model predictive control. The feedback law, available In parallel to sliding mode control [4] and maximum trans-
beforehand, is described in detail, together with its variation for
different plant conditions. The explicit controller is implemented missible torque estimation [5] algorithms, the recent literature
on a rapid control prototyping unit, which proves the real- (see [6]–[18]) on the topic of ABS and TC shows growing
time capability of the strategy, with computing times on the interest in model-based control, with focus on model predictive
order of microseconds. These are significantly lower than the control (MPC). For example, [6] discusses a gain scheduled
required time step for a TC application. Hence, the explicit linear quadratic regulator approach for ABS control, with
model predictive controller can run at the same frequency as a
simple TC system based on proportional integral (PI) technology. experimental results on an internal-combustion-engine-driven
High-fidelity model simulations provide: 1) a performance com- vehicle with electro-mechanical brakes. Boisvert et al. [7]
parison of the proposed explicit nonlinear model predictive and Anwar and Ashrafi [8] include different approaches to
controller (NMPC) with a benchmark PI-based traction con- ABS control, i.e., linear quadratic Gaussian regulation and
troller with gain scheduling and anti-windup features, and generalized predictive control, which is reproposed in [9]
2) a performance comparison among two explicit and one implicit
NMPCs based on different internal models, with and without for a TC implementation. A linear MPC strategy is devel-
consideration of transient tire behavior and load transfers. oped in [10], where the ABS slip regulation is achieved
Experimental test results on an electric vehicle demonstrator are through torque blending between the friction brakes and
shown for one of the explicit NMPC formulations. in-wheel motors. Similarly, [11]–[13] combine ABS control
Index Terms— Electric vehicle, in-wheel motors, model and torque blending, by using linear MPC formulations.
predictive control (MPC), proportional integral (PI) control, Yoo and Wang [14] present an MPC-based ABS, with test
traction control (TC), wheel slip. results on a hardware-in-the-loop rig. The internal model
includes a tire force dynamics formulation; however, its effect
on the controller performance is not discussed in this paper,
I. I NTRODUCTION
nor, to the authors’ knowledge, in any other study in the

T HE adoption of electric drivetrains, and in particular of

in-wheel motor layouts, has the potential of significantly
enhancing the performance of wheel slip control systems, i.e.,
literature. Yuan et al. [15] present a nonlinear model predic-
tive controller (NMPC) for ABS and TC. The formulation
considers all four wheels in the same internal model.
antilock braking systems (ABS) and traction control (TC) Reference tracking is not used, since the slip ratio is solely
systems [1]. This is caused by the higher control bandwidth controlled through the constraints of the NMPC formulation.
and precision in torque modulation that electric drivetrains Moreover, the tire-road friction coefficient is considered to
can offer, with respect to the more conventional internal be known a priori, which introduces some challenges for
combustion engines and hydraulic/electro-hydraulic braking a real vehicle implementation. For an internal-combustion-
units. Murata [2] and Ivanov et al. [3] include experimentally engine-driven vehicle, [16] introduces four linear MPC
measured reductions in stopping distances and acceleration TC strategies that are compared with a hybrid explicit MPC.
times, achieved through the continuous modulation of the The hybrid design adopts a piecewise linear approximation of
electric drivetrain torques. However, further work can be done the nonlinear longitudinal tire force characteristic as a function
in terms of control design to enhance the slip ratio tracking of the slip ratio. Simulation and experimental results show the
performance enhancement of the hybrid strategy with respect
Manuscript received March 1, 2018; accepted May 2, 2018. Manuscript to the linear approaches.
received in final form May 12, 2018. This work was supported by the
European Union’s Horizon 2020 Program under Grant 653861 (SilverStream In the case of implicit NMPC, a nonlinear program-
Project, Social Innovation and Light electric VEhicle Revolution on STREets ming (NLP) problem is solved on-line at each sampling
and AMbient). Recommended by Associate Editor A. G. Stefanopoulou. time. The resulting computational load makes implicit
(Corresponding author: Aldo Sorniotti.)
The authors are with the Centre for Automotive Engineering, University NMPC difficult to implement in real automotive applications,
of Surrey, Guildford GU2 7XH, U.K. (e-mail: [email protected]; if the required sampling frequency is high. In this respect,
[email protected]; [email protected]; [email protected]). Basrah et al. [17] provide an example of real-time capable
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. NMPC for an ABS with torque blending, including a compar-
Digital Object Identifier 10.1109/TCST.2018.2837097 ison with a linear MPC approach. The results show that the
1063-6536 © 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted,
but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

computational time of the implicit NMPC, i.e., 3–4 ms on a II. E XPLICIT N ONLINEAR M ODEL P REDICTIVE C ONTROL
desktop personal computer, is within the selected time step of A. Problem Formulation
5 ms. In [15], the implicit NMPC strategy is run on a rapid
Similarly to the NMPC, the explicit NMPC requires the
control prototyping unit, with a computational time of 4–5 ms
formulation of an optimization problem, potentially including
and an implementation time step of 10 ms.
constraints on the control inputs and system states. A generic
The study of this paper presents an explicit NMPC
nonlinear optimal control problem for a finite horizon in the
(eNMPC in the remainder) for TC on electric vehicles with
time interval [tk , t f ] can be defined as the minimization of the
in-wheel drivetrains. The explicit solution is computed off-line
following cost function:
by using a multiparametric (mp) quadratic programming (QP)
approximation of the mp-NLP problem. The control action V (x[tk , t f ], u[tk , t f ], p(tk ), ν[tk , t f ])
is evaluated on-line at each sampling time starting from  tf
the current values of system states and parameters, and the  L(x(t), u(t), p(tk ), ν(t))dt + M(x(t f ), p(tk ), t f )
tk
off-line explicit solution, stored in the memory of the control (1)
unit. This drastically reduces the required computational load.
The other advantage is that the complete feedback law is where x, u, p, and ν are the state, input, parameter, and slack
available beforehand in its explicit form, which allows its variable vectors, respectively. L is the stage cost, and M is the
analysis for the range of states and reference parameters. terminal cost. The problem is subject to inequality constraints
Another important aspect is the performance comparison of the form
and critical analysis of different TC implementations. In this x min ≤ x(t) ≤ x max (2)
respect, [16] claims that the performance of the proposed MPC
u min ≤ u(t) ≤ u max (3)
“is comparable with that of a well-tuned PID” controller.
The same authors state that “the simulation and test results g(x(t),u(t), p(tk ),ν(t), t) ≤ 0. (4)
demonstrated that the l1 -optimal hybrid controller used in The ordinary differential equations (ODEs) describing the
this problem can lead to about 20% reduction in peak slip system dynamics represent the equality constraints:
amplitudes and corresponding spin duration when compared
d
to best case linear MPC counterparts.” Similarly, [17] shows x(t) = f (x(t), u(t), ps (tk ), t) (5)
the superiority of NMPC over linear MPC in terms of slip dt
control performance. The necessity of “objective benchmark- where ps is the vector of the system parameters. The initial
ing technologies” in the field of ABS/TC was pointed out conditions x(tk ) are assigned to the state vector.
in the survey study in [19]. In order to understand where The infinite-dimensional optimal control problem in (1)–(5)
the strategies of different papers stand with respect to each is discretized and parametrized, thus becoming an NLP
other, a comparison is well needed. De Pinto et al. [20] problem, which is solved through numerical methods. This
partially cover this knowledge gap, but limit the analysis approach is known as direct method [21]. In this operation,
to on-board electric drivetrains, characterized by significant the equality constraints (5) are represented as finite approxi-
torsional dynamics. Satzger and de Castro [13] include also an mations. The infinite-dimensional unknown solution, u[tk , t f ],
MPC-PI experimental comparison, but for an ABS application and the slack variables, ν[tk , t f ], are replaced by a finite
combined with torque blending. number of decision variables. The prediction horizon t p =
Based on the previous discussion, the points of novelty of t f − tk is defined as t p = N p ts , where N p is the number
this paper are as follows. of prediction steps and ts is the characteristic discretization
1) The design of TC systems based on explicit NMPC, interval of the internal model. The input signal, u[tk , t f ],
implementable at the same time step as a typical is assumed to be piecewise constant along the horizon. It is
PI controller for TC, but with better tracking calculated through the function χ and is parameterized through
performance. the vector of control parameters U such that u(t) = χ(t, U ).
2) The study of the explicit feedback control law, and Similarly, the piecewise constant slack variable trajectory is
its dependency on the vector of parameters from the parameterized through the vector of slack variables, N.
plant. The technique known as direct single shooting
3) The simulation-based analysis of the performance (see [21], [22]) is used for the management of the equality
advantages of the proposed eNMPC compared to a well- constraints. It consists of eliminating the ODE equality
tuned benchmark PI TC system with gain scheduling and constraints by substituting their discretized numerical solution
anti-windup features. into the cost function and constraint formulations. Starting
4) The sensitivity analysis of the performance of TC algo- from the continuous constraint equations (5), the numerical
rithms with respect to their time step. solution is derived by discretization and integration of
5) The discussion of the benefit of considering transient the ODEs
tire response and vertical load transfers in the internal x(tk+ j ) = φ(x(tk ), U, ps (tk ), tk+ j ), j = 1, . . . , N p . (6)
model for the NMPC formulation.
To obtain the function φ, an explicit integration scheme is
6) The presentation of experimental test results based on
selected
explicit NMPC applied to a fully electric vehicle proto-
type with in-wheel drivetrains. x(tk+ j +1 ) = F(x(tk+ j ), χ(tk+ j , U ), ps (tk ), tk+ j ) (7)
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

TAVERNINI et al.: EXPLICIT NONLINEAR MODEL PREDICTIVE CONTROL FOR ELECTRIC VEHICLE TRACTION CONTROL 3

with given initial conditions x(tk ). If the whole horizon is formulation (17)
considered, the state trajectories are all mapped into a single
1
function, and the system dynamics do not appear any more as V0 (z, x p )  (z − z 0 )T H0(z − z 0 )
equality constraints 2
+ (D0 + (x p − x p,0 )T F0 )(z − z 0 ) + Y0 (x p )
x(tk+ j ) = F(x(tk+ j −1 ), χ(tk+ j −1 , U ), ps (tk ), tk+ j −1 ) (15)
x(tk+2 ) 1
   Y0 (x p )  (x p − x p,0 ) ∇x p x p V (z 0 , x p,0 )(x p − x p,0 )
T 2
2
=F(F(. . . F(F(x(tk), . . . ,tk), . . . ,tk+1), . . . ,tk+j −2), . . .,tk+ j −1 ).
   + (∇x p V (z 0 , x p,0 ))T (x p − x p,0 ) + V (z 0 , x p,0 )
x(tk+1 )
   (16)
x(tk+j −1)
G 0 (z − z 0 ) ≤ E 0 (x p − x p,0 ) + T0 . (17)
(8)
The different terms are computed as follows and evaluated at
The optimal control problem is now in its generic mp-NLP the linearization point (z 0 , x p,0 ):
form
2
H0  ∇zz V (z 0 , x p,0 )
V ∗ (x(tk ), p(tk )) = minV (x(tk ), U, p(tk ), N) (9) D0  (∇z V (z 0 , x p,0 ))T
U,N
G 0  (∇z G(z 0 , x p,0 ))T
subject to
E 0  −(∇x p G(z 0 , x p,0 ))T
G(x(tk ), U, p(tk ), N) ≤ 0 (10)
T0  −G(z 0 , x p,0 )
1
where p includes the system and controller parameters, which F0  ((∇zx
2
p
V (z 0 , x p,0 ))T + ∇x2p z V (z 0 , x p,0 )). (18)
are considered constant for the duration of the prediction 2
horizon. Two additional vectors are defined: 1) the vector The mp-QP formulation is employed to compute local
of parameters x p (tk ) ∈ Rn p , where n p = n + d, i.e., n p is approximations of the original mp-NLP problem in the
the sum of the number of states n and the number of exploration space. This is represented as a number of
parameters d hyper-rectangles, on which single mp-QP problems are
  solved. Each hyper-rectangle is further partitioned into poly-
x(tk ) hedra, i.e., the critical regions for the mp-QP problem.
x p (tk ) = (11)
p(tk ) Finally, the mp-QP solution is represented as a piecewise
affine function that is continuous across the boundaries
and 2) the vector of decision variables, z ∈ Rs among different polyhedra, but discontinuous across the
  hyper-rectangles.
U In this paper, the mp-QP problems are computed by means
z= . (12)
N of Multi-Parametric Toolbox 3.0 [25]. The solution is eval-
uated in points of interest within each hyper-rectangle and
Based on (11) and (12) it is possible to reformulate the compared with the solution of the NLP problem at the same
optimization problem as points, where the initial state conditions are the coordinates
of the points themselves. The NLPs are computed by means
V ∗ (x p (tk )) = minV (z, x p (tk )) (13) of IPOPT, a software package for nonlinear optimization [26].
z
s.t. G(z, x p (tk )) ≤ 0. (14) Based on the maximum error between the evaluated mp-QP
and computed NLP solutions for all the points, a decision
The minimization is performed with respect to z and is is made whether to subpartition the hyper-rectangle into
parameterized with x p (tk ). smaller ones, or to stop the process and accept the mp-QP
approximating solution. The algorithm in [21] that implements
this concept is summarized. For all the unexplored hyper-
rectangles the following steps are implemented.
B. Off-Line Solution
1) Compute the hyper-rectangle volume. (A minimum
The mp-NLP problem is not solved directly, but through volume is defined to decide whether the hyper-rectangle
its approximation (see [23]). In this paper, an mp-QP for- can be further split.)
mulation is adopted, as suggested in [21] and implemented 2) Compute the NLP solution (or recover it from previous
in [24]. The mp-NLP in (13) and (14) is linearized around steps) at the points of interest.
a predefined point (z 0 , x p,0 ) by means of Taylor series 3) Compute the mp-QP solution on the whole hyper-
expansion (with z 0 being the optimal solution at x p,0 ), such rectangle, using the NLP solution at the Chebyshev
that the cost function is approximated with a quadratic center plus its coordinates, as the linearization point for
function (15)–(16) and the constraints assume a linear the terms in (15)–(18).
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

4) Evaluate the mp-QP solution for all the aforementioned TABLE I

points. I NTERNAL M ODEL PARAMETERS
5) Calculate the maximum error between the
NLP-computed solutions and the mp-QP-evaluated
solutions.
Based on this information each hyper-rectangle is either
stored or marked “to be split” with a heuristic splitting
rule similar to the one in [21]. When all the tolerances are
fulfilled or the minimum allowed volume has been reached,
the algorithm terminates and the solution is available for any
point inside each hyper-rectangle.
With respect to the stability of the resulting controller,
common schemes in the literature for implicit MPC include
stabilizing terminal constraints or terminal costs, which need
to satisfy Lyapunov function-type conditions (see [27], [28]).
Alternatively, Grüne [29], and Reble and Allgöwer [30]
present a stability and performance analysis technique for
unconstrained (with respect to stability preserving constraints)
implicit NMPC schemes. However, all these approaches are
for implicit cases. To the best of the authors’ knowledge,
there is no comparable practical NMPC theory in the literature
addressing the stability for explicit NMPC. Therefore, in this
Fig. 1. Simplified architecture of the implemented TC strategy.
paper the eNMPC stability will be verified empirically through
the simulated scenarios and experimental test results of usually limited (i.e., <100 for this TC application), which
Sections V and VI. makes both methods viable in terms of processing burden and
C. Implementation of the Explicit Solution for searching time.
Real-Time Applications III. T RACTION C ONTROL D ESIGN
Once the solution is computed off-line, the next step is This section discusses the structure and formulation of the
to define the most efficient way to access it on-line. This proposed model predictive TC strategies, first by deriving the
is performed through point location and piecewise control internal model, and then by formulating the optimal control
function evaluation. In particular, the former problem becomes problem. In particular, three internal models with increasing
challenging if the total number of regions composing the final complexity are proposed and used with the same cost function
solution is large (>1000–2000). Two families of methods are and constraints.
available. The values of the main vehicle data used for internal model
1) Sequential search methods, which, in the worst case, parameterization are reported in Table I. They refer to the
may check every region to identify the one containing electric vehicle simulated in Section V.
the considered point and
2) Binary-search-tree methods [31], providing a fast solu- A. Traction Control Structure
tion for the location of the point with a limited number
of mathematical operations, which is logarithmic in the Fig. 1 shows the TC structure. The torque-vectoring con-
number of regions, for a balanced tree. As a drawback, troller of the electric vehicle calculates the total reference
binary-search-tree methods require significant off-line wheel torque and the reference yaw moment. The control
processing, which makes them unsuitable for a large allocation (CA) algorithm outputs the individual wheel torques
number of regions [31]. for the in-wheel motors, indicated as TCA , to achieve the
The specific application (i.e., the 4-D case of Section III-B) references. A state predictor (SP) compensates for the system
has a total number of 85 hyper-rectangles, obtained with the delays on the states, e.g., caused by the CAN bus. The pre-
selected approximation tolerances for the values of the cost dicted parameter vector with the updated states, x̂ p , is provided
function, the normalized solution and the maximum normal- to the core block of the TC, i.e., the on-line implementation
ized constraint violation. A two-layer solution is proposed. of the eNMPC, which outputs the torque correction T , to be
The top layer includes a binary-search-tree to determine the subtracted from TCA .
index of the hyper-rectangle the measured/estimated point
lies in. This information is then passed to the bottom layer, B. 4-D Problem: Internal Model
consisting of functions, one for each hyper-rectangle, which The controlled variable is the wheel slip velocity s
identify the correct critical region within the hyper-rectangle,
s = ωr − V (19)
and evaluate the piecewise control function. In the bottom
layer either binary-search-tree or sequential search methods where ω is the angular wheel speed, r is the rolling radius of
can be used, as the number of polyhedral critical regions is the wheel, and V is the linear speed of the vehicle, so that the
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

TAVERNINI et al.: EXPLICIT NONLINEAR MODEL PREDICTIVE CONTROL FOR ELECTRIC VEHICLE TRACTION CONTROL 5

slip ratio is indicated as u in the remainder. The 4-D eNMPC will be

ωr − V s referred to as eNMPC4 . During the control system design, the
σx = = . (20) individual components of x p and z are normalized through
ωr ωr
division by their maximum expected value.
The time derivative of (19) is given by
d d d
s(t) = r ω(t) − V (t). (21) C. 5-D Problem (a): Internal Model
dt dt dt
The model of Section III-B considers instantaneous genera-
The first term on the right-hand side results from the wheel tion of the longitudinal tire force. In this section, the model is
moment balance enhanced to account for the tire force dynamics, by includ-
d 1 ing the concept of tire relaxation length, σ . A first-order
ω(t) = (TCA −T (t) − Fx r ) (22)
dt Jw differential equation calculates the slip ratio for the MF
where Jw is the wheel mass moment of inertia. TCA is kept in (26) and (28), starting from the wheel speed and vehicle
constant over the prediction horizon, and thus is a system speed. By assuming a linear dependency between the longi-
parameter. Fx is the longitudinal tire force, estimated through tudinal tire force and vertical load the first-order longitudinal
a simplified version of the Pacejka magic formula (MF) [33] tire force dynamics are implemented. The resulting internal
model is described by the differential equations (29)–(32):
Fx = μx Fz (23) 2

d r 1   
μx = D sin(C arctan(Bσx )) (24) s(t) = − − Dsin Carctan Bσxrel (t) Fz
dt Jw m
where Fz is the vertical tire load, considered as a constant, (TCA −T (t))r
and μx is the longitudinal tire force coefficient, with B, C, + (29)
Jw
and D being the MF parameters [33]. The longitudinal vehicle d
dynamics are modeled by considering a mass m equal to a eint (t) = s(t) − σxref ω(t)r (30)
dt
quarter of the total vehicle mass d 1     
ω(t) = TCA −T (t)− Dsin Carctan Bσxrel (t) Fz r
d 1 dt Jw
V (t) = Fx . (25)
dt m (31)

By substituting (22) and (25) into (21) the wheel slip d rel (ω(t)r − s(t)) s(t)
σ (t) = − σxrel (t) . (32)
dynamic equation, i.e., the first equation of the internal model, dt x σ ω(t)r
is obtained In this case, the state vector, input vector, and parameter
2

d r 1 Bs(t) vector are, respectively, x = [s, eint , ω, σxrel ], u = [T ],

s(t) = − − Dsin Carctan Fz
dt Jw m ω(t)r and p = [TCA ]. The problem includes five parameters
(TCA − T (t))r (5-D problem), i.e., x p = [s(tk ), eint (tk ), ω(tk ), σxrel (tk ),
+ . (26) TCA (tk )], and five decision variables, i.e., z = [T (tk ),
Jw
T (tk+1 ), T (tk+2 ), T (tk+3 ), ν(tk )]. The respective expli-
An integral action is incorporated to tackle the steady-state cit controller will be called eNMPC5a in the remainder.
error and model uncertainties. This considers the integral of the
error eint between the actual slip velocity s and the reference
slip velocity computed from the target value σxref of the slip D. 5-D Problem (b): Internal Model
ratio. The respective differential equation, i.e., the second The model of Section III-B considers a constant value of
equation of the internal model, is the vertical tire load. In this section, a more accurate case
d is considered, where the vertical tire load is computed as
eint (t) = s(t) − σxref ω(t)r. (27) a function of the vehicle longitudinal and lateral accelera-
dt
tions. The estimated vertical load value becomes a slowly
By substituting (23) and (24) into (22), the third equation of varying parameter for the control problem, thus increasing its
the internal model is obtained dimension.

d 1 Bs(t) In this case, the equations of the system are exactly

ω (t) = TCA −T (t)− Dsin Carctan Fz r .
dt Jw ω(t)r the same as in Section III-B, but the state vector, input
(28) vector, and parameter vector are, respectively, x = [s, eint , ω],
u = [T ] and p = [TCA , Fz ]. The problem now includes
The model state vector, input vector and parameter vector
five parameters (5-D problem), i.e., x p = [s(tk ), eint (tk ), ω(tk ),
are, respectively, x = [s, eint , ω], u = [T ] and p = [TCA ].
TCA (tk ), Fz (tk )], and five decision variables, i.e., z = [T (tk ),
Unless otherwise specified, in the following analyses the
T (tk+1 ), T (tk+2 ), T (tk+3 ), ν(tk )]. The respective impli-
explicit solution is reported for N p = 4 and ts = 2 ms. The
cit controller will be called NMPC5b in the remainder.
parametric problem includes four parameters (4-D problem),
i.e., x p = [s(tk ), eint (tk ), ω(tk ), TCA (tk )], and five deci-
sion variables, i.e., z = [T (tk ), T (tk+1 ), T (tk+2 ), E. Control Problem Formulation
T (tk+3 ), ν(tk )]. The receding horizon control input that The three internal models of Sections III-B–III-D share the
is applied to the system is u(tk ) = T (tk ), which will be same optimal control problem formulation. The continuous
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

form of the cost function is

 tf
qx1 q x2
V = (s(t) − σxref ω(t)r )2 + 2 eint (t)2
tk w x1 wx2
2
ru rν p x1
+ 2 T (t) + 2 ν(tk ) dt + 2 (s(t f )
2 2
wu wν wx1
p x2
−σxref ω(t f )r )2 + 2 eint (t f )2 (33)
wx2
where qx1, q x2 , ru , rν , px1, and px2 are the weights for the
different terms, and the notations wi indicate scaling factors.
As a consequence, a tracking problem is set for the first state
s and a regulating problem is set for the second state, eint .
The choice of adopting the slip velocity s as state and
tracking variable, rather than the more commonly used slip
ratio σx , finds its motivation in the algorithm for the com-
putation of the explicit solution. In fact, the adoption of σx
would lead to a feedback law that is scaled with the angular Fig. 2. Normalized control action u for the 4-D problem as a function
wheel speed. The higher variability of the feedback control law of x p (1) (normalized wheel slip velocity) and x p (4) (normalized torque
would imply a finer partition of the space, to reach a good demand from the driver).
approximation of the nonlinear problem. Hence, the choice
of different internal models, although equivalent from the
viewpoint of the represented physics, influences the efficiency
of the generation of the explicit solution. Careful consideration
of this aspect in the design phase leads to a reduction of
the off-line computational burden and the on-line memory
requirement.
The minimization of (33) is subject to state and input bound
constraints

smin − ν ≤ s ≤ smax + ν (34)

0 ≤ T ≤ TCA . (35)

IV. E NMPC-BASED TC I MPLEMENTATION

An advantage of eNMPC with respect to implicit NMPC is
the availability of the feedback control law beforehand. This
allows the analysis of the control action for any value of the
vector of parameters.
The solution of the eNMPC4 , i.e., the 4-D eNMPC Fig. 3. Effect of x p (2) and x p (3) on u for the 4-D problem. (a) Negative
(see Section III-B), is presented in Fig. 2. To plot the 3-D value of x p (2). (b) Positive value of x p (2). (c) Low value of x p (3). (d) High
surface in Fig. 2, two parameters have been fixed, i.e., the value of x p (3).
normalized integral of the wheel slip error, x p (2), which is
set to zero, and the normalized wheel angular velocity, x p (3),
which is set to 0.85. The red line “reference” indicates the The analysis of the control action shows that no regulation is
wheel slip velocity corresponding to the reference slip ratio applied until the reference slip is reached, if the normalized
for the specific x p (3). x p (4) is the normalized torque demand torque demand is small. On the other hand, for high values
from the CA. of x p (4), a regulation is prescribed even before reaching
The solution essentially consists of three planes: the reference, based on the prediction available to the con-
1) a plateau of zero control action for low values of slip troller. Beyond the reference a regulation that is below the
velocity, indicated as “input lower constraint” in Fig. 2. maximum possible value is applied for the whole range of
According to (35), the TC torque correction must be torque demands, as long as the slip velocity is lower than
positive; a specific nonconstant value (see the surface “non-saturated
2) an inclined plane, parallel to the x p (1)-axis, indicated feedback law”). Above this value the regulating control action
as “input upper constraint” in Fig. 2, which expresses is equal to the torque demand, i.e., u = x p (4).
that, according to (35), the regulating torque cannot be The effect of the normalized integral of the slip velocity
larger than the torque demand; error, x p (2), is presented in Fig. 3(a) and (b), corresponding
3) another inclined plane, i.e., the “non-saturated feedback to a negative value and a positive value of x p (2), respectively.
law,” which is saturated by the previous two. The whole surface of the feedback law shifts along the
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

TAVERNINI et al.: EXPLICIT NONLINEAR MODEL PREDICTIVE CONTROL FOR ELECTRIC VEHICLE TRACTION CONTROL 7

TABLE II
M AIN PARAMETERS OF THE S IMULATION M ODEL

V. S IMULATION R ESULTS
A. Test Scenario and Evaluation Metrics
Fig. 4. Normalized control action with the corresponding region indication.
x p (2) and x p (3) have been fixed. The simulation analysis was carried out with a high fidelity
vehicle simulation model implemented with the software
IPG CarMaker. The vehicle data (see Table II) are those
x p (1)-axis, while the reference does not move. This acts of an electric quadricycle prototype with a front-wheel-drive
as a compensation for the initial positive or negative value topology, based on two in-wheel motors (direct drive) with
of x p (2). Fig. 3(c) and (d) shows the variation of the feedback a peak torque of 500 Nm each. Given the low mass of the
law with the normalized wheel speed, x p (3). Although the vehicle, the available torque is sufficient to provoke front
shape of the surface does not change, it translates with the wheel spinning even in high tire-road friction conditions.
reference slip velocity along the x p (1)-axis. Fig. 4 shows The tire model is the MF (ver. 5.2), and includes the
that the piecewise affine feedback law is actually evaluated variation of the longitudinal and lateral relaxation lengths as
from a number of different regions of the parametric problem, functions of the vertical load. The electric motor dynamics
i.e., hyper-rectangles and polyhedral critical regions, despite are modeled through a first order transfer function and a
the control action mainly consists of only three planes. The pure time delay. A pure time delay is also considered on the
analysis of Figs. 2–4 suggests that the whole feedback law controller output to model the CAN bus [32]. Unless otherwise
could be realized as a ruled-based strategy that defines dif- specified, in the remainder the implementation time step of the
ferent planes intersections and translations, given the input controllers, t S,I , is of 2 ms.
measurements from the plant. Alternatively, a rigorous method The considered acceleration test scenario is based on a
for the reduction of the memory requirements of explicit model straight road with varying tire-road friction coefficient, μ. The
predictive controllers is presented in [34]. values of μ are modified in steps, according to the sequence
During the implementation phase of the eNMPC, as shown 0.9–0.15–0.9–0.45–0.9. This provides a real challenge to the
in Fig. 1, a specific strategy was applied for the compensation TC, which has to regulate the slip ratio to a constant reference
of δm and δCAN , i.e., the pure time delays associated with value of 0.10, while the vehicle is accelerating from an initial
the electric motor drive and the CAN bus, respectively. The speed of 5 km/h, at which a fast torque demand ramp up to
adopted technique is based on the concept used in [16] for the drivetrain peak torque is imposed.
a hybrid explicit MPC implementation of a TC. A state To objectively assess the TC performance, a set of perfor-
predictor, employing the same model formulation described mance indicators is identified based on [20].
in Section III-B, and a buffer, containing part of the past 1) The root-mean square value of the slip ratio error, i.e., a
control history, are used to predict the trajectory of the input tracking performance indicator
parameters to the eNMPC, for a horizon length of δm + δCAN .
Thus, the inputs to the controller are projected into the future,  te
1
and the control action is computed based on this prediction. RMSE = (σx (t) − σxref )2 dt (36)
te − ti ti
The solution of the eNMPC4 was tested on a dSPACE
MicroAutobox II (900 MHz, 16 MB) rapid control prototyping where σx (t) is the actual value of the slip ratio during
unit. An exploration of the parameter space was performed the relevant part of the test, defined by the initial and
to assess the computational time for a fine and comprehen- final times ti and te .
sive grid of possible inputs. The computational time for the 2) The final value of vehicle velocity, V f , i.e., an acceler-
combination of the two function evaluation layers was in ation performance indicator.
the range of ∼5–25 μs. These values are very low com- 3) The normalized integral of the absolute value of the
pared to the implementation time step of 2 ms, which is control action, which gives an indication of the required
not achievable with more conventional implicit NMPC tech- control effort
nology on the same hardware. Hence, the eNMPC can run  te
in real-time at any frequency within the range typical of 1
IACA = |T (t)| dt. (37)
TC applications. te − ti ti
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

Fig. 5. NMPC4 and eNMPC4 comparison: actual and reference slip ratios Fig. 7. PI and eNMPC4 comparison: actual and reference slip ratios of the
of the front left wheel. front left wheel.

regions are used in the simulated complex scenario. More-

over, the crossings of different hyper-rectangle boundaries,
which imply discontinuities in the solution, do not bring
any significant degradation of the explicit feedback control
action.

C. eNMPC4 and Proportional Integral (PI) Controller

The results of the eNMPC4 are compared with those
obtained through a simple yet effective PI-based TC system,
with gain scheduling on vehicle speed and including anti-
windup features on its integral contribution.
A frequency response-based initial design of the PI gains
was performed with a linearized plant model for different
vehicle speeds. This was followed by an empirical fine tuning
through simulations in the time domain with the CarMaker
Fig. 6. eNMPC4 : hyper-rectangle index (top) and polyhedral critical region model. The gains obtained with this process were finally
index (bottom) with the vertical lines indicating the hyper-rectangle switching reassessed by employing the linearized plant to verify gain and
times. Each hyper-rectangle has an independent numbering of its polyhedra. phase margins, as well as the sensitivity and complementary
sensitivity functions.
B. eNMPC4 Benchmarking The comparison of the controller results in terms of slip
To prove the effectiveness of the local quadratic approxima- ratios is reported in Fig. 7. For both the PI and the eNMPC4 the
tions of the multi-parametric nonlinear problem, the simulation TC is activated in proximity of the reference slip value, σxref .
results for the described scenario are reported in Fig. 5, with an The response of the two controllers to the initial wheel
overlap between the eNMPC4 solution and the corresponding torque demand application presents visible differences. The
implicit one. The implicit strategy for the 4-D case (NMPC4 ) PI overshoots σxref , and then reaches the desired value with
is implemented by solving on-line the same nonlinear optimal a damped oscillatory response. The eNMPC4 presents an
control problem with the same solver, IPOPT, employed for initial undershoot caused by the controller activation and the
the generation of the explicit solution. The implicit strategy, discrepancy between the tire-road friction coefficients of the
which is not real-time capable, represents the optimal solution, plant and the internal model. This is promptly recovered
because of the absence of the local quadratic approximations. by the integral action. Afterwards, the eNMPC4 approaches
Fig. 5 shows that the solutions of the NMPC4 and eNMPC4 the reference more gently, with a lower overshoot and less
are indistinguishable. As this is confirmed by all the sim- oscillations. The reason for this behavior is that the design of
ulations that were performed during the study, the level of the eNMPC4 TC is based on tire characteristics for μ = 0.45.
sub-optimality of the eNMPC4 implementation is considered Hence, when the controller operates in higher tire-road friction
satisfactory. conditions (e.g., at μ = 0.9), it tends to be conservative.
Fig. 6 reports the index of the hyper-rectangles that are Nevertheless, σxref is reached at approximately the same
used by the eNMPC4 in the considered scenario, and the time as in the PI case. The transition between μ = 0.9
index of the polyhedral critical regions that are employed and μ = 0.15 is very demanding for the controllers. The
within each hyper-rectangle. The figure reveals that only a few PI responds with an overshoot that is maintained until the slip
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

TAVERNINI et al.: EXPLICIT NONLINEAR MODEL PREDICTIVE CONTROL FOR ELECTRIC VEHICLE TRACTION CONTROL 9

Fig. 8. PI and eNMPC4 comparison: torques before and after the front left
TC block. Fig. 10. PI and eNMPC4 comparison: longitudinal acceleration of the vehicle.

TABLE III
S UMMARY OF P ERFORMANCE I NDICATORS AND R ESPECTIVE VARIATIONS

Fig. 9. PI and eNMPC4 comparison: front left wheel speed multiplied by

the wheel radius, VFL , and vehicle velocity, Vvhl .

ratio reaches σxref . For the eNMPC4 the overshoot presents

a smaller peak and a faster response leading to the reference.
This is followed by a promptly recovered undershoot. The next
difficult transition is the one that leads back to μ = 0.9. In this severity of the selected scenario. The longitudinal acceleration
case, both controllers present undershoots followed by a few does not significantly differ among the two controllers.
oscillations with similar duration. The oscillations have higher Table III reports the objective performance indicators
amplitudes for the PI. In the final μ-transitions the overshoots defined in Section V-A for Cases i–vi.
and undershoots are relatively small and of similar magnitude Case i: The PI TC running at t S,I = 2 ms. During the
for the two controllers, although slightly higher for the PI, implementation phase of the controller it was veri-
which also exhibits a slower response. fied that a further reduction of t S,I within reasonable
Fig. 8 shows the wheel torques before and after the limits would not have brought substantial benefits.
TC block. Similarly to the slip ratios, the time histories high- Case ii: The PI TC running at 4 and 8 ms, with the same
light the marginally faster response of the eNMPC4 , together gains as for Case i, apart from the anti-windup
with the more quickly damped oscillations of its control gain. The variation of the anti-windup gain was
action. Fig. 9 shows the angular speed of the front left wheel, necessary to provide control system stability in the
multiplied by the wheel rolling radius, and the vehicle speed. selected test, especially immediately after the first
The time histories of the longitudinal vehicle acceleration are μ-transition.
reported in Fig. 10. The wide range of values, i.e., from Case iii: The PI TC running at 4 and 8 ms with optimized
∼0 to ∼4.5 m/s2 during the relevant part of the test, together gains for those time steps. The PI gain optimization
with their abrupt variations, confirms the high level of was based on CarMaker simulations of the selected
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

maneuver, and was aimed at the minimization of the

slip ratio RMSE.
Case iv: The eNMPC4 TC running at 2 ms.
Case v: The eNMPC4 TC running at 4 and 8 ms, with
the same weights of the cost function, the same
discretization interval ts of the internal model and
the same prediction horizon t p as for Case iv. In the
4-ms subcase, in the eNMPC4 off-line process it
is imposed T (tk ) = T (tk+1 ) and T (tk+2 ) =
T (tk+3 ), while in the 8-ms subcase it is imposed
T (tk ) = T (tk+1 ) = T (tk+2 ) = T (tk+3 ).
Case vi: The eNMPC4 TC running at 4 and 8 ms, with a fine-
tuning of the weights of its cost function. Similarly
to Case iii, the eNMPC4 tuning process consisted
of CarMaker model simulations and iterative com-
putations of the slip ratio RMSE.
The comparison between Case i and Case iv shows a 9.2% Fig. 11. eNMPC4 and eNMPC5a comparison: reference and actual slip
ratios.
reduction of the RMSE for the eNMPC4 TC compared to
the PI TC, together with a negligible increment on the final is of ∼10 ms. In particular, the eNMPC4 TCs at 4 and 8 ms,
velocity and IACA. Both the PI TC and eNMPC4 TC are respectively, provide similar and worse results than the PI TC
subject to a significant decay of the respective tracking per- at 2 ms, which means that the appropriate selection of
formance, when they are implemented at 4 and 8 ms without t S,I should have higher priority in the TC design process
modifying their design with respect to the cases running with respect to the control structure selection, at least for
at 2 ms. In particular, the RMSE increase is of 15.4% and electric vehicles with very responsive in-wheel motors such
185.1% for the PI controller, while it is of 10.1% and 132.4% as that of this paper. Nonlinear MPC technology can be used
for the eNMPC4 . If the PI TC and eNMPC4 TC are retuned to enhance the TC performance, but, this is actually beneficial
for the time steps of 4 and 8 ms, the performance decay is still only if the NMPC is run at 2 ms. In such a condition the
significant, i.e., it amounts to 8% and 61.7% for the PI, and NMPC provides better results than the PI controller, that can
8.6% and 78.6% for the eNMPC4 . It is possible to observe be easily implemented with a very low time step. However,
the following. with the available computing hardware for automotive appli-
1) For the specific application significant retuning of the cations, an implicit NMPC does not currently run at 2 ms, and
controller is needed when changing the time step, which possibly not even safely at 4 ms, according to the literature
is an important outcome, not reported in the existing mentioned in Section I. This makes the implementation of the
TC literature to the knowledge of the authors; and eNMPC4 , rather than an implicit NMPC4 , necessary and ben-
2) The performance decay induced by the increase of t S,I eficial to achieve the potential vehicle performance benefits.
is relatively similar for the two control structures.
These results can be justified through the analysis of the D. Effect of Tire Force Dynamics Modeling
linearized model of the plant without TC, including consid-
eration of tire relaxation. The linearization was carried out This section evaluates the effect of considering the lon-
in proximity of the reference slip ratio. At a vehicle speed gitudinal tire force dynamics in the internal model for
of 2.5 m/s the slip ratio response to a motor torque step NMPC design. The simulation results for the eNMPC5a TC,
input has a rise time, Tr , of only ∼5 ms, which becomes derived from the internal model of Section III-C, are reported
∼11 and ∼26 ms, respectively, at 5 and 10 m/s. The very in Fig. 11 for the considered μ-varying scenario. The addition
fast response is related to the in-wheel layout of the spe- of the relaxation length does not bring any benefit in terms of
cific electric drivetrains. Based on the indications in [35], tracking performance. The reason is related to the relative fast
the implementation time step should range from 6% to 40% dynamics of the longitudinal tire force generation, especially,
of Tr . For the average speed of the simulated scenario, for higher vehicle velocities and a flat road surface. The
i.e., ∼5 m/s, this implies a recommended range of t S,I eNMPC5a implementation shows that a 5-D problem can also
from 0.7 to 4.4 ms. At the initial speed of the simulated be managed with this control methodology.
tests the recommended time step would be even significantly
lower. Therefore, the system rise time values are consistent E. Effect of Time-Varying Vertical Load Modeling
with the TC performance degradation for t S,I = 4 ms This section studies the effect of including the variable
and t S,I = 8 ms, where the latter is nearly twice the maximum vertical tire load in the internal model of the NMPC
recommended time step at 5 m/s. (see Section III-D). Since it has been proven that the generated
In summary, a low value of the implementation time step at explicit solution for the eNMPC4 shows no visible difference
which the TC is run guarantees a significant enhancement of from its corresponding implicit solution, i.e., the NMPC4 ,
the results, independently of the selected controller. It should the comparison for this particular internal modeling feature
be noted that in many practical TC applications the time step will be carried out only through the implicit strategy.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

TAVERNINI et al.: EXPLICIT NONLINEAR MODEL PREDICTIVE CONTROL FOR ELECTRIC VEHICLE TRACTION CONTROL 11

TABLE IV
ROBUSTNESS A SSESSMENT: V EHICLE PARAMETERS
VARIATION E FFECT ON T RACKING P ERFORMANCE

Fig. 12. NMPC4 and NMPC5b comparison: actual and reference slip ratios
of the front left wheel.

Fig. 12 shows the results of this comparison along the

simulated scenario. The performance of the two controllers
is very similar. In the first part of the scenario, when μ = 0.9,
the NMPC4 shows a slightly better response. In the rest
of the test the NMPC5b provides better tracking. Overall,
the difference is very limited, and it amounts to less than 0.5%
in terms of RMSE. It can be concluded that, in this application,
to increase the dimension of the problem by introducing a TABLE V
time-varying vertical load does not provide any major benefit ROBUSTNESS A SSESSMENT: N OISE I NJECTION
with respect to the 4-D problem with constant load. E FFECT ON T RACKING P ERFORMANCE
Future research will focus on the evaluation of alternative
selections of the fifth parameter of the controller. For example,
additional parameters could include a time-varying σxref , to
improve the lateral tire force capability, as shown in [36], or to
provide better performance when starting from standstill.

F. Robustness Assessment
The robustness against the variation of the tire-road friction
coefficient μ has already been assessed. In this section, further
simulations are performed with the eNMPC4 and the PI, with
t S,I = 2 ms. tive for a wide range of wheel characteristics. Finally, also
Three vehicle parameters have been identified to have a when K x is varied by +/−20% to consider different tire
potentially relevant effect on control system performance, properties, the RMSE variation is limited, and it amounts
namely: 1) the total vehicle mass, M; 2) the wheel mass to +4.7% and −6.2% for Cases xv and xvi (eNMPC4 ), and
moment of inertia, J ; and 3) the longitudinal slip stiffness of to +5.2% and −5.6% for Cases xvii and xviii (PI). In conclu-
the tires, K x . The results in terms of RMSE and corresponding sion, both controllers are robust for the considered reasonable
percentage variation with respect to the baseline condition of range of plant parameter variations, with a limited advantage
the controllers are reported in Table IV. of the eNMPC4 over the PI.
For a +/−15% variation of M, the results show that the Another aspect of control system robustness is the noise
RMSE increase/decrease for the eNMPC4 (Cases vii and viii) rejection performance. The sensor noise resulting from a real
is confined to +5.1% and −5.4%. The same applies to vehicle prototype test, presented later on in this paper, was
cases ix and x, i.e., to the PI TC, with +6.1% and −5.6%. analyzed. Gaussian white noise with different initial seeds is
Hence, the addition of a passenger or payload does not added to the simulated wheel speeds of each corner. These are
significantly affect the TC tracking performance. When the main input signals of the controller, which are used to com-
a +/−30% variation of J is imposed, the eNMPC4 pute the slip ratio. The results are reported in Table V, in terms
(Cases xi and xii) and the PI (Cases xiii and xiv) present of RMSE variation and maximum slip ratio throughout
the same very marginal performance degradation (i.e., by 0.4% the scenario. The comparison is made with respect to the same
and 2.5%). This means that the controllers will be effec- controllers without the noise injection.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

12 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

Fig. 14. Experimental tests: comparison of actual and reference slip ratios
for the vehicle with the eNMPC4 and the passive vehicle (TC off).

Fig. 13. Electric vehicle prototype during the TC and passive vehicle
experimental test session on the low-μ metal plates. The vehicle skids laterally
when the TC is deactivated (bottom).

In the Case xix, the eNMPC4 is still able to follow the

reference throughout different μ variations. The RMSE
increase is of 142.5%, and is mainly caused by oscillations
around the reference. The peak values of slip ratio remain
similar to the case without noise injection, with a maximum
increase of 28.3%. The PI presents a very different situation.
The controller is no longer able to follow the reference closely
in all friction conditions. This is evident from the 290.5%
RMSE increase, and the 92.3% increase of the maximum value
of σx . Although the PI controller is still able to eventually
recover the tracking of the reference slip ratio, the eNMPC4
presents much better noise rejection characteristics. It must Fig. 15. Experimental tests: comparison of vehicle speed (Vvhl ) and front
left angular wheel speed multiplied by the wheel radius (VFL ) for the vehicle
be noted that these results were obtained without retuning of with the eNMPC4 and the passive vehicle (TC off).
the controllers. This operation is recommended for obtaining
desirable performance in case of noisy signals.
constraints, to take into account the greater vehicle mass and
lower motor torque capability, with respect to the simulated
VI. E XPERIMENTAL R ESULTS scenarios.
An experimental testing session was conducted with the The slip ratio time histories for the vehicle with the
eNMPC4 TC on the electric quadricycle prototype of the eNMPC4 and the passive vehicle, i.e., the vehicle with deac-
European H2020 SilverStream project. The vehicle has a tivated TC, are presented in Fig. 14. In the passive vehicle σx
mass of 640 kg (driver excluded), and is equipped with four reaches values of almost 0.9. This affects the duration of the
in-wheel motors with a peak power of 4.2 kW and a peak maneuver, since the lateral force capability of the front tires
torque of 115 Nm each. The prototype is shown in Fig. 13. is drastically reduced, because of the coupling effect between
The tests were conducted in front-wheel-drive mode, on a longitudinal and lateral tire forces. Hence, the driver was not
series of smooth steel plates, which were lubricated to fur- able to drive the vehicle in a straight line. For the eNMPC4 ,
ther decrease the friction coefficient. This is estimated to after a first peak of 0.25, σx goes back to the reference value
be ∼0.09–0.10. Similarly to the simulation scenarios, the of 0.10 in the following 0.2 s. The good tracking performance
vehicle was driven on the metal plates at speeds of 5–7 km/h continues for the duration of the test with limited oscillations
and, then, the driver suddenly pressed the accelerator pedal to around the reference. Faster response and closer tracking were
demand the maximum available torque from the front in-wheel obtained with a different eNMPC4 tuning, at the expense of
motors. The pedal position was maintained until the end of the increased motor torque oscillations.
metal surface was reached. The eNMPC4 with t S,I = 4 ms Fig. 15 confirms the criticality of the friction conditions,
was updated in terms of internal model parameters and input with the front left tire of the passive vehicle that spins up
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

TAVERNINI et al.: EXPLICIT NONLINEAR MODEL PREDICTIVE CONTROL FOR ELECTRIC VEHICLE TRACTION CONTROL 13

2) The presented explicit nonlinear MPC implementa-

tions are characterized by on-line computational times
in the range of 5–25 μs on the adopted dSPACE
MicroAutoBox rapid control prototyping unit. This
means that the strategies could be potentially imple-
mented at any reasonable frequency for automotive
TC applications. On the contrary, based on the literature
it would not be possible to run an equivalent implicit
NMPC at the required time step of 2 ms.
3) The nonlinear model predictive controller allows a 9.2%
tracking performance improvement with respect to a PI
controller during the variable tire-road friction scenario,
simulated with a high fidelity vehicle model.
4) The local multi-parametric quadratic approximation of
the nonlinear problem, for the selected explicit nonlinear
model predictive control method, does not bring perceiv-
Fig. 16. Experimental tests: comparison of motor torque demand before able performance differences with respect to the corre-
(TCA ) and after the front left TC block for the vehicle with the eNMPC4 . sponding implicit nonlinear model predictive controller.
5) The consideration of tire force dynamics and vertical
load transfers in the internal model for MPC system
compared to the rear wheels, which provide the estimated design has negligible effects on the TC performance
vehicle speed. The vehicle velocity profiles with and without during the simulated scenario.
TC present similar trends. In fact, regardless of the considered 6) The interpretation of the explicit nonlinear model predic-
road surface, when the slip ratio moves beyond the peak tive control law provides useful information on the effect
of longitudinal tire force, the Fx reduction is limited and of different input parameters on the control action. The
the vehicle acceleration is not substantially affected. In these piecewise affine control law can be approximated with
conditions, the most important effect is the loss of lateral tire only three planes.
force capability, caused by the tire force coupling effect [37], 7) An explicit nonlinear model predictive control strategy
which makes the passive vehicle skid laterally, and go outside for TC has been successfully implemented on a fully
the metal stripes [see Fig. 13 (bottom)]. electric prototype vehicle and presented in the literature
Fig. 16 shows the electric motor torque regulation, with for the first time, to the best of the authors’ knowledge.
respect to the torque demand from the driver. The reduced Future developments of the research will evaluate as
torque settles at a value of ∼50 Nm, compared to the driver follows.
demand of 115 Nm, resulting in a ∼56% torque reduction. 1) The increase of the number of parameters of the explicit
The torque oscillations, also caused by the nonconstant tire nonlinear model predictive control problem, and the
friction properties along the metal stripes, are reasonable for implications in terms of memory requirements and per-
the specific implementation and the extreme testing conditions. formance benefits.
Lower peak-to-peak oscillatory responses were obtained for 2) The possibility of simpler strategies able to replicate a
higher tire-road friction levels during the experimental testing similar control pattern with reduced memory require-
session. ments for the on-line implementation of the controller.

VII. C ONCLUSION R EFERENCES

This paper presented traction controllers for electric vehicles [1] H. Fujimoto, J. Amada, and K. Maeda, “Review of traction and braking
with in-wheel motors, based on explicit nonlinear model pre- control for electric vehicle,” in Proc. IEEE Vehicle Power Propuls. Conf.,
Oct. 2012, pp. 1292–1299.
dictive control of the wheel slip velocity. These were compared [2] S. Murata, “Innovation by in-wheel-motor drive unit,” Vehicle Syst. Dyn.,
with more conventional TC strategies based on PI control. The vol. 50, no. 6, pp. 807–830, 2012.
novel conclusions are as follows. [3] V. Ivanov, D. Savitski, K. Augsburg, and P. Barber, “Electric vehi-
cles with individually controlled on-board motors: Revisiting the ABS
1) The implementation time step of the TC has a more design,” in Proc. IEEE Int. Conf. Mechatronics (ICM), Mar. 2015,
significant impact on the control system performance pp. 323–328.
than the selection of the control system technology. [4] M. Amodeo, A. Ferrara, R. Terzaghi, and C. Vecchio, “Wheel slip
control via second-order sliding-mode generation,” IEEE Trans. Intell.
Employing nonlinear MPC is not enough to provide Transp. Syst., vol. 11, no. 1, pp. 122–131, Mar. 2010.
better performance than that of a PI running at an appro- [5] D. Yin, S. Oh, and Y. Hori, “A novel traction control for EV based on
priate time step. To achieve a performance enhancement, maximum transmissible torque estimation,” IEEE Trans. Ind. Electron.,
vol. 56, no. 6, pp. 2086–2094, Jun. 2009.
for the case study TC application, time steps of ∼2 ms [6] T. A. Johansen, I. Petersen, J. Kalkkuhl, and J. Ludemann, “Gain-
are recommended, rather than of 4 or 8 ms. Both for scheduled wheel slip control in automotive brake systems,” IEEE Trans.
the PI TC and nonlinear MPC TC, the control system Control Syst. Technol., vol. 11, no. 6, pp. 799–811, Nov. 2003.
[7] M. Boisvert, P. Micheau, and J. Nadeau, “Nonlinear LQG slip controller
parameters have to be fine-tuned through tests in the based on an empirical model for a three wheel hybrid vehicle,” in Proc.
time domain for the selected time step. IEEE Veh. Power Propuls. Conf. (VPPC), 2014, pp. 1–5.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

14 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY

[8] S. Anwar and B. Ashrafi, “A predictive control algorithm for an anti-lock [33] H. B. Pacejka, Tire and Vehicle Dynamics. Oxford, U.K.: Butterworth,
braking system,” SAE Tech. Paper 2002-01-0302, 2002. 2012.
[9] S. Anwar, “Brake-based vehicle traction control via generalized predic- [34] T. Geyer, F. D. Torrisi, and M. Morari, “Optimal complexity reduc-
tive algorithm,” SAE Tech. Paper 2003-01-0323, 2003. tion of polyhedral piecewise affine systems,” Automatica, vol. 44,
[10] R. de Castro, R. E. Araujo, M. Tanelli, S. Savaresi, and D. Freitas, pp. 1728–1740, Jul. 2008.
“Torque blending and wheel slip control in EVs with in-wheel motors,” [35] M. S. Santina and A. R. Stubberud, Basics of Sampling and Quantiza-
Vehicle Syst. Dyn., Int. J. Vehicle Mech. Mobility, vol. 50, no. 1, tion. Handbook of Networked and Embedded Control Systems. Boston,
pp. 71–94, Jul. 2012. MA, USA: Birkhäuser, 2005.
[11] C. Satzger, R. de Castro, A. Knoblach, and J. Brembeck, “Design and [36] J. H. Park and C. Y. Kim, “Wheel slip control in traction control system
validation of an MPC-based torque blending and wheel slip control for vehicle stability,” Veh. Syst. Dyn., vol. 31, no. 4, pp. 263–278, 1999.
strategy,” in Proc. IEEE Intell. Veh. Symp., Jun. 2016, pp. 514–520. [37] R. W. Allen, T. J. Rosenthal, and J. P. Chrstos, “A vehicle dynamics tire
[12] C. Satzger and R. de Castro, “Combined wheel-slip control and torque model for both pavement and off-road conditions,” SAE Tech. Paper
blending using MPC,” in Proc. Int. Conf. Connected Veh. Expo (ICCVE), 970559, 1997.
Nov. 2014, pp. 618–624.
[13] C. Satzger and R. de Castro, “Predictive brake control for electric
vehicles,” IEEE Trans. Veh. Technol., vol. 67, no. 2, pp. 977–990,
Feb. 2018.
[14] D. Yoo and L. Wang, “Model based wheel slip control via constrained
optimal algorithm,” in Proc. Int. Conf. Control Appl., Oct. 2007, Davide Tavernini received the M.Sc. degree in
pp. 1239–1246. mechanical engineering and the Ph.D. in dynam-
[15] L. Yuan, H. Zhao, H. Chen, and B. Ren, “Nonlinear MPC-based slip ics and design of mechanical systems from the
control for electric vehicles with vehicle safety constraints,” Mechatron- University of Padova, Padua, Italy, in 2010 and
ics, vol. 38, pp. 1–15, Sep. 2016. 2014, respectively. During his Ph.D. he was with
[16] F. Borrelli, A. Bemporad, M. Fodor, and D. Hrovat, “An MPC/hybrid the Motorcycle Dynamics Research Group.
system approach to traction control,” IEEE Trans. Control Syst. Technol., He is currently a Lecturer in advanced vehicle
vol. 14, no. 3, pp. 541–552, May 2006. engineering with the University of Surrey, Guildford,
[17] M. S. Basrah, E. Siampis, E. Velenis, D. Cao, and S. Longo, “Wheel slip U.K. His current research interests include vehicle
control with torque blending using linear and nonlinear model predictive dynamics modeling and control, mostly applied to
control,” Vehicle Syst. Dyn., vol. 55, no. 11, pp. 1665–1685, 2017. electric and hybrid vehicles.
[18] F. Bottiglione, A. Sorniotti, and L. Shead, “The effect of half-shaft
torsion dynamics on the performance of a traction control system for
electric vehicles,” Proc. Inst. Mech. Eng. D, J. Automobile Eng., vol. 226,
no. 9, pp. 1145–1159, 2012.
[19] V. Ivanov, D. Savitski, and B. Shyrokau, “A survey of traction control
and antilock braking systems of full electric vehicles with individually Mathias Metzler (GS’17) received the Dipl.-Ing.
controlled electric motors,” IEEE Trans. Veh. Technol., vol. 64, no. 9, degree (summa cum laude) in mechanical engi-
pp. 3878–3896, Sep. 2014. neering from the Vienna University of Technology,
[20] S. De Pinto, C. Chatzikomis, A. Sorniotti, and G. Mantriota, “Compar- Vienna, Austria, in 2015. He is currently pursuing
ison of traction controllers for electric vehicles with on-board drive- the Ph.D. degree in advanced vehicle engineering
trains,” IEEE Trans. Veh. Technol., vol. 66, no. 8, pp. 6715–6727, with the University of Surrey, Guildford, U.K.
Aug. 2017. His current research interests include vehicle
[21] J. A. Grancharova and T. A. Johansen, Explicit Nonlinear Model dynamics control, model predictive control, opti-
Predictive Control: Theory and Applications, vol. 429. Berlin, Germany: mization, and nonlinear systems.
Springer, 2012.
[22] T. A. Johansen, “On multi-parametric nonlinear programming and
explicit nonlinear model predictive control,” in Proc. 41st IEEE Conf.
Decis. Control., vol. 3, Dec. 2002, pp. 2768–2773.
[23] L. F. Domínguez and E. N. Pistikopoulos, “Recent advances in explicit
multiparametric nonlinear model predictive control,” Ind. Eng. Chem.
Res., vol. 50, no. 2, pp. 609–619, 2011.
[24] P. Tøndel and T. A. Johansen, “Lateral vehicle stabilization using Patrick Gruber received the M.Sc. degree in motor-
constrained nonlinear control,” in Proc. Eur. Control Conf., no. 1, sport engineering and management from Cranfield
Sep. 2003, pp. 1887–1892. University, Cranfield, U.K., in 2005, and the Ph.D.
[25] M. Herceg, M. Kvasnica, C. N. Jones, and M. Morari, “Multi-parametric degree in mechanical engineering from the Univer-
toolbox 3.0,” in Proc. Eur. Control Conf., 2013, pp. 502–510. sity of Surrey, Guildford, U.K., in 2009.
[26] A. Wächter and L. T. Biegler, “On the implementation of a primal- He is currently a Senior Lecturer in advanced
dual interior point filter line search algorithm for large-scale nonlinear vehicle systems engineering with the University of
programming,” Math. Program., vol. 106, no. 1, pp. 25–57, Mar. 2006. Surrey. His current research interests include vehicle
[27] H. Chen and F. Allgöwer, “A quasi-infinite horizon nonlinear model dynamics and tire dynamics with special focus on
predictive control scheme with guaranteed stability,” Automatica, vol. 34, friction behavior.
no. 10, pp. 1205–1217, 1998.
[28] D. Q. Mayne, J. B. Rawlings, C. V. Rao, and P. O. M. Scokaert, “Con-
strained model predictive control: Stability and optimality,” Automatica,
vol. 36, no. 6, pp. 789–814, 2000.
[29] L. Grüne, “Analysis and design of unconstrained nonlinear MPC
schemes for finite and infinite dimensional systems,” SIAM J. Control
Optim., vol. 48, no. 2, pp. 1206–1228, 2009. Aldo Sorniotti (M’12) received the M.Sc. degree in
[30] M. Reble and F. Allgöwer, “Unconstrained model predictive control mechanical engineering and Ph.D. degree in applied
and suboptimality estimates for nonlinear continuous-time systems,” mechanics from the Politecnico di Torino, Turin,
Automatica, vol. 48, no. 8, pp. 1812–1817, 2012. Italy, in 2001 and 2005, respectively.
[31] P. Tøndel, T. A. Johansen, and A. Bemporad, “Evaluation of piecewise He is currently a Professor in advanced vehicle
affine control via binary search tree,” Automatica, vol. 39, no. 5, engineering with the University of Surrey, Guildford,
pp. 945–950, 2003. U.K., where he coordinates the Centre for Auto-
[32] T. Goggia et al., “Integral sliding mode for the torque-vectoring control motive Engineering. His current research interests
of fully electric vehicles: Theoretical design and experimental assess- include vehicle dynamics control and transmission
ment,” IEEE Trans. Veh. Technol., vol. 64, no. 5, pp. 1701–1715, systems for electric and hybrid vehicles.
May 2015.