1

Continuous Control Set Nonlinear Model Predictive

Control of Reluctance Synchronous Machines
Andrea Zanelli, Julian Kullick, Hisham Eldeeb, Gianluca Frison, Christoph Hackl, Moritz Diehl

Abstract—In this paper we describe the design and imple- measurements to be carried out online.
mentation of a current controller for a reluctance synchronous Regarding the control of RSMs, two main concepts have been
machine based on continuous set nonlinear model predictive pursued in the past: (i) Direct Torque Control (DTC) [4, 28]
control. A simplified experimentally identified grey box model
arXiv:1910.10681v1 [eess.SY] 23 Oct 2019

of the flux linkage map is employed in a tracking formulation and (ii) field-oriented control (FOC) [3, 31, 46]. While DTC
which is implemented using the high-performance framework for is known for its robustness and fast dynamics [5], it produces
nonlinear model predictive control acados. The resulting con- a high current distortion leading to torque ripples [6]. In
troller is validated in simulation and deployed on a dSPACE real- contrast, vector control improves the torque response [42] and
time system connected with a physical reluctance synchronous the efficiency of the system [26], but good knowledge of the
machine. Experimental results are presented where the proposed
implementation can reach sampling times in the range typical for system parameters is required for implementation. In [20], a
electrical drives and outperforms state-of-the-art classical control completely parameter-free adaptive PI controller is proposed
strategies. which guarantees tracking with prescribed transient accuracy.
Index Terms—predictive control, electric motors, nonlinear The controller is applied to current control of (reluctance) syn-
systems. chronous machines, but measurement results are not provided.
In [42] and [47], the inductances are tracked online in order to
adjust the current references thus achieving a higher control ac-
I. I NTRODUCTION
curacy. In [23], a FOC control scheme is proposed, where the
N recent years, reluctance synchronous machines (RSMs)
I have emerged as a competitive alternative to classical syn-
chronous machines (SMs) with permanent magnet (PMSM)
PI control parameters are continuously adapted to the actual
system state, which improves the overall current dynamics.
An alternative to classical control approaches is the use of
or direct current excitation. In addition to the favourable optimization-based control techniques such as model predic-
properties of SMs in general, e.g. high efficiency, reliability tive control (MPC). When using MPC, a parametric optimiza-
and compact design, RSMs are easy to manufacture and com- tion problem is formulated that exploits a model of the plant
parably cheap due to the absence of magnets. Moreover, their to be controlled and enforces constraints while minimizing
anisotropic magnetic path in the rotor, makes them particularly a certain objective function. Although MPC can in principle
suitable for saliency-based encoderless control [29, 30]. improve the control performance and ease the controller design
However, a major drawback of the RSM concerning control [18], meeting the required sampling times is in general a chal-
is its characteristic nonlinearity of the flux linkage, caused by lenging task due to the high computational burden associated
magnetic saturation and cross-coupling effects in the rotor. As with the solution of the underlying optimization problems.
a consequence, the machines’ inductances vary significantly In order to circumvent this difficulty, several algorithmic
with the stator currents. Additional coupling between the stator strategies have been proposed over the past decade that use
d- and q-currents is imposed by the cross-coupling inductances different approaches and (potentially) different formulations of
and the coupling of the nonlinear back electro-motive force the optimal control problems to be solved. Among the possible
in the synchronous reference frame, which requires further classifications of methods present in the literature, in the
fields of electrical drives and power electronics, a fundamental
This research was supported by the German Federal Ministry for Eco-
nomic Affairs and Energy (BMWi) via eco4wind (0324125B) and DyConPV distinction can be made between what is sometimes referred
(0324166B), by DFG via Research Unit FOR 2401 and by the EU via ITN- to as finite (FS-) and continuous control set (CS-) MPC [37],
AWESCO (642 682) [7].
Andrea Zanelli and Gianluca Frison are with the Systems Control and
Optimization Laboratory, Department of Microsystems Engineering, Univer- In FS-MPC, the switch positions of the power converter are
sity of Freiburg, Freiburg, Germany (email: [email protected], gian- regarded as optimization variables leading to mixed-integer
[email protected]). programs. In this way, the need for an external modulator
Julian Kullick is with the research group Control of Renewable Energy
Systems at the Munich School of Engineering, Technical University of is eliminated and the switching sequences are directly deter-
Munich, Munich, Germany (email: [email protected]). mined by the solution to the optimal control problem (hence
Hisham Eldeeb is with IAV GmbH, Munich, Germany. the name “direct” MPC used in some of the literature on MPC
Christoph M. Hackl is with the Department of Electrical Engineering
and Information Technology at the Munich University of Applied Sciences, for electrical drives and power converters [17]).
Munich, Germany. When using CS-MPC instead, we delegate the determination
Moritz Diehl is with the Systems Control and Optimization Laboratory, of switching sequences to an external modulator in order to
Department of Microsystems Engineering and Department of Mathematics,
University of Freiburg, Freiburg, Germany (email: [email protected] obtain a continuous optimization problem. For this reason,
freiburg.de). CS-MPC is sometimes referred to as “indirect” MPC [17].
 2

1.0 0.4
0.5 0.2

ψsd / W

ψsd / W
0.0 0.0
−0.5 −0.2
−0.4
−1.0

40 40
20 20
0 0
/A /A
−40 −40
−20 −20
0 −20 iqs 0 −20 iqs
isd / A 20 −40 isd / A 20 −40
40 40
(a) fitted grey box flux model - d-component (b) fitted grey box flux model - q-component
Figure 1: Nonlinear flux linkage of a real RSM (obtained from FEM data) and fitted grey box model. The worst-case relative error amounts to less than 10%.

Although the computation times associated with this latter II. BACKGROUND ON RSM S AND NMPC
approach scale favourably with prediction horizon length and In order to facilitate the discussion of the design and
number ˙ +
of control variables (typically complexity O N (n u implementation of the proposed controller, in the following,
3

nx ) can be achieved, where N , nu and nx represent horizon mathematical models of RSMs and voltage source inverters
length, number of inputs and states, respectively), for short (VSI) will be derived and numerical methods for NMPC will
horizons, strategies based e.g. on sphere decoding algorithms be introduced. Note that the argument (t), used to denote
applied to FS-MPC formulations can achieve sufficiently short dependence on time, is dropped for the sake of readability.
computation times. On the contrary, CS-MPC is generally
regarded as more computationally expensive and it is still, A. Generic model of the RSM
arguably for this reason, largely unexplored [17]. Among The machine model in the synchronously rotating (d, q)-
the experimental results in the literature obtained with CS- reference frame is given by [21, Chap. 14]
MPC, in [2] a DC-excited synchronous motor is controlled
=:J
using the real-time iteration method. In [14], a fixed-point z }| {
iteration scheme is used to control a permanent magnet syn- 0 −1
d

us = Rs is +ω ψs is + dt ψs i s , (1)
chronous machine (PMSM). Among applications leveraging 1 0
linear-quadratic CS-MPC we mention the work in [12] in d np h i
d
dt ω = mm (i s ) − m l , dt φ = ω,
which permanent magnet synchronous machines and induc- Θ
tion machines are controlled using explicit model predictive where us := (uds , uqs )> are the applied stator voltages, Rs is
control. the stator resistance, is := (ids , iqs )> are the stator currents
and ψs := (ψsd , ψsq )> are the stator flux linkages (functions of
is ). The (d, q)-reference frame rotates with electrical angular
A. Contribution frequency ω = np ωm of the rotor where np is the number of
In this paper, we describe the design and implementation pole pairs and ωm denotes the mechanical angular frequency
details together with experimental results of a nonlinear CS- of the machine. Furthermore, Θ is the total moment of inertia,
MPC controller (CS-NMPC) for an RSM. The contributions mm (is ) := 32 np (is )> Jψs is


(2)
of the present work are:
• We describe the design and implementation details of is the electro-magnetic machine torque, and ml represents an
a tracking CS-NMPC formulation that relies on the external (time-varying) bounded load torque.
software package acados, which is capable of achiev- In order to formulate an optimal control problem, the flux
ing timings in the microsecond time scale necessary to dynamics can be described, based on (1), with the following
control the electrical drive. differential algebraic equation (DAE):
We propose the use of a simple grey box model for the d
dt ψs = us − Rs is − ωJψs + v,
•
flux maps of RSMs that can be used for online applica- (3)
0 = ψs − Ψs (is ),
tions where computation times are of key importance.
• Finally, we present simulation and experimental results where Ψs := (Ψds , Ψqs )> : R2 → R2 represent the identified
that confirm the validity of the proposed control formula- flux maps and v := (v d , v q )> are additive disturbances which
tion and its implementation and its superior performance will be used in a offset-free NMPC formulation (see Section
in comparison with state-of-the-art methods from the II-F).
field of classical control. This is, to the best of the Based on the available flux maps computed through finite
authors’ knowledge, one the of the earliest experimentally element method (FEM), we obtained a continuously differ-
validated applications of CS-NMPC to an RSM. entiable model by fitting a simple grey box model. Due to
 3

β
b hexagon in the αβ-plane is obtained (see Figure 2), where
 
1 1 1 −1 0
2u 0 −
uss = κ udc 2 √3 1 −1 sabc
s s
u010 3 dc u110 2 
0 s (7)
0 2 0 −1 0 1
1u
3 dc s abc
p on the switching vector ss and the Clarke-factor κ ∈
depends
uref

s
u011 √1 u s s 1u s {2/3, 2/3} [21, Chap. 14]. Using space-vector modulation
3 dcu000 = u111 2 dc
u100
α (SVM) to generate the switching√vector, any voltage reference
− 2 udc
3
− 1 udc
3
0 1u
3 dc
2u
3 dc
a
within the circle of radius udc / 3 can be realized, with udc
− 1 udc
3
denoting the (assumed constant) DC link voltage. Finally, the
inverter output voltage is transformed into the rotating (d, q)-
s s
reference frame using the inverse Park transformation, i.e.
u001 − 2 udc u101
3
 d  
us cos(φ) sin(φ) s
us = q = u . (8)
c us − sin(φ) cos(φ) s
| {z }
Figure 2: Voltage hexagon of the two-level VSI. =:Tp (φ)
−1

From now on, since we will only refer to currents, fluxes

and voltages applied to the stator and in the (d-q)-frame, we
their low-dimensionality and simple structure, we propose the
will simplify the notation by dropping the associated subscript
following parametrization of the flux maps:
such that, for example, i = (id , iq ) denotes the stator currents
d2
!
d d q
d −i in the (d-q)-frame.
c
Ψ̂s (is , is , θd ) = q 0 2 exp s
atan(cd1 ids ) + cd2 ids ,
2πσq 2σq2
2!
C. Nonlinear model predictive control
q d q
q −ids NMPC is an optimization-based control strategy that allows
Ψ̂s (is , is , θq ) = √ c0
exp atan(cq1 iqs ) + cq2 iqs ,
2σd2
2
2πσd one to tackle control problems involving potentially nonlinear
dynamics, constraints and objectives by solving online a
with unknown coefficients θd := (cd0 , cd1 , cd2 , σd ) and θq := series of parametric nonlinear programs (NLP). Due to the
(cq0 , cq1 , cq2 , σq ). The numerical values of the coefficients can computational challenge of solving NLPs within the required
be computed by solving the following (decoupled) nonlinear sampling times, NMPC has initially found application in
least-squares problems: the chemical industry and in the field of process control
m X n  2 in general [43], where relatively slow dynamics allow for
Ψds (īds,j , īqs,k , θd ) − Ψ̂ds (īds,j , īqs,k )
X
min sufficiently long sampling times. In more recent times, due to
θd
j=1 k=1 the development of increasingly efficient numerical methods
m X
n  (4)
2 and software implementation and due to the growing computa-
Ψqs (īds,j , īqs,k , θq ) Ψ̂qs (īds,j , īqs,k )
X
min − tional power of embedded control units, NMPC has gradually
θq
j=1 k=1
become a viable approach for applications with much shorter
where īds,j , īqs,k , respectively, are the j-th and k-th data point of computation times. Among other recent works that reported
the current base vectors from the FEM data Ψ̂ds and Ψ̂qs . The on the successful application of MPC to control systems with
fitting problems have been solved with the MATLAB Curve sampling times in the range of milli- and microsecond we
Fitting Toolbox and the resulting fits are shown in Figure 1. mention [1, 48].
In this paper, we will regard the following standard tracking
formulation, where the squared deviation of fluxes ψs and
voltages us from properly defined steady-state references are
B. Model of the two-level VSI penalized:
N −1 2
The machine is supplied by a two-level voltage source 1 X ψi − ψ̄ + 1 kψN − ψ̄k2W

inverter (VSI), which – on average over one switching period min
ψ0 ,...,ψN 2 i=0 ui − ū W 2
N
u0 ,...,uN −1
Ts – translates a given voltage reference
s.t. ψ0 − ψe = 0,
uss,ref := (uα β
s,ref , us,ref ) (5)
g(ψi , ui , ωe , ve ) − ψi+1 = 0, i = 0, . . . , N − 1, (9)
(in the stationary s = (α, β)-reference frame) into the inverter 
udc
2
>
output voltage uss , i.e. ui ui ≤ √ , i = 0, . . . , N − 1,
3
uss (k Ts ) ≈ uss,ref ((k − 1)Ts ), k ∈ N. (6) Cui ≤ c, i = 0, . . . , N − 1,
Since a two-level voltage source inverter may produce a total where g describes the discretized dynamics obtained by in-
of eight unique switching vectors, i.e. sabc
s := (sas , sbs , scs )> ∈ tegrating the differential-algebraic model in (3) using the
{000, 001, 010, 100, 011, 101, 110, 111}, the typical voltage Gauss-Legendre collocation method of order 2 [24] assuming
 4

constant angular velocity ωe and disturbances ve . The variables via Newton-type iterations. In the context of NMPC, efficient
ψ̄ and ū denote the steady-state references computed for a strategies and tailored implementations are available that rely
given desired torque using a maximum-torque-per-Ampere on, for example, exploiting the structure of the model [16, 38],
(MTPA) criterion [13]. Given the flux maps obtained from on reuse of Jacobian factorizations and efficient sensitivity
FEM data in Figure 1, it is possible to compute off-line lookup generation [41] on lifting-based formulations [39] and on
tables (LUTs) that contain the MTPA reference fluxes and inexact iterations [40].
voltages for a finite number of values of the target torque in a Once the linearization is carried out, one generally needs to
specified range. The LUTs are then interpolated online in order solve structured linear systems that can be used to compute
to compute approximate values of ψ̄ and ū associated with the the solution to a quadratic program (QP) as in sequential
specified target torque m̄. We use the notation kxk2P = x> P x, quadratic programming (SQP), compute the update defined
for some positive definite matrix P , to denote the squared by an interior-point method or the one used by other various
P -weighted norm of the vector x. Finally, C and c define strategies, e.g. first-order methods. Since the description of
polytopic constraints (“safety” constraints later) that are meant the details of the different available approaches to solve (9)
to be always inactive at any local solution of (9), but can goes well beyond the scope of this work, we will focus, in the
mitigate constraint violation of intermediate SQP iterates. following, on the SQP strategy, which constitutes the basis for
the real-time iteration (RTI) method used in this application.
Remark 1. Notice that the actual dynamics of the system
involve a coupling of mechanical (ω) and electrical states When using SQP, a sequence of structured QPs of the
(ψ). It is however common, given the large difference between following form needs to be solved:
 >  
associated time constants, to assume a constant angular N −1 s s  >
1 X  i   i  1 sN
 
velocity ω when designing controllers. In our case, it will s
min ui H ui + HN N
allow to use much shorter prediction horizons since we do s0 ,...,sN 2 2 1 1
u0 ,...,uN −1 i=0 1 1
not require the OCP in (9) to steer the speed of the motor
s.t. s0 − x = 0, (10)
to the desired reference, but only fluxes which directly map to
currents and, for a given speed, to torques. si+1 = Asi + Bui + c, i = 0, . . . , N − 1,
Cui + Dxi + e ≤ 0, i = 0, . . . , N − 1,
Problem (9) is used to define an implicit feedback policy
that requires the solution of an instance of the parametric NLP DN xN + eN ≤ 0,
at every sampling time, where the value of the parameter x where A ∈ Rns ×ns , B ∈ Rns ×nu and c ∈ Rns define the
is given by the current estimate of the system’s state. The linearized dynamics obtained through numerical integration
resulting solutions are feasible with respect to the constraints and where C ∈ Rnπ ×nu , D ∈ Rnπ ×ns , e ∈ Rnπ and
and minimize (at least locally) the cost function. Nominal and CN ∈ Rnπ ×ns , eN ∈ Rnπ define the linearized constraints.
N N

inherently robust stability of the closed-loop system can be Finally the matrices
guaranteed in a neighborhood of a steady-state by properly
Q S q
 
choosing the terminal cost [43]. 
QN qN

>
H= S  R r  and HN = > , (11)
Remark 2. Notice that formulations more general than (9) qN 0
q> r> 0
can in principle be used in the framework of NMPC. Among
others, economic costs and more general nonlinear constraints with Q ∈ Rns ×ns , S ∈ Rns ×nu R ∈ Rnu ×nu r ∈ Rnu
and nonlinear cost terms, are features that can be included q ∈ Rns obtained through linearization of the cost, define the
in the problem in order to better capture control design cost of the QP. The matrices and vectors defining the QP (10)
requirements. However, for the application discussed in this are computed based on the linearization associated with the
paper, the nonlinear least-squares problem described in (9) is given current primal-dual iterate z k = (sk , uk , λk, µk ) (where
general enough. λ and µ represent the Lagrange multipliers associated with the
equality and inequality in (9), respectively) and, after solving
D. Numerical methods and software for NMPC (10), the iterate is updated:
In order to be able to solve problem (9) within the available z k+1 ← z k + α(zQP
k
− z k ), (12)
computation time, the use of efficient numerical methods is
k
fundamental. First, since (9) is obtained through a multiple- where zQPrepresents the primal-dual solution of the QP
shooting discretization strategy, an efficient way of computing associated with the linearization point z k and α > 0 is the
evaluations of the discretized dynamics f and of its first step size, which can be adjusted to achieve convergence. Under
(and eventually second order) needs to be available. This is standard assumptions [34], the iterates in (12) converge to a
commonly achieved by means of numerical integration of the local minimum of (9).
ordinary differential equation (ODE) or algebraic differential Due to the computational burden associated with the solu-
equations (DAE) describing the dynamics of the system. tion of QPs and re-linearization of the original NLP in (9), sev-
Although for ODEs explicit and implicit integration methods eral approximate strategies can be used that can significantly
can be used, for DAEs, as for the system under consideration reduce computation times (e.g. [49], [19], [15]). In this work,
(see Section II-A), implicit schemes are generally necessary we will use the RTI strategy [8–10], which relies on a single
that involve the solution of nonlinear root-finding problems SQP iteration in order to provide an approximate feedback
 5

Figure 4: Control diagram: the MTPA LUTs provide the reference flux ψ̄ and
voltage ū associated with a given reference torque m̄. The NMPC
controller computes the optimal control action based on the current
Figure 3: Laboratory setup with dSPACE real-time system, voltage-source state and disturbance estimate provided by an EKF.
inverters connected back-to-back, RSM and PMSM and torque
sensor.

and is considered as a constant-over-time parameter that is

law. The RTI has been implemented in the software packages updated at every sampling time. In [35, Theorem 14] a stream-
MUSCOD-II [11], ACADO [25] and in its successor acados lined version of the results from [32, 33, 36] is presented,
[44], which is shortly described in the next subsection. where under the assumptions, among others, of observability
of the augmented dynamics (13) and asymptotically constant
E. The acados framework disturbances v, the steady-state of the closed-loop system is
offset-free.
The high-performance software package acados [44] pro-
vides a modular framework for NMPC and moving horizon
estimation (MHE). It consists of a C library that implements III. I MPLEMENTATION AND SIMULATION RESULTS
building blocks needed to solve NLPs arising from NMPC and
MHE formulations. It relies on the high-performance linear An RTI strategy [9], where a single QP of an SQP algorithm
algebra package BLASFEO and on the quadratic program is carried out per sampling time, is used to solve (9). In par-
(QP) solver HPIPM and contains efficient implementations of ticular, the generalized Gauss-Newton Hessian approximation
explicit and implicit integration methods. Moreover, it inter- proposed in [44] is used. In this way, the (positive) curvature
faces a number of QP solvers such as qpOASES, qpDUNES contribution coming from the convex spherical constraints on
and OSQP and it provides high-level Python and MATLAB voltages can be exploited in order to improve the Hessian
interfaces. Through these interfaces, one can conveniently approximation used in the QP subproblems. Although in our
specify optimal control problems and code-generate a self- experience this improves a lot the convergence of the RTI
contained C library that implements the desired solver and iterates on this specific problem, the approximate feedback law
can be easily deployed onto embedded control units such as can, from time to time, be largely infeasible with respect to
dSPACE using the automatically generated C wrapper and S- the nonlinear spherical constraints (recall that the intermediate
Function. The code-generation takes place through templated full-step SQP iterates are feasible only with respect to linear
C code which is rendered by the Tera templating engine constraints). Since a-posteriori projection of the control actions
written in Rust. In this way, human-readable C code can be onto the feasible set can deteriorate control performance, we
generated that facilitates the deployment on the target unit. add extra polytopic “safety” constraints (defined by C and c in
(9)) around the spherical ones in order to ensure that constraint
violation will be bounded at any successfully computed iterate.
F. NMPC offset-free tracking formulation In order to be able to meet the short sampling times required
In order to achieve offset-free regulation, we adopt the to control the electrical drive, we use a prediction horizon of
standard strategies discussed, for example, in [36]. In partic- Th = 3.2 ms obtained with 2 shooting nodes (N = 2) and we
ular, we use the following augmented dynamics to design an use the QP solver qpOASES, which is particularly suited for
extended Kalman filter (EKF): problems with short horizons [27]. For both simulation and
d experimental results the controller is run at 4 kHz.
dt ψ = u − Ri − ωJψ + v,
In the following Sections, we will be discussing simulation
d
dt v = 0, (13) and experimental results obtained using the above described
0 = ψ − Ψ(i), RTI strategy to solve (9) with acados.
In order to validate, first in simulation and then experi-
where the disturbance state v is introduced and we assume that
mentally, the proposed approach, we regard a setting where
pseudo-measurements y are available through the interpolated
the RSM is connected to a permanent synchronous ma-
FEM flux maps:
chine (PMSM) which can be used to simulate different load
y = ψm = Ψ̂(im ), (14)
conditions. The CS-NMPC controller has been implemented
while current measurements im are physically carried out on in acados using its Python interface and integrated in a
the machine. An EKF is designed using (13) and (14) which Simulink model that makes use of a high-fidelity model of
uses flux measurements to estimate fluxes ψe and disturbances the system to be controlled including a model of the PMSM
ve . Notice that the angular velocity ωe is estimated externally and of the two-level VSI described in Section II. Moreover,
 6

300 300

200 200

100 100
uq / V

uq / V
0 0

−100 −100

−200 −200

−300 −300

−200 0 200 −200 0 200

ud / V ud / V

Figure 5: Current steps at 165 rads

: simulation results obtained using the CS-NMPC (left) and the gain-scheduled PI controller (right). The voltage spherical
constraints are directly included into the control formulation using the SCQP strategy in [45]. Additionally, a “safety” polytopic constraint is included
which, due to its linearity, is always satisfied exactly.

we have implemented an EKF based on the augmented model dSPACE ControlDesk 6.1p4 for rapid-prototyping, data
(13) using the implicit integrators available in acados. acquisition and evaluation, the custom-built 9,6 kW RSM as
We set the PMSM such that it maintains a constant rota- device under test and a 14,5 kW SEW PMSM as load machine.
tional speed and we change the torque reference fed to the The DR2212 torque sensor allows to measure the mechanical
RSM’s controller in order to assess the tracking performance torque too, but it was not used. The controller based on the
of the proposed controller. We compare the closed-loop tra- formulation described in Section III and implemented using
jectories obtained with the ones achieved when using instead the acados framework has been deployed on the dSPACE
the gain-scheduled PI controller with anti-windup presented unit connected to the physical RSM.
in [22]. The current trajectories obtained with the CS-NMPC Two different experiments have been carried out. In the first
and PI controller are reported in Figure 6 (similarly for input case we used the PMSM to maintain the nominal rotational
trajectories in Figure 5). It can be clearly seen from the speed of the rotor (157 rad/ sec) and different torque refer-
snapshots of the trajectories reported in Figure 7 that the ences have been fed to the RSM controllers under analysis. In
transient can be drastically improved when operating near the the second case, on the contrary, a fixed torque reference was
boundaries of the feasible control set. fed to the CS-NMPC and PI controller and the speed of the
shaft was changed by the PMSM.
The closed-loop trajectories for the conducted experiments
IV. E XPERIMENTAL R ESULTS
are reported in Figure 8-9. Similarly to the results obtained
The presented NMPC scheme has been implemented and in simulation, when tracking torque steps at a fixed speed (in
verified experimentally on a custom-built 9,6 kW RSM (Cour- Figure 8), the proposed CS-NMPC controller achieves better
tesy of Prof. Maarten Kamper, Stellenbosch University, South tracking performance than the gain-scheduled PI controller.
Africa) with the parameters Notice that there is a substantial discrepancy between simula-
rad tion and experimental results right after the third torque step,
Rs = 0,4 Ω, ωm,nom = 157,07 , at t = 0,75 s, due to a drop in the DC-link voltage. In fact, in
s
mm,nom = 61 N m, ı̂s,max = 29,7 A, (15) the presence of a sudden change in the torque reference, the
voltage of the DC-link capacitor can drop if the recharging
ûs,max = 580 V,
rate is slower the discharging rate (behavior not modelled in
and the nonlinear flux linkage maps as depicted in Figure 1 simulation). In the second experiment, in Figure 9, similar
(maps were obtained from FEM). The overall laboratory conclusions can be drawn since the desired currents can be
setup is depicted in Figure 3 and comprises the dSPACE more closely tracked when using CS-NMPC.
real-time system with processor board DS1007 and various
extensions and I/O boards, two 22 kW SEW inverters in back- V. C ONCLUSIONS
to-back configuration sharing a common DC-link, the HOST- In this paper we present simulation and experimental results
PC running MATLAB/Simulink with RCPHIL R2017 and obtained with a CS-NMPC torque controller for RSMs. As
 7

40 40

30 30

20 20
currents (in A)

currents (in A)
10 10

0 0

−10 −10

−20 −20
id id
−30 −30
iq iq
−40 −40
0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0
time (in sec) time (in sec)

rad
Figure 6: Current steps at 157 s
: simulation results obtained using the CS-NMPC (left) and gain-scheduled PI controller (right).

25 25

20 20
currents (in A)

currents (in A)

15 15

10 10

5 CS-NMPC (sim.) 5 CS-NMPC (exp.)

PI (sim.) PI (exp.)
0 0
1.26 1.28 1.30 1.32 1.34 1.36 1.38 1.40 1.26 1.28 1.30 1.32 1.34 1.36 1.38 1.40
time (in sec) time (in sec)

Figure 7: Current steps at 157 rad s

: simulation (left) and experimental (right) results obtained using CS-NMPC and the gain scheduled PI controller: actual
(thin) and filtered (thick). Using the proposed CS-NMPC controller, the settling time can be substantially reduced.

opposed to most successful implementations present in the R EFERENCES

literature, that use instead FS-MPC/NMPC, we show the
effectiveness and real-time feasibility of the continuous control [1] T. Albin, D. Ritter, N. Liberda, R. Quirynen, and
set approach. We show that, using the software implementa- M. Diehl. In-vehicle realization of nonlinear MPC for
tion of the real-time iteration method for NMPC available gasoline two-stage turbocharging airpath control. IEEE
in the software package acados, it is possible to deploy Transactions on Control Systems Technology, pages 1–
the proposed controller on embedded hardware and to meet 13, 2017.
the challenging sampling times typically required to control [2] Thomas J. Besselmann, Sture Van de moortel, Stefan
electrical drives. We discuss implementation details and report Almér, Pieder Jörg, and Hans Joachim Ferreau. Model
on simulation as well as experimental results which show predictive control in the multi-megawatt range. IEEE
that the proposed approach can outperform state-of-art control Transactions on Industrial Electronics, 63(7):4641–4648,
methods. 2015.
[3] R.E. Betz, R. Lagerquist, M. Jovanovic, T.J.E. Miller,
and R.H. Middleton. Control of synchronous reluctance
machines. IEEE Transactions on Industry Applications,
Future research will involve the investigation of novel 29(6):1110–1122, 1993.
numerical methods, e.g. the real-time first-order methods pro- [4] I Boldea, ZX Fu, and SA Nasar. Torque vector con-
posed in [50], to speed up the computation times, which are trol (TVC) of axially-laminated anisotropic (ALA) rotor
currently still rather long and do not easily allow for extensions reluctance synchronous motors. Electric machines and
of the optimal control formulations (e.g longer horizons, state power systems, 19(3):381–398, 1991.
or input spaces of higher dimension, etc). [5] S. Bolognani, L. Peretti, and M. Zigliotto. Online
 8

30 30

20 20
currents (in A)

currents (in A)
10 10

0 0

−10 −10

−20 −20
id (PI, exp.) id (CS-NMPC, exp.)
−30 −30
i (PI, exp.)
q
iq (CS-NMPC, exp.)

0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0
time (in sec) time (in sec)

rad
Figure 8: Torque steps at 157 s
: experimental results obtained using the proposed CS-NMPC controller (left) and gain-scheduled PI controller (right).

40 40

35 35

30 30
currents (in A)

currents (in A)

25 25

20 20

15 15
i (CS-NMPC, exp.)
d
id (PI, exp.)
10 10
i (CS-NMPC, exp.)
q
iq (PI, exp.)
5 5
0 1 2 3 4 0 1 2 3 4
time (in sec) time (in sec)

Figure 9: Speed steps at 61 N m: experimental results obtained using the proposed CS-NMPC controller (left) and gain-scheduled PI controller (right).

3 × 10−4 [6] Abdesselam Chikhi, Mohamed Djarallah, and Khaled

Chikhi. A comparative study of field-oriented control
and direct-torque control of induction motors using an
adaptive flux observer. Serbian Journal of Electrical
CPU time (in sec)

2 × 10−4 Engineering, 7(1):41–55, 2010.

[7] P. Cortes, M.P. Kazmierkowski, R.M. Kennel, D.E.
Quevedo, and J. Rodriguez. Predictive control in power
electronics and drives. IEEE Transactions on Industrial
Electronics, 55(12):4312–4324, 2008.
[8] M. Diehl. Real-Time Optimization for Large Scale Non-
linear Processes. PhD thesis, University of Heidelberg,
10−4
0.0 0.5 1.0 1.5 2.0 2001.
time (in sec) [9] M. Diehl, H.G. Bock, and J.P. Schlöder. Newton-
type methods for the approximate solution of nonlinear
Figure 10: Torque steps at 157 rad
s
: overall control loop turnaround time (in programming problems in real-time. In G. Di Pillo
black) obtained with the CS-NMPC controller using acados with
qpOASES (available computation time of 250 µs in red).
and A. Murli, editors, High Performance Algorithms and
Software for Nonlinear Optimization, pages 177–200.
Kluwer Academic Publishers B.V., 2002.
[10] M. Diehl, R. Findeisen, and F. Allgöwer. A stabilizing
MTPA control strategy for DTC synchronous-reluctance-
real-time implementation of nonlinear model predictive
motor drives. IEEE Transactions on Power Electronics,
control. In L. Biegler, O. Ghattas, M. Heinkenschloss,
26(1):20–28, 2011.
 9

D. Keyes, and B. van Bloemen Waanders, editors, Real- Differential Equations II: Stiff and Differential–Algebraic
Time and Online PDE-Constrained Optimization, pages Problems. Springer Series in Computational Mathemat-
23–52. SIAM, 2007. ics. Springer-Verlag, Berlin, 2nd revised edition, 2004.
[11] M. Diehl, D.B. Leineweber, and A.A.S. Schäfer. [25] B. Houska, H. J. Ferreau, and M. Diehl. ACADO toolkit
MUSCOD-II Users’ Manual. IWR-Preprint 2001-25, – an open source framework for automatic control and
University of Heidelberg, 2001. dynamic optimization. Optimal Control Applications and
[12] A. Domahidi, S. Mariethoz, and M. Morari. High- Methods, 32(3):298–312, 2011.
bandwidth explicit model predictive control of electrical [26] Marten J. Kamper, F.S. van der Merwe, and
drives. IEEE Transactions on Industry Applications, S. Williamson. Direct finite element design optimisation
48(6):1980–1992, November 2012. of the cageless reluctance synchronous machine. IEEE
[13] Hisham Eldeeb, Christoph M. Hackl, Lorenz Horlbeck, Transactions on Power Conversion, 11(3):547–555,
and Julian Kullick. A unified theory for optimal feed- 1996.
forward torque control of anisotropic synchronous ma- [27] D. Kouzoupis, G. Frison, A. Zanelli, and M. Diehl.
chines. International Journal of Control, 91(10):2273– Recent advances in quadratic programming algorithms
2302, 2017. for nonlinear model predictive control. Vietnam Journal
[14] T. Englert and K. Graichen. A fixed-point iteration of Mathematics, 46(4):863–882, 2018.
scheme for model predictive torque control of pmsms. [28] R. Lagerquist, I. Boldea, and T.J.E. Miller. Sensorless-
In Proceedings of the 6th IFAC Conference on Nonlinear control of the synchronous reluctance motor. IEEE
Model Predictive Control, volume 51, pages 568–573, Transactions on Industry Applications, 30(3):673–682,
Madison, Wisconsin, USA, August 2018. 1994.
[15] C. Feller and C. Ebenbauer. Relaxed logarithmic bar- [29] Peter Landsmann, Ralph Kennel, Hugo W. de Kock,
rier function based model predictive control of linear and Marten J. Kamper. Fundamental saliency based
systems. IEEE Transactions on Automatic Control, encoderless control for reluctance synchronous machines.
62(3):1223–1238, March 2017. In Proceedings of the XIX International Conference on
[16] Jonathan Frey, Rien Quirynen, Dimitris Kouzoupis, Gi- Electrical Machines (ICEM), pages 1–7, 2010.
anluca Frison, Jens Geisler, Axel Schild, and Moritz [30] Peter Landsmann, Dirk Paulus, Peter Stolze, and Ralph
Diehl. Detecting and exploiting Generalized Nonlinear Kennel. Reducing the parameter dependency of encoder-
Static Feedback structures in DAE systems for MPC. In less predictive torque control for reluctance machines.
Proceedings of the European Control Conference (ECC), In Proceedings of the IEEE International Symposium on
2019. Sensorless Control for Electrical Drives (SLED), pages
[17] Tobias Geyer. Model Predictive Control of High Power 93–99, 2010.
Converters and Industrial Drives. John Wiley & Sons, [31] T. Matsuo and T.A. Lipo. Field oriented control of
2016. synchronous reluctance machine. In Power Electronics
[18] Tobias Geyer, Georgios Papafotiou, and Manfred Morari. Specialists Conference, 1993. PESC ’93 Record., 24th
Model predictive direct torque control – Part I: Concept, Annual IEEE, pages 425–431, Jun 1993.
algorithm, and analysis. IEEE Transaction on Industrial [32] Manfred Morari and Urban Maeder. Nonlinear offset-free
Electronics, 56(6):1894–1905, 2009. model predictive control. Automatica, 48(9):2059–2067,
[19] K. Graichen and A. Kugi. Stability and incremen- 2012.
tal improvement of suboptimal mpc without terminal [33] K. R. Muske and T. A. Badgwell. Disturbance modeling
constraints. IEEE Transactions on Automatic Control, for offset-free linear model predictive control. Journal
55(11):2576–2580, 2010. of Process Control, 12:617–632, 2002.
[20] Christoph M. Hackl. Current PI-funnel control with anti- [34] J. Nocedal and S. J. Wright. Numerical Optimization.
windup for synchronous machines. In Proceedings of the Springer Series in Operations Research and Financial
54th IEEE Conference on Decision and Control, pages Engineering. Springer, 2 edition, 2006.
1997–2004, 2015. [35] G. Pannocchia, M. Gabiccini, and A. Artoni. Offset-free
[21] Christoph M. Hackl. Non-identifier Based Adaptive Con- MPC explained: novelties, subtleties, and applications. In
trol in Mechatronics. Springer International Publishing, Proceedings of the 5th IFAC Conference on Nonlinear
2017. Model Predictive Control, volume 48, pages 342–251,
[22] Christoph M. Hackl, Maarten J. Kamper, Julian 2015.
Kullick, and Joshua Mitchell. Nonlinear PI cur- [36] G. Pannocchia and J.B. Rawlings. Disturbance Models
rent control of reluctance synchronous machines. for Offset-Free Model-Predictive Control. AIChE Jour-
arxiv.org/pdf/1512.09301.v1, 2015. nal, 49:426–437, 2003.
[23] C.M. Hackl, M.J. Kamper, J. Kullick, and J. Mitchell. [37] D. E. Quevedo, R. Aguilera, and T. Geyer. Model
Current control of reluctance synchronous machines with predictive control for power electronics applications. In
online adjustment of the controller parameters. In 2016 S. V. Rakovic and W. S. Levine, editors, Handbook of
IEEE 25th International Symposium on Industrial Elec- Model Predictive Control, pages 551–580. Birkhäuser,
tronics (ISIE). IEEE, jun 2016. Cham, 2019.
[24] Ernst Hairer and Gerhard Wanner. Solving Ordinary [38] R. Quirynen, S. Gros, and M. Diehl. Efficient NMPC for
 10

nonlinear models with linear subsystems. In Proceedings

of the IEEE Conference on Decision and Control (CDC),
pages 5101–5106, 2013.
[39] R. Quirynen, S. Gros, and M. Diehl. Lifted implicit
integrators for direct optimal control. In Proceedings of
the IEEE Conference on Decision and Control (CDC),
pages 3212–3217, 2015.
[40] R. Quirynen, S. Gros, and M. Diehl. Inexact Newton-type
optimization with iterated sensitivities. SIAM Journal on
Optimization, 28(1):74–95, 2018.
[41] R. Quirynen, M. Vukov, and M. Diehl. Auto genera-
tion of implicit integrators for embedded NMPC with
microsecond sampling times. In Mircea Lazar and Frank
Allgöwer, editors, Proceedings of the 4th IFAC Nonlinear
Model Predictive Control Conference, pages 175–180,
2012.
[42] E.M. Rashad, T.S. Radwan, and M.A Rahman. A
maximum torque per ampere vector control strategy for
synchronous reluctance motors considering saturation
and iron losses. In Conference Record of the 2004
IEEE Industry Applications Conference (39th IAS Annual
Meeting), volume 4, pages 2411–2417, 2004.
[43] J. B. Rawlings, D. Q. Mayne, and M. M. Diehl. Model
Predictive Control: Theory, Computation, and Design.
Nob Hill, 2nd edition edition, 2017.
[44] Robin Verschueren, Gianluca Frison, Dimitris
Kouzoupis, Niels van Duijkeren, Andrea Zanelli,
Rien Quirynen, and Moritz Diehl. Towards a modular
software package for embedded optimization. In
Proceedings of the IFAC Conference on Nonlinear
Model Predictive Control (NMPC), 2018.
[45] Robin Verschueren, Niels van Duijkeren, Rien Quirynen,
and Moritz Diehl. Exploiting convexity in direct opti-
mal control: a sequential convex quadratic programming
method. In Proceedings of the IEEE Conference on
Decision and Control (CDC), 2016.
[46] Longya Xu, Xingyi Xu, Thomas A. Lipo, and Donald W.
Novotny. Vector control of a synchronous reluctance
motor including saturation and iron losses. IEEE Trans-
actions on Industrial Applications, 27(5):977–985, 1991.
[47] S. Yamamoto, J.B. Adawey, and T. Ara. Maximum
efficiency drives of synchronous reluctance motors by
a novel loss minimization controller considering cross-
magnetic saturation. In Proceedings of the 2009 IEEE
Energy Conversion Congress and Exposition, pages 288–
293, 2009.
[48] A. Zanelli, G. Horn, G. Frison, and Moritz Diehl. Nonlin-
ear model predictive control of a human-sized quadrotor.
In Proceedings of the European Control Conference
(ECC), pages 1542–1547, 2018.
[49] A. Zanelli, R. Quirynen, and M. Diehl. Efficient zero-
order nmpc with feasibility and stability guarantees. In
Proceedings of the European Control Conference (ECC),
Naples, Italy, June 2019.
[50] A. Zanelli, Q. Tran-Dinh, and M. Diehl. Contraction
estimates for abstract real-time algorithms for nmpc. In
Proceedings of the IEEE Conference on Decision and
Control (CDC), Nice, France, Dec 2019. (accepted).