See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/357819237

Data-Driven Continuous-Set Predictive Current Control for Synchronous Motor

Drives

Article in IEEE Transactions on Power Electronics · January 2022

DOI: 10.1109/TPEL.2022.3142244

CITATIONS READS
88 434

4 authors:

Paolo Gherardo Carlet Andrea Favato

University of Padua University of Padua
31 PUBLICATIONS 621 CITATIONS 18 PUBLICATIONS 312 CITATIONS

SEE PROFILE SEE PROFILE

Saverio Bolognani Florian Dörfler

ETH Zurich ETH Zurich
142 PUBLICATIONS 4,668 CITATIONS 365 PUBLICATIONS 21,988 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Paolo Gherardo Carlet on 01 February 2022.

The user has requested enhancement of the downloaded file.

 Data-Driven Continuous-Set Predictive Current
Control for Synchronous Motor Drives
Paolo Gherardo Carlet1 , Andrea Favato1 , Saverio Bolognani2 , Florian Dörfler2
1
Electric Drives Laboratory, Dept. of Industrial Engineering, University of Padova, Padova (Italy)
2
Automatic Control Laboratory, ETH Zurich, 8092 Zurich (Switzerland)
paologherardo.carlet@unipd.it, andrea.favato@phd.unipd.it, bsaverio@ethz.ch, dorfler@ethz.ch

Abstract—Optimization-based control strategies are an affirmed power electronics control tasks for two decades, reaching an
research topic in the area of electric motors drives. These methods industrial and commercial level [2].
typically rely on an accurate parametric representation of the Continuous control set (CCS) MPC methods for PMSM
motor equations. In this paper, we present the transition from
model-based towards data-driven optimal control strategies. We current control, which is the focus of this paper, rely on a state-
start from the model predictive control paradigm which uses space model of the motor to build the predictive controller [3]–
the voltage balance model of the motor. Second, we discuss the [5]. The model is commissioned by performing an experimental
prediction error method, where a state-space model is identified characterization of specific parameters. These procedures often
from data, without a parametrization. Moving toward data-driven include many different tests and they require specific measuring
controls, we present the Subspace Predictive Control, where
a reduced model is constructed based on the singular value devices and proper test-bed setups. Then, the resulting accurate
decomposition of raw data. The final step is represented by a model can be exploited in real-time by means of look-up-tables.
complete data-driven approach, named data-enabled predictive Alternatively, parameters could be estimated via offline [6]
control, in which raw data is not encoded into a model but or online [7] procedures. Self-commissioning and auto tuning
directly used in the controller. The theory behind these techniques techniques are also consolidated strategies. In [8], an exhaustive
is reviewed and applied for the first time to the design of the
current controller of synchronous permanent magnet motor drives. survey of research and state-of-art parameter identification and
Design guidelines are provided to practitioners for the proposed self-commissioning methods for AC motor drives is discussed.
application and a way to address offset-free tracking is discussed. In particular, these approaches are of interest when high
Experimental results demonstrate the feasibility of the real-time performance control is required with sensorless applications.
implementation and provide comparisons between model-based Finally, many methods have been proposed in literature to
and data-driven controls.
improve the robustness against parameter variations [9]–[13],
Index Terms—Data driven control, model predictive control although most of these strategies are implemented for finite-set
(MPC), permanent magnet synchronous motor (PMSM), predic- MPCs.
tion error method (PEM), subspace predictive control (SPC),
data-enabled predictive control (DeePC) The key idea behind data-driven predictive controllers is
to avoid the model identification stage entirely, and design
the controller directly from collected input/output (I/O) data,
I. I NTRODUCTION e.g. voltage/current samples. This approach overcomes the
challenges of model selection and identification, resulting of par-
The interest in data analysis is constantly growing, supported ticular interest for many industrial applications [14]. However,
by an unprecedented availability of computational power and there are just a few examples of data-driven control applications
memory storage, as well as advances in optimization, statistics for electric motor drives. In [15], an observer is coupled to
and machine learning. This leads to an increasing attention an MPC to update the PMSM model, improving its reliability.
towards data-enabled methods in all branches of science and However, this approach still relies on a parametric model.
engineering. This revolution has a significant impact on the A controller design procedure was proposed by Wallscheid
control engineering too. Data-driven control design consists in et al. in [16], based on deep reinforcement learning. The
synthesizing a controller using the data collected on the real solution guarantees the benefits of an optimal controller, without
system, without defining and identifying a parametric model for requiring expensive computations. Many effective techniques
the plant [1]. This is in contrast with model-based approaches, have been presented which go toward the data-driven paradigm,
which rely on plant modeling and identification procedures. named model-free [17]–[19] or parameter-free [20] algorithms.
The epitome of this model-based paradigm is arguably the In particular, [17] and [18] propose to online update a parameter-
model predictive control (MPC), which has been applied to free model, but they rely on the hypothesis that there are no
data available for guessing an initial controller, which might
A preliminary version of part of the results contained in this manuscript has
been previously presented at the 2020 IEEE Energy Conversion Congress and be too restrictive.
Exposition (ECCE), 11-15 Oct. 2020, Detroit, MI, USA, entitled ”Data-driven In this work we show a transition from model-based to
predictive current control for synchronous motor drives”. data-driven control design, considering as application the
This research was supported by Fondazione “Ing. Aldo Gini” – via Portello,
15 – 35129 Padova (Italy), by the University of Padova (Project SID 2017 current control of PMSMs. This control task serves as a well-
-BIRD175428), and by ETH Zurich funds. understood benchmark for new methods, despite the fact that
other traditional non-data-driven methods yield satisfactory techniques can impact the electric drives field. Differently
results for this application. We consider optimization-based from other machine-learning based solution that can be found
control scheme, i.e. MPC-type solutions. We first present a in literature, the proposed schemes are more computationally
state-of-the-art CCS-MPC, whose model is obtained through a efficient, less data hungry and more suitable to rigorous stability
previous characterization of the motor parameters. Then, we and robustness analysis [28].
move step by step towards more data-driven control designs,
exploiting just voltage and current measurements collected from II. M ODEL -BASED MPC OF PMSM C URRENTS
the motor. First, the prediction error method (PEM) technique
coupled with MPC is presented, which is a consolidate solution According to the MPC paradigm, the future control input
for identifying a state-space model from data [21]. A further sequence u = [u(k), u(k + 1), . . . u(k + N − 1)]T is optimized
step is represented by the subspace predictive control (SPC) in order to steer the predicted future output y = [y(k+1), y(k+
[22], where the collected data are processed offline by means of 2), . . . y(k + N )]T to a desired reference r = [r(k + 1), r(k +
a least-square program, and the resulting auto-regressive model 2), . . . r(k + N )]T . Only the first optimal input of the sequence
with exogeneous inputs (ARX) is de-noised by singular value u(k) is applied to the plant (receding horizon principle). Thus,
thresholding. This pseudo-identification procedure is used to the following optimization problem is solved at each control
build a linear predictor for the currents dynamics. Finally, a period:
 
completely data-driven control algorithm is presented, named min ∥y − r∥2Q + ∥u∥2R (1a)
the data-enabled predictive control (DeePC) [23], [24], where u,x,y
the system identification process is completely avoided and the subject to x(k + 1) = Ax(k) + Bu(k), y = Cx, (1b)
collected data are directly used in the controller. This technique u(k) ∈ U, k = 0, . . . , N − 1 (1c)
has already found application in power electronics [25]–[27].
The contributions of this work are manifold: where N is the prediction horizon, Q ≥ 0 and R > 0 are
two weighting matrices, A and B represent the state space
• we illustrate the perspective of data-driven control design
model used to predict the output y = Cx, and U is the
using a predictive control framework; input feasible set. Considering the specific application, U is
• we demonstrate the practical real-time implementation
a polytope [29]. If the set U is neglected, the problem is
of data-driven methods, which is not trivial since data- referred to unconstrained, and it has a closed-from solution of
driven methods are expensive in terms of computation reduced computational burden. On the opposite, if the polytopic
and samples; constraints are included, the optimization problem becomes a
• we show that data-driven paradigm can be a systematic
quadratic program (QP) which requires an online QP solver
design tool for PMSM current controllers; like qpOASES, as in [30], but it is still usually solvable in
• we compare the computational aspects of the presented
real-time.
control strategies;
In the context of PMSM currents control, future currents
• as a technical contribution, we address the problem of the
are estimated by exploiting a parametric model, based on
offset-free tracking for the SPC and DeePC methods;
the PMSM voltage balance equations, represented in the dq
• we provide guidelines for the choice of the control
reference frame, synchronous with the rotor flux. The equations
parameters and excitation input signals for this application.
are arranged in a state-space form and discretized using the
A relevant advantage of data-driven strategies is that they explicit Euler approximation technique:
can be easily implemented as automatic procedures that excite
idq (k + 1) = Aidq (k) + Budq (k) + Bh(k)
the system with predefined input signals, perform offline
calculations, and deliver a ready-to-use control law. No special
 Ts Lq  T
s

1 − Rs ωe Ts 0 , (2)
skills or specialized commissioning personnel are required Ld Ld  Ld
A= Ts  , B = 
  
Ld Ts 
to set up the procedure. This approach could be interesting −ωe Ts 1 − Rs 0
for some industrial challenges. For example, in compressor Lq Lq Lq
for refrigeration equipment or submersible pumps, offline where Rs is the stator winding resistance, Ts is the sampling
characterizations cannot be performed when PMSMs are period, ωe is the electric angular speed and Ld and Lq are
inaccessible. Another case of interest is multi-purpose drives, the d and q-axis inductances, respectively. Moreover, idq and
where algorithms suitable for different PMSM topologies are udq are the dq currents and voltages, respectively. udq are the
needed. In addition, PMSM and inverter manufacturers are inputs of the system whereas idq are the states. Finally, h =
often different companies and they were never meant to be [0 − ωe Λpm ]T is the back-electro motive force (back-EMF) due
integrated in the same application. Moreover, if the motor drive to the permanent magnet flux linkage Λpm . In the considered
needs to be manually re-tuned during its life-cycle, data-driven application, the full state, i.e. motor currents, is measurable.
procedures represent a simple and reliable method to adapt the This model neglects the cross-saturation phenomena, as well
initial design. as iron-saturation and back-EMF harmonics effects. Thus, the
The goal of this paper is not to demonstrate the superiority model can result as oversimplified for some PMSM topologies,
of the data-driven paradigm over the model-based approach. such as pure reluctance motors. However, many CCS-MPCs
Instead, this work provides some concrete, although prelimi- proposed in literature work with even more simplified models,
nary, answer to the contemporary question how data-driven obtaining indeed good results. In particular, the dependence of
matrix A on the operating speed ωe is neglected, preferring ud
a constant A matrix for the real-time implementation [29]. PMSM
INV
We will see later that the data-driven paradigm overcomes uq
iα iβ
the problem of selecting a model structure. Moreover, it is
αβ θe p
worth noticing that the Euler discretization does not introduce dq
significant errors because of the high sampling rates, which id iq
are typical of the power electronics area.
An integral action is included in the MPC formulation by uc yc
means of the velocity form of the MPC problem (1) [31], in
order to achieve an unbiased current reference tracking. The Offline design
discussion about the offset-free data-driven control is given in
Sec. III-C. (a) Scheme of the data collection process.

III. T OWARDS DATA -D RIVEN C ONTROL OF PMSM D RIVES Memory

Buffer
A data-driven controller for PMSM solves essentially a Tini
uini 1
current reference tracking problem analogous to the one z
i∗dq uαβ
presented in (1). However, in contrast to the parametric model DATA-DRIVEN dq PWM
αβ PMSM
(2) used in the MPC solution, a non-parametric model is MPC udq INV
adopted, consisting of raw measurements arranged in a matrix yini θe p
representation. The construction of this model happens offline,
Memory
therefore it is not an adaptive controller. A data-driven controller Buffer
idq dq iabc
abc
design procedure consists of two steps: Tini
• A data collection step, followed by offline rearrangement
(b) Block scheme of the online controller.
of the voltages/currents samples into proper matrices;
• An online program, when the tracking problem is solved, Fig. 1: Overview of the data collection step and the online program.
with the voltages/currents samples matrices acting as a
constraint. In this online step, the controller has access to
the latest I/O (voltage/current) samples and optimizes the SPC — In the SPC algorithm, the state-space model (1b)
predictions over an horizon of N steps. is replaced by a different algebraic constraint that relates the
future currents trajectory with the past Tini voltages/currents
A. Data Collection and Offline Computations samples and the future N input voltage samples.
All the considered data-driven designs begin from the To obtain this model, two Hankel matrices H (uc ) and
collection of a T -long sequence of I/O voltages uc and currents H (y c ) are built using the collected sequences uc and y c :
y c measurements (Fig. 1(a)), where the superscript c stands
 
u1 u2 · · · uT −Tini −N +1
for collected. The sequence uc = [uc1 ; uc2 ; ...; ucT ] ∈ R2T  u2 u3 · · · uT −Tini −N +2 
H (uc ) :=  .  . (4)
 
contains the inverter reference voltages and it fulfills the .. ..
 .. . . 
persistency of excitation requirement [32, Corollary 2], that is
the Hankel matrix of inputs in (4) needs to have full raw rank. uTini +N uTini +N +1 · · · uT
The selection of the input signal is further discussed in Sec. The number of columns of a Hankel matrix is hereinafter
IV-A. The resulting output sequence contains the dq currents denoted as L. Given T data and the horizon lengths N and
y c = [y1c ; y2c ; . . . ; yTc ] ∈ R2T . Tini , L can be computed as: L = T − Tini − N + 1. The
output matrix H (y c ) is built in an analogous way from the
PEM-MPC — In the PEM-MPC method the state-space
samples y c . Then, the matrices are partitioned in Past and
model (1b) is used, similarly to a standard MPC approach.
Future sub-blocks:
However, the coefficients of the state-space matrices A and B    
used in (1b) are inferred from data by means of an ordinary UP YP
:= H (uc
), := H (y c ), (5)
least-square problem1 that involves the sequence uc and yc : UF YF
T
X −1 where UP contains the first Tini block rows of H (uc ), i.e. 2Tini
min ∥xc (k + 1) − Axc (k) − Buc (k)∥2 . (3) rows, and UF the remaining N block rows. The block Hankel
A,B
k=1 matrices YP and YF are similarly obtained. The dimensions
The main difference between the resulting model and the of all the presented matrices are summarized in Table I for
parametric voltage balance equation (2) is that the latter convenience. The I/O block Hankel matrices UP , UF , YP and
inherently requires the ad-hoc identification procedures to YF are used in the SPC design to construct an ARX predictor
identify all the electric parameters (Rs , Ld , Lq , Λpm ). The [33]:  
PEM method, instead, does not enforce any parametrization uini
y = Pw + Pu u, (6)
of the model and the resulting matrices can, in general, have a yini
structure that is different from the one of (2).
where uini , yini ∈ R2Tini are the past dq voltage and current
1 We refer to [25] for a discussion on how to solve this problem numerically. samples, respectively, u, y ∈ R2N are the future ones. The
 TABLE I: Overview of matrices dimensions for the considered PMSM current control application.

Matrix H (uc ) H (y c ) UP YP UF YF Pw Pu M Φ A B
rows 2(Tini + N ) 2(Tini + N ) 2Tini 2Tini 2N 2N 2N 2N L L 2 2
columns L L L L L L 4Tini 2Tini 4Tini 2N 2 2

matrices Pw and Pu are computed solving the least-square PEM-MPC — PEM-MPC algorithm is completely analo-
problem gous to a standard model-based MPC, from the point of view of
" # 2
UP the online program. It is worth remembering that two possible
min YF − [Pw | Pu ] YP , (7) online controllers can be obtained, depending on the presence or
Pw ,Pu UF not of the constraints (1c). If the problem is unconstrained ((1c)
  is absent), the PEM-MPC yields a linear feedback controller
UP
where Pw multiplies the two blocks and Pu multiplies [5] of the form u = Kr r + Kx x(k). On the other hand, the
YP
the block UF . The matrix Pw is exploited in (6) to set the QP problem requires an iterative solver as in [30], if input
initial condition of the prediction, i.e. to compute the term constraints are included. In both situations, the complexity of
Pw (uini , yini )T . A singular value decomposition (SVD) of the the PEM-MPC is the same of a standard model-based MPC,
initial trajectory predictor Pw can be performed to mitigate which is amenable for real-time implementation on adequate
the noise effect in the data [22]. Only the dominant singular hardware. The dimension of 2N the decision variable coincides
values are used to construct a reduced-rank matrix. with the dimension of u ∈ R , thus it scales linearly with
the prediction horizon. In the considered application, the full
DeePC — The design of a DeePC controller is purely data- state of the system is available, but in general the PEM-MPC
driven, as the data block Hankel matrices defined in (5) are requires a state estimator. The other two data-driven methods,
used in their raw form in the controller. This method is based on SPC and DeePC, do not require a state estimator, since they
the so called Fundamental Lemma of behavioral system theory naturally work with the plant outputs.
[32], which guarantees that (under persistency-of-excitation
assumptions on uc ) any trajectory of the system needs to satisfy, SPC — The SPC algorithm solves the same tracking problem
L
for a unique g ∈ R , the linear equations (1) as in MPC or PEM-MPC, but with the state-space model
    (1b) replaced by the predictor (6):
UP uini
YP   yini 
UF  g =  u  . (8) min ||y − r||2Q + ||u||2R (10a)
u,y
YF y
 
u
subject to y = Pw ini + Pu u (10b)
Implicitly, (8) serves as a predictor of the future N -long I/O yini
voltages/currents trajectory (u, y) based on Tini -long I/O initial u(k) ∈ U, k = 0, . . . , N − 1. (10c)
trajectory (uini , yini ). If we consider (u, y) as free optimization
Similarly to the PEM-MPC, if the constraints (10c) are not
variables, the vector g that satisfies the first two block-equations
of (8) can be expressed explicitly as present, then we can solve the problem in closed-form by
 †     substituting the predictor equation (10b) into (10a) and by
UP uini u setting the gradient of the resulting convex quadratic cost to
g= + Φz = M ini + Φz, (9)
YP yini yini zero. The resulting online controller is a linear feedback of
′
where † denotes the Moore-Penrose pseudo-inverse operator, the form u = Kr r + Kini [uini , yini ]T . If the constraints (10c)
and Φ, such that [UP , YP ]T \Φ = 0, represents a basis of the are present, the minimization program can be solved online,
kernel of M. Both Φ and M can be computed offline using at the same computational complexity of the PEM-MPC one.
standard linear algebra routines. This decomposition allows In fact, the computational burden depends on the length of u.
expressing the future trajectory as a function of the lower- We remind that this property does not hold for those systems
dimensional variable z, and turns out to be useful in the online whose states and outputs have different dimensions.
phase of the unconstrained control problem, as explained in
DeePC — The DeePC algorithm, because of the implicit
the next subsection.
form of the algebraic constraint, requires the minimization over
B. Computational Aspects of the Online Stage the decision variables g, u, y:
In the online stage, the MPC tracking problem (1) is solved, min ||y − r||2Q + ||u||2R + λg ||g||2 (11a)
g,u,y
but with different representations in place of (1b) depending    
UP uini
on the adopted data-driven method. Both the unconstrained YP   yini 
and constrained solutions are now discussed for each data- UF  g =  u  , u(k) ∈ U, k = 0, ..., N − 1, (11b)
s.t.    
driven method, clarifying the practicality of their real-time
YF y
implementation from the computational burden point of view.
A complete overview of the data-driven design procedures is where λg adds a regularization on the decision variable g.
provided in Fig. 2, where the differences between the presented In fact, if noisy data are used, the Hankel matrices are full
methods are highlighted both for the offline and online stage. raw rank, but the realized control error in (11a) could be
different from the predicted one. Thus, the term λg ||g||2 helps Select T
Apply the voltages uc , collect the resulting currents y c
to robustify the control problem [34, Section III.C]. In the (uc , y c )
unconstrained case, the problem can be solved directly using the Build Hankel matrices H(uc ) and H(y c )
null-space representation presented in (9). The future currents H(uc ), H(y c )

and voltages sequences u and y are replaced in (11a) with Split the blocks UP , UF , YP and YF Select Tini , N

UF g and YF g, respectively, obtaining

Control selection

   
uini u
min ||YF (M + Φz) − r||2Q + ||UF (M ini + Φz)||2R . PEM-MPC SPC DeePC
z yini yini Identify A,B Compute Pw , Pu Compute M, Φ
(12) Pw , P u
The solution of the problem is available in closed form as SVD of Pw
z opt = H -1 dT , where the Hessian matrix H and the linear term (Pw , Pu )
d are defined as: (A, B) (M, Φ)
Constraints

H := ΦT YTF QYF Φ + ΦT UTF RUF Φ No Yes

(13)
   

u T 
u T
Compute K r , K ini Select Q, R
d := r − YF M ini QΦ − UF M ini RΦ. or K r , K x
yini yini OFFLINE
(K r , K ini )
ONLINE Set-up problem (1a) + (1c) Select Q, R
u = K r r + K ini [uini , yini ]T
The Hessian inversion can be evaluated offline with proper u Equality constr.
numerical techniques, further reducing the complexity of the Project u in U
scheme. More details on the closed-form solution of the
unconstrained DeePC can be found in [25]. Starting from PEM-MPC SPC DeePC
(1b) with Equation (16) Equation (8)
the optimal value of z opt , (9) is used to compute g opt , and, identified A,B
finally, the sequence of optimal input uopt . It is still possible to
condense this controller in a feedback law similar to the SPC, QP problem solution Choose solver

with a decision variable that scales linearly with the prediction

horizon length. The constrained solution of (12) would instead Optimal control input
require an online QP-solver. However, the dimension of the
Fig. 2: Overview of the three proposed data-driven controller design
decision variable g can be large, as it depends on the number procedures.
of samples used in (8). Thus, the real-time implementation of
the DeePC algorithm is still a challenging problem. -0.8 1.1
meas meas
In conclusion, three main aspects differentiate the SPC ref q-current (p.u.) ref
d-current (p.u)

and DeePC methods [34]: the way the predictor is built, the -0.9 1

underlying prediction model and the variables over which the

-1 0.9
QP problem is solved. In fact, the SPC forces a least-square
fit to a linear time invariant (LTI) system model, whereas the -1.1 0.8
DeePC does not. Thus, SPC is more suited for LTI systems 0.95 1 1.05 0.95 1 1.05
time (s) time (s)
or linear parameter varying ones. On the other hand, DeePC
exhibits interesting features also when applied to non-linear (a) d-current (b) q-current
system, even if the fundamental lemma requires in theory a Fig. 3: Offset-free tracking error: simulation of DeePC algorithm
linear time invariant plant, e.g. the grid connected inverter with no integral action (before t = 1s) and with the offset-free
application shown in [25]. Finally, SPC solves the tracking implementation after t = 1s.
problem in the input u, whereas the DeePC in g.
e.g. ∆y = y(k) − y(k − 1). For instance, the DeePC problem
in (11) is written as follows:
C. Integral Action ||∆y − r′ ||2Q + ||∆u||2R + λg ||g||2
min
g,∆u,∆y
 ′ 
UP ∆uini
 
An integral action is needed to avoid bias errors in the
Y′P   ∆yini 
currents reference tracking for the SPC and DeePC algorithms. subject to  ′  g =  (14)
UF ∆u 
For instance, the back-EMF induced by the magnets acts as a Y′F ∆y
constant disturbance in the voltage equation, inducing a steady- u(k) = u(k − 1) + ∆u(k) ∈ U, k = 0, 1..., N − 1
state error in reference tracking. Following this principle, we r(k)′ = r(k) − y(k), k = 1, ..., N.
introduce this framework also for data-driven controllers. The
integral action can be included by formulating the optimization [U′P Y′P U′F Y′F ]T are the Hankel matrices filled with incremental
problem in its velocity-form [31]. The idea is to perform the data. The optimization problem (14) is solved for g, then ∆u(k)
data collection stage filling the matrices with incremental data, is found.
 max(||yc||)
1
||yc||

||idq|| (p.u.)
0.8
0.6
0.4
0.2
0
15 30 45 60 75 90
Excitation voltage (V)

(a) Output currents during the excitation tests for different uexc
values: mean and maximum output currents values.
1

SV magnitude (log10)
0
15V
-1 30V
45V
60V
Load motor Test motor Load panel -2 75V
90V
Load inverter dSPACE Test inverter -3
MicroLabBox 1 2 3 4
Singular Value (SV)
Fig. 4: Test-bed layout.
(b) Singular values analysis of Pw for different uexc values.
TABLE II: Overview of the drive parameters.
Fig. 5: Overview of some key parameters of the data collection test.
Parameter Symbol Value
Pole pairs p 3
Phase resistance Rs 1Ω choice, the rotor is not required to be locked or to be maintained
d-axis inductance Ld 0.010 H
q-axis inductance Lq 0.014 H at standstill by another motor. The selected zero-mean voltage
PM flux-linkage Λpm 0.26 V s sequence induces zero-mean currents and, consequently, a zero
Nominal current IN 6.2 Arms mean torque. Since the mechanical dynamic is much slower
Nominal d current IN,d −1.1 A
Nominal q current IN,q 8.7 A than the electric one, the rotor remain at standstill even if
Nominal speed ΩN 1000 1/min instantaneously the torque is not zero. In addition non-linear
DC bus voltage UDC 300 V frictions help to avoid rotations of the motor. If the mechanical
Sampling period Ts 100 µs
inertia of the system is too low or the frictions are negligible
the rotor could slightly move, as happen for other injection-
The effectiveness of the proposed solution is shown in Fig. 3. based commissioning techniques [35]. Designers should be
The nominal current reference has been set and steady state aware of this potential issue in some specific applications.
is reached, while the motor is kept at nominal speed. Before The criteria to select the voltages amplitude is here discussed.
time t = 1s, the standard data-driven formulation is considered The motor is driven by a two-level voltage source inverter with
as controller. As can be seen, a bias appears in the tracking. a DC bus voltage of 300 V. The voltage sequence is generated
At time t = 1s, the controller designed with incremental data by picking the values from a uniform probability distribution
is selected and the bias is removed. in the interval [−uexc , uexc ]. We propose a test to analyze the
effects of uexc on the sequence y c and the data-driven design.
IV. E XPERIMENTAL VALIDATION Fig. 5(a) refers to several excitation tests, characterized by
different values of uexc . On one hand, the maximum excitation
The authors propose the experimental validation on an
voltage should be limited to avoid over-currents, preserving a
interior permanent magnet motor. The nominal parameters
safe motor operation. The figure, in fact, shows that the mean
of the considered machine are reported in Table II, while
value of the currents samples are quite low with respect to the
the test-bench layout is shown in Fig. 4. All the algorithms,
nominal value. However, the nominal current value, for the
i.e. the MPC, PEM-MPC, SPC and DeePC, are real-time
proposed motor, is achieved using uexc = 90 V, i.e. the 30%
implemented on the dSPACE MicroLabBox at a sample rate
of the DC bus voltage. Higher excitation voltages should be
of Ts = 100 µs. The MPC nominal model is commissioned
avoided. On the other hand, a too low voltage excitation could
by means of standard tests and offline post-processing. The
lead at least to current sampling issues, due to small signal
motor under test is not significantly affected by the magnetic
to noise ratios. Moreover, we need to take into consideration
non-linearity.
also other problems, i.e. if the information carried by the data
is rich enough to describe the current dynamics. The PWM
A. Data Acquisition Step synthesis of low voltages could emphasize some inverter non-
The test designed to collect I/O data from the interior linearities, e.g. not properly compensated dead-times, that are
permanent magnet (IPM) motor consists of excitation with a not of interest of our identification. In order to evaluate if the
random (detailed below) dq voltage vectors sequence uc and the data are collected properly, the dominant singular values of the
measurement of the dq currents via LEM sensors. Thanks to this matrix Pw are analyzed (see the logarithmic plot in Fig. 5(b)).
 TABLE III: Matrices dimensions resulting from the design choices: N = 3, Tini = 1 and T = 100 samples.

Matrix H (uc ) H (yc ) UP YP UF YF Pw Pu M Φ A B

rows 8 8 2 2 6 6 6 6 97 97 2 2
columns 97 97 97 97 97 97 4 2 4 6 2 2

The number of dominant values should be coherent with the 0.015

measured predicted
anticipated dimension of the state, see Table I. Two dominant 0.01

iq (A)
values characterize the considered dynamic, as expected. 0.005

-0.005

B. Parameters Selection -0.01

1 1.01 1.02 1.03 1.04 1.05
time (s)
In this section we address the problem of parameters (a) MPC
selection for designing the data-driven controls. The prediction
0.015
horizon length N is chosen according to the MPC framework, measured predicted
0.01
i.e. N = 3. This value is a good trade-off between accuracy

iq (A)
and computational effort for this application [29]. Moreover, all 0.005

the these controllers share the same cost function; thus, equal 0
weighting matrices Q and R are chosen. In particular, Q is the -0.005
identity matrix, whereas R is the identity scaled by a factor -0.01
1 1.01 1.02 1.03 1.04 1.05
0.0001. We consider the robust formulation of the DeePC, and time (s)
the related parameter in (11) has been set to λg = 0.1.
(b) DeePC
Two parameters that characterize the data-driven algorithms
are the length of the initial trajectory Tini and the number Fig. 6: Accuracy of the data-driven predictors in the estimation of the
q-axis current variation.
of samples T . The trajectory [uini , yini ]T replaces the initial
condition for the prediction. Thus, it determines the inherent
system state, and the parameter Tini provides a complexity for
200 1500
the model. In [32], the system lag2 l is used to find a lower
PDF [1/A]

PDF [1/A]
1000
bound for Tini . In particular, if Tini ≥ l the system prediction is 100
500
uniquely determined. Thank to this criteria, the value of Tini
0 0
can be chosen even without knowing the system dimension, but DeePC DeePC
SPC 0.02 SPC 0.005
using an estimate of it. Since the system lag is known for the PEM
MPC 0
PEM
MPC 0
considered application (i.e l = 1), we set Tini = 1. The length -0.02
res [A]
-0.005
res [A]
T of the recorded I/O vectors should be long enough to make
(a) d-axis current (b) q-axis current
sure that the Hankel matrices have full rank. The Fundamental
Lemma in [32] gives a lower bound for T , whose value for Fig. 7: Residual analysis of the prediction error: probability distribution
function of the residuals for the presented predictors.
the considered application is T ≥ 3(Tini + N + 2) − 1. We
selected T = 100 samples, which satisfies the inequality.

D. Accuracy of the Data-Driven Predictor

C. Complexity of the online program The accuracy of the data-driven predictors is investigated in
this subsection, taking the model-based MPC as benchmark.
The design choices described in the previous subsection This analysis is performed during steady-state operation, when
set all the dimensions of the matrices presented in TA- the motor is working at the nominal maximum-torque-per-
BLE III. All the controllers have been implemented in ampere current point (see Table II) at standstill. During the tests
their unconstrained version, i.e. a feedback law of the form the currents are regulated by standard PI controllers. We are
u = K r r + K ini [uini , yini ]T or u = K r r + K x x(k). This means interested on the open-loop prediction accuracy of the methods.
that the turn-around-time of all the controllers are similar. This means that the predictors are fed by current measurements
A slightly higher computation time is required for the first and the reference voltages computed by the PIs. A first
feedback law. The turn-around-time of each predictive control qualitative information on the accuracy is provided by Fig. 6(a)
scheme is about 9.6 µs - 9.7 µs, depending on the specific and Fig. 6(b). The figures show the comparison between the
feedback-law. The dSPACE MicroLabBox is equipped with measured q-axis current increments and the predicted ones,
a 2 GHz NXP QorIQ P5020 microprocessor. The number of using respectively the predictor obtained with (3) and (12). We
computations required by the feedback laws scales linearly
with respect to the length of the prediction horizon N and the 2 The lag l of a linear system is the smallest integer value for which the
length of the initial trajectory Tini . observability matrix O = [C CA...CAl−1 ]T has full rank.
 TABLE IV: Residual analysis of the prediction error: mean and standard deviation of the residuals.
Controller MPC PEM SPC DeePC
d-axis q-axis d-axis q-axis d-axis q-axis d-axis q-axis
Mean (µA) 2.4 −0.4 −1.2 −0.2 −2.9 −0.4 −2.8 −0.4
Standard deviation (µA) 3100 500 2100 350 3800 590 3800 590

observe a good correspondence between measurements and TABLE V: Total harmonic distortion analysis at the nominal point.
predictions for both the controllers. Controller MPC PEM SPC DeePC
The residuals between measured and estimated currents are THD 0.32% 0.25% 0.36% 0.36%
considered as performance index, as suggested in [15]. The
results of this analysis are reported in Fig. 7. The figures
show the estimated probability density function of the d and q The distortion is quite low, as expected from a continuous-
residuals for all the described predictors. From literature [15], set MPC, which encapsulates a modulator in the controller
we expect a zero mean normal distribution of the residuals, structure. The PEM-MPC seems again the preferable structure,
which is coherent with the obtained results. The PEM-MPC as observed also in the previous section. Nevertheless, both SPC
predictor appears the most accurate one, proving that using and DeePC grant the same harmonic distortion of a benchmark
data to validate the commissioning tests is an interesting tool. (model-based) velocity-form MPC.
Finally, the closed-loop cost analysis is proposed to provide
E. Online Unconstrained Controller a different insight on the steady-state performance. The closed-
loop cost is computed for each control step using (14), where
In this subsection we provide a comparison between model- the weight matrices are designed as described in Sec. IV-B.
based and data-driven designed controllers in terms of step The test considers the machine at standstill (Fig. 8) and at
current reference response. In particular, the reference r is nominal speed (Fig. 9), when steady-state condition is reached.
changed from zero to the nominal maximum-torque-per-ampere Results are described in Fig. 10, where the cost is reported in
current. The model-based MPC adopts the motor parameters decibel. The closed-loop cost achieved by the data-driven and
which were previously obtained by means of characterization model-based controllers are not significantly different, both at
procedures (see Table II). All the data-driven controllers are standstill Fig. 10(a) and at nominal speed Fig. 10(b). Thus, a
designed from the same data recording, in particular the one purely data-driven design assure the same closed-loop cost of
defined by a uexc = 50 V. a standard model-based procedure, even if no assumption on
The step responses are compared at standstill in Fig. 8. It is the model structure is required.
interesting to notice that the data-driven designs allow achieving
similar performances with respect to the model-based controller.
In fact, the commissioning effort of all the proposed algorithms V. C ONCLUSION
in terms of measurement apparatus, number of carried out tests In this work we present a transition path from model-based
and their complexity and duration is much lighter compared to to data-driven design of PMSM current controllers. Different
the characterization required to build an accurate model-based data-driven algorithms are considered: the prediction error
controller [8]. Among data driven controllers, the DeePC is method model predictive control, the subspace and the data-
considered the most data-oriented algorithm, because it uses raw enabled predictive controls (DeePC). All the algorithms were
data without any pre-processing. Despite the direct exploitation online implemented in the unconstrained version, proving
of raw data, it has almost the same performance as the others. their online feasibility. Similar accuracy between model-based
We reported the same step response analysis also at nominal and data-driven predictors is demonstrated with experimental
speed rate in Fig. 9. This test confirms the effectiveness of the data. Experimental results show that all these controllers have
integral action included in the data-driven control framework. comparable performance, considering the MPC with an accurate
The back-EMF and the state transition matrix A of the IPM model as benchmark. Moreover, among data-driven controllers,
motor model (2) depend on the operating speed. Thus, a bias the DeePC performs well both in steady state and dynamics.
in the current tracking should be observed if the integral action There are several challenges to address in the future. First,
is missing (as in Fig. 3). We underline that the proposed data- a comparison between data-driven and self-commissioning
driven methods seems very effective for the q-axis current. techniques would be valuable. This could help to design
Moreover, other tools can be used to further improve their effective strategies for the excitation voltage signals. Second,
behavior, in particular the one of the DeePC (see [23]). In the extension of data-driven methods for nonlinear system
addition, accordingly to [36], a feed-forward term can be nested is at the beginning. The possibility to automatically include
in the controller to improve disturbance rejection performances, the motor nonlinearities in the control law, e.g. magnetic
without penalizing the overshoot in the dynamics. We therefore cross saturation effects, is of particular interest. Third, finding
believe that there is much unexplored potential to improve the computationally efficient methods for implementing high-
performance to data-driven controllers. dimensional data-driven methods that include constraints in
Concerning the steady-state behaviors, a total harmonic real-time is still an open challenge. Finally, other future research
distortion analysis is performed on the motor currents at will focus on online adaptation of data-driven control structures
nominal speed. The results are briefly summarized in TABLE V. and applications to other drives problems.
 1 1 1 1
current (p.u.)

current (p.u.)

current (p.u.)

current (p.u.)
0.5 d-ref
0.5 d-ref
0.5 d-ref
0.5 d-ref
q-ref q-ref q-ref q-ref
0 d-curr
0 d-curr
0 d-curr
0 d-curr
q-curr q-curr q-curr q-curr
-0.5 -0.5 -0.5 -0.5

-1 -1 -1 -1
0 0.005 0.01 0.015 0.02 0 0.005 0.01 0.015 0.02 0 0.005 0.01 0.015 0.02 0 0.005 0.01 0.015 0.02
time (s) time (s) time (s) time (s)

(a) MPC (b) PEM-MPC (c) SPC (d) DeePC

Fig. 8: Comparison of the step responses of model-based and data-driven current controllers at standstill.

1 1 1 1
current (p.u.)

current (p.u.)

current (p.u.)

current (p.u.)
0.5 d-ref
0.5 d-ref
0.5 d-ref
0.5 d-ref
q-ref q-ref q-ref q-ref
0 d-curr
0 d-curr
0 d-curr
0 d-curr
q-curr q-curr q-curr q-curr
-0.5 -0.5 -0.5 -0.5

-1 -1 -1 -1
0 0.005 0.01 0.015 0.02 0 0.005 0.01 0.015 0.02 0 0.005 0.01 0.015 0.02 0 0.005 0.01 0.015 0.02
time (s) time (s) time (s) time (s)

(a) MPC (b) PEM-MPC (c) SPC (d) DeePC

Fig. 9: Comparison of the step responses of model-based and data-driven current controllers at nominal speed.

-20 [9] F. Wang, K. Zuo, P. Tao, and J. Rodrı́guez, “High performance model
-20
-40
predictive control for PMSM by using stator current mathematical model
Closed-loop cost

Closed-loop cost

-40
self-regulation technique,” IEEE Trans. Power Electron., vol. 35, no. 12,
-60 pp. 13 652–13 662, 2020.
-80 -60 [10] X. Zhang, L. Zhang, and Y. Zhang, “Model predictive current control
for PMSM drives with parameter robustness improvement,” IEEE Trans.
-100
-80 Power Electron., vol. 34, no. 2, pp. 1645–1657, 2019.
-120 [11] X. Liu, L. Zhou, J. Wang, X. Gao, Z. Li, and Z. Zhang, “Robust predictive
-100
-140 current control of permanent-magnet synchronous motors with newly
MPC PEM SPC DeePC MPC PEM SPC DeePC designed cost function,” IEEE Trans. Power Electron., vol. 35, no. 10,
pp. 10 778–10 788, 2020.
(a) Standstill (b) Nominal speed
[12] M. Yang, X. Lang, J. Long, and D. Xu, “Flux immunity robust predictive
Fig. 10: Closed-loop cost analysis at steady state. current control with incremental model and extended state observer
for PMSM drive,” IEEE Trans. Power Electron., vol. 32, no. 12, pp.
9267–9279, 2017.
[13] M. De Soricellis, D. Da Rù, and S. Bolognani, “A robust current
R EFERENCES control based on proportional-integral observers for permanent magnet
[1] Z. S. Hou and Z. Wang, “From model-based control to data-driven synchronous machines,” IEEE Trans. Ind. Appl., vol. 54, no. 2, pp.
control: Survey, classification and perspective,” Information Sciences, 1437–1447, 2018.
vol. 235, pp. 3–35, 2013. [14] S. Yin, H. Gao, and O. Kaynak, “Data-driven control and process
[2] S. Vazquez, J. Rodriguez, M. Rivera, L. G. Franquelo, and M. No- monitoring for industrial applications - Part I,” IEEE Trans. Ind. Electron.,
rambuena, “Model predictive control for power converters and drives: vol. 61, no. 11, pp. 6356–6359, 2014.
Advances and trends,” IEEE Trans. Ind. Electron., vol. 64, no. 2, pp. [15] A. Brosch, S. Hanke, O. Wallscheid, and J. Böcker, “Data-driven recursive
935–947, Feb 2017. least squares estimation for model predictive current control of permanent
[3] S. Hanke, O. Wallscheid, and J. Böcker, “Continuous-control-set model magnet synchronous motors,” IEEE Trans. Power Electron., vol. 36, no. 2,
predictive control with integrated modulator in permanent magnet pp. 2179–2190, 2021.
synchronous motor applications,” in Int. Electric Machines Drives Conf., [16] M. Schenke, W. Kirchgässner, and O. Wallscheid, “Controller design for
2019. electrical drives by deep reinforcement learning: A proof of concept,”
[4] S. Bolognani, R. Kennel, S. Kuehl, and G. Paccagnella, “Speed and IEEE Transactions on Industrial Informatics, vol. 16, no. 7, pp. 4650–
current model predictive control of an IPM synchronous motor drive,” 4658, 2020.
in Int. Electric Machines Drives Conf. (IEMDC), 2011. [17] Y. Zhang, J. Jin, and L. Huang, “Model-free predictive current control of
[5] A. Favato, P. G. Carlet, F. Toso, and S. Bolognani, “A model predictive PMSM drives based on extended state observer using ultralocal model,”
control for synchronous motor drive with integral action,” in 44th Annu. IEEE Trans. Ind. Electron., vol. 68, no. 2, pp. 993–1003, 2021.
Conf. of the IEEE Ind. Electron. Soc. (IECON), 2018.
[18] F. Tinazzi, P. G. Carlet, S. Bolognani, and M. Zigliotto, “Motor parameter-
[6] Q. Wang, G. Zhang, G. Wang, C. Li, and D. Xu, “Offline parameter
free predictive current control of synchronous motors by recursive least-
self-learning method for general-purpose PMSM drives with estimation
square self-commissioning model,” IEEE Trans. Ind. Electron., vol. 67,
error compensation,” IEEE Trans. Power Electron., vol. 34, no. 11, pp.
no. 11, pp. 9093–9100, 2020.
11 103–11 115, 2019.
[7] A. Boglietti, A. Cavagnino, and M. Lazzari, “Experimental high- [19] S. Aghaei Hashjin, S. Pang, E. H. Miliani, K. Ait-Abderrahim, and
frequency parameter identification of ac electrical motors,” IEEE Trans. B. Nahid-Mobarakeh, “Data-driven model-free adaptive current control
Ind. Appl., vol. 43, no. 1, pp. 23–29, 2007. of a wound rotor synchronous machine drive system,” IEEE Trans.
[8] S. A. Odhano, P. Pescetto, H. A. A. Awan, M. Hinkkanen, G. Pellegrino, Transport. Electrific., vol. 6, no. 3, pp. 1146–1156, 2020.
and R. Bojoi, “Parameter identification and self-commissioning in AC [20] N. A. Losic and L. D. Varga, “A current-free and parameter-free control
motor drives: A technology status review,” IEEE Trans. Power Electron., algorithm,” IEEE Trans. Ind. Appl., vol. 30, no. 2, pp. 324–332, 1994.
vol. 34, no. 4, pp. 3603–3614, Apr. 2019. [21] J. B. Jørgensen, J. K. Huusom, and J. B. Rawlings, “Finite horizon
 MPC for systems in innovation form,” in 50th IEEE Conf. Decision and
Control and Eur. Control Conf., 2011.
[22] W. Favoreel, B. De Moor, and M. Gevers, “Spc: Subspace predictive
control,” IFAC Proc. Volumes, vol. 32, no. 2, pp. 4004 – 4009, 1999.
[23] J. Coulson, J. Lygeros, and F. Dörfler, “Data-enabled predictive control:
In the shallows of the deepc,” 18th Eur. Control Conf. (ECC), 2018.
[24] J. Coulson, J. Lygeros, and F. Dörfler, “Distributionally robust chance
constrained data-enabled predictive control,” 2020.
[25] L. Huang, J. Coulson, J. Lygeros, and F. Dörfler, “Data-enabled predictive
control for grid-connected power converters,” in 58th Conf. Decision
and Control (CDC), 2019.
[26] L. Huang, J. Zhen, J. Lygeros, and F. Dörfler, “Quadratic regularization
of data-enabled predictive control: Theory and application to power
converter experiments,” 2020.
[27] P. G. Carlet, A. Favato, S. Bolognani, and F. Dörfler, “Data-driven
predictive current control for synchronous motor drives,” in IEEE Energy
Convers. Congr. and Expo. (ECCE), 2020.
[28] C. De Persis and P. Tesi, “Formulas for data-driven control: Stabilization,
optimality, and robustness,” IEEE Transactions on Automatic Control,
vol. 65, no. 3, pp. 909–924, 2020.
[29] G. Cimini, D. Bernardini, S. Levijoki, and A. Bemporad, “Embedded
model predictive control with certified real-time optimization for syn-
chronous motors,” IEEE Trans. Control Syst. Technol., pp. 1–8, 2020.
[30] F. Toso, P. G. Carlet, A. Favato, and S. Bolognani, “On-line continuous
control set MPC for PMSM drives current loops at high sampling rate
using qpOASES,” in IEEE Energy Convers. Congr. and Expo. (ECCE),
2019.
[31] G. Pannocchia, M. Gabiccini, and A. Artoni, “Offset-free MPC explained:
novelties, subtleties, and applications,” IFAC-PapersOnLine, vol. 48,
no. 23, pp. 342 – 351, 2015.
[32] J. Willems, P. Rapisarda, I. Markovsky, and B. De Moor, “A note on
persistency of excitation,” Syst. Control Lett., vol. 54, Apr. 2005.
[33] B. Huang and R. Kadali, Dynamic Modeling, Predictive Control and
Performance Monitoring: A Data-driven Subspace Approach. Springer-
Verlag London, 2008.
[34] F. Dörfler, J. Coulson, and I. Markovsky, “Bridging direct & indirect
data-driven control formulations via regularizations and relaxations,”
2021.
[35] A. Varatharajan, P. Pescetto, and G. Pellegrino, “Sensorless self-
commissioning of synchronous reluctance machine with rotor self-locking
mechanism,” in 2019 IEEE Energy Conversion Congress and Exposition
(ECCE), 2019, pp. 812–817.
[36] R. Kadali, B. Huang, and A. Rossiter, “A data driven subspace approach
to predictive controller design,” Control Engineering Practice, vol. 11,
no. 3, pp. 261–278, 2003.

View publication stats