An optimization formulation of converter control and its

general solution for the four-leg two-level inverter

Abdelkader Bouarfa, Marc Bodson, Maurice Fadel

To cite this version:

Abdelkader Bouarfa, Marc Bodson, Maurice Fadel. An optimization formulation of converter control
and its general solution for the four-leg two-level inverter. IEEE Transactions on Control Systems
Technology, 2018, 26 (5), pp.1901 - 1908. �10.1109/TCST.2017.2738608�. �hal-01898582�

HAL Id: hal-01898582

[Link]
Submitted on 18 Oct 2018

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 An optimization formulation of converter con-
trol and its general solution for the 4-leg 2-level
inverter
Abdelkader Bouarfa, Marc Bodson, Fellow, IEEE, and Maurice Fadel, Member, IEEE

Abstract—The paper proposes an optimization formulation of creases rapidly as the number of switches grows, especially
the control problem for power electronic converters. A benefit of with the widely-used space-vector representation. For the 4-leg
the approach is a systematic method for the control of high 2-level inverter, the addition of the fourth-leg raises the num-
switch-count static converters. In the case of the 3-phase, 4-leg, 2-
ber of possible locations of a reference voltage vector from a
level inverter, the framework provides a characterization of all
the possible solutions that yield a maximal extension of the invert- 2D plane divided into 6 sectors [4], [5] to 24 tetrahedrons in a
er linearity range. The method makes it possible to recover well- 3D representation [12], [19]–[21].
known modulation strategies, as well as to discover some new Starting from these observations, the paper proposes a new
ones having different properties and distinct advantages. The control strategy different from traditional PWM control and
characteristics resulting from different design choices are evalu- based on on-line numerical optimization using linear pro-
ated in simulations, with consideration being given to the linearity
gramming techniques. This strategy is expected to be more
range, total harmonic distortion, and switching losses. Key prin-
ciples of extension of the proposed method to multilevel, multileg generic, to be less dependent on the switch count, and to avoid
converters are given, as well as motivations for an FPGA-based difficulties encountered with a geometric approach. The con-
hardware implementation enabling real-time PWM control. trol problem of the converter is formulated as an under-
determined constrained optimization problem. Such problem
Index Terms—Voltage source inverter, 4-leg 2-level inverter, is remarkably similar (although not identical) to the control
voltage control, PWM, control allocation, optimization, simplex allocation problem studied previously in flight control and
algorithm, median voltage injection.
marine applications [22], [23]. In this paper, the focus is
placed on the 4-leg 2-level inverter, where a complete charac-
I. INTRODUCTION terization of the solutions is possible without the use of numer-
ical techniques, thus enabling a good understanding of the pos-
F IXED high switching-frequency pulse-width modulation
(PWM) is an essential class of modulation methods for
power converters [1], [2]. For the widely-studied 3-leg 2-level
sibilities opened by this new formulation.
A previous implementation of the control allocation ap-
proach method for the 4-leg inverter was based on the descrip-
inverter, two popular PWM methods are: (1) carrier-based tion of its active voltage vectors [24] and the regular-sampled
pulse-width modulation (CBPWM) [1]–[3], using low- symmetric PWM (RSPWM) [3]. The correct non-zero vector
frequency modulating signals and a high-frequency carrier sequence was computed using a numerical optimization meth-
wave, and (2) space vector theory, using a 3-phase geometric od based on the simplex algorithm, removing the need to iden-
vector representation of the inverter and leading to the well- tify a reference tetrahedron, as was done in earlier work.
established space vector modulation (SVM) techniques [1], Considering switching cells as available resources subjected
[2], [4], [5]. The main theoretical differences between to ranges of operation, a different control allocation method is
CBPWM and space vector theory are related to the way the introduced here, reducing the size of the problem considerably.
degree of freedom left in the inverter is exploited [1], [6]–[12]. Moreover, the paper proposes an analytical solution of the
Typically, this degree of freedom is used to increase the in- optimization problem that reduces the control of the inverter to
verter linearity range [4]–[7], [13], [14], to reduce switching the computation of the median of a special series. Furthermore,
losses [15], [16], or to mitigate current harmonics using subop- CBPWM-equivalent formulas are given for modulation laws
timal modulating solutions [17]–[18]. derived from the new control allocation problem.
However, the difficulty of developing control strategies in- As constraints are taken into account in the optimization
problem, solutions resulting from control allocation methods
A. Bouarfa (corresponding author) and M. Fadel are with LAPLACE, for power converters naturally yield a maximal extension of
Université de Toulouse, CNRS, INPT, UPS, France: 2 rue Charles Camichel, the linearity range of the inverter [24], [25]. Specific choices
F-31071 Toulouse, France (bouarfa@[Link], fadel@[Link]-
[Link]).
of the algorithm’s parameters produce modulation laws with
M. Bodson is with University of Utah, Dept. Electrical and Computer En- different properties, all implementable with few computations.
gineering, 50 S Central Campus DR RM 2110, Salt Lake City UT 84112- Under the general umbrella of the proposed strategy, not only
9206, USA (bodson@[Link]).
 iDC TS
TS /2
DA Triangular
1 carrier
SA SB SC SN
EDC iA RA LA 0
A t
2 SA
iB RB LB N 1
B
iC RC LC
0 t
C DATS DATS
RN LN
EDC N Fig. 2. Illustration of the realization of gating pulse on switch A of width
2 DA·TS and centered over the switching period TS (RSPWM).
SA SB SC SN vNG
K  A, B, C,  VKN  TS  E DC DK  DN   E DC DKN  
G
Fig. 1. Illustration of the 4-leg 2-level inverter connected to a three-phase
load with impedances ZA, ZB, ZC. The linearity range of the converter is defined here as the
range of amplitude of the 3-phase sinusoidal reference voltage
are previously-proposed modulation laws found as special system for which the converter delivers a voltage system with
cases, but new interesting modulation schemes are discovered the desired values. Because of the star-connection, with classic
as well. Simulation results illustrate the properties of the dif- sinusoidal PWM (SPWM), the linearity range is restricted to
ferent methods obtained. Switching losses and total harmonic EDC/2. A common modification extends this range by varying
distortion (THD) on voltages and currents are evaluated in the neutral potential of the load. When the desired voltages
computations for comparison. are balanced and three-phase, the maximal output amplitude
achievable linearly is 1/3·EDC, as for SVM [4], [5] or SPWM
II. 4-LEG 2-LEVEL INVERTER ASSUMPTIONS FOR PWM CONTROL with 1/6 third-harmonic injection (THIPWM1/6) [13], [14].
A. Inverter and load
III. CONTROL ALLOCATION FOR INVERTER CONTROL
The 4-leg 2-level inverter is a well-proven solution for 3-
phase systems with neutral wire, for applications like distribut- A. Control allocation methods
ed generation, active power filtering, common-mode active Control allocation methods were developed in aerospace
filtering or fault-tolerant operation of electric drives [12], problems as a solution for over-actuated systems subjected to
[19]–[21]. Fig. 1 illustrates the inverter, star-connected to a constraints [22], [23]. It is assumed that the process to be con-
load with per-phase resistances and inductances values RK, LK, trolled can be modeled by a constrained system of m control
K{A,B,C,N}. The load can be unbalanced or nonlinear. The equations with n unknowns, with m<n:
fourth-leg (N) is connected to the neutral point of the load in
order to control the neutral voltage and also handle possible  B  u  a, umin  u  umax  
unbalanced currents. Here, RN and LN are neglected, and the
The matrix B, of size m×n, specifies the effectiveness of the
ground reference G is the lower DC bus in order to facilitate
actuators. B is assumed to be of rank m (rows are linearly in-
the derivation of control equations. EDC is the DC bus voltage.
dependent). The vector u, of length n (column), is the control
B. Switching cells vector that specifies the chosen use of actuators. The vector a,
A switching cell is taken to be an association of two func- of length m (column), is the resulting output vector. Denote
tional switches working at complementary binary states. Dead umin and umax (vectors) as the minimum and maximum bounds
times are neglected. Each leg is comprised of one switching of each component of the control vector u, respectively. One
cell. For K{A,B,C,N}, the state of the switching cell K is the wants to find a control vector u to obtain a given desired out-
state of the corresponding upper switch, SK, equal to 0 when put vector ad. Considering the constraints, the problem (2) may
the switch is off and 1 when the switch is on. have zero, one, or an infinity of solutions. Methods have been
developed to extract a solution from the whole set of feasible
C. Output voltages, load voltages solutions by formulating an optimization problem.
The 4 output voltages VKG are referred to the ground G. Re- The objective of the 4-leg 2-level inverter control is to ob-
garding the star-connection, it is useful to define for each of tain a desired reference vector ad = Vref = (VANref, VBNref, VCNref)T
the three first legs a relative switching state SKN = SK − SN. of 3 load voltages on average over the switching period TS. In
Then, the 3 independent load voltages, referred to the neutral order to scale the voltages with respect to EDC, define the
point N, are given by VKN = EDC SKN, for K{A,B,C}. scaled reference vector as ΔDref = Vref / EDC = (DANref, DBNref,
DCNref)T. Here, ΔDref is similar to the difference of duty cycles
D. Pulse-width modulation
between the three first legs and the fourth one, see (1). Then,
The mean value DK =<SK>Ts of SK over the switching period the system of equations of (2) depends on the chosen control
TS defines the duty cycle of the switching cell K, DK[0,1]. variables and the converter control requirements.
Mean-value references voltages are obtained with RSPWM by
determining the gating pulses SK from the duty cycles DK as B. Switching cell formulation
illustrated on Fig. 2. The load voltage mean values are The control problem can be formulated as finding duty cycle
commands DS to the cells such that  K  A, B, C, DK  DKN ref  DN  
 Dref  M S DS , 0  DS  1  
If there is at least one positive and one negative voltage in the
 1 0 0 1 reference vector (it is the case if the reference voltage vector is
 
 M S   0 1 0 1  DS  D A DB DC DN   
T
three-phase), bounds DNmin and DNmax can be defined as
 0 0 1 1
  DN min   min Dref , DN max  1  max Dref  

The switching cell formulation (3) is a simpler alternative to and the whole set of feasible solutions that achieve the refer-
the space vector formulation in [24], introducing less control ence vector is described by constraints on DN alone:
variables and a smaller matrix. In order to find a control solu-
tion to the under-determined constrained control problem (3),  DN DN min ,DN max , J ctrl DN  0  
the following optimization problem is defined:
Therefore, as long as the constraint set is not empty, i.e., as
min J , J  M S  DS  Dref   0    DS  DS pref long as DNmin ≤ DNmax, any signal resulting from solutions of
DS 1   (5) will achieve a maximal linearity range.
subject to 0  DS  1
B. Reduction of the solution to a single decision variable
The primary criterion Jctrl(DS) = ||MSDS − ΔDref||1 is a control The secondary criterion to be minimized is
error based on the control equation in (3). Because the solu-
tion is often not unique, the secondary criterion Jpref(DS) =  J pref ( DS )   K
K A, B ,C , N 
DK  DK pref  
ε·|DS − DSpref| is a deviation error added to specify preferred
uses of the switches, where | | denotes the application of the By defining a series dmod of elements dmodK as
element-wise absolute value. The column vector DSpref =
(DApref, DBpref, DCpref, DNpref)T corresponds to the preferred duty  K A,B,C, d mod K  DK pref  DKN ref  
cycle values. The row vector ε = (εA, εB, εC, εN) helps to give  d mod N  D N pref d mod  d mod K KA, B ,C , N   
further preference to one or more switches, which makes it
possible to affect the weight of deviation errors from preferred and by substitution in (25), the criterion Jpref can be now re-
positions in Jpref. Although the parameters εK can be real num- formulated as an explicit function of DN:
bers, we will restrict them to be integers, with little loss of
generality. This choice will result in a simple solution of the  J pref ( DN )   K
K  A , B ,C , N 

DN  d mod K  
optimization problem. Finally, the small real parameter ε0
gives priority to the minimization of the control error Jctrl [22]. that is a sum of distances on the real-axis from DN to the points
The optimal solutions are determined by the choice of the dmodK and weighted by the parameters εK. The single decision
parameters DSpref and ε, which produce a large number of pos- variable DN solely represents the available functional degree of
sible configurations. We will discuss a few choices in the next freedom. The optimization problem is considerably simplified.
sections. The associated control diagram is shown on Fig. 3.
C. Solution of the optimization problem
IV. CHARACTERIZATION OF ALL FEASIBLE SOLUTIONS WHEN The minimization of sums of weighted absolute deviations
THE REFERENCE IS ACHIEVABLE from a unique number (here DN) to other fixed numbers (here
dmodK) is a well-known mathematical problem [26]–[30]. Here,
A. Constraint set when the reference is achievable the solution is introduced thanks to the definition of the special
When the reference vector is achievable, the optimization series dmodε, for which each dmodK appears εK times. From [26]–
problem can be simplified. Solutions can be determined ana- [30], it can be deduced that, without considering constraints,
lytically, depending on the parameters DSpref, and ε. This leads the solution of the minimization of (13) reduces to the deter-
to relatively simple expressions of the modulating signals. mination of the median (denoted med) of the series dmodε for
Due to the particular form of MS, the solutions that achieve all values of the design parameters.
the reference can be characterized by the single variable DN: The sum of all the εK gives the size (or number of elements)
of dmodε. In practice, two distinct cases are encountered: (1) the
CURRENT LOOP VOLTAGE REFS
series dmodε is of odd size: the optimal value of DN denoted
PARAMETERS
(OPTIONAL) VANref
VBNref
DSpref ε DUTY CYCLES DNopt is the midpoint of the series and is unique; (2) the series
Iref VCNref DS
Current
Reg. Control Allocation MODULATOR
dmodε is of even size: DNopt is a set of values of DN belonging to
3 3 4
Imeas 3
CONTROL PART
RELAY
1
the segment delimited by the two midpoints of the series.
S -
+ DS
0
0
Fig. 4 illustrates the optimal solutions depending on the na-
VAN PHYSICAL PART TRIANGULAR
VBN CARRIER ture of the size of dmodε. The criterion Jpref is a piecewise linear
Load
VCN
4-leg 2-level inverter
S
function due to the use of absolute values. The “even” case in
3 4 SWITCHING STATES
OUTPUT VOLTAGES Fig. 4 (a) is inconvenient because there is an infinite number of
Fig. 3. Diagram of the proposed control allocation method. solutions, and control is not uniquely determined. In practice, a
slight adjustment of the problem statement is sufficient to means DSpref = (0.5 0.5 0.5 0.5)T and ε = (1 1 1 1).
make the optimal solution unique, as shown in Fig. 4 (b). Determination of the median: As all the εK are identical, all
Finally, the median set of the series dmodε has to be con- the dmodK have the same occurrence in the series dmodε:

 
strained by the bounds DNmin and DNmax in order to obtain fea-
sible solutions. Therefore, optimal solutions of the constrained d mod   0.5  DAN ref , 0.5  DBN ref , 0.5  DCN ref , 0.5  
problem are the elements DNopt belonging to the intersection of
The size is even. Therefore, optimal solutions DNopt belong to
the median set of dmodε and of the segment [DNmin, DNmax].
the segment defined by the two midpoints of dmodε. In the case
D. Special choices of parameters where the reference system is three-phase, 0.5 is necessarily
The choice of parameters DSpref and ε has a direct and signif- one of the two midpoints of dmodε. The second midpoint is
icant effect on the nature of dmodε, on the optimal set of DN, and linked to ΔDmed. Consequently,

med d mod  0.5Dmed , 0.5 

on the resulting modulation strategy. Sets of parameters having  
interesting practical implications are given as examples.
The adjustment of DSpref translates into preferred duty cycle Remark: If ΔDmed is negative, the segment [0.5−ΔDmed, 0.5]
values of for each switch. The most common choices for DSpref has to be replaced by [0.5, 0.5−ΔDmed]. For simplicity, this
would be DSpref {(0…0)T, (1…1)T, (0.5…0.5)T}. trivial adjustment will be assumed to be made if necessary.
The first obvious choice of ε is a vector of ones, implying Optimal solutions: Finally, optimal solutions are any value
no discrimination between switches. More generally, to avoid belonging to [0.5−ΔDmed, 0.5] and saturated by [DNmin, DNmax].
non-symmetric behavior with respect to the three phases, the Fig. 5 gives an illustration of the domain of the DNopt wave-
first three weights should be identical. This yields the follow- forms. The bounds DNmin and DNmax are represented by the ex-
ing propositions for adjusting ε: {(1…1), (1 1 1 0), (0 0 0 1)}. tremal bold lines, delimiting the saturation domain. As the
reference amplitude increases, the saturation domain becomes
V. EXAMPLES & SIMULATION RESULTS narrower. The scaled value of 1/3 is the limit where for some
Specific solutions are derived and illustrated for the most in- instants =k/3, k{0,...,6}, the two extremal bold lines meet,
teresting choices of parameters DSpref and ε, as summarized in meaning that DNmin is equal to DNmax, see Fig. 5 (c).
Table I. Simulations are carried out in the MATLAB Simulink In Fig. 5, the two plain midlines draw the bounds of the op-
environment. Evaluations of switching losses, THD on voltag- timal domain of DN (hatched area). As the scaled amplitude
es and THD on currents are given later for comparison. increases, the median segment grows wider and eventually
For the following figures showing duty cycles, the scaled encounters the saturation domain (from (b)). The two lines
reference amplitude value increases from left (a) to right (c) labelled (1) and (2) correspond to the maximum and minimum
until the limit of 1/3 is reached. A whole fundamental period bounds of the nonsaturated optimal solution set, respectively.
Fig. 6 shows the optimal domain of the DA waveforms, de-
is illustrated. Reference voltages are chosen to be three-phase
termined from the optimal domain of DN by using (6). The
and balanced for simplicity, but the method also works with
fundamental wave is represented by the dashed gray midline
unbalanced references. As it will be useful for determining
(when visible). As for Fig. 5, the optimal domain of DA ob-
optimal solutions, we denote ΔDmed = med ΔDref the median of tained if DN’s is not saturated is represented by the lines la-
the scaled reference vector. belled (1) and (2). The parts of any optimal saturated DN mod-
A. Configuration 1 ulating signal that encounters the bounds DNmin or DNmax corre-
spond to periods where a leg is continuously clamped to the
Selected configuration: First of all, consider the natural
DC bus or to the ground, as observed on Fig. 6 (b)–(c).
choice where all preferred duty cycles are set to 0.5, which
When DN is not saturated, a certain regular waveform of DA
DIS TRIBUTIO N OF ARBITRARY dmod is identified, which stays proportional to the reference ampli-
DNmin DNmax
dmod B dmod C
dmod N
DN dmod A tude until saturation appears. Then, as a result of the satura-
DN tion, the optimal domain of DA changes form compared to
0.5
0
GENERAL TREND OF Jpref WITH ε=(111 1)
1 Fig. 6 (a) in order to achieve the correct fundamental wave.
Jpref This particular property is shared by most of modulation strat-
_ + egies obtained from the proposed control allocation method.
med dmodε TABLE I
Jpref,opt
DN CONFIGURATION EXAMPLES OF THE CONTROL ALLOCATION METHOD
0 dmod B dmod C dmod N dmod A 1
DNopt n° DSpref T ε Resulting PWM laws
GENERAL TREND OF Jpref WITH ε=(1211) 1 (0.5 0.5 0.5 0.5) (1 1 1 1) undetermined
Jpref
_ + 2 (0.5 0.5 0.5 0.5) (1 1 1 0) OMIPWM [25]
med dmodε 3 (0.5 0.5 0.5 0.5) (0 0 0 1) ASPWM
Jpref,opt
dmod C
DN 4 (1 1 1 1) (1 1 1 1) DPWMmax [1], [9], [11]
0 dmod B dmod N dmod A 1
DNopt ~ (0 0 0 0) (1 1 1 1) DPWMmin [1], [9], [11]
Fig. 4. Illustration of the effect of parameters on the solving of the simpli-
fied optimization problem with arbitrary ΔDref. DSpref = (0.5 0.5 0.5 0.5)T. OMIPWM: opposite median-voltage injection PWM; ASPWM: adaptive sinus PWM; DPWM: discon-
tinuous PWM.
 DNmax (1) DNmax (1)

DNmax max DNopt max DNopt = min DNopt

max DNopt

min DNopt
min DNopt
(2) DNmin
DNmin (2) DNmin

(a) Vref = 0.3EDC (b) Vref = 0.5EDC (c) Vref = 1/3EDC

Fig. 5. Optimal waveforms of DN with DSpref = (0.5 0.5 0.5 0.5)T, ε = (1 1 1 1). Labels (1) and (2) indicate nonsaturated lines for max, min DNopt, respectively.

(1)
max DAopt max DAopt (1)
out of (2) max DAopt = min DAopt
frame
(1)
(2) out of
frame
(1)
min DAopt min DAopt (2)
(2)

(a) Vref = 0.3EDC (b) Vref = 0.5EDC (c) Vref = 1/3EDC

Fig. 6. Optimal waveforms of DA with DSpref = (0.5 0.5 0.5 0.5)T, ε = (1 1 1 1). Labels (1) and (2) indicate nonsaturated lines for max, min DAopt, respectively.

B. Configuration 2 C. Configuration 3
Selected configuration: With the previous configuration, the Selected configuration: another possibility is to place a
optimal solution was not unique. In order to force the series to preference for the fourth duty cycle at 0.5, which means keep-
be of odd cardinal, a first possibility is to remove the prefer- ing the neutral potential nearest to half of the DC bus voltage:
ence on the fourth duty cycle that controls the neutral potential DSpref = (0.5 0.5 0.5 0.5)T and ε = (0 0 0 1).
while keeping the symmetry between the three phases. This Determination of the median: This time, only εN is non-null.
leads to DSpref = (0.5 0.5 0.5 0.5)T and ε = (1 1 1 0). dmodN = 0.5 is the only element of dmodε, thus, med dmodε = 0.5.
Determination of the median: dmodN does not appear in the Optimal solution: DNopt is unique and corresponds to the
series dmodε and, as the other εK are identical, value 0.5 saturated by [DNmin, DNmax].

 
Fig. 7 shows the resulting optimal waveform of DN. In fact,
d mod   0.5  DAN ref , 0.5  DBN ref , 0.5  DCN ref   most of the time, the optimal solution corresponds to SPWM,
as shown on Fig. 8. However, SPWM has a linearity range
Thus, the median is directly given by med dmodε = 0.5–ΔDmed.
limited to the scaled amplitude value of 0.5. As can be de-
Optimal solution: the optimal value DNopt is unique and is
duced from Fig. 7 (b), the constant DN solution value of 0.5
equal to 0.5–ΔDmed saturated by [DNmin, DNmax]. It corresponds
starts to be saturated by the bounds DNmin and DNmax. It gives
to parts of the optimal domain given by the previous configu-
the minimal adjustment of the neutral potential needed to
ration and farthest from 0.5 in Fig. 5. Thanks to the analytical
achieve the reference. A benefit of the optimal solution given
solution, a carrier-based equivalent method can be given. It
by configuration 3 is that, unlike for SPWM, the maximal limit
consists in injecting the opposite of the median reference in
of 1/3 is reached. It appears that this solution has not been
the waveform of DN as a zero-sequence signal. This new mod-
proposed before in the literature. The associated modulation
ulation method was recently proposed in [25] under the name
law is called here adaptive sinus PWM (ASPWM).
of opposite median-voltage injection PWM (OMIPWM). Un-
like SVM, where the zero-sequence signal always corresponds D. Configuration 4
to the value of the midpoint between DNmin and DNmax (see [6], Selected configuration: it is also possible to adjust DSpref in
[8]–[11], [16], [25]), DN will tend to reach the bounds DNmin order to specify a preferred value of duty cycle. For configura-
and DNmax for OMIPWM. From a certain reference amplitude tion 4, we set DSpref = (1 1 1 1)T and ε = (1 1 1 1).
value, discontinuous modulation appears with increasing dis- Determination of the median: the median of the series
continuous periods. This feature is highlighted in Fig. 5 (b)–(c) med dmodε is not reduced to a unique element. It corresponds to
at =/6+k/3, k{0,…,5}. OMIPWM is a new method, easy [1–ΔDmed, 1] and it is always out of the saturation domain.
to implement, which automatically ensures the transition from Optimal solution: With the selected vector DSpref, the opti-
a continuous to a discontinuous modulation scheme as a func- mal solution is DNmax, whatever the reference amplitude value
tion of the reference amplitude [25]. is (of course comprised between 0 and 1/3). This solution
 DNmax DNmax
DNopt
DNmax DNopt
DNopt

nonsaturated DNopt DNmin

DNmin DNmin

(a) Vref = 0.3EDC (b) Vref = 0.5EDC (c) Vref = 1/3EDC

Fig. 7. Optimal waveforms of DN with DSpref = (0.5 0.5 0.5 0.5)T, ε = (0 0 0 1), corresponding to ASPWM.

DAopt
DAopt DAopt
fundamental
= nonsaturated
DAopt

(a) Vref = 0.3EDC (b) Vref = 0.5EDC (c) Vref = 1/3EDC

Fig. 8. Optimal waveforms of DA with DSpref = (0.5 0.5 0.5 0.5)T, ε =(0 0 0 1), corresponding to ASPWM.
corresponds to the DPWMmax technique [1], [8]–[11], [16]. the DPWM’s. As the amplitude increases, the THD on cur-
The same reasoning can be carried out with a preferred duty rents for OMIPWM converges to DPWM THD’s, which have
cycle value of 0. This choice leads to the optimal solution dropped and are closer to the SVM THD’s. The same trend is
DNmin, which corresponds to the DPWMmin technique. found for ASPWM, but in a more mitigated manner, as
ASPWM remains a continuous PWM method most of the time.
E. Comparison of performances
OMIPWM presents better harmonic quality at low reference
Performance criteria of importance to power electronic con- amplitude and lower switching losses at high reference ampli-
verters were evaluated in simulations to gain a better under- tude. It constitutes an advantageous compromise between con-
standing of the differences between the various solutions. Sim- tinuous and discontinuous modulation [25]. ASPWM has the
ulation parameters are given in Table II. particularity of moving the neutral potential as little as possible
Basic estimation of switching-loss energies over the simula- from half the DC voltage, and moreover of not having an ef-
tion time were obtained by using simple parametric models fect on the neutral potential when not needed.
derived from curves available on datasheets. Only general
trends should be considered. Switching losses are shown in VI. EXTENSION TO MULTILEVEL, MULTILEG CONVERTERS
Fig. 9 (a). Two quasi-parallel lines, one higher than the other,
In this paper, we introduced a new control allocation meth-
describe the evolution of the switching losses, for SVM and
od for the voltage control of the 4-leg 2-level inverter. Howev-
for DPWMmin or DPWMmax. The trend for OMIPWM is
er, the general approach is also applicable to multilevel, multi-
remarkable since the curve follows the SVM curve, and then
leg converters, as long as one derives control equations linking
switches to the DPWM curve as a function of the amplitude
duty cycles of switching cells to desired outputs. As the nu-
[25]. It is also the case for ASPWM, but the transition occurs
merical optimization is based on the simplex algorithm, con-
for a much greater reference signal, because ASPWM is
trol equations must be linear (or linearized).
equivalent to SPWM until the scaled value of 0.5 is reached.
We already developed control allocation methods for the 3-
Fig. 9 (b) shows the evolution of the THD on the load volt-
phase multilevel flying capacitor inverter [31] and the 3-phase
age VAN. THD was evaluated using the function thd of the Sig-
modular multilevel inverter [32]. To ensure safe and efficient
nal Processing Toolbox in MATLAB. THD trends on voltages
operation of these converters, additional issues must be ad-
are very similar for all methods, but some differences are
dressed, like the active-balancing of quantities related to ener-
found for currents that highlight differences in the spectral
gy-storing elements, like capacitor voltages. Linear discrete
distributions, see in Fig. 9 (c). As for switching losses, only
control equations can be derived from differential equations of
general trends should be considered since the precise value of
capacitor voltages and/or output currents by sampling and
THD depends on the load impedances. The transition effect of
holding state values and by performing 1-order predictions.
OMIPWM is also observed with the THD on the currents. At
Then, a LP problem is formulated. The problem is updated and
low amplitudes, the THD for OMIPWM on the currents is
solved for each control period. As a result of the optimized
higher than SVM, but fairly close compared to the THD for
control, the method in [31] led to a fast active-balancing of
 (a) (b) (c)
Fig. 9. Performance evaluation in simulation for proposed configurations. (a) Switching losses. (b) THD on load voltage VAN. (c) THD on load current IA.

capacitor voltages. Finally, 4-leg inverters are already multileg ly for flight control. The control objective was thus formulated
inverters, but, in the same way, the approach is expected to be as the distribution of the 3 reference voltages to the available
also applicable to converters with more than 4 legs. redundant actuators, which consisted of the duty cycles of the
switching cells. The control allocation theory presents an al-
VII. ABOUT THE HARDWARE IMPLEMENTATION OF THE ternative solution to converter control taking a step back from
CONTROL ALLOCATION METHOD the known electronic or geometric viewpoints. The available
To achieve an efficient PWM operation, a high control fre- degree of freedom is used in an optimized manner as the volt-
quency is needed. Typical switching periods are around few age constraint set is fully taken into account. A major interest
hundreds of microseconds. FPGA implementation is a solution of this framework is its possible application to problems with
for complying with these strong time constraints, thanks to large number of switching cells [31], [32]. In the context of
short memory read/write access times and advantages of hard- this paper, however, where a small size, 4-leg 2-level inverter
wired logic like parallelism and pipelining. Up to now, the was considered, the main result was the development of an
problem of the FPGA implementation of the simplex algorithm approach providing a whole range of solutions giving a maxi-
has been rarely addressed, and the first implementation known mal extension of the inverter linearity range, all under the um-
to the authors was proposed in 2006 [33] with 18-bit data pre- brella of a single mathematical formulation. Investigations on
cision. It led to a gain in computing times up to 20 times those optimization parameters revealed several possible choices with
obtained with optimized commercial solvers on PC’s [33]. some corresponding to existing solutions, and others providing
We are currently working on an FPGA implementation of new approaches with distinct advantages, like OMIPWM [25]
the simplex algorithm dedicated to control allocation methods and ASPWM. An interesting feature of the analytic solutions
for power converters, using a Terasic TR4 development board obtained was that the resulting algorithms all reduced to carri-
(Altera Stratix IV FPGA core), with 32-bit floating-point data er-based equivalent PWM methods. Moreover, they are appli-
representation. A first validation of our implementation was cable to the classic 3-leg 2-level inverter connected to a three-
carried out with small-size arbitrary linear programs thanks to phase balanced load [25].
the ModelSim Altera software. For the control problem of the As the control allocation method optimizes the use of avail-
4-leg 2-level inverter, we estimated computation times around able resources with a general problem formulation, the ap-
50 µs, which is appropriate for switching frequencies up to proach seems promising for implementing a fault tolerant ca-
10 kHz. Depending on the control objectives, the delay in- pability. However, this feature was not studied in this paper as
duced by the computing time should be taken into account by the 4-leg 2-level inverter is not adequate for a meaningful in-
the control strategy. A solution is to perform a two-step predic- vestigation of redundancy.
tion of the system states, and optimal control solutions will be
applied only for the next control period. REFERENCES
[1] D. G. Holmes and T. A. Lipo, Pulse Width Modulation for Power Con-
VIII. CONCLUSIONS verters: principles and practice, New York: Wiley-IEEE Press, 2003.
[2] J. Holtz, “Pulsewidth Modulation for Electronic Power Conversion,”
In this paper, a new formulation of the control of a 4-leg 2- Proc. IEEE, vol. 82, no. 8, pp. 1194–1214, Aug. 1994.
level inverter was proposed. The problem was stated in terms [3] S. R. Bowes, “New sinusoidal pulsewidth-modulated invertor,” Proc.
Inst. Elect. Eng., vol. 122, pp. 1279–1285, Nov. 1975.
of an optimization problem similar to the control allocation [4] J. Holtz, P. Lammert and W. Lotzkat, “High-Speed Drive System with
problem that was solved earlier in other applications, especial- Ultrasonic MOSFET PWM Inverter and Single-Chip Microprocessor
Control,” IEEE Trans. Indus. Appl., vol. IA-23, no. 6, pp. 1010–1015,
TABLE II
Nov. 1987.
SIMULATION PARAMETERS
[5] H. W. Van der Broeck, H.-C. Skudelny and G. V. Stanke, “Analysis and
Symbol Meaning Values Realization of a Pulsewidth Modulator Based on Voltage Space Vec-
TS Switching period 100 µs tors,” IEEE Trans. Indus. Appl., vol. 24, no. 1, pp. 142–150, Jan.–Feb.
R, L Load parameters 0.5 Ω, 10 mH 1988.
f Fundamental frequency 50 Hz [6] D. G. Holmes, “The significance of zero space vector placement for
carrier-based PWM schemes,”, IEEE Trans. Indus. Appl., vol. 32, no. 5,
EDC DC bus voltage 400 V pp. 1122–1129, 1996.
[7] S. R. Bowes and Y.-S. Lai, “The relationship between space-vector [30] Z. Drezner and H. W. Hamacher, Facility Location: Applications and
modulation and regular-sampled PWM,” IEEE Trans. Indus. Electron., Theory. Springer Science & Business Media, 2004.
vol. 44, no. 5, pp. 670–679, Oct. 1997. [31] A. Bouarfa, M. Bodson, and M. Fadel, “A fast active-balancing method
[8] C. B. Jacobina, A. M. N. Lima, E. R. C. da Silva, R. N. C. Alves and P. for the 3-phase multi-level flying capacitor inverter derived from control
F. Seixas, “Digital scalar pulse-width modulation: a simple approach to allocation theory,” in 2017 20th IFAC World Congress, Toulouse,
introduce nonsinusoidal modulating waveforms,” IEEE Trans. Power France, July 2017.
Electron., vol. 16, no. 3, pp. 351–359, May 2001. [32] A. Bouarfa, M. Fadel, and M. Bodson, “A numerical optimization
[9] E. R. C. da Silva, E. C. dos Santos, and C. B. Jacobina, “Pulsewidth method using the simplex algorithm for control of modular multilevel
Modulation Strategies,” IEEE Indus. Electron. Mag., vol. 5, no. 2, pp. converters,” in ELECTRIMACS 2017, Toulouse, France, July 2017.
37–45, June 2011. [33] S. Bayliss, C. Bouganis, G. A. Constantinides, and W. Luk, “An FPGA
[10] J.-H. Kim and S.-K. Sul, “A Carrier-Based PWM Method for Three- implementation of the simplex algorithm,” in 2006 IEEE International
Phase Four-Leg Voltage Source Converters,” IEEE Trans. Power Elec- Conference on Field Programmable Technology, 2006, pp. 49–56.
tron., vol. 19, no. 1, pp. 66–75, Jan. 2004.
[11] N.-Y. Dai, M.-C. Wong, F. Ng and Y.-D. Han, “A FPGA-Based Gener-
alized Pulse Width Modulator for Three-Leg Center-Split and Four-Leg Abdelkader Bouarfa was born near Paris, France, in
Voltage Source Inverters,” IEEE Trans. Power Electron., vol. 23, no. 3, 1991. He received the degree of engineer in electrical engi-
pp. 1472–1484, May 2008. neering and automation from the Ecole Nationale Supéri-
[12] X. Li, Z. Deng, Z. Chen and Q. Fei, “Analysis and Simplification of eure d'Electrotechnique, d'Electronique, d'Informatique,
Three-Dimensional Space Vector PWM for Three-Phase Four-Leg In- d'Hydraulique et des Télécommunications (ENSEEIHT),
verters,” IEEE Trans. Indus. Electron., vol. 58, no. 2, pp. 450–464, Feb. engineer school member of the National Polytechnic Insti-
2011. tute of Toulouse (INPT), Toulouse, France, in 2014.
[13] G. Buja and G. Indri, “Improvement of Pulse Width Modulation Tech- He is currently preparing a Ph.D. thesis in control of
niques,” Archiv für Elektrotechnik, vol. 57, pp. 281–289, 1975. power converters at the Laboratory of Plasma and Energy Conversion,
[14] J. A. Houldsworth and D. A. Grant, “The Use of Harmonic Distortion to LAPLACE, Université de Toulouse, mixed unit of research CNRS-INPT-
Increase the Output Voltage of a Three-Phase PWM Inverter,” IEEE UPS. His research interests encompass control of energy conversion systems.
Trans. Indus. Appl., vol. IA–20, no. 5, pp. 1224–1228, Sept.–Oct. 1984.
[15] S. Ogasawara, H. Akagi, and A. Nabae, “A novel PWM scheme of
voltage source inverters based on space vector theory,” Archiv für El- Maurice Fadel (M’09) was born in Toulouse, France. He
ektrotechnik, vol. 74, no. 1, pp. 33–41, Jan. 1990. got the PhD degree at the Institut National Polytechnique
[16] A. M. Hava, R. J. Kerkman, and T. A. Lipo, “A high-performance gen- de Toulouse in 1988, in the domain of the Control in Elec-
eralized discontinuous PWM algorithm,” IEEE Transactions on Indus- tric Engineering.
try Applications, vol. 34, no. 5, pp. 1059–1071, Sep. 1998. He is currently a Professor in the Ecole Nationale Supé-
[17] J. W. Kolar, H. Ertl and F. C. Zach, “Minimizing the current harmonics rieure d'Electrotechnique, d'Electronique, d'Informatique,
RMS value of three-phase PWM converter systems by optimal and d'Hydraulique et des Télécommunications of Toulouse
suboptimal transition between continuous and discontinuous modula- (ENSEEIHT). In 1985 he as integrated the Laboratory of
tion,” in 22nd Annual IEEE PESC, 1991, pp. 372–381. Electrotechnics and Industrial electronics (LEEI), mixed
[18] J. Holtz and B. Beyer, “Optimal pulsewidth modulation for AC servos unit of research (CNRS-INPT-Univ. Toulouse III). He was leading of the
and low-cost industrial drives,” IEEE Trans. Indus. Appl., vol. 30, no. 4, LEEI laboratory in 2005. From January 2007 until December 2015 he was
pp. 1039–1047, Jul. 1994. Deputy Director of the Laboratoire Plasma et Conversion d’Energie (Labora-
[19] R. Zhang, V. H. Prasad, D. Boroyevich and F. C. Lee, “Three- tory of Plasma and Energy Conversion, LAPLACE). This laboratory count
Dimensional Space Vector Modulation for Four-Leg Voltage-Source about hundred permanent researchers, more than hundred PhD students cov-
Converters,” IEEE Trans. Power Electron., vol. 17, no. 3, pp. 314–326, ering the continuum of specialty, materials, plasmas and systems to the ser-
May 2002. vice of the conversion and the treatment of the electric energy. The field of
[20] O. Ojo and P. M. Kshirsagar, “Concise Modulation Strategies for Four- scientific interest of Pr. Maurice FADEL concerns the modeling and the con-
Leg Voltage Source Inverters,” IEEE Trans. Power Electron., vol. 19, trol of the electric systems more especially of the synchronous machine, the
no. 1, pp. 46–53, Jan. 2004. control law of the static converters with the help of direct predictive controls
[21] M. A. Perales, M. M. Prats, R. Portillo, J. L. Mora, J. I. Leon and L. G. approach and the definition of control strategies for cooperative systems.
Franquelo, “Three-Dimensional Space Vector Modulation in abc Coor-
dinates for Four-Leg Voltage Source Converters,” IEEE Power Elec-
tron. Letters, vol. 1, no. 4, pp. 104–109, Dec. 2003. Marc Bodson (F’06) received a Ph.D. degree in Electrical
[22] M. Bodson, “Evaluation of Optimization Methods for Control Alloca- Engineering and Computer Science from the University of
tion,” Journal of Guidance, Control and Dynamics, vol. 25, no. 4, pp. California, Berkeley, in 1986. He obtained two M.S. de-
703–711, July–Aug. 2002. grees - one in Electrical Engineering and Computer Sci-
[23] T. A. Johansen and T. I. Fossen, “Control Allocation–A survey,” Auto- ence and the other in Aeronautics and Astronautics - from
matica, vol. 49, pp. 1087–1103, 2013. the Massachusetts Institute of Technology, Cambridge MA,
[24] A. Bouarfa, M. Fadel, M. Bodson and J. Lin, “A New Control Alloca- in 1982. In 1980, he received the degree of Ingénieur Civil
tion Method for Power Converters and its Application to the Four-Leg Mécanicien et Electricien from the Université Libre de
Two-Level Inverter,” in IEEE 23rd Mediterranean Conf. Control Autom. Bruxelles, Belgium. His research interests are in adaptive control, with appli-
(MED), Torremolinos, Spain, pp. 1020–1026, 16–19 June 2015. cations to electromechanical systems and aerospace.
[25] A. Bouarfa, M. Fadel, and M. Bodson, “A new PWM method for a 3- Currently, he is a Professor of Electrical & Computer Engineering at the
phase 4-leg inverter based on the injection of the opposite median refer- University of Utah in Salt Lake City. He was Chair of the department between
ence voltage,” in 2016 Int. Symp. Power Electron. Electr. Drives Au- 2003 and 2009, and he was an Assistant Professor and an Associate Professor
tom. Motion (SPEEDAM), Anacapri, Italy, pp. 791–796, June 2016. at Carnegie Mellon University, Pittsburgh, PA, between 1987 and 1993. He
[26] C. G. Small, “A Survey of Multidimensional Medians,” International was a Belgian American Educational Foundation Fellow in 1980 and a Lady
Statistical Review, vol. 58, no. 3, pp. 263–277, 1990. Davis Fellow at the Technion, Haifa, Israel, in 1990. He is coauthor, with S.
[27] J. A. Hanley, L. Joseph, R. W. Platt, M. K. Chung, and P. Belisle, “Vis- Sastry, of the book Adaptive Control: Stability, Convergence, and Robust-
ualizing the Median as the Minimum-Deviation Location,” The Ameri- ness, published by Prentice-Hall in 1989 and reprinted by Dover in 2011.
can Statistician, vol. 55, no. 2, pp. 150–152, May 2001. Professor Bodson was the Editor-in-Chief of IEEE Trans. on Control Sys-
[28] D. W. Stroock, Probability Theory: An Analytic View. Cambridge Uni- tems Technology from 2000 to 2003. He was elected Fellow of the IEEE in
versity Press, 2010. 2006, and Associate Fellow of the American Institute of Aeronautics and
[29] J.-B. Hiriart-Urruty and C. Lemaréchal, Fundamentals of Convex Anal- Astronautics in 2013. He received the Engineering Educator of the Year
ysis. Berlin, Heidelberg: Springer Berlin Heidelberg, 2001. award from the Utah Engineers Council in 2007.