Basic Principles

of MPC for Power

Converters

©istockphoto.com/ayzek

Bridging the Gap

Between Theory and Practice

CARLOS BORDONS and CARLOS MONTERO

T
his article presents the for power converters [continuous- proportional-integral controller com-
basic principles of op- control-set MPC (CCS-MPC) and fi- bined with pulse-width modulation
eration for model pre- nite-control-set MPC (FCS-MPC)] (PWM) to newer control methods, such
dictive control (MPC), are described, and their application as fuzzy, sliding mode, or adaptive
a control methodology to a voltage-source inverter (VSI) is control, which need much more cal-
that opens a new world shown to illustrate their capabilities. culation power. Recently, MPC has
of opportunities. MPC This article tries to bridge the gap attracted the interest of the research
is a powerful technique that can ful- between the powerful but sometimes and academic communities, thanks to
fill the increased performance and abstract techniques developed by re- its optimization capabilities. MPC is a
higher efficiency demands of power searchers in the control community well-known control strategy that has
converters today. The main features and the empirical approach of power been used in several fields of engin-
of this technique are presented as electronics practitioners. eering. Although MPC was born in the
well as the MPC strategy and basic el- The control of power converters framework of industrial process con-
ements. The two main MPC methods and drives has become a relevant trol [1], in recent years, it has been ap-
topic in the last decade. The existing plied to a wide variety of fields such as
Digital Object Identifier 10.1109/MIE.2014.2356600 control methodologies range from energy, bioengineering, robotics, and
Date of publication: 21 September 2015 those based on the conventional aerospace [2]. Today, its application to

1932-4529/15©2015IEEE september 2015  ■  IEEE industrial electronics magazine  31

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on March 25,2020 at 19:51:01 UTC from IEEE Xplore. Restrictions apply.
 MPC presents a series of advantag-
es over other methods, among which
MPC has attracted the interest of the research the following stand out:
and academic communities, thanks to its ■■ The concepts are very intuitive, and, at
the same time, tuning is relatively easy.
optimization capabilities. ■■ It can be used to control a great variety
of processes, from those with relative-
ly simple dynamics to more complex
electronic power converters and elec- are continuous ones. Despite this com- ones, including nonlinear systems,
trical drives is increasing [3], [4] since plexity, one important feature of power nonminimum phase, or unstable ones.
it can help fulfill the growing perform- converters is that the models are avail- ■■ The appropriate choice of the cost
ance, efficiency, and safety demands able, accurate, and often relatively well function may achieve the optimi-
and meet the standards and operation- known. This allows the predictive con- zation of different performance
al limits demanded by the evolving in- troller to take into account the complex criteria, including operational con-
dustry applications. dynamics while several design criteria straints in the design process.
The term MPC does not designate and constraints are considered to com- ■■ It introduces feedforward control
a specific control strategy but rather pute the best control action at every in a natural way to compensate for
an ample range of control methods sampling time. measurable disturbances.
that makes explicit use of a model of The objective of this article is to pre- ■■ The resulting controller (once com-
the process to obtain the control sig- sent the main features of MPC in a sim- puted) is an easy-to-implement con-
nal by minimizing an objective func- ple and intuitive way and to show how trol law.
tion. The MPC paradigm is based on it can be applied in power electronics. ■■ Its extension to the multivariable
selecting the best among all feasible The two wide categories of MPC used in case (multiple-input, multiple-­output
input sequences over a future horizon power electronics (CCS- and FCS-MPC) systems) is conceptually simple.
according to some criterion. The first are described, and the way to imple- ■■ It is very useful when future references
input of this sequence is applied to the ment them in the unconstrained case (robotics or converters) are known.
plant, and the scheme is repeated in a is presented. To illustrate the concepts As is logical, however, MPC also has
receding horizon fashion at every sam- and the implementation, some experi- its drawbacks. The greatest one is that
pling time when new state information ments on a VSI are included. its derivation is more complex than that
is available. of classical controllers. If the process
MPC presents several features that Basic Principles of MPC dynamic is known and does not change,
make it suitable for controlling power The term MPC designates an ample range as is the general case in power convert-
converters. Apart from being intuitive of controllers, which make use of a model ers, the derivation of the controller can
and easy to understand, constraints, of the plant to obtain the control action be done beforehand, provided that the
nonlinearities, and the multivariable by minimizing an objective function. The tuning parameters do not change. The
case can easily be included in the for- basic principles of predictive controllers solution of the optimization problems
mulation. However, since an open-loop are described in the following. is given by an explicit control law that
optimal problem is solved at each sam- is easy to implement and requires lit-
pling instant, the computational cost is Main Features tle computation. But when constraints
high compared to ordinary linear con- The ideas, appearing in greater or are considered, the computation has to
trol schemes. This point is of crucial lesser degree in the predictive control be carried out at every sampling time.
importance in the case of power con- family, are basically: The price paid by the use of a predict-
verters; therefore, different varieties of ■■ explicit use of a model to predict ive controller is the large amount of
MPC have been proposed in the litera- the process output at future time calculations required, especially in the
ture to cope with this problem [5]. The instants (horizon) constrained case or when using long
advent of powerful microprocessors ■■ calculation of a control sequence horizons, since an optimization prob-
with increased computational capabili- minimizing an objective function lem is being solved at every sampling
ties is enabling its application in this ■■ receding strategy, so that at each in- time. However, digital signal processing
field with significant success. stant the horizon is displaced toward has experienced an explosive evolution
MPC uses a mathematical model of the future, which involves the appli- in recent decades, and the advent of
the plant that allows the controller to cation of the first control signal of the immensely powerful microprocessors
predict the impact of its control actions. sequence calculated at each step. with increased computational capabil-
Power converters are nonlinear systems The various MPC algorithms only ities is enabling its successful applica-
exhibiting hybrid behavior since the differ among themselves in the model tion in the field of power electronics.
inputs are discrete signals that com- used to represent the plant, the cost Another important issue is the need
mand the transitions of the switches, function to be minimized, and the way for an appropriate model of the process
and the outputs (currents, voltages) the optimization is performed. to be available. The design algorithm is

32  IEEE industrial electronics magazine  ■  september 2015

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on March 25,2020 at 19:51:01 UTC from IEEE Xplore. Restrictions apply.
based on prior knowledge of the model
and is independent of it, but it is obvi- u (t + k|t)
ous that the benefits obtained will be
affected by the discrepancies exist- u(t)
ing between the real process and the
model used. This is not a big problem
in power electronics, where the math- ^ + k|t)
y(t
ematical model of the plant is often y(t)
available and relatively well known, but N
it can be of crucial importance in other
fields such as the process industry.

t –1 t t +1 ... t+k ... t+N

MPC Strategy
The methodology of all controllers
FIGURE 1 – The MPC strategy.
belonging to the MPC family is char-
acterized by the following strategy,
represented in Figure 1: u (t + 1 ; t + 1) is calculated [which, instant, and the procedure is repeated
1) The future outputs for a determined in principle, will be different from for the next control decision in a re-
horizon N, called the prediction hori- the u (t + 1 ; t) because of the new ceding horizon fashion. Different opti-
zon, are predicted at each sampling information available] using the re- mization criteria are possible, leading
instant t using the dynamic model of ceding horizon concept. to different results. For example, if the
the plant. These predicted outputs This strategy is implemented using driver desires the shortest duration of
y (t + k ; t) (the notation indicates the basic structure shown in Figure 2. the trip, he will act more rapidly than
the value of the variable at the in- A model is used to predict the future if fuel consumption is to be minimized.
stant t + k calculated at the current plant outputs based on past and cur-
instant t) for k = 1f N depend rent values and on the proposed op- Basic Elements
on the known values up to instant timal future control actions. These This section describes the elements
t (past inputs and outputs) and on actions are calculated by the optimiz- that are common to all MPCs, show-
the future control signals u (t + k ; t), er taking into account the cost func- ing the various alternatives used in
k = 0 f N - 1, which are those to tion as well as the constraints. different implementations. All MPC al-
be computed and sent to the system. Notice that the MPC strategy is very gorithms possess common elements,
2) The set of future control signals is similar to the control strategy used in and different options can be chosen
calculated by optimizing a deter- driving a car (see Figure 3). The driver for each element, giving rise to differ-
mined criterion, which, in general, knows the desired reference trajec- ent algorithms. The following sections
will try to keep the output as close as tory for a finite control horizon and, describe these elements.
possible to the reference trajectory by taking into account the car char-
w (t + k) (which can be the setpoint acteristics (mental model of the car), Model
itself or a close approximation of it). decides which control actions (accel- The model is the cornerstone of MPC.
This criterion usually takes the form erator, brakes, and steering) to take to Consequently, the plant model plays a
of a quadratic function of the errors track the desired trajectory. Only the decisive role in the controller. The cho-
between the predicted output signal first control actions are taken at each sen model must be able to capture the
and the predicted reference trajec-
tory, and it can include the neces-
sary control effort. Although the Reference
Euclidian norm is the most used, the Trajectory
first or the infinity norms can also be
considered in the cost function, as State (Past Predicted
Inputs and Outputs +
will be shown in detail in the “Basic
Outputs) Model
Elements” section. –
3) The control signal u (t ; t) is sent to
Future Inputs
the process while the next control (Control Actions)
signals calculated are rejected be-
Optimizer
cause, at the next sampling instant, Future Errors
y (t + 1) is already known (feedback
action). Step 1 is repeated with this
Cost Function Constraints
new value, and all of the sequences
are brought up to date. Thus, the FIGURE 2 – The basic structure of MPC.

september 2015  ■  IEEE industrial electronics magazine  33

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on March 25,2020 at 19:51:01 UTC from IEEE Xplore. Restrictions apply.
 where U is a nonlinear mapping, y is
Future the output, u is the input, and e is the
Past
noise input. The integer values n y, n u,
and n e are the number of past values of
outputs, inputs, and noise, respectively,
t–2 t–1 t ... t+N
that are considered in the regression.
The suitability of this model depends
FIGURE 3 – An MPC analogy.
on the choice of the function U and
the order parameters. Notice that this
process dynamics to precisely predict of stability and robustness criteria. It equation covers a wide range of de-
the future outputs and be simple to im- can also be extended to deal with non- scriptions, depending mainly on func-
plement and understand. The use of the linear systems. In the linear case, it tion U. Volterra and related models
plant model is determined by the neces- has the following representation: (Hammerstein and Wiener), local mod-
sity to calculate the predicted output at x (t) = Ax (t - 1) + Bu (t - 1), el networks, or neural networks can be
future instants yt (t + k ; t) . The different  used to capture the nonlinearities.
y (t) = Cx (t), (2)
strategies of MPC can use various mod- A typical characteristic of most
els to represent the relationship between where x (t) is the state (column vec- MPC for linear systems is the use of
the outputs and the measurable inputs, tor of dimension n, being the system free and forced response concepts. The
some of which are manipulated vari- order) and A, B, and C are the ma- idea is to express the control sequence
ables and others that can be considered trices of the system, input, and output, as the addition of the two signals:
measurable disturbances, which can be respectively. This type of model repre- u (t) = u f (t) + u c (t) . The signal u f (t)
compensated for by feedforward action. sentation is very useful in power elec- corresponds to the past inputs and is
As MPC is not a unique technique but tronics since it can be directly derived kept constant and equal to the last val-
rather a set of different methodologies, by discretization of the differential ue of the manipulated variable in future
there are many types of models used in equations that describe the converter time instants. The signal u c (t) is made
various formulations. For linear plants, dynamics. In many occasions, the pow- equal to zero in the past and equal to
the transfer function model is, perhaps, er converter is a multivariable system, the next control moves in the future.
the most widespread in the academic with m inputs and p outputs. Notice The prediction of the output se-
community and is used in most control that (2) can still be used, but, in this quence is separated into two parts, as
design methods as it is a representation case, y (t) is a column vector of dimen- can be seen in Figure 4. One of them
that requires only a few parameters and sion p and u (t) is of dimension m. ^ y f (t) h, the free response, corresponds
is valid for all kinds of processes. The Nonlinear plants can be repre- to the prediction of the output when the
transfer function G = B/A is used so sented as a nonlinear autoregressive process-manipulated variable is made
that the discrete-time output is given by moving average model with exog- equal to u f (t); and the other, the forced
enous input (narmax), which can be response ^ y c (t)h, corresponds to the
A (z -1) y (t) = B (z -1) u (t), (1) considered an extension of the trans- prediction of the process output when
fer function concept. For single-input, the control sequence is made equal to
A (z -1) and B (z -1) being polynomials single-output processes, it is given by u c (t) . The free response corresponds
in the backward shift operator z -1 . the general equation to the evolution of the process due to its
The state-space model is also used present state, while the forced response
in some formulations, as it can eas- y (t) = U [y (t - 1), f, y (t - n y), is due to future control moves. For lin-
ily describe multivariable processes u (t - 1), f, u (t - n u), ear systems, the entire future system
and it allows for an easier expression e (t), f, e (t - n e + 1)], (3) behavior can be calculated as the sum
of the free and forced response using
the superposition principle. This simple
concept will be very useful when per-
u y
Process forming the optimization since the con-
troller will compute the future control
t t actions u c (t + j) so that, knowing the
given free response, the system output
will track the reference.
uf uc yf yc
This concept can be used indepen-
+ + dently of the type of model employed
to describe the system’s dynamics.
t t t t In case a transfer function model is
used, the vector y of future output
FIGURE 4 – The free and forced responses. predictions along the horizon can be

34  IEEE industrial electronics magazine  ■  september 2015

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on March 25,2020 at 19:51:01 UTC from IEEE Xplore. Restrictions apply.
expressed as a function of the vector
of future control signals u, the dynam-
ic matrix G, and the vector of future
Long prediction horizons lead to a significant
free responses f as performance improvement at steady-state operating
y = Gu + f. (4) conditions, lowering the current distortions and/or
If a state-space model is used, the pre- the switching frequency.
diction also depends on the state, which
must be estimated when the state vec-
tor is not accessible by means of an
observer. Notice that the state-space desirable for the output to track the ref- refer to the seminal paper [8] for fur-
model given by (2) can be used for both erence. Thus, if a high value of N 1 is tak- ther details. For motor drives, it has
monovariable and multivariable sys- en, it is because it is of no importance been shown that long prediction hori-
tems. In this situation, the prediction if there are errors in the first instants. zons lead to a significant performance
can be expressed in vector form as This will originate a smooth response improvement at steady-state operating
of the process. Also, if the process is conditions, lowering the current distor-
y = Hu + Fxt (t) . (5) nonminimum phase, this parameter tions and/or the switching frequency
will allow the first instants of inverse [9]. By using long prediction horizons,
The calculation of G, H, f, and F response to be eliminated from the the closed-loop stability margin and the
can be found in [2]. Notice that, inde- objective function. In the simple case performance are enhanced. However,
pendently of the type of model used, that N 1 = N 2 = 1, the optimization is the price to pay is an increase in the
predictions can always be separated performed only for the next sampling computational complexity.
in two terms: one depending on the instant. The coefficient m (j) is a weight-
decision variable (u) and one that ing sequence that considers the future Optimization Procedure
is independent of the control actions behavior; usually, constant values or The optimizer is another fundamental
taken along the horizon. exponential sequences are considered. part of the strategy as it provides the
A different norm can be used to meas- control actions. If the cost function is
Objective Function ure the error value. In (6), a quadratic quadratic, its minimum can be obtained.
The various MPC algorithms propose (Euclidean) norm is used, but the 1-norm The optimization procedure consists of
different cost functions for obtain- (absolute value) or 3-norm (maximum finding the minimum of the cost function
ing the control law. The general aim value) is considered in other formula- (generally quadratic), which can be done
is that the future output ^ y h on the tions. The Euclidean norm produces an in two ways. First, if the control set is
considered horizon should track a de- overproportional cost (in powers of two) continuous, the minimum, in the absence
termined reference signal ^w h, and, at compared to the 1-norm, giving a higher of constraints, is obtained as an explicit
the same time, the control effort ^Tu h penalization of bigger errors compared (linear) function of past inputs and out-
necessary for doing so should be pe- with smaller ones. This can be used to puts and the future reference trajectory
nalized. The general expression for control variables closer to the reference by making the gradient of the cost func-
such an objective function will be and reduce the ripple amplitude. The tion equal to zero. Second, if the control
3-norm can be used when uncertain- set is finite, then the optimization can be
N2
J (N 1, N 2, N u) = / [yt (t + j ; t) - w (t + j)] 2 ties exist, so the control problem can be done by evaluating the cost function for
j = N1 solved for the worst case. all possible values of the inputs (which
Nu
The cost-function formulation can be are the states of the switches). This last
+ / m (j) [Tu (t + j - 1)] 2 .
(6)
j=1 very general, allowing the optimization procedure is conceptually simple, but
of several important parameters such it leads to a large amount of calculation
In some methods, the second term, as the number of switches, switching when the horizon is long.
which considers the control effort, is losses, reactive power control, and mo- For the CCS case, the optimization
not taken into account, and in some, tor torque ripple minimization. Several is done by imposing that the gradient
the values of the control signal (not examples of different cost functions for of the cost function with respect to
its increments) appear directly. The power converters can be found in [3], the decision variable is equal to zero.
parameters in the cost function are [6], and [7]. That is, (2J) / (2u) = 0.
that N 1 and N 2 are the minimum and One important parameter in the ob- Since the prediction of the outputs
maximum prediction horizons and N u jective function is the maximum pre- is done using the free/forced response
is the control horizon, which does not diction horizon. Generally speaking, a concept, the cost function (6) only de-
necessarily have to coincide with the long prediction horizon is required for pends on u, and it can be written in
maximum horizon. The meaning of N 1 improved stability and plant perform- vector form as
and N 2 is rather intuitive; they mark ance. Although stability of MPC is not
the limits of the instants in which it is addressed in this article, readers can J = (y - w) T (y - w) + mu T u. (7)

september 2015  ■  IEEE industrial electronics magazine  35

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on March 25,2020 at 19:51:01 UTC from IEEE Xplore. Restrictions apply.
In the transfer function case, using the on the variables of concern, the plant Notice that the cost function is a parab-
value of the prediction, this function is able to operate at its physical limits oloid and the constraints are given by
takes the form without violating them. Thus, the most planes. In the unconstrained case, the
favorable operation can be obtained minimum is located at the point in the
J = (Gu + f - w) T (Gu + f - w) while the operational limits of the plant curve where the derivative equals zero.

+ mu T u,  (8) are fully respected. Normally, bounds Many methods have been proposed
in the amplitude and in the slew rate of to reduce the computation time of solv-
whose derivation gives rise to the the control signal as well as limits in the ing a QP problem. It has been shown that
vector of future control actions along output will be considered: MPC for these cases can be considered
the horizon a multiparametric quadratic [10] or LP
problem [11] and that the MPC solution
u min # u (t) # u max,
u = (G T G + mI) -1 G T (w - f). (9) turns out to be a relatively easy-to-imple-
du min # u (t) - u (t - 1) # du max, 
ment piecewise affine controller. The so-
y min # y (t) # y max 6t . (12) lution can be precomputed offline for the
In the state-space case, the cost func-
tion can be written as space of all possible states, and the imple-
By adding these linear constraints to mentation consists of searching for the
J = (Hu + Fxt (t) - w) T
 the objective function, the minimization right solution depending on the system
(Hu + Fxt (t) - w) + mu T u. (10)
becomes more complex. So the solu- state at every sampling time. However,
If the are no constraints, an analytical solu- tion cannot be obtained explicitly as the obtained explicit solution may be ex-
tion exists that provides the optimum as in the unconstrained case. It is a quad- cessively complex for medium- to large-
ratic programming (QP) problem whose scale systems, in which case approximate
u = (H T H + mI) -1 H T (w - Fxt (t)) . (11) solution has to be obtained by more explicit solutions can be computed. An
computationally taxing numerical algo- application of this technique to a PWM
Notice that the solution for different rithms. If the cost function is a 1-norm inverter with an inductor–capacitor–in-
types of models (transfer function and or 3-norm type, MPC results in a linear ductor (LCL) filter is presented in [12].
state space), given by (9) and (11), is programming (LP) problem. The size of The authors propose alternative (piece-
similar. For multivariable systems, the the optimization problem depends on wise affine) models that account for
state-space model of (2) can be directly the number of variables and the pre- the switched behavior of the converter.
used, and, therefore, the computation diction horizons used, and it usually Based on these improved models, an ex-
of the control signal is as given before. turns out to be a relatively modest opti- plicit MPC scheme is derived to provide a
In this case, the number of decision mization problem that does not require fast response, making it very suitable for
variables is larger than in the mono- solving sophisticated computer codes. applications such as active filtering where
variable case since there are as many However, the amount of time needed a large bandwidth is required.
control actions as the control horizon for the constrained case can be various When the explicit solution is not ap-
multiplied by the dimension of u (t) . orders of magnitude higher than for the propriate, online optimization methods
An important feature of MPC is its unconstrained case. Figure 5 shows the can be used. Ritcher et al. [13] present
capability of considering constraints in optimization in the case of two decision a practical implementation of the fast
the optimization. By imposing constraints variables (control horizon equal to two). gradient method for the control of an

× 105
× 105
Constraint
10 10
9
8 8
Cost Function 7 Cost Function
6 6
5
J

4 4
3
2 2
1
0 100 0
0 0 100
20 20 80
40 50 40 60
60 60 40 u (t + 1)
80 u (t + 1) 80 20
u(t ) 100 0 u (t ) 100 0
(a) (b)

FIGURE 5 – The minimization of a quadratic function: (a) the unconstrained case and (b) the constrained case.

36  IEEE industrial electronics magazine  ■  september 2015

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on March 25,2020 at 19:51:01 UTC from IEEE Xplore. Restrictions apply.
ac–dc power converter, with compu-
tation times as small as a few tens of
­microseconds, making the approach ide-
The PWM block not only controls but also avoids the
al for power electronics applications. An use of a switching model, allows the fixed frequency
application of real-time optimization for
boost converters can be found in [14]. operation, and guarantees a decoupling between the
This method applies the results of re- switching and sampling time.
cently developed hybrid optimal MPC to
determine switching through real-time
minimization of a user-defined perfor- is used to calculate the required ref- the corresponding system response is
mance index. Similar results for a Cuk erence voltage to reach the desired predicted, the cost function is evalu-
converter is shown in [15]. If the model reference value for a certain variable ated, and, consequently, the sequence
is nonlinear, the optimization is even (usually the current). The predicted that yields the minimal cost is select-
more costly, giving rise to a nonlinear reference voltage is later generated by ed. This approach is known as FCS-
programming problem. An interesting the converter via a modulation stage. MPC [19] since the possible control
application of nonlinear MPC to a dc–dc This scheme has been applied in sev- actions (switching states) are finite.
converter can be found in [16]. eral power electronics applications Notice that a modulator is not needed.
One conceptually simple way of solv- such as rectifiers, inverters, active fil- This method has been successfully ap-
ing the optimization problem is the so- ters, and power supplies [17]. plied to a wide range of power convert-
called direct MPC or finite-state MPC, The detailed classifications of predic- ers and drive applications [6], [20].
which is used in many applications of tive controllers subdivided into several The methodology is simple and intui-
power electronics. The optimization categories, depending on the operating tive, as shown in Figure 6, where each of
problem is solved using an exhaus- principle and other characteristics, can the M admissible switching actions S i
tive search, where the set of switching be found in [3] and [18]. However, not all gives rise to a different value of the sys-
sequences is enumerated, the system of those methods are part of the family tem state in the next sampling instant
output is predicted, the cost function of MPCs since they do not make use of ^ xt i (t + 1) h . If the horizon is bigger than
is evaluated for each sequence, and the features described in the “Basic Prin- one, the procedure must be replicated
the switching sequence that yields the ciples of MPC” section. for each value of xt i (t + 1); that is, from
minimal cost is chosen as the optimal One partition of MPC methods can this state, each of the M switching ac-
one. This is a straightforward way of be made according to the type of control tions will lead to a new value of the state
computing the control action, which is principle: those that control the inverter in t + 2 and so on. This strategy is very
not computationally expensive for short directly (calculate the optimal switching intuitive and easy to implement, pro-
horizons. In the following section, this states) and those strategies that deter- vided the horizon is small, because it
method and other MPC methods ap- mine a continuous control signal that is basically consists of evaluating the cost
plied to power converters are reviewed. synthesized by a modulator. Notice that, function for each possible sequence.
in the first case, the problem to be solved However, this approach for imple-
MPC Methods for combines continuous variables (output menting MPC for power converters is
Power Converters voltages and currents) with discrete currently limited to short horizons (usu-
Over the years, many control strategies variables (those that take only discrete ally one or two) because the solution is
have been proposed for power convert- values, such as switches), giving rise to a obtained by evaluating a cost function
ers that are mainly based on the classi- mixed integer program that can be com- in a kind of exhaustive search. The op-
cal proportional-integral-derivative (PID) putationally expensive, especially for long timization problem can be very time
controller combined with PWM. The horizons. Taking into account the MPC consuming, except for short horizons. In
PWM block not only controls but also paradigm presented in this article, two fact, as the prediction horizon is enlarged
avoids the use of a switching model, al- main branches can be considered and will and the number of decision variables is
lows the fixed frequency operation, and be analyzed below: 1) direct or FCS-MPC increased, the computational complexity
guarantees a decoupling between the and 2) CCS-MPC. grows exponentially and, thus, cannot be
switching and sampling time. bounded by a polynomial (see also [21]).
There are several methods that Direct or FCS-MPC Despite the good performance that
use a model of the plant to compute Since power converters have a finite FCS-MPC can offer, there remain several
the control action, such as hysteresis- number of switching states, the MPC open problems such as the lack of stabil-
based control, sliding-mode control, or optimization problem can be simplified ity guarantees. In the context of MPC, the
dead-beat control, which is considered and reduced to the prediction of the infinite-horizon case, in general, ensures
by many researchers to be the earliest system behavior only for those pos- closed-loop stability provided that a solu-
predictive controller. Dead-beat elimi- sible switching states. Then, at each tion with a finite cost exists [22]. However,
nates the classic linear controller by sampling instant, the set of admissible in power electronics, short horizons (usu-
using a model of the system, which switching sequences is enumerated, ally equal to one) are preferred because

september 2015  ■  IEEE industrial electronics magazine  37

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on March 25,2020 at 19:51:01 UTC from IEEE Xplore. Restrictions apply.
 [31]. If a state-space model is employed, a
similar formulation is obtained [2].
The total harmonic distortion of the output voltages The derivation of GPC involves sev-
and the RMS value of the error between the eral computations that can be found in
the original article or in [18]. It uses a
reference and the actual output voltages have been transfer function model of the plant with
selected as quality performance indices. an integrated white noise (the so-called
controlled auto-regressive integrated
moving average model) and a quadratic
cost function as given in (6). In the uncon-
of practical limitations. For FCS-MPC, the the power flow through a VSI under a strained case, the optimization is solved
stability analysis problem is difficult to constant switching frequency operation. analytically at each sampling instant,
address since this MPC strategy, in gen- The application times are computed, im- supplying the proposed set of changes
eral, does not provide an explicit solution. posing a constant switching frequency in the control actions along the horizon
Recently, the use of concepts such as the as well as other objectives; for example, u =[Tu (t), Tu (t +1), f, Tu (t + N u -1)] .
move-blocking strategy [23] or extrapola- power losses are reduced and fast tran- GPC uses the concept of free and forced
tion [24] allows for the use of longer ho- sient dynamic behavior is obtained. response to obtain the vector y of pre-
rizons in FCS-MPC. A review of FCS-MPC dicted outputs as a function of the free re-
strategies that achieve long prediction MPC with CCS sponse ^ f h, which only depends on the
intervals was recently published in [25]. In this case, the control actions are con- known state of the plant, and the forced
In general, FCS-MPC does not keep a tinuous-time signals that are sent to a response, which depends on the control
constant switching frequency, which is modulator, and the optimization problem sequence ^ Gu h through matrix G, ob-
an important issue in commercial appli- is solved analytically by setting the de- tained from the model parameters. Since
cations of power electronic converters. rivative of the cost function equal to zero the plant is linear, y = Gu + f, the cost
However, fixed switching frequency can in the unconstrained case (if constraints function is given by (8), whose minimum
also be obtained in FCS-MPC, as in [26], are included, a QP must be solved, as de- gives the solution already presented
where the transistor switching times for scribed in the “Basic Elements” section). in (9), assuming there are no constraints
an ac/dc converter are computed mini- Since an analytical solution is provided, on the control signals.
mizing a cost function that considers long horizons can be employed. The most Notice that the control signal that
active and reactive power errors and extended method of this type is general- is actually sent to the process is the
imposing a constant value for the sum of ized predictive control (GPC) [28], so this first element of vector u, given by
the three active voltage vectors. Another branch is also known as GPC-type control;
example is predictive direct-power con- although other MPC methods can be used, Tu (t) = K (w - f), (13)
trol, presented in [27], which selects the such as dynamic matrix control [29], ex-
best voltage-vector sequences and com- tended prediction self adaptive control where K is the first row of matrix
putes their application times to control [30], and predictive functional control (G T G + mI) -1 G T , which can be com-
puted beforehand from the transfer
function. This has a clear meaning that
can easily be derived from Figure 7: if
there are no future predicted errors,
that is, if w - f = 0, then there is no
control move because the objective
State

x^i1(t + 2) will be fulfilled with the free evolution

S1
S1 x^1(t + 1) of the process. However, in the other
case, there will be an increment in the
x^ii(t + 2) control action proportional (with a
x^ (t ) Si Si factor K) to that future error. Notice
that the action is taken with respect to
x^i (t + 1) future errors, not past errors, as is the
case in conventional PID controllers.
SM SM x^iM(t + 2)
x^M (t + 1)
Illustrative Example:
Application to a VSI
t t+1 t+2 Several industrial and critical ap-
Discrete Time plications include loads that need a
constant power supply to ensure the
FIGURE 6 – The FCS-MPC operating principle. correct working of the system. These

38  IEEE industrial electronics magazine  ■  september 2015

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on March 25,2020 at 19:51:01 UTC from IEEE Xplore. Restrictions apply.
critical loads are usually supported by
an uninterruptible power supply (UPS), Reference
Plant Output
where a VSI provides a high-quality Trajectory ∆u(t)
W y(t)
output sinusoidal voltage to the loads. +
K
In this section, the two methods, FCS –
and CCS, which represent two different
Free
approaches to the problem, are devel-
Response
oped and applied to this converter, and Feedback
f Model and
their performances are assessed. Past Data

System Description
A three-phase, two-level power converter FIGURE 7 – A block interpretation of the GPC control law.
used as a VSI is depicted in ­Figure 8. The
load is connected to the VSI through an
inductor–capacitor (LC) filter to remove that can generate a combinatorial used with N u = 1, prediction horizons
the high-order harmonic components explosion if the horizon is too long. N 1 = 1, N 2 = 5, and a control-weight-
in the converter output voltage and pro- Therefore, the cost function used is ing factor of m = 0.5. The values of N 2
vide a high-quality sinusoidal voltage to and m will be changed in the experi-
the load. The behavior of the system is J = [v *ca - v ca (t + 1)] 2 ments to assess the system’s perfor-

defined by the dynamic equations of the + [v c*b - v cb (t + 1)] 2, (15) mance. To derive the control law, the
output filter inductor currents and the system is represented using its trans-
output filter capacitor voltages as follows: where v ca and v cb are the real and fer function, which can be written in
imaginary parts of the output voltage. the Laplace domain and in the z-do-
dv C, abc
i L, abc = C + i O, abc, v I, abc The steps needed to implement the main for a sampling time of 83, 33 ns
dt
FCS-MPC are as follows: (since the sampling frequency is
di L, abc
=L + v C, abc . (14) 1) At each sampling time t, measure 12 kHz) as follows:
dt
v c (t ) .
In this example, v dc = 400 V, L = 2 mH, 2) For each of the seven S i, evaluate the
C = 20 nF, and the load is resistive cost function J (15) using the discrete Y (s) 1
H (s) = = ,
U (s) LCs 2 + L s + 1
with R l = 60 X. The control objective model and choose the S i that gives the Rl
)
is to generate a three-phase sinusoidal smallest values of J. Make S = S i .
) H (z) = 0.085 + 0.083z -1 . (16)
output voltage (v C ) tracking a desired 3) Apply S and go to 1. 1 - 1.765z -1 + 0.932z -2
reference of 120-V root mean square
(RMS) and 50 Hz. Derivation of GPC
The dynamics in (14) are functions of The discrete transfer function is
Derivation of FCS-MPC the VSI output voltages that depend used to compute the predictions as
In the case of a three-phase UPS invert- on the power semiconductor’s switch- described in “MPC with CCS” section.
er, the voltage (control action) can take ing functions. Therefore, the VSI out- For the chosen tuning parameters,
seven different values at each sampling put voltages, v I, abc, have been chosen the controller gain of (13), shown
instant. Therefore, the prediction of the as the control signal (input), u (t), and in Figure 7, is given by K = [0.022,
system output along the horizon must the final switching sequence will be 0.081, 0.166, 0.260, 0.347], which is com-
be done considering that this voltage generated through a PWM technique. puted beforehand. During execution,
can change in the future. Therefore, the A cost function, as described in (6), is the only computation that has to be
previously described FCS-MPC can be
developed with seven possible control
actions. To do that, the model of the in-
verter that gives the predicted output S1 S2 S3
along the horizon ^v C (t + j) h as a func-
tion of the switching states S i is needed. iLa L a iO a
Vdc r
This model can be found in [32] and is iLb L b iO b
s Load
not described here. It provides a discrete iLc L c iO c
t
equation that supplies the inverter out- iC a iC b iC c
put as a function of the switching signals S4 S5 S6
VI, abc C C C VC, abc
and the measured variables at each sam-
pling instant.
O
The prediction horizon is set to one
to prevent excessive computations FIGURE 8 – A scheme of a VSI connected to an output LC filter.

september 2015  ■  IEEE industrial electronics magazine  39

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on March 25,2020 at 19:51:01 UTC from IEEE Xplore. Restrictions apply.
done is the evaluation of the increment
in the control action TABLE 1 – THE MODEL PARAMETERS. TABLE 2 – THE EXPERIMENTAL PARAMETERS.
PARAMETER DESCRIPTION VALUE PARAMETER DESCRIPTION VALUE
L Output filter 2 mH L Output filter 2 mH
Du (t) = - 0.005Du (t - 1) - 7.544y (t) inductance inductance
+12.315y (t -1) - 5.647y (t - 2) C Output filter 50 nF C Output filter 50 nF
+ 0.021w (t + 1) capacitance capacitance
+ 0.081w (t + 2) Rl Output resistor load 60 X
Rl Output resistor load 15 X
+ 0.166w (t + 3)
Ll Output inductor load 10 mH
+ 0.260w (t + 4) (RL) load. The laboratory system is
+ 0.347w (t + 5), (17) v dc dc-link voltage 400 V
shown in Figure 9. In both cases, the
 V *
RMS value of the 120 V
target is to generate a three-phase RMS
voltage reference
sinusoidal output voltage, tracking a
where w (t + j) is the reference tra- desired reference of 120 V RMS and
jectory, which is known in advance. 50 Hz. The sampling frequency is switching frequency is smaller than the
A detailed description of all the GPC set to 12 kHz in the GPC case and is sampling frequency, making it necessary
matrices and parameters as well as changed between 20 and 40 kHz in the to work with a high sampling frequency
the experimental validation can be FCS-MPC case. For all of the labora- to get a similar performance to GPC.
found in [7]. tory tests, the experiments are per- In FCS-MPC experiments with a
The steps needed to implement the formed with a mismatch between the sampling frequency of fs = 20 kHz, an
GPC are as follows: model (Table 1) and the actual load effective switching frequency of 3 kHz
1) Compute the value of gain K before (Table 2). The total harmonic distor- is achieved. To try to make a com-
starting the controller. tion (THD) of the output voltages and parison as honest as possible, the
2) At each sampling time t, measure the RMS value of the error between sampling frequency will be increased
v c (t ) . the reference and the actual output as much as possible. In this case, the
3) Compute the control action u (t) voltages have been selected as quality limit is imposed by the microcontroller
using (17). performance indices. TMS320F28335 digital signal controller
4) Apply this action to the PWM and As described earlier, the control (DSC), which runs the control algo-
go to 2. signal in the GPC is generated through rithms and has an instruction cycle of
a pulse-width modulator. So the sam- 6.6 ns. For executing this control algo-
Experiments and Assessment pling frequency matches the switching rithm properly, the microcontroller is
This section shows some illustra- frequency, which, therefore, can be set able to reach a maximum sampling fre-
tive experiments that can help to to a specific value. However, for the quency of 40 kHz, and it results in an
understand the previously explained FCS-MPC, this is not the case since the effective switching frequency of about
algorithms. The experiment results switches are directly manipulated and 4.5 kHz. Notice that the usual range of
are obtained by applying both MPC the switching frequency is variable, reported FCS-MPC implementations is
algorithms to a VSI connected to an depending on the operating point and between 15 (see [33] for an asymmetric
LC filter supplying a resistor–inductor the sampling frequency. The effective flying capacitor converter) and 40 kHz

(a) (b)

FIGURE 10 – The experimental results for an FCS-MPC controller with fs = 20 kHz (effective
FIGURE 9 – An experimental VSI. switching frequency of 3 kHz): (a) output voltages and (b) harmonics and THD values.

40  IEEE industrial electronics magazine  ■  september 2015

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on March 25,2020 at 19:51:01 UTC from IEEE Xplore. Restrictions apply.
(see [3] for a three-phase inverter). On known, the performance of the GPC An assessment of the effect of the tun-
the other hand, in the GPC case, the strategy depends on the proper choice ing parameters on the system perfor-
limit is imposed by a design decision, of the controller parameters, the pre- mance can be found in [7].
and it is fixed only at 12 kHz because diction horizon, and the weighting fac- In Table 3, the performance of both
this sampling frequency is enough to tor. Figure 14(b) shows how the GPC is methods is compared for different set-
obtain good performance. able to handle the problem of mismatch tings. Note that it is not easy to compare
Figure 10 shows experimental wave- between the model and the actual load. both methods in the same conditions.
forms of the three-phase output capac-
itor voltages for an FCS-MPC controller,
with a prediction horizon equal to one,
captured with the Fluke 434 three-
phase power analyzer. Figure 10(a) pre-
sents the output capacitor voltages for
a 20-kHz sampling frequency FCS-MPC
controller, and Figure 10(b) presents
the corresponding THD value. A bet-
ter performance of this controller
is achieved by increasing the sam-
pling frequency to the limit of 40 kHz,
which results in an approximately (a) (b)
4.5-kHz effective switching frequency.
Figure 11 shows how the performance FIGURE 11 – The experimental results for an FCS-MPC controller with fs = 40 kHz (effective
is improved. The output tracks the switching frequency of 4.5 kHz): (a) the output voltages and (b) the harmonics and THD values.
reference voltage with good behav-
ior despite the model mismatch. To
have a clear idea of this improvement,
Figure 12 displays the internal DSC var- 250 250
iables corresponding to the values of 200 200
150 150
the output voltage of one of the three 100 100
phases and its reference voltage and 50 50
VaO (V)

VaO (V)

error value between both signals. No- 0 0

−50 −50
tice that experiments with smaller sam- −100 −100
pling frequencies are not shown here −150 −150
since the results are very poor. In fact, −200 −200
−250 −250
40 kHz can be considered a common 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
value for the sampling frequency in Time (ms) Time (ms)
(a) (b)
FCS-MPC implementations. This is the
* *
chosen value in [3] for a similar VSI VaO VaO VaO – VaO
since a smaller value of 15 kHz showed
bad results even in simulations. FIGURE 12 – The system performance for an FCS-MPC controller with different values of sam-
It is very difficult to increase the pling f­requencies: (a) 20 kHz and (b) 40 kHz.
horizon in FCS-MPC since the compu-
tational problem is untractable in the
available time imposed by the micro-
controller computation limit. Howev-
er, it can be easily augmented in GPC.
A value of N = 6 provides the results
shown in Figure 13, with a sampling
(and switching) frequency of 12 kHz.
Figure 14 displays the internal DSC
variables for two experimental results
with a GPC controller. Figure 14(a)
shows the results when the control-
ler design parameters are N = 5 and (a) (b)
m = 0. To improve the tracking of the
reference signal, these values are tuned FIGURE 13 – The experimental results for a GPC controller with N = 6 and m = 1.05: (a) the
to N = 6 and m = 1.05. As is well output v­ oltages and (b) the harmonics and THD values.

september 2015  ■  IEEE industrial electronics magazine  41

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on March 25,2020 at 19:51:01 UTC from IEEE Xplore. Restrictions apply.
 been stated before, GPC provides an
200 200
explicit solution for the unconstrained
150 150
MPC strategy, whose complexity (and,
100 100
therefore, computing time) is almost
50 50
independent of the prediction horizon.
VaO (V)

VaO (V)
0 0
Note that most of the computation of
−50 −50
this unconstrained case can be done be-
−100 −100
forehand, and the computational cost is
−150 −150
small. Both methods are able to handle
−200 −200
0 10 20 30 40 0 10 20 30 40 the problem of mismatch between the
Time (ms) Time (ms) model and the actual load. The experi-
(a) (b)
mental results [7] have revealed that the
* *
VaO VaO VaO – VaO prediction horizon length and control-
weighting factor value directly affect
FIGURE 14 – The system performance for a GPC controller with different values of N and m: the performance of the system. Note
(a) N = 5 and m = 0 and (b) N = 6 and m = 1.05. that FCS-MPC needs a high sampling
frequency (which means that the micro-
controller is working near its operation
limit) to behave like GPC (which can
TABLE 3 – A PERFORMANCE COMPARISON. easily work with longer horizons and
CONTROLLER N THD (%) ERRORRMS (%) smaller sampling frequencies).
fs (kHz) fesw (kHz) m
This illustrative case has been
FCS–MPC 1 20 3 – 5.7 15.4 presented as an example of the appli-
FCS–MPC 1 40 4.5 – 7.8 cation of both methods to a real con-
1.9
verter. Further comparisons among
GPC 5 12 12 1.05 2.2 4.2
controllers can be found in published
GPC 6 12 12 1.05 2.0 2.6
papers, such as a comparison among
several MPC schemes and techniques
based on PWM, space vector modula-
tion, and optimized pulse patterns for
the control of a medium-voltage drive
TABLE 4 – A CONTROLLER COMPARISON.
presented in [9]. A comparative as-
COMPUTATIONAL THEORETICAL
CONTROLLER MODULATOR HORIZON DERIVATION COST RESULTS sessment of FCS-MPC with a linear cur-
FS–MPC No Short Simple Low (N=1) Scarce
rent controller in two-level VSIs is also
presented in [35].
GPC Yes Any Complex Low (unconstrained) Well established

Conclusions
This article has presented the main
The main issue is that, in one case (GPC), and RMS value of the error between features of MPC for power convert-
the switching frequency is fixed and the measured and reference signals. ers. It has been shown that MPC is a
equal to the sampling frequency; while, The proposed GPC control strategy powerful technique whose capabili-
in the FCS-MPC, this value changes. This provides high performance with a low ties are just starting to be exploited
means that, to get an appropriate switch- THD of the output voltages. The FCS- in this field. The main elements of any
ing frequency, the sampling frequency in MPC presents similar THD results for predictive controller have been pre-
FCS-MPC must be considerably higher fs = 40 kHz (effective switching fre- sented, and the concepts have been
than GPC (in our case, 40 versus 12 kHz). quency of 4.5 kHz), but the RMS value illustrated on a VSI. FCS-MPC and CCS-
These are the frequencies chosen for of the error between the measured MPC, which use different approaches
comparison since the THD obtained in and reference signals is worse for this to the problem, have been analyzed
both cases is similar. controller. Note that this is mainly and applied to the example, showing
The best results for GPC running due to the steady-state tracking error, the effect of the choice of the predic-
at fs = 12 kHz are obtained for a set which can be significant in the FCS- tion horizon and the control-weight-
with values of N = 6 and m = 1.05 MPC, especially when operating with ing factor on the THD and RMS.
(a GPC with N = 1 is also tested to lower switching frequencies or small There are many issues that could
show that a worse performance is current reference amplitudes, as ana- not be addressed in this article. One of
obtained for short horizons indepen- lyzed in [34]. them is stability. The developed meth-
dently of the method). The choice Some general features of both meth- ods must guarantee stability, and this is
of these parameters affects the THD ods are depicted in Table 4. As has still an open issue, mainly in FCS-MPC.

42  IEEE industrial electronics magazine  ■  september 2015

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on March 25,2020 at 19:51:01 UTC from IEEE Xplore. Restrictions apply.
Concepts such as terminal region and of Science and Innovation. Currently, he [18] A. Linder, R. Kanchan, R. Kennel, and P. Stolze,
Model-Based Predictive Control of Electric Drives.
terminal cost that are used in other is pursuing his Ph.D. degree at the same Göttingen: Cuvillier Verlag, 2012.
MPC formulations should be adapted university. His research interests in- [19] S. Kouro, P. Cortes, R. Vargas, U. Ammann, and
to this approach. Other open fields are clude MPC strategies for power convert- J.  Rodriguez, “Model predictive control—A
simple and powerful method to control power
the reduction of the computational ef- ers and electric vehicles. converters,” IEEE Trans. Ind. Electron., vol. 56,
fort (mainly in the constrained case and pp. 1826–1838, June 2009.
[20] R. Vargas, P. Cortes, U. Ammann, J. Rodriguez,
long horizons) or parameter tuning. References and J.  Pontt, “Predictive control of a three-
The popular and intuitive FCS-MPC [1] J.  Richalet, A.  Rault, J.  Testud, and J.  Papon, phase neutral-point-clamped inverter,” IEEE
“Model predictive heuristic control: Appli- Trans. Ind. Electron., vol. 54, pp. 2697–2705, Oct.
has paved the way to the implemen- cation to industrial processes,” Automatica, 2007.
tation of MPC in demanding power vol. 14, no. 2, pp. 413–428, 1978. [21] D.  E. Quevedo, G.  C. Goodwin, and J.  A.  D.
[2] E. Camacho and C. Bordons, Model Predictive Dona, “Finite constraint set receding horizon
electronics applications. The great Control. New York: Springer, 2004. quadratic control,” Robust Nonlinear Control,
potential of MPC will flourish in the [3] J.  Rodriguez and P.  Cortes, Predictive Control vol. 14, no. 4, pp. 355–377, 2004.
of Power Converters and Electrical Drives. New [22] J. Rawlings and D. Mayne, Model Predictive Con-
coming years, thanks to the expected Jersey: Wiley, 2012. trol: Theory and Design. Nob Hill Publishing, Madi-
theoretical contributions and to the [4] S. Vazquez, J. Leon, L. Franquelo, J. Rodriguez, son, WI, 2009.
H. Young, A. Marquez, and P. Zanchetta, “Mod- [23] R.  Cagienard, P.  Grieder, E.  C. Kerrigan, and
increasing computational capabilities el predictive control: A review of its applica- M. Morari, “Move blocking strategies in reced-
of control hardware. tions in power electronics,” IEEE Ind. Electron. ing horizon control,” J. Process Control, vol. 17,
Mag., vol. 8, no. 1, pp. 16–31, 2014. no. 6, pp. 563–570, 2007.
[5] P. Cortes, M. Kazmierkowski, R. Kennel, D. Que- [24] T. Geyer, G. Papafotiou, and M. Morari, “Model
Acknowledgments vedo, and J. Rodriguez, “Predictive control in predictive direct torque control. Part I: Con-
power electronics and drives,” IEEE Trans. Ind. cept, algorithm and analysis,” IEEE Trans. Ind.
This work was partially supported by Electron., vol. 55, pp. 4312–4324, Dec. 2008. Electron., vol. 56, p. 18941905, June 2009.
the Spanish Ministry of Economy and [6] J. Rodriguez, J. Pontt, C. Silva, P. Correa, P. Lezana, [25] P. Karamamakos, T. Geyer, N. Oikonomou, F. D.
P.  Cortes, and U.  Ammann, “Predictive current Kieferndork, and S.  Manias, “Direct model
Competitiveness under Grant DPI2013- control of a voltage source inverter,” IEEE Trans. predictive control: A review of strategies that
46912. We want to thank Sergio Vazquez Ind. Electron., vol. 54, pp. 495–503, Feb. 2007. achieve long prediction intervals for power
[7] S.  Vazquez, C.  Montero, C.  Bordons, and electronics,” IEEE Ind. Electron. Mag., vol.  8,
and Abraham Marquez for their help in L.  Franquelo, “Design and experimental no. 1, pp. 32–43, 2014.
carrying out the experiments. validation of a model predictive control [26] P. Antoniewicz and M. Kazmierkowski, “Virtu-
strategy for a VSI with long prediction hori- al-flux-based predictive direct power control
zon,” in Proc. IECON Annu. Conf. IEEE Indus- of AC/DC converters with online inductance
Biographies trial Electronics Society, Vienna, Austria, 2013, estimation,” IEEE Trans. Ind. Electron., vol. 55,
pp. 5786–5791. no. 12, pp. 4381–4390, 2008.
Carlos Bordons (bordons@us.es) re- [8] D. Mayne, J. Rawlings, C. Rao, and P. Scokaert,
[27] S.  Aurtenechea, M.  Rodriguez, E.  Oyarbide,
ceived his Ph.D. degree in electrical “Constrained model predictive control: Sta-
and J.  Torrealday, “Predictive direct power
bility and optimality,” Automatica, vol.  36,
engineering from the University of Sev- pp. 789–814, 2000.
control—A new control strategy for DC/AC
converters,” in Proc. IEEE 32nd Annual Conf.
ille, Spain, in 1994. He is currently a full [9] T. Geyer, “A comparison of control and modu-
Industrial Electronics (IECON), Paris, pp. 1661–
lation schemes for medium-voltage drives:
professor and head of the Department Emerging predictive control concepts versus
1666.
[28] D. Clarke, C. Mohtadi, and P. Tuffs, “Generalized
of Systems Engineering and Automatic PWM-based schemes,” IEEE Trans. Ind. Appli-
predictive control. Part I. The basic algorithm,”
cat., vol. 47, pp. 1380–1389, May 2011.
Control. He has been a Member of the [10] A. Bemporad, M. Morari, V. Dua, and E. Pistiko- Automatica, vol. 23, no. 2, pp. 137–148, 1987.
IEEE since 1997. He was elected as a poulos, “The explicit linear quadratic regu- [29] C. Cutler and B. Ramaker, “Dynamic matrix con-
lator for constrained systems,” Automatica, trol—A computer control algorithm,” in Proc. Au-
European Union Control Association vol. 38, no. 1, pp. 3–20, 2002. tomatic Control Conf. (ACC), San Francisco, 1980.
Council member in 2007, and he is an [11] A. Bemporad, F. Borrelli, and M. Morari, “Model [30] R. D. Keyser and A. V. Cuawenberghe, “Extend-
predictive control based on linear programming. ed prediction self-adaptive control,” in Proc.
associate editor of IEEE Transactions on The explicit solution,” IEEE Trans. Autom. Contr., IFAC Symp. Identification and System Parameter
Industrial Electronics and Control Engin- vol. 47, no. 12, pp. 1974–1985, 2002. Estimation, York, UK, 1985, pp. 1317–1322.
[12] S.  Mariethoz and M.  Morari, “Explicit model- [31] J. Richalet, S. A. el Ata-Doss, C. Arber, H. Kun-
eering Practice. He is the coauthor of the predictive control of a PWM inverter with an tze, A.  Jacubash, and W.  Schill, “Predictive
book Model Predictive Control (Spring- LCL filter,” IEEE Trans. Ind. Electron., vol.  56, functional control. Application to fast and ac-
pp. 389–399, Feb. 2009. curate robots,” in Proc. 10th IFAC Congr., Munich,
er-Verlag, first and second editions). [13] S. Richter, S. Mariethoz, and M. Morari, “High- 1987.
His current research interests include speed online MPC based on a fast gradient [32] P.  Cortes, G.  Ortiz, J.  Yuz, J.  Rodriguez,
method applied to power converter control,” S.  Vazquez, and L.  Franquelo, “Model predic-
advanced control, especially model pre- in Proc. American Control Conf. (ACC), 2010, tive control of an inverter with output LC filter
dictive control and its application to fuel- pp. 4737–4743. for UPS applications,” IEEE Trans. Ind. Elec-
[14] J.  Neely, S.  Pekarek, R.  DeCarlo, and N.  Vaks, tron., vol. 56, no. 6, pp. 1875–1883. 2009.
cell-based systems, power converters, “Real-time hybrid model predictive control [33] P.  Lezana, R.  P. Aguilera, and D.  E. Quevedo,
and microgrids. of a boost converter with constant power “Model predictive control of an asymmetric
load,” in Proc. 25th Annu. IEEE Applied Power flying capacitor converter,” IEEE Trans. Ind.
Carlos Montero (cmontero1@us.es) Electronics Conf. Exposition (APEC), Feb. 2010, Electron., vol. 56, no. 6, pp. 1839–1846, 2009.
received his B.Eng. degree in telecom- pp. 480–490. [34] R.  P. Aguilera, P.  Lezana, and D.  E. Quevedo,
munications engineering and his M.Sc. [15] J. Neely, R. DeCarlo, and S. Pekarek, “Real-time “Finite-control-set model predictive control
model predictive control of the Cuk convert- with improved steady-state performance,” IEEE
degree in automation, robotics, and er,” in Proc. IEEE 12th Workshop on Control and Trans. Ind. Inform., vol.  9, no.  2, pp.  658–667,
telematics from the University of Sev- Modeling for Power Electronics (COMPEL), June 2013.
2010, pp. 1–8. [35] H.  Young, M.  Perez, J.  Rodriguez, and H.  Abu-
ille, Spain, in 2009 and 2011, respective- [16] J. Bonilla and R. D. Keyser, “Nonlinear predic- Rub, “Assessing finite-control-set model pre-
ly. He has been with the Department of tive control of a DC-DC converter: A NEPSAC dictive control: A comparison with a linear
approach,” in Proc. European Control Conf., current controller in two-level voltage source
Systems Engineering and Automatic Kos, Greece, July 2007, pp. 1753–1758. inverters,” IEEE Ind. Electron. Mag., vol. 8, no. 1,
Control since 2006, first as a student on [17] P.  Mattavelli, “An improved deadbeat control pp. 44–52, 2014.
for UPS using disturbance observers,” IEEE
a scholarship and now as a technical ex- Trans. Ind. Electron., vol. 52, no. 1, pp. 206–212, 
pert support cofunded by the Ministry 2005.

september 2015  ■  IEEE industrial electronics magazine  43

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on March 25,2020 at 19:51:01 UTC from IEEE Xplore. Restrictions apply.