You are on page 1of 21

1

Alternative Control Methods for a Four-level


Three-cell DC-DC Converter
Diego Patino, Mihai Baja , Herve Cormerais , Pierre Riedinger, Jean Buisson, Claude Iung

Abstract
This paper proposes three synthesis methods for controlling power converters The three control strategies yield
easy to implement state feedback control laws. The first one is a stabilization approach, based on energetic principles
and the notion of Lyapunov function. The second one is an optimal control approach based on the minimum principle.
The last one is a neural predictive approach which uses model predictive control to track a given optimal stable limit
cycle. This method allows a proper control of the waveform. Excepting for the predictive approach, the system
stability for the others methods are guaranteed. The four-level three-cell DC-DC Converter is used as a benchmark to
test these strategies. Simulation and experimental results show that the methods have a good performance even with
load perturbations.
Index Terms
Stabilizing control, optimal control, neural predictive control, switched systems, power converters.

I. I NTRODUCTION
Power supplies are currently embedded in computers, electric drives, and cellular phones and generally in all
electric devices. Their aim is to convert an electrical energy shape (voltage /current / frequency) to another one.
Applications in power systems frequently use DC-DC converters such as boost, buck or Cuk converters. In the
case of industrial applications with power of a few megawatts, the switching components voltage becomes very
high (several kilovolts). Therefore, the switching frequency must be maintained to a low value and bulky filters
are needed for obtaining an appropriated output [1]. To palliate this drawback, a new class of power electronic
converters has appeared, called multicellular or multilevel converters. These structures consist of a series connection
of switching devices with passive storage elements, which are used to generate intermediate voltage levels levels
[2]. The control law for these structures needs to maintain the intermediate voltage levels at some constant values
and to regulate the voltage or the load current.

Corresponding author. Email:herve.cormerais@supelec.fr

Mihai Baja, Herve Cormerais and Jean Buisson are with the Hybrid System Control Team, SUPELEC / IETR, CS 47601, 35576 CessonSevigne, France.
Diego Patino, Pierre Riedinger and Claude Iung are with the CRAN (Centre de Recherche en Automatique de Nancy) Nancy-Universite,
CNRS, 2, avenue de la foret de Haye, 54516 Vandoeuvre-l`es-Nancy Cedex, France.

The two main advantages of the multilevel converter are: i) it allows reducing the voltage throughout the switches
by splitting and distributing it on the intermediate levels, ii) the intermediate voltages take discrete values between
0 and the full input voltage E. These voltage levels applied to the load produce a reduction of the output voltage
harmonics. The downside of the listed advantages is that, excepting simpler DC-DC converters, the control of
multilevel converters is more complex [3], [4].
The intermediate voltages on the capacitors must be drastically regulated for two reasons: firstly, to ensure a
good waveform in the load [5] and secondly to protect the switches against high voltages. Usually, for an n-level
multicellular converter, this can be done regulating the intermediate voltages to the following reference values:
(n1)E
E
n,...,
n

[6].

From a control point of view, this is a non-linear problem. Several techniques have been used to solve it. The
first solutions were open-loop self-balancing strategies using Pulse Width Modulation (PWM), a fixed duty cycle
and a constant phase shift between the levels [7]. The simplicity of these approaches leads to low performances
particularly concerning transient.
The first close loop solutions employed independent PI controllers on each intermediate voltage in order to control
each switching cell [8]. There is no systematic method to tune the PI parameters and performances decrease in
presence of parameters or operating point variations. Recently, more complex methods have been developed such
as sliding mode control, which uses a state feedback method with a Lyapunov criterion to synthesize a switching
function [9].
Another approach consists in the system linearization. The control is then achieved using a linear synthesis [10].
In [11] a control strategy based on a decision tree is employed for the control of a multilevel converter, based on
the requirements imposed by a Direct Torque Control (DTC) strategy, which is needed for the drive of an induction
motor. There exist also predictive approaches which are based on the cost function minimisation over a given time
horizon [12]. With the predictive approaches, system nonlinearities and constraints related to the control objective
can be explicitly taken into account. Although these methods are suitable for the control of multilevel converters
[13], [14] they are difficult to implement and sometimes the computing time becomes high with respect to the
system dynamics.
The aim of this paper is to introduce and to compare three state-feedback control techniques on a common
benchmark defined on a four-level three-cell converter. The control objective is to regulate the average value of the
current in a R-L load and the average voltages in the capacitors to fixed values as specified above. The design part
must take into account some constraints concerning the switching frequency and also the robustness must be tested
with step change in the load and power input. The three-hybrid controls (the first one developed by the team at
SUPELEC and the two others by CRAN) are: :
1) A stabilization approach: It is based on energetic principles. The first step of the method consists in the
definition of a common Lyapunov candidate function taking into account the control objective. Its expression
is directly related to the storage energy. It uses the Hamiltonian Port formalism. In the second step, the
control, which is under a static state feedback, is chosen to ensure at each sample time the negativity of the

time-derivative expression of the common Lyapunov candidate function guaranteeing the system stability [15].
One of the advantages for this approach is that it can be easily extended to the case of DCM (Discontinuous
Conduction Mode).
2) An optimal control approach: using the minimum principle and singular control theory, it is shown that it is
possible to synthesize an optimal state feedback control law. The optimisation of a quadratic criterion is a
pledge for system stability. The synthesis implies determining the singular optimal solutions that arise in the
optimal problem. An algorithm is proposed to generate all optimal trajectories on a given state space area.
Then, a neural network is used to interpolate all the optimal solutions. Indeed, the optimal solutions are learnt
off-line and the real time implementation is ensured with a simple evaluation of the activation functions at
sample time.
3) A neural predictive control approach: the development of such a method is mainly motivated by the fact
that in the power converters field, classical control methods are usually designed with a constant reference
which represents an average value of the steady state. Consequently, the cycle is not properly controlled and
it may present unexpected harmonic contents. In this part of the article, we propose to track an optimal limit
cycle defined as a reference in the steady state. This cycle represents the best cycle in the sense given by
a criterion. For instance, this criterion can be defined in order to reduce the quadratic norm of the tracking
error or to preserve the system against undesired harmonic content. The tracking control is obtained using
a predictive controller, which can take into account constraints on the switching frequency related to the
devices in the design part. The optimal solutions are learnt through a neural network that yields a state
feedback control. This state feedback is given in the form of a simple input-output relation in such a way
that real time implementation can be guaranteed. The predictive control has already been used as a control
method for the power converters but not with the goal of optimising the limit cycle [13].
The paper is organized as follows: in section II, the physical model of the four-level three-cell DC-DC converter is
introduced. Like most of the DC-DC converters, it is shown that the model enters into the class of affine-switched
system. Section III states the control problem and its constraints. In section IV, the three control methods are
introduced and a short description of the control design is given. In section V, the experimental setup used to test
the methods is presented. Section VI shows experimental and simulation results. Finally, section VII provides some
comparisons between the results obtained by the proposed approaches.
II. P HYSICAL

MODEL OF THE FOUR - LEVEL THREE - CELL

DC-DC

CONVERTER

Multilevel converters and more generally power converters represent a particular class of dynamic systems called
switching physical systems. They include semiconductor switching devices evolving much faster than the time scale
at which the system behavior needs to be analysed. Therefore, the switching times are negligible and for modelling
purposes the switching devices can be assumed to be ideal. The matrix representation of a switching system in a
standard PCH (Port Control Hamiltonian) formulation has the following expression [15]:
x = [J() R()]

H(x)
+ G()u
x

(1)

3
C2

Fig. 1.

1 3

vC2 C1

vC1

1 2

1 1

Lch
iL

Rch

The four-level three-cell DC-DC converter

where x Rn is the state vector, {0, 1}p is the boolean control variable and u Rm is the power input vector.
Matrices J and R are called structure matrices. Matrix J is skew-symmetric, J = J T and it corresponds to the
power interconnections in the model. Matrix R is nonnegative and represents the dissipating part of the system.
G is the power input matrix. H represents the stored energy in the system. This is also called the Hamiltonian of
the system. If the constitutive relations of the storage elements are linear, which is most often the case in power
converters, the Hamiltonian of the system is such that:
H (x)
= Fx = z
x

(2)

where F = F T > 0 and in the simple cases, it is a diagonal matrix. Thus, the Hamiltonian of the system can be
represented as:
H(x) =

1 T
x Fx
2

(3)

Furthermore, it has been proved that the matrices J(), R() and G() are affine with respect to the boolean control
variable [16] and they can be written as:
J () = J0 +

p
X

j Jj ,

R () = R0 +

p
X

j Rj ,

G () = G0 +

j Gj

(4)

j=1

j=1

j=1

p
X

where j are the components of the control vector .


The circuit topology of the four-level three-cell DC-DC converter is shown in Fig.1. Three switching cells can
be isolated, each one containing two switches that operate dually. The behavior of each cell can be described using
only one boolean control variable i {0, 1} with i = 1, 2, 3. i = 1 means that the upper switch is closed and the
lower switch is open whereas i = 0 means that the upper switch is open and the lower switch is closed. Taking
x = [qC1 , qC2 , pL ]T where qC1 , qC2 represent the charge in each capacitor, pL the magnetic flux in the inductance
and u = E as the input voltage, the matrices J, R, G

0
0
0
2 1

J() = 0
0
3 2 , R() = 0

0
1 2 2 3
0

and F are given by:

1
0
0
0


C1

0
0 , G() = 0 , F () = 0

0 Rch
3
0

0
1
C2

0 (5)

1
Lch

The voltages and the current are simply deduced from relation z = F x i.e.
vC1 =

1
qC1
C1

vC2 =

1
qC2
C2

iL =

1
pL
Lch

(6)

An equivalent state space equation in function of the voltages vC1 , vC2 on the capacitors C1 and C2 and the
current iL in the load is:

v C1

CiL1


v C2 = 0 + 0


vC1
ch
i L
R
Lch iL
Lch

1

iL

C 2 2
EvC2
3
Lch

iL
C1
CiL2
vC2 vC1
Lch

(7)

In this model, the boolean control (1 , 2 and 3 ) appears explicitly in an affine form.
In this article and for convenience, the stabilization control approach uses model (1) while the neural predictive
and optimal control approaches deal with the state equation (7).
Remark 1: For the four-level three cell converter, a DCM appears when vC2 vC1 due to the use of MOSFET
as switching devices. In this article, the DCM is not taken into account. We assume that the operating area is far
from this boundary. Therefore, the switching devices are fully controlled.

III. T HE C ONTROL P ROBLEM


Power converters are made to adapt the energy source to an electric load. This goal is achieved by high switching
frequency of the semiconductor devices and leads to a cyclic behavior in steady state. The desired operating point
can be defined as the state average value of the cycle. Let us note by z the state of (7) (or (1)), it is clear that these
equations can be written in the following form:
z = f (z) + g(z)

(8)

with T = [1 , , p ], p N. The operating points set are generally expressed in term of the average state z. z
is given by the convolution product:
z(t) = Tp

1
z(t) =
Tp

z( )d

(9)

tTp

where Tp is the cycle period and Tp is a rectangular window function.


The dynamical model of z is obtained differentiating (9). However, the derivative is generally intractable or
unusable because of its nonlinear form. A solution consists in defining the average state model
z = f (
z ) + g(
z )

(10)

i [0 1], i = 1, , p

(11)

by relaxing the control set to its convex hull:

which gives an approximation of the dynamic of z. is the average value of on the cycle. It has been proven that
z z and z when Tp 0).
ze is close to z and z when Tp is small with respect to the system dynamics [17] (e
The operating points are then defined as the equilibrium points of the average state model:
Z0 = {z0 Rn : f (z0 ) + g(z0 )ref = 0, ref [0 1]p }

(12)

Therefore, there exist equilibrium points z0 whose associated control ref belongs to ]0 1[p . Since no mode allows
to hold the given reference z0 with ref {0, 1}p, the switched system enters into a cyclic behavior around the
reference z0 . Consequently, ref is nearly the average value of on the cycle.
The control objective for the multilevel converter is to supply a fixed average current value iL0 to the load. Using
(7) and (12), one can compute the admissible value of iL0 . This is iL0 = (E/Rch )3 , with an average value for
the control fixed to i [0 1], i = 1, 2, 3 with 1 = 2 = 3 The capacitor voltages can be freely chosen. As
we already mentioned in the introduction, for improving the spectral quality and to equally distribute the switches
stresses. The reference voltage values for the capacitors are:
vC20 =

2
E
3

vC10 =

1
E
3

(13)

The proposed benchmark is subject to the following performance specifications:


1) The current in the load must be maintained to the reference value iL0 with the best accuracy and small
oscillations. The transient response must be as fast as possible without overshoot.
2) The controllers must be robust to the load variations and to parametric uncertainties with respect to the
physical parameters of the model.
3) For physical reasons, the switching law must also respect a dwell time between successive switches of the
same device.
IV. C ONTROL

APPROACHES

A. Stabilization Approach
1) General method: Usually, the approaches in the literature which are based on Lyapunov function consider,
linear systems with a common equilibrium point for each configuration around which the system is stabilized [18],
[19]. In the case of power converters, for each value of the boolean control variable , the system represented by
the state equation (1) may have different equilibrium points. However, for the class of studied systems, a common
Lyapunov function can be established taking:
1
2

(x x0 ) F (x x0 )

(14)

0 = (J (0 ) R (0 )) z0 + G (0 ) u.

(15)

V (x, x0 ) =
where x0 is an equilibrium point defined by

This is equivalent to (12). Because the matrix F is unique for all the modes of the system, V is positive and
continuous for any x. V is also zero only in x0 . Using (1), (2), (4) and (15) the derivative of V is given by:
T
V = (z z0 ) R () (z z0 ) +

p
X

(z z0 ) ((Jj Rj ) z0 + Gj u) (j 0j )

(16)

j=1

with z0 = F x0 . Due to the fact that R() is a non-negative matrix, the first term is always negative. Because
0 0j 1 the sum can be negative by choosing an appropriate value for each boolean j such that each product

(z z0 )T ((Jj Rj )z0 + Gj u)(j 0j ) is negative. If such control law is applied, then V is a Lyapunov function
for the closed loop system, which converges asymptotically toward x0 . Multiple state feedback control strategies
can be considered for attaining this goal [15]:

The maximum descent strategy which consists in choosing, at each time, the value of such that all the terms
in the sum are negative or equal to zero. If R is constant, this yields the minimum value of V . From (16),
switching surfaces are p hyperplanes defined by
Tj = (z z0 )T ((Jj Rj )z0 + Gj u) = 0

j = 1, . . . , p

(17)

This strategy will lead to a sliding motion on the switching surfaces (17).

The minimum switching strategy, which may be used in order to decrease the number of switchings, consists
in holding the same value of until the trajectory hits the switching surface defined by V = 0 and then, in
choosing on that surface a new value of such that V < 0. Even if it does not lead to zeno phenomena, this
approach will lead to faster and faster switching when getting closer to the admissible reference.

Both strategies require an infinite bandwidth. In practice, switching frequency is limited. Multiple strategies to
limit the switching frequency can be devised:

A dead-zone can be created with the help of a parameter . The new switching surfaces are defined by V =
or Tj = . The period and the oscillations amplitude around the reference depend on this parameter.

A delay is introduced between switching instants.

Switches occur at discrete periodic instances. This is the case when using a discrete time controller.

A ball is defined around the reference using the Lyapunov candidate function and commutations occur when
the trajectory hits the surface of the ball.

2) Application to the four-level three-cell DC-DC converter: The Lyapunov candidate function from (14)
becomes:
V (x, x0 ) =

(qC2 qC20 )2
(pL pL0 )2
(qC1 qC10 )2
+
+
2C1
2C2
2Lch

(18)

Further on, a minimum switching strategy will be used. At switching time, several strategies can be applied to
decide the new value for the boolean control variable, which will lead to different behaviors. The one which yields
the switching frequency for the transistors is when the smallest number of boolean control variables change at a
given switching time. The least recent commutation is prioritized and the first configuration with a value for the
Lyapunov candidate function derivative smaller than 0 is chosen.

B. Optimal control
1) Problem formulation and model: In this part of the article, we investigate the capability of optimal techniques
from a theoretical view point for high performance design of power electronic devices.

We are concerned with hybrid optimal control problems, which involve once again a switched affine system:
Z tf
L(z z0 )dt
min

s.t.

z = f (z) + g (z)

(19)

z(t1 ) = z1 {0, 1}p


L(.) : Rn R is the performance function.
In order to face the problem of solving (19), we can use the necessary conditions given by the minimum principle
(MP) [20][22]. We want to highlight that it is possible to determine the appropriate control laws among those
which satisfy the necessary conditions of the MP in the case where the following assumption is satisfied:
Assumption 2: All optimal solutions of the average state model reach the equilibrium in finite time or even in
infinite time in case of infinite time criteria.
The assumption is obviously satisfied both for time optimal criterion or for quadratic criterion in infinite time.
However, it is not hold for quadratic criterion in finite time.

One of the difficulties encountered in this type of optimal control problem (Hamiltonian function have an affine
form w.r.t. the control) is the existence of what is named singular arcs. We will explain this case in the sequel.
Nevertheless, for low order system, we show that it is possible to determine final conditions such that a backwardtime integration generate all trajectories ending at the equilibrium. In that case, optimal control appears as function
of the state.
In order to cover a given entire area of the state space we interpolate the solution using a neural network (NN)
to obtain a map which determines the control as a function of z. Indeed, since the solution of the optimal control
problem is in open-loop, a relation between the state z and the control is highly desired. The high mapping
capabilities of NNs establishes this relation. This synthesis is made off-line.
As example, the procedure is presented when a quadratic criterion in infinite time is used. This methodology
is particularly useful for the control synthesis because it allows to design proper control laws without excessive
computational time for the system (8).
The Hamiltonian H of the problem (19) has an affine form in the control variable :
H = H(0 , (t), z(t), (t))
(20)
= 0 L(z(t) z0 ) + T (t)(f (z) + g (z) (t))
where 0 0 and (t) Rn is the adjoint variable (Lagrange multiplier). We do not mention the dependency with
respect to the time for the different variables.
Since the case where the control dimension p > 1 is quite long to explain and it is not the aim of the article,
we will assume that p = 1.
It follows the minimum principle (MP) to control the affine switched system (19):

Theorem 3: Let (z , ) solve the problem (19). There exists an absolutely continuous function : {0, tf } Rn
and a positive constant 0 , (, 0 ) non identically equivalent to 0, with:
H
=
z

z =

(21)

for almost all t [0, tf ], such that the following conditions are satisfied:
1) The minimum condition on the Hamiltonian:
H = H(0 , , z , ) =

inf

{0,1}

H(0 , , z , )

(22)

The following transversality conditions:


2) Transversality condition 1: For all t [0, tf ]
H(t) = cst

(23)

cst is a constant with cst = 0 if tf is not specified.


3) Transversality condition 2: The initial and final condition:
(0) free

(tf ) = 0

(24)

2) Singular trajectories: This subsection address the problem of singular arcs [23], [24] encountered for solving
the problem (19) using the necessary conditions from the Theorem 3. Following the MP two possible solutions to
(19) can be found:

{0, 1} for almost all t: The solution to the optimal control problem (19) is a bang-bang solution.

There exists a set of time T with nonzero measure such that for all t T , ]0, 1[. The solution to (19) is
not bang bang. This is known as singular arc.

Although the singular trajectory does not solve the problem (19), this is a Fillipov solution that can be
approximated by a sliding motion of the switched systems [25]. This reason explains why it is necessary to
compute singular arcs.
Formally, let us take some basic concepts from the singular optimal control theory [26], [27], [28]:
Definition 4: Since H is linear in , an arc ( , z , ) will be singular on (a, b) if
H
( , z , ) 0

(25)

holds for every t (a, b).


Definition 5: Let (19) be given. Then (t) :=

is called the switching function.

Definition 6: Let (19) be given. The problem order is the smallest integer q such that

d(2q)
dt(2q)

contains explicitly,

where after each differentiation z and are replaced by their expression in (21). Thus:
d(2q)
= A(z, ) + B(z, ).
dt(2q)
If this number does not exist, then q :=

(26)

10

Definition 7: Let (z , ) be a singular arc on (a, b). If q < , the arc order r is the smallest integer such that
on (a, b):

d2r H
dt2r

( , z , ) 6= 0

(27)

If it doesnt exist then r = .


Contrary to the problem order, the arc order depends on the choice of the singular arc and we have obviously r q.
From all these definitions, the necessary conditions to establish optimality have been developed. Here, we use
two well-known conditions:
1) The first order condition or the linear condition can be easily deduced from the definition of equation (25).
It follows that all the derivatives from (z, ) with respect to t must vanish:
d
( ,z , ) = 0
dt
d2
( ,z , ) = 0
dt2
d3
( ,z , ) = 0
dt3

(28)

...
2) A second order condition which has been proven by Robbins [27] is known as the Generalized Legendre Clebsch condition (GLC).
Theorem 8: Let (z , ) be an optimal singular arc and r = q, then (1)q B(z, ) 0 for each t ]a, b[.
The arcs which satisfy these two necessary conditions are candidates to optimality.
When p > 1, the definition of problem order and arc order have a more complex form but there are also equivalent
definitions. See details in [26].
3) Control synthesis from a singular arc: The control objective is to reach an equilibrium point z0 given by the
average model (i.e. control in the interval [0,1]) and to maintain the system at this equilibrium point. Then, two
possibilities may occur:
i) A bang bang control allows to reach z0 in finite time tr < (tr is finite because for any value of the state,
there exists always a mode for which the vector field f (z) + f (z) does not vanish). Therefore, (t) = 0 or 1 for
all t < tr . Since the state cannot be maintained (with z = 0) at the equilibrium z = z0 with = 0 or 1, must
switch to a singular arc at t = tr .
ii) A singular arc tends asymptotically to z0 . The bang-bang solution does not exist. However, it can be
approximated by chattering creating a sliding motion.
The procedure to obtain all the trajectories ending to the final state z0 can be computed backward in time with
the following 4-step scheme:
1) The set of values for (, z, ) which defines the singular arcs is computed from the equation set (28). Indeed,
this equation set forms (finite for q < ) a non linear algebraic system whose solutions are candidates to
optimality. In particular, the control is given by the equation A(z, ) + B(z, ) = 0. See equation (26).
Notice that for low order system, this equation set is sufficient to determine as a state feedback.

11

2) We check the necessary conditions given by Theorem 8 for each possible candidate (, z, ). Thus, we obtain
a smaller set, .
3) As we have mentioned above (point 1), the equilibrium . Using a backward-time integration of the
Hamiltonian system (21) from this final state and from singular arcs ending to this final state, it is possible
to obtain a dense set of the optimal trajectories.
It is important to notice that bifurcations must be taken into account along the singular arcs by switching
to 0 or 1. Moreover, the optimal conditions given by MP must always be checked along each trajectory. It
may arise that two paths intersect with each other in the state space. In that case, the potential conflict could
rise by assessing the cost function.
4) Using the values of z and obtained from the previous step, an interpolation is made in order to obtain a
state feedback (z) on the entire state space area. A NN is used in this phase.
The type of neural network (NN) used to learn the optimal policy is a feedforward NN. Training is obtained by
classical backpropagation procedure, see [29]. The inputs of the NN are the tracking error (z(t) z0 (t)) and the
operating point reference z0 . The output is the optimal policy obtained .
Remark 9: NNs are promising solution for real time implementation since the computation effort is really small.
Moreover, it is possible to implement adaptive training considering the influence of exogenous effects such as load
variations or parameter variations.
4) Application to the four-level three-cell DC-DC converter: The criterion for this case is a quadratic criterion
with L(z) =k z z0 k2Q and Q = diag[1, 1, 1000]. The 4-step scheme mentioned above can be applied to obtain
the feedback control. The used NN has a feedforward structure with 10 neurons in the hidden layer.
Remark 10: Since the trajectories are generated backward in time from the equilibrium point, by construction
the stability is ensured. Moreover, for all the examples treated until now there is no a big difference between the
direct solution and the NN. Thus, the optimality is not lost due to the control discrete nature. The NN learns the
state feedback and an approximation is observed only on the border of the state space partition.
C. Predictive control Approach
1) Method Description: In the field of DC-DC converters, classical control methods are usually designed with a
constant reference which represents a mean value of the steady state (see section III). Consequently, the switches
are not necessarily controlled around the equilibrium and the steady state may show unexpected harmonic content.
In order to avoid this uncontrolled behavior, we propose to design a state feedback control law, which is able to
track an optimal limit cycle near the operating point instead of a mean value.
Indeed, for a given operating point, an optimal cycle can optimize a criterion. This criterion is tuned in the design
part. It might define a quadratic norm reduction of the tracking error or could be designed as a filter to penalize
undesired frequencies.
Once the optimal cycle is defined, the tracking control is achieved using a predictive controller. This method is
composed by two-stage strategy: near the equilibrium, the optimal cycle is tracking. Far from the equilibrium, only

12

the mean value of the cycle is used as reference. Additionally, the predictive control design takes into account time
constraints due to the physical nature of the switches. These constraints allow realistic and achievable switching
laws to be applied.
As for the optimal approach and in order to avoid excessive computation time, the real time implementation of
the controller is ensured by the use of a NN. Indeed, the optimal policy that gives the mode to be used from the
tracking error is learnt off-line by a NN and applied on-line evaluation the NN activation functions once per sample
time.
Now having described the basic of the proposed methodology, we will detail each point.
2) Closed loop design: Consider the class of affine switched systems described by (8). Once a relation between
control values and operating modes is given (i.e. a one to one map from {0, 1}p {1, 2, , 2p }), we can define:
Definition 11: A switching sequence is a finite sequence represented by:
(T , I)s = {(t1 , i1 ), (t2 , i2 ), . . . , (tj , ij ), . . . , (ts , is )}

(29)

where

T = {0 = t1 , t2 , . . . , tj , . . . , ts } is a strictly increasing time sequence composed by the time values when a


mode is switched on.

I = {i1 , i2 , . . . , ij , . . . , is }, ij {1, . . . , 2p } for j = 1, . . . , s is the mode sequence. A mode ij is switched


on at time tj , j = 1, . . . , s.

s is the (finite) length of the sequence.


a) Determination of the optimal steady state: The aim is to determine the best cycle in steady state. It means

the best switching sequence (T , I)s (1 < s < smax ) which optimizes the quadratic criterion:

J ((T , I )s ) = min

s,T ,I

tf

k z z0 k2Q dt

(30)

subject to the constraints:


z(0) = z(tf ) (periodic)

(31)

tf Tp,max (maximal duration)

(32)

k (tj ) tmin |k (tj ) k (tj+1 )| j = 1, . . . , s


k (tj ) = 1 k = 1, . . . , p

(dwell time)

(33)

k (tj+1 ) = 0 if |k (tj ) k (tj+1 )| =


6 0
where k k2Q is a quadratic norm associated to a symmetric positive definite matrix Q, z0 is the average reference,
tf is a free final time (tf = ts+1 ) which is bounded by Tp,max . Equation (33) imposes a minimum duration equals
to tmin , between two activations of the same switch. k is the time elapsed from the last activation. This equations
set is indeed an integrator with a reset for each switch.

13

Remark 12: In order to reduce the undesired harmonic contents a stable filter can be added to the quadratic term
k z z0 k2Q in (30). If this filter is appropriately designed, then it is possible to concentrate the load current or
voltage spectrum [13].
With (s, I) fixed, the length and the mode of the sequence, the solution of (30) is determined using nonlinear
programming. The procedure is repeated until all admissibles values for (s, I) are tested. We remark that the method
Psmax s
p times. This optimisation is obviously
could be computationally slow since problem (30) need to be solved s=1
performed off-line.

The solution of (30) gives an optimal sequence (T , I )s and the optimal reference R0 (t) for the closed loop
in steady state.
b) Neural Predictive Controller: As mentioned above, the design control and the data for training the neural
network are obtained in the following two-stages:

Far from the optimal limit cycle R0 (t): since the behavior of the system is in a transitory phase, (T , I) are
optimized and we choose to fix s and the receding horizon tf to (s , tf ). The following cost function is
minimized on a grid:
J(z, z0 ) = min
T ,I

t
f

k z z0 k2Q dt

(34)

The first mode and its duration are used for training the network.

Near the limit cycle R0 (t). I and s are fixed to the reference values I and s . The optimization is only done
with respect to the time sequence. The cost function becomes:
Z tf
min
kz R0 k2Q dt
T

(35)

As in the optimal approach, the inputs of the NN are the tracking error ((t) = z(t) R(t)) and the operating
point reference z0 . The output is the optimal policy obtained from the optimization problem (34) or (35) with
R = z0 or R = R0 respectively.
3) Application to the four-level three-cell DC-DC converter: We consider the state equation given by (7) with
z = [vC1 , vC2 , iL ]T . For example, if we solve the optimization problem (30) with Tp,max = 3ms, smax = 12,
tmin = 0.25ms, Q = diag[10, 1, 1000] and taking as operating point z0 = [10, 20, 0.6]T , we find the following
optimal sequence:

(T , I )s = {(0.4ms, 2), (0.65ms, 4), (0.9ms, 6), (1.150ms, 3), (1.4ms, 6), (1.625ms, 3),
(36)
(1.9ms, 6), (2.150ms, 3)}
with an initial condition z(0) = [16.65, 5.26, 0.594]T .
The table I gives the control value with respect to the mode
The matrix Q has been tuned in order to reduce the current oscillation around the mean value.
It can be noticed, from equation (36) that the optimal period which minimizes the oscillations is Tp = 2.150 ms.
The dwell time constraints tmin for each switching component is also verified.

14

TABLE I
TABLE OF MODES

ij

In order to generate all the optimal trajectories, a variation of the initial condition has to be considered. These
trajectories are used to train the NN. For this example, the grid of initial conditions is composed by 1000 points
in each variable. The ranges of the variables are 0 and 2iL0 for the current, 0 and 2vC10 and 0 and 2vC20 for the
capacitor voltages.
The NN interpolates the solution with 20 neurones in the hidden layer with a back-propagation training algorithm
and sigmoid functions. After training, the NN is tested with a set of known solutions not used for the training phase
(Validation set). In the case when the error with the validation set is not acceptable, the number of neurons is
increased in the hidden layer until all validation sets lead to a good result.

V. I MPLEMENTATION
A. The four-level three-cells multi-level converter
The control methods presented in section IV have been tested on a real power converter. The power converter
was made at SUPELEC [30] as a three-phased four-cell power converter. Only one branch was used from the
initial structure and it was reduced to a three-cell power converter. The switches are MOSFET IRFP360 transistors
and the value of each capacitor is 45F . The R L load has been created using a rheostat and an inductance.
Their nominal values are 30 and 0.5H. The input voltage is E = 30V and the reference for the load current is
iL0 = 0.6A.

B. Measurements
The structure contains a built-in circuit which allows the capacitor voltages measurement. These measurements
are made with isolated sensors and the output voltage is obtained via resistive voltage dividers. Due to a calibration
needed for higher voltages than those used in the present experiment, the measurements are affected by noise.
The current in the load is measured using the voltage on the load resistor.

15

C. Controller hardware setup


For the methods implementation, a computer with a dSPACE DS1102 controller board, built around a Texas
Instruments TMS320C1 32 bit floating-point Digital Signal Processor (DSP) is used. The control programs were
codified with Matlab Simulink using the dSPACE toolbox and Real Time Workshop. dSPACE ControlDesk allowed
the on-line access to program parameters and data acquisition.
The control algorithm was implemented in Simulink using S-functions written in C. The program contains also I/O
blocks from the dSPACE toolbox and elementary Simulink blocs for the parameters and references. The sampling
frequency is 4kHz for all methods. The inputs in the program are the measures for capacitors voltage and the load
current. The outputs are the boolean values for the switches 1 , 2 , and 3 . Only three outputs are used, namely
the control values for the top switches. These control signals are inverted for obtaining the control signals for the
bottom switches.

VI. E XPERIMENTAL

RESULTS

In order to validate and to compare the control strategies, a few relevant scenarios have been selected for the
benchmarks tests. These benchmark tests have been implemented in simulation and on a real process.
1) Start-up from zero initial conditions: the control objective is to regulate the voltages on the capacitors and
the current in the load to the reference values specified above using nominal physical parameters.
2) Response to load variations: the converter is in steady state with the nominal physical parameters, excepting
the resistive load, which is equal to Rch =
component is applied to Rch =

4 nom
3 Rch

2 nom
3 Rch

= 20. At the instance t = 0, a step in the resistive

= 40.

The experimental and simulation results are displayed in Fig.2, 3 and 4. In each figure, the top half rows
contain the experimental results and the bottom half rows contain the simulation results. Experimental results on
the benchmark show that the three methods achieve the control objective for the load current and the capacitor
voltages.
Fig.2 contains the start-up evolution. All control methods globally fulfill the objective. They quickly reach the
desired reference. The methods do not display differences regarding the settling time in the load current. It is about
40ms (caused by the high value of the load inductance).
The amplitude of the oscillations in the steady state is directly related to the sampling time. Considering the
average values of iL , vC1 and vC2 , the error is near to zero.
The three methods have been tested using a sampling frequency of 4kHz. It is observed that the oscillations for
the stabilization controller have a bigger amplitude than for the other methods. The reason is that the stabilization
method is depending on the parameter . Indeed, the control must verify that V , otherwise the control must
switch. Since the measurements are subject to quantification noise, an error in the Lyapunov function is introduced
and it affects the switching frequency of the MOSFET. An apparent frequency of 200Hz is observed from the
experimental results.

16

TABLE II
O SCILLATIONS AMPLITUDE OF THE ERROR WITH RESPECT TO THE AVERAGE REFERENCE VALUE IN THE STEADY STATE .

[vC1 , vC1 ]

[vC2 , vC2 ]

[iL , iL ]

Stabilization

[4.85, 5]

[4.8, 3.35]

[0.024, 0.024]

Predictive

[4.75, 2.5]

[5.2, 4.05]

[0.01, 0.013]

Optimal

[2.51, 1.15]

[3.5, 3.27]

[0.013, 0.01]

Method

Fig.3 is a zoom on the steady state showing 15 periods of the control signal and the corresponding state variables
evolution. The time axis is not the same because of the differences in the oscillation frequency of the three
methods. In the table II, the minimum and the maximum oscillation values of the error with respect to the average
reference is shown where vC1 = min (vC1 vC10 ), vC1 = max (vC1 vC10 ), vC2 = min (vC2 vC20 ), vC2 =
max (vC2 vC20 ), iL = min (iL iL0 ), iL = max (iL iL0 ) in the cycle.
For the capacitor voltages, it is observed a particular waveform (cf. Fig. 4) creating the effect of a carrying signal
for the stabilization and the optimal control strategies. This is produced because these methods are asynchronous
with respect to the sampling time.
Considering load variations (more precisely a variation of the resistor component) in average values, the three
methods achieve the control objective. Although, for the predictive and optimal approaches a transitory is produced,
it is not very significant compared to the amplitude of the steady state oscillations in the case of the stabilization
approach.
Regarding the implementation, the stabilization approach is very flexible and very easy to implement. The
predictive and the optimal methods require a big off-line effort and the control law is harder to obtain. Nevertheless,
the predictive control has the advantage to take into account time constrains in design part.
VII. C ONCLUSIONS
In this paper three original control strategies have been implemented on the same platform : a four-level three-cell
serial DC-DC converter.
The neural predictive approach allows to track an optimal limit cycle and to control the waveform. Although the
stability is not guaranteed, this method is a flexible and easy-to-tune approach which may be used to improve
the spectral quality of the output signal. The stabilization approach, based on energetic principles, determines the
control values ensuring the asymptotic stability of the system combined with a control objective. In the case of the
optimal strategy, since the solution is given by the optimization of a quadratic criterion, robustness and stability
is clearly achieved and guaranteed. For these last two methods, the asymptotic stability means that the trajectory
enters in a ball centered on the equilibrium whose radius is depending of the employed switching strategy.
From a methodological point of view, the stabilization approach has the advantage of being easier to synthesize
while other proposed methods imply off-line computation. However, all the methods are simple to implement. At
each sampling time, only a few functions have to be evaluated. On the other hand, the optimal and predictive

17

approaches, even if they are more complex, they seem to present an increased robustness with respect to the
measurement noises and quantification errors.
All these approaches are direct (the control values are Boolean) and asynchronous that is why the fixed sampling
frequency imposed by the experimental process lead to a degradation of the performances. Nevertheless, results are
good and the control objective is achieved for the two investigated scenarios. A higher sampling frequency would
give better results.
Regarding the comparison with classical existing approaches, such as the self-balancing method [3], the
performances of the three approaches presented in this article are superior, in particularly with respect to the
transient. All the proposed control designs are clearly multi-variable and dedicated to switched affine systems. Two
of them ensure the stability which cannot be done using simpler controllers such as PI controller. Moreover, the
tuning phases are easier and the resulting state feedbacks are a pledge for robustness. Consequently, the proposed
methods may be useful and viewed as alternative methods when hard performances involving stability are required.
Further works will consist in the improvement of the methods. For the stabilization approach, an extension to the
case of discontinuous conduction modes (DCM) has to be studied. On the other hand, a modification of the generic
Lyapunov function by the introduction of a weighting matrix would allow the achievement of better performances.
In our opinion, the predictive method can be improved by taking into account the sample frequency in the
control design and in the selected optimal limit cycle. In that case, the switching times will be synchronous with
the sampling period ensuring a better tracking.
Moreover, in a practical point of view, the load must be considered as an unknown parameter. Thus, a load
observer is necessary to guarantee the robustness of the method.
Concerning the optimal control approach, it is fundamentally an asynchronous method, since the switching law
is given from a partition of the state space. A better chattering can be obtained using techniques from sliding modes
[25].
ACKNOWLEDGMENTS

This work was supported by the European Commission research project FP6-IST-511368 Hybrid Control
(HYCON).

The authors would want to thank Prof. Amir Arzandeh from Supelec in Gif-sur-Yvette, France, for his help
in obtaining the experimental results.
R EFERENCES

[1] L. Delmas, T. A. Meynard, and G. Gateau, Computed pwm for flying capacitors multicell converters, in Proceedings of the 2004 IEEE
International Symposium on Industrial Electronics, vol. 2, (Ajaccio, France), pp. 953956, 2004.
[2] T. A. Meynard, M. Fadel, and N. Aouda, Modeling of multilevel converters, IEEE Transactions on Industrial Electronics, vol. 44, pp. 356
364, june 1997.
[3] T. A. Meynard, H. Foch, P. Thomas, J. Courault, R. Jakob, and M. Nahrstaedt, Multicell converters: basic concepts and industry
applications, IEEE Transactions on Industrial Electronics, vol. 49, pp. 955 964, october 2002.

18

[4] J.-S. Lai and F. Z. Peng, Multilevel converters-a new breed of power converters, IEEE Transactions on Industrial Applications, vol. 32,
no. 3, 1996.
[5] M. Perez, J. Rodriguez, J. Pontt, and S. Kouro, Power distribution in hybrid multi-cell converter with nearest level modulation, in
Proceedings of IEEE International Symposium on Industrial Electronics, pp. 736741, 2007.
[6] B. P. McGrath and D. G. Holmes, Analytical modelling of voltage balance dynamics for a flying capacitor multilevel converter, in
Proceedings of 2007 IEEE Power Electronics Specialists Conference, (Orlando), pp. 18101816, 2007.
[7] T. A. Meynard, H. Foch, P. Thomas, J. Courault, R. Jakob, and M. Nahrtaedt, Multicell converters: Basic concepts and industry
applications, IEEE Transactions on Industrial Electronics, vol. 49, pp. 955 964, october 2002.
[8] B. M. Song, J. Lai, J. Chang-Yong, and Y. Dong-Wook, A soft-switching high-voltage active power filter with flyingcapacitors for urban
maglev system applications, in Proceedings of 2007 IEEE Industry Applications Conference, pp. 14611468, 2001.
[9] D. Pinon, M. Fadel, and T. Meynard, Commande par mode de glissement dun hacheur a` deux cellules, Modelisation et commande des
entranements e lectriques EPE99, vol. 5, no. 1, pp. 147 162, 2002.
[10] G. Gateau, M. Fadel, P. Maussion, R. Bensaid, and T. Meynard, Multicell converters: Active control and observation of flying capacitor
voltages, IEEE Transactions on Industrial Electronics, vol. 49, pp. 998 1008, october 2002.
[11] M. F. Escalande, J. C. Vannier, and A. Arzande, Flying capacitor multilevel inverters and dtc motor drive applications, IEEE Transactions
on Industrial Electronics, vol. 49, pp. 809 815, august 2002.
[12] J. M. Maciejowski, Predictive control with constraints. Essex, England: Prentice Hall, 2002.
[13] E. I. Silva, B. P. McGrath, D. E. Quevedo, and G. C. Goodwin, Predictive control of a flying cpacitor converter, in Proceedings of the
2007 American Control Conference, (New York, USA), 2007.
[14] G. Bornard, New control law for capacitor voltage balance in multilevel inverter with switching rate control, in IEEE Annual Meeting
and World Conference on Industrial Applications of Electrical Energy, (Rome), 2000.
[15] J. Buisson, H. Cormerais, and P. Richard, On the stabilization of switching electrical power converters, in Hybrid Systems: Computation
and Control, (Zurich), March 2005.
[16] H. Cormerais, J. Buisson, P. Y. Richard, and C. Morvan, Modelling and passivity based control of switched systems from bond graph
formalism: Application to multicellular converters, Journal of The Franklin Institute, no. 345, pp. 468 488, 2008.
[17] J. A. Sanders and F. Verhulst, Averaging Methods in Nonlinear Dynamical Systems. Springer-Verlag, 1985.
[18] R. DeCarlo, M. Branicky, S. Pettreson, and B. Lennartson, Perspectives and results on the stability and stabilisability of hybrid systems,
in Perspectives and results on the stability and stabilisability of hybrid systems, vol. 88, pp. 10691082, July 2000.
[19] D. Liberzon and A. Morse, Basic problems in stability and design of switched systems, IEEE Control Systems Magazine, vol. 19,
pp. 5970, 1999.
[20] P. Riedinger, C. Iung, and F. Kratz, An optimal control approach for hybrid systems, European Journal of Control, vol. 3, 2003.
[21] H. Sussmann, A maximum principle for hybrid optimal control problems, in Proceedings of the IEEE 42nd Conference on Decision and
Control, (Maui, Hawaii), 2004.
[22] M. Shaikh and P. Caines, On the optimal control of hybrid systems: Analysis and algorithms for trajectory and schedule optimization,
in Proceedings of the IEEE 42nd Conference on Decision and Control, 2003.
[23] B. Bonnard, Singular Trajectories, Feedback Equivalence and the Time Optimal Control Problem. Geometry of feedback and optimal
control, Marcel DEKKER, 1998.
[24] F. J. Y. Chitour and E. Trelat, Genericity results for singular trajectories, Journal of differential Geometry, vol. 73, 2006.
[25] V. Utkin, Sliding modes in control and optimization. , Springer Verlag. Sprinder-Verlag, 1992.
[26] M. Volker, Singular optimal control - the state of the art, vol. CAS-29, 1996.
[27] H. Robbins, A generalised legendre-clebsch-conditions for the singular cases of optimal control, IBM Journal Res. Develop., vol. 11,
1967.
[28] W. Powers, On the order of singular optimal control problems, Journal of Optimization Theory and Applications, vol. 32, 1980.
[29] P. Patterson, Artificial Neural Networks. Singapore: Prentice Hall, 1996.
[30] M. F. Escalante-Gutierrez, Contribution a` la definition de structures optimales donduleurs pour la commande de machines a` courant
alternative par DTC. PhD thesis, Universite de Paris VI, 2001.

19

20

15
10
5
0.15

0.2

0.05

0.1
Time (s)

0.15

15
10
5
0

0.2

20

20

20

15
10
5
0

0.05

0.1
Time (s)

0.15

0.2

Voltage Vc1

25

15
10
5
0

0.05

0.1
Time (s)

0.15

0.05

0.1
Time (s)

0.15

0.2

0.05

0.1
Time (s)

0.15

0.2

0.05

0.1
Time (s)

0.15

0.2

10
5
0

0.2

15

0.8

0.7

0.7

0.6

0.6

0.5

0.5
Current i

0.8

0.4

0.4

0.3

0.3

0.2

0.2

0.1

0.1
0
0

0.05

0.1
Time (s)

0.15

0.2

25

20

20

Voltage Vc2

25

15
10
5
0

vC [V]

0.05

0.1
Time (s)

0.15

15
10
5
0

0.2

25

20

20

Voltage Vc1

25

15
10
5
0

0.05

0.1
Time (s)

0.15

0.05

0.1
Time (s)

0.15

0.2

0.05

0.1
Time (s)

0.15

0.2

0.05

0.1
Time (s)

0.15

0.2

15
10
5
0

0.2

0.8

0.7

0.7

0.6

0.6

0.5

0.5
L

0.8

0.4

0.4

0.3

0.3

0.2

0.2

0.1
0

Fig. 2.

Current i

Voltage Vc1
Current iL

Voltage Vc2

0.1
Time (s)

10

25

Voltage Vc1

0.05

15

25

Current iL

Voltage Vc2

20

Voltage Vc2

20

vC [V]

iL [A]

Predictive approach
25

iL [A]

Optimal approach
25

Voltage Vc1

Voltage Vc2

Stabilization approach
25

0.1

0.05

0.1
Time (s)

0.15

0.2

Comparison between real and simulated response at startup, for the nominal load Rch = 30.

20

10
0.1

0.12

0.14

0.16

0.18

0.2

10
0.1

Voltage Vc1

Voltage Vc

states

10
0
0.1

0.12

0.14

0.16

0.18

Current iL

Current iL

0.8
0.6
0.4
0.1

0.12

0.14

0.16

0.18

0.105

0.11
Time (s)

0.115

0.12

20
10
0
0.1

0.2

Time (s)

0.12

20

Time (s)
20

Voltage Vc2

15

0.105

0.11
Time (s)

0.115

0.8
0.6
0.4
0.1

0.105

0.2

0.11
Time (s)

0.115

0.12

Current iL

20

Predictive approach

Voltage Vc1

Optimal approach
Voltage Vc2

Voltage Vc

Stabilization approach
25

25
20
15
10
0.1

0.105

0.11
Time (s)

0.115

0.12

0.105

0.11
Time (s)

0.115

0.12

0.105

0.11
Time (s)

0.115

0.12

0.11

0.115

0.12

0.115

0.12

0.115

0.12

20
10
0
0.1
0.8
0.6
0.4
0.1

Time (s)
1

0
0.1

0.12

0.14

0.16

0.18

0.5
0
0.1

0.2

0.5

0.105

Time (s)

Time (s)

0.14

0.16

0.18

0.5
0
0.1

0.2

2
0.12

0.105

0.105

0.11

0.115

0.5
0
0.1

0.12

0.105

Time (s)

Time (s)

0.11

Time (s)

1
3

0
0.1

0.12

0.5
0
0.1

0.115

Time (s)

control

0.11

0.5

0.5

0.5

0.5
0
0.1

0
0.1

0.12

0.14

0.16

0.18

0.2

0.16

0.18

0.2

0.105

0.11

0.115

0
0.1

0.12

0.105

Time (s)

0.11

Time (s)

10
0.1

0.12

0.14

Voltage Vc1

states

Voltage Vc

Time (s)
20
10
0
0.1

0.12

0.14

0.16

0.18

0.2

Current iL

Current iL

Time (s)
0.8
0.6
0.4
0.1

0.12

0.14

0.16

0.18

20
10
0
0.1

0.105

0.11
Time (s)

0.115

0.12

Voltage Vc2

15

0.12

Voltage Vc2

20

20
10
0
0.1

0.105

0.11
Time (s)

0.115

0.8
0.6
0.4
0.1

0.105

0.2

0.11
Time (s)

0.115

0.12

Current iL

25

Voltage Vc2

Voltage Vc

Time (s)

20
10
0.1

0.105

0.11
Time (s)

0.115

0.12

0.105

0.11
Time (s)

0.115

0.12

0.105

0.11
Time (s)

0.115

0.12

0.11

0.115

0.12

0.115

0.12

0.115

0.12

20
10
0
0.1
0.8
0.6
0.4
0.1

Time (s)

0.12

0.14

0.16

0.18

0.2

Time (s)

control

1
0.5
0
0.1

0.12

0.14

0.16

0.18

0.2

1
0.5
0
0.1

1
0.5
0
0.1

0.105

0
0.1

0.12

0.14

0.16

0.18

0.2

0.105

Time (s)

0.105

0.11
Time (s)

0.115

0.5
0
0.1

0.12

0.105

0.11

Time (s)
1

0.105

0.11
Time (s)

0.115

Time (s)

Fig. 3.

0
0.1

0.12

3
3

0.5

1
0.5
0
0.1

0.115

0.5

Time (s)
1

0.11
Time (s)

0
0.1

1
0.5

Steady state evolution for states and command, for the nominal load Rch = 30.

0.12

0.5
0
0.1

0.105

0.11

Time (s)

21

Stabilization approach

Optimal approach

Predictive approach

Voltage Vc2

30
25
20
15
10
0.1

vC [V]

0.05

0
Time (s)

0.05

0.1

0.05

0
Time (s)

0.05

0.1

15
10
5

0.8

0.8

0.8

0.75

0.75

0.75
0.7
0.65

0.6
0.55

0.7
0.65
L

0.7
0.65

Current i

iL [A]

Current iL

0
0.1

Current i

Voltage Vc1

20

0.6
0.55

0.6
0.55

0.5

0.5

0.5

0.45

0.45

0.45

0.4
0.1

0.05

0
Time (s)

0.05

0.1

0.4
0.1

0.05

0
Time (s)

0.05

0.1

0.4
0.1

0.05

Voltage Vc2

30
25
20
15
10
0.1

vC [V]

0.05

0
Time (s)

0.05

0.1

0.05

0
Time (s)

0.05

0.1

0.05

0.1

Voltage Vc1

20
15
10
5
0
0.1

0.8
0.75
0.7

iL [A]

Current iL

0.65
0.6
0.55
0.5
0.45
0.4
0.1

Fig. 4.

0.05

0
Time (s)

Comparison between real and simulated responses at a load variation from R = 20 to R = 40.

0
Time (s)

0.05

0.1

You might also like