Beruflich Dokumente
Kultur Dokumente
1
L
+
R
C
L(R
C
+R
load
)
_
v
C
+
v
in
L
d
1
(2.5)
dv
C
dt
=
_
R
load
C (R
load
+R
C
)
_
i
L
_
1
C (R
load
+R
C
)
_
v
C
(2.6)
For the output the following equation holds in both switching positions (no actual
averaging is therefore required):
v
out
= v
C
+C
dv
C
dt
R
C
(2.7)
17
CHAPTER 2. MODELS OF THE CIRCUIT
Equation 2.7 can be expanded by substituting
dv
C
dt
fromEquation 2.6, and after some
algebraic manipulation we obtain:
v
out
= v
C
+
i
L
R
load
v
C
R
load
+R
C
R
C
(2.8)
Equations 2.5-2.6 and 2.8 are the averaged model. It is evident that a matrix version
of these equations in the standard linear state-space description can be obtained
without the need of an additional linearization because, as can be observed (and
as explained previously), the equations are linear in the states and in the inputs by
themselves. The standard matrix formulation is:
x = Ax +Bu
y = Cx +Du
(2.9)
with:
x =
_
i
L
v
C
_
u = d
1
(2.10)
and:
A =
_
R
L
L
+
R
C
R
load
L(R
C
+R
load
)
1
L
+
R
C
L(R
C
+R
load
)
R
load
C(R
load
+R
C
)
1
C(R
load
+R
C
)
_
B =
_
v
in
/L
0
_
C =
_
R
load
R
C
R
load
+R
C
1
R
C
R
load
+R
C
_
D =
_
0
(2.11)
The equivalent buck topology which is achieved by setting d
2
= 1 lacks of the
degrees of freedom which are given by the full (d
1
, d
2
, ) coordinates when they can
be chosen independently in [0, 1]
3
. This is why there is no optimization with respect
to losses for this topology: because there is only one possible d
1
which is able to
reach a given v
out
demand.
This is also why, as it can be veried, a simple feedback approach (as opposed
to a combined feedforward and feedback approach, as discussed later) is enough to
control this scheme.
2.3.2 Boost Model
The circuit in Figure 2.1 becomes equivalent to a synchronous boost converter if the
rst switch, T1, is always on, i.e. if d
1
= 1.
It turns out that this version does not have linear differential equations coming
automatically out of the averaging procedure.
18
2.3. BASIC ANALYTICAL MODELS: FULL-BUCK AND FULL-BOOST
This is one of the most important difculties encountered while developing a
controller for this mode of the circuit and for the whole buck-boost topology: the
set of differential equations governing them is a set of non-linear equations (non-
linear both in the states and in the inputs). It is therefore mandatory to make
some sort of linearization at some point (being it a single linearization or multi-
ple linearizations arising in a PWA approach) in order to apply standard control
techniques. Control of a non-linearized plant should not be undertaken in a rst
approach for the control strategy.
Thus, it is clear that more care is needed to gain the model this time compared
to the buck version, while the procedure is basically the same: rst determine the
set of differential equations valid in each of the two circuit states emerging from the
two switching positions of T3 and T4 (step 1). Then multiply them by the portion of
time in which they are active, and then sum them together (step 2). This is the set
of non-linear differential equations described in the previous paragraph. This set of
equations is then linearized around some operating point, and only then a standard
matrix formulation of the state-space model can be formulated (step 3).
It is to be noted that formulating the set of non-linear equations in terms of some
matrices prior to linearization is possible to a certain extent, but can be confusing
and is therefore avoided here. Also note that the equations are rst merged and
only after that they are linearized.
(step 1) The equations are again gained with a combination of KVL and KCL,
with a procedure basically equal to the previous case; after algebraic manipulations,
we achieve the equations valid while the switch T3 is on (and T4 is off)
di
L
dt
=
1
L
_
v
in
_
i
L
R
L
+v
C
+R
C
R
load
i
L
v
C
R
load
+R
c
__
(2.12)
dv
C
dt
=
1
C
R
load
i
L
v
C
R
load
+R
C
(2.13)
v
out
= v
C
+R
C
R
load
i
L
v
C
R
load
+R
C
(2.14)
On the other hand, while T3 is off (and T4 is on), we have:
di
L
dt
=
1
L
(v
in
i
L
R
L
) (2.15)
dv
C
dt
=
v
C
C (R
load
+R
C
)
(2.16)
v
out
= v
C
+v
C
R
C
R
load
+R
C
(2.17)
19
CHAPTER 2. MODELS OF THE CIRCUIT
(step 2) Now we can merge the equations taking their weighted sum. The mul-
tiplicative weight for the rst set of equations is d
2
, while the weight for the second
set is 1 d
2
. Summing them afterward results in the following set of averaged non-
linear differential equations:
di
L
dt
=
1
L
_
v
in
_
i
L
R
L
+d
2
v
C
+d
2
R
C
R
load
i
L
v
C
R
load
+R
C
__
(2.18)
dv
C
dt
=
1
C
d
2
R
load
i
L
v
C
R
load
+R
C
(2.19)
v
out
= v
C
+R
C
d
2
R
load
i
L
v
C
R
load
+R
C
(2.20)
Notice from this equations the evident non-linearity, which is due to terms such as
d
2
v
C
.
(step 3) The set can now be linearized, and we choose as linearization point the
dummy point x
0
= [i
L,0
, v
C,0
] and u
0
= [d
2,0
]. The result is the following set of
equations:
di
L
dt
=
1
L
_
R
L
+d
2,0
R
C
R
load
R
C
+R
load
_
i
L
+
d
2,0
L
_
1
R
C
R
load
+R
C
_
v
C
+...
... +
1
L
_
v
C,0
+R
C
R
load
i
L,0
v
C,0
R
load
+R
C
_
d
2
(2.21)
dv
C
dt
=
1
C
_
d
2,0
R
load
R
load
+R
C
_
i
L
+
1
C
_
1
R
load
+R
C
_
v
C
+...
... +
1
C
_
R
load
i
L,0
R
load
+R
C
_
d
2
(2.22)
v
out
= d
2,0
_
R
C
R
load
R
load
+R
C
_
i
L
+
_
1
R
C
R
load
+R
C
_
v
C
(2.23)
which can clearly be expressed in the standard state-space matrix form, and used
once numerical values for x
0
and u
0
are chosen. Explicitly, this results in the follow-
ing matrix description
x = Ax +Bu
y = Cx +Du
(2.24)
with:
x =
_
i
L
v
C
_
u = d
2
(2.25)
20
2.4. BUCK-BOOST MODEL
and:
A =
_
_
1
L
_
R
L
+d
2,0
R
C
R
load
R
C
+R
load
_
d
2,0
L
_
1
R
C
R
load
+R
C
_
1
C
_
d
2,0
R
load
R
load
+R
C
_
1
C
_
1
R
load
+R
C
_
_
_
(2.26)
B =
_
_
1
L
_
v
C,0
+R
C
R
load
i
L,0
v
C,0
R
load
+R
C
_
1
C
_
R
load
i
L,0
R
load
+R
C
_
_
_
(2.27)
C =
_
d
2,0
R
C
R
load
R
load
+R
C
1
R
C
R
load
+R
C
_
(2.28)
D =
__
R
C
R
load
R
load
+R
C
_
i
L,0
_
(2.29)
This concludes the introductory discussion on the basic full-boost and full-buck
topologies. In the next Section, the discussion is extended to the complete buck-
boost functionality and the results are going to be shown.
2.4 Buck-Boost Model
T
aking the discussion a step further, in this section the development of the com-
bined buck-boost topology is presented. As noted previously, when exploiting
the degrees of freedom of this circuit, a handful of new features arises which are not
to be encountered in the previous special cases. Among others, these can be cited:
1. Operating the switches as discussed in Section 1.2 gives rise to four feasible
combinations of switching states 2.5 compared to the previous cases where we
only had two intermitting states. This is the reason why this converter is called
a four-state converter.
2. This also directly relates to the fact that the equations for the state-times (the
time portions, during which one of the four state is active each cycle) are more
complex than before; the solution to this problem is given below, see Section
2.4.3.
3. Previously the choice of the duty cycles d
1
(for the buck) and d
2
(for the boost)
was constrained by the v
out
output demand. For instance, in the buck mode, if
the demand was v
out
= 10V with a given v
in
= 20V , then there was only one
feasible d
1
which could achieve that demand (which is d
1
= 0.5 in the ideal
case). This is no longer the case, as it is going to be shown, since an innity of
combinations of (d
1
, d
2
, ) exists that achieves a given v
out
demand.
21
CHAPTER 2. MODELS OF THE CIRCUIT
4. This additional degrees of freedom also enable us to have a better control on
the currents owing inside the circuit while keeping the same v
out
. Exploiting
this property is going to allow us to optimize the switching in order to achieve
better performance in terms of losses.
5. Another important aspect concerning the modeling is that two different mod-
eling approaches are actually going to be taken this time. The rst one is the
usual averaging-technique, which will deliver the model used for the feedback
MPC. The second one is an engineered simulation of the hybrid behavior of
the switched system, described further in Section 2.4.5; the true waveform of
i
L
(and not only its average) will be reproduced with this model, which will
be used as a reliable tool for assessing the losses behaviour.
In the following, the presentation of the different models needed for the evaluation
of the buck-boost functionality is organized as follows: in Section 2.4.1, the Equa-
tions valid for the mean values of the plants variables at steady state is briey de-
scribed; in Section 2.4.2, the Equations governing the circuit behavior in each of the
four possible states are presented, and in the subsequent Section 2.4.3, the algorithm
used in order to determine the state times (the weights of the average) is exposed.
The results of these two Sections will be combined in Section 2.4.4 in order to achieve
the average model, and in Section 2.4.5 in order to obtain the hybrid model.
2.4.1 Estimation Equations
This is a set of equations useful for determining the steady-state mean values of
different signals. They do not account for the presence of the parasitic ESRs which
are otherwise included throughout the rest of the report, but are nonetheless useful.
Output values:
v
out
= v
in
d
1
d
2
(2.30)
i
out
=
v
in
d
1
/d
2
R
load
(2.31)
Capacitance:
v
C
= v
out
(2.32)
i
C
= 0 (2.33)
22
2.4. BUCK-BOOST MODEL
Inductance:
v
L
= 0 (2.34)
i
L
=
v
in
d
1
/d
2
R
load
d
2
(2.35)
2.4.2 Equations for the four States
For each of the four states depicted in Figure 2.5 the corresponding sets of governing
differential equations is given here below. These differential equations are obtained,
as usual, by applying KVL and KCL laws and then by algebraically reorganizing
them, similar to what has been done in Sections 2.3.1 and 2.3.2. The denotations are
to be understood as follows: State 13 means that switches T1 and T3 are currently
on; State 23 means T2 and T3 are on, and so on.
a)
L
RC
C
Rload
L
RC
C
V
in
Rload
L
RC
C
V
in
Rload
L
RC
C
Rload
RL
RL
RL
RL
a)
b) d)
c)
Figure 2.5: Pictures of the circuit in its four possible states.
23
CHAPTER 2. MODELS OF THE CIRCUIT
State 13 (see Figure 2.5.a)
di
L
dt
=
1
L
_
v
in
i
L
_
R
L
+
R
C
1 +R
C
/R
load
_
v
C
_
1
R
C
R
load
+R
C
__
(2.36)
dv
C
dt
=
_
i
L
v
C
R
load
__
1
C (1 +R
C
/R
load
)
_
(2.37)
v
out
= v
C
+R
C
i
L
R
load
v
C
(R
load
+R
C
)
(2.38)
State 14 (see Figure 2.5.b)
di
L
dt
=
1
L
[v
in
R
L
i
L
] (2.39)
dv
C
dt
=
v
C
C (R
C
+R
load
)
(2.40)
v
out
= v
C
R
C
v
C
(R
load
+R
C
)
(2.41)
State 23
4
(see Figure 2.5.c)
di
L
dt
=
1
L
_
i
L
_
R
L
+
R
C
1 +R
C
/R
load
_
v
C
_
1
R
C
R
load
+R
C
__
(2.42)
dv
C
dt
=
_
i
L
v
C
R
load
__
1
C (1 +R
C
/R
load
)
_
(2.43)
v
out
= v
C
+R
C
i
L
R
load
v
C
(R
load
+R
C
)
(2.44)
State 24
5
(see Figure 2.5.d)
di
L
dt
=
R
L
L
i
L
(2.45)
dv
C
dt
=
v
C
C (R
C
+R
load
)
(2.46)
v
out
= v
C
R
C
v
C
(R
load
+R
C
)
(2.47)
2.4.3 Equations for the State-Times
The following equations will model the time the circuit spends in each of the four
states for a given (d
1
, d
2
, ) coordinate. Notice that these equations also respect the
4
Notice that is basically the same set of equations as the previous one, if v
in
is taken to be zero.
5
Again, this is the same set of equations of State 14 if v
in
is set to zero.
24
2.4. BUCK-BOOST MODEL
(relative) order in which these states appear
6
.
Notice how a partitioning in cases was necessary to model the state-times cor-
rectly: a case-choice is to be made each time in order to fetch the valid set of equa-
tions (in gure 2.6 the correspondence between case and switching condition is de-
picted).
a)
b)
c)
d)
e)
f)
Figure 2.6: Diagrams of the different possible gate signals.
The symbols are understood as follows: t
13
is the time spent while T1 and T3 are
on (therefore T2 and T4 are off). Similarly, t
23
means T2 and T3 are on, and so on;
interpret the following as a MATLAB-code snippet (i.e. nested if routines where
necessary).
Case 1 (Figure 2.6.a):
if (phase<=d1) && (mod(phase+d2,1)<= d1)
if ((phase+d2 - mod(phase+d2,1)) == 0)
t14 = phase;
t13 = d2;
6
While the relative sequence of states is not a relevant information for the averaged model, it is in fact
fundamental for properly simulating the switching behavior which is done in Section 2.4.5 in order to
gain the hybrid model
25
CHAPTER 2. MODELS OF THE CIRCUIT
t14 = d1 - (d2+phase);
t24 = 1-d1;
end
Case 2 (Figure 2.6.b):
if ((phase+d2 - mod(phase+d2,1)) = 0)
t14 = phase - (1-d2);
t13 = 1-d2;
t23 = d1-phase;
t24 = 1-d1;
end
end
Case 3 (Figure 2.6.c):
if (phase>=d1) && (mod(phase+d2,1)>= d1)
if ((phase+d2 - mod(phase+d2,1)) == 0)
t14 = d1;
t24 = phase-d1;
t23 = d2;
t24 = 1-phase-d2;
end
Case 4 (Figure 2.6.d):
if ((phase+d2 - mod(phase+d2,1)) = 0)
t13 = d1;
t14 = phase-(1-d2)-d1;
t24 = 1-d2;
t23 = 1-phase;
end
end
Case 5 (Figure 2.6.e):
if (phase<=d1) && (mod(phase+d2,1)>= d1)
t13 = phase;
t14 = d1-phase;
t13 = phase+d2-d1;
t23 = 1-phase-d2;
end
26
2.4. BUCK-BOOST MODEL
Case 6 (Figure 2.6.f):
if (phase>=d1) && (mod(phase+d2,1)<= d1)
t13 = phase-(1-d2);
t23 = d1-(phase-(1-d2));
t24 = phase-d1;
t23 = 1-phase;
end
Based on the results of these two rst Subsections, it is now possible to estabilish
the equations for the averaged and hybrid models.
2.4.4 The Averaged Model
Multiplying the state-equations (2.38)-(2.47) with the corresponding state-times al-
ready provides the set of non-linear differential equations used to get the averaged
model. Explicitly, this results in:
di
L
dt
=
1
L
_
v
in
i
L
_
R
L
+
R
C
1 +R
C
/R
load
_
v
C
_
1
R
C
R
load
+R
C
__
t
13
+...
... +
1
L
[v
in
R
L
i
L
] t
14
+...
... +
1
L
_
i
L
_
R
L
+
R
C
1 +R
C
/R
load
_
v
C
_
1
R
C
R
load
+R
C
__
t
23
+...
... +
R
L
L
i
L
t
24
(2.48)
dv
C
dt
=
_
i
L
v
C
R
load
__
1
C (1 +R
C
/R
load
)
_
t
13
+...
... +
v
C
C (R
C
+R
load
)
t
14
+...
... +
_
i
L
v
C
R
load
__
1
C (1 +R
C
/R
load
)
_
t
23
+...
... +
v
C
C (R
C
+R
load
)
t
24
(2.49)
v
out
=
_
v
C
+R
C
i
L
R
load
v
C
(R
load
+R
C
)
_
(t
13
+t
23
) +...
... +
_
v
C
R
C
v
C
(R
load
+R
C
)
_
(t
14
+t
24
)
(2.50)
27
CHAPTER 2. MODELS OF THE CIRCUIT
Equations 2.48-2.49 can be rewritten in a more compact form as follows:
_
i
L
v
C
_
= F(i
L
, v
C
, v
in
, d
1
, d
2
, ) =
=
_
F
13
_
i
L
v
C
_
+
v
in
L
_
t
13
(d
1
, d
2
, ) +
_
F
14
_
i
L
v
C
_
+
v
in
L
_
t
14
(d
1
, d
2
, ) +...
... +F
23
_
i
L
v
C
_
t
23
(d
1
, d
2
, ) +F
24
_
i
L
v
C
_
t
24
(d
1
, d
2
, )
(2.51)
with:
F
13
= F
23
=
_
_
1
L
_
R
L
+
R
C
1+R
C
/R
load
_
1
L
_
1
R
C
R
load
+R
C
_
_
1
C(1+R
C
/R
load
)
1
R
load
_
1
C(1+R
C
/R
load
)
_
_
(2.52)
F
14
= F
24
=
_
R
L
L
0
_
(2.53)
Since this model is non-linear, like the boost version discussed in Section 2.3.2,
a linearization around some operating point, denoted here again x
0
= (i
L,0
, v
C,0
)
and u
0
= (d
1,0
, d
2,0
,
0
), needs to be made. Such linearization can be written in the
standard state-space matrix form yielding, for A:
A =
F(i
L
, v
C
, v
in
, d
1
, d
2
, )
x(i
L
, v
C
)
x
0
,u
0
=
= F
13
t
13
(d
1,0
, d
2,0
,
0
) +F
14
t
14
(d
1,0
, d
2,0
,
0
) +...
... +F
23
t
23
(d
1,0
, d
2,0
,
0
) +F
24
t
24
(d
1,0
, d
2,0
,
0
) R
22
(2.54)
The values t
13
(d
1,0
, d
2,0
,
0
), t
14
(d
1,0
, d
2,0
,
0
), ..., which are still left undetermined
in this last result have to be determined with the right case-choice from the if-
structure presented in the previous Section.
28
2.4. BUCK-BOOST MODEL
For B, we do partial derivation with respect to the inputs, yielding:
B =
F(i
L
, v
C
, v
in
, d
1
, d
2
, )
u(d
1
, d
2
, )
x
0
,u
0
=
=
__
F
13
_
i
L
v
C
_
+
v
in
L
_
t
13
(d
1
, d
2
, )
u
_
x
0
,u
0
+...
... +
__
F
14
_
i
L
v
C
_
+
v
in
L
_
t
14
(d
1
, d
2
, )
u
_
x
0
,u
0
+...
... +
_
F
23
_
i
L
v
C
_
t
23
(d
1
, d
2
, )
u
_
x
0
,u
0
+...
... +
_
F
24
_
i
L
v
C
_
t
24
(d
1
, d
2
, )
u
_
x
0
,u
0
=
=
_
0
i
L,0
C(1+R
C
/R
load
)
0
v
in
L
R
load
v
C,0
L(R
load
+R
C
)
+
R
C
i
L,0
L(1+R
C
/R
load
)
0
_
R
23
(2.55)
Notice that the last equality in Equation (2.55) is not trivial. First, only after
computation of the partial derivative of the state-times with respect to the inputs,
i.e.:
t
13
(d
1
, d
2
, )
u
x
0
,u
0
,
t
14
(d
1
, d
2
, )
u
x
0
,u
0
,
t
23
(d
1
, d
2
, )
u
x
0
,u
0
,
t
24
(d
1
, d
2
, )
u
x
0
,u
0
it turns out that for each if-case the B matrix looks always the same, while poten-
tially six different matrices could have arisen for each of the six cases. Second, notice
how the last column of B contains only zeros; this conrms the statement made in
Section 1.4, where it was asserted that the mean of the output voltage does not de-
pend on . Based on this model, it actually turns out that none of the mean values
(i.e. also the internal states mean values) changes if the phase is varied
7
.
For the C and D matrices, the derivation of the output Equation with respect to
the states and the inputs has to be made, leading to
8
:
C =
_
R
load
R
C
R
load
+R
C
(t
13
+t
23
) 1
R
C
R
load
+R
C
_
x
0
,u
0
R
12
(2.56)
D =
_
i
L,0
R
load
R
C
R
load
+R
C
(t
13
+t
23
)
u
_
x
0
,u
0
R
13
(2.57)
7
The fact that the output mean value is not inuenced is in fact a consequence of this fact, and therefore
only indirectly a consequence of the fact that the last column of B only contains zeros.
8
In the MATLAB routines attached to this document, approximations for these matrices have usually
been made. In particular, C has been approximated to be C [0 1] and D [0 0 0]
29
CHAPTER 2. MODELS OF THE CIRCUIT
This ends the discussion about the models needed for the feedback part of the
control scheme, discussed in Section 4.4.2.
2.4.5 The Hybrid Model
Hybrid model might sound exotic, but it is developed here in the (probably) most
intuitive way without resorting to tools such as HYSDEL
9
. This is done by using the
information on the state times and their sequence from the previous Sections 2.4.2-
2.4.3: basically, the sequence in which the states appears is known, and their active
time is known as well. The voltage across the inductor is easily estimated with the
following equations:
State 13:
v
L
= v
in
_
v
C
+
_
i
L
v
out
R
load
__
i
L
R
L
(2.58)
State 23:
v
L
=
_
v
C
+
_
i
L
v
out
R
load
__
i
L
R
L
(2.59)
State 14:
v
L
= v
in
i
L
R
L
(2.60)
State 24:
v
L
= i
L
R
L
(2.61)
Then, integrating
di
L
=
v
L
L
dt (2.62)
over the period of time during which each state is active (thereby approximating
with the assumption that v
L
be constant) readily gives the desired waveform with
the small addition that the offset still needs to be adjusted. This can be done either
by using the approximative equation for the mean i
L
at steady state (2.35), which is
done here (because of its simplicity) with good results, or by enforcing the constraint
that the current at the beginning of each period must be equal the current at the end
of it.
9
http://control.ee.ethz.ch/ hybrid/hysdel/
30
2.4. BUCK-BOOST MODEL
Figure 2.2 shows for one given (d
1
, d
2
, ) coordinate that the two approaches
agree up to 0.1%; it can be veried that the prediction for i
L
obtained this way is
accurate also for other choices of u.
As stated previously, the results from this hybrid model are used afterward to
make assessment regarding the behavior of switching and conduction losses; this is
going to be the main topic of Chapter 3.
31
Chapter 3
Power Losses
Ploetzliche Regenfaelle koennen zum Betreten einer
Buchhandlung zwingen.
Loriot
3.1 Introduction
O
ur objective is to drive the circuit while minimizing the losses occurring inside
of it. In order to do this, models for the behavior of these losses are necessary;
i.e. the aim of this Chapter is to present how this mapping, P
loss
= f(d
1
, d
2
, ), can
obtained.
3.2 Losses Description
T
here are two different types of losses occurring inside the circuit: Conduction
Losses (P
Conduction
) and Switching Losses (P
Switching
); in the following, these
two types of losses are going to be shortly described.
Conduction Losses
These are losses of resistive type, and, for the particular circuit that is investi-
gated, they are produced because of current owing through the following resistive
media:
MOSFETs channel resistance R
DS,on
MOSFETs body diode
Capacitances ESR (Equivalent Series Resistance)
Inductances ESR
33
CHAPTER 3. POWER LOSSES
Being these losses of resistive type, they can be modeled to be proportional to the
square of the RMS value of the current owing inside the circuit, i.e.:
P
Conduction
i
2
L,rms
(3.1)
Note: this estimation is already an approximation, since higher losses will usu-
ally cause a rise of temperature, which is going to affect the values of the parasitive
resistances, which in turn is going to affect the losses.
Switching Losses
The mechanisms involved in the production of switching losses are more com-
plicated than the previous ones. They are produced by the action of turning on
and off active devices on the powers path, therefore they only happen at discrete
times t
j
(where j indexes all the times at which switchings of a given MOSFET
occur) and for a short period; they occur under the following circumstances (Mohan
et al. n.d.):
switching of power currents (turning on and off currents in the presence of
voltage)
parasitic drain capacitance charge and discharge
gate drive losses
body diode reverse recovery
In the part of investigation conducted on switching losses, only the rst contribution
(i.e. switching losses due to switching of power currents) has been considered. An
estimate of their magnitude can be obtained as follows:
If the MOSFET is turning on, and the current was not owing through its
body diode, then the switching loss can be estimated as being proportional to
the product of the current that will start to ow through the MOSFET and the
voltage across the MOSFET prior to switching, i.e.:
P
Switching,j
v(t
j
)i
L
(t
+
j
) (3.2)
If the MOSFET is turning off, on the contrary, and the current will not be able
to ow through its body diode, then the switching loss can be estimated as
being proportional to the product of the current that was owing through it
and the voltage that will be applied to the MOSFET after, i.e.:
P
Switching,j
v(t
+
j
)i
L
(t
j
) (3.3)
34
3.3. MODELING POWER LOSSES
because of this, if the current i
L
is positive (owing from the input stage to the out-
put stage), then switching losses will occur only at switches T1 and T4. Conversely,
if i
L
is negative, then switching losses will occur in switches T2 and T3.
On a side note, it can be noted that since these losses occur at switching times,
the more switchings there are, the higher the switching losses will be (if the same
MOSFETs are used), i.e. switching losses grow proportionally to the switching fre-
quency. Therefore, on one hand, switching frequency should not be chosen to be
arbitrarily high. But on the other hand, switching frequency should not be chosen
too low either because that would cause higher ripples on the output voltage.
Also, it is of critical importance to note at this point that during the simulations
described further in this chapter, the magnitude of the losses is estimated using these
very equations. But since these equations only give results that are proportional to the
exact values, their shape will describe the general behavior of the losses properly,
but their magnitude will need to be corrected by an adequate multiplicative correc-
tion constant. This constant will strongly depend on the choice of components that
is going to be made. This aspect is discussed more in detail in the next Section.
3.3 Modeling Power Losses
I
tis possible to achieve a characterization of the switching and conduction losses
using the equations of above, by estimating their value for each choice of (d
1
, d
2
, ).
This is accomplished by using the model (2.58)-(2.61) discussed in Section 2.4.5,
which reproduces the hybrid behavior of the plant; explicitly, for each combination
of (d
1
, d
2
, ), the shape of i
L
is calculated. This enables the calculation of the con-
duction losses. Then, the equations for the state-times exposed in Section 2.4.3 are
used to fetch the sequence of switchings occurring during each cycle, and the times
at which each switching occurs. This information is sufcient in order to calculate
the switching losses.
Notice that such a detailed description (in contrast to a simplied version given
by the averaged models) is needed in order to assess losses, rst because the RMS
value of the current is needed (which can only be calculated if the correct shape of
the i
L
current is known), and second because the losses also depend on the sequence
of switchings which is adopted, which therefore needs to be known (what switch is
switched when, so that the current can be multiplied with the correct voltage).
The losses have been investigated for many v
out
/v
in
ratios, i.e. the set of (d
1
, d
2
, )
combinations giving the same output voltage has been gridded and then simulated
for each of the resulting knots. Typical results for lowoutput references and for high
output references are shown in Figure 3.3 and 3.3.
35
CHAPTER 3. POWER LOSSES
Figure 3.1: Switching and Conduction Losses for a low v
out
/v
in
demand. Values are not
scaled properly, but do give an idea of the percentages. Notice that the switching losses have
a pretty marked behavior in this case; this is probably due to resonant occurrences inside the
circuit which should be investigated further. This image has been selected mainly because it
shows well howconduction and switching loss minima-locations do not necessarily coincide.
A fundamental result of this study is shown in Figure 3.3: it is evident from the
two pictures that the positions at which conduction losses and switching losses
are minimal do not necessarily coincide. The same result can be veried in Figure
36
3.3. MODELING POWER LOSSES
Figure 3.2: Switching and Conduction Losses for a high v
out
/v
in
demand. Values are not
scaled properly, but do give an idea of the percentages. The switching losses behavior de-
picted for this case is more representative of what is usually encountered.
3.3. This poses a challenge in the following sense: if the two minimal positions
would coincide, then the minimum of the sum of the two losses would also be at
that same place; but since they do not, it is important to sum up together the correct
values for conduction and switching losses, in order to achieve the correct position
of the minimum of the total. On one side, this means that sometimes there needs to
37
CHAPTER 3. POWER LOSSES
be an optimized trade-off between these two types of losses. On the other hand, this
also means that the exact values of the two losses will inuence the position of the
minimum of the sum. Or stated another way round: we saw that the exact values
of switching and conduction losses still depend on the value of the multiplicative
correction terms that need to be applied; these values in turn depend on the choice
of components, and as a consequence the value and position of the minimum total
loss will depend on the choice of components.
2.96 2.964 2.968 2.972 2.976 2.98
1.1
1.11
1.12
1.13
1.14
1.15
1.16
1.17
Conduction Losses vs
Swiching Losses
time (ms)
i
L
(
A
)
2.96 2.964 2.968 2.972 2.976 2.98
0
5
10
15
20
25
30
time (ms)
S
w
i
t
c
h
i
n
g
L
o
s
s
(
~
)
phase = 0.8
phase = 0.2
Figure 3.3: Comparison between conduction (upper graph) and switching losses (lower
graph) for two different (d
1
, d
2
, ) combinations giving both v
out
/v
in
= 0.2 (both have
d
1
= 0.14, d
2
= 0.7). On the rst graph, it is clear that the blue curve, with = 0.2 has
overall lower conuction losses, because its RMS value is lower than the red one. On the lower
graph, however, where switching losses are depicted (as losses occurring only at discrete
times) it is clear that the blue inputs combination delivers higher switching losses. Therefore,
it is veried that conduction and switching losses do not necessarily have coinciding minimal
points, since lowering one might cause an increase in the other.
For the successive development of the controller, which will depend on these
38
3.3. MODELING POWER LOSSES
results, a certain choice for the corrective weights has been made, which is in no
way bound to any actual choice of components - as it should. This choice has been
made for prototyping purposes, and appropriate weight still need to be determined
from a proper selection of components.
39
Chapter 4
Controller Design
Short cuts make long delays.
Hobbit Proverb
4.1 Introduction
B
ased on the research done on the models in Chapter 2 and the Losses in Chapter
3, it is now possible to start developing an efcient control for the plant. As a
reminder, our task is to control the duty cycles of each pair of transistors and their
phase, so as to ensure:
rst and most important: reaching of and stabilizing around a given output
voltage demand;
reaching the target steady state should happen in the desired manner, i.e. the
controller needs to handle transients properly;
the controller also needs to be able to reject disturbances (usually encountered
on the load and on the input voltage source v
in
);
while doing all this, the controller (in the full buck-boost mode) needs to
choose among the innite possibilities of inputs, that would satisfy the above
conditions, those that will cause the least losses.
Notice that, as explained previously, minimization of losses is only possible if all
the degrees of freedom of the circuit are exploited, i.e. there is no optimization of
control strategy towards minimal losses for the buck or boost modes alone.
The presentation of the development of such a controller is going to be handled
in the next Sections as follows: rst the choice of MPC as control approach for the
feedback is motivated; in Section 4.3 the control techniques deployed for the basic
versions of the circuit (full-buck and full-boost) are discussed, based on which the
more complex control for the whole buck-boost functionality is nally going to be
discussed, Section 4.4.
41
CHAPTER 4. CONTROLLER DESIGN
4.2 Why MPC
MPC was chosen based on the following reasons:
its ability to efciently deal with constraints
because standard control techniques such as a PID implementation delivered
unsatisfactory results (Frehner 2007)
because of its robustness to model uncertainties
and because it works for multiple inputs plants exactly the same way as it
works for the SISO ones, thereby allowing analogous implementations for the
buck/boost (one input) and buck+boost (three inputs) modes; on the contrary,
traditional implementations using control techniques such as PIDs present
many challenges when applied to multiple inputs plants, especially when it
comes to the assessment of control performances such as stability, robustness
and disturbances rejection.
Also, while MPC was previously only applied to slow-dynamics plants, be-
cause the control action needed to be calculated from an optimization problem each
time (on-line control), recent developments made it possible to move the burden of
calculating the control moves off-line, through what is called a multi-parametric
(quadratic in this case) program approach (mpQP), thus allowing to apply MPCs
also to plants with fast dynamics (IfA Website
1
).
Calculating the input move on-line would consist in solving a quadratic program
(QP) each time (if 2-norm cost functions are taken - as it is usually the case), and a
multiparametric approach just means that this optimization problemis parametrized
and solved for a certain set of parameters x (this is the set of all the xs contained
in a predened constrained set X), and during runtime the correct input is easily
computed by plugging in the parametrized solution the appropriate current state x.
Multiparameter problems allow to calculate so-called explicit MPCs. An explitic
MPCis therefore nothing more than a look-up table that will return the control move
to be taken given some current state x. Notice that this is the same control move that
would be applied if an on-line MPC strategy would be chosen instead of the explicit
one.
Since the control moves are pre-calculated moving around the whole feasible
U X set, it is necessary to constrain it in order to make this procedure end in nite
time. This is why this technique can be applied only to constrained problems. For
this particular plant, the inputs, which are basically the duty cycles and the phase,
1
see (IfA n.d.) http://control.ee.ethz.ch/ hybrid/control.php
42
4.3. BASIC CONTROL STRATEGIES
need to be bounded between 0 and 1. This results in the inputs set U to be delimited
to:
U =
_
u R
3
| u [0, 1]
3
_
(4.1)
X, on the other hand, is only loosely constrained, i.e. it is taken to be much larger
then the actual values of the states are ever going to be driven at, just to make the
computation feasible. Notice that other restrictions could be made, in particular a
restriction on the inductor current to be non-negative (if discontinuous mode has
to be considered) and other restrictions to ensure soft-start requirements could be
made
2
.
4.3 Basic Control Strategies
4.3.1 Buck, the Simplest Mode
A
s discussed in Chapter 2.3.1, the model obtained with the averaging technique
is linear. There is only one variable being controlled (d
1
) and there is no opti-
mization of controls towards least losses. This is why a simple feedback approach
(as opposed to a combined feedforward and feedback approach, as discussed later)
is enough to control this scheme.
The feedback control is implemented as an explicit MPC. The MPC is to be fed
with a discrete time model which is easily obtained from the continuous model of
Equations (2.24)-(2.29) with the c2d() routine in MATLAB . Thus, starting fromthese
equations and plugging in testing values for the components reported in Table 4.1,
the following discrete-time model results:
x
k+1
= Ax
k
+Bu
k
(4.2)
y
k
= Cx
k
+Du
k
(4.3)
with:
A =
_
0.9992 0.0498
0.03321 0.9925
_
B =
_
0.9997
0.01662
_
C =
_
0.9998 0.0009998
D = [0];
2
http://control.ee.ethz.ch/ hybrid/applications.php#powerElectronics
43
CHAPTER 4. CONTROLLER DESIGN
Full-Buck, Full Boost v
in
20V
L 2e
5
H
C 3e
5
F
R
load
5
R
L
0.001
R
C
0.001
Buck-Boost v
in
20V
L 2e
4
H
C 3e
5
F
R
load
5
R
L
0.001
R
C
0.001
Sampling Frequency T
S
100kHz
Table 4.1: Values of the components used to test the behavior of the controller.
A typical start-up response of the controlled output can be seen in Figure 4.1.
The details of this implementation, such as the inuence of the weights and of the
prediction horizon, are not discussed further, since this is a special case of the com-
plete buck-boost topology; such discussions are going to be made for the general
case in Section 2.4.
4.3.2 Boost Mode and the Gain Scheduling Technique
A
nalogous to the Buck operation mode, only one variable is being controlled
(d
2
) and no optimization towards least losses can be done. Therefore also in
this case a simple feedback controlling strategy (i.e. not a combined feedforward +
feedback) is taken. And again, the feedback is based on a MPC approach, and the
discrete-time model is obtained with c2d() like before.
But the boost operation mode presents a challenge which was not encountered
for the buck mode: since the model obtained with the averaging technique is not
linear, the matrices obtained linearizing around x
0
and u
0
still depend on the nu-
merical values of x
0
and u
0
, as can be seen from Equations (2.24)-(2.29). This could
indeed be a major problem because basically it means that different steady-state op-
erating points have different dynamics. Notice that this challenge is to be faced in
the full buck-boost operation mode as well.
Fortunately, it turns out that while the plant driven in boost mode is indeed non-
linear, it still behaves well (see below for what is meant with well).
This fact has been exploited both for the boost and the buck-boost mode. The
reasoning behind it is the following: the fact that the model is non-linear is a reality
which cannot be avoided and needs to be accounted for; the most basic approach to
44
4.3. BASIC CONTROL STRATEGIES
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
0
1
2
3
4
5
6
7
Full-Buck Output Performance
time (ms)
v
o
u
t
(
V
)
Figure 4.1: Startup performance of the controller developed for the full-buck operation, with
v
out,ref
= 6V . Small ripples are visible at steady state, since noise was added to v
in
in order
to make a rst assessment of noise rejection performance.
face this problem from a control perspective is to apply a Gain Scheduling Tech-
nique
3
: this technique basically consists in linearizing the system around as many
steady-state points as possible (and at least so many as to ensure the basic require-
ments of controlling) and then to design a control for each of these linearizations; af-
terwards switching policies between these multiple controls need to be established,
thereby obtaining a control that works for the whole region.
In the boost case, due to its low degree of non-linearity, only one of such lin-
earizations is strictly necessary in order to ensure reaching of the target steady-state
v
out
and stabilization around that point. Of course, the best performance is obtained
when the target steady-state point is at or near to the point around which the model
of the plant was linearized. Taking output demands far from that point result in a
degradation of performance during transients. The extent of this performance loss
depends on the choice of components, but this has not been investigated further;
yet, later in Section 2.4 the discussion on the whole buck-boost functionality is go-
ing to lead to this same topic again, and it is going to be shown that indeed the plant
behaves well.
3
http://control.ee.ethz.ch/ apnoco/Lectures/lec08.pdf
45
CHAPTER 4. CONTROLLER DESIGN
Pluggin numerical values for the components in the continuous model (2.24)-
(2.29), choosing as a linearization point v
out
/v
in
= 1.5 and then obtaining the discrete-
time model from c2d() results in the following matrix description:
x
k+1
= Ax
k
+Bu
k
(4.4)
y
k
= Cx
k
+Du
k
(4.5)
with:
A =
_
0.9998 0.01993
0.01329 0.9932
_
B =
_
2.509
0.8137
_
C =
_
0.0003999 0.9998
D = [0.025];
The control scheme deployed for this mode can be seen in Figure 4.2. A typical
start-up response of the controlled output can be seen in Figure 4.3. Also in this
case the details of the implementation, such as the inuence of the weights and
of the prediction horizon, are not discussed further since this is a special case of
the complete buck-boost topology; such discussions are going to be made for the
general case in the next Section.
4.4 Buck-Boost Operation
T
he buck-boost implementation is similar to the boost one in that non-linearity is
still present. Other than that, it turns out that exploiting the possibilities given
by the full buck-boost operation requires additional care because:
there are now multiple inputs
control actions also need to drive the plant while ensuring least possible losses
While going from a single input to a multiple inputs plant is done smoothly
with an MPC approach, losses control require additional care. In this work, the way
the handling of losses has been achieved is by applying a combined feedforward
and feedback control structure (rather than a feedback control alone, as done pre-
viously). A simplied abstraction of the control scheme that enlightens the control
ow can be seen in the block diagram in Figure 4.4. The more detailed Simulink
implementation, including all the subsystems used, can be seen in Figure 4.4.
46
4.4. BUCK-BOOST OPERATION
Figure 4.2: Control scheme deployed for the full-boost operation mode. The feedforward ele-
ment is nothing more than an implementation of the Equations 2.4.1 used to achieve approx-
imatively the right output voltage. The ner control towards the exact value is undertaken
by the feedback MPC. The addition of noise on v
in
is also visible. The PLECS circuit was
substituted with an implementation of the Equatins in 2.3.2.
4.4.1 The Feedforward
The feedforward control (denoted as precalculated look-up table in the block dia-
gram) is responsible for the major contribution to the signal u. It is by far the sim-
plest approach with which complications such as proper treatment of losses can be
handled.
The feedforward table is constructed as follows: recall that in Chapter 3 the in-
vestigations led to two surfaces for each v
out
/v
in
demand telling us how switching
and conduction losses behave with respect to the inputs
4
. The information from
these surfaces has been used here as follows: the total loss for each given point is a
weighted sum of switching and conduction losses. The weights can be adjusted and
can potentially lead to different feedforward tables, but the essence of the procedure
is the same. In this work, the weighting has been the following
5
:
P
loss,Total
= 0.7P
switching
+ 0.3P
conduction
(4.6)
These total losses are then compared between each other for each v
out
/v
in
demand
and the combination of inputs that leads to the least total loss is chosen and inserted
4
Notice that all the losses mentioned in the current discussion are losses calculated at steady state.
5
As stated, the weights could be different based on the true value of the components chosen (and they
do not necessarily need to add up to 1).
47
CHAPTER 4. CONTROLLER DESIGN
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
0
5
10
15
20
25
30
35
40
Full-Boost Output Performance
time (ms)
v
o
u
t
(
V
)
Figure 4.3: Startup performance of the controller developed for the full-boost operation, with
v
out,ref
= 25V . Again, small ripples are visible at steady state, since noise was added to
v
in
in order to make a rst assessment of noise rejection performance. Further, the initial
overshoot can be regulated with a proper choice of the weights, which is not shown but that
can be veried.
v
in
v
out,ref
Precalculated
Steady State Controls
(look-up table)
+
MPC-Based Feedback
PWM
Circuit Model
(PLECS)
v
x
x
0
u
0
u
out
Figure 4.4: Block diagram of the control scheme deployed for the full buck-boost operation
mode.
in the feedforward lookup table.
It is to be noted again that, under these circumstances, changing the weights can
and will change the optimal u to be plugged into the feedforward table.
Feedforward control therefore ensures that among all the innite possibilities of
48
4.4. BUCK-BOOST OPERATION
Global Scheme
Feedback and Feedforward Blocks
PWM Block
Figure 4.5: Detailed scheme of the control scheme deployed for the whole buck-boost func-
tionality.
49
CHAPTER 4. CONTROLLER DESIGN
(d
1
, d
2
, ) control combinations that will steer the plant to the desired output de-
mand at steady state, only the ones around that one which will cause the least losses
are going to be selected.
In general, feedforward control ensures that the plant will be steered to (at least
approximately) the right state even if the feedback were to be switched off. It is
straightforward in that no mathematical model needs to be elaborated and then
(somehow) patched together with the MPC, which would considerably increase its
complexity and size (in memory).
The typical content of a feedfoward table is depicted in Figure 4.4; notice how,
for the specic choice of weights of above, the control usually tries to drive the plant
with the highest d
2
possible. This is directly related with the fact that the current in
the inductor i
L
(which increases with decreasing d
2
, see Equations in Section 2.4.1)
is a direct measure for the conduction losses.
Figure 4.6: Figure depicting the typical content of a feedforward table. Each point corre-
sponds to a certain v
out
/v
in
ratio.
50
4.4. BUCK-BOOST OPERATION
4.4.2 The Feedback
The feedback part is going to be similar to the previous implementations (Section
4.3.2), i.e. it is the part of the controlling structure that will ensure reaching of the
desired steady-state output and rejection of disturbances.
The difference from the previous cases is that this time the MPC will steer the
plant to x = 0 R
n
(therefore y = 0 R). So instead of handling controls towards
absolute values for the output, the feedback will correct signals deviating from the
desired steady state, i.e. the input to the MPC is not going to be x anymore (remem-
ber, MPC is a state-feedback control) but rather x x
0
, a strategy usually referred
to as control towards origin for obvious reasons. For this reason, one can expect
that the control actions from the MPC are going to be zero once the steady state has
been achieved, if the feedforward control was precise enough and if there are no
disturbances.
Since the plant is non-linear, the gain scheduling technique is going to be used
also here. This time, however, two different linearizations are going to be made
instead of one. This is due to the plant working in two regions where different set
of equations for the state times hold (see Section 2.4.3). Explicitly,
For the range v
out
/v
in
[0.05, 0.4], the set of equations for the state-times that
is valid is the 5th
For the range v
out
/v
in
[0.4, 4], the set of valid equatins is the 2nd one instead.
The simplest approach to deal with this aspect is therefore to linearize once around
some point contained in the rst range, and then linearize once more around an-
other point in the second range. For this work, the two linearization points chosen
are v
out
/v
in
= 0.2 for the lower linearization, and v
out
/v
in
= 1.5 for the upper
one. The values of u
0
and x
0
deriving from these choices, which are what is actu-
ally needed, are going to be fetched from the feedforward control component (see
below).
The policy adopted to switch between these two MPCs is straightforward: if
v
out
/v
in
is smaller than 0.4 then choose the rst one, if v
out
/v
in
is larger than 4, then
choose the second one. Two things need to be noted here:
output demands out of the actual ranges where the equations are valid (i.e.
v
out
/v
in
smaller than 0.05 or larger than 4) are still treated with the correspond-
ing one of these two MPCs, because they shouldnt be requested to start with,
and even if they are, the output performance is still decent (see Figure 4.12) in
part due to the feedforward action.
disturbances in v
in
should also be considered. If the input voltage v
in
is mea-
sured, then the correct v
out
/v
in
ratio is calculated and the correct lineariztion
51
CHAPTER 4. CONTROLLER DESIGN
is chosen, therefore no problem arises. If instead v
in
cannot be measured, and
disturbances to it would actually cause a switch of used controller, but this
switches are not accomplished because these disturbances are not measured,
then the need to rely on the robustness of the control arises. This is an as-
pect that should be investigated further (along with a precise assessment of
the controllers robustness and stability).
The controller has been designed with the help of the mpt studio routine in-
cluded in the mpt toolboox for MATLAB (R)
6
. Other than choosing a control to-
wards origin policy for the reasons explained above, the feedback MPC has been
designed as follows: to calculate the cost function, the 2-Norm has been chosen;
the weights on the states and on the output and the prediction horizon have been
determined iteratively; the procedure has been the following: rst, look at the in-
uence to the output curve given by the output weight; since the output voltage is
basically the same signal as v
C
, this weight is also going to be the weight to the v
C
state (therefore, a weight of zero is going to be applied to the state v
C
, for otherwise
deviations of v
C
from the desired reference would be accounted twice for); for the
lower MPC implementation a weight on the output of 10 has been chosen, see
Figure 4.4.2.
The prediction horizon has also iteratively been chosen to be 5, and its inuence
can be seen in Figure 4.4.2. In general, it is desirable to have the highest prediction
horizon possible. High prediction horizons improve stabilization characteristics of
MPCs and in general provide better output performances; yet they cannot be chosen
arbitrarily high, since computation of the control self increases dramatically with
increasing prediction horizon, and its size (in bytes) increases considerably as well,
thus potentially making impossible their implementation on platforms with limited
storage capacities such as DSPs. After this step, the inuence of the weight on the i
L
state has been considered, and as can be seen in Figure 4.4.2, even while varying this
parameter by many orders of magnitude, there seems to be no appreciable inuence
to the output voltage signal once the other two weights are established.
The same procedure has been applied also for the upper MPC, see Figure 4.4.2.
4.4.3 Results
It can be argued that if this precalculated lookup table does indeed contain the best
values the plant (circuit) can be driven at steady state, then the contribution fromthe
MPC feedback can be avoided. This is of course not the case: rst, looking at Figure
4.9, it is clear that the contribution from the MPC boosts the performance during
6
http://control.ee.ethz.ch/ mpt/
52
4.4. BUCK-BOOST OPERATION
the initial transient. Furthermore, a feedback action is always desired in any control
scheme, in order to ensure the ability to reject disturbances and model uncertainties.
The two contributions to the u signal coming from the feedback and from the
feedfoward part can be seen in Figure 4.10. As it can be seen, the MPC supplies the
plant with a contribution different than zero only during the transient. As soon as
the transient has settled, it contribution goes to zero and stays there; this is always
the case as long as no disturbances or other external inuences affect the circuit; if
disturbances are indeed applied, then the MPC control is going to counter those and
its contribution is going to be different than zero.
A typical disturbance rejection done by the controller can be seen in Figure 4.11:
the blue bottom curve depicts the perturbation (in percentage) affecting the input
voltage v
in
, while in the upper graph, the red curve shows how this perturbation
affects the output voltage if no feedback action is taken, and the green one shows
the output if rejections are countered by the MPC.
The resulting output start-up performance for a set of different output references
can be seen in Figure 4.12; notice that the controller is indeed able to drive the circuit
both in its buck mode and boost mode, as specied in the objectives for this
project. Further, notice that the control is indeed able to properly drive the plant also
towards steady states different than those around which the models were linearized,
thus showing its well behavior.
53
CHAPTER 4. CONTROLLER DESIGN
0 0.2 0.4 0.6 0.8 1 1.2
-5
0
5
10
15
Output Weight Dependence
time (ms)
v
o
u
t
(
V
)
0 0.2 0.4 0.6 0.8 1 1.2
0
0.5
1
1.5
2
2.5
3
time (ms)
i
L
(
A
)
0.1
1
100
0.1
1
100
0 0.2 0.4 0.6 0.8 1 1.2
0
0.5
1
1.5
2
2.5
3
time (ms)
i
L
(
A
)
0 0.2 0.4 0.6 0.8 1 1.2
-5
0
5
10
15
i
L
State Weight Dependence
time (ms)
v
o
u
t
(
V
)
0 0.2 0.4 0.6 0.8 1 1.2
-5
0
5
10
15
Prediction Horizon Dependence
time (ms)
v
o
u
t
(
V
)
0 0.2 0.4 0.6 0.8 1 1.2
0
0.5
1
1.5
2
2.5
3
3.5
time(ms)
i
L
(
A
)
1
2
5
10
1
2
5
10
Figure 4.7: These Figures show the change in output response and in the i
L
behavior with
respect to changes to weightings for the lower part of the controller.
54
4.4. BUCK-BOOST OPERATION
0 0.2 0.4 0.6 0.8 1 1.2
-10
0
10
20
30
40
Output Weight Dependency
time (ms)
v
o
u
t
(
V
)
0 0.2 0.4 0.6 0.8 1 1.2
0
2
4
6
8
10
12
time (ms)
i
L
(
A
)
0.01
0.1
1
100
0.01
0.1
1
100
0 0.2 0.4 0.6 0.8 1 1.2
-10
0
10
20
30
40
i Weight Dependency
time (ms)
v
o
u
t
(
V
)
0 0.2 0.4 0.6 0.8 1 1.2
0
2
4
6
8
10
12
time (ms)
i
L
(
A
)
0.01
0.1
1
100
0.01
0.1
1
100
L
0 0.2 0.4 0.6 0.8 1 1.2
-10
0
10
20
30
40
Prediction Horizon Dependence
time (ms)
v
o
u
t
(
V
)
0 0.2 0.4 0.6 0.8 1 1.2
0
5
10
15
time (ms)
i
L
(
A
)
1
2
5
10
1
2
5
10
Figure 4.8: These Figures show the change in output response and in the i
L
behaviour with
respect to changes to weightings for the upper part of the controller. As before, there is no
noticeable inuence from the i
L
weight.
55
CHAPTER 4. CONTROLLER DESIGN
0 0.5 1 1.5 2 2.5 3 3.5
-5
0
5
10
15
20
25
30
35
40
Feedforward vs
Feedforward + Feedback
time (ms)
v
o
u
t
(
V
)
feedforward
feedforward + feedback
Figure 4.9: Difference of output behaviour if the MPC is deactivated.
0 0.2 0.4 0.6 0.8 1 1.2
0.4
0.6
0.8
1
Typical Control Action
time (ms)
0 0.2 0.4 0.6 0.8 1 1.2
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
time (ms)
d1
0
d2
0
phase
0
d1*
d2*
phase*
Figure 4.10: Typical control actions of the two parts of the controller. The upper graph shows
the control actions supplied by the precalculated look-up table (which are therefore constant
signals) and the lower graph shows the signals that are superimposed generated by the MPC.
56
4.4. BUCK-BOOST OPERATION
1.5 2 2.5 3 3.5 4 4.5 5
-5
-3.33
-1.67
0
1.67
3.33
5
6.67
8.33
v
in
Perturbations Rejection
time (ms)
v
o
u
t
d
e
v
i
a
t
i
o
n
%
1.5 2 2.5 3 3.5 4 4.5 5
-4
-3
-2
-1
0
1
2
3
time (ms)
v
i
n
p
e
r
t
u
r
b
a
t
i
o
n
%
without Feedback
with Feedback
Figure 4.11: Disturbances rejection behavior, with and without the MPC action.
57
CHAPTER 4. CONTROLLER DESIGN
0 0.2 0.4 0.6 0.8 1 1.2
0
5
10
15
20
25
30
35
40
Start-Up Performance
time (ms)
v
o
u
t
(
V
)
2V
5V
10V
15V
20V
25V
30V
40V
Figure 4.12: Start-up performance for different output demands.
58
Chapter 5
Conclusion
The present work is a study on the circuit depicted in Figure 1.2 which is used to
achieve DC-DC power conversion.
In the rst part of the work (Chapters 1-3), different models for its behaviour
have been developed, including a state-space averaged model and an hybrid one.
Based on these models simulations have been conducted in order to assess the losses
occurring inside of it. These simulations reveal that it is in general not possible to
drive the circuit while minimizing simultaneously both conduction and switching
losses. Rather, in order to drive the circuit in the most efcent way, an optimized
balance between these two losses needs to be made. Further, this balance depends
on the specic choice of components used.
In the second part of the work (Chapter 4), for a specic choice of components,
the implementation of a controller for this circuit is discussed. The controller has
been designed as working on the combined action of a precalculated look-up table
(feedforward action) and a Model Predictive Control (MPC) based feedback action.
The abilty to drive the circuit both in its boost as well its buck modes and its noise
rejection capabilty are the performance benchmarks for this controller which have
been studied.
Recommended extensions to this work include the renement of the models to
account for parasitics and non-ideal behaviours, so as to enable a subsequent con-
troller implementation based solely on MPC, and a more accurate evaluation of the
controllers stabilization capabilties.
59
Bibliography
Cuk, S. and Middlebrooks, R.: n.d., Modeling, analysis and design of switching converters.
Erickson, R.: n.d., Dc-dc power converters, Article in Wiley Encyclopedia of Electrical and Elec-
tronics Engineering .
Frehner, P.: 2007, Control of a buck-boost power converter, Semester Thesis, IFA, ETHZ .
IfA, W.: n.d., http://control.ee.ethz.ch/ hybrid/control.php, Institut fr Automatik Website, IFA,
ETHZ .
Lin, Y.-C. and Liaw, D.-C.: 2006, A method using an averaging technique for the analysis and
evaluation of real quasi-resonant converters, IEICE TRANS. ELECTGRON. .
Mohan, N., Undeland, T. M. and Robbins, W. P.: n.d., Power electronics: Converters, applica-
tions, and design.
Note, A.: n.d., Synchronous rectication aids low-voltage power supplies.
Plesnik, M.: 2006, Use of the state-space averaging technique in fast steady-state simulation
algorithms for switching power converters, IEEE CCECE/CCGEI .
Selders, R.: 2003, Synchrnous rectication in high-performance power converter design,
POWER designer .
Torrisi, F. D., Bemporad, A., Bertini, G., Hertach, P., Jost, D. and Mignone, D.: 2002, Hysdel
2.0.5 - user manual, IEEE CCECE/CCGEI .
Walter, S.: 2008, Design and control of a buck-boost dc-dc power converter, Semester Thesis,
IFA, ETHZ .
61