Sie sind auf Seite 1von 10

1

Multiobjective Switching State Selector for


Finite States Model Predictive Control based on
Fuzzy Decision Making in a Matrix Converter
Felipe Villarroel, José R. Espinoza, Member, IEEE, Christian A. Rojas, Member, IEEE,
Jose Rodriguez, Fellow Member, IEEE, Marco Rivera, Member, IEEE, and Daniel Sbárbaro, Senior Member, IEEE

Abstract—Finite States Model Predictive Control is a rising hybrid/switched systems is an appealing approach because
alternative in the control of power converters and drives. Success- it could allow higher dynamic performance [3]. Switched
ful application to different topologies and applications such as 2- systems have received increasing attention in the control
level voltage source inverters, neutral-point-clamped, cascaded H-
bridge inverters, and matrix converters have shown its potential community in past decades [4], [5], and some applications
in power converters. However, when multiple control objectives have been reported in power electronics [1], [2], [6].
are desired, weighting factors are required to appropriately select From the control techniques that consider power converters
the switching states. The selection of these factors is a time- as hybrid systems, Finite States Model Predictive Control (FS-
consuming and complex task. In this work, the standard selection MPC), also called Finite Control Set Model Predictive Control,
stage is replaced by a fuzzy decision making strategy, considering
as a case study the control of both load and supply currents in the has been rising as a promising control technique [7]. In this
Direct Matrix Converter. As a result, weighting factor selection control method the discrete system model is evaluated for
is avoided and a simple selection scheme is obtained for this every possible converter actuation and then compared with
application. In addition, a more natural design approach to the the signal reference in order to select the best voltage vector
state selection is opened for other applications. Simulation and [8]. Due to its basic concept, it is restricted to deterministic
experimental results are presented to validate the approach in
an experimental Direct Matrix Converter prototype. power conversion systems for which a model can be derived
through means of standard circuit laws. As a nonlinear control
Index Terms—AC-AC power conversion, Control systems,
strategy, it could consider linear and non-linear system models
Current Control, Digital control, Fuzzy logic, Predictive control.
and constraints. An important technological limitation is that
sampling frequencies are limited to be approximately less than
50kHz, given the computational power required to evaluate
I. I NTRODUCTION
the system states at every sampling time. On the other hand,

E NABLED by the actual computational power available,


new control techniques that previously were very difficult
to implement are increasingly being considered for power
it features fast dynamic response and flexibility [8], [9]. FS-
MPC’s potential has been shown by its application to different
topologies such as 2-level voltage source inverters, neutral-
converter and drives control. Controlling this kind of systems point-clamped, and cascaded H-bridge inverters, and matrix
is a complex task due to their inherent hybrid/switched na- converters [10]–[15]. Each application and converter topology
ture [1]. This characteristic appears because these systems has its own control objectives, but uses basically the same
are composed by a part with continuous states and another general control formulation [8].
with discrete ones, i.e. the switching matrix, implemented FS-MPC is based on cost functions that represent desired
through on-off switches. The desired behavior, reflected on the goals, e.g. current reference tracking, capacitor balancing,
continuous part, is obtained changing switching matrix states active and reactive power control, switching frequency, effi-
through a proper control strategy [2]. ciency, common-mode voltage. Then, when one control ob-
The conventional approach to control power converters is jective is desired, only one cost function must be minimized.
to discard the switched characteristic of the system through However, many applications have more than one control
modulation [1]. However, considering power converters as objective. To solve this, an aggregated objective function
Copyright (c) 2012 IEEE. Personal use of this material is permitted. (AOF) is constructed as a linear combination of individual cost
However, permission to use this material for any other purposes must be functions using so-called weighting factors. Every sampling
obtained from the IEEE by sending a request to pubs-permissions@ieee.org. time, a switching state is selected to be applied at the following
This work was supported in part by the Chilean National Fund of Scientific
and Technological Development (FONDECYT) under Grant 1100404 and one. The selected state minimizes the AOF two sampling times
1110794. Furthermore, the work of C. A. Rojas was supported by CONICYT ahead, in the set of all valid switching states [8], [16].
Scholarships for PhD studies in Chile. The main issue of using weighting factors is their correct
F. Villarroel, J. R. Espinoza, and D. Sbárbaro are with the Department of
Electrical Engineering, Universidad de Concepción, Concepción 160-C, Chile selection, because it is more complex than tuning PI parame-
(e-mail: fvillarroel@udec.cl; jose.espinoza@udec.cl; dsbarbar@udec.cl). ters or hysteresis bands on conventional controllers [17], [18].
C. A. Rojas, J. Rodriguez, and M. Rivera are with the Controller performance is directly related to weighting factor
Electronics Engineering Department, Universidad Técnica
Federico Santa Marı́a, Valparaı́so, 2390123 Chile (e-mail: selection [17]. Their main function is to try to model the
christian.rojas@usm.cl; jrp@usm.cl; marco.rivera@usm.cl). relative importance among desired control objectives [19]. In
2

contrast to the tuning of conventional controllers, FS-MPC A. FS-MPC scheme


currently lacks tools to systematically select its parameters. In FS-MPC, each desired control goal is represented by a
Although several methods to find appropriate weighting factors cost function gik+2 (Sk+1 ) in the sampling time k + 2 that is a
j
have been reported, they are based on empirical [12], [14], [19] function of the converter valid states Sk+1 in the time k + 1,
j
or offline procedures [17], and are time consuming and not with i ={1,...,n}, and j={1,...,m}, where n is the total number
systematic. In addition, they depend on the system operating of objectives, and m the number of valid converter switching
point and its parameters [18]. states. Each cost function is evaluated through a model of the
As tuning weighting factors is cumbersome and problem- system at every sampling time, and is considered as greater
atic, completely avoiding them could be an interesting option. than or equal to zero, so that the control goal is met when it
This drawback could be mitigated replacing the weighting fac- is minimized.
tor based stage with a multiobjective (MO) formulation [20]– The discrete-time prediction model allows the evaluation of
[22]. This basic concept gives the motivation for this paper. the state vector xk+1 at time k + 1 from its values at time k,
In particular, the authors propose using the well-known fuzzy through its dependence on the applied switching state Sk at
multicriteria decision making (FMCDM) [23]–[25] to avoid time k. The system model can be represented by the nonlinear
the weighting factors selection. This approach has already been state variables model f () and nonlinear output function h() as,
reported in Model Predictive Control (MPC) [26], [27], but
xk+1 = f xk , Sk , pk ,

not in the context of avoiding weighting factors in FS-MPC (1)
of power converters. The main contribution of this paper is yk = h xk , Sk , pk ,

(2)
to show the feasibility of such approach in FS-MPC using
standard digital platforms, which has as main challenge the where, pk is the perturbation vector at time k, and yk the
computational effort in a realtime implementation [28]. output vector at time k.
To assess the feasibility of the proposed method, the Direct
Matrix Converter (DMC) is considered as a case study, because B. State selection with multiple objective functions
this system requires at least two control objectives for proper In general, the ideal state to select at each sampling time is
operation [10]. As such, focus is placed on applications with the one that meets all the imposed criteria in the best possible
equally important control targets. The proposed method is way. This can be thought as the simultaneous minimization
compared with the conventional FS-MPC presented in [29], of all the proposed cost functions. Considering the delay
where a predictive current control strategy is applied to the compensation [30], [31], the MO optimization problem in the
DMC. Finally, a comparison will be performed through sim- FS-MPC scheme can be stated as,
ulation and tests on an experimental setup.
min Gk+2 Sk+1

This work is organized as follows: Section II introduces the j , (3)
Sk+1
j
∈S
basic concepts of state selection in FS-MPC and the need of
a decision making strategy. Section III introduces FMCDM in where Sk+1
j is the evaluated switching state to be applied at
FS-MPC. Section IV analyzes the application of the strategy in k + 1, and Gk+2 is a vector function of the switching states,
the DMC topology. Section V discusses experimental results, T
Gk+2 Sk+1
  k+2 k+1 k+2 k+1
and finally, conclusions are presented in Section VI. j = g1 (Sj ) g2 (Sj ) ... gnk+2 (Sk+1
j
) , (4)
with gik+2 , the i objective function in the k + 2 sampling time.
This simultaneous optimization can be interpreted as a vector
II. S TATE SELECTION IN THE FS-MPC SCHEME
optimum problem [24].
When multiple control goals are desired in FS-MPC, a In most cases, the simultaneous minimization of all objec-
strategy is needed to deal with the different nature of each tive functions is not possible. It is accepted that a proper
cost function. The simplest way is to convert the problem to a definition of vector optimality is given by the Pareto non
scalar optimization through an AOF. To this effect, weighting dominated solutions [32]. A solution Sp ∈ S is Pareto optimal
factors are required. The reasons for their presence are two: or a non dominated solution if,
try to model preference information in the state selector, and
∄Sq ∈ S : ∀i, gi (Sq ) ≤ gi (Sp ) ∧ ∃r : gr (Sq ) < gr (Sp ) (5)
render the cost functions comparable, i.e. with the same units
and scale. These factors should be appropriately selected for where, p, q ={1,...,m}, p 6= q; and S, the set of valid states.
correct operation [17]. For its design, some guidelines have In other words, a solution is better in at least one objective
been given in [14], [19], and promising methods have been than another, without being worse for all remaining ones. Note
developed in [17], [18]. that now we have, in general, a set of optimum solutions, from
The MO nature of the problem is made explicit in [17], which it is not possible a priori to determine which is better
however, weighting factors continue to be used in the scheme. than another. Selecting one solution from this set requires a
The difficulty of specifying appropriate weighting factors is certain kind of trade-off or preference among the objectives.
evident in the methods proposed in [17], [18]. Instead of trying The final selection is done through a decision making strategy.
to tune them, we propose a direct MO switching state selector To select a particular switching state there are two ways
as an alternative. In particular, FMCDM looks as an interesting to proceed. The first is obtaining the Pareto front, the set of
alternative in this application as will be shown in the following. all non dominated solutions, and then selecting one from this
3

µi (Sk+1
j
)
Measurement: xk
Step 1
replacements 1
Apply: Sk
opt
Step 2

Estimate: xk+1 (Sk


opt )
Step 3

k+2(min) k+2(max)
gi gi gik+2 f orj, i

Fig. 1. Linear membership function proposed for the control scheme. Predict: xk+2 (Sk+1
j
)

Evaluate: gik+2 (Sk+1


j
)
set through a decision strategy. The other way is to directly
j, i ≤ m, n
select a solution, without first determining the Pareto front. Step 4
Due to the computational complexity involved, the second
option seems more appropriate to realtime applications with Level detection: max, min gi
fast sampling times, such as a power converter. The algorithm
f orj, i
used in this work reflects this type of selection.
Membership: µi (Sk+1
j )
III. FS-MPC WITH FUZZY DECISION MAKING BASED
STATE SELECTION j, i ≤ m, n
Step 5
A decision making strategy that has been developed and
used with good results in various fields is FMCDM [23], f orj

[25], [33]. In particular, it has been applied extensively to the


solution of MO optimization problems, where multiple and Decision: µD (Sk+1
j
)
conflicting objectives must be met [24], [34], [35].
j ≤ m
FMCDM is characterized by the use of membership func-
tions. These functions represent the degree of attainment of
Select: Sk+1
opt
the goals for each solution. Then, a decision is obtained as Step 6
the intersection or confluence of them [23].
The application of FMCDM to MPC has been introduced Fig. 2. FDM-FS-MPC algorithm considering delay compensation.
in [26]. This application of fuzzy logic is a bit different
from the traditional approach used in control applications.
The conventional approach derives a control action from the the resulting values are compatible and could be directly
actual error and its change through a set of rules. In contrast, compared. The form of the membership functions could be
in FS-MPC we are concerned with the specification of the understood as the type of scale used for this mapping and
preferences of the state selector from a set of given goals. comparison. In particular, the linear mapping has the benefit
A previous application that is somewhat related to the of low computational requirements.
approach presented here is [36]. This work replaces the tradi- From the membership function definition it can be seen that
tional switching table of direct power control with a fuzzy rule- the mapping varies at each sampling time. In this way, the
based selection scheme. However, the main idea is completely scale used for each cost function is variable. Therefore, the
different to the one studied here. possible levels of achievement of each goal are considered.
These levels can vary significantly each sampling time, de-
A. Fuzzy decision making application to FS-MPC pending on the system characteristics and the cost functions
used. Then, it is possible to make a direct comparison among
To apply FMCDM in FS-MPC the form of the membership the achievable optimization levels for each objective.
functions and the type of confluence should be specified. A The final selection of the switching state is performed by
commonly used form of membership function is the linear the decision function. For its specification, the designer of
one, Fig. 1. For this application it will be defined as, the strategy sets the preference for each goal. Considering the
k+2(max) goals as equally important, two types of functions could be
− gik+2 Sk+1

k+1
 gi j
µi Sj = k+2(max) k+2(min)
, (6) used. These are the maximizing decision, and the Hamacher
gi − gi logical and. The maximizing decision is given by the maxi-
where, µi is the i goal membership function, gi
k+2(max)
is the mum of the intersection membership function defined by,
objective function gi maximum value at k + 2 instant, and µD Sk+1
 
k+2(min) j = min µ1 (Sk+1
j ), µ2 (Sk+1
j ), ..., µn (Sk+1
j ) . (7)
gi is the objective function gi minimum value at k + 2
instant. Each objective function has an associated membership Then, the best state to be applied at the next sampling time is
function at each sampling time. selected as the one with the maximum value of this function,
Each membership function constitutes a mapping from the Sk+1 k+1

opt = arg max µD Sj . (8)
space where the cost functions vary to the range [0, 1]. Then, Sk+1
j
∈S
4

TABLE I
On the other hand, the Hamacher logical and is defined by, S TATE SELECTION EXAMPLE
n
 Y
µD Sk+1 µi Sk+1

j = j . (9)
State g1 g2 u1 u2 uD1 uD2
i=1

Its difference from the maximizing decision point of view is 1 1.22 74.3 0.89 0 0 0
that it considers some degree of interaction among the objec- 2 8.24 10.4 0.14 1 0.14 0.14
3 9.57 40.3 0 0.53 0 0
tives. In contrast, the maximizing decision always maximizes 4 0.24 29 1 0.71 0.71 0.71
the fulfillment of the objective with the poorest achievement. 5 5.55 68.3 0.43 0.09 0.09 0.04
It could bring better performance for a higher number of goals. 6 7.7 71 0.20 0.05 0.05 0.01
More information about these operators can be found in [26], 7 7.1 33 0.26 0.65 0.26 0.17
[27].

B. Fuzzy decision making FS-MPC algorithm performance with respect to these optimized indices. Care
must be taken when selecting the cost functions to be used
The maximization of µD is done by exhaustive evaluation
for adequate performance.
for all feasible switching states, as in the standard FS-MPC ap-
The previous discussion about FMCDM was limited to
proach. Based on Fig. 2, the control strategy can be described
equally important objective functions, however it is not re-
in the following sequence:
stricted to these cases. Objective preference information could
Step 1 Measurement: Sample variables in order to get the
be modeled through the decision functions [26]. Using tools
actual state vector xk , perturbation vector pk , and output
from FMCDM a higher level design could be made. As a
vector yk , needed by the prediction model.
result, this could offer advantages when the parameters of
Step 2 Apply: Apply the optimal switches combination Skopt
a system vary and the model used is adjusted continuously
found in the previous loop iteration.
to track these deviations. An AOF based algorithm should
Step 3 Estimate: Use Skopt and the measured variables xk ,
require additional tuning for its factors, such as the continuous
p , and yk , to evaluate the model and predict xk+1 and yk+1 .
k
application of the methods proposed in [17], [18], whereas
Step 4 Evaluate: Predict xk+2 and yk+2 for every switch
the proposed scheme does not necessarily need to change its
combination. Calculate gik+2 for every objective and state.
decision law.
Step 5 Fuzzification: Using the gik+2 values, determine the
The computational aspect of the strategy requires a special
maximum and minimum levels for each gi at instant k + 2.
mention. Multiple objective control formulations specially
Calculate µi for every objective and state.
suffer from increased computational requirements [21]. It can
Step 6 Select: Obtain overall value µD for every state. Max-
be clearly seen that because the optimization and decision is
imize the decision function µD and select optimal switching
being carried out online, the proposed scheme requires more
state Sk+1
opt . computational effort. This implies the increase of the minimum
required sampling time. As the performance of FS-MPC is
C. Algorithm calculation example directly related to its sampling time, performance degradation
To illustrate the procedure, a calculation example is pre- could occur. However, the required computation time strongly
sented in Table I. The example considers seven switching depends on the implementation of the algorithm.
states and two objective functions. First, the values g1 and g2
are calculated for each state with the help of the system model.
Then, the maximum and minimum levels are determined for
(max) (min)
each objective obtaining g1 = 9.57, g1 = 0.24,
(max) (min)
and g2 = 74.3, g2 = 10.4. With these values, the
membership functions u1 , u2 can be calculated. After that, the
decision functions uD1 , uD2 are evaluated, where uD1 denotes
the maximizing decision function, and uD2 the product t-
norm, respectively. Finally, the state is chosen, in this example
it corresponds to state 4, for both decisions.

D. Comparison with the AOF approach


The proposed scheme as formulated in this work, only
considers the information of the cost functions at the present
sampling time. This is an important difference from the
standard FS-MPC approach. The AOF approach requires ad-
ditional performance indices in order to perform the tuning
of its weights, and in this way it could consider additional
information, such as in [17], [18]. From this, it is clear that
weighting factor-based methods can potentially yield better Fig. 3. Data flow chart of the FDM-FS-MPC algorithm.
5

E. Computational aspect of the proposed strategy is a challenge for a realtime application, and direct coupling
To show the dependence of the algorithm execution time between source and load variables.
on the implementation, the operation of the algorithm should
be studied. The algorithm operates as follows. First, each A. FS-MPC formulations for the DMC
cost function is evaluated for every switching state. Next, The DMC closed loop controller design is related to the
their maximum and minimum values are obtained. Then, desired control objectives, such as to achieve accurate control
membership functions are calculated for every switching state, of load currents and sinusoidal source currents with unitary
considering the previously determined levels. Finally, the deci- displacement power factor in the source side. These goals must
sion is performed. As the evaluation of the cost functions and be reached using the power converter finite set of states. FS-
membership functions does not depend on each other, some MPC implies knowledge of an accurate discrete-time system
stages are parallelizable. Thus, the calculation process could model to achieve desired objectives. From the control point
be divided in four sequential but internally parallel stages: cost of view, the strategy must achieve certain desired behavior
function calculation, level determination, membership function for the system outputs yk = [is k io k ]T , acting on the control
evaluation, and decision. The sequential characteristic of these inputs Sk = [S1k , ..., S9k ], as a function of the system references
stages can be seen from the data flow in Fig. 3, where a bi- y∗k = [is ∗k io ∗k ]T .
objective problem with 27 valid states is considered. 1) Discrete-time system model: We now turn to the deriva-
As previously noted, the internally parallel stages are not tion of a bilinear time-invariant discrete-time dynamic model
required to be evaluated in sequential form. If they are imple- for the DMC [43], [44]. Using elementary circuit analysis
mented in parallel, an increase of the algorithm throughput techniques in Fig. 4(a), it is possible to conclude that:
is possible. However, parallel implementations depend on
the choice of implementation technology. If the algorithm is xk+1 = Ad xk + Nd uk xk + Ed pk , (10)
implemented using a conventional DSP based system, these yk = Cd xk , (11)
operations need to be done sequentially [37], [38]. However, k k T
if the implementation considers the use of low cost FPGA where, xk = [vi k is io ] are the system state variables,
based systems, there is an opportunity to design specialized with vi k = [via k k
vib k T
vic ] the converter input voltage vector,
hardware to perform these tasks in parallel [28]. FPGAs have is k = [iksa iksb iksc ]T the source current vector, and io k =
already been used for power electronics control with excellent [ikoa ikob ikoc ]T the load current vector, all of them at time k.
results [38]–[40], and even FS-MPC algorithms have been The system outputs are yk = [is k io k ]T , while the measurable
implemented [41]. disturbance pk = vs k is given by the source voltage vector
If only a DSP is used to implement the algorithm, then vs k = [vsa k k
vsb k T
vsc ] . The control input matrix is
 k 
the first and second stages could be merged into one. The T 0 0
calculation time of this stage should not introduce significant uk =  Tk 0 0 , (12)
time with respect to the standard FS-MPC scheme. This is k T
0 0 (T )
because the number of cost function evaluations is essentially
the same as in standard FS-MPC. Two additional comparisons with, 0 ∈ R3x3 the null matrix and Tk the DMC direct
per cost function and switching state are added to determine commutation matrix given by,
 k
S1 S4k S7k

the maximum and minimum levels.
The greatest difference is in the third stage, where mem- Tk (Sk ) =  S2k S5k S8k  . (13)
bership functions are calculated, which does not exist in S3k S6k S9k
the standard scheme. Each cost function and state evaluation
requires a multiplication, which has greater computational cost
than additions and comparisons. In the opinion of the authors, DMC
this is the major factor for the increase of calculation time
vs Rf Ro
in the algorithm implementation on a DSP. Note that these is Lf S1 S2 S3 io Lo
vi
multiplications are independent and could be implemented in S4 S5 S6 n
vo
parallel if the proper computational resources are available, as ii S7 S8 S9
Cf
in a FPGA based system.
(a)
pk
IV. PARTICULAR CASE STUDY

This section addresses the study of the conventional FS- y∗k uk DMC Model yk
FS-MPC
MPC and Fuzzy Decision Making FS-MPC (FDM-FS-MPC)
Controller (10)-(11)
strategies applied to a Direct Matrix Converter (DMC) [10].
The DMC poses an intrinsic two objective control problem
because its input and output should be controlled simultane-
(b)
ously. An initial bi-objective MO formulation was done in
[42] for this application. The control of this topology is very Fig. 4. Three-phase to three-phase Direct Matrix Converter: (a) Power
interesting due to its high number of switching states, which Topology, (b) Closed loop FS-MPC block diagram.
6

(a) vsa /30, i sa (A)

(a) vsa /30, i sa (A)


5 vsa 5 vsa
isa isa
0 i∗sa 0 i∗sa

−5 −5
0 0.05 0.1 0.15 0.2 0.25 0 0.05 0.1 0.15 0.2 0.25
Time (s) Time (s)
5 5
io a io a
(b) io a (A)

(b) io a (A)
i ∗o a i ∗o a
0 0

−5 −5
0 0.05 0.1 0.15 0.2 0.25 0 0.05 0.1 0.15 0.2 0.25
Time (s) Time (s)
400 400
vo ab vo ab
(c) vo ab (V)

(c) vo ab (V)
200 200
0 0
−200 −200
−400 −400
0 0.05 0.1 0.15 0.2 0.25 0 0.05 0.1 0.15 0.2 0.25
Time (s) Time (s)

Fig. 5. Simulation results of conventional FS-MPC with load and source Fig. 7. Simulation results of FDM-FS-MPC with load and source current
current control (zero input weighting factor): (a) Source voltage vsa [V] and control: (a) Source voltage vsa [V] and current isa [A], (b) Load current ioa
current isa [A], (b) Load current ioa [A], (c) Load line-to-line voltage voab [A], (c) Load line-to-line voltage voab [V].
[V].

1.5
ks = 0
(a) µ̃ 1(S okpt)
(a) vsa /30, isa (A)

5 1 k so pt
vsa FDM
isa 0.5
0 i∗sa
0
0 0.02 0.04 0.06 0.08 0.1
−5
Time (s)
0 0.05 0.1 0.15 0.2 0.25 1.5
Time (s) ks = 0
(b) µ̃ 2(S okpt)

5 1 k so pt
io a FDM
(b) io a (A)

i ∗o a 0.5
0
0
0 0.02 0.04 0.06 0.08 0.1
−5
Time (s)
0 0.05 0.1 0.15 0.2 0.25
Time (s)
400 Fig. 8. Membership function values of the selected switching state (after
vo ab
(c) vo ab (V)

200 application of moving average filter with 20 samples): (a) Filtered membership
function µ1 associated to output current cost function g1 , (b) Filtered
0
membership function µ2 associated to input current cost function g2 .
−200
−400
0 0.05 0.1 0.15 0.2 0.25
Time (s) where, ∆io k+2 = io ∗k+2 − io k+2 corresponds to the output
current error, ∆is k+2 = is ∗k+2 − is k+2 to the source current
Fig. 6. Simulation results of conventional FS-MPC with load and source error, and ko , ks are the weighting factors. To find the nominal
current control (optimum weighting factor): (a) Source voltage vsa [V] and
current isa [A], (b) Load current ioa [A], (c) Load line-to-line voltage voab
weighting factor, the output current weight is fixed to ko = 1,
[V]. while the source current weight is calculated using the currents
nominal values [14], [19].
2
I nom

The commutation state vector Sk = [S1k , ..., S9k ] is the discrete ks = ko onom = 6.8678, (15)
Is
control input. State matrices {Ad , Nd , Ed , Cd } are given in
where, ko , Ionom , and Isnom are defined in Appendix - Table II.
the Appendix. The DMC predictive closed loop controller
This is a starting factor and can be offline optimized according
design is illustrated in Fig. 4(b).
to some performance metric through simulation. Optimizing
2) Aggregate cost function: Considering the system outputs its value for the nominal operating point with respect to input
as yk = [is k io k ]T , and using the prediction model (10) and and output current THDs, the resulting weighting factor is
(11) to obtain the variables for time k + 2, the AOF to be used
in the control scheme is [29], ks = ksopt = 23.708. (16)
3) Conventional FS-MPC simulation results: To verify the
Gk+2 = ko (∆io k+2 )T (∆io k+2 ) + ks (∆is k+2 )T (∆is k+2 ), (14) correct operation of the proposed scheme simulations are
7

vsa
vsa
isa isa

ioa
ioa

voab
voab

Fig. 9. Experimental results of Conventional FS-MPC with load and source Fig. 11. Experimental results of FDM-FS-MPC with load and source current
current control, steady state operation: (CH1) Source voltage vsa [V], (CH2) control, steady state operation: (CH1) Source voltage vsa [V], (CH2) Source
Source current isa [A], (CH3) Load current ioa [A], (CH4) Load line-to-line current isa [A], (CH3) Load current ioa [A], (CH4) Load line-to-line voltage
voltage voab [V]. voab [V].

vsa
vsa

isa
isa
ioa ioa

voab voab

Fig. 10. Experimental results of Conventional FS-MPC with load and source Fig. 12. Experimental results of FDM-FS-MPC with load and source current
current control, dynamic evolution: (CH1) Source voltage vsa [V], (CH2) control, dynamic evolution: (CH1) Source voltage vsa [V], (CH2) Source
Source current isa [A], (CH3) Load current ioa [A], (CH4) Load line-to-line current isa [A], (CH3) Load current ioa [A], (CH4) Load line-to-line voltage
voltage voab [V]. voab [V].

weighting factor is significantly different to the starting one.


carried out in the software PSim, using an external library
written in C language. Simulation and system parameters are
illustrated in Appendix - Table II, using a sampling time of B. FDM-FS-MPC control objectives for the DMC
40 [µs]. The case study considers an output current reference The control objectives defined for the FDM-FS-MPC ap-
change from 4[A] at fo =10[Hz] to 5[A] at fo =30[Hz]. proach are given by the tracking of the load and source
To illustrate the influence of the weighting factor on the cor- currents [29],
rect operation of the control scheme, a simulation is performed
setting ks = 0. As this disables the input current control, g1 = (∆io k+2 )T (∆io k+2 ), (17)
it can be seen From Fig. 5 that the input current is highly g2 = (∆is k+2 )T (∆is k+2 ). (18)
distorted and not in phase with the input voltage. On the other
hand, after setting an appropriate weighting factor, which in These cost functions are evaluated together with their asso-
fact is the optimum one from the point of view of input and ciated membership and decision functions for each possible
output current THDs, good performance is obtained, Fig. 6. commutation state of the DMC. Linear membership functions
From these results, it can be concluded that correct weighting are used for both load and source current control.
factor selection is crucial to good controller performance. Note 1) FDM-FS-MPC simulation results: The FDM-FS-MPC
that behind the final weighting factor selection there is a strategy is simulated using the same conditions of the conven-
cumbersome and time consuming iterative selection process tional FS-MPC seen before. Fig. 7 shows the operation of the
as seen in [29]. In fact, as (15) and (16) show, the optimum FDM-FS-MPC scheme under the same reference change as
8

performed in conventional FS-MPC. It can be seen that these and frequency fo =30[Hz], the same as used for the simulation
results are very similar to the ones obtained in Fig. 6, in which tests. The performance of the load and source current control
an offline optimized weighting factor was used. Note that no is very similar to the obtained through simulation with an
additional trial and error process is needed with the proposed optimized weighting factor. Similar results are obtained in
scheme. This is the main advantage of the strategy. Fig. 10, where a dynamic evolution is presented. The test
Basically, the idea behind the proposed scheme is to equally considered is the same as used for the simulation, an output
optimize the cost functions at every sampling time. This means current reference change from 4[A] at fo =10[Hz] to 5[A] at
that the degree of achievement of both controlled variables fo =30[Hz]. The resulting current THD for the 30[Hz] output
should be nearly the same. To verify this, Fig. 8 shows the reference is 15.77% for the input current, and 8.53% for the
membership function values associated to the selected opti- output current.
mum states after the application of a moving average filter with
a window of 20 samples. The filter is applied in order to make B. FDM-FS-MPC experimental results
it easier to see the evolution of the membership functions, Fig. 11 shows the FDM-FS-MPC steady state operation,
because the switching operation introduces a fair amount of with the same reference as the previous case. The results show
noise. Three cases are considered: zero input weighting factor that the control effort used for the load and source currents
(ks = 0), offline optimized weighting factor (ksopt ), and the is very similar, which can be seen from the ripple of the
proposed fuzzy scheme (FDM). Although the first two cases waveforms. Both currents show relatively low distortion for
are not fuzzy logic based, the membership functions of the this specific operating point. The resulting current THD for
selected switching state could be considered as a kind of fitness the 30[Hz] output reference is 14.38% for the input current,
functions. and 10.50% for the output current. The results are comparable
From Fig. 8 it is seen that when the input weighting factor with the ones from the FS-MPC strategy using an optimized
is zero, the amount of minimization that could be performed weighting factor.
regarding the output current is complete (the membership The dynamic performance of the FDM scheme is presented
function is unitary). However, the input current cost function in Fig. 12, using the same reference change as in standard FS-
is minimized to a much lesser amount, in fact the input fitness MPC. The scheme allows a successful output current reference
function mean value is approximately 0.5, instead of unitary. tracking. In addition, it successfully maintains control fairness
This shows that the optimization carried is not fair to both with different references between input and output currents,
goals. After the input weighting factor optimization, the fitness without the need of additional tuning.
functions mean values are nearly equal, which means that
both goals are being optimized to the same degree. Finally, VI. C ONCLUSION
as the fuzzy scheme tries to optimize both variables to the
A FMCDM-based switching state selector for FS-MPC has
same degree by design, the functions mean values show that
been proposed and tested in the predictive control of currents
is indeed the case. The results obtained by this simple heuristic
in the Direct Matrix Converter. FMCDM allows the design
are very similar to the optimized weighting factor case.
of the FS-MPC state selector from a higher level approach,
instead of tuning weighting factors as in the standard scheme.
V. E XPERIMENTAL R ESULTS The discussion has been limited to an algorithm which selects
A laboratory DMC prototype with bi-directional IGBTs switching states that optimize the required control objectives to
SK60GM123, with a control platform based on a DSP the same degree at each sampling time. Naturally, this imposes
TMS320C6713 and an Actel ProAsic A500 PQ208 FPGA, a fixed trade-off to the selection stage.
was used to validate the conventional and proposed methods. The simulation and experimental results shown indicate that
Only the DSP is actually used for algorithm implementation, the algorithm allows good performance in this application,
while the FPGA is in charge of firing signals distribution. compared to an optimized weighting factor scheme. In contrast
The conditions for steady state and dynamic evolution are to the conventional approach, the proposed strategy does not
considered equal to simulations to validate the previous results. require further tuning for correct operation. Its drawback is
The following results were obtained in a distorted and weak increased computational requirements. However, it has been
grid (as in most research laboratory outlets), some of its shown that the proposed algorithm can be implemented with
parameters are in Appendix - Table II. The sampling time current hardware technologies, and that it is particularly suited
of the experimental set-up is the same as the one used for to parallel implementations.
the simulations (40 [µs]). The calculation effort time of the
implementation is 14.36[µs] for standard FS-MPC, while for A PPENDIX A
FDM-FS-MPC is 31.32[µs], which shows that, at least for our S YSTEM PARAMETERS
DSP based implementation, the computation time required is The state matrices of (10) and (11) are calculated using
nearly doubled. an exact discretization of the input filter and a forward Euler
discretization of the load side [43],
A. Conventional FS-MPC experimental results 
C1 C2 0
 
0 0 C6

Fig. 9 shows the conventional FS-MPC steady state op- Ad =  C3 C4 0  , Nd =  0 0 C8  ,
eration, with an output current reference of amplitude 5[A] 0 0 C9 C10 0 0
9

TABLE II
S YSTEM PARAMETERS [9] K. V. Ling, W. K. Ho, Y. Feng, and B. Wu, “Integral-square-error
performance of multiplexed model predictive control,” IEEE Trans. Ind.
Inf., vol. 7, no. 2, pp. 196–203, May. 2011.
System Description Value [10] J. Rodriguez, M. Rivera, J. Kolar, and P. Wheeler, “A review of control
Vs : Voltage source rms line-to-line 130[V] and modulation methods for matrix converters,” IEEE Trans. on Ind.
Vs5 : 5th voltage source harmonic 5.1[%] Electron., vol. 59, no. 1, pp. 58 –70, jan. 2012.
Vs7 : 7th voltage source harmonic 2.2[%] [11] R. Vargas, P. Cortes, U. Ammann, J. Rodriguez, and J. Pontt, “Predictive
fvs : Source frequency 50[Hz] control of a three-phase neutral-point-clamped inverter,” IEEE Trans.
Cf : Filter capacitance (delta connection) 54[µF] Ind. Electron., vol. 54, no. 5, pp. 2697–2705, Oct. 2007.
Lf : Filter inductance 5[mH] [12] C. D. Townsend, T. J. Summers, and R. E. Betz, “Multigoal heuristic
Rf : Filter inductance parasitic resistor 0.5[Ω] model predictive control technique applied to a cascaded h-bridge
Rd : Passive damping resistor 50[Ω] statcom,” IEEE Trans. on Power Electronics, vol. 27, no. 3, pp. 1191–
Ro : Load resistance 10[Ω] 1200, Mar. 2012.
Lo : Load inductance 15[mH] [13] P. Lezana, R. Aguilera, and D. Quevedo, “Model predictive control of
Ionom : Peak load current 5[A] an asymmetric flying capacitor converter,” IEEE Trans. Ind. Electron.,
Isnom : Peak source current 1.90793[A] vol. 56, no. 6, pp. 1839–1846, Jun. 2009.
fo : Output frequency 10-30[Hz] [14] R. Vargas, U. Ammann, B. Hudoffsky, J. Rodriguez, and P. Wheeler,
ko : Output weighting factor 1.0000 “Predictive torque control of an induction machine fed by a matrix con-
ks : Source weighting factor 23.708 verter with reactive input power control,” IEEE Trans. Power Electron.,
Pn : Rated power 7.5[kW] vol. 25, no. 6, pp. 1426 –1438, Jun. 2010.
[15] M. Rivera, J. Rodriguez, J. Espinoza, and H. Abu-Rub, “Instantaneous
Reactive Power Minimization and Current Control for an Indirect Matrix
   Converter under a Distorted AC-Supply,” IEEE Trans. Ind. Inf., vol. 8,
  C5 0 0 I no. 3, 2012.
0 I 0
Cd = , Ed =  0 C7 0  I , [16] Y. Zhang, J. Zhu, and W. Xu, “Analysis of one step delay in direct torque
0 0 I control of permanent magnet synchronous motor and its remedies,” in
0 0 0 I
International Conference on Electrical Machines and Systems, Oct. 10-
where, I ∈ R3x3 is the identity matrix, Ci = ci I with ci , 13, 2010, pp. 792–797.
[17] P. Zanchetta, “Heuristic multi-objective optimization for cost function
i = 1, .., 8, the coefficients of the discrete-time input filter weights selection in finite states model predictive control,” in Proc. on
state space model given by, PRECEDE, Oct. 14-15, 2011, pp. 70–75.
    [18] S. Davari, D. Khaburi, and R. Kennel, “An improved FCS-MPC al-
c1 c2 Ac Ts c5 c6 gorithm for an induction motor with an imposed optimized weighting
=e , = Ac −1 (eAc Ts − I2x2 )Bc , factor,” IEEE Trans. Power Electron., vol. 27, no. 3, pp. 1540 –1551,
c3 c4 c7 c8
march 2012.
with Ts the sampling time and {Ac , Bc } the continuous-time [19] P. Cortes, S. Kouro, B. La Rocca, R. Vargas, J. Rodriguez, J. Leon,
S. Vazquez, and L. Franquelo, “Guidelines for weighting factors design
input filter state space model, in model predictive control of power converters and drives,” in IEEE
" 1
# " # Int. Conf. Ind. Technol. ICIT, Feb. 10-13, 2009, pp. 1–7.
0 Cf 0 − C1f [20] J. Maciejowski, Predictive Control with Constraints. Prentice Hall,
Ac = R , Bc = 1 .
− L1f − Lff Lf 0 2000.
[21] A. Gambier, “Mpc and pid control based on multi-objective optimiza-
For the load side the coefficients Ci = ci I with ci , i = 9, 10 tion,” in American Control Conference, Jun. 11-13, 2008, pp. 4727–
4732.
are given by [22] E. C. Kerrigan and J. M. Maciejowski, “Designing model predictive
controllers with prioritised constraints and objectives,” in IEEE Inter-
Ro Ts Ts
c9 = 1 − , c10 = . national Symposium on Computer Aided Control System Design, Sept.
Lo Lo 18-20, 2002, pp. 33–38.
[23] R. E. Bellman and L. A. Zadeh, “Decision-making in a fuzzy envi-
R EFERENCES ronment,” Application Series Management Science, vol. 17, no. 4, pp.
B141–B164, Dec. 1970.
[1] T. Geyer, G. Papafotiou, and M. Morari, “Model predictive control in [24] H. J. Zimmermann, “Fuzzy programming and linear programming with
power electronics: A hybrid systems approach,” in Proc. of the 44th several objective functions,” Fuzzy Sets and Systems, vol. 1, no. 1, pp.
IEEE Conf. on Decision and Control, Dec. 12-15, 2005, pp. 5606–5611. 45–55, Jan. 1978.
[2] K. Jezernik and R. Horvat, “Application of hybrid system theory to [25] C. Carlsson and R. Fullér, “Fuzzy multiple criteria decision making:
switching control of a three-phase inverter,” in IEEE International Recent developments,” Fuzzy Sets and Systems, vol. 78, no. 2, pp. 139–
Symposium on Industrial Electronics, June 27-30, 2011, pp. 1537–1542. 153, Mar. 1996.
[3] X. Lin-Shi, F. Morel, A. M. Llor, B. Allard, and J. M. Rétif, “Imple- [26] J. da Costa Sousa and U. Kaymak, “Model predictive control using fuzzy
mentation of hybrid control for motor drives,” IEEE Trans. on Industrial decision functions,” IEEE Trans. Systems, Man, and Cybernetics-Part B:
Electronics, vol. 54, no. 4, pp. 1946–1952, Aug. 2007. Cybernetics, vol. 31, no. 1, pp. 54–65, Feb. 2001.
[4] L. Zhang, C. Wang, and L. Chen, “Stability and stabilization of a class [27] J. Sousa and U. Kaymak, Fuzzy decision making in modeling and
of multimode linear discrete-time systems with polytopic uncertainties,” control. New Jersey: World Scientific, 2002.
IEEE Trans. on Ind. Electron., vol. 56, no. 9, pp. 3684–3692, Sept. 2009. [28] C. Buccella, C. Cecati, and H. Latafat, “Digital Control of Power
[5] L. Zhang and P. Shi, “Stability, l2 -gain and asynchronous h∞ control of Converters - A Survey,” IEEE Trans. Ind. Inf., vol. 8, no. 3, 2012.
discrete-time switched systems with average dwell time,” IEEE Trans. [29] M. Rivera, J. Rodrı́guez, P. Wheeler, C. Rojas, A. Wilson, and J. Es-
on Automatic Control, vol. 54, no. 9, pp. 2192–2199, Sept. 2009. pinoza, “Control of a matrix converter with imposed sinusoidal source
[6] M. Senesky, G. Eirea, and T. J. Koo, “Hybrid modeling and control currents,” IEEE Trans. on Ind. Electron., vol. 59, no. 4, pp. 1939–1949,
of power electronics,” in in Hybrid Systems: Computation and Control, Apr. 2012.
ser. Lecture Notes in Computer Science. Springer-Verlag, 2003, pp. [30] P. Cortes, J. Rodriguez, C. Silva, and A. Flores, “Delay Compensation
450–465. in Model Predictive Current Control of a Three-Phase Inverter,” IEEE
[7] J. H. Lee, “Model predictive control: Review of the three decades Trans. Ind. Electron., vol. 59, no. 2, pp. 1323–1325, 2012.
of development,” International Journal of Control, Automation, and [31] C. Xia, M. Wang, Z. Song, and T. Liu, “Robust model predictive
Systems, vol. 9, no. 3, pp. 415–424, 2011. current control of three-phase voltage source pwm rectifier with online
[8] S. Kouro, P. Cortes, R. Vargas, U. Ammann, and J. Rodriguez, “Model disturbance observation,” IEEE Trans. Ind. Inf., vol. in IEEE Early
predictive control - a simple and powerful method to control power Access Articles, 2012.
converters,” IEEE Trans. Ind. Electron., vol. 56, no. 6, pp. 1826–1838, [32] G. Liu, J. Yang, and J. Whidborne, Multiobjective Optimisation and
Jun. 2009. Control. Research Studies Press Ltd., 2003.
10

[33] M. Köppen and R. Vicente-Garcia, “A fuzzy scheme for the ranking of Christian A. Rojas (S’10-M’11) received the En-
multivariate data and its application,” in IEEE Annual Meeting of the gineer degree in electronic engineering from the
Fuzzy Information Processing Society, June 27-30, 2004, pp. 140–145. Universidad de Concepción, in Concepción, Chile,
[34] M. Sakawa and K. Kato, “An interactive fuzzy satisficing method for in 2009. He was awarded a scholarship from the
multiobjective structured linear programs and its application,” in IEEE Chilean Research Foundation CONICYT in 2010
Int. Conf. Systems, Man and Cybernetics ICSMC, vol. 5, Oct. 22-25, to pursue his PhD studies in power electronics at
1995, pp. 4045–4050. Universidad Técnica Federico Santa Marı́a, in Val-
[35] B. Reardon, “Fuzzy logic versus niched pareto multiobjective genetic paraiso, Chile.
algorithm optimization,” Modelling and Simulation in Materials Science His research interests include matrix converters,
and Engineering, vol. 6, no. 6, pp. 717–734, Jun. 1998. digital control, and model predictive control of
[36] A. Bouafia, F. Krim, and J. Gaubert, “Fuzzy-logic-based switching state power converters and drives.
selection for direct power control of three-phase pwm rectifier,” IEEE
Trans. on Ind. Electron., vol. 56, no. 6, pp. 1984–1992, June 2009.
[37] M. Kazmierkowski, M. Jasinski, and G. Wrona, “Dsp-based control of
grid-connected power converters operating under grid distortions,” IEEE
Trans. Ind. Inf., vol. 7, no. 2, pp. 204–211, May. 2011.
[38] T. Atalik, M. Deniz, E. Koc, C. Gercek, B. Gultekin, M. Ermis, and
I. Cadirci, “Multi-dsp and -fpga based fully-digital control system for
cascaded multilevel converters used in facts applications,” IEEE Trans. Jose Rodriguez (M’81-SM’94-F’10) received the
Ind. Inf., vol. in IEEE Early Access Articles, 2012. Engineer degree in electrical engineering from the
[39] E. Monmasson, L. Idkhajine, M. Cirstea, I. Bahri, A. Tisan, and Universidad Federico Santa Maria (UTFSM), Val-
M. Naouar, “Fpgas in industrial control applications,” IEEE Trans. Ind. paraiso, Chile, in 1977 and the Dr.-Ing. degree in
Inf., vol. 7, no. 2, pp. 224–243, may 2011. electrical engineering from the University of Er-
[40] T. Orlowska-Kowalska and M. Kaminski, “Fpga implementation of the langen, Erlangen, Germany, in 1985. He has been
multilayer neural network for the speed estimation of the two-mass drive with the Department of Electronics Engineering,
system,” IEEE Trans. Ind. Inf., vol. 7, no. 3, pp. 436 –445, Aug. 2011. University Federico Santa Maria, since 1977, where
[41] S. Thielemans, T. J. Vyncke, M. Jacxsens, and J. A. Melkebeek, “Fpga he is currently full Professor and Rector. He has
implementation of online finite-set model based predictive control for coauthored more than 350 journal and conference
power electronics,” in Proc. on PRECEDE, Oct. 14-15, 2011, pp. 63–69. papers. His main research interests include multi-
[42] F. Villarroel, J. Espinoza, C. Rojas, C. Molina, and E. Espinosa, level inverters, new converter topologies, control of power converters, and
“A multiobjective ranking based finite states model predictive control adjustable-speed drives. Dr. Rodriguez is member of the Chilean Academy of
scheme applied to a direct matrix converter,” in Proc. IEEE 36th Annual Engineering.
Conf. Ind. Electron. IECON, Nov. 7-10, 2010, pp. 2941–2946.
[43] C. Rojas, M. Rivera, J. Rodriguez, A. Wilson, J. Espinoza, F. Villarroel,
and P. Wheeler, “Predictive control of a direct matrix converter operating
under an unbalanced ac source,” in IEEE Int. Symp. on Ind. Electron.
ISIE, Jul. 4-7, 2010, pp. 3159–3164.
[44] B. Zhang, X. Wang, Y. Tang, and W. Zhang, “Predictive functional
control and stability analysis of mimo bilinear systems,” in American
Control Conference, Jun. 14-16, 2006, pp. 4854–4858.
Marco Rivera Marco Esteban Rivera Abarca (S09-
M11) received his B.Sc. in Electronics Engineering
and M.Sc. in Electrical Engineering from the Uni-
versidad de Concepción, Chile in 2007 and 2008,
respectively. He received the PhD degree at the
Felipe Villarroel received the B.Sc. and Engineer
Department of Electronics Engineering, Universi-
degree in Electronics Engineering (with first-class
dad Técnica Federico Santa Marı́a, in Valparaı́so,
honors) from the University of Concepción, Con-
Chile, in 2011 with a scholarship from the Chilean
cepción, Chile, in 2007 and 2009, respectively. He
Research Fund CONICYT. His research interests
is currently pursuing the M.Sc. degree in Electrical
include matrix converters, predictive and digital
Engineering at the same institution.
controls for high-power drives, four-leg converters
His research interests include the modeling, sim-
and development of high performance control platforms based on Field-
ulation and control of power converters, in partic-
Programmable Gate Arrays. Currently he is working on a Post Doctoral
ular emerging control techniques such as predictive
position and as part-time professor of Digital Signal Processors and Industrial
control, digital control and signal processing, and
Electronics at Universidad Técnica Federico Santa Marı́a.
embedded system development.

José R. Espinoza (S’92-M’97) received the Eng.


degree in electronic engineering and the M.Sc. de-
gree in electrical engineering from the University
of Concepción, Concepción, Chile, in 1989 and Daniel Sbárbaro (S’81-M’84-SM’01) was born in
1992, respectively, and the Ph.D. degree in electrical Concepción, Chile. He received the degree in elec-
engineering from Concordia University, Montreal, trical engineering from the Universidad de Con-
QC, Canada, in 1997. cepción, Concepción, Chile, in 1984, and the Ph.D.
Since 2006, he has been a Professor in the De- degree from Glasgow University, Glasgow, U.K., in
partment of Electrical Engineering, University of 1993. He was an Alexander von Humbolt Fellow at
Concepción, where he is engaged in teaching and the Control Engineering Laboratory, Ruhr Univer-
research in the areas of automatic control and power sity, Bochum, Germany, in 1998.
electronics. He has authored and coauthored more than 100 refereed journal He is currently a Professor with the Department of
and conference papers and contributed to one chapter in the Power Electronics Electrical Engineering, Universidad de Concepción.
Handbook published in 2011 by Academic Press. His current research interests include design of ob-
Prof. Espinoza is currently an Associate Editor of the IEEE TRANSAC- servers for nonlinear systems and the development and industrial application
TIONS ON INDUSTRIAL ELECTRONICS and Power Electronics. of nonlinear control algorithms.

Das könnte Ihnen auch gefallen