pinheiro_05472661

Upload: daltonvidor

Post on 03-Apr-2018

213 views

Category:

Documents


0 download

TRANSCRIPT

  • 7/28/2019 Pinheiro_05472661

    1/6

    Multiple Controllers for Boost Converters

    under Large Load Range:

    a GA and Fuzzy Logic Based ApproachFabrcio Hoff Dupont1, Vincius Foletto Montagner2, Jos Renes Pinheiro2, Humberto Pinheiro2,

    Srgio Vidal Garcia Oliveira1 and Adriano Pres11Universidade Regional de Blumenau (FURB)

    2Universidade Federal de Santa Maria (UFSM)

    [email protected]

    Abstract- This paper presents a new methodology to designcontrollers for boost converters operating in a wide range ofload variation. The proposed technique uses genetic algorithmto find local LQR controllers providing optimal performance interms of overshoot and ITSE criteria. Fuzzy logic strategy isthen used to combine these local controllers, yielding a strategythat guarantees good performance for a large set of loadconditions. The proposed approach is compared with a

    conventional technique based on a single controller and with astrategy of switching controllers. Simulation results illustratethe superior performance of the converter with the proposedcontroller.

    I. INTRODUCTIONBoost converters are used in large number of applications

    being of fundamental importance to industry as well as the

    subject of studies by the academia [1-4]. One difficulty to

    control these converters comes from the nonlinearity of the

    model. A way to overcome the nonlinearity is to linearize the

    model assuming small perturbations around some operating

    point. Controllers designed for this situation ensure stability

    and performance characteristics only locally. However, oftenone has to guarantee good performance for a wide load range,

    and in this case, controllers designed for a single operating

    point may present poor responses.

    The main goal of this paper is to propose a methodology to

    design a control system that guarantees good transient and

    steady responses for a boost converter operating in a large

    load range. The proposed methodology is based on the

    design of local linear quadratic regulators (LQR [5]) for

    several load conditions. For each load condition, genetic

    algorithm (GAs) [6-8] is used to select the LQR with better

    performance in terms of overshoot and ITSE criteria. These

    optimal local controllers are combined using fuzzy logic [9-12] to provide a global controller with good performance for

    the load range under consideration. Extensive simulations

    results demonstrate the good performance of the proposed

    controller which ensures a smooth transition between control

    actions thanks to the fuzzy logic combination of the multiple

    controllers

    In the sequence, Section II presents the model and the

    parameters of the boost converter used here. It also shows a

    conventional design of an LQR (derived for the nominal load

    condition) and the improvement of LQR performance with

    the aid of a GA. Section III presents the use of multiple

    controllers to obtain good system response under load

    variations. Finally, the results are compared and the

    performance superiority of the proposed strategy is

    demonstrated.

    II. BOOST CONVERTER AND LQRCONTROLLERThe design of a control system depends on a mathematical

    model that describes its dynamic behavior. Usually thesemodels are described in the form of differential equations and

    often exhibit nonlinear behavior. Assuming that the system

    will be perturbed by small signals around a given operating

    point, these models can be linearized, allowing to apply linear

    control techniques [13].

    A. State-Space Average ModelFig. 1 illustrates the circuit of the boost converter. The

    losses related to the parasitic resistance of the components are

    considered, except to the diode, which is considered as an

    ideal switch. When the boost converter operates in

    continuous conduction mode, CCM, there are two circuits,

    the first one represents the stage when the switch S is onwhile the second when it is off. Since that the capacitor

    voltage ripple and the inductor current ripple are usually kept

    small by a suitable converter design, it is reasonable for the

    controller design to consider the average value of variables in

    a switching period, which can be obtained by:

    1 t

    t Tss

    x t x t dtT

    (1)

    where x is the average value of a given variable x and sT is

    the PWM carrier period [1-2]. By Fig. 1, the description of

    the circuit for the period in which the switch S is on

    (0 t DTs is given by:

    0 1

    10 0

    0

    L DS

    L Li

    C C

    L C

    LLo

    CL C

    r rx xL

    VLx x

    C R r

    xRv

    xR r

    x x

    N1M1

    yxT1

    (2)

    where Lx and Cx are the states related to the inductor current

    and capacitor voltage, respectively. The rest of the symbols

    978-1-4244-5697-0/10/$25.00 2010 IEEE 105

  • 7/28/2019 Pinheiro_05472661

    2/6

    are represented in Fig. 1. For the period that the switch is off

    (DTs t< Ts), the description obtained is:

    2

    2

    1

    10

    L L C L C L

    L C L CL Li

    C CL

    L C L C

    LL C Lo

    CL C L C

    R r r r r R

    L R r L R rx xVL

    x xR

    C R r C R r

    xR r Rv

    xR r R r

    ux x

    N

    M2

    yxT

    (3)

    From (2) and (3), the state-space average model is given in (4):

    iV

    y

    x M x N

    T x

    (4)

    where the average matrices are defined by:

    1 2

    1 2

    1 2

    D D

    D D

    D D

    M M M

    N N N

    T T T

    (5)

    being D the nominal duty cycle and 1D D . Thus, the

    equilibrium point of the converter is determined from (6),where the sub index q highlights the operating point:

    1q iV

    X M N (6)

    For the model linearization, it is assumed that all the

    disturbances caused in the system are negligible, so that any

    product of the disturbance variables can also be neglected.

    Thus, the linearized model for small perturbations is obtained

    in (7):

    1 2 1 2

    1 2

    d

    y d

    BA F

    CE

    x M x M M x N N u N u

    T x T T x

    (7)

    where d is the duty cycle perturbation which represents the

    control signal to be applied, and ivu is the input voltage

    perturbation. In this analysis the input voltage is considered

    constant, so 0u . This equation describes the dynamic

    behavior of the circuit, representing the deviation of the

    variables with respect to the equilibrium point obtained in (6).

    B. Converter ParametersFor simulation and performance analysis in the following

    sections, a boost converter of 1.5 kW whose parameters are

    presented in Table I will be used.

    For this configuration, the converter will enter in the DCM

    only for power levels less than 2.5% of nominal value, case

    that is not of interest here. This paper will address the

    problem of controlling this converter with a load varying in alarge range of values.

    C. State space controlThere are several possible control techniques for systems,

    each one with its own complexity. In the state feedback, the

    control signal is a function of the state variables, which are

    supposed to be measured or estimated. Moreover, this type of

    strategy is very useful for optimal control techniques [13].

    Since the system described in (7) does not have integrators,

    the output voltage steady state error cannot be driven to zero

    with a simple state feedback control law. To ensure zero

    steady state error, an additional state ex is created, defined as:

    0 0

    t t

    ex t e d y d (8)which corresponds to the integral of error of the output

    voltage. Note that in this description the error of the output

    voltage is described as a disturbance in the output voltage

    with opposite sign ( e y ). Thus, the additional state can be

    written as:

    ex d C x E (9)

    Including (9) in (7), one has that the augmented system of

    state equations given by (10):

    0

    0 ee

    dxx

    y d

    A B

    x xA B

    C E

    C x E

    (10)

    where A and B denote the augmented matrices A e B ,

    respectively. For the circuit parameters shown in Table I, the

    system is controllable by means of the control law

    e

    dx

    xK

    ,

    (11)

    where K is the gain vector

    1 2 3K K KK , (12)

    to be determined [13].

    III. SINGLE CONTROLLERIn order to improve the system response, optimal control

    techniques are often used, where a given cost fucntion is

    minimized. The LQR (also known in the literature as 2H optimal control problem [5]) is a very popular optimal

    controller that can provide good stability margins and it will

    be used in this work.

    A. Conventional LQRIn case of system (10) be rewritten as

    d A B (13)

    +

    -

    +

    -

    i

    rL L

    S

    rDS

    rC

    C

    RL

    xC

    vo

    DB

    + -

    vL

    xL

    Fig. 1 Circuit of the boost converter with losses.

    TABLE I

    PARAMETERS OF THE BOOST CONVERTER USED.

    Parameter Value Parameter Value Parameter Value

    Vi 56 V L 602.11 H C 27 F

    Vo 200 V rL 5 m rC 50 m

    Fs 50 kHz rDS 10 m RL 26.666

    D 0.72

    106

  • 7/28/2019 Pinheiro_05472661

    3/6

    the LQR cost function to be minimized is given by

    0

    T TLQRJ d d dt

    Q R , (14)

    where the superscriptT denotes the matrix transpose and the

    weighting matrices Q and R are real, symmetric and

    positive defined, i.e. TQ Q , TR R , 0Q e 0R . This

    index is minimized when the equality (15) is obtained.1 T K R B P , (15)

    being the matrix P the solution of the Riccatis equation:1 0T T A P P A P B R B P Q (16)

    The computational solution of LQR is available in

    specialized programs such as MATLAB (lqr function), being

    a useful and easy to apply synthesis tool. However, the

    design of a good LQR is fundamentally based on the

    appropriate choice of Q and R matrices, that depends on the

    knowledge of the control system. In general, one does not

    have a specific method for the determination of them [8].

    The control of the proposed boost converter is

    accomplished from the block diagram illustrated in Fig. 2,

    where refV is the reference voltage and d t is the control

    signal and the control gains are obtained from the LQR As is

    conventionally done, the converter controller is designed for

    the nominal load condition, considered here as 100% of

    output power. In this situation, the operating point is defined

    in (17).

    100

    100

    26.503A

    197.9V

    Lq

    Cq

    X

    X

    (17)

    An acceptable response can be obtained using the

    following weighting values:

    100

    6

    3100

    1 0 0

    0 1 0

    0 0 1 10

    10 10R

    Q(18)

    which result in the control gain vector in (19):3 3

    100 46.7925 10 2.9557 10 10LQRK (19)

    Fig. 3 illustrates the simulation results for the controller

    with gains in (19). The reference voltage is set to 200 V and

    the converter starts up at maximum load until 20 ms, when

    the load is switched to 25% of nominal power, returning to

    100% in 35 ms. The observed overshoot in P1C is 41.8%, at

    20.77 ms. In P2C the undershoot is 29.55%, at 35.56 ms.

    B. Refinement of LQR Using Genetic AlgorithmsAs mentioned above, one of the main advantages of the

    state feedback control is the possibility of optimization

    through various methods. It was also highlighted that one of

    the difficulties in design of a LQR is the determination of the

    weighting matrices, usually obtained after many attempts.

    Even with the effort and knowledge of the designer, one

    cannot guarantee that the set of weights found is the best for

    the situation.

    The genetic algorithm (GA) is a technique of search and

    optimization based on the natural selection principle and may

    be the evolutionary computation technique most widelyknown and applied today [6, 8]. The algorithm starts with a

    random population of individuals (chromosomes), and

    through genetic processes similar to those occurring in nature,

    evolve under specified rules in order to minimize a cost

    function. Since the population is generated randomly, the GA

    is able to virtually search the entire solution space, and

    provide simultaneous searches at different points in this

    space. During the algorithm execution, the chromosomes that

    possess the best characteristics (lowest cost) generate

    offspring, improving the average cost value of the population

    as a whole.

    Due to the quantity of parameters and the stochastic nature

    of the process, the algorithm can converge to different resultseach run or be confined in a local minimum point. To avoid

    this problem, sufficiently large population values, rates of

    mutation and crossover can be adjusted to ensure a good

    population diversity. Moreover, elitism can be used to ensure

    the survival of the best chromosomes for the next generation.

    In this article, the GA objective is to determine Q and R

    matrices so that the LQR presents small overshoot in the

    event of a load disturbance. For this, each chromosome is

    composed of the genetic structure defined in Table II, which

    corresponds to elements of the weighting matrices given by

    11

    22

    33

    11

    0 0

    0 00 0

    GA

    GA

    q

    qq

    R r

    Q . (20)

    The cost function to be minimized by the genetic algorithm

    is defined by

    25 5te e

    GA tb b

    tJ ovr t e t dt ovr ITSE

    t , (21)

    where ovr is the percent overshoot, e the error with respect

    to the reference, bt and et are the times of beginning and end

    of the load disturbance, respectively.

    The solution process of the algorithm follows the steps:

    i. Randomly creates the initial population;ii. Determines the LQR gains for each chromosome;

    iii. Simulates the closed loop system for eachchromosome, reducing the load by 25% in 20 ms;

    iv. Calculates the cost function (21) for each simulationresult;

    v. Sort results based on their fitness values. If the stopcondition is met, skip to step vii;

    vi. Creates the next generation from changes in thechromosome of a single parent (mutation) or

    combination of genes from a pair of chromosomes

    (crossover). Return to step ii;

    Input current

    Output voltage

    K1

    K2

    K3

    ~d

    D

    XLq

    XCq

    Vref

    +

    +

    +++

    d(t)+

    +

    Fig. 2 Block diagram of the controller.

    107

  • 7/28/2019 Pinheiro_05472661

    4/6

    vii. Determines the best weighting matrices based on thebest chromosome;

    viii. Terminates the algorithm.In this article the population is kept fixed in 300

    chromosomes, limited to the range [1106; 1107], and the

    two best chromosomes of each generation have ensured its

    perpetuation (elitism). Of the remainder, excluding the

    elitism, each new generation is composed of 55% ofcrossover between pairs of chromosomes and the other 45%

    are generated by mutations.

    After about 150 generations the best chromosome found

    was the presented in Table III, leading to the gain vector

    defined in (23).3 3

    100 80.1262 10 9.4583 10 -29.433GAK (22)

    Fig. 3 presents the results using the gain vector (22) for the

    same conditions of the previous simulation. In P1G is

    observed an overshoot of 34.52% at 20.53 ms. In P2G the

    undershoot is 27.04% at 35.48 ms. As shown in Fig. 3, the

    LQR optimized by GA provided a superior performance

    when compared with the controller designed previously.IV. MULTIPLE CONTROLLERS

    As can be observed by the system model, the variation of

    load resistance directly changes the dynamics of the circuit

    and also the equilibrium point. Although shown above that

    the controllers were able to maintain a good performance

    even with a load variation of 100% to 25%, it can be expected

    that controllers designed to work in specific ranges of the

    power load will ensure a better performance, when properly

    selected or combined.

    The equilibrium points of the model, as well the load

    resistance and output current for different output power are

    given in Table IV.These four load situations will be used to design multiple

    controllers and to verify the control performance for the boost

    converter analyzed. Since the load resistance variation

    directly reflects the output current of the converter, this

    current will be used as decision variable for choosing the

    most appropriate controller for each situation.

    To design the LQR for each load condition, the augmented

    matrices in (10) are recalculated with the appropriate values

    of LR , and using the genetic algorithm exposed above, the

    control gain vectors from (24) were obtained:3 3

    100

    3 375

    3 350

    3 325

    80.126 10 9.458 10 29.433

    62.290 10 9.400 10 21.585

    85.903 10 20.263 10 20.509

    48.212 10 15.715 10 10.108

    GA

    GA

    GA

    GA

    K

    K

    K

    K

    (23)

    As will be seen in the following sections, besides a suitable

    design of the controllers, another crucial factor for the proper

    operation of a system with multiple controllers is the

    technique used to provide the selection of one of them or to

    do the combination of more than one of the controllers.

    A. Switching ControllersThe simplest technique for the application of multiple

    controllers is a direct switch from one controller to another.

    In this case, based on the decision variable, the signal from

    the most appropriate controller is selected to be used.

    For the case studied here, the output current is divided into

    four sections, one for each of the controllers obtained (23).

    The thresholds between each section are presented in Table V.

    The selection strategy of one of the controllers is carried

    out with comparators, as shown in Fig. 4. The signals 1Ctrl ,

    2Ctrl , 3Ctrl and 4Ctrl assume values 1 and 0 for active and

    inactive, respectively. The signals from each controller are:

    1d , 2d

    , 3d and 4d

    and the resulting control signal d is given

    by:1 1 2 2 3 3 4 4d Ctrl d Ctrl d Ctrl d Ctrl d

    (24)

    observe the signal 1 100GAd K , being obtained as the

    variables deviation in relation to the equilibrium point for the

    100% load condition. The signals 2d to 4d

    are obtained in a

    similar way.

    The results of this strategy are presented and compared to

    the other strategies investigated here in subsection C.

    B. Controllers Combined with Fuzzy LogicOne of the problems of switching between controllers is the

    discontinuity of the control signal caused by the sudden

    transition between different sets of gains. To minimize this

    problem and ensure a smooth control signal, a fuzzy logic

    20 25 30 35 40 45

    150

    200

    250

    300

    Outputvo

    ltage

    (V)

    Time (ms)

    LQR (conventional)

    LQR (GA)P1G

    P1C

    P2G P2C

    Fig. 3 Comparison between the conventional LQR and the LQR obtained byGA, both designed for 100% load situation.

    TABLE IIPARAMETERS OF GENES REPRESENTED BY CHROMOSOME.

    Gene 1 2 3 4

    Parameter 11q 22q 33q 11r

    TABLE III

    BEST CHROMOSOME FOUND FOR THE LQRFOR THE NOMINAL LOAD CONDITION.

    Gene 1 2 3 4

    Value 462.66103 155.85103 913.81103 1.054103

    TABLE IV

    EQUILIBRIUM POINTS OF THE MODEL FOR DIFFERENT OUTPUT POWER.

    Outputpower Lq

    X (A) CqX (V) LR () oI (A)

    100% 26.503 197.89 26.666 7.421

    75% 19.930 198.41 35.555 5.580

    50% 13.321 198.94 53.333 3.730

    25% 6.678 199.46 106.666 1.870

    108

  • 7/28/2019 Pinheiro_05472661

    5/6

    based supervisor is used.

    As shown in Fig. 5, the supervisor control uses two input

    variables. The first one, the output current ( oi ) is divided into

    four membership functions denominated L (low), ML

    (medium low), MH (medium high) and H (high), whichcorresponds to the four ranges of the output power used for

    each LQR 25%, 50%, 75% and 100% of the output power,

    respectively. The second one, the derivative of the output

    current ( oi ), is divided into five membership functions,

    denominated NB (negative big), NM (negative medium), Z

    (zero), PM (positive medium) and PB (positive big). The oi

    variable acts to anticipate the choice of most appropriate

    controller in the occurrence of significant load variations.

    Before being used by the inference process, the signals oi and

    oi are scaled, to adjust them to the variables universe of

    discourse. Such scaling gains are defined by:

    6

    0.1333

    16 10i

    i

    K

    K

    (25)

    being iK just a normalization factor for the maximum output

    current, and with iK is determined by trial and error.

    The inference process of Takagi-Sugeno has been used

    here. The defuzzification method is the weighted average of

    the rules activation, expressed in (27).

    1

    n

    i ii

    i

    w z

    Ctrlw

    (26)

    where represents one of the four outputs of the fuzzy

    supervisor, iz is a constant obtained from the rule base and

    iw is the value of the activation rule, given by:

    AND ,i o ow i i (27)

    The AND method uses the product of the membership

    values ( ) of the output current ( oi ) and its derivative ( oi ).

    One advantage of using a fuzzy supervisor is that the

    signals 1Ctrl , 2Ctrl , 3Ctrl and 4Ctrl now can assume any

    value between 0 and 1, and the control signal is now obtained

    by a combination of each LQR controller action in (24),

    resulting in d without sudden changes.

    The performance of this strategy compared to others is

    shown next.

    C. Simulation results and comparisonIn order to verify the results of each control strategy,

    simulations have been carried out for load variations

    according specifications given by (29):

    20 100%

    20 30 75%

    30 40 50%

    40 50 25%50 100%

    o

    o

    o

    o

    o

    t ms P

    ms t ms P

    ms t ms P

    ms t ms P t ms P

    (28)

    The simulation results are shown from Fig. 7 to Fig. 10, In

    these figures, LQRGA identifies the system output when a

    single controller (genetic algorithm based controller designed

    for 100% of power output) is applied, LQRGAsw represents the

    system output for switching between controllers, and

    LQRGAfuzzy is the system output for the fuzzy logic controllers

    combination.

    Despite of the simplicity of using a single controller, the

    limitation of this technique (LQRGA) becomes clear for large

    load disturbances. The multiple controller based techniques

    LQRGAsw and LQRGAfuzzy provide better performance than thesingle controller LQRGA and have similar performance when

    directly compared, as shown from Fig. 7 to Fig. 10. The

    superiority of the multiple controller based techniques is

    confirmed by Fig. 11, which presents a comparison between

    the ITSE performance criteria for each situation. In this

    figure, the ITSE is shown normalized with respect that of the

    single controller LQRGA.

    However, the strategy of switching between controllers

    present a serious problem when the decision variable takes

    values close to the thresholds. To illustrate this problem,

    simulations were performed considering load disturbances

    according specifications given by (30) and the results are shown

    in Fig. 12:20 100%

    20 30 87.5%

    30 40 62.5%

    40 50 37.5%

    50 100%

    o

    o

    o

    o

    o

    t ms P

    ms t ms P

    ms t ms P

    ms t ms P

    t ms P

    (29)

    It is clear that the combination of multiple controllers using

    fuzzy logic overcomes an important difficulty which affects

    the switching controller. In all cases, the fuzzy logic

    combination of multiple controllers demonstrates good

    TABLE VTHRESHOLDS FOR SWITCHING CONTROLLERS

    Thres

    ho

    ld

    Output power (%) Loadcurrent (A)

    87.5

    62.5

    37.5

    6.562

    4.687

    2.812

    Output

    current

    6.562

    4.687

    2.812

    Ctrl1

    Ctrl2

    Ctrl3

    Ctrl4

    +

    +

    +

    Fig. 4 Selection strategy used for switching the controllers.

    Ctrl1

    Ctrl2

    Ctrl3

    Ctrl4

    Fuzzy

    Superv

    isorio

    io

    Ki

    Ki

    1 0.8 0.5 0 0.5 0.8 10

    0.5

    1

    NB NM Z PM PB

    (io)

    io

    0 0.25 0.5 0.75 1

    0

    0.5

    1L ML MH H

    (i

    o)

    io

    Fig. 5 Block diagram and membership functions of fuzzy supervisor control.

    109

  • 7/28/2019 Pinheiro_05472661

    6/6

    behavior for the entire range of operation, with transients with

    fast accommodation, few oscillations (even in the case of

    operation near thresholds) and small overshoots and

    undershoots.

    V. CONCLUSIONSThis paper presented as a first contribution the use of a

    genetic algorithm to find LQR controllers (local controllers)

    suitable to optmise the performance of a boost converter in

    terms of overshoot and ITSE criteria for each load condition.

    Another contribution is the implementation of a fuzzy logic

    that combines the local controllers. This combination

    provides a global controller with good transient and steady

    state performances. A comparison of this controller with a

    conventional LQR controller and with a switched LQR

    controller showed the advantages of the proposed strategy,

    which can be extended to control other power converters.

    ACKNOWLEDGEMENTS

    This work is supported by PROCAD (Academic

    Cooperation Program) from CAPES (Brazilian Commissionfor Higher Education).

    REFERENCES

    [1] J. G. Kassakian, M. F. Schlecht, and G. C. Verghese, Principles ofpower electronics. Addison-Wesley, Reading, Mass., 1991.

    [2] R. W. Erickson and D. Maksimovic, Fundamentals of powerelectronics, 2nd ed. Kluwer Academic, Norwell, 2001.

    [3] B. K. Bose,Power Electronics and Motor Drives: advances and trends.Academic Press, Burlington, 2006.

    [4] G. Liping, J. Y. Hung, and R. M. Nelms, "Evaluation of DSP-BasedPID and Fuzzy Controllers for DC/DC Converters," in IEEETransactions on Industrial Electronics, vol. 56, pp. 2237-2248, 2009.

    [5] P. J. Antsaklis and A. N. Michel,A linear systems primer. Birkhuser,

    London, 2007.[6] R. L. Haupt and S. E. Haupt,Practical Genetic Algorithms, 2nd ed. JohnWiley and Sons, Hoboken, 2004.

    [7] C. Wongsathan and C. Sirima, "Application of GA to design LQRcontroller for an Inverted Pendulum System," inIEEE InternationalConference on Robotics and Biomimetics, 2008. ROBIO 2008.,pp. 951-954, 2009.

    [8] M. B. Poodeh, S. Eshtehardiha, A. Kiyoumarsi, and M. Ataei,"Optimizing LQR and pole placement to control buck converter bygenetic algorithm," inInternational Conference on Control, Automationand Systems, 2007. ICCAS '07,pp. 2195-2200, 2007.

    [9] K. Tanaka and H. O. Wang, Fuzzy Control Systems Design andAnalysis: a linear matrix inequality approach. John Wiley and Sons,New York, 2001.

    [10] P. Mattavelli, L. Rossetto, G. Spiazzi, and P. Tenti, "General-purposefuzzy controller for DC/DC converters," in Tenth Annual AppliedPower Electronics Conference and Exposition, 1995. APEC '95.

    Conference Proceedings 1995,vol. 2, pp. 723-730, 1995.[11] A. R. Ofoli and A. Rubaai, "Real-time implementation of a fuzzy logic

    controller for switch-mode power-stage DCDC converters," in IEEETransactions on Industry Applications, vol. 42, pp. 1367-1374, 2006.

    [12] V. F. Montagner, R. C. L. F. Oliveira, and P. L. D. Peres, "ConvergentLMI Relaxations for Quadratic Stabilizability and H Control ofTakagi-Sugeno Fuzzy Systems," in IEEE Transactions on FuzzySystems, vol. 17, pp. 863-873, 2009.

    [13] R. C. Dorf and R. H. Bishop, Modern control systems, 11th ed.Pearson/Prentice Hall, Upper Saddle River, NJ, 2008.

    20 21 22 23 24 25190

    195

    200

    205

    210

    215

    220

    225

    Time (ms)

    Outputvo

    ltage

    (V) LQRGA

    LQRGAswLQRGAfuzzy

    Fig. 6 Transient response to load disturbance from 100% to 75%.

    30 31 32 33 34 35195

    200

    205

    210

    215

    220

    225

    Time (ms)

    Outputvo

    ltage

    (V) LQRGA

    LQRGAswLQRGAfuzzy

    Fig. 7 Transient response to load disturbance from 75% to 50%.

    40 41 42 43 44 45190

    195

    200

    205

    210

    215

    220

    225

    Time (ms)

    Outputvo

    ltage

    (V) LQRGA

    LQRGAswLQRGAfuzzy

    Fig. 8 Transient response to load disturbance from 50% to 25%.

    50 51 52 53 54 55120

    140

    160

    180

    200

    220

    240

    Time (ms)

    Outputvo

    ltage

    (V)

    LQRGALQRGAsw

    LQRGAfuzzy

    Fig. 9 Transient response to load disturbance from 25% to 100%.

    100% t o 75% 75% t o 50% 50% to 2 5% 25% to 1 00%0

    0.2

    0.4

    0.6

    0.8

    1

    Noma

    lize

    dITSE

    0.3

    74

    0.4

    04

    0.1

    47

    0.1

    63

    0.0

    54

    0.0

    59

    0.7

    10

    0.6

    06

    LQRGALQRGAswLQRGAfuzzy

    Fig. 10 Normalized ITSE criteria for each load disturbance.

    20 25 30 35 40 45 50 55 60150

    160

    170

    180

    190

    200

    210

    220

    230

    Time (ms)

    Outputvo

    ltage

    (V)

    LQRGAswLQRGAfuzzy

    Fig. 11 Load perturbations to the threshold regions.

    110