Вы находитесь на странице: 1из 17

Applied Soft Computing 52 (2017) 418434

Contents lists available at ScienceDirect

Applied Soft Computing


journal homepage: www.elsevier.com/locate/asoc

A new controller for DC-DC converters based on particle swarm


optimization
J.B.L. Fermeiro, J.A.N. Pombo, M.R.A. Calado, S.J.P.S. Mariano
Instituto de Telecomunicaces and Department of Electromechanical Engineering, University of Beira Interior, Covilh, Portugal

a r t i c l e i n f o a b s t r a c t

Article history: We propose a new controller for DC-DC converters based on particle swarm optimization (PSO). This
Received 19 February 2016 new converter controller uses the PSO optimization method to directly control, by itself, the output
Received in revised form 14 October 2016 voltage of a boost DC-DC converter. In order to validate and qualify the proposed converter controller,
Accepted 15 October 2016
we analyzed and implemented some variants of the PSO algorithm, namely the standard PSO and the
Available online 21 October 2016
global local best PSO. The proposed converter controller was then compared with a variant of the classic
PI controller with anti-windup, for different operational conditions. The three controllers compared in
Keywords:
this work were implemented in the microcontroller TMS320F28027 by using the code composer studio
DC-DC converter
Voltage.mode control
from Texas Instruments. The results show that the proposed controller exhibits better behavior in terms
Particle swarm optimization controller of settling time and overshoot. Unlike most popular DC-DC converter controllers, the proposed controller
does not require any sort of optimal parameter determination.
2016 Elsevier B.V. All rights reserved.

1. Introduction algorithms used to determine the optimal parameters of the opti-


mal controller [18] and the PID controller [19], and which can also
The DC-DC converter is an electronic circuit that converts a be used to optimize fuzzy logic control in maximum power point
source of direct current (DC) from one voltage level to another, pro- tracking [20].
viding a regulated voltage output. Therefore, it is very important in Another optimization algorithm widely used to determine
most electronic systems, from consumer electronics to renewable optimal parameters is particle swarm optimization (PSO). This
energy generation systems. Control and modulation of power elec- bioinspired algorithm is an optimization technique that features
tronic systems has been one of the great challenges for researchers the self-adaptation of a population of particles according to a cost
and engineers working in the eld [1]. function to nd the best solution [21,22].
With varying complexity and popularity, several control strate- The PSO is used in many different situations. Using a model of DC
gies that provide a regulated voltage output can be found in the motor as the plant, and performance indices (e.g. integral squared
literature. In industry the most used strategy is the proportional error (ISE), integral absolute error (IAE), and integral time-weighted
integral and derivative (PID) controller [2,3] due to its simplicity. absolute error (ITAE)) as the objective/cost functions, it is possi-
It has three parameters that control each one of its parts. These ble to determine the optimal parameters of the PID controller [23].
parameters greatly inuence the behavior of the PID controller, Using an adaptive PSO [24], where the velocities and positions are
therefore its optimal values must be determined. Other strategies given by a Gaussian or Lvy distribution, the PID optimal param-
have particular relevance in the eld, for example the optimal con- eters with a ball and hoop model as the transfer function and the
trollers [47], sliding mode controllers [811], fuzzy controllers performance indices ISE, mean squared error (MSE) and IAE as the
[1214] and articial neural networks (ANN) controllers [1517]. objective functions were determined. An enhanced performance
Along with the PID controller, some of these controllers also require criterion is proposed in [25] based on the integral squared time-
tuning of their fundamental parameters using optimization algo- weighted error (ISTE), as the objective function of the parameter
rithms. Among this type of optimization, we highlight genetic optimization process (with PSO), in order to reduce the overshoot
of a robust PID controller. The PSO is used in PID controller tuning
for buck converters [26], and for automatic voltage regulators [27].
In all the aforementioned cases the authors concluded that tun-
Corresponding author. ing with this optimization method increases the performance of
E-mail addresses: fermeiro@ubi.pt (J.B.L. Fermeiro), d1232@ubi.pt the PID controller compared with other methods. This optimiza-
(J.A.N. Pombo), rc@ubi.pt (M.R.A. Calado), sm@ubi.pt (S.J.P.S. Mariano).

http://dx.doi.org/10.1016/j.asoc.2016.10.025
1568-4946/ 2016 Elsevier B.V. All rights reserved.
J.B.L. Fermeiro et al. / Applied Soft Computing 52 (2017) 418434 419

Nomenclature

Vi (t) Vector of the particle i velocity at iteration t


R1 and R2 Vector with random numbers
c1 , c2 and C Acceleration parameters
Pbesti Vector of the personal best position of particle i
Gbest Vector of the global best position
Xi (t) Vector of the position of particle i at iteration t
w Inertial weight
gbest Cost-function value of the global best position
pbestAV Cost-function average value of all the particles per- Fig. 1. Graphical representation of the particle evolution.
sonal best position
pbesti Cost-function value of the particle i personal best
position Lastly both implemented PSO controller algorithms (standard
RC Internal resistance of the capacitor and global local best) were compared with an anti-windup PI
RL Internal resistance of the inductor controller in different operation conditions.
d Converter modulation index (duty-cycle)
R Load resistance This paper is organized in the following manner: Section 2 pro-
Vout Output voltage vides a brief review of the particle swarm optimization algorithm,
Vin Input voltage its topologies and some relevant variants present in the litera-
Vref Reference voltage ture, and a description of the implementation of the proposed
Duty p1 Value of the particle 1 (duty-cycle) controller algorithm; Section 3 explores the development of the
Converter duty Current converter duty-cycle DC-DC converter used to perform the tests and to validate and
VAV p1 Particle 1 average voltage qualify the proposed controller; Section 4 presents and discusses
VAV Average of all the particles average voltage the test results of the proposed controller; Section 5 presents the
v p1 Particle 1 velocity conclusions of this work.
KP Proportional parameter
KI Integral parameter
2. Background and proposed approach

2.1. Particle swarm optimization


tion method is also used to determine the parameters of other
types of controller, namely fuzzy controllers [28,29] and sliding
The particle swarm optimization algorithm is inspired by coop-
mode controllers [30], or used in a different process, like maximum
eration and social behavior principles. This algorithm involves a
powerpoint tracking (MPPT) [3134].
population of particles, where each particle represents a possible
The performance, adaptability and simplicity of the control algo-
solution. Each particle has a related velocity, which is adjusted with
rithms used in converter controllers in the literature [35] vary to
an update equation that considers the history of individual and col-
different levels, e.g. the PID control algorithm has high simplicity
lective experiences [42,43]. The main idea is to change the position
but low adaptability, whereas the sliding mode algorithm has low
of each particle with certain velocities in order to nd the best solu-
simplicity and high performance. Additionally, the controller tun-
tion. At rst the particles are randomly or evenly distributed within
ing is a lengthy process as the method of trial and error, even with
the search space. Then, they are evaluated with a cost function
the help of optimization algorithms, requires time to determine the
and all the other particles update their positions in the direction
optimal parameters of the controller for specic system character-
of the particle with the best solution with a determined velocity.
istics. And when there are system changes, the controller requires
This velocity is affected by three main factors: the velocity of the
repeating this tuning process to ensure its response. The afore-
previous iteration, the particle best solution (pbest), and the global
mentioned reasons motivated us to search for a control algorithm
best solution (gbest) [22], Fig. 1.
that would require no or less tuning, that would adapt to system
An important aspect for the algorithms performance is the
changes, and also a controller with a better overall performance in
topology of the swarm, i.e. the way the particles communicate with
terms of settling time and overshoot.
each other. Among the many topologies in the literature, the most
In this work, we explore PSO characteristics, including imple-
common are:
mentation simplicity, robustness, self-adaptation and versatility
in order to apply this method in a way that will allow the PSO
based algorithm to directly control the output of a DC-DC con- Star topologyevery particle communicates with one another,
verter, rather than using it to determine the optimal parameters Fig. 2(A);
of a certain controller, i.e. as a controller itself. The algorithm of Ring topologyeach particle only communicates with a certain k
the proposed controller calculates the converter modulation index number of adjacent neighbors, Fig. 2(B);
(duty-cycle) required to match the output voltage to the reference Cluster topologyinside each cluster the particles communicate
voltage through the reduction of the error. In order to qualify and with one another and certain particles communicate between
validate the proposed controller: clusters, Fig. 2(C);
Von Neumann topologyparticles are connected such that parti-
A DC-DC boost converter was developed to perform the tests in cles in one extremity communicate with particles in the opposite
real conditions; extremity, Fig. 2(D).
Some PSO variants present in the literature [3640], addressed
in Section 2, were analysed and one was chosen to be imple- In 1995 Eberhart and Kennedy developed the original frame-
mented alongside the standard algorithm, called global local best work of particle swarm optimization, known as Standard PSO [21].
PSO [41]; There are two fundamental equations in the standard PSO. The rst
420 J.B.L. Fermeiro et al. / Applied Soft Computing 52 (2017) 418434

premature convergence before nding the true local minimum


(guaranteed convergence PSO) [4648]. The main idea is to intro-
duce an additional particle, which searches the region around the
current global best position. In that manner, the current global best
particle is treated also as a member of the swarm (e.g. particle) [47].
Lastly, in the variant global local best particle swarm optimization
(GLBest PSO) the inertia weight and the acceleration parameter are
calculated according to the equations
 gbest

W= 1.1
pbestAV
 gbest
 . (3)
C= 1+
pbesti

In this method the position update equation is the same as in


the standard variant, but the velocity equation is as follows

V i (t + 1) = WV i (t) + CR 1 (Pbest i + Gbest 2X i (t)) , (4)

where W represents the inertia weight; C represents the acceler-


ation parameter; R1 is a vector with random numbers uniformly
distributed between 0 and 1; gbest is the cost-function value of
the global best particle at the moment; pbesti is the cost-function
Fig. 2. Graphical representation of the most common PSO topologies, A) Star topol-
value of the personal best position for particle i; pbestAV is the cost-
ogy, B) Ring topology, C) Cluster topology, D) Von Newmann topology.
function average value of the personal best particle at iteration
t.
one is the velocity equation
V i (t + 1) = V i (t) + c1 R 1 (Pbest i X i (t)) + c2 R 2
2.2. Proposed approach
(Gbest X i (t)) , (1)
In the proposed approach the particles represent the converter
where variable t is the iteration number; variable i is the particle modulation index (duty-cycle) that varies within the range [0,1].
number; the vector Pbesti represents the best solution of particle Due to the intrinsic characteristics of the boost converter it is nec-
i at that point; the vector Gbest represents the best solution found essary to ensure the positioning of the particles within a certain
from all the particles at that moment; the positive constants c1 and range. In order to know this range we need a mathematical model
c2 represent the acceleration parameters; the vectors R1 and R2 for this type of converter (boost) [49], with which it is possible to
are random numbers uniformly distributed between 0 and 1; and relate the duty-cycle as a function of the quotient between Output
the symbol denotes element-by-element vector multiplication. Voltage and Input voltage (i.e. Gain). Considering ideal mosfet and
To ensure convergence the condition c1 + c2 4 must be respected. diode, the boost converter gain equation can be written as
The second is the position update equation
Vout R (R + RC ) (d 1)
X i (t + 1) = X i (t) + V i (t + 1) . (2) Gain = = .
Vin R (RC + RL ) + rC RL + R2 (1 2d) + dR (dR RC )
Many variants of this standard algorithm have been developed (5)
in order to improve performance, including for example the modi-
ed particle swarm optimization [37], which introduces an inertia
weight (W). This inertia weight improves the control over the par- Knowing that the internal resistance of the inductor is the
ticle movement, forcing the particle to travel more or less in the parameter that most inuences the boost converter gain and con-
search space. Other variants force the particle movement by deter- sidering an ideal capacitor (RC = 0) in the converter gain Eq. (5), we
mining an inertia weight given the number of iterations in the PSO can observe in Fig. 3 the variation of the gain value as a function
variant called linearly decreasing weight PSO (LDW-PSO) [36], or by of the duty-cycle (d), for different quotients between R and RL . This
determining the acceleration parameters C1 and C2 given the iter- allows the determination of the range for the duty-cycle (from 0 to
ation count in the PSO variant called particle swarm optimization the point of maximum gain) that ensures the stability of the system.
with time varying acceleration coefcients (PSO-TVAC) [38]. In the The controller should not work close to the point of maximum gain
case of the PSO variant called supervisor student model in particle because from this point on the increase of the duty-cycle reduces
swarm optimization (SSM-PSO) [39], the particles are prevented the output voltage, thus making the controller unstable and in the
from ying out of a search space, without checking the validity worst case scenario causes system failure. Therefore, we imple-
of the position at each iteration, by introducing a variable called mented the connement method Random Back, based on [50]. If
momentum factor. In the case of the variant parameter free PSO a minimum or maximum is reached in that iteration, the velocity
(pf-PSO) [40], the PSO parameters are removed and the local best is multiplied by a pseudorandom number and changes direction,
(Pbesti ) and global best (Gbest) positions are used to update the moving the particle in the opposite direction of the borders. In
position of the particles directly in the position update equation. closed loop the system transfer function only depends on the plant
Concerning the convergence, the use of the constriction coefcient model (DC-DC boost converter). This means that the connement
can be viewed as a recommendation to make the particles travel method guarantees system stability, conning the minimum and
in smaller steps and eventually converge. The convergence men- maximum values of the duty-cycle.
tioned here obviously implies that the particles velocity reaches With the Eq. (5) we are able to calculate a very approximate
zero. But the correspondent position is not necessarily the wanted value of duty-cycle required to achieve a certain gain. Since the
convergence point [44,45]. There are also PSO variants that avoid algorithm controls directly the duty-cycle, and to reduce the com-
J.B.L. Fermeiro et al. / Applied Soft Computing 52 (2017) 418434 421

Fig. 3. Boost Converter gain curves as a function of the duty-cycle, for different values of R/RL in the system with component characteristics, C = 200e6 F, L = 1.6e3 H.

Fig. 4. Flowchart of the particle swarm algorithm implemented in the microcontroller.


422 J.B.L. Fermeiro et al. / Applied Soft Computing 52 (2017) 418434

Fig. 5. CPU use of the implemented algorithm.

Fig. 6. Temporal diagram of the analogic to digital converter AD7367.

putation time, we consider all components ideal and the gain rupt routines. The CVNST interrupt routine triggers the external
equation can be written as analog-to-digital converter (ADC) for data conversion initializa-
tion by creating a low pulse (0 V) of 0.2 ms on the microcontroller
1
Gain = . (6) pin. The Read interrupt routine receives the binary data from the
(1 d)
external ADC (through SPI) and converts it into decimal data. The
This equation will allow us to position the particles selectively PSO algorithm interrupt routine executes the proposed algorithm
in a small range increasing the speed of the optimization process, online, positioning the particles, evaluating their performance with
i.e. position the particles more precisely and accurately. The parti- the cost-function and calculating the velocities and the new posi-
cles are positioned in the following manner: the central particle is tions of the particles. These interrupt routines are explained in more
positioned according the equation detail below.
  Fig. 5 illustrates the use of the CPU, where the interrupts are
Vout 1 Vref Vin triggered by the ePWM module, congured to have a frequency of
= d= , (7)
Vin (1 d) Vref 20 kHz, matching the commutation frequency of the DC-DC con-
verter. The Read and CVNST interrupt routines are programmed to
and the other particles are positioned within a congurable range
be triggered at every 2 periods of the PWM, meaning that they are
of the central particle.
triggered with a 10 kHz frequency. Lastly, the PSO algorithm rou-
With this positioning method we guarantee the convergence
tine is executed with a congurable frequency, in this case 100 Hz
at the expense of even faster convergence, because the particles
due to the used voltage average calculation method, as explained
arent positioned randomly or too far from each other. This also
below.
lowers the overshoot (the maximum peak value of the response
The CVNST interrupt routine is responsible for the analogic to
curve measured from the desired system voltage value at steady-
digital converter AD7367 trigger. Through the microcontroller gen-
state) and provides a more accurate output voltage.
eral purpose input/output number 6 (GPIO6), congured as digital
The algorithm was implemented in the microcontroller
output, it controls the ADC input Conversion Start as shown in Fig. 6.
TMS320F28027 with the Texas Instruments compiling program
The Read interrupt is a hardware interrupt routine associated
Code Composer version 5, in C language. The owchart of the con-
with the falling edge of the microcontroller pin GPIO12. This pin is
troller algorithm based on standard particle swarm optimization
congured as input and is connected to the Busy output of the ADC,
(proposed approach) is shown in Fig. 4. The algorithm starts by ini-
which indicates the end of the conversion and that data is available
tializing the peripheral modules: serial peripheral interface (SPI)
at the serial port, Fig. 6. In this routine the instant voltage value is
and enhanced pulse width modulator (ePWM). Then the interrupt
obtained via serial communication (SPI), and the average voltage
routines are dened and the system variables initialized. After this
value is calculated for use in the PSO interrupt routine. Since the
the algorithm remains in an innite loop waiting for the inter-
J.B.L. Fermeiro et al. / Applied Soft Computing 52 (2017) 418434 423

Fig. 7. Block representation of the proposed boost converter topology.

frequency of the Read interrupt routine is 10 kHz and the cong- linear congruential method [51,52]. Second, we calculate the par-
ured frequency for the PSO algorithm routine is 100 Hz the average ticle velocities and update their positions with the Eq. (2). We used
voltage value is calculated discarding the rst 50 voltage values, in c1 = 1 and c2 = 1 in the velocity Eq. (1) because it respects the conver-
order to reject possible reading errors from the recent transition, gence condition used in the standard PSO algorithm [21], and also
and using the next 50 voltage values for each particle. in order to show the scenario that requires no parameter tuning.
The PSO algorithm interrupt routine contains the proposed PSO With each iteration the output voltage will approximate the refer-
algorithm in the star topology, with a congurable population ence voltage value and eventually the particle velocities will reduce
(number of particles). These particles represent the DC-DC con- to zero, meaning that all particles will converge to the best value
verter modulation index (duty-cycle), meaning that our problem found by the algorithm. Having reached the convergence point,
is one-dimensional. A bigger population results in a more accurate particles will not move, thence the strong stability of the output
controller response, however this leads to longer computational voltage of the proposed controller. However, if the best achieved
time, so a tradeoff between the accuracy and the speed of the con- value does not produce an output approximate to the reference
troller response should be taken into account. This routine starts by voltage the process needs to be reinitialized in order to converge
position the particles as explained previously. After that the parti- to the wanted value. For the reinitialization process to begin there
cles need to be evaluated with the cost function (the error between are two conditions that need to be satised: the particle velocities
the average voltage of each particle with the reference voltage) in must be lower than 0.01, meaning that the particles need to be close
order to achieve the required output voltage. to the convergence point; and the difference between the output
After the particle evaluation, the next step is to calculate the voltage and the reference voltage must be higher than 0.4 V. If both
particle velocities with Eq. (1). In order to do so the pseudo- the previous conditions are satised, the particles are repositioned
random parameters R1 and R2 need to be determined. The most and the process reinitializes.
common technique for generating pseudorandom numbers is the
424 J.B.L. Fermeiro et al. / Applied Soft Computing 52 (2017) 418434

s
The pseudo-routine of the proposed algorithm is as follows:

In order to develop and test the proposed controller, a non-isolated DC-DC boost converter was designed and
in the next section.
J.B.L. Fermeiro et al. / Applied Soft Computing 52 (2017) 418434 425

Fig. 8. Acquisition system representation.

Fig. 9. Command system of the power circuit developed.

Fig. 10. Converter DC-DC boost circuit. C1 = 500 F, C2 = C3 = 3.7 F, C4 = 1000 F, L1 = 10 H, L2 = 6 mH, S is the mosfet 20N60C3 from Inneon and D is the diode C3D06060
from Cree.

3. Hardware bility of parallel processing. It has a frequency clock of 60 MHz, with


12 kb RAM memory, 2 kb ROM memory and 64 kb ash memory. It
The hardware developed for this work can be represented, Fig. 7, features 6 PWM channels and 812 bits ADC channels with a min-
in four different system blocks: imum conversion time of 333 ns. It also features different types of
communications like I2C, CAN and SPI.
Processing and control system
Acquisition system; 3.2. Acquisition system
Command system of the power circuit (power drive)
Power system; The acquisition system is based on the analogic digital converter
AD7367, it is a 14 bits 4 channel with simultaneous conversion. It
3.1. Processing and control system can be congured via the microcontroller to work with different
conversion amplitudes like +/5 V, +/10 V, +/12 V and 010 V,
The processing and control unit used in this work was the micro- granting good precision and exibility with a conversion time lower
controller TMS320F28027. It is a xed point 32 bits microcontroller, than 1.25 s and a transmission rate of 500 ksps. The microcon-
used in mathematical tasks with certain complexity, due to its capa- troller ADCs have low resolution and exibility so we opted for this
426 J.B.L. Fermeiro et al. / Applied Soft Computing 52 (2017) 418434

Fig. 11. Photograph of the DC-DC boost converter developed and used in this work in order to test the proposed controller.

Fig. 12. Anti-windup PI controller topology used to compare with the proposed approach controller.

Fig. 13. Transient response of the three controllers in terms of output voltage, to a reference voltage from 0 to 60 V, with an input voltage of 30 V where the x-axis represents
time in millissecond and the y-axis represents voltage in volt. A) Standard PSO algorithm (50 ms and 1 V per division), B) GLBest PSO algorithm (50 ms and 1 V per division)
and C) Anti-windup PI controller (50 ms and 2 V per division).
J.B.L. Fermeiro et al. / Applied Soft Computing 52 (2017) 418434 427

Fig. 14. Response in terms of output voltage of the reference voltage variation test, changing between 40 V and 80 V, with an input voltage of 30 V.

external ADC. In this work it was pre-programmed to work with An isolated DC-DC (TEN3-0523) was used to power the opto-
conversion amplitude from 0 to 10 V. The communication with the isolator, Fig. 9, with the purpose not only of isolating the logic
microcontroller is serial communication SPI. and power circuits, but also reducing the noise of the switching
The voltage sensor used was the sensor CYHVS25A, Fig. 8, with converter onto the processing and control system circuit.
a transformation relation of 2500:1000, a nominal current on the
primary of 10 mA and on the secondary of +/25 mA, exhibiting
3.4. Power system
precision and linearity errors lower than 1%. It was assembled for a
maximum input amplitude range of 200 V. The input resistor limits
The converter created in this work was a non-isolated boost
the current so that the nominal current over the primary does not
converter. This is a step-up converter meaning that it generates
exceed 10 mA. On the other hand, the measure resistor needs to
an output voltage level higher or equal to the input voltage. The
be sized so that the voltage levels are adequate for the analogic to
topology used is shown in Fig. 10.
digital converter. In this case with a 350 ohm resistor, the voltage
Fig. 11 depicts the experimental assembly of the different sys-
varies from 0 to 8.75 V.
tems of the converter mentioned above. The upper plate holds the
The current sensor used was the sensor CYHCS-B1-25, Fig. 8.
control and processing systems and the lower plate holds the power
This sensor grants maximum reading measures of 25 A. As with the
system.
voltage sensor, the measure resistance, Rm , must be sized so that
With the non-isolated DC-DC converter built, the implemented
the voltage levels are adequate for the analogic to digital converter.
algorithms were tested in different operational conditions. The dis-
A 350 ohm resistor (Rm ) was also used with this sensor, varying the
cussion and results of the tests are shown in the next section.
voltage from 0 to 8.75 V.

4. Discussion and test results


3.3. Command system of the power circuit (power drive)
In order to qualify and validate the PSO algorithm as a con-
In order to ensure the isolation between the microcontroller and troller, we not only implemented the standard algorithm, but also
the power circuit, a Toshiba photocoupler TLP152 was used. This one variant called GLBest PSO, discussed previously in the paper,
opto-isolator has a maximum commutation frequency of 250 kHz. and compared both in different operating conditions with an anti-
428 J.B.L. Fermeiro et al. / Applied Soft Computing 52 (2017) 418434

Fig. 15. Magnications of the graphics in Fig. 14 of the reference voltage variation test, at the transition from 40 V to 70 V, with an input voltage of 30 V.

Fig. 16. Response in terms of duty-cycle of the three controllers in the reference voltage variation test, changing between 40 V and 80 V, with an input voltage of 30 V.
J.B.L. Fermeiro et al. / Applied Soft Computing 52 (2017) 418434 429

Fig. 17. Response of the Standard PSO based controller to load variation, with a reference voltage of 60 V and an input voltage of 30 V. The controller response in terms of
output voltage in the top graphic, in terms of duty-cycle in the middle graphic and the current variation in the bottom graphic.

windup PI controller, a variant that improves the performance of Table 1


Performance comparison between the three controllers at operation start-up.
the classic PI controller. Its topology is shown in Fig. 12 and the opti-
mal parameters for this controller were determined by the method Settling time [ms] Overshoot [%]
of trial and error for the best performance of the tradeoff between Standard PSO 120
overshoot, settling time (the time required for the response curve to GLBest PSO 110
reach and stay within a certain range of the desired system voltage Anti-windup PI 250 33.3
value at steady-state) and system stability.
In the rst experiment we analyzed the transient response of the
ters, as needed by most controllers used in research, including the
three controllers to a single step input (voltage transition from 0 to
anti-windup PI used for comparison.
60 V). In this case, the PI controller parameters were determined for
In the second experiment we analyzed the transient response
this transition (KP = 0.003 and KI = 0.024), and the overshoot was
from the controllers to a prole (a temporal sequence of reference
affected by the initial conditions and the settling time. Compar-
voltages, starting at 60 V, going to 40 V, 70 V, 80 V, 70 V and nishing
ing both PSO based controllers the response had similar behavior,
with 50 V), Fig. 14. In this case, the controllers were submitted to
but when comparing with the anti-windup PI both PSO based con-
higher converter modulation index (duty-cycle), moving the con-
trollers show better response, not just in terms of overshoot but
verter working point in the direction of the point of maximum gain
also in terms of settling time, as seen in Fig. 13 and Table 1. Note
(to the right in Fig. 3). As a consequence, the PI controller parame-
that the captures were taken with 10 x scope probes and, in case of
ters were adjusted in order to maintain the best tradeoff between
the anti-windup PI controller, the scale for the voltage was set to
overshoot, settling time and the system stability in all the working
2.0 V per division in order to visualize the full range of the transient
points (KP = 0.003 and KI = 0.008). These parameters led to a slower
response.
PI controller response which, in those terms, reduced the overshoot
Another relevant advantage of the proposed controller is the fact
when comparing with the rst experiment. Note that the overshoot
that it does not require the determination of the optimal parame-
only happens in the rst transition, mainly inuenced by the initial
conditions. Also, we can see that in terms of settling time the PSO
430 J.B.L. Fermeiro et al. / Applied Soft Computing 52 (2017) 418434

Fig. 18. Response of the GLBest PSO based controller to load variation, with a reference voltage of 60 V and an input voltage of 30 V. The controller response in terms of
output voltage in the top graphic, in terms of duty-cycle in the middle graphic and the current variation in the bottom graphic.

based controllers show faster convergence responses even with a the values of the duty-cycle used in this prole never exceeded the
lower operation frequency compared with the anti-windup PI con- predened range that ensures the controller stability, as we can see
troller. Another important factor to be taken into account is the fact in the boost converter gain curve (Fig. 3).
that the anti-windup PI controller parameters were determined to In another experiment we studied the behavior of the three con-
obtain the best tradeoff for the specic characteristics of the sys- trollers in a situation of load variation. The controllers were set to a
tem. This implies that the controller performance could differ if reference voltage of 60 V (approximately Gain = 2) and its response
there were changes in operating conditions, whereas the PSO based to load variation is shown in Fig. 17 for the standard PSO controller,
approach wouldnt have this problem. The data in the following g- in Fig. 18 for the GLBest PSO controller, and in Fig. 19 for the anti-
ures were collected via Code Composer Studio Version 5 from Texas windup PI controller. From the results we can see that in terms of
Instruments with a sampling time of 100 ms (maximum sampling settling time and overshoot the controllers have similar behavior at
rate), which led to slightly unsynchronized data from the controller the transition zones, however in terms of steady state oscillations
responses. the anti-windup PI controller shows more oscillations than the PSO
The transient behavior of the controllers in the area highlighted controllers.
in Fig. 14 is illustrated in detail in Fig. 15. There are also magni- In order to make performance and statistical analysis, the con-
cations of the steady-state output voltage oscillations behavior of trollers were subjected to perform a specic prole, Fig. 20. With
the three controllers (between 5 s and 10 s). It is now clear that the the collected data on the behavior of each controller in 15 runs,
PSO controllers exhibit a better behavior in terms of settling time we started by creating a histogram with the error value, Fig. 21.
and output oscillation. This data reveals that the proposed approach has less overall error,
The response of the controller in terms of duty-cycle is shown as it shows more minor error values. This happens because there
in Fig. 16. We can see the effect of particle positioning in the con- are more occurrences in the range close to zero, whereas the anti-
troller response at the transient point, whereas the anti-windup PI windup PI controller has less occurrences in this tighter range but
controller shows a more gradual but slower transition response. It more occurrences in larger error values, meaning it has more over-
is also possible to see more oscillations at the steady-state from the all error.
anti-windup PI controller than from the PSO controllers. Note that
J.B.L. Fermeiro et al. / Applied Soft Computing 52 (2017) 418434 431

Fig. 19. Response of the Anti-windup PI controller to load variation, with a reference voltage of 60 V and an input voltage of 30 V. The controller response in terms of output
voltage in the top graphic, in terms of duty-cycle in the middle graphic and the current variation in the bottom graphic.

Table 2 will penalize larger errors more than smaller ones, and the PSO
Performance indices of the three controllers for the 15 runs with the prole men-
controllers performed better once again, with lower error. The last
tioned above Fig. 20.
indices are the integral time-weighted absolute error (ITAE) and
Performance indices Standard PSO GLBest PSO Anti-windup PI the integral time weighted squared error (ITSE). The ITAE weighs
IE 95 134 12 long time errors much more heavily than those at the start of the
IAE 951 1112 3430 response. The ITSE also weighs the error according to time, but in
ISE 10756 12371 21581 this case it penalizes larger errors more heavily. For both mea-
ITAE 25008 28202 88510
sures, the PSO controllers show less error than the eanti-windup
ITSE 277370 332169 541985
PI controller.
We calculated the average values of the error and the absolute
error, for each of the 15 runs, Table 3. These values were used in
The performance of the controllers was measured using the
order to calculate the average absolute error and standard devia-
indices shown in Table 2 [23,25,27]. The integral error (IE) stands for
tion for all the 15 runs, Table 4. This data shows that the standard
the cumulative sum error and may provide information about the
PSO controller has the lowest mean absolute error from the three
regularity of the controller response to a symmetric prole. Since
controllers and the highest standard deviation. Nevertheless, it has
the PSO controllers depend on particle tracking, which is somewhat
the lowest minimum and maximum values.
random, they could show a more irregular response. In this experi-
In order to determine whether there is a statistically signi-
ment, the standard PSO favored the positive transitions and GLBest
cant difference between standard PSO and anti-windup PI, and also
PSO favored the negative transitions, as we can see in the histogram
between GLBest PSO and anti-windup PI, a Wilcoxons test was per-
(Fig. 21). The anti-windup PI controller has a lower IE value because
formed for the 15 independent runs. The test was conducted with
it has a more regular behavior in both types of transitions (a higher
signicance level = 0.05, and results, in terms of p-value, h-value
IE absolute value indicates a more irregular controller response).
and z-value, are shown in Table 5 for absolute error.
The integral absolute error (IAE) shows the absolute error of the
The results obtained by the standard PSO and GLBest PSO are
controller. The PSO controllers show less overall/absolute error
signicantly different from the anti-windup PI.
compared with the anti-windup PI. The integral squared error (ISE)
432 J.B.L. Fermeiro et al. / Applied Soft Computing 52 (2017) 418434

Fig. 20. Reference voltage prole with four transitions, two with reference voltage increase and two with reference voltage decrease.

Fig. 21. Histogram of the error in all controllers.

5. Conclusion troller, which exhibits a better behavior in terms of settling time,


overshoot, and output oscillation for the initial transition, a bet-
In this work, we propose a new controller for DC-DC converters ter behavior in terms of settling time and output oscillation for
based on particle swarm optimization (PSO). This new converter reference voltage variation, and a similar behavior for load vari-
controller uses the PSO method to directly control, by itself, the ation. The calculated performance indices and statistical analysis
output voltage of a boost DC-DC converter. In order to validate conrm its better performance when compared to the classic con-
and qualify the proposed converter controller, some variants of the troller. Although all the controllers exhibit stabilized outputs, the
PSO algorithm were analyzed, namely the standard PSO and the proposed controller exhibits reduced output oscillations. It is also
variant global local best PSO were developed and implemented. important to mention that it does not require any sort of optimal
The proposed converter controller was then compared with an parameter determination, unlike most popular DC-DC converter
anti-windup PI controller, for different operational conditions. The controllers.
results show better overall performance by the PSO based con-
J.B.L. Fermeiro et al. / Applied Soft Computing 52 (2017) 418434 433

Table 3
Mean error values and mean absolute error values for each of the 15 performed runs with the prole mentioned in Fig. 20.

Run Mean error Mean absolute error

Standard GLBest Anti-windup PI Standard PSO GLBest PSO Anti-windup PI


PSO PSO

1 0.0011 0.0171 0.0467 0.2405 0.1525 0.5974


2 0.0206 0.0657 0.0202 0.0907 0.2058 0.6292
3 0.0943 0.1137 0.0173 0.1266 0.1898 0.5349
4 0.0567 0.0163 0.0105 0.2294 0.2272 0.5407
5 0.0684 0.0172 0.0545 0.1221 0.2106 0.5823
6 0.0079 0.0124 0.0128 0.2011 0.2655 0.5552
7 0.0023 0.0007 0.0176 0.2132 0.1889 0.4859
8 0.0152 0.0018 0.0345 0.2019 0.2228 0.5185
9 0.1160 0.0134 0.0447 0.2044 0.2539 0.5940
10 0.0108 0.1184 0.0004 0.0672 0.2058 0.5609
11 0.0179 0.0604 0.0313 0.1630 0.1563 0.5570
12 0.0347 0.0175 0.0198 0.1578 0.1967 0.5214
13 0.1861 0.0083 0.0182 0.0737 0.1928 0.5380
14 0.0606 0.0258 0.0091 0.1383 0.2098 0.5136
15 0.0748 0.0079 0.0006 0.1270 0.1795 0.5233

Table 4 [10] M. Bensaada, A.B. Stambouli, M. Bekhti, A. Bellar, L. Boukhris, General purpose
Mean value, standard deviation, minimum and maximum values of the absolute pulse width modulation based sliding mode controller for buck DC-DC, World
error for the 15 runs. Acad. Sci. Eng. Technol. 66 (2012).
[11] R.J. Wai, L.C. Shih, Design of voltage tracking control for DC-DC boost
Standard PSO GLBest PSO Anti-windup PI converter via total sliding-mode technique, IEEE Trans. Ind. Electron. 58
Mean absolute error 0.1571 0.2039 0.5502 (2011) 25022511.
[12] M.J. Er, Y.L. Sun, Hybrid fuzzy proportional-integral plus conventional
Standard deviation 0.0563 0.0309 0.0379
derivative control of linear and nonlinear systems, IEEE Trans. Ind. Electron.
Mininum value 0.0672 0.1525 0.4859
48 (2001) 11091117.
Maximum value 0.2405 0.2655 0.6292 [13] P. Siano, C. Citro, Designing fuzzy logic controllers for DC-DC converters using
multi-objective particle swarm optimization, Electr. Power Syst. Res. 112
Table 5 (2014) 7483.
[14] L.P. Guo, J.Y. Hung, R.M. Nelms, Comparative evaluation of sliding mode fuzzy
Comparison of standard PSO with anti-windup PI, and GLBest PSO with anti-windup
controller and PID controller for a boost converter, Electr. Power Syst. Res. 81
PI for the mean absolute error using Wilcoxons rank sum test at = 0.05.
(2011) 99106.
Standard PSO/PI GLBest PSO/PI [15] C.F. Hsu, Design of intelligent power controller for DC-DC converters using
CMAC neural network, Neural Comput. Appl. 18 (2009) 93103.
p-value 3.3918e-06 3.3833e-06 [16] J. Mahdavi, M.R. Nasiri, A. Agah, A. Emadi, Application of neural networks and
h-value 1 1 state-space averaging to DC/DC PWM converters in sliding-mode operation,
z-value 4.6455 4.6461 Mechatron. IEEE/ASME Trans. 10 (2005) 6067.
[17] R.J. Wai, L.C. Shih, Adaptive fuzzy-neural-network design for voltage tracking
control of a DC-DC boost converter, IEEE Trans. Power Electron. 27 (2012)
21042115.
Acknowledgement [18] M.B. Poodeh, S. Eshtehardiha, A. Kiyoumarsi, M. Ataei, Optimizing LQR and
pole placement to control buck converter by genetic algorithm, Control,
The rst author wishes to thank Instituto de Telecomunicaces Automation and Systems, 2007. ICCAS07. International Conference on (2007)
21952200.
for funding this work, with the investigation scholarship PEst- [19] M. Namnaba, M. Bayati Poodeh, S. Eshtehardiha, Comparison the control
OE/EEI/LA0008/2013. methods in improvement the performance of the DC-DC converter, Power
Electronics, 2007. ICPE07. 7th Internatonal Conference on (2007) 246251.
[20] C. Larbes, S.A. Cheikh, T. Obeidi, A. Zerguerras, Genetic algorithms optimized
References fuzzy logic control for the maximum power point tracking in photovoltaic
system, Renewable Energy 34 (2009) 20932100.
[1] R. Priewasser, M. Agostinelli, C. Unterrieder, S. Marsili, M. Huemer, Modeling, [21] R.C. Eberhart, J. Kennedy, A new optimizer using particle swarm theory,
control, and implementation of DC-DC converters for variable frequency Proceedings of the Sixth International Symposium on Micro Machine and
operation, IEEE Trans. Power Electron. 29 (January) (2014) 287301. Human Science (1995) 3943.
[2] E. Sahin, M.S. Ayas, I.H. Altas, A PSO optimized fractional-order PID controller [22] R. Poli, J. Kennedy, T. Blackwell, Particle swarm optimization, Swarm Intell. 1
for a PV system with DC-DC boost converter, 16th International Power (2007) 3357.
Electronics and Motion Control Conference and Exposition (PEMC) (2014) [23] M.I. Solihin, L.F. Tack, M.L. Kean, Tuning of PID controller using particle swarm
477481. optimization (PSO), Int. J. Adv. Sci. Eng. Inf. Technol. 1 (2011) 458461.
[3] H.M. Li, X. Ye, Sliding-Mode PID control of DC-DC converter, ICIEA 2010: [24] S. Morkos, H. Kamal, Optimal tuning of PID controller using adaptive hybrid
Proceedings of the 5th IEEE Conference on Industrial Electronics and particle swarm optimization algorithm, Int. J. Comput. Commun. Control 7
Applications vol. 2 (2010) 165169. (2012) 101114.
[4] F.H.F. Leung, P.K.S. Tam, C.K. Li, An improved Lqr-based controller for [25] J. Zhao, T.P. Li, J.X. Qian, Application of particle swarm optimization algorithm
switching Dc-Dc converters, IEEE Trans. Ind. Electron. 40 (1993) 521528. on robust PID controller tuning, Advances in Natural Computation, Pt 3,
[5] F.H. Dupont, V.F. Montagner, J.R. Pinheiro, H. Pinheiro, S.V.G. Oliveira, A. Peres, Proceedings vol. 3612 (2005) 948957.
Comparison of digital LQR techniques for DC-DC boost converters with large [26] S. Emami, M.B. Poudeh, S. Eshtehardiha, Particle Swarm Optimization for
load range, IEEE International Symposium on Circuits and Systems (ISCAS) improved performance of PID controller on Buck converter, Mechatronics and
(2011) 925928. Automation, 2008. ICMA 2008. IEEE International Conference on (2008)
[6] M. Habib, F. Khoucha, An improved LQR-based controller for PEMFC 520524.
interleaved DC-DC converter, Balkan J. Electr. Comput. Eng. 3 (2015). [27] H. Gozde, M.C. Taplamacioglu, M. Ari, Automatic voltage regulator (AVR)
[7] J.Y. Shi, C.L. Li, P.T. Niu, C.S. Wang, Dynamic controller based linear quadratic design with chaotic particle swarm optimization, Electronics, Computers and
optimal control for stabilization of zero-voltage-switching buck DC/DC Articial Intelligence (ECAI), 2014 6th International Conference on (2014)
converters, Mater. Sci. Inf. Technol. 433440 (18) (2012) 48694872. 2326.
[8] A. Romero, L. Martinez-Salamero, H. Valderrama, O. Pallas, E. Alarcon, General [28] Z. Bouchama, M. Harmas, Optimal robust adaptive fuzzy synergetic power
purpose sliding-mode controller for bidirectional switching converters, ISCAS system stabilizer design, Electr. Power Syst. Res. 83 (2012) 170175.
98Proceedings of the 1998 International Symposium on Circuits and [29] M. Ranjani, P. Murugesan, Optimal fuzzy controller parameters using PSO for
Systems vol. 16 (1998) E466E469. speed control of Quasi-Z Source DC/DC converter fed drive, Appl. Soft
[9] E.M. Navarro-Lpez, D. Corts, C. Castro, Design of practical sliding-mode Comput. 27 (2015) 332356.
controllers with constant switching frequency for power converters, Electr.
Power Syst. Res. 79 (2009) 796802.
434 J.B.L. Fermeiro et al. / Applied Soft Computing 52 (2017) 418434

[30] T. Vishnu, T. Shanavas, S. Patnaik, Modied PSO based sliding-mode controller [42] R. Mendes, M. Calado, S. Mariano, Particle swarm and Boxs complex
parameters for buck converter, Electrical, Electronics and Computer Science optimization methods to design linear tubular switched reluctance generators
(SCEECS), 2012 IEEE Students Conference on (2012) 14. for wave energy conversion, Swarm Evol. Comput. 28 (June) (2016) 2941.
[31] K. Ishaque, Z. Salam, A review of maximum power point tracking techniques [43] M. Zambrano-Bigiarini, M. Clerc, R. Rojas, Standard particle swarm
of PV system for uniform insolation and partial shading condition, Renewable optimisation 2011 at CEC-2013: a baseline for future pso improvements,
Sustainable Energy Rev. 19 (2013) 475488. Evolutionary Computation (CEC), 2013 IEEE Congress on (2013) 23372344.
[32] L.Q. Liu, X.L. Meng, C.X. Liu, A review of maximum power point tracking [44] M. Clerc, J. Kennedy, The particle swarm-explosion, stability, and convergence
methods of PV power system at uniform and partial shading, Renewable in a multidimensional complex space, IEEE Trans. Evol. Comput. 6 (2002)
Sustainable Energy Rev. 53 (2016) 15001507. 5873.
[33] Z. Salam, J. Ahmed, B.S. Merugu, The application of soft computing methods [45] R.C. Eberhart, Y. Shi, Comparing inertia weights and constriction factors in
for MPPT of PV system: a technological and status review, Appl. Energy 107 particle swarm optimization, Evolutionary Computation, 2000. Proceedings of
(2013) 135148. the 2000 Congress on (2000) 8488.
[34] K. Ishaque, Z. Salam, A. Shamsudin, Application of particle swarm [46] F. Van den Bergh, A.P. Engelbrecht, A convergence proof for the particle
optimization for maximum power point tracking of PV system with direct swarm optimiser, Fundam. Inform. 105 (2010) 341374.
control method, IECON 2011-37th Annual Conference on IEEE Industrial [47] P.K. Patel, V. Sharma, K. Gupta, Guaranteed convergence particle swarm
Electronics Society (2011) 12141219. optimization using personal best, Int. J. Comput. Appl. 73 (2013).
[35] A. Zulu, S. John, A review of control algorithms for autonomous quadrotors, [48] C. Leboucher, H.-S. Shin, P. Siarry, S. Le Mnec, R. Chelouah, A. Tsourdos,
Open J. Appl. Sci. 4 (2014) 547. Convergence proof of an enhanced particle swarm optimisation method
[36] S. Pattnaik, K. Bakwad, S. Devi, B. Panigrahi, S. Das, Parallel bacterial foraging integrated with evolutionary game theory, Inf. Sci. 346347 (2016) 389411.
optimization, in: Handbook of Swarm Intelligence, Springer, 2011, pp. [49] J. Fermeiro, J. Pombo, M. Calado, S. Mariano, Evaluation of a particle swarm
487502. optimization controller for Dc-Dc boost converters, Compatibility and Power
[37] Y.H. Shi, R. Eberhart, A modied particle swarm optimizer, IEEE International Electronics (CPE), 2015 9th International Conference on (2015) 179184.
Conference on Evolutionary ComputationProceedings (1998) 6973. [50] H. Mekki, O. Benzineb, D. Boukhetala, M. Tadjine, M. Benbouzid, Sliding mode
[38] A. Ratnaweera, S.K. Halgamuge, H.C. Watson, Self-organizing hierarchical based fault detection, reconstruction and fault tolerant control scheme for
particle swarm optimizer with time-varying acceleration coefcients, IEEE motor systems, ISA Trans. 57 (July) (2015) 340351.
Trans. Evol. Comput. 8 (2004) 240255. [51] J. Hstad, R. Impagliazzo, L.A. Levin, M. Luby, A pseudorandom generator from
[39] Y. Liu, Z. Qin, X.S. He, Supervisor-student model in particle swarm any one-way function, SIAM J. Comput. 28 (1999) 13641396.
optimization, CEC2004: Proceedings of the 2004 Congress on Evolutionary [52] R. Rastogi, S. Mittal, S. Shekhar, Linear algorithm for imbricate cryptography
Computation vol. 1 and 2 (2004) 542547. using pseudo random number generator, Computing for Sustainable Global
[40] G.R. Murthy, M.S. Arumugam, C.K. Loo, Hybrid particle swarm optimization Development (INDIACom), 2015 2nd International Conference on (2015)
algorithm with ne tuning operators, Int. J. Bio-Inspired Comput. 1 (2009) 8994.
1431.
[41] M.S. Arumugam, M.V.C. Rao, A. Chandramohan, A new and improved version
of particle swarm optimization algorithm with global-local best parameters,
Knowl. Inf. Syst. 16 (2008) 331357.

Вам также может понравиться