InternationalJournalonElectricalEngineeringandInformaticsVolume5,Number3,September2013

Constrained Model-Based Predictive Controller for a High-Frequency

Low-Power DC-DC Buck Converter
Ghulam Abbas1, Umar Farooq2, Jason Gu3 and Muhammad Usman Asad4
1,4
Department of Electrical Engineering, The University of Lahore, Lahore Pakistan
2
Department of Electrical Engineering, University of the Punjab, Lahore Pakistan
3
Department of Electrical and Computer Engineering, Dalhousie University Halifax, N. S. Canada
1
[email protected], [email protected], [email protected],
4
[email protected]

Abstract: This paper describes a very detailed and comprehensive description of a

model predictive controller (MPC) applied to the buck converter working in Continuous
Conduction Mode (CCM) to optimize the performance of the converter. The converter
is designed for a fixed switching frequency of 1 MHz. The proposed model predictive
control technique achieves improved set point tracking with minimal overshoot, offers
consistent closed loop performance and has short settling time. Although MPC rejects
unmeasured disturbances effectively, but it shows poor load and line regulation.
Solution has been suggested in order to robustify the load and line regulation. Further
neural network based MPC has also been designed. The impact of the MPC parameters
such as prediction and control horizons, performance weights, and constraints on the
performance is minutely observed. In order to investigate the effectiveness of the
proposed algorithms, MPC is compared with the classical well-known PWM-PID
controller. MATLAB/Simulink simulation results validate the control design procedure.

Keywords: DC-DC Buck Converter, Model Predictive Controller (MPC), PWM-PID,

Continuous Conduction Mode (CCM), Neural Network, Robustification

1. Introduction
Synchronous buck converters are being extensively used in regulated switched-mode power
supplies (SMPS) targeted for portable applications. However, toggling of the switching
converters among a set of linear or nonlinear circuit topologies during an operation,
nonlinearity caused by the switching action, chaosity which underlines the complexity, etc.
make the power switching converters quite complex and complicated. Due to its switching
characteristics, the control problems associated with such converters still pose theoretical and
practical problems and challenges to the control community [1]. Linear control techniques do
not always guarantee the required static and dynamic performance. The paper proposes the more
advanced robust control technique i.e. the constrained model predictive control (MPC) to
comply with the challenges.
The acceptable performance for the power switching converters could be achieved using
the well-established and most commonly used control techniques like PID and lead-lag [2, 3].
PID parameters normally need to be tuned efficiently to obtain the excellent transient response
and robustly ensure the system stability. Depending on tuning guidelines, the tuning process
sometimes becomes quite sticky and laborious to achieve a good trade-off between settling
time and percent overshoot, particularly under the load variation or input supply voltage
perturbation. Model predictive control, although its derivation is more intricate than that of the
PID controllers, however is far more intelligent than PID as it incorporates the constraints
within the optimization [4].
Although many other nonlinear control methods have been studied, such as fuzzy logic
control [5], neural network control [6, 7] and sliding mode control (SMC) [8] to resolve the
aforementioned problems and challenges, but we will apply a promising and in real sense an

Received: February 22nd 2013. Accepted: September 2nd, 2013

316
Ghulam Abbas, et al.

advanced model predictive control (MPC) technique to buck converter to address the issues.
Model predictive control (MPC), also known as receding-horizon control (RHC), makes
explicit use of the process model and its current measurements to calculate the control signal,
at each sampling instant over a finite prediction horizon, by minimizing the cost function.
Unlike the conventional control techniques where the designer tunes, often by trial and error,
the controller coefficients and gains to indirectly handle the constraints, the sophisticated RHC
control technique optimizes the performance index subject to the various hard and soft
constraints to meet the best dynamic performance [9] while manifestly amalgamate the
additional control information, like the future predicted outputs and the estimates of future
disturbances.
Although MPC suffers from the obvious drawbacks like more complex control algorithm
derivation as all the computation has to be carried out at every sampling time, need for an
appropriate model of the process, difficulties in its implementation although it is now possible
to implement MPC with shorter time steps [10, 11] and with the help of FPGA [12] and
microcontrollers [13] and its restriction to stable systems [14] but it offers potential advantages
including its ability and capability of systematically coping with hard constraints on inputs,
outputs and states, its suitability for directly addressing multivariable systems, the ability to
systematically handle MIMO control problems subject to constraints, its demonstrability,
simplicity, richness, practicality, the on-line solution of an open-loop optimal control problem,
need limited knowledge of control, compensation for the measurable disturbances and dead
times and its ability to handle non-minimal phase processes.
Historically model predictive control techniques have been extensively used to control
constrained multivariable process with complex dynamics [15]. [16] proposes an MPC
controller, globally asymptotically stable, that behaves just like a linear controller while
optimally dealing with constraints during transients. Distributed model predictive control [17]
is suggested to regulate the large scale nonlinear process outputs by considering several distinct
sets of manipulated inputs. Although MPC finds its significance for the multi-inputs, multi-
outputs (MIMO) complex and nonlinear processes, but it can equally be applied to single-
input, single-output (SISO) linear processes which also undergo inequality constraints. On the
basis of the discrete-time piecewise affine (PWA) modeling of the power converters where the
duty cycle ratio is considered as a constrained manipulated variable, the discrete-time model
predictive control (MPC) is familiarized particularly in [18, 19].
Sanders et al. showed the benefits of MPC compared to a standard PID controller [20].
Further the duty ratio is bounded in the interval [0, 1] in power converters. From an
optimization point of view this is a constrained optimization problem. Hence, it is natural to
apply the model predictive control (MPC). In addition, during the last decade, neural networks
found their extensive use in various control applications [21, 22, 23]. Nonlinear controllers
including MPC can be realized using neural networks which use their universal approximation
capabilities of the multilayer perceptron [24]. Several studies ensured the promising results of
the application of neural network based nonlinear model predictive control to successful
industrial applications [25, 26]. Thus this paper investigates comprehensively the design and
analysis of the model predictive controller.
Rather than following the complex mathematics to describe the MPC algorithm, the paper
mainly focuses on the systematic and detailed description of the design of not only the
generalized MPC controller but also of the MPC controller designed on the basis of neural
networks for the buck converter to achieve the fast dynamic response. The paper minutely
studies the effect of the parameters associated with MPC on the closed-loop performance as
well. The MPC controller is then compared with the classical PID to confirm its effectiveness.
The paper is organized as follows: The small-signal mathematical model of the buck converter
in the form of transfer function and state-space is reviewed in Section II. Section III describes
the detailed design procedure of the three control techniques: model predictive control, neural
network based model predictive control and PID controller. The simulation results obtained on

317
Constrained Model-Based Predictive Controller for a High-Frequency

the basis of these control techniques are given and compared in section IV. Conclusions are
drawn in Section V.

2. Buck Converter Modeling

Model predictive control requires a well-defined linear dynamic model of the plant. In this
section, we give the birds eye view of the buck converter model.
A buck converter converts a higher unregulated DC input voltage Vin into a lower regulated
output voltage Vout at a required voltage level. The block diagram incorporating a buck
converter and a feedback loop is illustrated in Figure 1. The converter consists of a power
switching network (transistors Q1 and Q2) and a second order LC low pass filter with parasitics
such as inductor DC resistance, RL, and capacitor equivalent series resistance, RC, denoted
explicitly. The capacitor ESR, RC, results in introducing a zero frequency to the transfer
function [27]. The typical values used for the converter are shown in Table 1.

Switching Circuit 2nd Order Low-Pass Filter Load

Q1 d.Vin RL L Vout

Q2 RC
Vin R
C

d Model Predictive Controller e -

(MPC) +

Vref
Figure 1. Closed loop analog buck converter system.

Table 1. Buck converter parameters with their typical values

Parameter Symbol Value
Load Resistance R 4.5
Series Resistance of Inductor RL 505 m
Series Resistance of Capacitor RC 5 m
Output Filter Inductor L 4.7 H
Output Filter Capacitor C 4.7 F
Input Voltage Vin 3.6 V
Output Voltage Vout 2V
Reference Voltage Vref 2V
Switching Frequency fs 1 MHz

318
Ghulam Abbas, et al.

The open-loop control-to-output voltage transfer function in s-domain is given by [28, 29]:

R
Vin ( s ) ( RC Cs + 1)
vout ( s )
= R + RL
d R + RC 2 L
LC s + + C ( R RL ) + RC C s + 1
R + RL R + RL
s
+1
R ZERO
= Vin ( s) 2
R + RL s + s + 1
2 Q
0 0
(1)

with
1 1 1 ;
0 = ; ZERO = ;Q=
R + RC R C L
LC C
0 + C ( R RL ) + RC C
R + RL R + RL
(2)

Where, 0, ZERO and Q represent the LC filter complex double pole frequency, output
capacitor ESR zero frequency and the quality factor, respectively.
The current flowing through the inductor (iL) and the voltage across the capacitor (vC) are
the two state-variables that characterize the buck converter. If we define the state variables as
x = [iL vC ] , the buck converter dynamics in the form of an LTI state-space can be
T

expressed as:
x& = Ax + Bu
(3)
y = Cx + Du

Where, u (duty cycle) is the control input, x is the state vector, and y (the output voltage) is
the measurement, and the various matrices are

( RRC + RRL + RC RL ) RC
C Vin
1 R + RC R + RC
A= B= L
LC RL L
0
R + RC R + RC
RRC R
C= D = [ 0]
R + RC R + RC

The continuous system is discretized by using the sampling period Ts = 1 s and a zero
order hold element. The discrete-time state-space dynamic buck converter system is described
by:
x( k + 1) = x(k ) + u (k )
(4)
y (k ) = hx(k ) + Ju (k )

319
Constrained Model-Based Predictive Controller for a High-Frequency

Where, x(k ) X n , u (k ) U m and y ( k ) Y p represent the system state,

control input and output variable, respectively. The matrices and depend on Ts, the
sampling period and are defined as:
Ts
= e A Ts ; = e A d B; h=C; (5)

0
With,
0.8766 -0.1952 0.7205
= =
0.1952 0.9328 0.0770
h = [ 0.0050 0.9989] J = [0]

This completes the description the buck converter in the form of transfer function and state-
space form. The MATLAB toolbox used in the paper may take all forms of the converter.

3. Controller Design
In this section, we describe the model predictive control and neural network based model
predictive control techniques. Main focus is on MPC while PID is just described for
comparison purpose.

A. Model Predictive Controller

MPC is a control algorithm, based on numerically computing an on-line finite horizon
open-loop optimal control signal, which solves the optimization problem typically a
constrained optimization problem, at each sampling instant, using the current state of the plant
as the initial state.
Although different design procedures depending on the presence of constraints may be
adopted to describe the MPC algorithm, but we will use the most generalized design procedure
as described in [30]. The algorithm calculates a train of the control signal adjustments
minimizing a cost function evaluating the difference between the reference signal and the
predicted process output in a given future horizon to take the process to the desired regime.

A.1 Basic Control Strategy

The control strategy applies a receding horizon principle which is schematically presented
in Figure 2 for a simplified single-input single-output (SISO) system and is summarized in the
following steps:

At each consecutive sampling instant k, a dynamic process model along with models
of constraints is applied to predict the future process outputs for an assumed
prediction horizon k 0, N p . These predicted outputs

{ y ( k + i k ) ; i = 0,..., N 1} not only depend on the known past values of control

p

input (manipulated variable) and output signals (controlled variable) up to instant k

but also on the future control signals {u ( k + i k ) ; i = 0,..., N m }
1 which are
applied to the plant. Based on the dynamic plant model, the optimizer generates the
future control trajectory as a possible solution, which incorporates the minimization of
the objective function while fulfilling the process constraints. The objective function J
(
usually returns the deviation of the predicted output trajectory y k + i k from the )

320
Ghulam Abbas, et al.

reference r ( k + i ) . J can also include a penalty for the consumption of control

strategy.
At each consecutive sampling instant k, a sequence of future control signals
u ( k ) = u ( k k ) , u ( k + 1 k ) ,..., u ( k + N m 1 k ) is computed, assuming

u ( k + h k ) = u ( k + N m 1 k ) ; h = N m ,K , N p , where N m is the control

( )
horizon. The employed notation u k + h k means the prediction of the control
input value for the future time k + h, performed at the time k. The series of control
inputs is calculated in such a way that it minimizes the differences between the
( )
predicted controlled outputs y k + i k and the some reference sequence r ( k + i )
for these outputs over the prediction horizon N p . The lower the difference is, the

( )
better the setpoint tracking is. Then, only the first element u k k of the calculated

(control) sequence is applied to control the process, i.e., u ( k ) = u k k . The( )

remaining control signals are discarded because at time k + 1 a new output y ( k + 1)
is not known. At the next sample (k + 1), there occurs a new measurement of the
process outputs and the controller then obtains new measurements and totally revises
its plan. This cycle repeats indefinitely with the prediction horizon of the same length,
but shifted by one step forward in time.

Figure 2. Illustration of working principle of MPC for SISO system.

A.2 Objective Function with Constraints

For a MIMO system, under the assumption that at each sampling instant k, a well-distincted
plant dynamic plant model and the estimates of the state variable vector x(ki ) and the

321
Constrained Model-Based Predictive Controller for a High-Frequency

measured disturbance vector xd (ki ) are available, the standard quadratic objective function
(cost function) J needs to be minimized may take the following general form:

N p 1 ny
{ (
i +1, j. y j ( k + i + 1 k ) rj ( k + i + 1) )}
y 2

i =0 j =1
N m 1 nu
{
J = + i , j.u j ( k + i k ) }
u 2
+ 2
i =0 j =1
N m 1 nu
{ (
. u j ( k + i k ) u jtarget ( k + i ) )}
2
+

u
i =0 j =1 i , j

min
N m 1

u = u ( k + i k ) , (6)
i =0

Where, N p is the prediction horizon and N m is the control horizon. N m is equal to the
number of maximal allowed changes of the manipulated variables. j denotes the jth component
( )
of a vector and " k + i k " denotes the value predicted for time k + i based on the
information available at time k. y j are the plant controlled or outputs variables (OVs), u j are
the plant inputs or manipulated variables (MVs) and u j = u j u j 1 is the rate of change of
the manipulated variable at the jth sampling instant . nu and ny are the number of plant inputs

y u u
and outputs, respectively. The variables, i +1, j
, i, j
and i, j
are the non-negative
performance weights to penalize the certain components of the variables y, u and u,
respectively at certain future time intervals. The smaller the value of the less important is the
role of the corresponding term in the cost function. The variable ( 0 ) is the slack
variable that is introduced to relax the constraints on u, u, y. The weight on the slack
variable penalizes the violation of the constraints. The larger the slack variable weight
compared to the manipulated and output variables weights, the more the constraint violation is
penalized. The tuning and, therefore, the behavior of the MPC can be adjusted by choosing the

u u u
appropriate weights i +1, j
, i, j
and i, j
.
Denoting the current sample of a reference for all the measured and unmeasured outputs by
r(k), subject to the model equations and to the inequality constraints:

y j ( k + i ) = f ( u j ( k ),K , u j (k + i 1) ) i = 0,..., N p 1
y j min ( i ) y j ( k + i + 1 k ) y j max ( i ) + i = 0,..., N p 1
u j min ( i ) u j ( k + i k ) u j max ( i ) i = 0,..., N m 1
u j min ( i ) u j ( k + i k ) u j max ( i ) i = 0,..., N m 1
u ( k + h k ) = 0 h = N m ,..., N p 1
0
(7)

322
Ghulam Abbas, et al.

Assuming all the input constraints are hard and all the output constraint are soft. In case the
reference r is not known in advance, the current reference r(k) is used over the whole
prediction horizon, i.e. r ( k + i + 1) = r ( k ) . The predicted y j ( k + i ) also depends on known
and estimated disturbances implicited in (7) as their effects remain constant.

Figure 3. Block diagram representing the basic structure of MPC.

As far as the buck converter system is concerned, it has one manipulated variable i.e. duty
cycle ratio (d) and the one output voltage (Vout) to be controlled. The buck converter is
essentially a single-input single output SISO system. The duty ratio, being a control signal, is a
sufficiently smooth signal rigidly limited in the closed interval [0, 1]. The vector
u jtarget ( k + i ) which signifies the setpoint for manipulated variable, usually incorporated into
the objective function when the number of MVs is greater than that of the OVs, can be
excluded from the objective function as we have not imposed any setpoint condition on the
duty cycle. The vector u jtarget ( k + i ) , in general, is just a sort of lower-priority setpoint.
As the output voltage cannot exceed the input voltage in step-down converter, consequently,
we have imposed a constraint on the output voltage i.e. 0 Vout Vin , unlike the constraint on
inductor current described in [31]. In this regard, our approach differs slightly as described in
[32]. However this additional soft constraint has negligible effect on the performance.
Maximum weight is assigned to the output voltage error in order to achieve the better setpoint
tracking.
As can be observed from the cost function that rather than computing the control signal u
directly, its gradient, u ( k | k ) is actually calculated at each step to calculate u.
Accomplishing this actually allows the control signal to reach the desired level to assist in
producing the required output voltage without worsening the performance index J. As already
remarked that all the remaining samples are rejected and a new optimization problem on the
basis of y ( k + 1) is constructed for the next sampling step k+1.
Since the main objective of the MPC is to obtain the fast setpoint tracking, high weight
factor is assigned to the error term in the cost function. The control signal variation term
usually bears a low weight factor value. In addition, as we have not applied any degree of
softness on any of the constraints, the slack variable may be dropped so does its weight. We
explicitly have used the hard and soft constraints.
To ensure a steady-state operation under a constant non-zero duty cycle, the difference
between two consecutive duty cycles, in the differential form, is computed as:

323
Constrained Model-Based Predictive Controller for a High-Frequency

d ( k ) = d ( k ) d (k 1) (8)

The objective function needs to be minimized in case of buck converter system (SISO) thus
reduces to:
N p 1
{ ( )}
2
i +1. y ( k + i + 1 k ) r ( k + i + 1)
y

(
J N p , N m , i +1, i
y u
) = i =N0 1
m


(9)

{ }
2
+ i .u ( k + i k )
u

i =0

The constraints on the control input and its gradient and the output signal may be described
by the following inequalities:
y j min ( i ) y j ( k + i + 1 k ) y j max ( i ) i = 0,..., N p 1
u j min ( i ) u j ( k + i k ) u j max ( i ) i = 0,..., N m 1 (10)

u j min ( i ) u j ( k + i k ) u j max ( i ) i = 0,..., N m 1

The above ISE (Integral of Squared Error) problem can be converted into an IAE (Integral
of Absolute Error) problem by replacing the squared terms with absolute values which place
less emphasis on the large deviations.
As can be seen from the optimization problem that the basic tuning parameters of the algorithm
are the prediction and control horizons and the non-negative performance weights in the cost
function. Some notions regarding the performance weights have already been highlighted.

A.3 Prediction and Control Horizon Selection

Regarding the choice of Np and Nm, there are no specific rules. Insensitivity of controllers
performance to small perturbations in horizons gives the notions of the Np and Nm values. The
selection of prediction horizon Np is highly dependent on the sampling period Ts. In order to
ensure the guarantee of the stability, the prediction horizon should be chosen in such a way that
it must incorporate the steady-state effect of all the previous control moves. In other words,
numerically, it must be equal to the plants open-loop settling time which is approximately, by
rule of thumb, 20 30 times the sampling period. The prediction horizon is usually taken equal
to the plants open-loop settling time. Relatively small value of control horizon as compared to
prediction horizon is selected. The optimization problem, if gives imprecise control
predictions, should be solved for other values of Np and Nm in order to obtain required
performance. An approach has been proposed in [33] to calculate the control input for
optimized performance even using the lower value of prediction horizon.
The parameters used by the model predictive controller for the nominal performance (to be
shown in simulations results section) are summarized in Table 2.

Table 2. Predictive control parameters

Parameter Value Parameter Value

Prediction Horizon (Np) 14 Tunning Weight 0.8

Control Horizon (Nm) 4 Control Interval (Ts) 1 s

A.4 Robustification
The simulation results reveal that MPC shows poor load and line regulation (see the
simulation results section), means that when there is a change in load current or input voltage,

324
Ghulam Abbas, et al.

the output voltage does not settle down to its steady-state value, showing some steady-state
error. MPC also suffers from the problem in the same way while dealing with the physical
limitations, model imperfections, parameter mismatch and unmeasured disturbances. Some
methods suggesting the addition of some sort of robustness to MPC can be found in [34, 35].
[36] employs Kalman filter to adjust the output voltage reference against the unmeasured
changes in the load resistor. However, we use the simple but effective method to deal with the
problem.
The shortcoming arises due to the fact that the control signal applied to the process depends
only on the current state, and not on the past history of the process. Consequently the MPC
does not take into account the changes in plant dynamics occurring due to the perturbation in
supply voltage or load current, thus creating output voltage convergence problem. The
controller thus does not provide any guarantee of steady-state zero offset against the load and
line transients.
The proposed MPC algorithm can be robustified by introducing an external loop with PI
compensator as shown in Figure 4. With the addition of an external loop, the predicted output
voltage error which is not available due to the offline realization of the optimization is
processed by the PI compensator before the MPC action to move the reference Vref .

Figure 4. Double-loop control system to ensure steady-state error free tracking.

'
The newly resulted reference signal Vref to MPC is expressed as:
t
Vref' = K p e(t ) + K i e(t )dt where e(t ) = (Vref Vout ) (11)
0

Surely now (Vref Vout ) is the manipulated variable to MPC under the new configuration.
'

It has been learned that MPC supported by PI controller ensures offset-free tracking. The PI
controller in the external loop of the system adjusts the output voltage reference of MPC in
such a way that it safeguards the output voltage tracking under the load and line variation.
We have used the same integral term of the PID compensator (to be discussed later). The
proportional constant is found by a hit and trial method. The integrator assists in eliminating
the steady-state tracking error whereas the proportional constant tends to ameliorate the
transient response without effectively disturbing the system dynamics. The PI compensator
employed in the external loop is:

Ki 101336.45
GPI ( s ) = K p + = 0.2 + (12)
s s

B. NN Model Predictive Controller [7]

Neural networks based model predictive controller can be realized by employing Neural
Network Toolbox [37] which uses a neural network model of a nonlinear plant. The future
plant performance can be predicted using the Network. The controller determines the control
input that assists in optimizing the plant performance over a specified future time horizon.

325
Constrained Model-Based Predictive Controller for a High-Frequency

There involves two steps, namely system identification and predictive control when using
neural networks for control of a system.
The first step involves the training of a neural network by model predictive control in order
to perform the system identification. Figure 5 depicts that the prediction error between the
plant output and the neural network plant output constitutes the neural network training
signal.

Figure 5. Complete closed-loop control system.

In order to predict the future values of the plant output, NN plant model uses previous
inputs and previous plant outputs. In Figure 6, the inputs to the neural network, namely yp(t)
and u(t) denote the current plant output and the control signal to the plant respectively whereas
ym(t+1) denotes the predicted plant output.

Figure 6. NN plant model [37] (Only two hidden layers are shown).

The output of the network, being the duty cycle, is bounded to values between 0 and 1. This
forces the output layer to use a sigmoid transfer function which follows the four hidden layers
of sigmoid neurons of feedforward networks in our case. An appropriate weight w is assigned
to each input. The sum of the weighted inputs and the bias constitutes the input to the transfer
function. Various activation functions are tabulated in Table 3.

Table 3. Activation functions

Activation Functions MATLAB Description Expression/Definition

Linear purelin f ( x) = x
1
Linear-Sigmoid logsig f ( x) =
1 + e x
e x e x
Tan-Sigmoid tansig f ( x) = x x
e +e

326
Ghulam Abbas, et al.

The network is trained offline in batch mode using Levenberg-Marquardt (trainlm) training
algorithm which optimizes the network weights and biases to minimize the cost function. The
cost function to be minimized for feedforward networks is mean square error (MSE) the
average squared error between the network outputs ym(t) and the target outputs yp(t) is
defined as:
N N 2
1 1
( ei ) ( y m (t ) y p (t ) )
2
F = MSE = = (13)
N i =1 N i =1

Any standard numerical optimization algorithm based either on the Jacobian of the network
errors with respect to the weights or the gradient of the network performance with respect to
the network weights can be employed to minimize the cost function. Backpropagation
algorithm is used to calculate the gradient [38].
The second step involves the determination of receding horizon technique [39] based
predictive controller which uses the plant model to predict future performance over a specified
time horizon. The numerical optimization program uses the predictions to calculate the control
signal that minimizes the following cost function:
N2 2 Nu 2

J= ( yr ( t + j ) - ym ( t + j ) ) + ( u ( t + j -1) - u ( t + j - 2 ) ) (14)
j = N1 =1 j =1

The optimization block which determines the values of the tentative control signal u' (input
to the plant) that minimizes the cost function J and the neural network plant model form the
controller (see Figure 7).

Figure 7. NN plant model with NN model predictive controller [37].

The parameters used by the neural network and predictive control for optimal performance,
are listed in Table 4.

Table 4. NN Predictive control parameters

Parameters Values Description Value
Size of Hidden Layers 4 Cost Horizon (N2) 14
Training Epochs 200 Control Horizon (Nu) 4
Training Function 'trainlm' Control Weighting Factor () 0.005
Training Samples 1000 Search Parameter () 0.001

327
Constrained Model-Based Predictive Controller for a High-Frequency

C. PID Controller
Major concentration has been devoted to MPC controllers. For the sake of comparison,
most widely used classical PID compensator with filter is presented in this section.
Assuming the derivative term is made implementable by converting it into a lead term, a
continuous-time parallel PID controller, the output of which is the sum of the proportional,
integral, and derivative actions, weighted according to the independent gain parameters K p ,
K i , and K d , is generally given by:
Ki Ns
GPID ( s ) = K p + + Kd (15)
s s+N

where K p , K i and K d are the proportional, integral and the derivative constants respectively.
The filter pole in the derivative action which lies at s = -N is determined by the filter
coefficient N. The tuning of the parameters, namely K p , K i , K d and N can be accomplished
automatically through SISO Design Tool [40] or Simulink Control Design toolbox. The PID
gains using the PID Tuner interface for optimal performance are calculated to be:

K p = 1.28, Ki = 1.01105 , K d = 3.6 106 , N = 1.98 106

4. Simulation Results
In order to investigate the behavior, performance and effectiveness of the constrained MPC
described in section III; simulation results using the MATLAB/Simulink environment are
presented. Simulation results demonstrate that the proposed MPC offers potential advantages.
For this purpose a voltage-mode controlled, synchronous buck converter system which
converts 3.6 V (unless otherwise stated) into 2 V, operating at a switching frequency of 1 MHz,
is tested and simulated. The buck converter uses the same component values summarized in
Table 1. All the measurement noises and the measured and unmeasured disturbances are
neglected. The performance of the suggested MPC is also compared with the well-tuned
traditional PWM-PID controller.

A. Nominal Results
Figure 8 exhibits the step response of the buck converter assuming all the initial conditions
to be zero. MPC shows an underdamped output voltage response and achieves the performance
specifications of maximum overshoot less than 1% and 7 s settling time whereas PWM-PID
shows 6% maximum overshoot and 15 s settling time. This shows that MPC, compared to the
PWM-PID, takes 2.14 times less time to settle down to its steady-state voltage value of 2 V.
NN-MPC almost shows the same performance as that of MPC (settling time = 87 s, overshoot
= 3.65%). The compensated system almost shows the zero steady-state error. The constraint
imposed particularly on the duty cycle ratio is efficiently respected. The proposed MPC
provides fast closed-loop dynamical performance for the startup without overshoots.
Regarding the NN-MPC, at epoch 200 the validation performance achieves its minimum
value of 1.05x10-10 (see Figure 9). A regression plot shown in Figure 9 clearly ensures the
exact linear relationship between the outputs and targets as R = 1.

328
Ghulam Abbas, et al.

2.5
2

V out (V )
1.5
MPC
1
0.5 PID
0
0 0.2 0.4 0.6 0.8 1
-4
x 10
2

iL (A )
1

0
0 0.2 0.4 0.6 0.8 1
-4
1.2 x 10
1
d

0.5

0
0 0.2 0.4 0.6 0.8 1
Time (s) -4
x 10
(a)

2.5
V out (V )

2
1.5
1
0.5
0
0 0.2 0.4 0.6 0.8 1
-4
x 10
2
1.5
iL (A )

1
0.5
0
0 0.2 0.4 0.6 0.8 1
-4
x 10
1.2
1
0.8
0.6
d

0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
Time (sec) -4
x 10

(b)

Figure 8. Nominal performance offered by (a) MPC and (b) NN-MPC.

329
Constrained Model-Based Predictive Controller for a High-Frequency

(a)

Figure 9. (a) Performance plot. (b) Regression plots.

B. Effect of Prediction and Control Horizons

The model predictive control strategy is evaluated and investigated using various predictive
horizon values. Nominal closed-loop performance improves as Np increases, provided that Nm
is sufficiently large. Alternatively, if the Np is large enough, increase in Nm may improve the
performance. However Nm is usually kept lower than Np. Simulation results depicted in Figure
10 remark that better response is achieved for the Np value of 14 and Nm value of 4. The
performance roughly remains the same even if we increase the Np value beyond 14, keeping the
Nm equal to 4. No significant improvement is observed even for higher values of Np due to the
process-model mismatch. We will use Np = 14 and Nm = 4 as the most efficient values for the
other simulation results in order to reduce the computational time as the algorithm calculates
the control signal at each sampling instant. MPC with horizon value of 14 results in near-
optimal control.

330
Ghulam Abbas, et al.

Output Voltage Response

2.5

1.5
P=14; M=4

Vout (V)
P=20; M=4
P=30; M=4
1
P=14; M=8

0.5

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Time (sec) -4
x 10
Output Voltage Response
2.5

1.5
P=14; M=4
Vout (V)

P=10; M=4

0.5

0
0 0.2 0.4 0.6 0.8 1
Time (sec) -4
x 10
Figure 10. Effect of prediction and control horizons by (left) MPC (right) NN-MPC.

C. Effect of Weight Tuning

As mentioned earlier, weighting is one of the tuning parameters to optimize the
performance. Well-adjusted tuning weights based controllers lead to optimal control.
The cost function expression incorporates the non-negative performance weights. This
clearly shows that the selection of the weights affects the closed-loop performance. The more
the weighting, the less active the changes are. As far as the buck converter under consideration
is concerned, we have given full weight to the output voltage error in order to ensure the
excellent setpoint tracking. Less weight has been assigned to the constrained manipulated
variable, ideally zero whereas gradient of the MV gets less weight. Using the nominal values of
the prediction and control horizon, the step response clearly shows that sufficiently-weighted
control signal drastically deteriorates the setpoint tracking. This observation needs some simple
insight. The performance index previously dominated by the steady-state output tracking error
where the optimization places all the emphasis on making tracking errors small as control
weighting was relatively small , does not remain the same when we give reasonable weight to
the gradient of the control signal. This results in poor setpoint tracking.

331
Constrained Model-Based Predictive Controller for a High-Frequency

2.5

Vout (V)
1.5

0.5

0
0 1 2 3 4 5 6 7 8
-5
x 10

0.8

0.6
d

0.4 Weight on MV: 0.0, OV=1.0

Weight on MV: 0.5, OV=0.5
0.2

0
0 1 2 3 4 5 6 7 8
Time (s) -5
x 10
Figure 11. Weight tuning effect shown by MPC.

D. Effect of Constraints
2.5

Constrained
Unconstrained
1.5
Vout (V)

0.5

0
0 0.2 0.4 0.6 0.8 1
Time (sec) -4
x 10

4
Constrained
Unconstrained
2
u

-2

-4
0 0.2 0.4 0.6 0.8 1
Time (sec) -4
x 10
Figure 12. Performance of MPC with and without constraints.

332
Ghulam Abbas, et al.

MPC tends to handle the active constraints, soft and hard, in a systematic way even by
ensuring, to some extent, the stability margins and closed-loop performance as that of the
unconstrained control law. The control algorithm accomplishes this by incorporating the
constraints represented by a set of linear inequalities into the optimization problem described
by a cost function. From Figure 12, it is obvious that constrained MPC shows reasonable
performance over the whole prediction horizon while meeting constrains imposed on the
control as well as on the output signal. However, the unconstrained MPC offers even faster
startup response. This is due to the reason that unconstrained control algorithm needs not to
calculate the constraints at each sampling time k as does the constrained algorithm. Sometimes,
we do not come with well-posed optimization under the given constraints. The constraints may
be relaxed gradually, in this case, to find the optimal solution.

E. Set Point Tracking

The results point out that MPC can be used effectively to keep good setpoint tracking.
Figure 13 shows that MPC follows the changes in the reference voltage when changed from 2
V to 3 V and then from 3 V to 2 V with negligible overshoot in the startup whereas PWM-PID
controller shows somewhat increased overshoot in the startup as well as at the time of
transients. Thus MPC shows good output voltage tracking characteristics. NN-MPC also
demonstrates good tracking of the reference trajectory with low control energy consumption
(simulations results are not shown for saving space purposes).

3.5 3.5

3 3

2.5 2.5

2 2

1.5 Vout (V) 1.5 Vref (V)

Vref (V) Vout (V)
1 1

0.5 0.5

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Time (sec) -3 Time (sec) -3
x 10 x 10

(a) (b)

3.5

2.5

(c) 1.5
Vout (V)
1 Vref (V)

0.5

0
0 0.2 0.4 0.6 0.8 1
Time (sec) -3
x 10
Figure 13. Setpoint tracking offered by (a) MPC, (b) PWM-PID, and (c) NN-MPC

333
Constrained Model-Based Predictive Controller for a High-Frequency

F. Load Regulation
The MPC algorithm surely depends on the load resistance. In order to investigate the
dynamic response of the system, the load resistance is changed from 4.5 to 9 and then
from 9 to 4.5 (50% dynamic load change). The simulation results dictate that the output
voltage does not converge exactly to the steady-state value, showing the steady state error for
the step-up and step-down load changes. However robustification provided to the MPC by
adding an external loop ensures the steady-state error free tracking (see Figure 14).
2.15 2.15

2.1 2.1

2.05 2.05
Vout (V)

Vout (V)
2 2

1.95 1.95

1.9 1.9

1.85 1.85
1 2 3 4 5 1 2 3 4 5
-4 -4
x 10 x 10

0.5 0.5

0.4 0.4
iL (A)

iL (A)
0.3 0.3

0.2 0.2

0.1 0.1

0 0
1 2 3 4 5 1 2 3 4 5
Time (sec) -4
Time (sec) -4
x 10 x 10

(a) (b)
2.15
2.1
2.05
Vout (V)

2
1.95
1.9
1.85
1 2 3 4 5
-4
x 10
0.6
0.5
0.4
iL (A)

0.3
0.2
0.1
0
1 2 3 4 5
Time (sec) -4
x 10

(c)
Figure 14. Load regulation offered by (a) MPC, (b) PWM-PID, and
(c) MPC supported by PI controller.

G. Line Regulation
The effect of variation in input voltage on the output performance is investigated in Figure
15. It is learned that like load regulation, MPC shows poor line regulation compared to PWM-
PID controller. The line regulation is evaluated by making a step change in the input voltage to
4.6 V from its nominal value of 3.6 V and then from 4.6 V to 3.6 V during the steady-state

334
Ghulam Abbas, et al.

operation. However, Figure 15 (c) shows that MPC supported by PI controller shows good line
regulation comparable to that of PID.

2.2
2.3
2.1
2.2

V o u t (V )
Vout (V)
2.1 2
2
1.9
1.9
1.8
1.8 1 2 3 4 5
1 2 3 4 5
-4 -4
x 10 x 10
5 5

4.5 4.5

V in (V )
Vin (V)

4 4

3.5
3.5

3
3
1 2 3 4 5
1 2 3 4 5
Time (sec) -4 Time (sec) -4
x 10 x 10

(a) (b)
2.2
2.15
2.1
V out (V )

2.05
2
1.95
1.9
1.85
1.8
1 2 3 4 5
-4
x 10
5

4.5
V in (V )

3.5

3
1 2 3 4 5
Time (sec) -4
x 10

(c)
Figure 15. Line regulation offered by (a) MPC, (b) PWM-PID, and
(c) MPC supported by PI controller.

Simulation results highly recommend the model predictive control as an advanced control
technique to adequately achieve the excellent static and dynamic performance when applied to
single-input, single-output (SISO) buck converter system.

335
Constrained Model-Based Predictive Controller for a High-Frequency

5. Conclusions
In this paper, a model-based predictive control method, based on the general theory and on
the neural network model, has been proposed and investigated in a very comprehensive way.
The proposed model predictive control scheme successfully minimizes the cost function to
achieve better trade-off between fast response and low overshoot for start-up as compared to
the PWM-PID. It not only eliminates the maximum overshoot faster but also takes the least
time to settle to its steady state value. Effects of constraints, horizons both prediction and
control, and performance weights on the closed-loop performance are investigated both
theoretically and practically. Excellent load and line regulation have been achieved by
introducing an external loop which contains a PI controller.
Compared to the LQR and PWM-PID controllers, MPC handles the constrained problems
with complex MIMO structure in an attractive way and shows the well-controlled tuning
characteristics. MATLAB/Simulink based simulation results validate the effectiveness of the
design.
MPC comes with the disadvantage that the potentially intensive computation has to be
performed to calculate the sequence of a control signal at each sampling instant for better
setpoint tracking. However the computational issue is no longer relevant as todays computer
power is at our disposal.
Future work involves the experimental verification of the simulation results and
implementation and testing of the MPC. The performance of the different cost functions can
also be investigated for achieving good performance and robustness.

References
[1] Hebertt J. Sira-Ramirez and Ramn Silva-Ortigoza, Control Design Techniques in
Power Electronics Devices, 1st Edition, ISBN: 978-1846284588, Springer, 2006.
[2] A. J. Forsyth and S. V. Mollov, "Modeling and Control of DC-DC Converters", Power
Engineering Journal, Vol. 12, No. 5, pp. 229236, 1998.
[3] J-H. Su, J-J. Chen, and D-S. Wu, Learning Feedback Controller Design of Switching
Converters Via MATLAB/Simulink, IEEE Trans. on Education, Vol. 45, No. 4,
pp.307315, 2002.
[4] E. F. Camacho and C. Bordons, Model Predictive Control, 2nd Edition, Springer-
Verlag, 2004.
[5] K. Lian, J. Liou, and C. Huang, LMI-based integral fuzzy control of DC-DC
converters, IEEE Transactions on Fuzzy Systems, 2006, pp. 7180.
[6] J. Mahdavi, M. R. Nasiri, A. Agah, and A. Emadi, "Application of Neural Networks and
State-Space Averaging to DC/DC PWM Converters in Sliding-Mode Operation",
IEEE/ASME Trans.on Mechatronics, Vol.10, No.1, pp. 6067, 2005.
[7] G. Abbas, U. Farooq, and M. U. Asad, Application of Neural Network Based Model
Predictive Controller to Power Switching Converters, International Conference and
Workshop on Current Trends in Information Technology (CTIT), pp. 132136, 2627
Oct. 2011, Dubai.
[8] J. M. Carraso, J. M. Quero, F. P. Ridao, M. A. Perales, and L.G. Franquelo, Sliding
Mode Control of a DC/DC PWM Converter with PFC Implemented by Neural
Network, IEEE Trans. on Circuit Sys. 1: FTA, Vol. 44, No. 8, 1997, pp.743749.
[9] J. M. Maciejowski, Predictive Control with Constraints. Englewood Cliffs, NJ: Prentice-
Hall, 2002.
[10] M. Morari and J. H. Lee, Model predictive control: past, present and future, Comput.
Chem. Eng., vol. 23, no. 4/5, pp. 667682, May 1999.
[11] P. Corts, M. P. Kazmierkowski, R. M. Kennel, D. E. Quevedo, and J. Rodrguez,
Predictive control in power electronics and drives, IEEE Trans. Ind. Electron., vol.
55, no. 12, pp. 43124324, Dec. 2008.
[12] K. V. Ling, S. P. Yue, and J. M. Maciejowski, An FPGA implementation of model
predictive control, American Control Conference, pp. 19301935, Minneapolis, 2006.

336
Ghulam Abbas, et al.

[13] A. K. Abbes, F. Bouani, and M. Ksouri, A Microcontroller Implementation of

Constrained Model Predictive Control, World Academy of Science, Engineering and
Technology 80, 2011.
[14] A. Zheng, Robust Control of Systems Subject to Constraints, Ph.D. Thesis, California
Institute of Technonlogy, Pasadena, California, 1995.
[15] A. Rusnk, M. Fikar, K. Najim, and A. Mszros, Generalized predictive control based
on neural networks, Neural Process, Lett. 4 (1996), pp. 107112.
[16] S. Di Cairano and A. Bemporad, Model predictive control tuning by controller
matching, IEEE Trans. Automat. Contr., vol. 55, no. 1, pp. 185190, 2010.
[17] J. Liu, X. Chen, D. Muoz de la Pea, and P. D. Christofides, Sequential and iterative
architectures for distributed model predictive control of nonlinear process systems,
AIChE Journal, pp. 21372149, 2010.
[18] T. Geyer, G. Papafotiou, and M. Morari Model predictive control in power electronics:
A hybrid systems approach, Proc. CDC-ECC05, pp. 56065611, 2005.
[19] A. G. Beccuti, G. Papafotiou, M. Morari, S. Almer, H. Fujioka, A. Wernrud, Rantzer,
M. Baja, H. Cormerais, and J. Buisson, Hybrid control techniques for switched-mode
DC-DC converters Part II: The step-up topology, Proc. ACC07, pp. 54645471, 2007.
[20] C. F. W. Sanders, M. J. Hounslow, and F. J. Doyle III, Model Predictive Control of
Wet Granulation Using an Experimentally Validated Population Balance Model,
AIChE Annual Meeting (2006), San Francisco.
[21] K. S. Narendra and K. Parthasarathy, Identification and control of dynamical systems
using neural networks, IEEE Trans. on Neural Networks, vol. 1, no. 1, pp. 427, 1990.
[22] G. W. Ng, Application of Neural Networks to Adaptive Control of Nonlinear
Systems, Research Studies Press Ltd., Taunton, Somerset, England, 1997.
[23] O. Omidvar and D. L. Elliot (Eds.), Neural Systems for Control, Academic Press, San
Diego, 1997.
[24] M. T. Hagan and H. B. Demuth, Neural Networks for Control, Proceedings of the
1999 American Control Conference, San Diego, CA, 1999, pp. 16421656.
[25] A. Muthuramalingam, S. Himavathi, and E. Srinivasan, Neural Network
Implementation Using FPGA: Issues and Application, International Journal of
Information and Communication Engineering, vol. 4, no. 2, Winter 2008.
[26] K. O. Temeng, P. D. Schnelle, and T. J. McAvoy, Model predictive control of an
industrial packed bed reactor using neural networks, J. Process Control 5 (1995), pp.
1927.
[27] Alejandro Oliva and Simon Ang, Power-Switching Converters, 3rd Edition, Taylor &
Francis Publisher, 2010, ISBN: 143981533X, 9781439815335.
[28] R. D. Middlebrook and S. Cuk, A General Unified Approach to Modeling Switching-
Converter Power Stages, IEEE Power Electronics Specialists Conference (PESC),
1976, pp. 1834.
[29] B. Johansson, DC-DC Converters, Dynamic Model Design and Experimental
Verification, Dissertation LUTEDX/(TEIE-1042)/1-194/(2004), Dep. of Industrial
Electrical Engineering and Automation, Lund University, Lund, 2004.
[30] A. Bemporad, M. Morari, and N. L. Ricker, "Model Predictive Control Toolbox 3 User's
Guide", The MathWorks, Inc. 20052010.
[31] A. Bemporad, N. Ricker, and J. Owen, Model predictive controlNew tools for design
and evaluation, in American Control Conference, Boston, MA, 2004, pp. 56225627.
[32] T. Geyer, G. Papafotiou, and M. Morari, On the optimal control of switch-mode DC-
DC converters, in Hybrid Systems: Computation and Control, vol. 2993, pp. 342356,
Lecture Notes in Computer Science. Springer-Verlag, March 2004.
[33] A. Zheng and W. Zhang, Nonlinear Predictive Control: Theory and Practice, Chapter
8: Computationally Efficient Nonlinear Model Predictive Control Algorithm for Control
of Constrained Nonlinear Systems. IEE Control Engineering Series, 2001.

337
Constrained Model-Based Predictive Controller for a High-Frequency

[34] G. Pannocchia an nd E. C. Kerrigan, Offset-frree receding hoorizon control of constrainedd

linnear systems, AIChE
A J., vol. 51, no. 12, pp. 31343146, 2005.
2
[35] G. Pannocchia, Robust
modell predictive conntrol with guarranteed setpoinnt tracking, J.
Prrocess Control,, vol. 14, no. 8,, pp. 927937, Dec. 2004.
[36] T. Geyer, G. Pap pafotiou, and M.M Morari, Hybrid model predictive contrrol of the step--
doown dc-dc conv verter, Automm. Control Labb., ETH Zurichh, Zurich, Swittzerland, Tech.
Reep. AUT08-01,, 2008 [Online]]. Available: htttp://www.control.ee.ethz.ch..
[37] Maark Hudson Beale,
B Martin T. Hagan, annd Howard B. Demuth, Neeural Networkk
Tooolbox 7 Useers Guide, Thhe MathWorkss, Inc. 199220010.
[38] R. Rojas, Neuraal Networks, A Systematic Introduction,
I Springer-Verllag, New York,,
19996.
[39] D. Soloway and P. J. Haley, Neural
Generaalized Predictiive Control, Proceedings
P off
thee 1996 IEEE Innternational Syymposium on Inntelligent Conttrol, 1996, pp. 277281.
[40] Naatick, MA. Co ontrol System Toolbox
T 9 : Useers Guide", Thhe Math Workss, Inc.

Ghhulam Abbas did d his B.Sc. inn Electrical Enngineering fromm University off
Eng gineering and Technology (UET)
( Lahore and M.Sc. andd Ph.D. both inn
Eleectrical Engineeering from thee Institut Natioonal des Sciencces Appliquess
de Lyon (INSA Lyon) France. He is currenntly working as a an Assistantt
Proofessor at the Department
D off Electrical Enggineering, Thee University off
Lahhore Pakistan. He has published a number of papers inn various IEEE E
connferences and international journals. His research
r interessts include thee
moodern and intellligent control of
o power switching converterrs.

Ummar Farooq did d his B.Sc. and M.Sc. booth in Electricaal Engineeringg
fro
om University of Enginering and Technoloogy Lahore, Paakistan in 20044
an
nd 2011 respeectively. He has publishedd a number of papers inn
intternational connferences and hash over eight years of teachhing exerience.
Hee is currently working
w as Asssisant Professoor at Departmennt of Electricall
Enngineering, Unnivesity of Thhe Punjab Lahhore. His reseearch interestss
incclude intelligennt control of noonlinear system
ms.

ason Gu receivved his bacheloors degree from

Ja m Electrical Enngineering andd
In
nformation Science at the University
U of Science
S and Technology
T off
Chhina in 1992, his Masters degree from Biomedical Engineering
E att
Shhanghai Jiaotonng University in 1995, and PhDP degree frrom Universityy
off Alberta in Caanada in 2001. He is currentlly professor in Electrical andd
Coomputer Enginneering at Dalhhousie University in Canadaa. He is also a
crooss-appointed professor in School of Biomedical
B Enngineering andd
Faaculty of Compputer Science for
f his multidissciplinary research work. Dr.
Gus reseearch areas incllude robotics, biomedical
b enggineering, rehaabilitation enginneering, neurall
networks,, and control. HeH has over seeventeen yearss research and teaching expeerience and hass
publishedd over 170 confference papers and articles. Twenty
T one maasters, five PhD D and four postt
doctoral fellows have completed thheir training under u his suppervision. He has been thee
Organizattion committeee for numerouss IEEE confereences. He has been the assocciate editor forr
followingg journals: Jou urnal of Conttrol and Intellligent System ms; Transactionns on CSME,,

338
Ghulam Abbas, et al.

Canada; IEEE Transaaction on Mechatronics;

M International Journal of Robotics andd
Automation; Unmanned
d Systems; Jouurnal of Engineeering and Emeerging Technoologies IEEE
E
Access.

Mu uhammad Usm man Asad didd his B.Sc. in Electrical Enggineering fromm
University of the Punjab Lahorre and is workking towards M.Sc.
M degree inn
Eleectrical Engineering from G. C. University Lahore.
L He waas the reciepentt
of gold and silveer medal awarrds for his pappers in 24th annd 26th IEEEP P
Inteernational Mullti topic Sysmmposium held ata Bahria Univversity Karachii
(20
009) and Pakisstan Navy Enggineering Colllege, National University off
Sciience and Techhnology (2011)) respectively. He is currentlyy working as a
Leccturer at Department of Electrical Enginneering, The University off
Lahore. His
H research interest
i includde the intelligent control off mobile roboots and powerr
systems.

339