Received April 16, 2018, accepted May 17, 2018, date of publication May 21, 2018, date of current

version June 29, 2018.

Digital Object Identifier 10.1109/ACCESS.2018.2839519

Machine Learning Based Adaptive Prediction

Horizon in Finite Control Set
Model Predictive Control
MUHAMMAD SALEH MURTAZA GARDEZI AND AMMAR HASAN
School of Electrical Engineering and Computer Science, National University of Sciences and Technology, Islamabad 44000, Pakistan
Corresponding author: Muhammad Saleh Murtaza Gardezi (mgardezi.msee15seecs@seecs.edu.pk)

ABSTRACT In this paper, an adaptive prediction horizon approach based on machine learning is
presented for the finite control set model predictive control (FCS-MPC) of power converters. Usually,
in FCS-MPC, the prediction horizon is kept constant. A large prediction horizon improves performance,
however, it significantly increases the computational cost. The prediction horizon is typically chosen to
be just large enough to give the required performance. We present a novel technique, where the prediction
horizon adapts to the states of the converter. We define a cyber-physical objective function that penalizes both
the error in converter performance and computational complexity. We perform several offline simulations
to find the optimal prediction horizon based on the instantaneous state of the converter, based on a cyber-
physical objective function. An artificial neural network is trained to calculate the optimal prediction horizon
in run time. The proposed scheme allows a varying prediction horizon that reduces the overall computational
complexity, while guaranteeing the required physical performance. The simulations and experimental results
of the proposed technique justify the usefulness of our approach.

INDEX TERMS Adaptive prediction horizon, finite control set model predictive control (FCS-MPC),
DC-DC boost converter.

I. INTRODUCTION
In recent years, the applications of DC-DC converters have
widened, ranging from DC power supplies, DC motor con-
trol, electric vehicles, hybrid electrical systems and renew-
able energy systems [1]. With the DC-DC converters comes
the need to control the converters to desired conditions. In this
aspect, linear control techniques like pulse-width-modulation
and PI have been applied on power converters [2]. However,
these control techniques pose problems such as extensive
manual tuning.
Ever since the dawn of fast processors, researchers have
been able to come up with newer control techniques that are
superior to classical control. Embedded systems and digital
processors have enabled the use of techniques like feed-
forward control [3], [4], fuzzy logic [5], [6], nonlinear tech-
niques [7], [8], sliding mode [9], [10] and H∞ [11]. One of the
control method that takes advantage of the processing power
is the model predictive control (MPC) [12].
Model predictive control (MPC) uses a discrete-time
model of the system to predict the future values of the system
based on the current state. In power electronic converters,
the control input is the on/off actuation of the switch. This
entails that the possible inputs are finite in number. The
variant of MPC that manipulates the switch position directly
is called the finite control set model predictive control
(FCS-MPC). The finite number of possible inputs produce
a finite number of possible switching sequences. The number
of prediction steps in the horizon, N , dictates the length of
the switching sequences and thus indicates how far ahead
the states are predicted. A cost function is defined and
FCS-MPC solves an optimization problem to find the optimal
input sequence. The optimization problem is solved online at
every sampling instant and the first actuation of the sequence
is then applied to the converter. This process is repeated at
every time step for new state values, which is known as the
receding horizon policy [13].
The main disadvantage of the MPC is the large computa-
tional cost associated. The time intervals in power electronic
converters are in the microseconds. The optimization prob-
lem can become too computationally intensive to solve in
real-time. The prediction horizon is an important factor in
the computational complexity of MPC. A larger prediction

2169-3536
2018 IEEE. Translations and content mining are permitted for academic research only.
32392 Personal use is also permitted, but republication/redistribution requires IEEE permission. VOLUME 6, 2018
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
M. S. M. Gardezi, A. Hasan: Machine Learning-Based Adaptive Prediction Horizon in FCS-MPC
horizon increases the performance of the system [14] but
at the same time, increases the computational cost. This is
due to the reason that the total number of sequences to be
searched increases exponentially with the prediction horizon.
However, the prediction horizon cannot be made too small as
it can affect both performance and stability, especially in case
of power electronic converters that exhibit a non-minimum
phase behavior.
Move-blocking scheme [15], [16] was introduced in MPC
by breaking the horizon into fine and coarse time steps.
For the immediate future, fine time intervals are used, while
coarse time intervals are used for the future far ahead. This
reduces the total number of time steps in the horizon, which
reduces the computational complexity.
Another approach is to reduce the computational cost by
efficiently solving the optimization problem. Techniques like
sphere decoding [17] were introduced in the realm of MPC
to replace the exhaustive search by solving the optimization
problem in a vector form. This converts the optimization
problem to an integer least square problem. Thus, the opti-
mal switching position is found in a more cost-effective
manner. However, this technique is limited to only a few
topologies of power converters. Other efficient solutions
to FCS-MPC involve techniques like event-based horizon
and extrapolation strategy [18]. However, all existing tech-
niques have a fixed number of prediction steps, N , in the
horizon.
In this paper, a novel technique is proposed where the
prediction horizon, N , is varying in an effort to further reduce
the computational cost of MPC. A large value of N results
in improved converter performance as a trade-off to higher
computational complexity. In case of varying N , we could
reduce its value in runtime to reduce computational complex-
ity. To find an optimal value of N , we can define a cost that
penalizes both the physical and cyber performance. A trade-
off between the physical performance and the cyber perfor-
mance (computational cost) could be achieved similar to a
cyber-physical system (CPS) in [19]–[21]. A CPS approach
to the attitude control of the CubeSat satellite is presented
in [21]. Due to the limitation of resources in the satellite,
the computational power is shared by all tasks being carried
out. The controller regulates both the physical and the cyber
resources of the satellite. Similarly, a cyber-physical cost
function (Jcp ) can be defined for an FCS-MPC controlled con-
verter that contains information of both the cyber utilization
and the physical performance of the converter.
However, solving the optimization with the cyber-physical
cost function will increase the overall computational burden
as it will have to find both the optimal input and optimal pre-
diction horizon. To avoid this, an artificial neural network is
trained to find the optimal N at run-time. Several simulations
are run with different values of converter states for a range of
N and the resulting cyber-physical cost, (Jcp ) is determined.
The value of N that gives the lowest cost is considered opti-
mal. A data set of these states and the corresponding optimal
N is created. An artificial neural network is then trained on
VOLUME 6, 2018
the data set to learn the behavior and predict the optimal N at
run-time.
The proposed method is applied on a boost converter. The
main objective of the controller is to regulate the output
voltage to a set reference. A discrete-time model for the
boost converter is used that incorporates both the continuous
conduction mode (CCM) and the discontinuous conduction
mode (DCM). A discrete Kalman filter [22] is included with
the control scheme to cater for the changes in load and provide
an offset-free voltage regulation.
The organization of the rest of the paper is as follows.
Section II briefly reviews the discrete-time model for the
boost converter which is suitable for both the CCM and
the DCM operations. The working of the FCS-MPC scheme
with Kalman filter is described in Section III. The proposed
method for determining an adaptive prediction horizon is
explained in Section IV. The simulation results and experi-
mental results to illustrate the performance and benefits of
the proposed control scheme are shown in Section V and VI,
respectively. In Section VII, the conclusions of the paper are
deduced.
II. MODEL OF THE BOOST CONVERTER
The discrete-time model of the DC-DC boost converter pre-
sented in [16] is used. The model is represented in the state
space form as:
x(k + 1) = EM x(k) + FM vs(k) (1a)
y(k) = Gx(k) (1b)
where x(k) = [il (k) vo (k)] is the state vector containing the
inductor current, il and the output voltage, vo . Vs is the supply
voltage of the converter. EM and FM are the matrix of the state
space that are dependent on the sampling interval Ts . k is used
to denote the current time step. M are the various modes of
operation of the boost converter and can take the value from
the set {1, 2, 3, 4} depending on the switch state {0, 1} and the
inductor current i.e. either zero or non-zero [16]. The output
of the model is denoted by y(k). The matrix G is the output
matrix of the system. Further details of the model of the boost
converter can be found in [16].
III. FINITE CONTROL SET MODEL PREDICTIVE CONTROL
A. CONTROL SCHEME
The MPC scheme proffered in [16] uses an enumeration tech-
nique that determines the actuation of the switch to control the
output voltage. The objective function is:
X−1
k+N
J= |vref − vo (l)| + λ|u(l − 1) − u(l)| (2)
l=k
where N is the prediction horizon over which the variables
of interest are penalized, vref is the reference voltage, and
u represents the switch position being either on or off.
The second term of the objective function is used to prevent
excessive switching and reduce the switching frequency.
32393
 M. S. M. Gardezi, A. Hasan: Machine Learning-Based Adaptive Prediction Horizon in FCS-MPC
The FCS-MPC determines the optimal sequence among all
possible switching sequences over the prediction horizon, N .
In enumeration, the objective function cost, J is calculated for
all these sequences and the sequence that returns the lowest
cost is the optimal sequence. The first entry of this optimal
sequence is the optimal switch position, which is then applied
to the converter. This whole process is carried out within the
sampling interval Ts and is repeated at every step according
to the newest values of the states of the model.
B. SIGNIFICANCE OF LARGE N
The boost converter exhibits a non-minimum phase behav-
ior that entails an initial drop in the output voltage when
the reference voltage is increased. This behavior requires a
prediction horizon large enough to see the eventual rise of
output voltage after it drops. The move-blocking strategy [16]
was devised for a seemingly large N with the computational
cost of a smaller prediction horizon. Furthermore, a large
prediction horizon also improves the physical performance
of the converter [14] in the transient state where it helps to
reach the reference voltage in a shorter time.
C. CHANGES IN LOAD
The load is often time-varying and unknown in most appli-
cations. Thus, a Kalman filter [16], [23] is added to the
control scheme to provide offset-free tracking of the reference
voltage during changes in load. In addition to the measured
states, il and vo , disturbance states, ie and ve are added to form
an augmented state vector.

xa = il vo ie ve
T
(3)
Therefore, the state space model changes to
xa (k + 1) = EaM x(k) + FaM vs (k) (4a)
ya (k) = Ga x(k) (4b)
 
  FM
E 0
EaM = M , FaM =  0  and
 
Ga = 1 1
0 1
0
(5)
where 1 and 0 are identity matrix and a null matrix of order
two respectively.
A discrete Kalman filter is designed with four different
Kalman gains, KM , according to the mode of operation. The
noise covariance matrices, Q and R, are assigned such that
high credibility is associated with the measured states. Thus,
the estimated state equation is defined as
x̂(k + 1) = EaM x̂a (k) + KM Ga (x̂a (k))FaM vs (k) (6)
The estimated disturbance state is used to eliminate the offset
from the reference by altering the reference voltage to
v̂ref = vo − v̂e (7)
Instead of the measured states, il and vo , the estimated states,
îl and v̂o , are used as input vector to the controller.
32394
IV. PROPOSED ADAPTIVE PREDICTION HORIZON
In the standard FCS-MPC scheme [16], the prediction hori-
zon is kept constant. An increase in prediction horizon
improves the physical performance but also increases the
computational complexity. The purpose of the proposed
method is to find a prediction horizon that provides a trade-
off between physical performance and low computational
complexity. We propose to have a varying prediction hori-
zon that adapts itself to the changing values of il , vo and
vo,err during disturbances in load, changes in output reference
voltage, transients and steady-state. The proposed method of
determining the adaptive prediction horizon is explained in
the following subsections.
A. CYBER-PHYSICAL COST FUNCTION
A cyber-physical cost function Jcp is defined as a metric to
determine the effect of different values of prediction steps in
the horizon, N , on the physical performance of the converter
and the cyber resource utilization. It is the sum of the physical
performance, Jp , and the weighted cyber utilization, Jc . The
physical performance cost is defined based on (2), which is
given below:
M
X
|vref − vo (l)| + λ|u(l − 1) − u(l)|


Jp := (8)
l=1
where λ is a weighting factor similar to equation (2) and
M is the number of samples to assess the physical perfor-
mance. M should be set sufficiently greater than the range of
values of prediction horizon to be used adaptively. Jp is part
of Jcp , which we use to identify how well the controller
behaves for different values of N . To compare like with like,
we have to choose a value of M that is greater than the largest
value of N.
It is well established in the MPC literature that the com-
putational complexity of MPC grows exponentially with the
prediction horizon [12]–[14]. In case of FCS-MPC the com-
putational complexity is proportional to 2N [15], [16]. There-
fore, we define the cyber utilization as
M
1 X N (l)
Jc := 2 . (9)
M
l=1
If a fixed prediction horizon is used, then Jc will simplify
to 2N .
We define the cyber-physical cost function as the weighted
sum of Jp and Jc
Jcp := Jp + γ Jc , (10)
where γ is a weighting factor between the physical cost and
the cyber cost. The choice of the value of γ is application
specific. It depends on the relative importance of the cyber
utilization as compared to the physical performance.
It can be seen from (8) that Jp is an accumulation of the
error for M time steps which captures the physical perfor-
mance of the controller. A larger value of N results in an
improved performance [15], [16]. Therefore, a larger N will
VOLUME 6, 2018
M. S. M. Gardezi, A. Hasan: Machine Learning-Based Adaptive Prediction Horizon in FCS-MPC

result in a smaller value of Jp . On the other hand, the value

of Jc increases with N . Thus, a trade off has to be found
between the physical performance and the cyber utilization.
The optimal prediction horizon, No , is the one that minimizes
the cost Jcp over N . We propose to find the optimal No by
using an artificial neural network.

B. STRUCTURE OF THE NEURAL NETWORK

An artificial neural network is a supervised machine learn-
ing technique that is analogous to function mapping. It is
inspired by the biological neural network where several FIGURE 2. Block diagram of FCS-MPC scheme with adaptive prediction
nodes combine themselves in a network to relate several horizon and Kalman filter.
features or inputs to an output. In the proposed case, the inputs
of the neural network are the states of the converter, il and The generation of training data set and the training of the
vo , and the error between the reference voltage and output neural network is done offline. Whereas, at every sampling
voltage, vo,err . The output of the neural network is the opti- interval, the function adaptiveN is computed online to find
mal prediction horizon, No , corresponding to the inputs. The the optimal prediction horizon, No . Further details on the
neural network forms a function, adaptiveN , as follows: generation of data set and training of the neural network are
provided in Section V.
No = adaptiveN (il , vo , vo,err ). (11)
C. FCS-MPC WITH ADAPTIVE PREDICTION HORIZON
A one hidden layer structure for the neural network has The control scheme proposed is summarized in Algorithm 1.
been used as shown in Fig. 1. We have used a feed forward Instead of fixed prediction horizon, the algorithm adapts
network with a sigmoid activation function. Back propaga- the horizon depending on il , vo and vo,err and the optimal
tion is utilized to find the relation between the inputs and prediction horizon, No , is determined by (11). We have used
the outputs; while Bayesian regularization is employed to move blocking in the algorithm [16]. The function f1 denotes
prevent over-fitting and provide a smooth function that maps a state model with fine sampling time, Ts , while f2 denotes a
the inputs to the outputs. The number of hidden units in state model with coarse sampling time. The block diagram of
the hidden layer should be specified large enough so that the proposed control method is shown in Fig. 2.
inaccuracies can be avoided in the output of the function,
adaptiveN . While training the neural network we found that Algorithm 1 FCS-MPC With Adaptive Prediction Horizon
20 or more units in the hidden layer give us a sufficient function MPC(il , vo , vo,err , x̂(k), u(k − 1))
accuracy. Therefore, we found it sufficient to keep 50 units in J ∗ (k) = ∞; u∗ (k) = ∅; x(k) = x̂(k)
the hidden layer of the neural network. The details of training No = adaptiveN (il , vo , vo,err )
are given in Section V. for all U over No do
J =0
for l = k to k + No + 1 do
if l < k + N1 then
x(l + 1) = f1 (x(l), u(l))
else
x(l + 1) = f2 (x(l), u(l))
end if
verr (l + 1) = v̂ref − vo (l + 1)
J = J + |verr (l + 1)| + λ|u(l) − u(l − 1)|
end for
if J < J ∗ (k) then
J ∗ (k) = J
u∗ (k) = U (1)
end if
end for
end function

V. SIMULATION RESULTS
The simulation results of the proposed control scheme
FIGURE 1. A representation of the artificial neural network. applied to the boost converter are presented in this section.

VOLUME 6, 2018 32395

 M. S. M. Gardezi, A. Hasan: Machine Learning-Based Adaptive Prediction Horizon in FCS-MPC

These results demonstrate that the physical performance of of the converter. We have tested training the neural network
the converter has not been compromised by adapting the for different sizes of data sets. For data sets of size 600 and
prediction horizon, however, there is a significant reduction above we do not see a profound effect on the accuracy of
in the computational cost. The performance of the proposed the training of the neural network or the performance of the
closed loop system to changes in reference voltage, input complete system. Therefore, we found it sufficient to have
voltage and load disturbances are shown and compared with 1330 points in the data set.
results for fixed prediction horizon.
The circuit parameters are Co = 220µF, L = 625 µH B. START-UP
and Rl = 1.2 . The input voltage vs = 12 V and the The start-up behavior of the converter under nominal con-
output reference voltage vref = 18 V. The load resistance is ditions is shown in Fig. 3. The reference voltage is set
R = 82 . as 18 V. It may be noticed that during the initial charging
The weighting factor in the objective function (2) is of the inductor, the prediction horizon is adapted to smaller
λ = 0.1, the sampling time interval is Ts = 5 µs. For the values because a larger prediction horizon would provide no
move-blocking strategy, the parameters are: number of fine benefit in the performance, hence, reducing the computa-
samples, N1 = 6, and the relative duration of coarse samples, tional cost (cyber utilization). However, when the inductor
ns = 4. As for the Kalman filter, the covariance matrices is charged, the prediction horizon adapts to the maximum
Q and R are set as diag([0.1, 0.1, 50, 50]) and diag([1, 1]), value at which the neural network was trained, to provide
respectively. the best possible physical performance. Eventually, when
the converter achieves steady-state, the optimal prediction
A. TRAINING THE NEURAL NETWORK horizon is adapted to a smaller value again.
For the generation of training data set, a range of values for
vo,err , il , vo and N are chosen. For every possible combination
of vo,err , il , vo and N , simulations are run and the corre-
sponding Jcp is determined. For every possible combination
of values of il , vo and vo,err , the corresponding N that gives
the lowest Jcp , is recorded as the optimal No . A training data
set is formed using this strategy where the output, the optimal
number of prediction steps in the horizon, No , is related to
every combination of the inputs, il , vo and vo,err .
For the collection of the data for training the neural net-
work, the have used the following range of values: vref =
[18, 24], vo = [0, 2, 4, 6, 8, 10, 12, 14, 15, 16, 17, 17.5,
17.9, 17.95, 17.97, 17.98, 17.99, 18, 18.01, 18.02, 18.05,
18.1, 18.11, 18.12, 18.5, 19, 21, 23, 23.5, 23.9, 23.95, 23.97,
23.98, 23.99, 24], il = [0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7,
0.8, 1, 1.3, 1.6, 1.8, 2, 2.2, 2.4, 3, 4, 5], and N = [8, 9, 10,
11, 12, 13]. The combination of these values creates a data
set that has 1330 samples. The range of values are chosen
which cover the whole range of operation of the converter.
The values of M = 100 and γ = 1/213 were used. MATLAB
neural network toolbox is used to train the artificial neural
network based on the training data set.
Due to excessive computational requirements, the genera-
tion of training data could be an issue in some applications.
However, the amount of training data required for successful
training of a neural network depends on the complexity of FIGURE 3. Simulation results of the start-up behavior for adaptive
the network structure, the number of inputs, and number of prediction horizon and fixed prediction horizon.
outputs [24]. In our case, we have a shallow structure with
only one hidden layer. As far as the inputs and outputs of A simulated start-up behavior for a fixed prediction hori-
the neural network are concerned, we have a single scalar zon of 8 and 13 is also shown. Comparing the behavior
output and three scalar inputs. Another important aspect in of a fixed horizon of 13 and adaptive prediction horizon,
successful training of the neural network is the data set should it is noted that the time taken to track the output reference
have a variety of combinations [24]. In applications of power voltage is same. This entails that the physical performance of
converters the nominal range of output voltages and induc- the converter is not compromised by adapting the prediction
tor currents are usually known. We use this information to horizon while the computational cost is significantly reduced.
generate a data set that covers the entire range of operation The ripple in the output voltage for adaptive N , fixed N = 13,

32396 VOLUME 6, 2018

M. S. M. Gardezi, A. Hasan: Machine Learning-Based Adaptive Prediction Horizon in FCS-MPC

and fixed N = 8 is 0.013 V, 0.012 V, and 0.014 V,

respectively.

C. STEP CHANGES IN THE OUTPUT REFERENCE VOLTAGE

A step-down change in the output reference voltage from
24 V to 18 V was provided as shown in Fig. 4. At the time
of transition, a large prediction horizon is desirable to which
the prediction horizon adapts and quickly settles for a smaller
prediction horizon. This is due to the reason that in step-
down changes, the output capacitor voltage is dropping off
and a large prediction horizon offers no benefit. The inductor
current also drops to zero as the capacitor looses its charge.
During the time interval where the capacitor is discharging
towards the new reference voltage, the minimum trained N
is opted. A large prediction horizon offers no advantage and
a smaller prediction horizon conserves the cyber resources.
Once the desired voltage is reached, the prediction horizon
adapts to the new states. The physical performance matches
that of a fixed prediction horizon of 13 but the computational
cost is reduced significantly. The MPC with fixed horizon
of 8 fails to converge and is unstable.

FIGURE 5. Simulation results of the step up behavior for adaptive

prediction horizon and fixed prediction horizon.

once the steady-state is reached. The physical performance

matches that of a fixed prediction horizon of 13 but the
computational cost is reduced significantly.

D. LOAD STEP CHANGE

FIGURE 4. Simulation results of the step down behavior for adaptive
prediction horizon and fixed prediction horizon.
The behavior of the output voltage for a step-up change
in the output reference voltage is shown in Fig. 5. A step-
up change from 18 V to 24 V was applied to the converter.
At this transition, due to the non-minimum phase behavior
of the boost converter, a large prediction horizon is required
and thus, the prediction horizon adapts and increases to its
maximum trained value before falling down to a smaller value
For the last case, a change in load is simulated. For the
converter operating at an input voltage of vs = 12 V and
the output reference voltage set at vref = 18 V, the load is
changed from R = 82 to R = 41 . The behavior of the
output voltage and the prediction horizon is shown in Fig. 6.
Initially, there is a slight drop in the output voltage during
which time the Kalman filter adjusts the reference output
voltage. Once the reference voltage is adjusted to a new value,
the output voltage starts creeping up to the desired value. The
adjustment in the output reference voltage provides an offset
free tracking i.e. no steady-state error. The prediction horizon
adapts itself to the changing states of output voltage and the
error between the output voltage and the reference voltage
without compromising the physical performance.
E. QUANTITATIVE COMPARISON
The simulation results of inductor current and capacitor volt-
age in Figures 3 to 6 illustrate that the physical performance
for the proposed adaptive horizon, No , and the fixed horizon,
N = 13, is quite similar. However, the results for fixed N = 8
have a markedly poor physical performance. To provide fur-
ther analysis of the simulation results, we quantitatively com-
pare both the physical performance and the cyber resource

VOLUME 6, 2018 32397

 M. S. M. Gardezi, A. Hasan: Machine Learning-Based Adaptive Prediction Horizon in FCS-MPC
TABLE 1. Quantitative comparison of physical performance and cyber resource utilization for No and N = 13.
FIGURE 6. Simulation results of the load change behavior for adaptive
prediction horizon and fixed prediction horizon.
utilization for No and N = 13. The comparison is given
in Table 1. For both adaptive No and fixed N = 13, Table 1
enlists the physical cost Jp and the cyber cost Jc for each
simulation in Figures 3 to 6. The table also includes the ratios
of physical costs and the cyber costs.
The ratio of physical costs i.e. Jp for No and Jp for N = 13,
shows that the physical performance is similar. Whereas,
the ratio of cyber costs i.e. Jc for No and Jc for N = 13,
shows that the cyber cost for the proposed adaptive prediction
horizon is considerably less than that of fixed N = 13. For
the case of voltage step down, Jc for No is only 5.9% of Jc
for N = 13. In the simulations, the highest cyber resource
utilization by the proposed algorithm is for the case of voltage
step up, when Jc for No is only 43.3% of Jc for N = 13.
As mentioned in Section IV.A, the cyber cost function Jc in
equation (9) reduces to 2N for the case of fixed N. Therefore,
the binary log of Jc is equal to N for fixed values of N.
In Table 1, we also also enlisted the binary log of Jc . In case
of fixed N = 13, the log2 Jc = 13 as expected. However, for
32398
FIGURE 7. Experimental results of the start-up behavior for adaptive
prediction horizon.
the proposed adaptive horizon No , the binary log of Jc shows
that for which value of fixed N the cyber resource utilization
would have been the same. For voltage step down with No the
value of log2 Jc = 8.91. Hence, the cyber resource utilization
for the proposed adaptive prediction horizon is even less than
that of fixed N = 9, while giving physical performance
similar to fixed N = 13.
VI. EXPERIMENTAL RESULTS
To further validate the results, a hardware was implemented
with a 32-bit microprocessor. A boost converter was built
using an IRF40 MOSFET as the active switch. The inductor,
capacitor and load were selected of the same values as men-
tioned in Section V. The input voltage and reference voltage
have the same values as in the simulation. However, due to
the computational limitations of the microprocessor, the pre-
diction horizon was reduced and the sampling interval was
increased. The maximum length of the steps in the prediction
horizon were taken as N = 7 while the sampling interval was
Ts = 10 µs. Move blocking was employed with number of
fine samples N1 = 3, number of coarse samples, N2 = 4, and
the relative duration of coarse samples, ns = 4.
Simulations were run again for the new training data
where the range of prediction horizon was changed to
N = [3, 4, 5, 6, 7]. The remaining values for inductor cur-
rent, il, output voltage, vo and the reference output voltage,
VOLUME 6, 2018
M. S. M. Gardezi, A. Hasan: Machine Learning-Based Adaptive Prediction Horizon in FCS-MPC
FIGURE 8. Experimental results of the step down behavior for adaptive
prediction horizon.
FIGURE 9. Experimental results of the step up behavior for adaptive
prediction horizon.
vo,err were kept the same as in the previous section. A new
training data set was formed and a neural network was trained
on the new data set to get a function for the experimen-
tal results. The experimental results were captured using a
Yokogawa DL9140 oscilloscope and have been plotted using
MATLAB.
A. START-UP
The start-up behavior of inductor current and output volt-
age under nominal conditions for a reference output voltage
of 18 V is shown in Fig. 7. Similar to the observation in the
simulation results, the prediction horizon adapts to small val-
ues during the initial stages. The prediction horizon increases
VOLUME 6, 2018
as the inductor is charged and until the reference output
voltage is attained. However, at steady state, the prediction
horizon settles at a smaller value.
B. STEP CHANGES IN THE OUTPUT REFERENCE VOLTAGES
A step-down change in the reference output voltage was
applied to observe the behavior of inductor current and output
voltage in case of adaptive prediction horizon. During the step
down change from 24 V to 18 V, the prediction horizon adapts
to smaller values when the capacitor is discharging as shown
in Fig. 8.
Similarly, a step-up change in the reference output voltage
from 18 V to 24 V causes the prediction horizon to adapt to
a large value during transients as shown in Fig. 9. Once the
steady state is reached, the prediction horizon falls back to a
smaller value.
VII. CONCLUSION
In this paper, we have proposed a varying prediction horizon
in FCS-MPC of power converters. The prediction horizon
is chosen according to a cyber-physical cost that allows to
reduce the computational cost as a trade off to physical perfor-
mance. The proposed technique has been applied to the boost
converter. It has been observed that during transients and dis-
turbances in the system, a large prediction horizon, N , is used
by the proposed approach, whereas, during the steady state
and the initial charging of inductor in the boost converter,
a smaller prediction horizon is used that is sufficient for
the required physical performance. Since the computational
complexity is an exponential function of N , the reduction of
N during steady state leads to significant reduction in the
overall computational complexity.
ACKNOWLEDGMENT
This research work has not been presented at any confer-
ence or submitted for publication elsewhere and is the original
work of the authors.
REFERENCES
[1] N. Mohan, T. M. Undeland, and W. P. Robbins, Power Electronics: Con-
verters, Applications, and Design, 3rd ed. Hoboken, NJ, USA: Wiley,
2003.
[2] J. Alvarez-Ramirez, I. Cervantes, G. Espinosa-Perez, P. Maya, and
A. Morales, ‘‘A stable design of PI control for DC–DC converters with an
RHS zero,’’ IEEE Trans. Circuits Syst. I, Fundam. Theory Appl., vol. 48,
no. 1, pp. 103–106, Jan. 2001.
[3] M. K. Kazimierczuk and A. Massarini, ‘‘Feedforward control of DC-
DC PWM boost converter,’’ IEEE Trans. Circuits Syst., vol. 44, no. 2,
pp. 143–149, Feb. 1997.
[4] M. K. Kazimierczuk and L. A. Starman, ‘‘Dynamic performance of
PWM DC-DC boost converter with input voltage feedforward control,’’
IEEE Trans. Circuits Syst. I, Fundam. Theory Appl., vol. 46, no. 12,
pp. 1473–1481, Dec. 1999.
[5] T. Gupta, R. R. Boudreaux, R. M. Nelms, and J. Y. Hung, ‘‘Implementation
of a fuzzy controller for DC-DC converters using an inexpensive 8-b
microcontroller,’’ IEEE Trans. Ind. Electron., vol. 44, no. 5, pp. 661–669,
Oct. 1997.
[6] P. Mattavelli, L. Rossetto, G. Spiazzi, and P. Tenti, ‘‘General-purpose fuzzy
controller for DC-DC converters,’’ IEEE Trans. Power Electron., vol. 12,
no. 1, pp. 79–86, Jan. 1997.
32399
 M. S. M. Gardezi, A. Hasan: Machine Learning-Based Adaptive Prediction Horizon in FCS-MPC
[7] H. J. Sira-Ramirez, ‘‘Nonlinear P-I controller design for switchmode DC-
to-DC power converters,’’ IEEE Trans. Circuits Syst., vol. 38, no. 4,
pp. 410–417, Apr. 1991.
[8] S. Hiti and D. Borojevic, ‘‘Robust nonlinear control for boost converter,’’
IEEE Trans. Power Electron., vol. 10, no. 6, pp. 651–658, Nov. 1995.
[9] P. Mattavelli, L. Rossetto, and G. Spiazzi, ‘‘Small-signal analysis of DC-
DC converters with sliding mode control,’’ IEEE Trans. Power Electron.,
vol. 12, no. 1, pp. 96–102, Jan. 1997.
[10] S.-C. Tan, Y. M. Lai, and C. K. Tse, ‘‘General design issues of sliding-mode
controllers in DC–DC Converters,’’ IEEE Trans. Ind. Electron., vol. 55,
no. 3, pp. 1160–1174, Mar. 2008.
[11] A. Kugi and K. Schlacher, ‘‘Nonlinear H∞ controller design for a DC-to-
DC power converter,’’ IEEE Trans. Control Syst. Technol., vol. 7, no. 2,
pp. 230–237, Mar. 1999.
[12] S. Kouro, P. Cortes, R. Vargas, U. Ammann, and J. Rodriguez, ‘‘Model pre-
dictive control—A simple and powerful method to control power convert-
ers,’’ IEEE Trans. Ind. Electron., vol. 56, no. 6, pp. 1826–1838, Jun. 2009.
[13] J. B. Rawlings and D. Q. Mayne, Model Predictive Control Theory and
Design. Madison, WI, USA: Nob Hill, 2009.
[14] T. Geyer, P. Karamanakos, and R. Kennel, ‘‘On the benefit of long-horizon
direct model predictive control for drives with LC filters,’’ in Proc. IEEE
Energy Convers. Congr. Exposit. (ECCE), Sep. 2014, pp. 3520–3527.
[15] R. Cagienard, P. Grieder, E. C. Kerrigan, and M. Morari, ‘‘Move blocking
strategies in receding horizon control,’’ J. Process Control, vol. 17, no. 6,
pp. 563–570, 2007.
[16] P. Karamanakos, T. Geyer, and S. Manias, ‘‘Direct voltage control of DC–
DC boost converters using enumeration-based model predictive control,’’
IEEE Trans. Power Electron., vol. 29, no. 2, pp. 968–978, Feb. 2014.
[17] T. Geyer and D. E. Quevedo, ‘‘Multistep finite control set model predictive
control for power electronics,’’ IEEE Trans. Power Electron., vol. 29,
no. 12, pp. 6836–6846, Dec. 2014.
[18] P. Karamanakos, T. Geyer, N. Oikonomou, F. D. Kieferndorf, and
S. Manias, ‘‘Direct model predictive control: A review of strategies that
achieve long prediction intervals for power electronics,’’ IEEE Ind. Elec-
tron. Mag., vol. 8, no. 1, pp. 32–43, Mar. 2014.
[19] J. Sztipanovits et al., ‘‘Toward a science of cyber–physical system integra-
tion,’’ Proc. IEEE, vol. 100, no. 1, pp. 29–44, Jan. 2012.
[20] J. M. Bradley and E. M. Atkins, ‘‘Toward continuous state–space regu-
lation of coupled cyber–physical systems,’’ Proc. IEEE, vol. 100, no. 1,
pp. 60–74, Jan. 2012.
[21] J. M. Bradley and E. M. Atkins, ‘‘Coupled cyber–physical system mod-
eling and coregulation of a CubeSat,’’ IEEE Trans. Robot., vol. 31, no. 2,
pp. 443–456, Apr. 2015.
32400
[22] T. Geyer, G. Papafotiou, and M. Morari, ‘‘Hybrid model predictive control
of the step-down DC–DC,’’ IEEE Trans. Control Syst. Technol., vol. 16,
no. 6, pp. 1112–1124, Nov. 2008.
[23] G. Pannocchia and J. B. Rawlings, ‘‘Disturbance models for offset-free
model-predictive control,’’ AIChE J., vol. 49, no. 2, pp. 426–437, 2003.
[24] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. Cambridge,
MA, USA: MIT Press, 2016. [Online]. Available: http://www.
deeplearningbook.org
MUHAMMAD SALEH MURTAZA GARDEZI
received the B.S. degree in electrical engineer-
ing from the University of Engineering and
Technology, Lahore, Pakistan, in 2015, and the
master’s degree in power electronics and control
systems from the National University of Sciences
and Technology, Islamabad, Pakistan, in 2017. His
research interests include model predictive con-
trol, control of power converters, and machine
learning-based control systems.
AMMAR HASAN was born in Pakistan in 1982.
He received the B.E. degree in electrical engi-
neering from the National University of Sciences
and Technology, Islamabad, Pakistan, in 2004, and
the master’s and Ph.D. degrees in control systems
from the Imperial College London, London, U.K.,
in 2008 and 2012, respectively.
Since 2012, he has been with the School of
Electrical Engineering and Computer Science,
National University of Sciences and Technology,
where he is currently an Assistant Professor. His current research interests
include model predictive control, control of power electronic converters,
optimization techniques for control systems, and numerical algorithms as
dynamical systems.
VOLUME 6, 2018