Applied Energy xxx (2014) xxx–xxx

Contents lists available at ScienceDirect

Applied Energy
journal homepage: www.elsevier.com/locate/apenergy

A load predictive energy management system for supercapacitor-battery

hybrid energy storage system in solar application using the Support
Vector Machine q
Yen Yee Chia a, Lam Hong Lee b, Niusha Shafiabady a,⇑, Dino Isa a
a
Department of Electrical and Electronic Engineering, University of Nottingham, Malaysia Campus, Malaysia
b
Faculty of Integrative Sciences and Technology, Quest International University Perak, Malaysia

h i g h l i g h t s

 A novel energy management system (EMS) for supercapacitor-battery hybrid energy storage system is implemented.
 It is a load predictive EMS which is implemented using Support Vector Machine (SVM).
 An optimum SVM load prediction model is obtained, which yields 100% accuracy in 0.004866 s of training time.
 The implemented load predictive EMS is compared with the conventional sequential programming control.
 This methodology reduces the number of power electronics used and prolong battery lifespan.

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

Article history: This paper presents the use of a Support Vector Machine load predictive energy management system to
Received 14 April 2014 control the energy flow between a solar energy source, a supercapacitor-battery hybrid energy storage
Received in revised form 18 July 2014 combination and the load. The supercapacitor-battery hybrid energy storage system is deployed in a solar
Accepted 8 September 2014
energy system to improve the reliability of delivered power. The combination of batteries and superca-
Available online xxxx
pacitors makes use of complementary characteristic that allow the overlapping of a battery’s high energy
density with a supercapacitors’ high power density. This hybrid system produces a straightforward
Keywords:
benefit over either individual system, by taking advantage of each characteristic. When the supercapac-
Supercapacitor
Battery
itor caters for the instantaneous peak power which prolongs the battery lifespan, it also minimizes the
Solar application system cost and ensures a greener system by reducing the number of batteries. The resulting perfor-
Load prediction mance is highly dependent on the energy controls implemented in the system to exploit the strengths
Support Vector Machine (SVM) of the energy storage devices and minimize its weaknesses. It is crucial to use energy from the superca-
pacitor and therefore minimize jeopardizing the power system reliability especially when there is a
sudden peak power demand. This study has been divided into two stages. The first stage is to obtain
the optimum SVM load prediction model, and the second stage carries out the performance comparison
of the proposed SVM-load predictive energy management system with conventional sequential program-
ming control (if-else condition). An optimized load prediction classification model is investigated and
implemented. This C-Support Vector Classification yields classification accuracy of 100% using 17 support
vectors in 0.004866 s of training time. The Polynomial kernel is the optimum kernel in our experiments
where the C and g values are 2 and 0.25 respectively. However, for the load profile regression model
which was implemented in the K-step ahead of load prediction, the radial basis function (RBF) kernel
was chosen due to the highest squared correlation coefficient and the lowest mean squared error. Results
obtained shows that the proposed SVM load predictive energy management system accurately identifies
and predicts the load demand. This has been justified by the supercapacitor charging and leading the
peak current demand by 200 ms for different load profiles with different optimized regression models.
This methodology optimizes the cost of the system by reducing the amount of power electronics within
the hybrid energy storage system, and also prolongs the batteries’ lifespan as previously mentioned.
 2014 Elsevier Ltd. All rights reserved.

q
This paper is included in the Special Issue of Energy Storage edited by Prof.
Anthony Roskilly, Prof. Phil Taylor and Prof. Yan.
⇑ Corresponding author.

http://dx.doi.org/10.1016/j.apenergy.2014.09.026
0306-2619/ 2014 Elsevier Ltd. All rights reserved.

Please cite this article in press as: Chia YY et al. A load predictive energy management system for supercapacitor-battery hybrid energy storage system in
solar application using the Support Vector Machine. Appl Energy (2014), http://dx.doi.org/10.1016/j.apenergy.2014.09.026
2 Y.Y. Chia et al. / Applied Energy xxx (2014) xxx–xxx
1. Introduction
A Support Vector Machine (SVM) load predictive energy man-
agement system is proposed in this paper. The proposed system
is deployed in a hybrid energy storage system (HESS) which con-
sists of lead acid battery and supercapacitor. The main rationale
of implementing this SVM-load predictive energy management
system is to accurately predict the load demand in advance for a
solar energy application. This proposal replaces part of the power
electronics with intelligent software which allows the improve-
ment of efficiency and lowers the cost of other components. The
reduction in cost can be achieved by eliminating the need for a
bi-directional DC-to-DC converter used for balancing the voltage
level between the supercapacitor and the battery. Research which
was carried out in [1] has proven that the system cost of an energy
management system using a software approach is reduced by
approximately 10.23% when compared with a conventional hard-
ware approach of energy management system which utilizes the
DC/DC converter. This is very crucial to ensure that the system
does not run in the risk of draining the battery, as the supercapac-
itor acts as an additional load if its voltage is lower than the batter-
ies nominal voltage, when both energy storage devices are in direct
parallel connection.
Solar energy source has been claimed as a clean and noise free
source of electricity. Even so, a reliable energy storage system is
required as an energy buffer to bridge the mismatch between
available and required energy. The proposed energy storage tech-
nology in this research is the integration of the lead-acid batteries
(that acts as a main energy storage device) and an auxiliary energy
storage device which is the supercapacitor. This hybrid energy
storage system is different from the conventional systems which
only consist of battery as the energy storage device.
The proposed hybrid energy storage system leads to system
cost reduction. This is accomplished by reducing the number of
batteries, and also the battery replacement costs by prolonging
battery lifespan. In a common household load profile, there is cer-
tain intermittent demand for high current such as when a motor
starts up. This can be 6–10 times of the normal operating current
of the motor and thus affects the battery life [2,3]. In a conven-
tional solar system, lead-acid batteries are always used to satisfy
peak current burst. Other than reducing battery life, the number
of lead-acid batteries in the conventional setup can be impractica-
ble large in order to match the peak current requirement. This
oversized number of battery for this purpose is proven to be costly
and not effective as the peak current demand might only need to
be met for a few seconds at a particular time [2].
The proposed SVM-load predictive energy management system
aims to solve the issues of coupling the battery and the superca-
pacitor in the hybrid energy storage system. In a hybrid energy
storage system, optimal design of the setup topologies between
different energy storage devices has been the subject of many
researches. Most conventional approaches consider a direct paral-
lel connection between the two storage banks [3], a bidirectional
DC/DC converter interfacing the two storage banks [2] and dual
input bidirectional DC/DC converter topology [5,6].
One of the onerous challenges in eliminating the number of
power electronics for the hybrid energy storage system is balanc-
ing different voltage level for each of the energy storage devices.
In this research, this is done by implementing a supervised learn-
ing machine – Support Vector Machine, MOSFET switches, sensors
and microcontroller to completely switch off battery during power
peak time. Therefore, the supercapacitor is turned on to cater for
the entire power peak without any delay in time response. Fig. 1
depicts the block diagram of the architecture of the proposed sys-
tem. This system controls the energy flow from the hybrid energy
storage system economically and accurately.
Please cite this article in press as: Chia YY et al. A load predictive energy management system for supercapacitor-battery hybrid energy storage system in
solar application using the Support Vector Machine. Appl Energy (2014), http://dx.doi.org/10.1016/j.apenergy.2014.09.026
Fig. 1. System architecture of the SVM load predictive energy management system.
Results obtained shows that this load predictive system allows
for accurate load classification and prediction. The supercapacitor
in the hybrid energy storage system is able to switch on to cater
for peak power without delay. This is crucial in maintaining an
optimized battery depth-of-discharge (DOD) in order to reduce
the rate of battery damage thru a degradation mechanism which
is caused from particular stress factors (especially sulphation on
the battery electrode and electrolyte stratification). The result
obtained from this research is compared with the result from the
energy management system which is coupled with sequential pro-
gramming (i.e. if-else conditional algorithm).
This paper is organized as follows: Section 1 is an introduction
of the proposed work. In Section 2, we present a literature review
on hybrid energy storage systems, batteries and supercapacitor
combination, and the system configuration. Section 3 shows the
research methodology on the implemented SVM-load predictive
energy management system for solar application. Section 4 pre-
sents the results obtained for the optimum load prediction model
using C-Support Vector Classification. Furthermore, the perfor-
mance of the implemented SVM-load predictive energy manage-
ment system is also compared with the conventional energy
management system using sequential programming in terms of
using the capacitor energy to address load requirements in a timely
manner. These results are analyzed and discussed in Section 4. The
conclusion for this research is presented in Section 5.
2. Literature review
In this section, the characteristics of battery and supercapacitor
are presented to explain the rationale of deploying the supercapac-
itor as an auxiliary energy storage device in battery systems
namely, the hybrid energy storage system (HESS).
2.1. Hybrid energy storage system
The batteries in a conventional stand-alone solar system are
replaced typically every 3–5 years, depending on the load demand
curve [2,3]. An oversized battery system is suggested to cater for
the peak power and also to prolong the battery lifespan. The deg-
radation in battery lifespan is due to inconsistent battery charging
by the solar energy source, as the output of the source is heavily
dependent on weather condition. The output of the solar energy
source fluctuates according to the intensity of the light, resulting
an inconsistent battery charging and discharging cycle. Heavy cur-
rent discharging due to heavy load requirements will also affect
battery life.
 Y.Y. Chia et al. / Applied Energy xxx (2014) xxx–xxx
The stress factor on the battery such as irregular discharging
rate and extensive time at the low state-of-charge (SOC) could
increase the rate of damage to the battery. The notable damage
mechanisms are related to battery electrolyte stratification and
also irreversible sulphation, which greatly shortens battery life-
time. Ideas and efforts have been put forward to extend battery
lifespan and reduce battery quantity used in the system, where
one solution is done by pairing batteries with super capacitors as
mentioned previously. When the batteries are paired with superca-
pacitors, the supercapacitors can act as a buffer, relieving the bat-
tery of pulsed or high power drain, as well as reducing the depth of
charge discharge cycles by means of buffering. This idea emerges
due to the fact that the supercapacitor has a greater power density
than the battery and this allows the supercapacitor to provide
more energy over a short period of time. Conversely, the battery
has a much higher energy density and this allows the battery to
store more energy and supply to the load over a longer period of
time. Hence, the role of supercapacitors is to supply sufficient
energy for peak power requirements while the role of battery is
to supply continuous power at a nominal rate.
2.2. Battery and supercapacitor
For solar charging application, the energy storage choice favors
batteries with deep cycle. Deep cycle batteries are designed to have
bigger plates and dissimilar chemistry to prevent the corrosive
effect when full capacity is frequently utilized [7]. Fig. 2 (which
was retrieved from [8]) shows the energy density versus power
density of various sources.
Lead acid batteries, which offer deep cycles, large capacity and
wide availability is typically the choice for this application [2]. It
has relatively high energy density; however, it does not have
capability of instantaneous charging and discharging as those of
a supercapacitor [8–10].
It could store much energy but it takes a longer time to
discharge and recharge. If it is driven at high C-rates especially in
an application which requires sudden load current burst, the
lifespan of the battery will be severely diminished. Besides this,
drawing a high current will also heat up the battery and possibly
cause a fire due to thermal runaway [11,12]. Also, the deep dis-
charge due to the large bust current drawn will also cause the bat-
tery SOC to be lowered. The low battery SOC will then lead to
stratification and sulphation of the battery, which reduce its life-
span [13,14].
Fig. 2. Ragone plot [8].
Please cite this article in press as: Chia YY et al. A load predictive energy management system for supercapacitor-battery hybrid energy storage system in
solar application using the Support Vector Machine. Appl Energy (2014), http://dx.doi.org/10.1016/j.apenergy.2014.09.026
3
Much literature have been presented to explain the motivation
of coupling supercapacitors with batteries to overcome the high
depth of discharge of battery in renewable energy system [15].
HESS employs a supercapacitor, which has high power density,
high rate of charging/discharging, no overcharging risk, and much
higher life cycles as compared with lead-acid battery. This mainly
benefits from how supercapacitor stores energy [16,17].
Electric double layer capacitor (EDLC) fabrication is an emerg-
ing technology and has already been associated with many appli-
cations [17] due to its unique ability to fill the void between
batteries and capacitors owing to its characteristics of higher
energy density than conventional capacitors and higher power
density than batteries [18,19] as demonstrated in the Ragone plot
as illustrated in Fig. 2.
Very high surface areas activated capacitors use a molecule-thin
layer of electrolyte as the dielectric to separate charge [10,17]. Sup-
ercapacitors resemble a regular capacitor except that it offers a
very high capacitance in a small package. Energy storage is by
means of static charge rather than of an electro-chemical process
inherent to the battery [9,10,20]. The supercapacitor is categorized
into two groups: the EDLC, and pseudo-capacitor, where both
groups differ in the way they store the charges [21]. The electro-
chemical supercapacitor stores the charge through the reversible
absorption of ions from an electrolyte to two porous electrodes
[22]: This creates an electric double layer at the electrode. The
EDLC is used in this project due to its cycle life being higher than
a pseudo-capacitor [17,23]. Additionally, the capacitance of sup-
ercapacitor is dependent primarily on the characteristic (surface
area and pore size of the distribution) of the electrode material
(such as carbon and activated carbon) [17,24]. These materials
are cheap due to its wide availability.
EDLC operation does not involve chemical reaction; it merely
absorbs the ion in physical means [12,17]. Energy is stored in the
double-layer capacitor as charge separation in the double-layer
formed at the interface between the solid electrode material sur-
face and the liquid electrolyte in the micropores of the electrodes
[9,10,17]. As shown in Fig. 3 [19], the EDLC can last up to millions
of cycles as the charging and discharging process does not affect
the electrode physically [25].
The ions displaced in forming the double-layers in the pores are
transferred between the electrodes by diffusion through electro-
lyte. The supercapacitor stores only a small amount of energy,
yet it could deliver a rapid power discharge which made it suitable
for high rate of charging and discharging operation.
The EDLC has high capacitance because its high power delivery
capability lies in the fact that it has much lower internal resistance
as compared to the battery [9,10]. Table 1 shows the internal resis-
tance of Lead Acid battery and Supercapacitor:
As shown in Table 1 [11], since the internal resistance of sup-
ercapacitor is much lower than the battery, the supercapacitor
can supply a large burst of current to the load while the battery
Fig. 3. Electrochemical double layer capacitor (EDLC).
4 Y.Y. Chia et al. / Applied Energy xxx (2014) xxx–xxx

Table 1
Internal resistance of lead acid battery and supercapacitor.

Type of energy storage device Internal resistances

Lead acid battery 2–30 mX
Supercapacitor 0.02–0.2 mX

will supply lower continuous power for a longer period of time.

These properties complement the battery’s limitation and allow
the combination of SLA battery and supercapacitor in HEESS to (a) Topology of the passive parallel connection
possess both high energy and power density [15]. Moreover, the
supercapacitor is very rugged and has a life expectancy of up to
50,000 h [2,3]. This makes it an ideal choice for remote solar appli-
cation where maintenance is difficult. Hence, HEESS will extend
the battery runtime, reduces the battery size, minimizes space
requirements, reduces the pulse current noise and improve load
balancing [15,26]
The charging time, self-discharging time, power and energy
densities and efficiency of the HESS formed by the supercapacitor
and battery are listed in Table 2 [2,9,27]. (b) Topology of the Bi-Directional DC/DC
Generally, supercapacitors are known as high power density Converter
storage devices that allow instantaneous delivery of power peak.
Due to the ability of supercapacitor to absorb and discharge large
amount of energy in a short period of time, supercapacitors make
a great secondary energy storage device especially in pulsed load
applications, where a large amount of energy is drawn in short
periods of time. Hence, supercapacitors are the suitable device to
couple with battery in this research.

2.3. System configuration of hybrid electrical energy storage system

(HEESS)

In the HEESS, optimal design of the topologies between the sup- (c) Topology of the Dual-input Bi-Directional
ercapacitor and battery has been a popular topic of research. Most DC/DC Converter
conventional approaches consider a direct parallel connection
Fig. 4. Topology of DC/DC converter.
between two storage banks [4], a bidirectional DC/DC converter
interfacing the two storage banks [2] and a dual input bidirectional
DC/DC converter topology as shown in Fig. 4 (which was retrieved Tatsuto Kinjo, Tomonobu Senjyu, Katsumi Uezato and Hideki
from [5,6]). Fujita examined the use of electrochemical double layer capacitors
One of the challenges of coupling supercapacitors and batteries (EDLCs) to stabilize the output of a wind power generation system
is the small voltage differences existing between real devices. A [30]. This study uses an energy system which consists of an EDLC
conventional way of coupling supercapacitor and battery is by con- and power electronic devices to compensate for the fluctuating
necting them in parallel [28]. However, this approach reduces the power of a photovoltaic system [29,30]. The power electronic
capacity of the supercapacitor [29]. This direct approach maintains devices include a buck boost DC/DC converter and a bi-directional
the same voltage over both sources but limits the power delivered inverter to complete the current source of the photovoltaic system.
from the supercapacitor [28]. Fig. 5 (which was retrieved from [29]) shows the current source of
The single DC/DC converter controls the output current of the the photovoltaic system:
battery and allows the supercapacitor to supply extra power In Fig. 5, GD1 (which acts as a gate turn-off thyristor) and D1
requirement to the load. Lastly, the dual input bi-directional DC/ (which acts as fast-recovery diode) operate as a boost converter
DC converter gives rise to highest efficiency, reliability and flexibil- to charge the EDLC bank when GD2 is turned on. Whereas, GD2
ity [5]. However, it involves the use of more costly DC/DC convert- (which is also a gate turn-off thyristor) and D2 (which is also a
ers. Hence, the tradeoff between these topologies is the cost of fast-recovery diode) operate as a buck converter to discharge the
power electronics and efficiency of the HEESS. EDLC bank when GD1 is turned off. The bi-directional inverter on
the other hand inverts the output from DC to AC. The charging
and discharging of the EDLC bank is controlled by using a PI control-
Table 2
Characteristics of battery and supercapacitor. ler [29,30]. This system uses power electronics, which are expen-
sive electronic components especially in high-power applications.
Battery Supercapacitor
In Harada, Sakau, Anan and Yamasaki’s research, an investiga-
Charging time Several hours Fraction of seconds to tion on the basic characteristic of the supercapacitors controlled
minutes
by non-isolated DC/DC converter has been carried out [30]. In this
Self-discharging time Weeks to few Hours to days
months research, the operating time, energy availability, input and output
Energy density 10–100 Wh/kg <5 Wh/kg voltage and current were analyzed.
Power density <500 W/kg >1000 W/kg From the result as shown in Fig. 6 (which was retrieved from
Charging/discharging 70–85% 85–98% [31]), it can be seen that the maximum ‘‘execution’’ or operating
efficiency
time of the buck-boost converter is the shortest among all the

Please cite this article in press as: Chia YY et al. A load predictive energy management system for supercapacitor-battery hybrid energy storage system in
solar application using the Support Vector Machine. Appl Energy (2014), http://dx.doi.org/10.1016/j.apenergy.2014.09.026
 Y.Y. Chia et al. / Applied Energy xxx (2014) xxx–xxx 5

Fig. 5. Current source of the photovoltaic system.

investigated converters. On the other hand, the buck boost con-

verter has the longest execution time under the similar conditions.
Fig. 7 (which was retrieved from [31]) shows the average output
power versus the output voltage of the non-isolated DC/DC con-
verters with supercapacitors for different energy availability.
Energy availability is calculated from the output power divided
by the input power of the converter. The buck and buck boost con-
verter’s output power increases as the output voltage increases,
whereas for boost converter, the output power is almost constant
although the output voltage increases.
From Harada, Sakau, Anan and Yamasaki’s research, it can be
seen that the energy availability, g of the buck converter is depen-
dent on the output voltage and the number of connected superca-
pacitors [31]. However, for the buck-boost converter, the energy
availability, g does not depend on the output voltage and the num-
ber of supercapacitors [31]. To obtain high output power from a
buck converter, many units of supercapacitors are needed to be
connected in series, in order to obtain a higher voltage from the
supercapacitor bank, to match the output voltage of the DC/DC
converter. This causes a problem of uneven charging due to the dis-
persion of capacitances of the supercapacitors [31]. This also leads
to a serious safety problem under British Standard of IEC 62391,
IEC 62391-2-1. To overcome this problem, an additional voltage
monitoring and current bypass circuit is needed for each superca-
pacitor to keep the balanced voltage [31].
M.E. Glavin, Paul K.W. Chan, S. Armstrong, and W.G. Hurley
examined the role of an electronic control unit (ECU) in a battery
supercapacitor hybrid energy storage system under different load
conditions [2]. The ECU is responsible in determining the State-
of-Charge (SOC) of supercapacitor and battery, and the supply of
energy to the load by either supercapacitor or battery with the
aid of various sensors [2,3]. The proposed ECU had a behavior as
shown in Table 3 [2].
The ECU in [2] is a proposal and it is not stated that whether the
ECU is software or hardware controlled. However, simulations
showed that the supercapacitor bank increased the SOC of battery
under peak and pulse current loads. This shows that the battery
lifespan is not prolonged with this strategy.
In [32], R.A. Dougal, Shengyi Liu and Ralph E. White investigated
the peak power enhancement and prolonged battery life of battery
– ultracapacitor hybrids over conventional systems. A simplified
model was created to analytically describe the performance of a
battery – ultracapacitor hybrid power source. Complementing a
battery power source with ultracapacitor banks was proven to Fig. 6. Input and output voltage versus operating time of (a) buck converters (b)
boost converters (c) buck-boost converters with supercapacitors.
greatly enhance peak power output, considerably reduce internal
losses and also prolong the discharge life of the battery. Fig. 8
(which was retrieved from [32]) shows the plot of the normalized
load current, battery current and supercapacitor current [32]. load off-state, the battery charges the supercapacitor. The peak
Based on the plot in Fig. 8 (which was retrieved from [32]), instantaneous current of the battery had been greatly reduced with
observation shows that during load on-state, both battery and sup- the assistance of the supercapacitor, which would otherwise have
ercapacitor provide the current consumed by the load. During the been the same as the output or load current. This greatly relieves

Please cite this article in press as: Chia YY et al. A load predictive energy management system for supercapacitor-battery hybrid energy storage system in
solar application using the Support Vector Machine. Appl Energy (2014), http://dx.doi.org/10.1016/j.apenergy.2014.09.026
6 Y.Y. Chia et al. / Applied Energy xxx (2014) xxx–xxx
Fig. 7. Average output power versus the output voltage of (a) buck converters (b)
boost converters (c) buck-boost converters with supercapacitors.
the battery off the peak stresses, reduces battery deterioration and
positively influences the performance of the system [32].
2.3.1. Energy management system for HEESS
J.C. Lima, A. Medeiros, V.M. Canalli, F. Antunes and F.S. Dos Reis
has implemented a neural network control system on a PIC micro-
controller to track the maximum power point transfer between the
load and PV system [33]. The neural network is used to control the
maximum output voltage in a DC/DC converter connected to solar
panels to obtain the maximum power transfer of the panels. The
power control is done in order to obtain the maximum power
transfer which extracts the best performance from the solar panel.
This can be achieved through the control of the duty cycle of the
DC/DC boost converter, which supplies an adapted voltage to a
3-phase inverter [33,34].
Fig. 9 (which was retrieved from [33]) illustrates the proposed
system in [33] which does not have an energy storage system
which can mitigate the problems associated with low-light condi-
tions. Furthermore, high power rating electronics and complexity
of designing bi-directional converter cause impracticability of
HEESS. In order to achieve an efficient HEESS at lower cost, soft-
ware based Energy Management System (EMS) has been
Please cite this article in press as: Chia YY et al. A load predictive energy management system for supercapacitor-battery hybrid energy storage system in
solar application using the Support Vector Machine. Appl Energy (2014), http://dx.doi.org/10.1016/j.apenergy.2014.09.026
researched into. L. Gao, R. Dougal and S. Liu presented a software
approach which offers modularity, design simplicity and reduces
the reliance on power electronics [35]. Previous research works
in [1,36] had shown a Sequential Programing based EMS which
adopts the ‘Monitor and Respond’ strategy but has shown unsatis-
fying results. This is due to the time required by the software
approach to measure, process and respond accordingly. A 70 ms
delay was recorded in the respond of the supercapacitor after a
peak load current is monitored. This does not meet the perfor-
mance requirement in small DC machines, as the starting current
is usually quite short in period. For example, it is measured that
a 12 V, 1A DC motor load has a starting current of only 5 ms in
duration [37]. Hence, the motor load will draw the large current
burst from the battery and causing the adverse effects on the
battery as mentioned earlier. In order to respond to this downfall
of the software approach, load prediction capability has to be inte-
grated into the EMS in order to match the respond of the superca-
pacitor with the performance required.
The Support Vector Machine (SVM) here is used for load
prediction, and it was applied and implemented to the supercapac-
itor-battery hybrid energy storage system. The SVM, a form of
supervised machine learning founded by V. Vapnik, is a non-para-
metric statistical model based on Structural Risk Minimization
(SRM) principle that offers excellent generalization and predictive
capability for limited sample size [38]. Since the load demand
could be predicted by using recursive time series, Support Vector
Regression (SVR) was chosen for the implementation in load pre-
diction due to its high generalization, single global minimum char-
acteristic and this will yield a good non-linear system model
[38,39].
In [40] Chang et al. proposed a SVR approach for the EUNITE Net-
work Competition which is the prediction of daily maximal electri-
cal load of January 1999 based on temperature and electricity load
demand (data used is from 1997 to 1998). It is interesting to men-
tion that there is a periodic component within the data set due to
the seasonal variation of consumer electricity demand (such as
‘holiday’ effects use of less electricity during major holidays) and
the impact of weather on electricity demand. Their inputs were sev-
eral attributes, including binary attributes for indicating which day
of the week or holiday, etc. [41]. From these attributes, they have
formulated the predicted maximum load in numerical form. They
concluded that the use of temperature data did not work due to
the reason that the inherent difficulty in predicting temperature.
They have also concluded that this SVR approach was feasible for
determining an accurate load prediction model. Chen et al. have
presented a SVM load forecasting approach which is the winning
approach in the EUNITE Network Competition [40,41]. It is interest-
ing to note from the design details that the use of temperature in
their model actually decreases the accuracy of the predictions. This
is due to the wide variance of the output hence resultsto an impro-
per temperature estimation [42]. Change et al. also experimented
their proposed approach with data inputs excluding the previous
(in time series) load data. The result obtained shows unsatisfied
performance. It is worth to mention that the inputs to the SVR are
restricted to time series load data [42].
In [43], Zhang et al. discussed the use of SVM for short-term
load forecasting. The authors stated that most of the linear models
such as Kalman Filter, Auto-Regressive (AR) and Auto-Regressive
Moving Average (ARMA) models are not typically sufficient to
model nonlinearities associated with short term load forecasting
processes. The use of SVR, with both electrical load data and corre-
sponding weather time series data, appears to outperform other
Neural Network (NN) based techniques including a back-propaga-
tion neural network (BPNN). The authors also used cross validation
to select the parameters for the RBF kernel function as well as the
regularization constant. The result obtained shows that Mean
 Y.Y. Chia et al. / Applied Energy xxx (2014) xxx–xxx 7

Table 3 Absolute Percentage Error (MAPE) of the SVM approach was lower
Behavior of ECU [2]. than that of the BPNN [43].
No. Condition Action Studies discussed above show the advantages of having
1 PV  PV supplies load software based energy control for renewable energy systems. The
power = load  No battery charging battery storage system could however be further enhanced using
(Battery SOC a buffering element, in this research, we propose a hybrid energy
High) storage system which consists of supercapacitor and battery, to
2 PV  PV supplies load prolong the battery lifespan and reduce maintenance and opera-
Power = Load  No battery charging tional costs at the same time.
(Battery SOC
Low)
In [44], the application of SVM is said to be motivated by good
generalization capability of the SVM. A load forecasting-energy
3 PV  PV supplies load
Power > Load  No battery charging
management system (using SVM) aided with the use of sensors
(Battery SOC could be used in the supercapacitor-battery hybrid energy storage
High) system (SB-HESS) to achieve maximum optimization in terms of
4 PV  PV supplies load system cost and power delivery. SVM implements classifiers of
Power > Load  PV charges battery an adjustable complexity, and controls the latter for optimal gener-
(Battery SOC alization ability, i.e. the performance for the future unknown sam-
Low)
ples [44]. From the literature study presented above, there is none
5 PV  PV supplies load of the energy management system that is implemented with SVM
Power < Load  Battery supplies load
load-predictive software to control the energy flow between the
(Battery SOC
High) hybrid energy storage devices and load.
6 PV  PV supplies load
Power < Load  Battery supplies load until minimum SOC is
3. Research methodology
(Battery SOC reached then shut down load
Low) This section consists of the background theories on SVM and
7 No PV Power  Battery supplies load Support Vector Regression (SVR) for load prediction and identifica-
(Battery SOC tion, and the steps in constructing a lab-scale prototype energy
High) management system. This prototype consists of a software control
8 No PV Power  Shut down load board which uses SVM and SVR for load forecasting.
(Battery SOC Pairing supercapacitors and batteries however requires expen-
Low)
sive and extensive power electronics, elevating the already high
costs associated with these hybrid photovoltaic systems [45].
There are however, other methods, which could be used in devel-
oping these systems. In this research, the wide availability and
affordability of microcontrollers nowadays allow these hybrid
systems to be controlled using purely software methods. The
microcontroller is embedded with the SVM pattern classifier to
decide the switch of energy sources, depending on the load
requirement. The supervised learning system in SVM allows the
prediction of load demand before it occurs.
This methodology aids in reducing the delay in delivering power
even when there are a few possible cases to be considered in con-
necting or disconnecting battery and supercapacitor to the load.
This would not only lower the operational cost, but at the same
time, allows the hybrid photovoltaic system to be flexible, which
comes in handy in places with different seasons and unpredictable
weather. The implementation of microcontroller also allows the
monitoring of multiple parameters, which may contribute to the
efficiency of the hybrid photovoltaic systems, optimizing the oper-
ation of these systems by taking appropriate actions when needed.

3.1. System architecture and integration

Fig. 8. Normalized load current, battery current and supercapacitor current.
Fig. 9. Block diagram of the power system suggested in [33].
Referring to Fig. 1, CPU + LCD represents the computer to run
the MATLAB R2011b. SLA BAT represents sealed-lead acid battery,
while S.C represents supercapacitors. MAXIM USB 6009 DAQ is
used to perform load current data acquisition, while Arduino
Mega2560 is used for voltage monitoring and switching control
with the Software Control Board. MATLAB R2011b allows the con-
trol of MAXIM USB 6009 DAQ Session interfaces through its DAQ
toolbox, Arduino Mega2560 through Arduino IO package, and LIB-
SVM through LIBSVM MATLAB extension. This interfacing allows
the system to perform load identification and load prediction. In
order to train and test the classification and load prediction accu-
racy of the SVM, the load profile has to be replicated by using a
programmable load (which is denoted as Motor and Motor Drive

Please cite this article in press as: Chia YY et al. A load predictive energy management system for supercapacitor-battery hybrid energy storage system in
solar application using the Support Vector Machine. Appl Energy (2014), http://dx.doi.org/10.1016/j.apenergy.2014.09.026
8 Y.Y. Chia et al. / Applied Energy xxx (2014) xxx–xxx

in Fig. 1). Several of load profiles are simulated and presented in

next methodology steps.
A schematic of software control circuit architecture which
enables the control of charging and discharging operation of the
batteries and supercapacitor bank is shown in Fig. 10.
This control circuit board allows the control of the charging and
discharging operation for both the battery and supercapacitor.
With the load predictive system using SVM and SVR, load demand
could be predicted multiple steps in advanced to allow the charg-
ing of the supercapacitor before the peak load occurs.

3.2. Load current prediction

In this research, K-step autoregressive (AR) model in which the

output variable dependent only on the historical values is applied.
In other words, the load forecasting technique used for this
research is a time series method – autoregressive model [46]. How-
ever, the AR output (dependent variable) is solely dependent line-
arly on its own previous values [47]. Hence, it is not capable of
dealing with nonlinear model representation in load forecast. Since
SVR performs well in the nonlinear domain, the concept of AR is
applied to SVR to approach the trend of load demand.
The modeling of time series could be considered as regression
problem in a high dimensional input space (more than one input
feature variable). The model which uses this prediction method
can be predicted by iteratively using one-step-ahead prediction
model [48], which is illustrated in Eq. (1):
Fig. 10. Schematic of software control board.
Assuming there are k training samples {xi, yi} where i = 1..., k,
and each sample has l input (xi e l) with an output class label of
(yi e 1, 1). The vector, w which is the vector perpendicular to
the hyperplane, parameterizes the hyperplanes in l and with a
constant, b as shown in Eq. (10) [51] in the later section.
For a training set as illustrated in Eq. (4),
K ¼ fðx1 ; y1 Þ; ðx2 ; y2 Þ . . . ðxn ; yn Þg ð4Þ

Xðt þ 1Þ ¼ F½XðtÞ; Xðt  1Þ . . . Xðt  m þ 1Þ ð1Þ and the linear function f(x) as shown in Eq. (5),

where F is the prediction model, m is the size of regressor, and t is f ðxÞ ¼ ðw  xÞ þ b ð5Þ
the time. hence solve the optimization problem as shown in Eq. (6),
For n iteration, the n-th prediction is illustrated in Eq. (2):
1X n
    X n
  
Xðt þ nÞ ¼ F½Xðt þ n  1Þ; Xðt þ n  2Þ . . . Xðt  m þ nÞ ð2Þ min ai  ai aj  aj ðxi  xj Þ þ e ai þ ai
aij Ri 2 i;j¼1 i¼1
The relationship between the SVR independent input variables and X
n
 
dependent output variable by using the K-step ahead approach is  yi ai þ ai ð6Þ
shown in Eq. (3): i¼1

)Yðt þ KÞ ¼ F½ðYðtÞ; Yðt  1Þ . . . Yðt  m þ 1Þ ð3Þ subject to

where K is the points in advanced of time to be predicted and m is X

n
 
the number of previous points to be used for the next step
ai  ai ¼ 0 ð7Þ
i¼1
prediction.
Based on the time response improvement requirement deter- where 0 6 ai ; ai 6 C; i ¼ 1; 2; . . . ; n.
mined empirically, the value of K is set to 7 and the value of m is By attaining the optimum value to obtain x and the constant b
selected to be 3. This K-step-ahead autoregressive strategy will X
n
 
yield a predicted output value which leads the actual load current x¼ ai  ai xi ð8Þ
i¼1
(K⁄ sampling time – processing time) in advanced. Fig 11 shows the
n 
X 
representation of the 7-steps ahead autoregression load prediction b ¼ yi  aj  aj ðxi  xj Þ þ e ð9Þ
model. j¼1

3.3. Support Vector Regression training and testing where C is the regularization parameter, a is the Gaussian kernel
function’s variance and e is the insensitive loss function.
To implement the K-step ahead prediction for each load profile,
the SVR model that describes the relationship between the previ-
ous 3 load current values and the 7 points ahead load current
value, has to be trained and optimized to ensure the model has
good descriptive accuracy. There are 3 steps to the modeling pro-
cess: data preprocessing, model selection, and cross-validation
and grid search.

3.3.1. Support Vector Regression

The SVM was originally developed for pattern recognition.
However, with the predictive capability of Vapnik’s e-insensitive
loss function, it has been extended into nonlinear regression esti-
mation domain [49,50]. It has been proven to exhibit excellent per-
formance for time series forecasting [46]. Fig. 11. The 7-steps ahead load prediction model.

Please cite this article in press as: Chia YY et al. A load predictive energy management system for supercapacitor-battery hybrid energy storage system in
solar application using the Support Vector Machine. Appl Energy (2014), http://dx.doi.org/10.1016/j.apenergy.2014.09.026
 Y.Y. Chia et al. / Applied Energy xxx (2014) xxx–xxx 9

From this, the linear regression function can be constructed. To Table 4

extend this linear condition to non-linear domain, the use of kernel Data adjusted for 7th point ahead forecast.

trick is introduced to transform the input data into a higher
dimensional feature space, which solves the problem of curse of
dimensionality [49,52].
3.3.2. Data preprocessing
This step involves data acquisition and data adjustment for the
implementation of the 7-steps ahead forecast technique. Data for-
mat is converted into the format as required for LIBSVM which is
called sparse format. Maxim USB 6009 DAQ is used to collect load
current data with a sampling rate of 10 Hz. The input vector is cre-
ated by making every 3 previous load current value as the input
feature variables. The output label linking each of the input vectors
will be the 7th point in advanced load current value as shown in
Table 4.
3.3.3. Model selection
This step involves the selection of the optimal kernel model
(linear, polynomial, radial basis function (RBF) or sigmoid). Unlike
linear kernel, RBF kernel could perform nonlinear mapping of the
samples into a higher dimensional feature space. This gives RBF
kernel a good nonlinear modeling capability which suits the
nonlinear time series forecast requirement. RBF kernel also has
fewer tuning parameter than polynomial kernel which makes the
selection of parameter process straightforward. Moreover, sigmoid
kernel faces some validity issues under certain parameter condi-
tion [53].
Thus, the RBF kernel has been selected as the primary choice.
Yet, all kernels are tested for every single load profile regression
by selecting the kernel which yields the highest squared correla-
tion coefficient and the lowest mean squared error. The squared
correlation coefficient (r2) shows the proportion of variance of out-
put that could be predicted by having the input. The mean squared
error (MSE) is a risk function that measure difference between the
predicted output and the true output [54].
3.3.4. Cross validation and grid search
This step involves the search for the best hyper/tuning parame-
ter for the model to predict real time load current data accurately
in every case of load profile. The hyperparameter serves to
optimize the loss function on the training dataset [55].
In the cross-validation stage, the dataset is split into v portions,
where (v1) portions are for training and one portion is for testing
and validation. The accuracy of prediction obtained from the model
by using the testing set reflects the model’s ability to predict
unknown dataset, which prevents overfitting problem.
For load profile regression, cross-validation is done using a dif-
ferent approach as compared to the v-fold cross validation method
used in classification problem [56]. This is due to the reason that
only one load profile dataset has to be included into the training
set for autoregressive time series prediction. Hence, the validation
is done by another similar set of load profile where there are slight
differences due to the voltage drops.
For parameters tuning, grid search is used along with cross
validation method. Pairs of hyper-parameter/tuning values in
exponentially growing sequence are tested and the pair which
yields lowest MSE is selected [57].
3.4. Load profile identification
For a real-life application, the load predictive energy manage-
ment system is integrated with the load profile identification.
The load profile identification has the capability to ensure the load
current forecasting is done based on the correct model.
Output label Input vector
Feature variable 1 Feature variable 2 Feature variable 3
1.250736 1.215048 1.207400 1.220146
1.243088 1.207400 1.220146 1.217597
1.230343 1.220146 1.217597 1.225244
1.220146 1.217597 1.225244 1.220146
1.215048 1.225244 1.220146 1.222695
The load profile identification by using SVM classification is done
through pattern recognition, as every load profile exhibits different
load demand pattern. This can be implemented with the use of
multiclass, multidimensional SVM. SVM is chosen for this applica-
tion as it has been proven with excellent generalization capability.
Since this research deals with real time peak load prediction, where
the forecasting task occurs after the identification of SVR load pro-
file model is made, it is assumed that the beginning part of the load
profiles holds a distinguish pattern which is recognizable.
3.5. Support Vector Machine for classification training and testing
To implement the load profile identification, a SVM model
which correctly classifies the actual load profile from the SVR load
profile models is trained and validated. This again, involves the 3
steps as mentioned above: data preprocessing, model selection,
and cross-validation and grid search.
3.5.1. Support Vector Machine
The foundation of SVM has been developed by Vladmir Vapnik
in 1995. SVM has been gaining popularity over the past 20 years
due to many potential features and good empirical performance.
SVM is based on Structural Risk Minimization (SRM) principle
which minimizes an upper bound on the expected risk [58] and
controls the complexity of the learning machine through a param-
eter called the VC dimension.
SVM classification functions by creating a hyperplane that sep-
arates a set of data containing two classes [59]. According to the
SRM principle [60], there is only one optimal separating hyper-
plane, which has the maximum distance to the closest data points
of each class. The data points which are closest to the optimal sep-
arating hyperplane, are called the Support Vectors (SVs). Fig. 12
illustrates a 2D SVM model.
Assuming there are k training samples {xi, yi} where i = 1..., k,
and each sample has l inputs ðxi eRl Þ with an output class label of
(yi e 1, 1). The vector, w which is the vector perpendicular to
the hyperplane, parameterizes the hyperplanes in Rl and with a
constant, b as shown in Eq. (10) [51]:
wxþb¼0 ð10Þ
Hence the function which classifies the training data is illustrated in
Eq (11) [57]:
f ðxÞ ¼ signðw  x þ bÞ ð11Þ
However, the hyperplane can also be expressed by all pairs of {kw,
kb} for k e R+. Hence, the canonical hyperplane is defined to separate
the data from the hyperplane by a minimum distance of 1 unit [39].
Hence,
yi ðxi  w þ bÞ P 1 8i ð12Þ
In a hyperplane, all pairs of {kw, kb} describe the same hyper-
plane, but all are different from each other in terms of the func-
tional distance to the data point. By normalizing the magnitude
of w, the geometric distance [57] is given by Eq. (13).

Please cite this article in press as: Chia YY et al. A load predictive energy management system for supercapacitor-battery hybrid energy storage system in
solar application using the Support Vector Machine. Appl Energy (2014), http://dx.doi.org/10.1016/j.apenergy.2014.09.026
10 Y.Y. Chia et al. / Applied Energy xxx (2014) xxx–xxx

yi ðxi  w þ bÞ 1
P ð13Þ ( )
kwk kwk ( ) ( )
( ) ( ) ( )
For good generalization capability, maximum geometric (.) ( ) ( )
( )
distance of the data point from the hyperplane is sought after. To ( ) ( )
( ) ( )
achieve this, the ||w|| is minimized by using Lagrange multiplier
( ) ( ) ( )
[57]. Hence the minimization is given by Eq. (14).
Minimize: ( )
( )
Input space Feature space
X
l
1X l X l
WðaÞ ¼  ai þ y y ai aj ðxi  xj Þ ð14Þ
i¼1
2 i¼1 j¼1 i j Fig. 13. Mapping of data in input space to feature space.

Subject to
Table 5
X
l
Data adjusted for load profile classification.
yi ai ¼ 0 ð15Þ
i¼1 Class label Input vector
F1 F2 F3 F4 F5 F6 F7 ...
where 0 6 ai 6 C "i.
a is the vector of l non-negative Lagrange multipliers, and by 1 1.31 1.32 1.34 1.36 1.38 1.40 1.41 ...
2 1.31 1.31 1.31 1.31 1.32 1.31 1.32 ...
defining the matrix of, this term is translated to Quadratic 3 1.29 1.30 1.30 1.31 1.31 1.32 1.33 ...
Programming (QP) Problem which can be solved by the method 4 1.30 1.30 1.30 1.30 1.31 1.32 1.31 ...
of Lagrange multipliers [51]: 5 1.31 1.31 1.31 1.32 1.31 1.32 1.32 ...
Minimize:

1
WðaÞ ¼ aT 1 þ aT Ha ð16Þ
2 Table 6
Performance of SVM on the implemented load predictive energy management
Subject to systems.

aT y ¼ 0
where 0 6 ai 6 C; i ¼ 1; . . . ; n; ð17Þ

Besides, from the dual formulation, the optimal separating

hyperplane is defined as Eq. (18).
X
w¼ ai yi xi ð18Þ
i

When the functional distance of a data point is >1, then ai = 0.

Hence, only the data points that lie on the margin have ai > 0,
which are required to define the optimal separating hyperplane.
They are the SVs. The ai is also a measure of the data point weight
in contribution to the hyperplane.

x
1
−

||w 2
b=

|| Table 7
-
·x

Performance of SVR on the implemented load predictive energy management system.

w

b=
-

1
·x

Load SVM Kernel Parameter value MAPE Training

−
w

b=

profile type (%) time (s)

C g p
-
·x
w

1 Epsilon -SVR Polynomial 100 0.33 0.01 11.6882 0.294515

2 Epsilon -SVR RBF 100 0.33 0.01 7.5650 0.018975
3 Epsilon -SVR RBF 10 0.33 0.01 5.7966 0.025643
4 Epsilon -SVR RBF 0.1 0.33 0.02 9.5908 0.022941
5 Epsilon -SVR RBF 10 0.33 0.01 4.9574 0.020495

To find b, the positive and negative SVs are used [40], and Eq.
(19) illustrates the function that results b.

1
b ¼  ðw  xþ þ w  x Þ ð19Þ
2
y
Constructing a separating hyperplane in this feature space leads to a
non-linear decision boundary in the input space [59]. Expensive cal-
culation of dot products in a high-dimensional space can be avoided
Fig. 12. 2D SVM model. by introducing a kernel function [61].

Please cite this article in press as: Chia YY et al. A load predictive energy management system for supercapacitor-battery hybrid energy storage system in
solar application using the Support Vector Machine. Appl Energy (2014), http://dx.doi.org/10.1016/j.apenergy.2014.09.026
 Y.Y. Chia et al. / Applied Energy xxx (2014) xxx–xxx 11

Load Blue = Actual load profile

Profile Green = Predicted load profile

2.8

2.6
(a)
2.4

current (A)
2.2

2
1.8

1.6
1.4
0 20 40 60 80 100 120 140 160 180 200
time (seconds)
2.3
2.2 (b)
2.1
2
current (A)

1.9
1.8
1.7
1.6
1.5
1.4
1.3
0 20 40 60 80 100 120 140 160 180 200
time (seconds)
2.1
2 (c)
1.9
current (A)

1.8

1.7
1.6

1.5
1.4
0 20 40 60 80 100 120 140 160 180 200
time (seconds)
2.4
2.3 (d)
2.2
2.1
current (A)

2
1.9
1.8
1.7
1.6
1.5
0 20 40 60 80 100 120 140 160 180 200
time (seconds)

2.1

2 (e)
1.9
current (A)

1.8

1.7

1.6

1.5
1.4
0 20 40 60 80 100 120 140 160 180 200
time (seconds)

Fig. 14. Predicted and actual load for each load profiles.

Please cite this article in press as: Chia YY et al. A load predictive energy management system for supercapacitor-battery hybrid energy storage system in
solar application using the Support Vector Machine. Appl Energy (2014), http://dx.doi.org/10.1016/j.apenergy.2014.09.026
12 Y.Y. Chia et al. / Applied Energy xxx (2014) xxx–xxx

If the data points are not linearly separable, they may be line-
arly separable in the higher dimensional feature space [62]. This
can be done through the introduction of kernel trick, where the
data points are mapped into a higher dimensional space, as shown
in Fig 13 (which was retrieved from [63]):
The kernel function, K(xi, xj), which defines the dot product in
the higher dimensional feature space as shown in Eq. (20) [51,57].
Kðxi ; xj Þ ¼ Uðxi Þ  Uðxj Þ ð20Þ
In practice, various kernel functions can be used. The kernel func-
tion reflects the geometric relationship between the input vector
and the SVs, or the similarities of the features of the faults.
Hence the classification model with the implementation of
Fig. 15. Time response of supercapacitor with sequential programming.
kernel function is shown as Eq. (21). This is appropriate to be used
in this research to identify the load pattern.
When the soft margin constant, C is infinite, the optimal !
separating hyperplane will be able to separate all the data points X
theoretically. Yet, with a finite C, the classifier is having a soft-mar-
f ðxÞ ¼ sign ai yi ðKðxi ; xÞÞ þ b ð21Þ
i
gin which allows trade-off between classifying all data correctly
with the complexity of the hyperplane model [57].

(a) (b)
Supercapacitor Supercapacitor

Actual Load
Actual Load
Threshold
Threshold

(c) (d)
Supercapacitor Supercapacitor

Actual Load Actual Load

Threshold
Threshold

(e)
Supercapacitor

Actual Load
Threshold

Fig. 16. Current level for supercapacitor and actual load.

Please cite this article in press as: Chia YY et al. A load predictive energy management system for supercapacitor-battery hybrid energy storage system in
solar application using the Support Vector Machine. Appl Energy (2014), http://dx.doi.org/10.1016/j.apenergy.2014.09.026
 Y.Y. Chia et al. / Applied Energy xxx (2014) xxx–xxx
Fig. 17. Zoom in view of load prediction.
3.5.2. Data preprocessing
This step involves data acquisition, data adjustment and format
conversion for the SVM classification model implemented by using
LIBSVM. The MAXIM USB6009 DAQ is used to acquire the data
points from each of the load profiles. These data are sampled at
100 Hz to speed up the load identification process. These data
points are then converted into the feature variables of the input
vector as shown in Table 5.
There are 8 sets of data which were obtained from each load
profile, where 5 sets of each are selected as the training sets and
the rest of 3 are used as the testing data for validation.
3.5.3. Model selection
This step deals with the selection of the optimal kernel to be
implemented in the classification model of the load profile. Various
kernels are evaluated and the one with optimized parameter which
yields the highest accuracy with lowest number of SVs is chosen.
3.5.4. Cross validation and grid search
The 8-fold cross validation is carried out, but instead of one,
three sets of data are used as the testing data while the other 5 sets
are used for training purpose. This allows the classification accu-
racy to better reflect the model’s ability in classifying new data.
4. Experiments and evaluations
Results obtained from the experiments are presented and
discussed below. This section is divided into two parts; first part
shows the optimum load prediction classification model using C-
support vector classification and the second part presents the per-
formance for the optimized regression models of the implemented
load predictive energy management system.
4.1. Optimum load prediction model
Mean Average Percentage Error (MAPE) is the accuracy measure
for techniques used in modeling time series applications. It is
expressed as the average of the total difference in percentage as
shown in Eq. (22) [64]:
Pn LR LP
i¼1
i i
LR
MAPE ¼ 100  i
ð22Þ
n
where LR is the actual value of load current and LP is the value of
predicted load current.
Please cite this article in press as: Chia YY et al. A load predictive energy management system for supercapacitor-battery hybrid energy storage system in
solar application using the Support Vector Machine. Appl Energy (2014), http://dx.doi.org/10.1016/j.apenergy.2014.09.026
13
Besides that, n is the number of points to be compared. Hence,
by using Eq. (22), the accuracy of the predicted load current can be
measured.
The optimized load profile classification model is implemented
by using C-SVC (C-Support Vector Classification). From Table 6, it
can be observed that linear, polynomial and RBF kernels with opti-
mized function yield excellent classification accuracy. Hence, the
model with the least SVs and training time (as highlighted) is
selected for better and faster classification as the SVM computa-
tional complexity is proportional to the number of SVs.
The optimized regression models for each load profile are
shown as in Table 7. Table 7 shows the accuracy percentage of
MAPE and training time for the five simulated load profiles. A
few runs of experiments are done to select the appropriate kernel
for different load profiles and obtain the optimized parameters
value. The predicted and actual load profiles by using the
optimized models are shown in Fig. 14. The blue lines in Fig. 14
represent the actual load profile, while the green lines represent
the predicted load profile. Y-axis represents the current of the load
and x-axis represents the time in seconds. As can be seen from
Fig. 14, the predicted load profiles were predicted beforehand with
similar pattern to the actual load.
4.2. Performance comparison
By running the sequential programming with the if-else condi-
tion, it is found that the time delay of the supercapacitor charging
corresponding to the peak load has around a 200 ms delay, as
shown in Fig. 15. This also means the supercapacitor was turned
ON 200 ms after the load current had met the peak load threshold
value of 1.5 A which means that the battery took the load.
Results obtained from the oscilloscope using the kernel and
parameters mentioned above are shown in Fig. 16.
The square pulses in Fig. 16 represents the time when superca-
pacitor turns on; while the varying graph is the load current. The
horizontal line (label ‘threshold’ in Fig. 16) shows the current
threshold where the supercapacitor should be turned ON when
the load current exceeds the threshold value.
If the predicted current is greater than the threshold which is
determined by considering the optimal state-of-charge (SOC) of
the lead acid battery to avoid deep discharging the battery; the
supercapacitor is then turned on before the load current exceeds
the current threshold. In this case, the optimal state-of-charge of
the lead acid battery used is 80% which is also equivalent to
11.52 V.
This prolongs battery lifespan based on the Power SonicLead
Acid Battery data sheet [65]. Hence, with the correct SVR load pro-
file model used for the regression task, the SVM-load predictive
energy management system shows a good time response as it is
able to predict the peak load current and turn on the supercapac-
itor in advanced for all load profiles. Fig. 17 shows the zoom in
view of the load prediction results.
As shown in Fig. 15, supercapacitor is turned on after the peak
load occurred. On the other hand, when the load profile is success-
fully identified using the SVM model, the supercapacitor response
leads the peak load current by 200 ms as shown in Fig. 17. This
allows the supercapacitor to deliver power rapidly. This aids in
reducing the rate of damage of battery and evidently prolongs bat-
tery lifespan.
5. Conclusion
The proposed SVM-load predictive energy management system
is implemented on an energy storage system containing
supercapacitor and battery. Results obtained showed that this load
14 Y.Y. Chia et al. / Applied Energy xxx (2014) xxx–xxx

predictive system is capable of accurately predicting load require- [26] Karden E, Shinn P, Bostock P, Cunningham J, Schoultz E, Kok D. Requirement
for future automotive batteries – a snapshot. J Power Sources
ment. It shows 100% accuracy with classification time of
2005;144(2):505–12.
0.004866 s. This means that the supercapacitor in the hybrid [27] Govindaraj A, King M, Lukic SM. Performance characterization and
energy storage system is able to deliver energy to cater for peak optimization of various circuit topologies to combine batteries and ultra-
power requirement without delay. This is crucial in maintaining capacitors. In: Procs of the 36th annual conference on IEEE industrial
electronics society IECON; 2010. p. 1850–7.
an optimized battery depth-of-discharge (DOD) in order to reduce [28] Yu A, Chabot Victor, Zhang Jiujun. Electrochemical supercapacitors for energy
the rate of battery damage through a degradation mechanism storage and delivery: fundamentals and applications (electrochemical energy
which is caused by particular stress factors especially sulphation storage and conversion). America: CRC Press; 2013.
[29] Kinjo T, Senjyu T, Urasaki N, Fujita H. Output levelling of renewable energy by
on the battery electrode and electrolyte stratification. This pro- electric double-layer capacitor applied for energy storage system. Energy
posed SVM load predictive system also allows the improvement Convers IEEE Trans 2006;21(1):221–7.
of efficiency and lowers the cost of other components by eliminat- [30] Kinjo Tatsuto, Senjyu Tomonobu, Uezato Katsumi, Fujita Hideki. Output
leveling of wind power generation system by EDLC energy storage system.
ing the need for a bi-directional DC/DC converter used in balancing Issue Electr Eng Japan 2006;154(4):34–41.
the voltage level between supercapacitor and battery. [31] Harada K, Sakai E, Anan F, Yamasaki K. Basic characteristics of electric double-
layer capacitors controlled by switching converters. In: Telecommunications
energy conference 18th international INTELEC ’96. Boston, MA; 1996.
References [32] Dougal R, Liu Shengyi, White RE. Power and life extension of battery-
ultracapacitor hybrids. Components and Packaging Technologies. IEEE Trans
Comp Packag Technol 2002;25(1):120–31.
[1] Chia Yen Yee. Integrating supercapacitor into a hybrid energy storage system
[33] Lima J, Corleta JM, Medeiros A, Canalli VM, Antunes F, Libano FB, Dos Reis FS. A
to reduce overall costs using the genetic algorithm (GA) and support vector
PIC controller for grid connected PV system using a FPGA based inverter. In:
machine (SVM). Malaysia: The University of Nottingham Malaysia Campus;
Industrial electronics, ISIE 2000. Proceedings of the 2000 IEEE international
2014.
symposium, Cholula, Puebla; 2000.
[2] Glavin ME, Chan PKW, Armstrong S, Hurley WG. A stand-alone photovoltaic
[34] Chen Y, Smedley K, Brouwer J. A cost-effective three-phase grid-connected
supercapacitor battery hybrid energy storage system. In: Proc. of power
inverter with maximum power point tracking. In: Industry applications
electronics and motion control conference EPE-PEMC; 2008. p. 1688–95.
conference, 2006. 41st IAS annual meeting. Conference record of the 2006
[3] Glavin ME, Hurley WG. Ultracapacitor/battery hybrid for solar energy storage.
IEEE, Tampa, FL.
In: Proc. of universities power engineering conference UPEC; 2007. p. 791–5.
[35] Gao L, Dougal RA, Liu Shengyi. Power enhancement of an actively controlled
[4] Solero L, Lidozzi A, Pomilio JA. Design of multiple-input power converter for
battery/ultracapacitor hybrid. IEEE Trans Power Electr 2005;20(1):236–43.
hybrid vehicles. IEEE Trans Power Electr 2005;20(5):1007–16.
[36] Ju-Jinn L. Intelligent energy management system for battery supercapacitor
[5] Nielson G, Emadi A. Hybrid energy storage systems for high-performance
hybrid storage in PV systems. Malaysia: The University of Nottingham,
hybrid electric vehicles. In: Proc. of IEEE vehicle power and propulsion
Malaysia Campus; 2011.
conference VPPC; 2011. p. 1–6.
[37] Dinesh S. Utilization of Support Vector Machines (SVM) to forecast load
[6] Saplin A, Meintz A, Ferdowsi M. Parametric study of alternative EV1
requirements in a hybrid battery – supercapacitor energy storage
powertrains. In: Proc. of IEEE vehicle power and propulsion conference
system. Malaysia Campus: The University of Nottingham; 2012.
VPPC; 2008. p. 1–5.
[38] Wang J, Shuyi Sun. Predictive control based on support vector machine model.
[7] Lukic S, Wirasingha SG, Rodriguez F, Cao J, Emadi A. Power management of an
WCICA 2006 In intelligent control and automation. Dalian; 2006.
ultracapacitor/battery hybrid energy storage system in an HEV. In: Proc. of
[39] Hsu CW, Lin Chih-Jen. A comparison of methods for multiclass support vector
IEEE vehicle power and propulsion conference VPPC; 2006. p. 1–6.
machines. IEEE Trans Neural Networks 2002;13(2):415–25.
[8] Lee Sang C, Kwon Osung, Thomas Sobi, Park Sam, Choi Gyeung-Ho. Graphical
[40] Chen B-J, Chang Ming-Wei, Lin Chih-Jen. Load forecasting using support vector
and mathematical analysis of fuel cell/battery passive hybridization with K
machines: a study on EUNITE competition 2001. IEEE Trans Power Syst
factors. Appl Energy 2014;114:135–45.
2004;19(4):1821–30.
[9] Conway BE. Electrochemical supercapacitors: scientific fundamentals and
[41] Chen BJ, Lin CJ, Chang Ming-Wei. EUNITE network competition: electricity load
technological applications. Springer; 1999.
forecasting. Taiwan: Department of Computer Science and Information
[10] Francois Beguin EF. Supercapacitors: materials, systems and applications. John
Engineering National Taiwan University; 2001.
Wiley & Sons; 2013.
[42] Sapankevych N, Sankar Ravi. Time series prediction using support vector
[11] Cao J, Emadi A. A new battery/ultracapacitor hybrid energy storage system for
machines: a survey. IEEE Comput Intell Mag 2009;4(2):24–38.
electric, hybrid, and plug-in hybrid electric vehicles. IEEE Trans Power Electr
[43] Zhang B-L, Dong Zhao-Yang. An adaptive neural-wavelet model for short term
2012;27(1):122–32.
load forecasting. Elect Power Syst Res 2001;59(2):121–9.
[12] Khaligh A, Li Z. Battery, ultracapacitor, fuel cell, and hybrid energy storage
[44] Esen Hikmet, Inalli Mustafa, Sengur Abdulkadir, Esen Mehmet. Modeling a
systems for electric, hybrid electric, fuel cell, and plug-in hybrid electric
ground-coupled heat pump system by a support vector machine. Renew
vehicles: state of the art. IEEE Trans Vehicular Technol 2010;59(6):2806–14.
Energy 2008;33:1814–23.
[13] Bindner H, Cronin T, Lundsager P, Manwell JF, Abdul Wahid U, Baring-Gould I.
[45] Jeong JU, Lee HD, Kim CS, Choi HS, Cho BH. A development of an energy storage
Lifetime modeling of lead acid batteries. Technical University of Denmark;
system for hybrid electric vehicles using supercapacitor. Online technical
2005.
report 2002. URL: http://pearlx.snu.ac.kr/Publication/IBHFESE02.pdf.
[14] Dunlop J, Farhi B. Recommendations for maximizing battery life in
[accessed November 2013].
photovoltaic systems: a review of lessons learned. In: Solar Energy: The
[46] Sapankevych N, Sankar R. Time series prediction using support vector
Power to Choose. Washington DC; 2001.
machines: a survey. IEEE Comput Intell Mag 2009;4(2):24–38.
[15] Yu A, Chabot V, Zhang J. Electrochemical supercapacitors for energy storage
[47] Stojanović Miloš B, Božić Miloš M, Stanković Milena M. Mid-term load
and delivery: fundamentals and applications. In: Electrochemical Energy
forecasting using recursive time series prediction strategy with support vector
Storage and Conversion. America: CRC Press; 2013.
machines. Facta universitatis – series: electronics and energetics
[16] Sharma P, Bhatti TS. A review on electrochemical double-layer capacitors.
2010;23(3):287–98.
Energy Convers Manage 2010;51(12):2901–12.
[48] Cheng H, Tan PN, Gao J, Scripps J. Multistep-ahead time series prediction. Adv
[17] Burke A. Ultracapacitors: why, how, and where is the technology. J Power
Knowledge Discovery Data Mining, Lect Notes n Comput Sci
Sources 2000;91(1):37–50.
2006;3918:765–74.
[18] Christen T, Carlen M. Theory of Ragone plots. J Power Sources
[49] Li Q, Meng Q, Cai J, Yoshino H, Mochida A. Predicting hourly cooling load in the
2000;91(2):210–6.
building: a comparison of support vector machine and different artificial
[19] Zhao X, Sánchez B, Dobson P, Grant P. The role of nanomaterials in redox-based
neural networks. Energy Convers Manage 2009;50(1):90–6.
supercapacitors for next generation energy storage devices. Nanoscale
[50] Li Q, Meng Q, Cai J, Yoshino H, Mochida A. Applying support vector machine to
2011;3:839–55.
predict hourly cooling load in the building. Appl Energy 2009;86(10):2249–56.
[20] Aurbach PD, De-Leon S. USA batteries & fuel cells seminar. San Diego: Plug-
[51] Haykin S. Neural networks: a comprehensive foundation. Prentice Hall
Volt. LLC; 2012.
International; 1999.
[21] Pawan Sharma TB. A review on electrochemical double-layer capacitors.
[52] Han Z, Zhu X, Yang X. Training Method of support vector regression based on
Energy Convers Manage 2010;51(12):2901–12.
multi-dimensional feature and research on forecast model of vibration time
[22] Barker P. Ultracapacitors for use in power quality and distributed resource
series. In: Proc. of international conference on intelligent computation
applications. In: Proc. of IEEE power engineering society summer meeting;
technology and automation (ICICTA); 2010. p. 1087–90.
2002. p. 316–20.
[53] David Sánchez AV. Advanced support vector machines and kernel methods.
[23] Miller JR, Simon P. Patrice electrochemical capacitors for energy management.
Neurocomputing 2003;55(1-2):5–20.
Sci Mag (Open Archive Toulouse Archive Ouverte – OATAO) 2008;321:651–2.
[54] Smola A, Schölkopf B. A tutorial on support vector regression. Statist Comput
[24] Huggins RA. Supercapacitors and electrochemical pulse sources. Solid State Ion
2004;14(3):199–222.
2000;134(1–2):179–95.
[55] Suykens JAK, De Brabanter J, Lukas L, Vandewalle J. Weighted least squares
[25] Zhao Haoran, Wu Qiuwei, Hu Shuju, Xu Honghua, Rasmussen Claus Nygaard.
support vector machines: robustness and sparse approximation.
Review of energy storage system for wind power integration support. Appl
Neurocomputing 2002;48(1–4):85–105.
Energy 2014.

Please cite this article in press as: Chia YY et al. A load predictive energy management system for supercapacitor-battery hybrid energy storage system in
solar application using the Support Vector Machine. Appl Energy (2014), http://dx.doi.org/10.1016/j.apenergy.2014.09.026
 Y.Y. Chia et al. / Applied Energy xxx (2014) xxx–xxx 15

[56] Guyon I, Elisseeff A. An introduction to variable and feature selection. J Mach
Learn Res 2003;3:1157–82.
[57] Cristianini N, Shawe-Taylor J. An introduction to support vector machines
and other kernel-based learning methods. Cambridge University Press;
2000.
[58] Vieira DAG, Takahashi RHC, Palade V, Vasconcelos JA. The Q-norm complexity
measure and the minimum gradient method: a novel approach to the machine
learning structural risk minimization problem. IEEE Trans Neural Networks
2008;19(8):1415–30.
[59] Isa D, Rajkumar R. Pipeline defect prediction using support vector machines.
Appl Artificial Intell 2009;23(8):758–71.
[60] Justino E, Bortolozzi F, Sabourin R. A comparison of SVM and HMM classifiers
in the off-line signature verification. Pattern Recog Lett 2005;26(9):1377–85.
[61] Hassan N, Rajkumar R, Isa D, Arelhi R. Pipeline defect classification by using
non-destructive testing and improved support vector machine classification.
Int J Eng Innovative Technol (IJEIT) 2013;2(7):85–93.
[62] Chapelle O, Haffner P, Vapnik VN. Support vector machines for histogram-
based image classification. IEEE Trans Neural Networks 1999;10(5):1055–64.
[63] Isa D, Lee LH, Kallimani VP, Rajkumar R. Text document preprocessing with the
Bayes formula for classification using the support vector machine. IEEE Trans
Knowledge Data Eng 2008;20(9):1264–72.
[64] Wu C, Chen P. Application of support vector machines in debt to GDP ratio
forecasting. In: Proc. of the international conference in machine learning and
cybernetics; 2006. p. 3412–5.
[65] Sonic PP. Sealed lead acid battery technical handbook power sonic; 2009.
<http://www.batteryweb.com/manuals/techman.pdf>.

Please cite this article in press as: Chia YY et al. A load predictive energy management system for supercapacitor-battery hybrid energy storage system in
solar application using the Support Vector Machine. Appl Energy (2014), http://dx.doi.org/10.1016/j.apenergy.2014.09.026