Available

Available online
online at
at www.sciencedirect.com
www.sciencedirect.com

ScienceDirect
Availableonline
Available onlineatatwww.sciencedirect.com
www.sciencedirect.com
Energy
Energy Procedia
Procedia 00
00 (2017)
(2017) 000–000
000–000

ScienceDirect
ScienceDirect
www.elsevier.com/locate/procedia
www.elsevier.com/locate/procedia

Energy
EnergyProcedia 142
Procedia 00(2017)
(2017)3002–3007
000–000
www.elsevier.com/locate/procedia

9th International Conference on Applied Energy, ICAE2017, 21-24 August 2017, Cardiff, UK

Quality Prediction of Refined Bleached Deodorized Palm Oil

The 15th International Symposium on District Heating and Cooling
(RBDPO) Using Partial Least Square Regression Technique
MohdAssessing
Aiman Mohdthe feasibility
Noor aa ofMohd
, Nur Atikah usingRosely
the aaheat
, Nor demand-outdoor
Adhihah Rashidaa, Yuhan
temperature
Moh aa
function
, Azmer Shamsuddin a,b
for a Kamaruddin
a,b, Mohd long-termAbd.
district
Hamidheat
aa
demand
, Kamarul Asri forecast
Ibrahimaa*
aa
Process
Process System
System Engineering
a,b,c
Engineering Centre
Centre (PROSPECT),
a
(PROSPECT), Universiti
a
Universiti Teknologi
Teknologi Malaysia,
b
Malaysia, 81310
81310 UTM
UTM Johor
Johor cBahru,
Bahru, Johor,
Johor, Malaysia
Malaysiac
bbI. Andrić
Lahad Datu
Lahad
Datu Edible
Edible Oils *, A. Pina , P. Ferrão , J. Fournier ., B. Lacarrière , O. Le Corre
Oils Sdn
Sdn Bhd,
Bhd, KM
KM 2,2, Jalan
Jalan Minyak
Minyak off
off Jalan
Jalan POIC,
POIC, Locked
Locked Bag
Bag 16,
16, 91109
91109 Lahad
Lahad Datu,
Datu, Sabah,
Sabah, Malaysia
Malaysia
a
IN+ Center for Innovation, Technology and Policy Research - Instituto Superior Técnico, Av. Rovisco Pais 1, 1049-001 Lisbon, Portugal
b
Veolia Recherche & Innovation, 291 Avenue Dreyfous Daniel, 78520 Limay, France
c
Abstract Département Systèmes Énergétiques et Environnement - IMT Atlantique, 4 rue Alfred Kastler, 44300 Nantes, France
Abstract

Out
Out ofof specification
specification in in the
the production
production of of Refined
Refined Bleached
Bleached Deodorized
Deodorized Palm Palm Oil
Oil (RBDPO)
(RBDPO) will will cause
cause recycling
recycling in in palm
palm oil oil
refining
refining
Abstract process
process as well as zero production during that period. Hence, the ability to predict the output quality variables so
as well as zero production during that period. Hence, the ability to predict the output quality variables so that
that
appropriate
appropriate adjustment
adjustment can can be be made
made in in thethe process
process beforehand
beforehand is is very
very useful
useful in in minimizing
minimizing the the production
production costcost and
and time
time
consumption.
District heating networks are commonly addressed in the literature as one of the most effective solutions for decreasing to
consumption. In
In this
this study,
study, Multivariate
Multivariate Statistical
Statistical Process
Process Control
Control (MSPC)
(MSPC) is
is used
used to
to develop
develop a
a prediction
prediction tool
tool in
in order
order to
the
predict
predict the
the output
greenhouse output quality
quality variables
gas emissions from thebefore
variables before
building the process
the sector.
processTheseeven
even began
began
systems and
and eventually
eventually
require reduce
reduce the
high investments the utility
whichcost
utility areas
cost as compare
compare
returned to
to traditional
traditional
through the heat
method
method of
of using
sales. Due using tacit knowledge
tacitchanged
to the knowledge which is
is not
not very
whichconditions
climate veryand practical.
practical.
building Knowing
Knowing product
product
renovation quality
qualityheat
policies, in
in advance,
demand adjustment
advance, adjustment
in the futurecan be
be made
can could made at
at the
the
decrease,
specific
specific unit
unit operation
operation and
and also
also
prolonging the investment return period. to
to identify
identify in
in which
which standard
standard the
the quality
quality variables
variables belongs
belongs to,
to, be
be it
it Palm
Palm Oil
Oil Refiners
Refiners Association
Association
Malaysia
The main(PORAM)
Malaysia (PORAM) standard,
paper China
standard,
scope of this China or
or Vietnam.
is to assess Vietnam. The
The MSPC
the feasibility MSPC method
method
of using used in
in this
useddemand
the heat this study
study is
is Partial
– outdoor Partial Least
Least Squares
temperature Squares
functionRegression
Regression (PLS)
(PLS)
for heat demand
which
which capable
forecast. The in
capable in finding
finding
district of the
the relationships
relationships
Alvalade, located between
between
in Lisbon the
the process
process and quality
and was
(Portugal), quality usedvariables
variables
as a case of
of the
the palm
palmThe
study. oil refining
refiningisprocess
oil district process assuming
assuming
consisted of 665
that
that the
the data
buildings are
datathat linear.
arevary in In
linear. In simulation,
simulation,
both constructiondata used
used are
data period areandobtained
obtained
typology. from Lahad
fromThree
Lahad Datu
Datu Edible
weather Edible
scenariosOil
Oil Sdn.
(low,Bhd.
Sdn. Bhd. (LDEO)
(LDEO)
medium, with
andpercentage
with
high) percentage
three districtof
of
fatty
fatty acid
acid (FFA),
(FFA), moisture
moisture content,
content, and
and iodine
iodine value
value as
as both
both the
the process
process and
and quality
quality variable
variable
renovation scenarios were developed (shallow, intermediate, deep). To estimate the error, obtained heat demand values were which
which undergone
undergone pre-screening
pre-screening and
and
pre-processing
compared withof
pre-processing data.
ofresults This
This study
data. from study
a dynamicalso considers
also heat
considers
demand aa constant
constant sampling
sampling time
model, previously time to
to ensure
developedensureand the
the randomness
by the of
randomness
validated of the
the data
authors.data as
as well
well asas the
the
residence
residence time
time in
in determining
determining the
the relationships
relationships of
of the
the quality
quality variables
variables corresponded
corresponded to
to the
the
The results showed that when only weather change is considered, the margin of error could be acceptable for some applications actual
actual process
process variables.
variables. Simulation
Simulation
results
results obtained
obtained
(the error from developed
fromdemand
in annual developed wascorrelation
lower thancoefficient
correlation coefficient
20% for all are evaluated
areweather
evaluated and
and compared
scenarios compared
considered). using Mean
Mean Squared
usingHowever, Squared Error
Error (MSE).
after introducing(MSE). MSE
MSE
renovation
results
results shows
shows that
that PLS
PLS method
method is
is capable
capable in
in predicting
predicting the
the output
output quality
quality variables.
variables.
scenarios, the error value increased up to 59.5% (depending on the weather and renovation scenarios combination considered).
© 2017
©The
2017 The Authors.
The of
value slope Published
Authors. Published
coefficientby by
by Elsevier
Elsevier on
increased Ltd.
Ltd.
© 2017 The Authors.
Peer-review under Published
responsibility of Elsevier
the Ltd.average
scientific committee
within the range of 3.8% up to 8% per decade, that corresponds to the
of the 9th International Conference on Applied Energy.
Peer-review
decrease inunder
the responsibility
number of heatingof the scientific
hours
Peer-review under responsibility of the scientific of committee
22-139h
committee of
during the 9th
of the 9th International
heating Conferenceon
season (depending
International Conference on
ontheApplied Energy.of weather and
combination
Applied Energy.
renovation scenarios considered). On the other hand, function intercept increased for 7.8-12.7% per decade (depending on the
Keywords:
Keywords: palm
palm oil;
oil; quality
quality prediction tool;
tool; partial least
least square regression
coupled scenarios). Theprediction
values suggested partialcould square
be used regression
to modify the function parameters for the scenarios considered, and
improve the accuracy of heat demand estimations.

© 2017 The Authors. Published by Elsevier Ltd.

Peer-review under responsibility of the Scientific Committee of The 15th International Symposium on District Heating and
** Corresponding
Corresponding author.
Cooling. author. Kamarul
Kamarul ‘Asri
‘Asri Ibrahim
Ibrahim
E-mail
E-mail address:
address: asri@utm.my
asri@utm.my
Keywords: Heat demand; Forecast; Climate change
1876-6102 ©
1876-6102 © 2017
2017 The
The Authors.
Authors. Published
Published by
by Elsevier
Elsevier Ltd.
Ltd.
Peer-review
Peer-review under
under responsibility
responsibility of
of the
the scientific
scientific committee
committee of
of the
the 9th
9th International
International Conference
Conference on
on Applied
Applied Energy.
Energy.

1876-6102 © 2017 The Authors. Published by Elsevier Ltd.

Peer-review under responsibility of the Scientific Committee of The 15th International Symposium on District Heating and Cooling.
1876-6102 © 2017 The Authors. Published by Elsevier Ltd.
Peer-review under responsibility of the scientific committee of the 9th International Conference on Applied Energy.
10.1016/j.egypro.2017.12.391
 Mohd Aiman Mohd Noor et al. / Energy Procedia 142 (2017) 3002–3007 3003
2 Nur Atiqah Mohd Rosely/ Energy Procedia 00 (2017) 000–000

1. Introduction

Statistical process control concepts and methods have become very important in the manufacturing and process
industries. Their objective is to improve the processes and product quality by studying the relations between the
dependent and independent variables of the process [3].
PLS regression is a recent technique that generalizes and combines features from principal component analysis and
multiple regression. Its goal is to predict or analyze a set of dependent variables from a set of independent variables
or predictors. Many applications of PLS regression have appeared over the years for prediction and estimation in the
manufacturing industry including chemical, pharmaceutical, and semiconductor processes [1][2].
The recycle process of out-spec RBDPO in the refinery plant is required when product specification are not met. The
recycle process not only resulted in profit loss due to no production with estimated of nearly RM 10,000 per month
to ensure the product quality meets the customer standard. Theoretically, if a statistical model is created to predict
the quality of the end product before the process starts, the possibility of recycle process requirement can be reduced
or removed as recycle process is an insignificant process and an extra expenditure for the company. Their objective
is to improve the processes and product quality by studying the relations between the process (input) and quality
(output) variables of the process [3]. Multivariate statistical process control (MSPC) was used to predict the RBDPO
quality. There are three major standard quality of RBDPO which are PORAM, CHINA and VIETNAM as shown in
Table 1.

Table 1. Standard quality of CPO and RBDPO

Properties CPO RBDPO
CHINA PORAM VIETNAM
Free Fatty Acid, FFA (%) ≤5 ≤0.07 ≤0.10 ≤0.10
Moisture, MOIST (%) ≤0.25 ≤0.10 ≤0.10 ≤0.10
Iodine Value, IV (Wijs) 50-55 50-55 50-55 50-55

Nomenclature

CPO Crude palm oil

FFA Free fatty acid
IV Iodine value
MOIST Moisture Content
PLS Partial Least Square
RBDPO Refine bleached deodorized palm oil

2. Methodology

2.1. Data Collection and Preparation

A total of 148 data with 4 hours(h) time interval were collected from Lahad Datu Edible Oil Company and
prepared as training data set. Standardization of data is done to ensure all the variables have a uniform range of scale
and proceeded with box-plotting in which to find the best sample size from the whole set of data. From the boxplot,
the normality and skewness of the data was checked based on the central limit theorem. Best sample size data was
used for autocorrelation plot to determine process optimum sampling time. Data sorting is conducted according to
the optimum sampling time. The sorted data were used to determine the optimum process resident time using
staggered cross-correlation plot [7].
3004 Mohd Aiman Mohd Noor et al. / Energy Procedia 142 (2017) 3002–3007
Nor Atiqah Mohd Rosely/ Energy Procedia 00 (2017) 000–000 3

2.2. Partial Least Square

Partial least squares (PLS) is a data-driven modeling technique that relates two data matrices (X and Y) by a
multivariate linear model. As a predictive analysis, the multiple linear regressions are used to explain the
relationship between one dependent variable and two or more independent variables. Generally, multiple linear
regressions can be expressed as in Equation (1).

Y  k1 X 1  k 2 X 2    k n X n  e (1)

Where Y is the dependent variable (RBDPO), X is the independent variable (CPO), k is the linear regression or
the relationship coefficients and e is the estimated error. However, in the presence of multicollinear data, PCA will
be applied. PCA reduced the collinearity of data and retained the relevant independent variables for prediction
analysis [8].
Contrary to PLS, the response variable matrix Y is not decomposed in PCA and is regressed with the singular
vectors of X instead to determine a regressive model with the most explained covariance in X, not in Y. PLS
regression remedies PCA low X-Y covariance by decomposing both X and Y matrices into scores and loadings
simultaneously and searching for common PCs that explains as much as possible the covariance between X and Y
and expressed as in Equation (2) and Equation (3).

X CPO  TCPO  PCPO (2)

YRBDPO  U RBDPO  QRBDPO (3)

Where XCPO is the independent variable (CPO), TCPO is score matrix of XCPO and PCPO is loading matrix of XCPO
while YRBDPO is the dependent variable in RBDPO, URBDPO is score matrix of YRBDPO and QRBDPO is loading matrix of
YRBDPO. The scores and loading vectors can be estimated via the singular value decomposition, hence its linear
nature.
In Equation (4), PLS matrix of regression coefficient is the score matrix relationship between CPO and RBDPO.
PLS linear regression coefficient findings is expressed in Equation (5) similar to the general multiple linear
regression equation requirement in Equation (1).

U RBDPO  TCPO  BPLS (4)

k PLS  PCPO  BPLS  PRBDPO (5)

Where kPLS is correlation coefficient using PLS regression method, PCPO is loading matrix of XCPO and BPLS is
matrix of regression coefficient.

2.3. Development of prediction model

The prediction models operated with predictor coefficient, k in forecasting the values of RBDPO properties by
only referring to the properties of CPO as expressed in Equation (6).

_
Y PLS  k PLS X (6)

Where Y is the predicted dependent variable (RBDPO), X is the independent variable (CPO) and kPLS is
correlation coefficient using PLS regression method.
 Mohd Aiman Mohd Noor et al. / Energy Procedia 142 (2017) 3002–3007 3005
4 Nur Atiqah Mohd Rosely/ Energy Procedia 00 (2017) 000–000

3. Results and Discussion

The best sample size was found to be 25 sample size which providing the most similar distributed trend and least
outliers (Figure 1a). From the boxplot, there were a few potential outliers from the FFA input, MOIST input and
output and IV input. All the quality variables were autocorrelated to identify the optimum sampling time by
inspecting the longest lag that fell below the threshold level among the autocorrelation graphs. Based on Figure 1(b),
it was the first lag that fell below the threshold level. So, by multiplying the lag with the time interval of the data
collected, the optimum sampling time observed was 4 h. The staggered cross-correlation technique allowed us to
identify the processing time. By inspecting the staggered cross-correlation graph (Figure 1c), the longest lag that
intercept with the zero y-axis, 3 lags were multiplied to 4 h of optimum sampling time resulted in 12 h optimum
processing time. The data was staggered according to the processing time lag.

Fig. 1. (a) Box plots of 25 sample size; (b) Autocorrelation graph of 25 sample size; (c) Cross-correlation graph of 25 sample size.

The prediction coefficients, k, was determined with the aid of PLS model. This prediction coefficient was utilized
to predict the value of FFA, MOIST and IV of RBDPO. Mean squared error (MSE) technique is a competent tool to
compare the effectiveness and efficiency of the prediction model. MSE is commonly used to measure the quality of
a prediction, the value of the MSE closer to zero is better. The MSE of prediction for training data was shown in
Figure 2. It showed that the MSE of the predicted value was lower and closest to zero. Therefore, it was valid to
conclude that PLS regression method was an adequate analyzing tool to predict the quality of RBDPO. To conduct a
further testing on the PLS regression prediction model, the model was ran using testing data.
3006 Nor Atiqah
Mohd Mohd
Aiman Rosely/
Mohd NoorEnergy
et al. Procedia 00 (2017)142
/ Energy Procedia 000–000
(2017) 3002–3007 5

Fig. 2. MSE of training data and testing data for (a) FFA; (b) MOIST; (c) IV.

The prediction model was validated using the 120 industrial data (testing data) from Lahad Datu Edible Oils Sdn.
Bhd. to validate the prediction model developed. The testing data MSE was expected to have greater values than that
of the training data MSE as the prediction model was developed based on the training data (Figure 2). Even though
the MSE for FFA and IV was slightly higher for testing data, the value is close to zero, thus it was proven that PLS
method was effective in predicting the quality of RBDPO . Control charts were plotted to determine which standard
quality the RBDPO belong to and to observe the process behavior. The control charts were plotted for the actual
output and predicted output of testing data as shown in Figure 3. Despite of having some out-control process; the
prediction was acceptable as long as the value did not exceed the standard limit. As shown in the figure 3(a), there
was no point fall outside the process control limit (red solid line), but the predicted quality was belong to the
CHINA standard quality (<0.07) in contrast to the actual quality which belong to the PORAM and VIETNAM
standard quality. With implementation of these control chart, refinery will able to avoid recycle process due to
ability to predict possible standard quality and monitors process performance.

Fig. 3. Control chart of predicted and actual value of RBDPO (a) FFA; (b) MOIST; (c) IV.
 Mohd Aiman Mohd Noor et al. / Energy Procedia 142 (2017) 3002–3007 3007
6 Nur Atiqah Mohd Rosely/ Energy Procedia 00 (2017) 000–000

Conclusion

In this study, a PLS-based prediction model was developed. 25 best sample size with optimum sampling time and
processing time of 4 h and 12 h, was used to develop the prediction coefficient. The MSE result for both training
data and testing data were fairly good and comparable. Therefore, the prediction of RBDPO quality using PLS was
valid. By imposing the control and quality prediction in palm oil refinery plants, many kinds of quality and practical
problems can be solved as stated before. These features will benefits industry in the form of monetary and energy
savings if it were to be implemented. However, the model was developed with the assumption of the process to be
linear, hence non-linear model will be studied in future work.

Acknowledgements

The financial support from Universiti Teknologi Malaysia (RUGS Tier 1 Q.J130000.2546.12H67) and (RUGS
Tier 2 Q.J130000.2646.12J39) are highly acknowledged.

References

[1] Madanhire, I. and Mbohwa, C. (2016). Application of Statistical Process Control (SPC) in Manufacturing Industry in a Developing Country.
Procedia CIRP. 40, 580-583.
[2] McAvoy, T. (2002). Model predictive statistical process control of chemical plants. Industrial & engineering chemistry research. 41 (25),
6337-6344.
[3] Nakamura, N., Takahashi, S., Kusumoto, S. and Nakatsuka, K. (2011). Approach to Introducing a Statistical Quality Control. Proceedings of
the 2011 Software Measurement, 2011 Joint Conference of the 21st Int'l Workshop on and 6th Int'l Conference on Software Process and
Product Measurement (IWSM-MENSURA), 297-301.
[4] Reiss, R., Wojsznis, W., Wojewodka, R., 2010. Partial least squares confidence interval calculation for industrial end-of-batch quality
prediction. Chemometrics and Intelligent Laboratory Systems, 100(2), pp.75-82.
[5] Vanlaer, J., Gins, G., Van Impe, J., 2013. Quality assessment of a variance estimator for Partial Least Squares prediction of batch-end quality.
Computers & Chemical Engineering, 52, pp.230-239.
[6] Kemalbay, G., Korkmazoglu, O., 2012. Effects of Multicollinearity on Electricity Consumption Forecasting using Partial Least Squares
Regression. Procedia - Social and Behavioral Sciences, 62, pp.1150-1154.
[7] Rosely N.A.M., Rashid N.A., Noor M.A.M., Hawi N.D.A, Sepuan S.Q., Shamsuddin A., Ibrahim K.A. and Hamid M.K.A., Product sampling
time and process residence time prediction of palm oil refining process, Chemical Engineering Transactions, 2017;56 , 1411- 1416
DOI:10.3303/CET1756236.
[8] Kumar, A. and Goyal, P. Forecasting of air quality in Delhi using principal component regression technique. Atmospheric Pollution
Research,2011; 2(4) pp.436-444.