Computers and Chemical Engineering 103 (2017) 201–209

Contents lists available at ScienceDirect

Computers and Chemical Engineering

journal homepage: [Link]/locate/compchemeng

Full Length Article

Thermodynamic optimization of atmospheric distillation unit

Funmilayo N. Osuolale a,b,∗ , Jie Zhang a
a
School of Chemical Engineering and Advanced Materials, Newcastle University, Newcastle upon Tyne, NE1 7RU, UK
b
Department of Chemical Engineering, Ladoke Akintola University of Technology, Ogbomoso, Nigeria

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

Article history: This paper presents a methodology for optimising the exergy efficiency of atmospheric distillation unit
Received 12 September 2016 without trading off the products qualities and process throughput. The presented method incorporates
Received in revised form 27 March 2017 the second law of thermodynamics in data driven models. Bootstrap aggregated neural networks (BANN)
Accepted 29 March 2017
are used for enhanced model accuracy and reliability. The standard error of the individual neural network
Available online 31 March 2017
predictions is taken as the indication of model prediction reliability and is incorporated in the optimiza-
tion objective function. The economic analysis of the recoverable energy (sum of internal and external
Keywords:
exergy losses) reveals the energy saving potential of the proposed method, which will aid the design and
Atmospheric distillation
Exergy
operation of energy efficient atmospheric distillation columns.
Optimization © 2017 Elsevier Ltd. All rights reserved.
Neural networks

1. Introduction been applied on a number of processes Li et al. (2012) with a com-

bination of differing objectives (Osuolale and Zhang, 2015). It has
Distillation process has always attracted the interest of also find application in improving the energy efficiency of the crude
researchers and quite a number of work in the literatures are distillation unit albeit in terms of the utility consumption alone
focused on improving its energy efficiency either for binary system (Ochoa-Estopier et al., 2013).
Osuolale and Zhang (2014) or multi-component system (Al-Mutairi Previous studies on the energy efficiency of crude distilla-
and Babaqi, 2014). Most success in improving the efficiency of tion units assert the need for the energy improvements methods
the column are the development of energy integrated schemes of the column to be based on second law efficiency Nur Izyan
such as the petyluk column (Amminudin et al., 2001), internally and Shuhaimi (2014) and exergy loss profiles of the column
heat integrated column (Nakaiwa et al., 2003), heat integrated (Tarighaleslami et al., 2012). This is because efficiency based on
distillation column HIDC (Amiya, 2010), thermally coupled divid- second law provides means of understanding the improvement
ing wall column Caballero and Grossmann (2013) and intensified capability of a system while identifying and quantifying the sources
distillation column (Vazquez-Castillo et al., 2009). However, the of inefficiency within the system. For example, past works on the
conventional columns are still dominantly in use in most chemical second law analysis (exergy) reveals low efficiency of some crude
and petrochemical industries. Most industries will rather maximize distillation systems. Cornelissen (Cornelissen, 1997) performed
available resources for maximum profitability in place of large scale analysis on a crude distillation unit with an exergy efficiency of
expansion. One tool that has gained widespread application in this 0.27 for an atmosphere distillation unit (ADU), while Al-Muslim
regard is optimization. Optimization is a major quantitative tool and Dincer (Al-Muslim and Dincer, 2005) came out with a result
in decision making for the process industries (Osuolale and Zhang, of 0.433 for an ADU. The first law analysis (energy efficiency) are
2016). In recent years optimization of crude distillation system has always higher than the second law analysis in all cases. This could
received considerable research interests. This is because the major imply that improvement of the efficiency of the column will be bet-
cost of refinery operation second only to the cost of crude is energy ter off when it is based on the second law. Even though there have
and 35% of these is consumed in the crude distillation unit. Opti- been optimization procedures for the energy improvement of the
mization is a well-developed field in chemical engineering and has crude distillation process they are often formulated by using energy
consumption Kiss and Ignat (2013) and/or plant economic Xu et al.
(2013) as the objective functions. This might give a fictitious sense
of accomplishment since energy based approach relies only on the
∗ Corresponding author at: Department of Chemical Engineering, Ladoke Akintola
first law of thermodynamics. The challenge is to develop ways of
University of Technology, Ogbomosho, Nigeria.
E-mail addresses: fnosuolale@[Link], funmikorede@[Link] incorporating exergy in an optimization frame work especially for
(F.N. Osuolale), [Link]@[Link] (J. Zhang). real time optimization. This work presents a data driven modelling

[Link]
0098-1354/© 2017 Elsevier Ltd. All rights reserved.
202 F.N. Osuolale, J. Zhang / Computers and Chemical Engineering 103 (2017) 201–209

Table 1
Nomenclature TBP distillation curve.

Assay Percent Temperature (K)

API American petroleum institute
2.68 309
Ex Exergy rate, kJ h−1
7.2 366.5
H Total enthalpy, kJ h−1 15 422
m Molar flow, k mol h−1 24.5 477.6
PA Pump around 33.31 533.1
PAT Change in pump around temperature 44.70 588.7
49.60 616.5
Po Reference pressure, kpa
59.14 672
S Specific entropy, kJ h−1 K−1 75.22 783
SS Side stripping 84.46 866.5
TBP True boiling point 95.81 1023
T0 Reference temperature, K
Wb Network weight for bootstrap sample
xi Neural network input
catalytic cracking unit respectively. This paper is focused on the
␴e Standard prediction error
atmospheric distillation unit highlighted in Fig. 1.
␸ Exergy efficiency, %

Subscript 2.2. HYSYS simulation

o Reference conditions
phy Physical A process simulation software HYSYS is used for the modelling
chem Chemical and simulation of the atmospheric distillation unit. The procedure
for the simulation is further highlighted as follows. The crude is
described by the TBP distillation curve given in Table 1 and the
approach for the optimization of an atmospheric distillation col- light ends properties in Table 2. The light ends basis in the assay is
umn based on first and second laws analysis aimed at minimizing 16.8%.
the inefficiencies without compromising the qualities of the prod-
ucts. In addition, in order to improve the optimization reliability
especially for application to real time optimization, minimization 2.2.1. Modelling and simulation of the ADU
of model prediction confidence bounds is incorporated in the objec- HYSYS contains a number of components in its data bank. The
tive function in addition to maximize the exergy efficiency of the components are well defined with its thermodynamic and physi-
column. cal properties, temperature dependent properties such as enthalpy
The paper is organized as follows: Section 2 gives a description and critical properties. The light end components of the refineries
of the atmospheric distillation unit and methods of analysis of the were inputted here. It should be noted that the given components
system such as exergy analysis and economic analysis. Section 3 are from the laboratory analysis of the raw crude. Other unknown
deals with bootstrap aggregated neural network (BANN) modeling components of the crude are determined from the crude char-
of the ADU and it also gives the optimization technique. Section 4 acterization in HYSYS. The property package in HYSYS includes
concludes the paper. Equation of states (EOSs), activity models, Chao Seadre models
and vapour pressure models. One of the property package in EOS
is Peng-Robinson. It is chosen as it is properly suited crude oil
2. Atmospheric distillation unit
analysis. The Peng-Robinson method utilises EOS in its enthalpy
calculations. Crude oil is a mixture of many identified chemi-
2.1. Description of the system
cal components and pseudo-components whose chemical identity
might be difficult and sometimes impractical to determine. Hence,
Crude at its raw state is a relatively low value material but when
there is a need for the characterizations of the crude. The crude
refined could yield products whose value is many times that of
was characterized using experimental assay that include the bulk
the original crude. Refining of crude is in two stages of distillation:
crude properties, light end volume percent, ASTM distillation, API
the atmospheric distillation unit and the vacuum distillation unit.
gravity and TBP distillation. The result of the characterization is
The crude distillation unit under consideration is diagrammatically
detailed chemical compositions of the identified components and
represented in Fig. 1 García-Herreros and Gómez (2013) with a
the pseudo-components. The complete and definitive analysis of a
pre-flash column incorporated into the system. Preheated crude at
crude oil is called crude assay. This is more detailed than a crude
343 ◦ C and 344 kPa is introduced at its flash point to the ADU. Super-
TBP curve.
heated steam is injected at the bottom to enhance vaporization and
separation of the crude. The column processes 85,000 barrel/day of
crude into 6 products: Light Naphtha, Heavy Naphtha, Kerosene,
Table 2
Diesel, Atmospheric gas oil (AGO) and Residue. Three side stream
Light ends assay.
products are taken from the side strippers. The ADU is equipped
with three pump around that recover heat for the preheat trains. Light ends Composition (mass%)
The bottom product (residue) is the residual liquid material which Propane 0.7595
could not be vaporized under the existing operating conditions of i-Butane 0.5622
temperature and pressure in the tower. It is further fed to another n-Butane 0.1567
i-Pentane 1.173
tower which operates at subatmospheric pressure; the vacuum n-Hexane 4.203
tower (Kaes, 2000). The side products of the vacuum distillation n-Heptane 1.308
unit (VDU) are Light vacuum gas oil (LVGO) and Heavy vacuum gas n-Octane 5.475
oil (HVGO). The overhead product and the bottom product could n-Nonane 2.939
n-Pentane 0.2167
serve as inlets to other processing units such as stabilisation and
 F.N. Osuolale, J. Zhang / Computers and Chemical Engineering 103 (2017) 201–209 203

Fig. 1. Schematic diagram of the atmospheric distillation unit.

Table 3 2.3. Analysis of system

ADU product specifications.

Products Specifications (K) 2.3.1. Exergy analysis

Light Naphtha ASTM D86 95% = 384 The total exergy of a stream is given as
Heavy Naphtha ASTM D86 95% = 455
Kerosene ASTM D86 95% = 541 Extotal = Exphy + Exchem (1)
Diesel ASTM D86 95% = 599
AGO ASTM D86 95% = 710 where Exphy and Exchem are physical exergy and chemical exergy
respectively.
The physical exergy is calculated as

Exphy = H − H0 − T0 (S − S0 ) (2)

2.2.2. Building the flow sheet where H is the total enthalpy, S is the total entropy and T is the
The modelling of the crude distillation units were done in the temperature. The subscript 0 denotes reference conditions
HYSYS environment using their operating and design parameters. For the crude stream considered, standard molar chemical
The simulation was done to be prototype of the actual process exergy Exchem is calculated from the standard molar chemical exer-
as much as possible in terms of these parameters: the number of gies of all identified components and pseudo-components as
trays, feed tray, feed temperature, feed flow rates, heat exchangers    
supply and target temperatures, product specifications, steam flow Exchem = m bqk + bchk + RT0 lnak (3)
rates, and pump around flow rates. Data such as entropy, enthalpy,
temperatures, pressures, compositions and stream flow rates were Where bchk is the chemical exergy for component k, bqk is the
extracted from the simulation for exergy analysis. chemical exergy for pseudo component k, and ak is the activ-
In the HYSYS simulation, the specifications of the products are ity coefficient of component k. The standard chemical exergy for
added to the streams as given in Table 3 and the quality of the pseudo-components can be determined for heuristic empirical
ADU product is maintained by ensuring the temperature difference expression as a function of the elementary composition and their
between the 95% vol. and the 5% vol. of the ASTM D86 of two con- heating values (Szargut et al., 1988).
secutive products is within the acceptable limit (Jones et al., 1999).
The ASTM curve for the product of ADU is shown in Fig. 2. bqk = ϑk Ck (4)
204 F.N. Osuolale, J. Zhang / Computers and Chemical Engineering 103 (2017) 201–209

800
AGO
Diesel
700
Kerosene

ASTM D86 Temperature (oC

Heavy Naphtha
600
Residue
Light Naphtha
500

400

300

200

100

-100
0 10 20 30 40 50 60 70 80 90 100
Cut point (%)

Fig. 2. ASTM D86 of end products of the ADU.

Where ϑk is the regression equation to express the ratio H/C, N/C, Table 4 gives the results of the physical exergy of the ADU.
O/C and S/C for pseudo-component k, Ck is the net calorific heating
value of the pseudo-component k (Rivero et al., 1999). From Fig. 1
2.3.2. Economic analysis
the exergy for the inlet and outlet streams for the ADU are given as
The total cost of the column is given as

Exout = ExL Naphtha + ExH naphtha + Exkero + Exdiesel Total cost = Energy cost + Capital cost (9)

+ExAGO + Exresidue (5) In the above equation,


Energy cost = Qx Cx (10)

Exin = ExPreflash liquid + Exsteam + Exfurnace + Exkero steam
where Qx is the duty of utility x and Cx is the unit cost of utility x.
+Exdiesel steam + ExAGOsteam (6) While assuming that the retrofit of the atmospheric distillation
unit might only involve adding or removing heat exchangers as the
case may be, the capital cost estimation of the ADU is based on the
The exergy efficiency and the irreversibility are then calculated heat exchanger cost. Hence the capital cost is given as (Al-Mutairi
as:
 and Babaqi, 2014)
Ex
ϕ=  out (7) Capital cost = Heat exchanger cost = A + B(area)c (11)
Exin
  where A is the fixed cost of installation and B is exchanger cost per
I= Exin − Exout (8)
unit area.

Table 4
Simulation data for exergy analysis.

Stream Name h(kJ/kmol) h0 (kJ/h) s(kJ/h K) s0 (kJ/h K) m(kmol/h) Ex

Inlet streams
Crude inlet −271082 −463266 730.2 309.4 2227 1.487 × 108
Crude Steam −230832 −286232 186.5 53.66 125.9 1.993 × 106
Kero Steam −230832 −286232 186.5 53.66 88.13 1.395 × 106
Diesel Steam −230832 −286232 186.5 53.66 62.95 9.964 × 105
AGO steam −230832 −286232 186.5 53.66 18.88 2.988 × 105
Furnace duty 6.792 × 107

TOTAL IN 2.213 × 108

Outlet streams
L Naphtha −162826 −194812 187.17 96.83 594.99 3.013 × 106
H Naphtha −219225 −261920 218.21 103.10 504.00 4.228 × 106
Kerosene −286630 −356746 338.35 158.68 423.00 7.010 × 106
Diesel −346029 −476298 578.41 267.78 177.00 6.674 × 106
AGO −386917 −599476 871.93 396.66 184.00 1.305 × 107
Residue −586313 −992013 1677.13 805.88 479.27 7.001 × 107

TOTAL OUT 1.040 × 108

 F.N. Osuolale, J. Zhang / Computers and Chemical Engineering 103 (2017) 201–209 205

Table 5 data. The network giving the lowest sum of squared errors (SSE) on
Feed, products and utility prices.
the testing data is considered as having the appropriate number of
Item Cost Unit hidden neurons. The Levenberg-Marquardt optimization trainning
Crude oil 80 $/bbl algorithm with regularisation and cross validation based “early
Light Naphtha 44.3 $/bbl stopping” was used to train the network. The final developed neural
Heavy Naphtha 136 $/bbl network model is evaluated on the unseen validation data. Because
Kerosene 122.7 $/bbl of the different magnitudes of the input and output data of the
Diesel 121.7 $/bbl
model, the data for the training, testing and validating were scaled
Atmospheric gas oil 95.29 $/bbl
Residue 89.71 $/bbl to be in the range [−1,1].
Fired heating 150 $/kJ For enhanced model accuracy, boostrap aggregated neural net-
Cooling water 5.25 $/kJ work Zhang (1999) is used in this study. In a bootstrap aggregated
Stripping steam 0.14 $/kmol
neural network model, several neural network models are devel-
oped to model the same relationship. Individual neural network
For a stainless steel shell and tube heat exchanger, models are developed from bootstrap re-sampling replications of
the original training data. Instead of selecting a single neural work
Heat exchanger cost = 33422 + 1784(area)0.81 (12) that is considered to be the “best”, several networks are combined
The method of improvement being proposed here does not together to improve model accuracy and robutsness. These mod-
include a change in any of the equipment and hence the capital els can be developed on different parts of the data set. One further
cost remains the same. Basically the economic analysis is based on advantage of BANN is that model prediction confidence bounds can
the operating profit of the column and expressed mathematically be calculated from individual network predictions. The confidence
as interval gives an indication of the prediction reliablity. The standard
error of the ith predicted value is calculated as

n

n
Profit = Mj Cj − [Mcrude Ccrude + Msteam Csteam + Qx Cx ] (13)  1

1  
n 2
j=1 x=1 2
= y xi ; W b − y (xi ; ) (14)
n−1
Where Mj and Cj are the flow rate and cost of the jth product, Msteam b=1
and Csteam are the flow rate and cost of steam respectively, Qx and
Cx are the heat requirement of utility and the cost of utility respec- 
n

tively. The calculation is based on the assumption of 8600 h per where y (xi ) = y xi ; W b /n and n is the number of neural net-
year. In Table 5, the feed, products and utility prices are shown b=1
(Energy Information Administration, 2014). works. The 95% prediction confidence bounds can be calculated as
y (xi ; ) ± 1.96. It indicates a 95% certainty level that the interval
3. Modelling of the atmospheric distillation unit will contain the true process output with a probabilty of 0.95. A nar-
rower confidence bounds indicates the associated model prediction
3.1. Data collection and processing is more reliable.

The inputs for the data driven model are the flow rates of all 3.2. BANN model of the column
the ADU products, the three side strippers flow rates, the pump
around rates and change in pump around temperatures. The input Bootstrap aggregated neural network models each containing
are chosen on the basis of being variables that can be manipulated 30 neural networks were developed for predicting the exergy effi-
and also impact the heat recovery of the process and hence the ciency and the product qualities of the ADU. Each network has a
energy efficiency. The crude flow rate and hence its stripping steam single hidden layer with 20 hidden neurons. It has been proved
were not considered as model inputs since they are constant for the that a single hidden layer neural network with the sigmoid activa-
case being studied. If the data for the modelling are to be taken from tion function in the hidden layer can approximate any continuous
actual plant operating conditions and are found to be varied, then nonlinear function as long as the number of hidden neurons is suffi-
they should be considered as additional model inputs. The outputs cient (Cybenko, 1989). Sigmoid activation function was used for the
are the products qualities and the exergy efficiency of the system. hidden neurons and linear activation function for the output neu-
Data for ANN models can be either obtained from process oper- ron. The 30 neural networks are different in that they are trained
ation parameters Bahar et al. (2004) or from other models (Basak with different training data sets, which are obtained as bootstrap
et al., 2002). In this case study, simulated process operational data re-sampling replications of the original training data set, and they
were generated from HYSYS (V8.2). Variables were varied within have different initial weights during training. Fig. 3 shows the pre-
their lower and upper bounds in a nested loop. A total of 2048 data dicted and the actual values. The predicted and the actual values are
samples were generated. The chemical and physical exergies of all super imposed on each other. This shows the BANN model almost
the inlets and outlets streams for each generated data were calcu- perfectly mimic the actual data for the training, testing and vali-
lated (Eqs. (1)–(4)) and the corresponding exergy loss and exergy dation. This gives a measure of confidence on the reliability of the
efficiency were evaluated (Eqs. (7) and (8)). The data were divided model. In Fig. 4, the prediction error between BANN model and
into training data (50%), testing data (30%), and unseen validation actual process data is shown for the training, testing and validation
data (20%). The training data is used for network training and the data set. BANN model predictions can be said to be very accurate.
testing data is used for network structure selection (number of hid- Fig. 4 shows the scaled MSE of the individual networks. It can be
den neurons) and “early stopping” in network training. With the seen from Fig. 4 that a network with small training MSE may have
“early stopping” mechanism, neural network prediction errors on quite large MSE on the validation data. This indicates inconsistency
the testing data are continuously monitored during training and and non-robust nature of the individual networks. Table 6 shows
training is terminated when the prediction errors on the testing the MSE of the bootstrap aggregated neural network models and
data do not futher reduce. The number of hidden neurons was the standard errors from the individual network predictions on the
determined by building a number of neural networks with differ- training and validation data sets. The model accuracy is seen to be
ent numbers of hidden neurons and testing them on the testing improved by combining the imperfect models (Fig. 5).
206 F.N. Osuolale, J. Zhang / Computers and Chemical Engineering 103 (2017) 201–209

Fig. 3. BANN model of the ADU.

-3
x 10
0.8 5

4
0.7

3
0.6
Predicted exergy efficiency

2
Prediction errors

0.5 1

0.4 0

-1
0.3
-2

0.2
-3

0.1 -4
0 0.2 0.4 0.6 0.8 0 500 1000 1500 2000
Actual exergy efficiency Samples

Fig. 4. BANN predicted vs. actual exergy efficiency (left) and prediction errors (right) for training, testing and validation data.

Table 6 3.3. Optimization techniques

Bootstrap aggregated neural network prediction accuracy for the ADU.

Mean square error Standard prediction error 3.3.1. Optimization of exergy efficiency based on BANN model
Testing data 4 × 10−6 6 × 10−6 The objective of the optimization here is to maximize the exergy
Validation data 1.6 × 10−5 1.9 × 10−5 efficiency of the atmospheric distillation unit. There are a number
of decision variables in the ADU which finding their optimal can
significantly improve the exergy efficiency of the ADU. The model
It can be seen that individual networks give inconsistent per- prediction confidence bound, which depends on model input vari-
formance on the training and unseen validation data indicating the ables, is incorporated in the optimization problem. By penalizing
non-robust nature of single networks. BANNs give consistent and wide model prediction confidence bounds, the optimization results
more accurate prediction performance on the training and unseen are forced into the region where model predictions are reliable.
validation data sets.
 F.N. Osuolale, J. Zhang / Computers and Chemical Engineering 103 (2017) 201–209 207

-5
x 10
2

MSE(training)
1

0
0 5 10 15 20 25 30 35
-5
x 10
4
MSE(testing)

0
0 5 10 15 20 25 30 35
-4
x 10
MSE(Validation)

0.5

0
0 5 10 15 20 25 30 35
Network No

Fig. 5. Model error of individual networks for the crude distillation unit.

The optimization formulation is given as Table 7

Summary of optimization results with different weightings on confidence bounds.
max
J = ϕ − ˇ (15) Items Base ␤=0 ␤=1 ␤ = 0.5 LB UB Unit
x
AGO 184 170 170 170 165 190 barrel/h
s.t. Diesel 177 160 160 160 160 184 barrel/h
Kerosene 423 410 410 410 405 435 barrel/h
␸ = U(x) L Naphtha 595 585 585 585 585 605 barrel/h
H Naphtha 504 490 490 490 480 510 barrel/h
where J is the objective function, ϕ is the exergy efficiency, U is PA1 1588 1587 1587 1454.7 1349 1826 barrel/h
◦
the BANN model of the ADU, x = (x1, x2, . . . xn )is a vector of decision PA1T 50 40 55 40 40 60 C
PA2 997 995.7 997.4 848 848 1147 barrel/h
variables which are the flow rates of light naphtha, heavy naphtha, ◦
PA2T 70 60 76.2 60 60 80 C
kerosene, diesel, AGO, residue, AGO SS steam, Diesel SS steam, Kero PA3 451 449.8 450 383 383 518 barrel/h
SS steam, PA1, PA2, PA3, and PA1T, PA2T, PA3T, ␴ is standard PA3 T 70 60 60 60 60 80 ◦
C
prediction error for BANN prediction, and ␤ is the weighting factor AGO steam 18.8 15 19 15 25 kmol/h
of ␴. Here PA represents pump around and PA1T represents the Diesel steam 62.9 55 57 55 70 kmol/h
Kero steam 88.1 80 80 92 95 kmol/h
change in temperature in the pump around.
Optimum efficiency 79.83 69.01 83.67 %
The optimisation problem was solved using the sequential HYSYS validated 46.99 77.47 69.61 83.04 %
quadratic programing (SQP) implemented by the function “fmin- Error 0.0304 0.0087 0.0075
con” in MATLAB Optimisation Toolbox. Table 7 shows the optimum
and HYSYS validated results with varying weighting factors of the
confidence bounds. The approach in this work is to improve model tradeoff between the two terms can be achieved. One could start
prediction reliability using BANN and to provide a model predic- with a small value and gradually increase it until the reduction in
tion confidence bounds which is then penalized in the optimization the standard error become small.
objective function. The effect of penalization of the wide model pre- Table 7 Crude distillation operations are often bounded by prod-
diction confidence bounds during the optimization can be clearly uct quality specifications. An optimal efficiency procedure without
seen in Table 7. The result of the optimization without including consideration to the quality specification might not be feasible in
the confidence bounds is included for the purpose of comparison. practice. For petrochemical system where it is not possible to give a
The relative error is calculated as the difference between BANN discrete component specifications as a measure of the product qual-
and HYSYS model predictions divided by the HYSYS model predic- ity, the 95% vol and 5% vol of the ASTM distillation is often use as a
tion. The method results in much less relative error between the
BANN model and HYSYS model. This indicates the reliability of the
Table 8
proposed model because the performance on the actual process ASTM distillation of ADU products.
(HYSYS simulation) is close to that predicted by the neural net-
work model. The improved reliability of the optimization results A5% ◦ C D5% ◦ C K5% ◦ C L5% ◦ C H5% ◦ C R5% ◦ C

is due to the incorporation of model prediction confidence bounds Base case 266.3 230.9 155.5 −46.62 93.67 775
in the optimization objective function. The selection of the weight- Optimized 259.5 225 153.8 −46.84 86 772
ing parameter, ␤, depends on the relative magnitudes of the two A, D, K, L, H, R are AGO, Diesel, Kerosene, Light Naphtha, Heavy Naptha and Residue
terms in the objective function. It should be selected so that a respectively.
208 F.N. Osuolale, J. Zhang / Computers and Chemical Engineering 103 (2017) 201–209

Table 9
Optimization results (efficiency and profit).

Exergy (%) Irreversibility (kJ/h) Steam cost Energy cost ($/yr) Profit ($/yr)

Base case 46.99 1.174 × 108 3.562 × 105 3.456 × 106 3.030 × 109
Optimum 83.04 6.478 × 107 3.322 × 105 3.193 × 106 3.055 × 109

guide (Jones, 1999). The quality of the products as%vol of the ASTM conditions. Product quality constraints on the column also intro-
distillation for the base case and optimized case are given in Table 8. duced a measure of penalization on the optimization result to
Quite often, optimization of distillation process is performed sub- give as close as possible to what obtains in reality. The proposed
ject to one or two distillate qualities (Gadalla et al., 2013). The technique can significantly improve the second law efficiency of
method proffered here allows for inclusion of as many product the system with an additional economic advantage. Most methods
qualities as desired. A number of methods exist in literature for found in the literature are based only on the first law analysis of the
online monitoring of chemical processes Shahnovsky et al. (2012) column. This might give a fictitious sense of achievement. The pro-
especially crude distillation process qualities (Shang et al., 2014). posed modeling and optimization strategy can aid in the operation
The method as proposed here could predict the product qualities and design of energy efficient column. The proposed method can
and as well predict the optimum operating exergy efficiency of the also be applied to other systems as long as the operational data of
column. Stringent requirement of petroleum quality demands the the systems are available and the exergy efficiency of the systems
need to monitor and control the quality at all times. One approach is can be modelled and then optimized.
to use off-line laboratory analysis at periodical intervals. This could
result in massive time delay and sometimes manual efforts. The
other approach of online analyser could be expensive and difficult Acknowledgement
to predict its exactness. These side effects necessitate the need for
data driven models for monitoring the product quality. There is a The work was supported by the Commonwealth scholarships
possibility of developing this method further to monitor the prod- commission in the UK and partially supported by the EU FP7 (Ref:
uct quality with the added advantage of predicting the efficiency PIRSES-GA-2013-612230).
of the system. It could have applications in process monitoring,
advanced control and fault diagnosis. In Table 8, the 95% vol and
5% vol of the ASTM distillation for the base case and the optimized References
cases are shown.
Al-Muslim, H., Dincer, I., 2005. Thermodynamic analysis of crude oil distillation
In Table 7, the least relative prediction error is at when the systems. Int. J Energy Res. 29, 637–655.
weighting on confidence bounds is 0.5. Using this optimum case Al-Mutairi, E.M., Babaqi, B.S., 2014. Energy optimization of integrated atmospheric
for the economy analysis of the optimization result, the optimiza- and vacuum crude distillation units in oil refinery with light crude. Asia-Pacific
J. Chem. Eng. 9, 181–195.
tion resulted in 76.71% increment of the exergy efficiency as shown Amiya, K.J., 2010. Heat integrated distillation operation. Appl. Energy 87,
in Table 9. This translates to 44.8% decrease in irreversibility loss in 1477–1494.
the column and 7.6% reduction in energy costs of the column with Amminudin, K.A., Smith, R., Thong, D.Y.C., Towler, G.P., 2001. Design and
optimization of fully thermally coupled distillation columns: part 1:
reference to their initial values. Every real process has an element of
preliminary design and optimization methodology. Chem. Eng. Res Des. 79,
irreversibility and often time the performance of engineering sys- 701–715.
tem is degraded by their presence. With the methodology proffered Bahar, A., Özgen, C., Leblebicioǧlu, K., Halıcı, U., 2004. Artificial neural network
estimator design for the inferential model predictive control of an industrial
from this study, ways of considerably reducing the irreversibility
distillation column. Ind. Eng. Chem. Res. 43, 6102–6111.
of the system as well as determining the efficiency of the process is Basak, K., Abhilash, K.S., Ganguly, S., Saraf, D.N., 2002. On-line optimization of a
made easy. This will be a good tool in the hand of process and design crude distillation unit with constraints on product properties. Ind. Eng. Chem.
engineers for the operation of energy efficient column. It could Res. 41, 1557–1568.
Caballero, J.A., Grossmann, I.E., 2013. Synthesis of complex thermally coupled
equally find relevance in the control of the column for improving distillation systems including divided wall columns. AIChE J. 59, 1139–1159.
efficiency. Cornelissen, R.L., 1997. Thermodynamics and Sustainable Development, The Use Of
The total profit is increased by $26.3 × 106 /year. The increment Exergy Analysis And The Reduction Of Irreversibility. PhD. University Of
Twente.
is mainly due to the optimum operating conditions from the exergy Cybenko, G., 1989. Approximation by superpositions of a sigmoidal function. Math.
based analysis of the column. The reduction in the cost of energy Control Signal Syst. 1989 (2), 303–314.
contributed to about 0.8% in the total profit. The results show that Energy Information Administration, 2014. Energy Information Administration.
[Link].
considerable economic benefit of the column can be achieved at no Gadalla, M., Kamel, D., Ashour, F., El Din, H.N., 2013. A new optimisation based
additional cost of equipments. retrofit approach for revamping an egyptian crude oil distillation unit. Energy
Proc., 454–464.
García-Herreros, P., Gómez, J.M., 2013. Modeling and optimization of a crude
4. Conclusions distillation unit: a case study for undergraduate students. Comput. Appl. Eng.
Educ. 21, 276–286.
Jones, A., O’Donnell, M., Terndrup, H., 1999. Crude unit control and optimisation at
The challenge for the design and optimization of crude distil- Grangemouth refinery. Comput. Control Eng. J. 10, 209–213.
lation system involves developing models that are reliable and Jones, D.S.J., 1999. Elements Of Petroleum Processing. Wiley.
computationally efficient. ANN models of distillation column has Kaes, G.L., 2000. Refinery Process Modeling: A Practical Guide To Steady State
Modeling of Petroleum Processes: Using Commercial Simulators. Athens
been developed, however, the reliability might be questionable
Printing Company.
especially for real time optimization of the process. The BANN mod- Kiss, A.A., Ignat, R.M., 2013. Revamping dimethyl ether separation to a single-step
els were found to predict optimum operating conditions of the ADU. process. Chem. Eng. Technol 36, 1261–1267.
Li, J., Misener, R., Floudas, C.A., 2012. Continuous-time modeling and global
The historic data of a plant can be utilized to build a neural network
optimization approach for scheduling of crude oil operations. AIChE J. 58,
model. If a system is generated based on real conditions of oper- 205–226.
ation of a plant, the system can be constantly improved. This is Nakaiwa, M., Huang, K., Endo, A., Ohmori, T., Akiya, T., Takamatsu, T., 2003.
because new operating conditions from the plant can be used to Internally heat-integrated distillation columns: a review. Chem. Eng. Res. Des.
81, 162–177.
train and adjust the neural network. This will guide operators and Nur Izyan, Z., Shuhaimi, M., 2014. Exergy analysis for fuel reduction strategies in
design engineer to operate the plant at the most energy efficient crude distillation unit. Energy 66, 891–897.
 F.N. Osuolale, J. Zhang / Computers and Chemical Engineering 103 (2017) 201–209 209

Ochoa-Estopier, L.M., Jobson, M., Smith, R., 2013. Operational optimization of Szargut, J., Morris, D.R., Steward, F.R., 1988. Exergy Analysis of Thermal, Chemical
crude oil distillation systems using artificial neural networks. Comput. Chem. and Metallurgical Processes. Hemisphere Publishing Corporation, New York.
Eng. 59, 178–185. Tarighaleslami, A.H., Omidkhah, M.R., Ghannadzadeh, A., Hesas, R.H., 2012.
Osuolale, F.N., Zhang, J., 2014. Energy efficient control and optimisation of Thermodynamic evaluation of distillation columns using exergy loss profiles: a
distillation column using artificial neural network. Chem. Eng. Trans. 39, 37–42. case study on the crude oil atmospheric distillation column. Clean Technol.
Osuolale, F.N., Zhang, J., 2015. Multiobjective optimisation of atmospheric crude Environ. Policy 14, 381–387.
distillation system operations based on bootstrap aggregated neural network Vazquez–Castillo, J.A., Venegas–Sánchez, J.A., Segovia–Hernández, J.G.,
models. Comput.-Aided Chem. Eng. 37, 671–676. Hernández-Escoto, H., Hernández, S., Gutiérrez–Antonio, C., Briones–Ramírez,
Osuolale, F.N., Zhang, J., 2016. Energy efficiency optimisation for distillation A., 2009. Design and optimization, using genetic algorithms, of intensified
column using artificial neural network models. Energy 106, 562–578. distillation systems for a class of quaternary mixtures. Comput. Chem. Eng. 33,
Rivero, R., Rendon, C., Monroy, L., 1999. The exergy of crude oil mixtures and 1841–1850.
petroleum fractions: calculation and application. Int. J. Thermodyn. 2, 115–123. Xu, L., Yuan, X., Chen, D., Luo, Y., Yu, K., 2013. Reversibility analysis for design
Shahnovsky, G., Cohen, T., Mcmurray, R., 2012. Integrated monitoring for optimization of an internally heat-integrated distillation column. Chem. Eng.
optimising crude distillation. Pet. Technol. Q. 17, 43–49. Technol. 36, 1147–1156.
Shang, C., Yang, F., Huang, D., Lyu, W., 2014. Data-driven soft sensor development Zhang, J., 1999. Developing robust non-linear models through bootstrap aggegated
based on deep learning technique. J. Process Control 24, 223–233. neural networks. Neurocomputing 25, 93–113.