Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Xin Xiao
Institute of Process Engineering, Chinese Academy of Sciences, Beijing 100190, P. R. China
In this work we develop a novel modeling and global optimization-based planning formulation, which predicts product
yields and properties for all of the production units within a highly integrated refinery-petrochemical complex. Distillation is modeled using swing-cut theory, while data-based nonlinear models are developed for other processing units.
The parameters of the postulated models are globally optimized based on a large data set of daily production. Property
indices in blending units are linearly additive and they are calculated on a weight or volume basis. Binary variables
are introduced to denote unit and operation modes selection. The planning model is a large-scale non-convex mixed
integer nonlinear optimization model, which is solved to e-global optimality. Computational results for multiple case
studies indicate that we achieve a significant profit increase (3765%) using the proposed data-driven global optimization framework. Finally, a user-friendly interface is presented which enables automated updating of demand, specificaC 2016 American Institute of Chemical Engineers AIChE J, 62: 30203040, 2016
tion, and cost parameters. V
Keywords: refinery, petrochemical, planning, Big-Data, global optimization
Introduction
In the last 20 years, the petrochemical industry has succeeded by creating markets and supplying them with suitable
products used to create goods such as plastics, cosmetics,
lubricants, and paints. Petrochemical production begins in a
refinery that separates crude oils mainly into lighter components such as naphtha, light naphtha, top oil, and liquid fuels
including gasoline, diesel, and jet fuel. The lighter components
are further processed into various petrochemicals such as ethylene, propylene, butadiene, benzene, toluene, xylol, and some
other high-valued products via cracking, butadiene extraction,
hydrotreating, etherification, and polymerization processes.
Additional Supporting Information may be found in the online version of this
article.
*Contributed equally to this work.
Current address of J. Li: School of Chemical Engineering and Analytical Science, The University of Manchester, Manchester.
Correspondence concerning this article should be addressed to X. Xiao and
C. A. Floudas at xxiao@ipe.ac.cn and floudas@tamu.edu.
C 2016 American Institute of Chemical Engineers
V
3020
AIChE Journal
and final products. Mathematical modeling of processing, production, pooling, and blending operations may introduce bilinear, quadratic, polynomial, signomial, exponential, and higher
order terms. Conversely, the selection of parallel production
units and production modes introduces binary variables.
Hence, the overall problem is a large-scale non-convex mixed
integer nonlinear optimization (MINLP) problem.
The refinery planning problem has received considerable
attention since the introduction of linear programming in
1950s.810 Research focused on developing different models
and algorithms to solve large-scale industrial problems, leading to commercial software such as RPMS (Refinery and
Petrochemical Modeling System),11 PIMS (Process Industry
Modeling System),12 GRTMPS (Haverly Systems).13 The
commercially available software can be extended to model
and optimize integrated petrochemical processes, however,
inaccuracy caused by non-rigorous linear models and approximate algorithms may reduce the overall profitability or sacrifice product quality.
Nonlinear models and specialized algorithms have also
been proposed for refinery planning problems.5,1424 For
instance, Pinto and Moro15 developed a nonlinear planning
model for production planning which allows for the implementation of nonlinear process models and blending relations.
Pinto et al.5 proposed a planning and scheduling model for
refinery operations. They presented a formulation based on
discretization of time for production and distribution scheduling and their model included features such as sequence
dependent transition cost of products within an oil pipeline. Li
et al.17 presented a refinery planning model that utilizes simplified empirical nonlinear process models with an considerations for crude characteristics, product yields and qualities.
Alhajri et al.18 developed a nonlinear model to address the
refinery planning problem. Alattas et al.21 developed a fractionation index based nonlinear model for crude distillation units
(CDUs) and integrated it into the linear refinery planning
model, solving it with nonlinear programming (NLP) solvers
without guaranteeing global optimality. Mouret et al.25
addressed the problem of the integration of refinery planning
and crude-oil scheduling. Menezes et al.23 used the improved
swing cut approach26 to develop a single-period nonlinear
optimization model for an oil-refinery production planning
model to predict the national overall capacity for different oil
refinery units in Brazil considering four future market scenarios. In all of the literature dealing with nonlinear models thus
far, no global optimality is guaranteed. A comprehensive
review on refinery planning can be found in Shah et al.6
Another level of complexity which has not been dealt with
sufficiently in the literature insofar is the integration and interaction between refineries and chemical plants. Refineries are
typically integrated with chemical plants through exchange of
intermediate streams both from refinery to chemicals and from
chemicals to refinery. Consequently, the operation of the former highly affects the operation of the latter and vice
versa,2732 and hence planning of the refinery operations and
chemical operations separately, will not lead to the globally
operating points of the entire integrated petrochemical complex. In Al-Qahtani et al.27,28 the formulation solved for the
coordination between multisite refineries and chemical plants
is a MINLP problem; in Swaty32 a refinery and ethylene plant
are integrated through exchange of intermediate streams using
linear programming; in Gonzalo et al.30 the advantages of the
integration between a refinery and a single hydrocracking unit
AIChE Journal
DOI 10.1002/aic
3021
Figure 1. A schematic diagram of a refinery-petrochemical complex with multiple crude oil sources.
[Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
Problem Definition
Figure 1 illustrates a representation of a refinerypetrochemical complex which we model in this work; a more
detailed representation of the refinery is shown in Figure 2a
and a representation of the chemical plants is shown in Figure
2b. Figures 2a, b show the main connections and processes
present in the complex studied in this work. Typically there
are C crude oils (c 2 f1; 2; . . . Cg) which are processed in a
refinery, where each one has different availability, consistency, properties, and price depending on its source. In this
work, a maximum of three different types of crudes are used.
In the entire petrochemical plant, there are total U
(u 2 f1; 2; . . . U g) units including processing units, pools, and
blenders. The processing units, which are denoted as UPRO
include CDUs, delayed coking units, hydrotreating units,
hydrocracking units, catalytic cracking units, reforming units,
extraction units, hydrogen generation units, sulfur production
3022
DOI 10.1002/aic
AIChE Journal
Figure 2. (a) A simplified refinery process in a real petrochemical plant. (b) A simplified chemical process in a real
petrochemical plant.
[Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
DOI 10.1002/aic
3023
DOI 10.1002/aic
Industrial Data
This work is enabled by the availability of industrial data
which is used to develop hybrid or purely data-based input
output models for all of the processing units. Specifically, the
data which was provided for this work consists of daily average flow rate and property measurements for the entire set of
streams in all units of operation of the integrated refinerypetrochemical complex. In addition, crude amounts and crude
AIChE Journal
Abbreviation
Addition basis
Specific gravity
Sulfur
Research octane number
Motor octane number
Nitrogen
Fe Content
Carbon residue
Freeze point
Flash point
Smoke point
Reid vapor pressure
Aromatics
Olefins
Viscosity at 208C, 1008C
Cetane number
Cold filter plugging point
SPG
SUL
RON
MON
N2
Fe
CCR
FRP
FLP
SMP
RVP
ARO
OLE
V20,V100
CET
CFPP
Volume
Weight
Volume
Volume
Weight
Weight
Weight
Volume
Weight
Volume
Volume
Volume
Volume
Volume
Volume
Volume
assay data is provided for the three crude materials used. Handling, analysis and use of large data-set poses great challenges,
but also creates opportunities for efficient decision making in
todays Big-Data era.33 First, we will describe the raw data
which was provided and then we will discuss the challenges,
advantages and limitations of a data-based approach.
The plant which is studied in this work uses three type of
crudes (C1, C2, C3) for which the daily processed amount and
crude assay data analysis is provided. The crude assay analysis
provides information such as specific gravity, sulfur content,
nitrogen content, flash point, viscosity, freezing point, and
Reid vapor pressure (Table 1) for different crudes and boiling
ranges of distillates. The flow rate and property data of the
crudes are specifically important for modeling the CDU unit
using a cutting-point temperature method described in the next
section.
The big data set provided also consists of daily measurements of flow rates of each input and output stream for every
unit u present in the entire petrochemical complex. The flow
rate data provided depicts in detail, not only the amount that is
produced on a daily basis for each unit, but also the fractions
of each stream that are sent to different processing units,
blending units, pooling units, inventory, and sales on a daily
basis. The flow rate data set does not contain missing points,
since a measurement is provided for each day for a total number of N days within the whole period. Daily averages are used
as data points, which in many cases comprise of measurements
taken at different time points during the day. These averages
alleviate the effect of the time delay which may be caused for
a material to flow from crude feeding to downstream processing and blending. In addition, our final goal is the optimization
of the planning operations over a monthly time horizon, thus
such delays are assumed to be negligible. First, we collect and
group the flow rate data into submatrices,each corresponding
out
to one unit u: FdataU 5 N3 jSin
U j1jSU j , where j j represents the cardinality of the corresponding set. Moreover, we
are provided with a large set of property data for certain input
and output streams within each unit of the petrochemical plant.
Preprocessing of the property data set is more challenging
since measurements are not provided for every day of operation. This creates missing data gaps and inconsistencies
between connections of different units which we need to overcome to solve the necessary parameter estimation problems.
The first step is to group all of the input properties and output
properties corresponding to all of the input and output streams
AIChE Journal
of a unit u into a matrix: EdataU 5 N3 jEin
U s; uj1
jEout
U s; uj, which contains the same number of rows as
matrix FdataU .
Due to measurement errors caused by uncertain tank measurements, inconsistencies between measurements at different
locations in the plant, and human error, there is a need for preprocessing of the flow rate and property data prior to its use
for parameter estimation. First, unreliable measurements were
flagged as erroneous by the company and were immediately
removed from the data set. Following, we also perform additional processing of the data on a per-unit basis. First, we combine the two data matrices FdataU and EdataU into one
matrix TdataU 5 FdataU EdataU (Figure 4). Then, we calculate the daily mass balance by subtracting the total sum of
flow rates of the outlet streams from the total sum of flow rates
of the inlet streams. If this amount exceeds a small percentage
threshold which is accepted as loss for each unit, then the
entire row of matrix TdataU is removed. After the above filtering steps, we further removed measurements beyond 63 sd
from the average of the data set. Due to the amount of the data
available for each stream, we assume that the calculation of
average properties and yields is accurate. Also, the occurrence
DOI 10.1002/aic
3025
of significant outliers was rare (i.e., less than 0.1% of the data
set, if any), thus we believe that this preprocessing was sufficient to alleviate any undesirable effects to the fitted parameters of the models. Any removed data measurements are
treated as missing elements, without removal of the entire row
of the data matrix. Finally, the TdataU matrix contains several
missing elements, which are imputed using nearest-neighbor
imputation (knnimpute function in Matlab) to enable the procedure of parameter estimation. This approach uses the nearest
measurement, based on the Euclidean distance of points
among the data set. Using this interpolation approach we
assume that missing property measurements are equal to existing measurements which are the most similar to the missing
element row-wise. In other words, missing property measurements are assumed to be equal to existing measurements from
days of operation which are the most similar based on both
property and flow rate information. This assumption has
shown to be very successful given the amount of data which
we have available for over a 2-year operation, but also due to
the realization that a plant measurement was typically not
made during days for which the operation is stable and consistent with prior days. Consequently, it is highly probable that
a very similar operation point exists in the data set to impute a
reasonable property measurement.
Filtering, sorting, and preprocessing of the data requires a
significant amount of time and effort, however, it is crucial to
ensure that the quality and reliability of the data is improved.
The processed data contains a reasonable amount of noise due
to uncertainty and variability in data-collection by different
operators and equipment throughout the plant. However, due
to the fact that we have in our possession a significant amount
of daily data, we will later show that the trends and correlations between flow rates, properties, and yields can be captured by postulating empirical nonlinear models and
performing parameter estimation using the processed data. In
fact, the fitted parametric models have the ability to smoothout the noise in the data, while capturing the true trends of the
data, when the parameter estimation is performed correctly
and overfitting is avoided.
There are several limitations of the data-based approach
which is followed in this work which should be mentioned
here. First, the data-based models have a good predictive ability which is bounded by the upper and lower bounds of the
experimental data which was used for their development. Consequently, if there is a significant change in the operation, the
flowsheet superstructure, or the material properties of the
investigated plant, then the model will not be able to capture
the performance of the plant unless new data is collected and
used to update the models. Second, due to certain missing
information regarding properties and operating conditions of
the processes, the data-based models are based only on the
available inputs. However, the aim of this work is to develop a
fast decision-making tool for the petrochemical plant which
captures the nonlinear behavior of the plant based on the information which is available to the decision makers at all times,
and which will be used for decisions at the planning level of
the plant. Due to the latter, we have found that a data-driven
approach is very powerful for petrochemical planning operations. Lastly, we must mention that all of the fitted models
were validated against data that was provided for a recent
monthly operation of the plant, which was not used in the
training data set. This validation procedure was performed in
collaboration with the operators of the plant. Depending on
3026
DOI 10.1002/aic
AIChE Journal
8u 2 UCDU ; c 2 Sin
u
(1)
Figure 6. Crude oil TBP curve for yield with swing cuts
illustration (based on crude oil assay data
from the real petrochemical plant).
[Color figure can be viewed in the online issue, which is
available at wileyonlinelibrary.com.]
m
X
an; c Tcn
8u 2 UCDU ; c 2 Sin
u
n50
where, a(n,c) is a parameter array which needs to be determined. Parameter a(n,c) is determined using a parameter estimation approach, which will be introduced and explained
later.
Since the cutting temperature is the same for all crude-oil
feedstock, we define Tcut(s) to be the cutting temperature for
distillate s and Yield(s,c,u) as the yield of distillate s contributed from crude c in CDU u.
X
Yields; c; u5f T cut s2
yields0 ; c; u
out
(2)
s0 <s;s0 2Su
in
;
c
2
S
8u 2 UCDU ; s 2 Sout
u
u
For the example in Figure 6, the yield of distillate s contributed from crude c in CDU u is presented as follows,
Yields; c; u5
m
X
yields0 ; c; u
s0 <s;s0 2Sout
u
n50
in
8u 2 UCDU ; s 2 Sout
u ; c 2 Su
0
0
Fin
U s ; u Yields; s ; u
P
0
Fin
U s ; u
s0 2Sin
u
CDU
out
8u 2 U
; s 2 Su
Yields; u5
s0 2Sin
u
(3)
DOI 10.1002/aic
3027
Ee; s; c; u5
4
X
am; e; c
m50
m
s0 <s;s0 2Sout
u
in
out
8u 2 UCDU ; s 2 Sout
u ; c 2 Su ; e 2 EU s; u
(6)
c2Sin
u
Fin
U c; u
Yields; c; u Ee; s; c; u
(7)
c22Sin
u
out
8u 2 UCDU ; s 2 Sout
u ; e 2 EU s; u
(4)
CDU
;s
s0 <s;s0 2Sout
u
2 Sout
u ;c
out
2 Sin
u ; e 2 EU s; u
(5)
Usually, the function f is a polynomial, exponential, and their
combinations. Figure 7 illustrates the relationship between
specific gravity to accumulated yield from one crude assay
dataset. It is found that we can use a polynomial function to
accurately describe this correlation.
3028
DOI 10.1002/aic
AIChE Journal
e 2
0
Ein
U s ; u; e
(9)
Eout
U s; u
the flow rates of all the inlet and outlet streams and the properties of all the inlet streams. The set of outputs to be modeled is
the set of yields for all outlet streams: Yieldsout ; u and the set
out
of properties for all outlet streams: Eout
U e; s ; u. Each yield
and property equation is a function of a subset of xin
U based on
the following rules:
1. all input flow rates are used as input variables in the
yield prediction and property prediction models;
2. all input properties of all input streams are used as
input variables in yield prediction models;
3. only the outlet flow rate corresponding to the same
stream of the outlet property prediction is considered
as an input variable to that property prediction
model;
4. a subset of input properties is used as an input variable
in property prediction models based on theoretical
knowledge, prior experience, and statistical analysis.
For example, if the outlet property to be predicted is
SPG, then only the inlet SPG is found to be sufficient
for a prediction. If the outlet property to be predicted is
viscosity, then inlet viscosity and SPG are found to be
important toward its prediction; and
5. if a unit can be operated in different modes, then the
data set is divided into as many modes of operation
and a different model is developed for each mode of
operation of the unit.
Once the set of input variables for Eqs. 8 and 9 are identified, the appropriate model must be postulated to describe the
data. The set of input variables for each of the yield and property prediction equations of forms based on Eqs. 8 and 9
depends on the amount of streams and available inlet properties of the streams of the specific unit. We denote the set of
inputs for models based on Eq. 8 as:
x yieldiin sout ; u xin
i51; . . .; Iyield , and the general
U
quadratic form is shown in the following equation:
Iyield
X
Yieldsout ; u5byield;0 sout ; u1 byield;i sout ; u x yieldiin sout ; u
i51
Iyield X
Iyield
X
i51
j52
ji
in 0
in 0
0
8u 2 UPRO ; u 62 UCDU ; s 2 Sout
u ; s 2 Su ; e 2 EU s ; u
(10)
Equation 10 contains bilinear and quadratic terms between the
selected inlet properties and inlet flow rates. Similarly, if the
set of inputs of Eq. 9 are summarized by variables
out
in
x ein
i51; . . .; Ie , then the general quadratic
i s ; u xU
model which any model of Eq. 9 may take is shown in
Eq. 11.
out
out
Eout
U e; s ; u5be;0 e; s ; u1
Ie
X
be;i e; sout ; u x ei sout ; u
i51
Ie X
Ie
X
i51
j52
ji
in 0
in 0
out
0
8u 2 UPRO ; u 62 UCDU ; s 2 Sout
u ; s 2 Su ; e 2 EU s ; u; e 2 EU s; u
(11)
In Eqs. 10 and 11, parameters byield and be are those which
must be estimated in order for the developed models to best
describe the data. This is achieved by least-squares
DOI 10.1002/aic
3029
Figure 9. Yield data (a) Raw industrial data with outliers, (b) Filtered data, (c) Industrial data vs. predicted data
after parameter estimation.
[Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
minimization between the data and model predictions. However, it is realized that each of the equations of a unit u are not
independent due to mass balance constraints and intercorrelations. Consequently, it is not sufficient to estimate the
parameters of each of these models individually through simple parameter estimation. Once each set of inputs and their
postulated equation forms are identified for each of the outlet
streams, simultaneous parameter estimation is performed for
all of the equations of form (10) and (11) associated with u,
subject to the mass balance constraints. Moreover, the model
predictions are restricted within the lower and upper bounds
identified by physical constraints and/or the experimental data.
The parameter estimation optimization problem has the objective of minimization of the sum of all the mean squared errors
(MSE) of each output, between experimental measurements
and predictions (Problem 12). The parameter estimation problems are nonlinear optimization problems (NLP) due to the
objective function in Problem (12), and are solved to global
optimality using solver ANTIGONE.39 The optimal parameters and bounds of each of the input and output variables for
each unit stream are stored in text files in an appropriate format so that they can be used by the planning model which will
be described subsequently.
X
X
min
MSEyield sout ; u1
MSEe e; sout ; u
be ;byield
e2sout
sout
in
in in
Yield out sout ; u5f Ein
U s ; u; FU s ; u; byield
out
in in
out out
in
in
Eout
U e; s ; u5f FU s ; u; FU s ; u; EU e; s ; u; be
X
MSEyield sout ; u5 yieldN sout ; u2yieldN;exp sout ; u2
N
X
2
out
out
MSEe e; sout ; u5 Eout
U;N e; sout ; u2EU;N;exp s ; u
X
N
out
sout
X
Fout sout ; u5 Yieldsout ; u
Fin sin ; u
sin
(12)
3030
DOI 10.1002/aic
Once the optimal parameters are identified through the solution of the nonlinear optimization problem (12), it is imperative to add a set of equations to the unit model for
completeness and integration within the planning model.
These additional constraints are described in the next section.
To further clarify the above procedure and also show the
type of data which this work was based on, we describe here
an example of a property prediction and a yield prediction
model for a catalytic cracking unit. In this unit there is only
one inlet stream, five inlet properties describing the inlet
stream and six outlet streams. Thus the input variable set
out
in
out
out
out
out
contains 12 variables: xin
U 5f 1U ; f 1U ; f 2U ; f 3U ; f 4U ; f 5U ;
in
in
out in
in
in
f 6U ; eU e1; s; u; eU e2; s; u; eU e3; s; u; eU e4; s; u; eU e5; s; u.
The model for the prediction of a yield of one of the outlet
streams is a function of six variables following the rules
described above (inlet flow rate and inlet properties). A linear
model is not found to be accurate for this prediction, thus a
full quadratic model (Eq. 10) is used to predict the data as
shown in Figure 9.
The model for the prediction of the specific gravity of one
of the outlet streams is a function of the three input variables:
the inlet flow rate, the outlet flow rate of the same stream, and
the input property corresponding to SPG. The function represented by Eq. 11 is used with three inputs and all of the nonlinear terms to produce an accurate prediction for this property.
The data and prediction of this model are shown in Figure 10.
It should be noted that the optimal parameters of both of the
predictions shown in Figures 9 and 10 are estimated through
one parameter estimation model of form 12 for the catalytic
cracking unit, along with all of the other yield and property
prediction functions of the same unit.
AIChE Journal
Figure 10. Specific Gravity data (a) Raw industrial data with outliers, (b) Filtered data, (c) Industrial data vs. predicted data after parameter estimation.
[Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
Fin
U s; u
(13)
s0 2Sout
u
Unit selection
There may be several parallel processing units to produce
the same products. To model the selection of those parallel
units, we define a binary variable Y(u) as follows,
(15)
(18)
where, Ein
U e; s; u is the blending index of property e in inlet
stream s to blending or pooling unit u. Note that Eq. 16 is for
volume-based indices, and Eq. 17 for weight-based indices.
Equations 1517 introduce bilinear terms.
AIChE Journal
DOI 10.1002/aic
3031
(
Yu5
If unit u is selected
otherwise
8u 2 UPAR
The total inlet flow rate to each unit must satisfy its minimum
[CAPLU u] and maximum [CAPU
U u] capacity. Then we have,
X
U
Fin
CAPLU u Yu
U s; u CAPU u Yu
in
(19)
s2Su
8u 2 UPAR
X
U
PAR
Fin
(20)
CAPLU u
U s; u CAPU u 8u 62 U
s2Sin
u
Mode selection
In some processing units, several operational modes may be
involved during the entire planning horizon. At each time
(e.g., each day) only one production mode is involved. To
model the selection of different operational modes, the following binary variable x(m,u) is defined.
(
1 If operational mode m in unit u is selected
xm; u5
0 otherwise
8u 2 Umod ; m 2 Mu
Note that the optimization model developed is used for the
global optimization of an average daily profit during the entire
planning horizon. Therefore, we need to know how many days
are used for mode m in the planning horizon. To achieve this,
we introduce the following auxiliary binary variable,
(
1 If mode m in unit u is operated in day d
zm; u; d5
0 otherwise
mod
8u 2 U ; m 2 Mu ; d
The total number of days for operational mode m in unit u,
which is defined as dm(m,u), can be calculated using the
equivalent mathematical form from Floudas (1995) given as
follows,
dmm; u5
D
X
d51
(21)
where, parameter D is used to denote the maximum number of
days to represent the total number of days in the planning horizon. For instance, D is assigned to 5 which is enough to represent 31 days in the planning horizon.
If mode m in unit u is operated in day d, then mode m in
unit u should be operated.
xm; u zm; u; d 8u 2 Umod ; m 2 Mu ; m < M; d D
(22)
Alternatively, if mode m in unit u is not operated in any day d,
then mode m in unit u should not be operated.
D
X
mod
(24)
m2Mu ;m51
8u 2 Umod ; m5M
(25)
mod
(26)
8u 2 U
; m5M
(27)
where, Inv0(s,u) denotes the initial inventory for stream s of a
unit u.
For some processing units such as hydrocracking and hydrotreating units, H2 is one of the feedstocks. The inlet flow rate
of H2 must satisfy its minimum [roLH s; u] and maximum
[roU
H s; u] ratio requirements.
X
X
0
in
U
0
Fin
Fin
roLH s; u
U s ; u F s; u roH s; u
U s ; u
s0 2Sin
u
8u;
s2Sin
u;
s0 2Sin
u
0
s 5 f H2 g
; m 2 Mu ; m < M (23)
(28)
xm; u
zm; u; d 8u 2 U
d51
3032
DOI 10.1002/aic
AIChE Journal
0 0
in
PAR
Eout
; s0 ; u0 ; s; u 2 UC
U e; s ; u 5EU e; s; u 8u 62 U
(29)
in;L
0 0
in
Eout
U e; s ; u EU e; s; u2EU e; s; u 12Yu
PAR
0 0
8u 2 U ; s ; u ; s; u 2 UC
(30)
out;U
0 0
in
e; s0 ; u0 12Yu
Eout
U e; s ; u EU e; s; u1EU
(31)
PAR
0 0
8u 2 U ; s ; u ; s; u 2 UC
(32)
in;U
in
Ein;L
U e; s; u Yu EU e; s; u EU e; s; u Yu (33)
8u 2 UPAR ; s0 ; u0 ; s; u 2 UC
The yield prediction model for those units without or with different operational modes can be developed by solving the parameter estimation problem (12). For completeness, the generic
correlations for yield and property predictions are given here.
0 0
in 0
mod
Yields; u5f Ein
U e ; s ; u; FU s ; u 8u 62 U
in 0 0
0
Yields; u; m5f EU e ; s ; u; Fin
U s ; u ;
0 0
in 0
where Ein
U e ; s ; u; FU s ; u 2 m
8u 2 Umod
(34)
(35)
Note that for the units with multiple modes, the property and
flow rate data is divided into different sets, each corresponding
to a different mode of operation. The summation of product
yields must be equal to 1 as follows:
X
Yields; u1YieldLossu51 8u 62 Umod
(36)
s2Sout
u
hP
s0 2Sin
u
i
0
Fin
U s ; u Yields; u; m dmm; u
(39)
Ie
X
be;i e; sout ; ux ei sout ; u
i51
Ie X
Ie
X
i51
8u 62 Umod
j52
ji
(40)
AIChE Journal
Objective function
The final objective for the entire planning model is to maximize the total profit, which consists of revenue from the sale
of final products, raw material purchase cost, and operational
cost. The operational cost for each unit u is assumed to be linear with the inlet or outlet stream flow rates of products of this
unit. The inlet and outlet streams involved in their operational
cost are denoted as SuUTOC;in and SUTOC;out
, respectively. We
u
define Price(s) as the price of each final product s ($/ton),
Cost(s) as the purchase cost for raw material s ($/ton), and
UTOC(u) as the operational cost for each production unit u.
The total profit is calculated by,
8
2
X X<
X
5
Prices 4
:
SAL
in
u0
u2U
s2Su
8
X X<
u2U
PURC
s2Sout
u
8
X <
PRO
Costs
0
UTOCu @
X
s0 2Sout
:s0 ;u0 ;s;u2UC
u0
u0
s0 s0 2Sin
:s0 ;u0 ;s;u2UC
u0
X
s2SuUTOC;in
Fin
U s; u1
39
=
Fs ; u ; s; u5
;
0
Fs; u; s0 ; u0
9
=
;
19
=
A
Fout
s;
u
U
;
UTOC;out
X
s2Su
(43)
At this point, we have completed the description of our entire
single-period model which comprises of Eqs. 1343. Detailed
representations of the forms of the models and parameter values are provided in Supporting Information.
Computational Platform
DH
8u 2 UPRO \ Umod ; s 2 Sout
u
out
out
Eout
U e; s ; u5be;0 e; s ; u1
Each final product s has its own demand during the planning
horizon. The minimum [DmdSL s] and maximum [DmdSU s]
demand requirements for each final product s can be constrained as follows,
X X
Fs0 ; u0 ; s; u DmdSU s
DmdSL s
0
0
0
0
s u :s ;u 2UC
(42)
FP
;
s
2
S
8u 2 USAL ; s 2 Sin
u
u2U
m2Mu ;m51
Demand constraints
PROFIT
s2Sout
u
Fout
U s; u
PM
DOI 10.1002/aic
3033
Computational Results
Figure 13. Schematic representation of procedure followed by the user using the Excel computational platform.
[Color figure can be viewed in the online issue, which
is available at wileyonlinelibrary.com.]
We solve three cases studies (CS1CS3) from a real petrochemical plant to illustrate the predictive power and superiority of the developed data-driven and global optimization
framework and compare our results with those based on an
empirical decision-making approach from the actual operation. In these three case studies, 36 raw materials including
three different crudes are used to produce 200273 products
including dry gas, LPG, naphtha, gasoline, diesel, Jet fuel,
benzene, toluene, xylol, sulfur, ethylene, styrene, methane,
and many more proprietary products. There are 34 processing
units, 5 pools, and 5 blenders. Two parallel hydrocracking
units (denoted as HC1 and HC2) and two parallel catalytic
cracking units (denoted as FC1 and FC2) with different
capacities are used to produce the same products. While two
production modes such as Naphtha and Reforming
Feedstock mode can be operated in the CDU unit, Light
Diesel and Jet Fuel production modes can be operated in
HC1 and HC2. These two different production modes cannot
be operated during the same day. Several important properties
including SPG, SUL, CCR, RON, V20, V100, FLP, FRP,
CFPP, CEN, ARO, OLE, BEN, SMP, RVP, and N2 are considered (Table 1). The formulated mixed integer nonlinear
optimization problem contains on average 11 binary variables,
1150 continuous variables, 1100 equations, and 1050 nonlinear terms. The breakdown of the model statistics for the integrated refinery and chemical plant complex is shown in Figure
14. The values of Figure 14 change slightly for different case
studies depending on the number of parallel units as well as
the number of variables which are fixed to constant values.
The main differences of these three case studies are crude
amounts and cost, final product price and demands, and raw
material cost. All examples are solved using ANTIGONE
1.13841 in GAMS 24.2.2 on Dell OPTIPLEX 960 (IntelV
XeonR CPU 3.00 GHz, 2 GB RAM memory) running Linux.
The default optimality gap of 1e-04 is used for convergence
and the solutions reported did not require the specification of
an initial point.
R
DOI 10.1002/aic
AIChE Journal
Table 2. Model statistics, computational cost (CPU time) for obtaining the global optimum solution, and comparison of actual
vs. optimized operation for CS1, CS2, CS3 in US $Million
CS1
Total equations
Total continuous variables
Total binary variables
Total nonlinear terms
Total CPU cost for GO (s)
Total income
Total material cost
Total operating cost
Total profit
Rel. profit increase (%)
CS2
CS3
Actual
operation
Optimized
operation
Actual
operation
Optimized
operation
Actual
operation
Optimized
operation
721.57
646.17
39.67
35.73
1067
1117
11
1026
18
742.5
655.74
37.536
49.22
600.80
586.84
28.20
214.23
1100
1149
11
1053
195
662.27
611.52
30.02
20.73
747.94
669.99
40.30
37.64
1118
1160
11
1072
212
769.57
668.67
39.08
61.83
37.75
245.65
64.25
ation and the optimized operation for all of the three case
studies.
Table 2 shows that the global optimization framework was
able to identify significantly improved feasible solutions for
the entire petrochemical planning problem for the three different case studies. For CS1 the total income is increased in the
optimized operation by $13.5 million, the material cost is also
increased by $9.57 million but the operating cost is decreased
by $2.1 million. In other words, a higher profit is made by purchasing more raw materials to make more sales, while
decreasing the operating cost. In the case of CS2, the total
income is dramatically increased by $62 million while the
material cost is increased by $24 million and the operating
cost is increased slightly by $2 million. For the operation of
this month, the global optimization framework suggests that
more expenses should be made to achieve a much higher
profit. The estimation of the actual profit for this month based
on the information that was provided, suggests that there was
no profit made during this month. This can be explained by
the low amount of crude materials that were processed during
this month. Finally, in CS3 the total income is actually
increased by $22 million by reducing the amount of raw materials purchased by $1 million and the operating cost by $1
Figure 15. Actual vs. Optimized relative income contributions (%) for CS1.
[Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
AIChE Journal
DOI 10.1002/aic
3035
Figure 16. Actual vs. Optimized relative cost contributions (%) for CS1.
[Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
million. In this case, market demands were satisfied by reducing the purchase and operating cost to achieve a higher profit.
Table 2 also summarizes the model statistics and the CPU
time required to obtain global optimality.
Following, we provide a more detailed comparison of the
actual operation vs. the optimized operation for each case
study separately. To analyze the optimized results and verify
their feasibility, it is important to identify the main sources of
the improved performance in the total profit when comparing
the actual operation to the optimized operation. For this we
have calculated the fractional percent contributions of each
stream, raw material and unit to the total income, total material cost, and total operating cost, respectively. Then, we rank
these contributions with decreasing order and compare them
to identify the main differences between the actual and the
optimized operation. Due to proprietary information restrictions, we are not able to disclose the names of the chemicals
produced by the petrochemical plant.
Figure 17. Actual vs. Optimized relative income contributions (%) for CS2.
[Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
3036
DOI 10.1002/aic
AIChE Journal
Figure 18. Actual vs. Optimized relative cost contributions (%) for CS2.
[Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
Figure 19. Actual vs. Optimized relative income contribution (%) for CS3.
[Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
AIChE Journal
DOI 10.1002/aic
3037
Figure 20. Actual vs. Optimized relative cost contribution (%) for CS3.
[Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
Acknowledgments
The authors would like to acknowledge financial support
from the Chinese Academy of Sciences Visiting Professorship for Senior International Scientists (Grant No.
2010T2G34-2012T1GY11-Continue to March, 2014) with
Science Research and Technology Development Program of
PetroChina Company Limited (No. 2012D-3202-0313),
National Science Foundation (CBET-0827907 and CMMI08856021). Jie Li is further thankful from the National Natural Science Foundation of China (21206174).
Conclusions
This work is a novel Big-Data application for developing
and optimizing a MINLP model for the planning operations of
a large petrochemical plant. The abundance of data allowed
for the development of inputoutput models for all of the processing units of the petrochemical plant by postulating theoretical or empirical models and then identifying the optimal
parameters of the models to best describe the existing data. All
of the developed models are integrated within a large-scale
planning model by introducing additional mass balance constraints, stream connections, blending and pooling unit equations. Discrete variables are introduced due to units which can
be operated at different modes, and also due to the existence
of multiple parallel units which can be used to process the
same materials to produce the same products. The entire planning model is highly non-convex nonlinear and was solved to
global optimality using global optimization software ANTIGONE. In this work we briefly introduce the development of a
user-friendly computational platform which is developed to
allow for the easy and fast automated updating, solution and
results analysis of this complex model through Excel. Results
shown for three case studies show that through our approach
we can obtain significant increase in profit, when compared to
an empirical decision-making practice. This the first complete
3038
DOI 10.1002/aic
Notation
CDU =
LPG =
MINLP =
FC1 =
FC2 =
HC1 =
HC2 =
HKU =
GHU =
GRU =
ARU =
DCU =
DHU =
ET1 =
ET2 =
HTU =
BU1 =
BU2 =
ER1 =
ER2 =
HT2 =
ST2 =
Sets
AIChE Journal
Umod =
USAL =
UPAR =
Sin
u =
Sout
u =
SFP =
Ein
U s; u =
Eout
U s; u =
Mu =
UC =
Subscripts
u = unit
s = stream
e = property specification
Superscripts
U=
L=
in =
out =
UTOC =
upper limit
lower limit
inlet
outlet
involved in the calculation of operational cost
Parameters
H=
N=
DH =
D=
Price(s) =
Cost(s) =
UTOC(u) =
DmdSL s =
DmdSU s =
A(n,s) =
ELU e; s; u =
EU
U e; s; u =
CAPLU u =
CAPU
U u =
Inv0(s,u) =
roLH s; u =
roU
H s; u =
planning horizon
total number of days for which industrial data is available
total number of days in the planning horizon
a parameter that is used to represent the total number of
days in the planning horizon
price ($/ton) of final product s
cost ($/ton) of raw material s
operational cost ($/ton) of unit u
minimum demand requirement for product s
maximum demand requirement for product s
a parameter in the correlation for yield prediction
minimum requirement for property e in stream s from unit u
maximum requirement for property e in stream s from unit
u
minimum capacity
maximum capacity
initial inventory of stream s to (from) unit u
minimum ratio requirement for stream s to unit u
maximum ratio requirement for stream s to unit u
Binary variables
Y(u) = 1 if unit u is selected
z(m,u,d) = 1 if mode m in unit u is operated in day d
01 Continuous variables
x(m,u) = 1 if production mode m in unit u is selected
Continuous variables
F(s,u, s0 , u0 ) = mass flow rate of stream s from unit u to unit u0 as
stream s0
xf(s,u,s0 ,u0 ) = volumetric flow rate of stream s into unit u that are used
to produce product s0
Fin
U s; u = inlet mass flow rate of stream s to unit u
Fout
U s; u = outlet mass flow rate of stream s from unit u
Yield(s,u) = yield of outlet stream s from unit u
YieldLoss(u) = loss yield of a unit u
Ein
U e; s; u = property e in the inlet stream s to unit u
Eout
U e; s; u = property e of outlet stream s from unit u
Tcut(s) = cutting temperature for distillate s
Yield(s,s0 ,u) = yield of distillate s contributed from crude s0 in the crude
distillation unit u
E(e,s,s0 ,u) = property e of distillate s contributed from crude s0 in the
crude distillation unit u
dm(m,u) = total number of days for operational mode m in unit u
Profit = total profit ($)
Literature Cited
1. Jia Z, Ierapetritou M. Mixed-integer linear programming model for
gasoline blending and distribution scheduling. Ind Eng Chem Res.
2003;42(4):825835.
AIChE Journal
2. Li J, Karimi IA. Scheduling gasoline blending operations from recipe determination to shipping using unit slots. Ind Eng Chem Res.
2011;50(15):91569174.
3. Li J, Karimi IA, Srinivasan R. Recipe determination and scheduling
of gasoline blending operations. AIChE J. 2010;56(2):441465.
4. Li J, Li W, Karimi IA, Srinivasan R. Improving the robustness and
efficiency of crude scheduling algorithms. AIChE J. 2007;53(10):
26592680.
5. Pinto JM, Joly M, Moro LFL. Planning and scheduling models for
refinery operations. Comp Chem Eng. 2000;24(910):22592276.
6. Shah NK, Li Z, Ierapetritou MG. Petroleum refining operations: key
issues, advances, and opportunities. Ind Eng Chem Res. 2011;50(3):
11611170.
7. Bengtsson J, Nonas S-L. Refinery planning and scheduling: an overview. In: Bjrndal E, Bjrndal M, Pardalos PM, R
onnqvist M, editors. Energy, Natural Resources and Environmental Economics.
Berlin: Springer, 2010:115130.
8. Symonds GH. Linear Programming: The Solution of Refinery Problems, Vol 8. New York: Esso Standard Oil, 1955.
9. Aronofsky JS, Dutton JM, Tayyabkhan MT. Managerial Planning
with Linear Programming: In Process Industry Applications. Michigan: Wiley; 1978.
10. Pelham R, Pharris C. Refinery operation and control: a future vision.
Hydrocarb Proces. 1996;75(7):8994.
11. Bonner & Moore. RPMS (Refinery and Petrochemical Modeling System): A System Description [Computer Program]. Houston: Bonner
& Moore Management Science; 1979.
12. ASPEN Technology Inc. ASPEN P.I.M.S. System Reference (v7.2.)
[computer program]. ASPEN Technology Inc.; 2010.
13. Haverly Systems. GRTMPS [computer program]. Haverly Systems;
2015.
14. Moro LFL, Zanin AC, Pinto JM. A planning model for refinery diesel production. Comp Chem Eng. 1998;22:S1039S1042.
15. Pinto JM, Moro LFL. A planning model for petroleum refineries.
Brazil J Chem Eng. 2000;17:575586.
16. Kelly JD. Formulating production planning models. Chem Eng Prog.
2004;100:4350.
17. Li W, Hui C-W, Li A. Integrating CDU, FCC and product blending
models into refinery planning. Comp Chem Eng. 2005;29(9):20102028.
18. Alhajri I, Elkamel A, Albahri T, Douglas PL. A nonlinear programming model for refinery planning and optimisation with rigorous
process models and product quality specifications. Int J Oil Gas
Coal Technol. 2008;1(3):283307.
19. Zhang BJ, Hua B. Effective MILP model for oil refinery-wide production planning and better energy utilization. J Clean Prod. 2007;
15(5):439448.
20. Guyonnet P, Grant FH, Bagajewicz MJ. Integrated model for refinery planning, oil procuring, and product distribution. Ind Eng Chem
Res. 2009;48(1):463482.
21. Alattas AM, Grossmann IE, Palou-Rivera I. Integration of nonlinear
crude distillation unit models in refinery planning optimization. Ind
Eng Chem Res. 2011;50(11):68606870.
22. Mahalec V, Sanchez Y. Inferential monitoring and optimization of
crude separation units via hybrid models. Comp Chem Eng. 2012;45:
1526.
23. Menezes BC, Moro LFL, Lin WO, Medronho RA, Pessoa FLP. Nonlinear production planning of oil-refinery units for the future fuel
market in Brazil: process design scenario-based model. Ind Eng
Chem Res. 2014;53(11):43524365.
24. Zhang BJ, Liu K, Luo XL, Chen QL, Li WK. A multi-period mathematical model for simultaneous optimization of materials and energy
on the refining site scale. Appl Energ. 2015;143:238250.
25. Mouret S, Grossmann IE, Pestiaux P. A new Lagrangian decomposition approach applied to the integration of refinery planning and
crude-oil scheduling. Comp Chem Eng. 2011;35(12):27502766.
26. Menezes BC, Kelly JD, Grossmann IE. Improved swing-cut modeling for planning and scheduling of oil-refinery distillation units. Ind
Eng Chem Res. 2013;52(51):1832418333.
27. Al-Qahtani K, Elkamel A. Multisite facility network integration
design and coordination: an application to the refining industry.
Comp Chem Eng. 2008;32(10):21892202.
28. Al-Qahtani K, Elkamel A. Multisite refinery and petrochemical network design: optimal integration and coordination. Ind Eng Chem
Res. 2009;48(2):814826.
29. Baulin ES, Boronin AB, Khokhlov AS. Rolling detailed short-term
planning of oil refineries and petrochemical complexes and optimization model updating. Autom Remote Control. 2015;76(1):139148.
DOI 10.1002/aic
3039
30. Gonzalo MF, Balseyro IG, Bonnardot J, Morel F, Sarrazin P. Consider integrating refining and petrochemical operations. Hydrocarb
Process. 2004;83:6165.
31. Sadhukhan J, Zhang N, Zhu XX. Analytical optimisation of industrial systems and applications to refineries, petrochemicals. Chem
Eng Sci. 2004;59(20):41694192.
32. Swaty TE. Consider over-the-fence product stream swapping to raise
profitability. Hydrocarb Process. 2002;81:3742.
33. Qin SJ. Process data analytics in the era of big data. AIChE J. 2014;
60(9):30923100.
34. Jaeckle CM, Macgregor JF. Product design through multivariate statistical analysis of process data. AIChE J. 1998;44(5):11051118.
35. Kourti T, Lee J, Macgregor JF. Experiences with industrial applications of projection methods for multivariate statistical process control. Comp Chem Eng. 1996;20(Suppl. 1):S745S750.
36. Kourti T, MacGregor JF. Process analysis, monitoring and diagnosis,
using multivariate projection methods. Chemomet Intell Lab Syst.
1995;28(1):321.
37. Zhang Y, Li S. Modeling and monitoring of nonlinear multi-mode
processes. Control Eng Pract. 2014;22:194204.
38. Misener R, Floudas C. GloMIQO: global mixed-integer quadratic
optimizer. J Glob Optim. 2013;57(1):350.
39. Misener R, Floudas C. ANTIGONE: algorithms for coNTinuous/
integer global optimization of nonlinear equations. J Glob Optim.
2014;59(23):503526.
40. Misener R, Floudas C. A framework for globally optimizing mixedinteger signomial programs. J Optim Theory Appl. 2014;161(3):905932.
3040
DOI 10.1002/aic
41. Misener R, Smadbeck JB, Floudas CA. Dynamically generated cutting planes for mixed-integer quadratically constrained quadratic programs and their incorporation into GloMIQO 2. Optim Method
Software. 2014;30(1):215249.
42. Packie J. Distillation equipment in the oil refining industry. AIChE
Trans. 1941;37(51):5178.
43. Watkins RN. Petroleum Refinery Distillation. Houston, TX: Gulf
Publishing Co., 1979.
44. Brooks RW, Van-Walsem FD, Drury J. Choosing cutpoints to optimize product yields: refining developments: special report. Hydrocarb Process. 1999;78(11):5360.
45. Guerra OJ, Le Roux GAC. Improvements in petroleum refinery planning: 1. formulation of process models. Ind Eng Chem Res. 2011;
50(23):1340313418.
46. Guerra OJ, Le Roux GAC. Improvements in petroleum refinery
planning: 2. Case studies. Ind Eng Chem Res. 2011;50(23):
1341913426.
47. Zhang J, Zhu XX, Towler GP. A simultaneous optimization strategy
for overall integration in refinery planning. Ind Eng Chem Res.
2001;40(12):26402653.
48. Geddes RL. A general index of fractional distillation power for
hydrocarbon mixtures. AIChE J. 1958;4(4):389392.
49. Alattas AM, Grossmann IE, Palou-Rivera I. Refinery production
planning: multiperiod MINLP with nonlinear CDU model. Ind Eng
Chem Res. 2012;51(39):1285212861.
Manuscript received Dec. 20, 2015, and revision received Feb. 9, 2016.
AIChE Journal