Download as pdf or txt
Download as pdf or txt
You are on page 1of 19

Modeling and Optimization of Fluid Catalytic Cracking units

Hamza Fawzi
advised by Larry Biegler (CMU) and Philippe Ricoux (Total)

Contents
1 Introduction 1
1.1 Oil refinery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Fluidized catalytic cracking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Problem statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2 Modeling of a Fluidized Catalytic Cracking unit 3


2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Model equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.3 Discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.4 Nonlinear optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.4.1 Unconstrained optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.4.2 Constrained optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

3 Model Simulation and Optimization 8


3.1 Base case operating conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.2 Simulation results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.3 Optimization of LCO output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.3.1 Effect of operating conditions on LCO output . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.3.2 Optimization with uniform temperature . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.3.3 Optimization with non-uniform temperature . . . . . . . . . . . . . . . . . . . . . . . . . 11

4 Conclusion 12

A Two-Dimensional model 14
A.1 Model equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
A.2 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
About the internship
For my internship, I worked on a research project for TOTAL at Carnegie Mellon University (Pittsburgh,
USA). There, I was hosted by the Center for Advanced Process Decision-making (CAPD) within the Chemical
Engineering Department and worked under professor Larry Biegler. My advisor at TOTAL was Dr. Philippe
Ricoux from the Applied Mathematics and Statistics group in the Scientific Division.
The Scientific Division at TOTAL works relatively closely with the CAPD group at CMU, mainly with
professors Larry Biegler (process control) and Ignacio Grossman (optimization for scheduling and planning). In
fact, the CAPD group collaborates a lot with the industry and its partners include TOTAL, BP, ExxonMobil,
Dow Chemicals, Pfizer, and others. Several graduate students in the group are directly funded by industrial
companies.
This internship has been a very rewarding experience for me and I learned a lot about the modeling and
control of chemical processes and about process systems engineering in general. Even though I was alone on the
project, I had the opportunity to discuss and exchange ideas with Biegler’s PhD students with whom I shared
the office.
The first part of my internship was devoted to literature review to understand and get acquainted with the
FCC process and choose an adequate model that is suitable for our purposes. Once the model was chosen, it
was implemented using the AMPL modeling language, and used as a base for the study. We simulated the
FCC process under different operating conditions and used a solver to optimize the production of certain fuel
components. The solver we used is called IPOPT and was originally developed at CMU by Biegler and his
students. I also had the time in the last weeks of the internship to work on extending the model to take into
account some physical effects in the reactor neglected by the previous model.

1 Introduction
1.1 Oil refinery
Oil refining is the process by which crude oil is refined into more useful petroleum products, such as gasoline,
diesel fuel, kerosene, and others. Crude oil consists of a complex mixture of hydrocarbons with different chain
lengths, plus other organic compounds. The oil refining process starts with a fractional distillation column
that separates the hydrocarbon components of the crude oil according to the boiling points. Figure 1 shows a
schematic representation of the fractionating column.

20 °C Butane
& Propane)
150°C

Gasoline
200 °C
Kerosene
300°C

Crude Oil Diesel


370 °C

Fuel Oil

400 °C

The oil is
heated in a Vacuum Gas Oil
furnace Lubriacting oil.,
Parrafin Wax,
Asphalt

Figure 1. Crude oil is separated into more valuable products by fractional distillation.
The fractions at the top of the column have lower boiling points than the fractions at the
bottom. [27]

After distillation, the heavy fractions of the crude oil at the bottom of the column are further processed and
cracked into lighter, more valuable products such as gasoline, diesel, LPG and others. In most refineries, this

1
Figure 2. Schematic diagram of an FCC unit

conversion is done by catalytic cracking in a unit called FCC (Fluid Catalytic Cracking). The FCC unit is in
fact one of the most important processes in an oil refinery and is sometimes called the workhorse of modern
refinery[2]. Indeed, FCC increase global production of refineries by converting residual oil into high-octane
gasoline and other valuable products. FCC units are present in approximately 400 [26] out of 700 oil refineries
in the world [3]. TOTAL operates a total of 17 FCC units worldwide.

1.2 Fluidized catalytic cracking


A fluidized catalytic cracking process is a unit that converts the vacuum gas oil to gasoline and other middle
distillates using cracking catalyst. The two basic units comprising an FCC are:
• the riser, a tubular reactor (diameter ≈ 1m, height ≈ 35m) where the cracking reaction occurs, and
• the regenerator which regenerates the spent catalyst used in the riser for the cracking reaction.
The vacuum gas oil is injected into the bottom of the riser where it is brought in contact with the hot catalyst
flowing from the regenerator. Thereupon, the feed (vacuum gas oil) is vaporized. The vaporized feed then goes
through endothermic catalytic cracking reaction on its way up through the riser. Lighter hydrocarbons are
produced as main cracking products along with by-product coke, which deposits on the catalyst surface and
lowers the catalyst activity. The riser reactor ends on a disengager which separates the catalyst from the vapor
products. Vapor products enter a fractionator where they are separated into various boiling point fractions.
The (spent) catalyst goes to the regenerator where a simple combustion reaction is used to burn off and remove
the coke from the catalyst surface. The hot regenerated catalyst then flows back to the bottom riser. The high
temperature of the catalyst particles supply the heat required for the vaporization and cracking reaction of the
feed in the reactor.
To give some orders of magnitude, about 45 wt-% of the feed is converted into gasoline (C5 − boiling point
220◦ C), 15 wt-% into diesel (220 − 360◦ C), 25 wt-% into light gases (H2 , C1 − C4 ), 5 wt-% into coke, and 10
wt-% is unconverted. As we can see, the major produce of an FCC unit is gasoline. Actually FCC units were
specifically designed to increase gasoline yield of oil refineries [21], because the gasoline produced alone by the
fractionating column was not sufficient. Moreover, in the 1930’s new higher-compression gasoline engines were
developed and required higher-octane gasoline. The FCC process was then developed in the 1940’s to process
the residual oil of the distillation column and convert it to high octane gasoline [21].

1.3 Problem statement


The demand for gasoline is high in North American countries where motor engines rely mainly on gasoline.
However this is not the case in Europe, where motor engines are mostly diesel engines. With the FCC technology
designed mainly to produce gasoline, TOTAL produces more gasoline than what the European market needs and
sells the excess of gasoline to North America, while in the same time, it buys diesel from foreign oil companies
to meet the demand of the local market.
The question is thus whether it is possible to shift FCC’s production from gasoline to diesel. The problem
is economically very important, and an increase of 3% diesel yield at the expense of gasoline can increase profit
by hundreds of thousands of dollars a day. Other oil companies like PetroChina for example have changed

2
the design of the unit in order to achieve higher diesel yield at the expense of gasoline [18]. However, and as
mentioned earlier, TOTAL already operates 17 FCC units, and it is evident that the modification of the design
would be extremely costly to do on the existing FCCs. The objective of the study is therefore to assess whether
it is possible to increase diesel yield at the expense of gasoline yield, only by changing the operating conditions
of the unit (pressure, temperature, catalyst amount, etc.), which would be relatively easy to implement on
currently operating FCCs.
Section 2 deals with the mathematical modeling of FCC riser reactors. In section 3, the simulation results
as well as the optimization results are presented and interpreted.

2 Modeling of a Fluidized Catalytic Cracking unit


2.1 Introduction
In this study, we are concerned with the yields of the different products; the regeneration of spent catalyst is
not in the scope of the study. We will therefore focus only on the modeling of the riser reactor.
The combination of different physical and chemical phenomena (heat transfer, cracking reaction, multi-phase
flow, turbulence phenomenon) make the FCC riser reactors very complex systems that are difficult to model
and simulate. A lot of effort has been devoted to better understand the complexity of the riser reactor, and the
literature is abundant on the subject.
The initial attempts to model FCC riser reactors were primarily oriented towards the kinetics of cracking,
and assumed very simple hydrodynamics. The authors of these kinetic models grouped the molecules of the
flowing gas in distillation cuts (e.g., gasoline, diesel, LPG, etc.) and considered pseudo-chemical reactions
between these groups (also called lumps). The first kinetic model based on this lumping scheme is the 3-lumps
model by Weekman [25]. This model was then further developed to incorporate more lumps. Widely used
models are the 4-lumps model by Lee et al. [14], the 5-lumps model by Takatsuka et al. [19], and the 10-lumps
model by Jacob et al. [12]. Pitault et al. also developed a more detailed and complex model involving 19 lumps
[17].
More recent models of the riser reactor incorporate more or less complex hydrodynamics with the cracking
kinetics to account for various effects such as non-uniformity of velocity, temperature and pressure along the
riser, radial dispersion of gas components, and others. These models span a wide range of complexity and go
from the relatively simple to the very complex. Indeed a few very complex models were developed to capture
the 3D geometry of the riser, turbulence phenomenon, and the complexity of the injection zone [1, 7, 20]. Such
very complex models can however only be used for simulation purposes and are not adequate for control and
optimization.

2.2 Model equations


The FCC riser model used in this study is the one presented in [6] and [5]. In [6] the authors provide a detailed
model of an FCC unit, and in [5], this model is validated against industrial data collected from a plant in Sines
(Portugal) and operated by Galpenergia.
The cracking kinetic model used is the six-lumps scheme originally developed by Takatsuka et al. in [19]
and adopted by Fernandes et al. in [6]. The lumps considered are:
• Vacuum Gas Oil or VGO (> 360◦ C)
• Light Cycle Oil or LCO (220 − 360◦ C)
• Gasoline or GLN (C5 – 220◦ C)

• Light Petroleum Gas or LPG (C3 and C4)


• Fuel Gas or FG (H2, C1, C2 and H2S)
This model can predict the output of diesel through the presence of the LCO lump.
The reactions between the different lumps are presented in Figure 3; the kinetics of these reactions are
presented in Table 1.

3
Figure 3. Cracking reactions in six-lumps model [5]

Table 1. Reactions kinetics [5]

Reaction Rate of reaction* ri→j [kg.m-3 .s-1 ]


VGO → LCO 2.09 × 1010 MMI
−0.43
API−1.5 Φ(Yck ) exp(−84.2 × 103 /RT )yVGO
2

VGO → GLN 3.98 × 10 MMI API Φ(Yck ) exp(−84.2 × 10 /RT )yVGO


6 −0.42 1.62 3 2

VGO → LPG 1.58 × 106 MMI


−0.52
API1.7 Φ(Yck ) exp(−84.2 × 103 /RT )yVGO
2

VGO → FG 7.94 × 10 MMI S Φ(Yck ) exp(−84.2 × 10 /RT )yVGO


6 −0.22 0.14 3 2

VGO → CK 3.98 × 107 MMI


−0.43
Φ(Yck ) exp(−84.2 × 103 /RT )yVGO
2

LCO → GLN 5.01 × 10 MMI Φ(Yck ) exp(−77.1 × 10 /RT )yLCO


7 −0.42 3

LCO → CK 1.58 × 107 MMI


−0.42
Φ(Yck ) exp(−77.1 × 103 /RT )yLCO
GLN → LPG 1.58 × 10 MMI Φ(Yck ) exp(−146.0 × 103 /RT )yGLN
12 −0.52

GLN → FG 1.00 × 1010 MMI


−0.22
Φ(Yck ) exp(−146.0 × 103 /RT )yGLN
LPG → FG 1.00 × 10 MMI Φ(Yck ) exp(−193.0 × 103 /RT )yLPG
13 −0.22

*
Catalyst deactivation function Φ(Yck ) = (1 + 477Yck )−1.66

In the rate of reactions equations in table 1, Φ is the catalyst deactivation and is a function of Yck which
is the amount of coke on the catalyst. Indeed, as mentioned previously, the presence of coke on the catalyst
deactivates the catalyst. The function Φ quantifies this deactivation. Φ is thus a decreasing function of Yck
(the more there is coke on the catalyst, the less it is active) and is equal to 1 for Yck = 0, and is equal to 0 for
Yck = ∞.
The riser was modeled as a one-dimensional plug flow reactor without radial dispersion. The riser was further
assumed to be in steady state. The flow in the riser consists of two phases : a gas phase (hydrocarbons) and
a solid phase (catalyst and coke). The velocity profiles along the riser of both phases are accounted for in the
model. In fact, the velocity of the gas phase increases significantly along the riser due to molar expansion (VGO
is cracked into lighter products with lower molecular weights, which increases gas mole content and decreases
gas density; lower density increases gas velocity because of conservation of mass). The catalyst (solid) phase
velocity is derived from Newton’s law applied on a cluster of catalyst particles. The gas phase is assumed to
satisfy ideal gas law. No energy balance equation was considered since temperature was taken to be a control
variable in our study. The model equations are given below:
Conservation of mass
Fg dyi

 (1 − ε )S dz = Ri for i 6= CK


g
(1)
Fc dYck
= Rck


(1 − εg )S dz

where Ri is the formation rate of lump i:


X X
Ri = rj→i − ri→j .
j j

Momentum balance in solid phase 1

dvc 3 ρg (vg − vc )2
ρ c vc = CD + (ρg − ρc )g with CD = 0.44 (2)
dz 4 dcl
24
1 In [6], the drag coefficient is : CD = Rep
(1 + 0.15Rep0.687 ) if Rep < 1000, and 0.44 otherwise where Rep = |vg − vc |dcl ρg εg /µg
is the Reynolds number of the gas phase (µg being the gas viscosity). We assumed in our model that CD = 0.44 always; we have
observed no major difference with the more complex expression for the drag coefficient CD .

4
collocation points

z0 z1 z2

h1 = z 1 − z 0

Figure 4. Finite elements and collocation points

Overall momentum balance on gas and solid phase


dvg dvc dP
ε g ρ g vg + (1 − εg )ρc vc =− − (εg ρg + (1 − εg )ρc )g (3)
dz dz dz
Ideal gas law X yi
P = ρg RT (4)
i
Mi
Gas volume fraction
Fc
εg = 1 − (5)
vc ρ c S
Gas flow
Fg
=1 (6)
vg ρg εg S
There are 11 unknown variables in the model: yi (i = VGO, LCO, GLN, FG, LPG, CK), vg , vc , P, ρg , and εg ,
and there are 11 independent equations. Therefore the model is closed.
The initial conditions for the differential equations are given by:
(
(0) 1 i = VGO
yi =
0 i = others

Fg Fc
vg(0) = (0) (0)
, vc(0) = (0) (0)
,
ε g ρg S (1 − εg )ρg S
where !−1
P (0) X y (0)
ρ(0)
g = i
, with P (0) = 310 kPa [5]
RT i
Mi
and
(0)
Fg /ρg
ε0g = (0)
Fg /ρg + Fc /ρc

2.3 Discretization
The model was fully discretized using the method of orthogonal collocation on finite elements. In this method,
the axial coordinate is divided into a number of finite elements 0 = z0 < z1 < . . . < zNE = Hris , and the
variables are approximated by polynomials on each finite elements.
We use the so-called monomial basis representation for the differential state profiles: If x is a differential
variable, we use the Lagrange representation on the collocation points for the derivative dx
dz :

K
z − zi−1 dx
 
dx X
(z) = ψq for zi−1 6 z 6 zi
dz q=1
hi dz i,q

with ψq the Lagrange polynomial that verifies ψq (sq ) = 1 and ψq (sj ) = 0 for j 6= q:
K
Y s − sj
ψq (s) = .
sq − sj
j=1,j6=q

5
The sj , j = 1, . . . , K are the collocation points. Writing the monomial basis representation in terms of the
differential variable x directly gives:
K
z − zi−1
 
X dx
x(z) = x(zi−1 ) + Ωq for zi−1 6 z 6 zi
q=1
hi dz i,q

where Z ρ
Ωq (ρ) = ψq (s)ds for 0 6 ρ 6 1.
0

It can be shown that this discretization scheme corresponds to an Implicit Runge Kutta method [22].
In our study, a total of 40 finite elements were used to discretize the axial coordinate of the riser, with 3
collocation points per finite element. The collocation points were chosen to be the roots of some polynomial
1 dK
of degree 3, namely, P̃3 − P̃2 where P̃K is the shifted Legendre polynomial of order K (P̃K (τ ) = K! dτ K
[(τ 2 −
1) ]). This choice of collocation points is known as the Radau quadrature scheme, and is widely used in the
K

discretization of systems of Differential and Algebraic Equations. The finite elements were chosen to be more
dense at the first section of the riser than at the end, for the dynamics of the riser occur mainly in the first few
meters of the riser. Specifically, the ith finite element is at position:
1h i
zi = (1 + 2Hris )(i−1)/NE − 1 for i = 1, . . . , NE
2
Figure 5 depicts this distribution of finite elements.

40

35

30
Finite Element Number

25

20

15

10

0
0 5 10 15 20 25 30 35
Riser height (m)

Figure 5. Distribution of the finite elements

The model was then implemented in the AMPL modeling language and solved using IPOPT [24], an open
source software package for large-scale nonlinear optimization originally developed at CMU by Biegler and his
students.

2.4 Nonlinear optimization


In nonlinear optimization one seeks to find a local (or if possible global) solution of a constrained optimization
problem of the form
minimize f (x)
subject to g(x) ≥ 0 (7)
h(x) = 0,
where x ∈ R , f : R → R is the objective function, g : Rn → RnI are the inequality constraints, and
n n

h : Rn → RnE the equality constraints.


IPOPT uses a so called primal-dual interior-point method to solve this type of optimization problems. In
this section I will briefly describe the general ideas behind the algorithm used by IPOPT. The original IPOPT
paper contains a detailed explanation of the algorithm [24].

2.4.1 Unconstrained optimization


If the problem is unconstrained, that is, if we are simply looking to minimize a function f over Rn , then a
necessary condition for a point x? to be a solution is ∇f (x? ) = 0. One way to approach the unconstrained

6
optimization problem would thus be to solve the equation ∇f (x∗ ) = 0 using Newton’s method for example.
This would yield the following iterations scheme:

xk+1 = xk − [∇2 f (xk )]−1 ∇f (xk )

If the initial point x0 is well chosen, this sequence xk will converge to a point x? satisfying ∇f (x? ) = 0. One
could then use sufficient conditions for optimality to guarantee that x? is indeed a local minimum for f (e.g.
∇2 f (x? ) positive definite).

2.4.2 Constrained optimization


In the previous unconstrained case, we used the necessary condition ∇f = 0 based on the gradient of f to
search for potential solutions of the problem. In the constrained problem, there also exists first-order necessary
conditions that have to be met by any solution of the problem. These conditions are known as the Karush-
Kuhn-Tucker (KKT) conditions:
Theorem 1. (KKT first-order necessary conditions, [16]) Let x? be some point in Rn . Assuming x? verifies
(LICQ), then if x?P
is a local solution
Pof 7, then there exists vectors λ? ∈ RnI and ν ? ∈ RnE such that:
nI nE
(i) ∇f (x ) − i=1 λi gi (x ) − j=1 νj? hj (x? ) = 0
? ? ?

(ii) g(x? ) ≥ 0 (primal feasibility)

(iii) h(x? ) = 0 (primal feasibility)

(iv) λ?i ≥ 0 (dual feasibility)

(v) λ?i gi (x? ) = 0 for all i = 1, . . . , nI (complementary slackness)


The λ and µ are the Lagrange multipliers. LICQ (Linear Constraint Qualification) is a regularity condition
for KKT theorem to apply: x? is said to verify LICQ if the vectors ∇hj (x? ) for j = 1, . . . , nE and ∇gi (x∗ ) for
i such that gi (x∗ ) = 0 (active constraints), are linearly independent.
The main purpose of an optimization algorithm will (in general) be to look for points x? that satisfy the
KKT conditions. The problem here is harder than in the previous unconstrained case where we only had to solve
an equation, ∇f = 0. Indeed the KKT conditions in the general case contain both equations and inequations,
and it is in fact the inequations (ii) and (iv) that make the task much more difficult.
The algorithm used by IPOPT is called an interior point method. Interior-point methods look at solving
the system of equations and inequations (i)-(v) through a Newton method2 , while ensuring that each iterate of
the Newton method satisfies, strictly, the two inequations (ii) and (iv).
Let F (x, λ, ν) be the set of equations of the KKT conditions (F : Rn+nI +nE → Rn+nE +nI ):
PnI PnE
∇f (x) − i=1 λi gi (x) − j=1 νj hj (x)
 

F (x, λ, ν) =  h(x) ,
λ • g(x)
def
where λ • g(x) = (λi gi (x))1≤i≤nI ∈ RnI . One idea to find a point satisfying the KKT conditions is to apply
Newton’s method to solve F = 0 and modify the step length at each iteration so that the two inequations λ ≥ 0
and g(x) ≥ 0 are satisfied strictly. This algorithm could be written as:

1. Start with x0 , λ0 , ν0 .
2. While not converged do
• Calculate the Newton search direction to solve F = 0. Call this direction (∆xk , ∆λk , ∆νk ).
• Search for αk (step length) such that (xk , λk , νk ) + αk (∆xk , ∆λk , ∆νk ) satisfies the inequations (ii)
and (iv) strictly.
• Let (xk+1 , λk+1 , νk+1 ) = (xk , λk , νk ) + αk (∆xk , ∆λk , ∆νk ).

In practice the above algorithm does not perform well, because the step lengths αk are often very small.
The direction provided by Newton’s method on F = 0 does not allow us to take long steps when we look at
satisfying the inequations (ii) and (iv). Most interior-point methods will thus apply Newton’s method on a
modified set of equations and not on F = 0 exactly. The condition (v) is then replaced by

(v’µ ) λ?i gi (x? ) = µ for all i = 1, . . . , nI


2 Actually a modified Newton method...

7
where µ is some positive scalar. The modified equations can be written Fµ = 0 where Fµ is:
PnI PnE
∇f (x) − i=1 λi gi (x) − j=1 νj hj (x)
 

Fµ (x, λ, ν) =  h(x) .
λ • g(x) − µ

By applying Newton’s method on this modified set of equations instead of F = 0, it is shown in practice
that one can take much longer step lengths that still satisfy the inequations (ii) and (iv). Indeed, by putting
λi gi (x) = µ > 0 instead of λi gi (x) = 0, the Newton search direction will be somehow biased towards the set
of λ ≥ 0 and x such that gi (x) ≥ 0, and we will thus be able to take longer steps when we look to satisfy the
inequations g(x) ≥ 0 and λ ≥ 0 [28]. The parameter µ is generally decreased to 0 the closer we get to a solution,
so that we recover the original system of equations F = 0 at the end. This gives the following algorithm

1. Start with x0 , λ0 , ν0 , and µ0 .


2. While not converged do
• Calculate the Newton search direction to solve Fµk = 0. Call this direction (∆xk , ∆λk , ∆νk ).
• Search for αk (step length) such that (xk , λk , νk ) + αk (∆xk , ∆λk , ∆νk ) satisfies the inequations (ii)
and (iv) strictly.
• Let (xk+1 , λk+1 , νk+1 ) = (xk , λk , νk ) + αk (∆xk , ∆λk , ∆νk ).
• Update µk .

3 Model Simulation and Optimization


3.1 Base case operating conditions
The base case operating conditions of the riser are given in [5]. They are summarized here in Table 2.

Fresh feed flowrate Fg 70.65 kg/s


Catalyst/Oil ratio 7.1
Temperature 800 K

Table 2. Base case operating conditions [5]

The Catalyst/Oil Ratio (amount of catalyst injected in the reactor / amount of feed oil injected in the
reactor, ratio also written as C/O or COR) is not explicitly given in [5], but we could infer an approximate
value from the simulation results presented there. In particular in Figure 11c of [5] we can see that for a feed
flowrate of 70.65 kg/s, the catalyst flow rate is slightly above 500 kg/s; a C/O ratio of 500/70.65 ≈ 7.1 was
therefore adopted.
Furthermore, we have assumed that the temperature is constant equal to 800 K throughout the riser:
Simulation results presented in [5] show indeed that one can consider the temperature to be nearly uniform in
the riser (at least after the first 2 meters where the temperature is decreasing due to the endothermicity of the
cracking reaction).

3.2 Simulation results


Figure 6 shows the profiles of the different variables along the riser axial coordinate. The graphs match the
simulation results presented in [5]. We see that the cracking reactions mainly occur in the first few meters of
the riser and after that the gas composition is quite uniform. This is expected because as more and more coke
is deposited on the catalyst, the catalyst is deactivated and the cracking reaction is much slower. Because VGO
gets cracked into lighter products, there is significant molar expansion (especially in the first few meters of the
riser) which causes the gas velocity to increase. The catalyst velocity increases too because of the interaction
force between the two phases. The model adequately predicts a slip factor (ratio gas velocity / catalyst velocity)
higher than 1, as reported in [1] for example. Also due to the high molar expansion, the gas volume fraction
increases throughout the riser (for nearly uniform pressure) which can be seen in the figure as well.

8
a) 1 b) 14

12
0.8 Gas
Mass fractions [wt%]

10

Velocity (m/s)
0.6
GLN
8
0.4 Catalyst
6
VGO LG
0.2
LCO 4

CK
0 2
0 5 10 15 20 25 30 35 0 5 10 15 20 25 30 35
Riser height (m) Riser height (m)
5
x 10
c) 3.1 d) 0.09

0.08
3.05

Catalyst volume fraction


0.07
Pressure (Pa)

3 0.06

0.05
2.95
0.04
2.9
0.03

2.85 0.02
0 5 10 15 20 25 30 35 0 5 10 15 20 25 30 35
Riser height (m) Riser height (m)

Figure 6. Simulation results: profiles along riser height for base case operating conditions
(in graph a/, LG – for Light Gases – is the sum of LPG and FG)

3.3 Optimization of LCO output


3.3.1 Effect of operating conditions on LCO output
We studied the effects of C/O ratio, temperature and feed flowrate on LCO and other lumps’ output using the
model presented in the previous section. Figure 7 shows the effect of each of these operating conditions on the
gas composition at the riser outlet.
In Figure 7(a), the effect of C/O ratio on LCO was examined. We first observe that the base case (C/O
= 7.1) corresponds to the maximum gasoline production (indeed FCC’s main purpose is to produce gasoline).
If we increase the C/O ratio from its nominal value of 7.1, the LCO decreases. Indeed increasing C/O ratio
induces higher cracking rates and therefore more light products (FG, LPG, GLN) will be created at the expense
of heavier products (LCO). However if we decrease C/O ratio from its base case value, the LCO increases since
we reduce the cracking activity, but the VGO also naturally increases (lower global conversion).
Figure 7(b) shows the effect of temperature. The graph is quite similar to the previous one. Indeed tem-
perature and catalyst amount are both agents that control cracking reaction rate in the same way. Higher
temperature means higher cracking rate and therefore more light product and less heavy product. Low temper-
ature means lower cracking rate and therefore more heavy products and less light products.
Finally the effect of feed flowrate on LCO is considered in Figure 7(c). In this figure we observe that
increasing the feed flowrate increases the amount of LCO at the riser outlet. Indeed when the feed flowrate is
increased the residence time is decreased and we have less cracking, and therefore we have more LCO.

3.3.2 Optimization with uniform temperature


In this section we examine the results of the optimization of LCO output with the three control variables (C/O,
Fg , T ). The results of this optimization are given in Table 3 (case 1).
We observe an increase of nearly 2% in LCO content compared to the base case. Concerning other compo-
nents we see that gasoline content has slightly decreased, and that light products have decreased significantly as
well as the global conversion (VGO content went from 8.5% to 17.4%). At the optimal point, temperature has
not really changed from its nominal value (800 K). C/O ratio has significantly decreased however (as suggested
by 7(a)), but feed flowrate has decreased from its base case value, contrary to the observation done in the
previous section with Figure 7(c) (higher Fg ⇒ more LCO).

9
Fg = 70.65 kg/s, T = 800 K Fg = 70.65 kg/s, C/O = 7.1
0.16 0.6 0.16 0.8

GLN
LCO
0.14 0.6
LG

VGO, GLN, LG [wt%]

VGO, GLN, LG [wt%]


0.14 0.4
LCO [wt%]

LCO [wt%]
LCO GLN
LG 0.12 0.4

0.12 0.2
0.1 0.2
VGO
VGO

0.1 0 0.08 0
3 4 5 6 7 8 9 10 600 650 700 750 800 850 900 950 1000
Catalyt/Oil ratio Temperature [K]

(a) (b)

C/O = 7.1, T = 800 K

0.15 0.5
GLN

0.14 0.4
LCO

VGO, GLN, LG [wt%]


LCO [wt%]

0.13 0.3

LG
0.12 0.2

VGO
0.11 0.1

0.1 0
50 100 150 200 250 300
Feed flowrate [kg/s]

(c)

Figure 7. Effect of C/O ratio, temperature and feed flowrate on gas composition at
riser outlet. A vertical dashed line on each graph indicates the base case value for each
parameter.

If we plot the curves of LCO content versus feed flowrate (as in Figure 7(c)) for different values of C/O ratio
we obtain Figure 8. We see that the thick line that corresponds to the base case (C/O=7.1) moves to the left
as the C/O ratio decreases, and that the optimal value of LCO content is nearly the same for all curves.
Indeed, one can show that, in a first approximation, LCO content is characterized by the ratio of C/O by
Fg only and not by the individual values of C/O and Fg : We know that the LCO content (and any other lump
content) is a function of τcat , the amount of time the catalyst is in contact with the feed. This is given by:

Wcat
τcat = ,
Fg

where Wcat is amount of catalyst (in kilograms) in the riser. If tcat is the residence time of catalyst in the riser,
then (with COR being the C/O ratio):

Hris Hris
Wcat = Fc · tcat = COR · Fg · tcat = COR · Fg · ≈ COR · Fg · (assuming vc ∝ vg )
vc Kvg

therefore
Hris COR
τcat = COR · ∝
Kvg Fg
since vg ∝ Fg through density. This shows that to a first approximation, the maxima of the different curves
shown in Figure 8 correspond to a same ratio COR / Fg . In Table 4, we have reported the value of these
maxima and their corresponding COR / Fg ratio. We can see that they are roughly centered around 0.040.
Figure 9 also shows a scatter plot of LCO versus COR/Fg that is obtained through multiple simulations with
different values of COR and Fg . The plot looks approximately like a function (for a same value of COR/Fg ,

10
Table 3. Optimization on 1D model. Underlined Figures are fixed (not optimized).
Emphasized Figures are optimized variables

Operating conditions Mass fractions at riser outlet


Fgo [kg/s] C/O T [K] LCO GLN LG* VGO Yck
Base case 70.6 7.1 800 12.4% 49.9% 24.2% 8.5% 0.8%
Case 1 59.0 2.8 797 14.3% 47.3% 17.1% 17.4% 1.2%
Case 2a 114 22 a T ? (z)b 15.3% 10.7% 63.5% 5.9% 0.1%
Case 2b 171 22 a T ? (z)c 15.1% 11.6% 58.4% 10.8% 0.1%
*
LG = Light Gases = Fuel Gas + Light Petroleum Gas
a
Upper bound constraint of 22 reached
b
The optimal temperature profile T ? (z) is a bang-bang type profile
where T is 1220 K on the first 7 centimeters and 670 K (lower bound)
on the rest of the riser.
c
T ? (z) has the same bang-bang shape (as in case 2) with the long low-
temperature plateau being at 700 K (instead of 670 K for case 2).

T = 800 K
0.15

0.145 C/O = 5

0.14

0.135 C/O = 4
C/O = 3
LCO [wt%]

0.13
C/O = 7

0.125

0.12

0.115

0.11

0.105
60 80 100 120 140 160 180 200 220 240 260 280
Feed flowrate [kg/s]

Figure 8. Mass fraction of LCO at riser outlet as a function of Fg for different values of
C/O

LCO content is same), which is adequate with the previous observation. This function is first increasing then
decreasing: At first, increasing τcat increases the cracking rate which produces more LCO. When we increase
τcat after a certain point however, the amount of LCO decreases due to secondary cracking of LCO into lighter
products (namely gasoline and coke, as shown in reactions diagram Figure 3).

Table 4. Relation between LCO content and COR/Fg ratio

COR 2 3 4 5 6 7 8 9 10
COR/Fg at optimal point 0.040 0.050 0.044 0.040 0.036 0.036 0.036 0.037 0.038
LCO content 0.143 0.144 0.144 0.143 0.143 0.142 0.142 0.141 0.140

3.3.3 Optimization with non-uniform temperature


We will now optimize LCO output with the same three control variables (C/O, Fg , and T ) but unlike in the
previous section where temperature was uniform in the riser, we will allow here the temperature to have a
profile along the riser’s axial coordinate. The results of this optimization are given in Table 3 (case 2a). An
increase in LCO by nearly 3% was achieved compared to the base case, while gasoline decreased significantly at
the expense of light gases, which went up by nearly 40%. Total conversion went up as well with a lower VGO
content (5.9%) than for the base case (8.5%).
The optimal temperature profile was a bang-bang type profile where T is 1220 K (upper bound 1270 K) on
the first 7 centimeters of the riser, and then 670 K (lower bound 670 K) for the rest of the riser. The C/O ratio
also reached its upper bound of 22 that was imposed during the optimization: In the first few centimeters of

11
T = 800K
0.16

0.14

LCO [wt%]
0.12

0.1

0.08
0 0.05 0.1 0.15 0.2
COR/Fg

Figure 9. Scatter plot of LCO versus COR/Fg ratio. The points lie approximately on a
function curve, that is for a same ratio COR/Fg but for different values of COR and Fg ,
the LCO content is the same.

the riser, the temperature and catalyst are very high and so the cracking rate is very high too. A lot of LCO
is produced, along with gasoline, light gases and coke. After enough LCO is produced, the reaction scheme
in Figure 3 shows that LCO should be cracked into lighter products (namely gasoline, and coke). The sudden
decrease of temperature to 670 K after the first few centimeters is meant to freeze these cracking reactions that
would produce light products at the expense of LCO. If we change the lower bound of 670 K to 700 K and
re-run the optimization, we get the results shown in line ‘case 2b’. We observe that we have less LCO and more
gasoline, because the reaction LCO→GLN was “less frozen” than in the previous case.

Case 1 vs. Case 2 If we compare the optimization results of cases 1 and 2 (uniform temperature vs.
temperature profile), we see that gas composition are significantly different. With a uniform temperature (case
1) the increase in LCO content is achieved at the expense of light gases and vacuum gas oil, which is undesirable.
In this case, we still have a lot of gasoline, and we have less LG and more VGO, which are all undesirable effects.
However, imposing a bang-bang type temperature profile (case 2) achieves higher LCO output while reducing
gasoline significantly and increasing light gases and global conversion, which are all very desirable effects.

4 Conclusion
The study has revealed some promising results and has shown that by imposing a certain temperature profile,
one can increase the FCC diesel yield by 3 points while reducing gasoline production significantly. These results
were obtained using a relatively simple five-lumps model with 1D hydrodynamics, and the next step should be
of course to validate the results using a more complex and detailed FCC model. We allude here to some of the
points one should consider in the validation step:
For the cracking kinetics, one can observe that the five-lumps model used in this study does not accurately
describe the activation energies of the reactions involved (see figure 3). Indeed looking at table 1, we see that
the activation energies are assumed to depend only on the type of reactant, that is for a reaction i → j, the
activation energy is a function of i only and not of j. By looking at other cracking kinetic model, we see that
this assumption does not hold in general, and that the activation energies can be quite different for two reactions
where reactant is same. For instance, in the four-lumps model used by Han et al. [9], we see that there is a
factor 2 between the activation energies of VGO → GLN and of VGO → CK. The activation energies would
need to be revised when validating the results, and this is especially important here since our results are based
on temperature profiles.
The hydrodynamic model also does not take into account several complex phenomena. All quantities are
assumed to be uniform radially, even though it was proved experimentally that this it is not the case [15, 4, 23].
It was shown in these references that catalyst concentration has a “core-annulus” structure i.e. is approximately
constant in the core, and then increases very rapidly near the wall. Velocity profiles as well as other quantities
were also shown to be radially-dependent. In fact, during the internship we worked on extending the 1D model
presented in this study to a 2D model that takes into account the radial dependency of the different quantities.
The 2D model was implemented and simulated against a base case. Simulation results looked satisfying, but
due to lack of time, we did not have time to arrive at concluding and valid optimization results with this model.
The 2D model along with the simulation results is included in appendix A.
Another complex phenomenon one might need to account for in the validation step is the feed injection zone
which is known to be a very turbulent area where feed is vaporized when brought in contact with catalyst.
The modeling of this zone is especially important here, because of the shape of the optimal temperature profile

12
(bang-bang).
Finally, one also has to make sure that the model discretization is still valid for such sharp variations as the
ones observed in the optimal temperature profile.

Notations
εg Gas volume fraction (= volume of gas +volume of gas
volume of catalyst particles ) [-]
ρc Catalyst density [1382 kg/m ] 3

ρg Gas density [kg/m3 ]


C/O, COR Catalyst to Oil ratio [-]
dcl Cluster diameter [2.0 × 10−3 m]
Fc Catalyst flowrate [kg/s]
Fg Feed flowrate [kg/s]
Hris Riser height [37.11 m]
Mi Molar weight of lump i, given in Table 5 [kg/mol]
P Pressure [Pa]
R Ideal gas law constant [8.31451 J/mol/K]
D Riser diameter [1.39 m]
S Riser cross-section [m2 ]
T Temperature [K]
vc Catalyst velocity [m/s]
vg Gas velocity [m/s]
yi Weight fraction of lump i [-]
Yck Coke content on catalyst [kg-coke/kg-catalyst]

Table 5. Average Molecular Weights of the Lumps in the Kinetic Model [5, 11]

Lump Molecular weight [kg/mol]


Vacuum Gas Oil, VGO 0.4
Light Cycle Oil, LCO 0.2
Gasoline, GLN 0.1
Liquefied Petroleum Gases, LPG 0.05
Fuel Gas, FG 0.025
Coke, CK 0.4

13
A Two-Dimensional model
In this section, we present the extension of the 1D model presented previously to a 2D model where radial
dependence is taken into account. All quantities are now radially dependent, and a radial dispersion term was
added in the mass balance equation. Furthermore, the velocity radial profile at the riser bottom is taken from
(0)
[15], and is given by vg (r, z = 0) = vg,av 73 1 − (r/Rr )1.5 .


A.1 Model equations


Conservation of mass
(
Ri is the formation rate of lump i, and
 
∂yi Dr ∂ ∂ρg yi
= r + (1 − εg )Ri , where (8)
∂z r ∂r ∂r Dr the radial dispersion coefficient

Momentum balance in solid phase

∂vc 3 ρg (vg − vc )2
ρ c vc = CD + (ρg − ρc )g (9)
∂z 4 dcl
Momentum balance for gas+solid phase
∂vg ∂vc ∂P
ε g ρ g vg + (1 − εg )ρc vc =− − (εg ρg + (1 − εg )ρc )g (10)
∂z ∂z ∂z
Ideal gas law X yi
P = ρg RT (11)
i
Mi
Gas volume fraction
Fc
εg = 1 − (12)
vc ρ c S
Gas flow
Fg
=1 (13)
vg ρg εg S
Riser inlet conditions (
1 i = VGO
yi (r, z = 0) = (14)
0 i = others
7
vg (r, z = 0) = vg,av
(0)
1 − (r/Rr )1.5 (15)

3
(0) 7
vc (r, z = 0) = vc,av 1 − (r/Rr )1.5 (16)

3
where
Fg Fc
(0)
vg,av = (0) (0)
, (0)
vc,av = (0) (0)
,
ε g ρg S (1 − εg )ρg S
where !−1
X yi (z = 0) (0)
P (0) Fg /ρg
ρ(0)
g = , with P (0) = 310kPa, and ε(0)
g = (0)
RT i
Mi Fg /ρg + Fc /ρc
Boundary conditions
Riser center :
∂yi ∂P
(r = 0, z) = 0 (r = 0, z) = 0 (17)
∂r ∂r
Riser wall :
∂yi ∂P
(r = Rr , z) = 0 (r = Rr , z) = 0 (18)
∂r ∂r

14
A.2 Simulation
The model was fully discretized like the 1D model. Two finite elements and two collocation points per finite
element were used for the radial direction.

T = 800 K, Fg = 70.65 kg/s, CO = 6.0


0.8
VGO
0.7 LCO
GLN
0.6 FG+LPG
CK
Weight fractions

0.5

0.4

0.3

0.2

0.1

0
0 5 10 15 20 25 30 35 40
Riser height (m)

Figure 10. Yields of lumps vs. riser height position

z=0.0m z=1.5m z=5.2m z=14.3m z=37.1m


0.25 20 15
Solid phase velocity (m/s)
Catalyst volume fraction

0.2
15
Gas velocity (m/s)

10
0.15
10
0.1
5
5
0.05

0 0 0
0 0.5 1 0 0.5 1 0 0.5 1
r/R r/R r/R

Figure 11. Radial profiles of catalyst concentration (left), gas velocity (center), solid
phase velocity (right). These match the experimentally found profiles in [15] and [4].
The catalyst concentration exhibit a “core-annulus” structure, and we can see that the
concentration is much higher near the wall than in the center of the riser.

15
z=0.0m z=1.5m z=5.2m z=14.3m z=37.1m
1 0.2 0.8
VGO weight fraction

LCO weight fraction

GLN weight fraction


0.8
0.15 0.6
0.6
0.1 0.4
0.4
0.05 0.2
0.2

0 0 0
0 0.5 1 0 0.5 1 0 0.5 1
r/R r/R r/R
−3
x 10
0.2 0.04 8
LPG weight fraction

FG weight fraction

CK weight fraction
0.15 0.03 6

0.1 0.02 4

0.05 0.01 2

0 0 0
0 0.5 1 0 0.5 1 0 0.5 1
r/R r/R r/R

Figure 12. Radial profiles of lumps’ concentrations. The catalyst concentration is higher
near the wall, and therefore cracking reaction rates are higher there. This explains why
the concentration of light products is higher in the annulus near the wall than in the core.

The three following graphs show the effects of temperature, Catalyst/Oil ratio, and feed flowrate on the
yields of the different components. These graphs follow the same trends as in the 1D model (see figures 7(a),
7(b), and 7(c)).

C/O = 6.0, Fg = 70.65kg/s


0.13 0.7
GLN
VGO
0.125 FG+LPG 0.6
LCO
GLN, VGO, FG+LPG (wt%)

0.12 0.5
LCO (wt%)

0.115 0.4

0.11 0.3

0.105 0.2

0.1 0.1

0.095 0
600 650 700 750 800 850 900 950
Temperature (K)

Figure 13. Effect of temperature

16
T = 800 K, Fg = 70.65 kg/s
0.13 0.6
GLN
VGO
FG+LPG
LCO

GLN, VGO, FG+LPG (wt%)


0.12 0.4

LCO (wt%)

0.11 0.2

0.1 0
2 3 4 5 6 7 8 9 10
Catalyst/Oil ratio

Figure 14. Effect of Catalyst/Oil ratio

C/O = 6.0, T = 800K


0.15 0.5
LCO
GLN
VGO
FG+LPG

GLN, VGO, FG+LPG (wt%)


0.14 0.4
LCO (wt%)

0.13 0.3

0.12 0.2

0.11 0.1
50 100 150 200 250 300
Feed flowrate (kg/s)

Figure 15. Effect of feed flowrate

Optimization on the 2D model exhibited results that were quite different from the optimization results
obtained for the 1D model presented above. Due to lack of time, we did not have time to investigate and
validate further the 2D model to understand where the differences originated from.

References
[1] D. A.K., B. E., M. G.B., and H. G.J. Three-dimensional simulation of a fluid catalytic cracking riser
reactor. Industrial & Engineering Chemistry Research, 42:2602–2617, 2003.
[2] A. Arbel, Z. Huang, I. Rinard, R. Shinnar, and A. Sapre. Dynamic and control of fluidized catalytic
crackers. I: Modeling of the current generation of FCC’s. Industrial & engineering chemistry research,
34(4):1228–1243, 1995.
[3] I. BILLEGE. 700 Refineries Supply Oil Products to the World. NAFTA, 60(2):401–403, 2009.
[4] C. Derouin, D. Nevicato, M. Forissier, G. Wild, and J. Bernard. Hydrodynamics of riser units and their
impact on FCC operation. Ind. Eng. Chem. Res, 36(11):4504–4515, 1997.

17
[5] J. L. Fernandes, C. I. C. Pinheiro, N. M. C. Oliveira, J. Inverno, and F. R. Ribeiro. Model Development
and Validation of an Industrial UOP Fluid Catalytic Cracking Unit with a High-Efficiency Regenerator.
Industrial & Engineering Chemistry Research, 47(3):850–866, 2008.
[6] J. L. Fernandes, J. J. Verstraete, C. I. Pinheiro, N. M. Oliveira, and F. R. Ribeiro. Dynamic modelling of
an industrial R2R FCC unit. Chemical Engineering Science, 62(4):1184 – 1198, 2007.

[7] J. Gao, C. Xu, S. Lin, G. Yang, and Y. Guo. Advanced model for turbulent gas–solid flow and reaction in
FCC riser reactors. AIChE Journal, 45(5):1095–1113, 1999.
[8] I. Han and C. Chung. Dynamic modeling and simulation of a fluidized catalytic cracking process. Part I:
Process modeling. Chemical Engineering Science, 56(5):1951–1971, 2001.

[9] I. Han and C. Chung. Dynamic modeling and simulation of a fluidized catalytic cracking process. Part II:
Property estimation and simulation. Chemical Engineering Science, 56(5):1973–1990, 2001.
[10] HowStuffWorks. Oil refining. http://science.howstuffworks.com/oil-refining.htm.
[11] M. F. Isabelle Pitault, David Nevicato and J.-R. Bernard. Kinetic Model based on a Molecular Description
for Catalytic Cracking of Vacuum Gas Oil. Chemical Engineering Science, 49:4249–4262, 1994.
[12] S. Jacob, B. Gross, S. Voltz, and V. Weekman. A lumping and reaction scheme for catalytic cracking.
AIChE Journal, 22(4):701–713, 1976.
[13] S. Kameswaram and L. T. Biegler. Convergence rates for direct transcription of optimal control problems
using collocation at Radau points. Computational Optimization and its Applications, 41:81–126, 2008.

[14] L. L.S., C. Y.W., and H. T.N. Four-lumped kinetic model for fluid catalytic cracking process. Canadian
Journal of Chemical Engineering, 67, 1989.
[15] M. Martin, C. Derouin, P. Turlier, M. Forissier, G. Wild, and J. Bernard. Catalytic cracking in riser
reactors: core-annulus and elbow effects. Chemical engineering science, 47(9-11):2319–2324, 1992.

[16] J. Nocedal and S. Wright. Numerical optimization. Springer, 2000.


[17] I. Pitault, D. Nevicato, M. Forissier, and J. Bernard. Kinetic model based on a molecular description for
catalytic cracking of vacuum gas oil. Chemical Engineering Science, 49:4249–4249, 1994.
[18] H. Shan, W. Zhao, C. He, J. Zhang, and C. Yang. Maximum FCC diesel yield with TSRFCC technology.
Prepr. Pap.-Am. Chem. Soc., Div. Fuel Chem, 48(2):710, 2003.
[19] T. T., S. S., M. Y., and H. H. A reaction model for fluidized-bed catalytic cracking of residual oil.
International Chemical Engineering, 27:107–115, 1987.
[20] K. Theologos and N. Markatos. Advanced modeling of fluid catalytic cracking riser-type reactors. AIChE
Journal, 39(6):1007–1017, 1993.

[21] United States Department of Labor. Petroleum refining process. http://www.osha.gov/dts/osta/otm/


otm_iv/otm_iv_2.html.
[22] R. D. R. Uri M. Ascher, Robert M. M. Mattheij. Numerical Solution of Boundary Value Problems for
Ordinary Differential Equations. SIAM, 1995.

[23] F. Van Landeghem, D. Nevicato, I. Pitault, M. Forissier, P. Turlier, C. Derouin, and J. Bernard. Fluid
catalytic cracking: modelling of an industrial riser. Applied Catalysis A, General, 138(2):381–405, 1996.
[24] A. Wächter and L. T. Biegler. On the Implementation of a Primal-Dual Interior Point Filter Line Search
Algorithm for Large-Scale Nonlinear Programming. Mathematical Programming, 106(1):25–57, 2006.

[25] V. Weekman Jr. Model of Catalytic Cracking Conversion in Fixed, Moving, and Fluid-Bed Reactors.
Industrial & Engineering Chemistry Process Design and Development, 7(1):90–95, 1968.
[26] Wikipedia. Fluid catalytic cracking. http://en.wikipedia.org/wiki/Fluid_catalytic_cracking.
[27] Wikipedia. Oil refinery. http://en.wikipedia.org/wiki/Oil_refinery.

[28] S. Wright. Primal-dual interior-point methods. Society for Industrial Mathematics, 1997.

18

You might also like