Computers and Chemical Engineering 40 (2012) 117–131

Contents lists available at SciVerse ScienceDirect

Computers and Chemical Engineering

journal homepage: www.elsevier.com/locate/compchemeng

Process dynamics, control and monitoring

Modeling, simulation and advanced control of methanol production from

variable synthesis gas feed
Sidharth Abrol a,b,∗ , Courtland M. Hilton a,c
a
Sundrop Fuels Inc., 1722 Boxelder Street, Louisville, CO 80027, United States
b
Department of Chemical Engineering, University of Delaware, Newark, DE 19716, United States
c
Hilton Consultancy, 4795 West 127th Street, Broomfield, CO 80020, United States

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

Article history: A process relying on feedstock produced using a non-continuous energy source such as solar energy needs
Received 30 January 2011 detailed understanding of the dynamics to define procedures for cycling and robust control design. In
Received in revised form 6 January 2012 this work, methanol production from synthesis gas (or syngas) is studied. The syngas is (assumed to be)
Accepted 7 February 2012
obtained from an intermittent upstream biomass-gasification process run by solar energy. Development
Available online 28 February 2012
of the methanol synthesis recycle-loop model is described in detail, along with several case studies
performed using the steady-state and dynamic models for better understanding of the process behavior.
Keywords:
A linear model predictive controller is designed to reject disturbances from a varying upstream syngas
Modeling
Kinetics
production process and track set point changes in the desired variables for methanol synthesis.
Sequential-modular © 2012 Elsevier Ltd. All rights reserved.
Non-continuous
System identification
Model predictive control

1. Introduction and designing a linear feedback controller for a solar-driven reactor

with the aim of maintaining uniform product composition during
Methanol synthesis is an established process technology in the such transients. Research effort has also been made to use advanced
chemical industry for many years now. The feedstock to the process or adaptive controllers employing model predictive techniques to
consists of synthesis gas or syngas (a blend of CO2 , CO and H2 ) which manipulate mirror fields or vary feed flow rates to adapt to the
can be derived from various sources such as multi-step reform- changing solar irradiation (Camacho, Berenguel, & Rubio, 1997;
ing of natural gas or higher hydrocarbons, coal gasification as part Silva, Lemos, & Rato, 2003).
of the integrated gasification combined cycle (IGCC) technology, Nonetheless, any control strategy for the solar biomass gasifica-
or biomass gasification. Biomass can be gasified using energy pro- tion process will not be able to maintain a constant flow of syngas
duced from burning a conventional fuel like coal or from alternate even if the composition (of syngas) is controlled at the desired
renewable sources like solar energy to meet the ever increasing values. To make the downstream processes (which can include a
energy demand world over. process like methanol synthesis) continuous, the variable rate of
Solar energy with its diurnal variation from high to zero, and syngas production can be mitigated by having a syngas storage or
even during peak periods, disturbances due to cloud cover, winds a buffer. But having such a storage large enough to hold the syngas
or other weather events, makes the synthesis gas production highly can be expensive, both in terms of capital costs and operating costs
variable. In the last decade, a lot of work has been done on devel- for compression (Apt, Newcomer, Lave, Douglas, & Dunn, 2008).
oping robust satellite-based models to predict such events. Most Also, if underground storage is used, the appropriate storage loca-
of these models use data provided by geostationary satellites and tion may not coincide with the location best suited for maximum
involve semi-empirical calculations based on statistical methods solar irradiance.
to predict such events (Mueller et al., 2004). Recently, Petrasch, In order to understand the dynamics of the downstream
Osch, and Steinfeld (2009) looked at developing a dynamic model methanol synthesis process with the possibility of operating at dif-
ferent set points or targets for varying syngas feed, and for robust
control design, development of a dynamic methanol loop model is
∗ Corresponding author at: Sundrop Fuels Inc., 1722 Boxelder Street, Louisville, required. A few authors have previously worked on developing a
CO 80027, United States. mathematical model for the methanol reactor (Rezaie, Jahanmiri,
E-mail address: abrolsidharth@gmail.com (S. Abrol). Moghtaderi, & Rahimpour, 2005; Shahrokhi & Baghmisheh, 2005),

0098-1354/$ – see front matter © 2012 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compchemeng.2012.02.005
118 S. Abrol, C.M. Hilton / Computers and Chemical Engineering 40 (2012) 117–131

Nomenclature
[H2 O]liquid,flash mole fraction of water in crude methanol
a activity of catalyst H298 standard heat of formation at 298 K
a0 quadratic mixing term Hf,i heat of formation of component i at operating tem-
a1 asymmetric polar term perature T
A cross-sectional area of reactor tube  catalyst effectiveness factor
Ci molar concentration of component i cat density of catalyst bed
Cm total mass concentration  tortuosity of catalyst
Cp specific heat at constant pressure
Ct total molar concentration
Di internal diameter of reactor tube
and for the methanol loop (Parvasi, Rahimpour, & Jahanmiri, 2008).
Deff,i effective diffusivity of component i
But most of the previous work focuses on (and are therefore devel-
Dk,i Knudsen diffusion coefficient of component i
oped for) studying performance of a continuous process with long
ḟt total mass flow rate per tube term catalyst deactivation in the reactor. The kinetic model used
F total molar flow rate in these models are based on the experimental work of Graaf,
Fbypass molar flow rate of bypass stream Scholtens, Stamhuis, and Beenackers (1990) which more recently
Fin,reactor molar flow rate of gas going into the reactor has been replaced by a model from Bussche and Froment (1996),
Fl total molar liquid flow rate from flash which addresses some of the shortcomings of the previous kinetic
FMUG molar flow rate of make-up gas models.
Fpurge molar flow rate of purge stream In our work, we describe in detail the development of a math-
Fv total molar vapor flow rate from flash ematical (dynamic) model of the methanol reactor-separator loop,
ki kinetic factor for reaction i followed by variable-space simulation runs using the steady-state
kij binary interaction parameter for pair i and j loop model to assess different operating regimes for the process.
ki pseudo rate constant for reaction i Simulation runs from the dynamic model are used to analyze
K1 , K2 , K3 adsorption equilibrium constants loop performance for variable flow of syngas, and for an isolated
Ki k-value for component i methanol loop or reactor during downtime. The data from the
Keq i equilibrium constants for reaction i dynamic model is then used to fit a linear model to be used in
i
Keq pseudo equilibrium constant for reaction i a linear model predictive controller (MPC) for regulatory control.
li composition of component i in liquid from flash In the end, a few suggestions on designing a robust distributed
ṁt total mass flow rate per tube controller for methanol synthesis from solar-derived syngas are
MWi molecular weight of component i discussed.
N number of components
p pressure 2. Methanol synthesis
pi partial pressure of component i
Pc critical pressure Low-pressure methanol synthesis processes with a typical oper-
ri rate of reaction for component i ating pressure between 50 and 100 bar became popular in the 1960s
rp radius of catalyst pellet replacing the high-pressure technology used earlier over a Cr-based
R universal gas law constant catalyst. Low-pressure gas-phase methanol synthesis can be car-
t time ried out using a variety of reactor styles including a boiling-water
T0 initial temperature at t = 0 shell with catalyst-packed tubes which was first introduced by
T temperature Lurgi (Meyer, 1984).
Tc critical temperature A typical boiling-water methanol reactor is designed as a shell
Tout,reactor temperature of reactor effluent and tube heat exchanger, in which the catalyst is packed in vertical
Tss temperature at steady-state tubes surrounded by boiling water in the shell. The temperature of
Tshell temperature of boiling water on the shell side the reactor is controlled by the pressure of the boiling water. Typical
Ushell overall heat transfer coefficient from reactor tube- catalyst used for the low-pressure process is CuO/ZnO/Al2 O3 . Con-
side to shell-side version to methanol is limited by equilibrium of the reactions taking
vi composition of component i in vapor from flash place. The reactor effluent is cooled to a temperature low enough
wi acentric factor to separate out liquid crude methanol product in a high-pressure
xi mass fraction of component i separator, and the vapors from the separator are compressed to
yi mole fraction of component i the reactor pressure and recycled back. A small purge is allowed
yiss mass fraction of component i at steady-state to keep the inerts from building up in the loop. Fig. 1 shows the
Y first-principles model output process flow diagram for the synthesis loop.
Ŷ transfer function model output The important reactions taking place in a typical methanol reac-
Ȳ mean of first-principles model output tor are the hydrogenation of CO and CO2 and the reverse water-gas
z axial direction shift reaction.
ˇ equilibrium separation in flash
εcat void fraction of catalyst bed CO + 2H2 ↔ CH3 OH, H 298 = −90.55 kJ mol−1 (1)
εs
 porosity of catalyst
−1
vi diffusion volume of component i CO2 + 3H2 ↔ CH3 OH + H2 O, H 298 = −49.33 kJ mol (2)
l,i fugacity coefficient of component i in liquid
CO2 + H2 ↔ CO + H2 O, H 298 = 41.12 kJ mol−1 (3)
M,i Thiele modulus for reaction i
v,i fugacity coefficient of component i in vapor The three reactions are not independent, and any one of them
can be expressed as a linear combination of the other two. One of
the design specifications on the make-up gas (MUG) or fresh syngas
 S. Abrol, C.M. Hilton / Computers and Chemical Engineering 40 (2012) 117–131 119

Fig. 1. Methanol synthesis loop.

feed is the stoichiometric number (SN) which is given by Eq. (4) can be calculated using the method described in the section that
(Lovik, 2001). follows.
∂xi
 ṁt
 ∂x i ri cat a
[H2 ] − [CO2 ] =− i
+ MWi (5)
SN = (4) ∂t Aεcat Cm ∂z εcat Cm
[CO] + [CO2 ]
 
where [·] is the molar concentration of the components of syn- ∂T ḟt ∂T
Di
=− + U (T − T)
gas. For a SN of 2 the reactants are present in exact stoichiometric ∂t Aεcat Ct ∂z Aεcat Ct Cp shell shell
amounts, but for a typical industrial methanol synthesis process
cat a 
the SN is anywhere between 2.01 and 2.1, implying a little excess N

of H2 . For a reforming process like steam-methane reforming (SMR) + i ri (−Hf,i ) (6)

εcat Ct Cp
or auto-thermal reforming (ATR), the composition of syngas can be i=1
controlled and adjusted to meet the specification on the SN (Tindall
where subscript i is for each of the N components CH3 OH, CO2 , CO,
& Crews, 1995). But that may not be always true for a gasifica-
H2 O, H2 , and inerts CH4 and N2 , Ct and Cm are the total molar and
tion process, in which a water-gas shift reactor and/or an acid gas
mass concentrations, which vary with pressure and temperature
removal system may have to be included downstream to drive the
inside the reactor, and ḟt and ṁt are the total molar and mass flow
SN to the desired value.
rates per tube. The initial and boundary conditions for the mass
fractions, xi and temperature, T of the reactor are as follows:
3. Reactor-separator loop model
t = 0, xi = xiss , T = T ss (7)
Development of a mathematically rigorous dynamic model and
for any process is necessary for a wide range of applications
including control design, operability studies, optimization, start- z = 0, xi = xi0 , T = T 0 (8)
up/shut-down procedures or fault diagnosis and monitoring. A
Both the steady-state composition, xissand temperature profiles,
homogenous (gas-phase) dynamic model of the reactor is formu-
Tss along the reactor length, z are determined by solving the steady-
lated here based on mass and energy balances. For the separator
state problem with differentials with respect to time in Eqs. (5) and
a pressure-temperature flash based on the Soave–Redlich–Kwong
(6) set equal to zero. Expressions for variation of heat of forma-
(SRK) equation of state is used. The recycle stream is torn, and
tion with temperature, and that of specific heat with composition
Wegstein algorithm is used for faster convergence of the loop.
and temperature have been taken from Smith, Van Ness, and Abott
The following sections deal with each of these steps in detail. The
(2003) and are given in Appendix A. While solving Eqs. (5) and (6) or
heat exchangers and recycle compressor are not modeled and it is
the derived steady-state equations, it is important to keep a check
assumed that the output temperature and pressure for these units
on the mass balance. Because of the stoichiometry of the reactions,
are at the desired values.
the effluent from the reactor shows a decrease in the total number
of moles, and use of mass fractions, xi instead of mole fractions is
3.1. Reactor model one way to resolve this issue.

As mentioned earlier, the reactions taking place inside the reac- 3.1.1. Kinetics
tor are given by Eqs. (1)–(3), and circulating boiling water in the Various mechanisms have been studied in the past to define
shell side is used to remove the exothermic heat of reaction. Mass the rate expressions of the reactions in methanol synthesis. Lovik
and energy balances across the catalytic reactor can be defined (2001) and Bussche and Froment (1996) have given an overview
by a set of partial differential equations (PDEs) given by Eqs. (5) of the various researched mechanisms and their limitations. Two
and (6). In developing these equations the diffusion limitations widely used rate expressions are based on the experimental work
in the catalyst are included as an effectiveness factor, i , which of Graaf et al. (1990) and Bussche and Froment (1996). The kinetics
120 S. Abrol, C.M. Hilton / Computers and Chemical Engineering 40 (2012) 117–131

Table 1 gas present at a particular length inside the reactor. The original
Constants for Bussche and Froment (1996) kinetic model.
work by Graaf et al. (1990) used a dusty-gas model defined for
k = Aexp(B/RT) A B a spherical pellet to estimate the effectiveness factor. The model
kMeOH 1.07 36,696 described can be reduced to a set of ordinary differential equations
kRWGS 1.22 × 1010 −94,765 (ODEs) which define a boundary value problem for the pellet.
K1 3453.38 – Lommerts, Graaf, and Beenackers (2000) compared the per-
K2 0.499 17,197 formance of five different diffusion models (with different
K3 6.62 × 10−11 124,119
computation times) to determine the effectiveness factor. The most
rigorous method was based on the dusty-gas model, and the least
Table 2 computationally intensive one involved use of the Thiele modu-
Reaction equilibrium constants. lus. It was concluded in their work that a simple approach like that
Keq = 10(A/T−B) A B involving the Thiele Modulus can be used to reduce the compu-
1
tation effort without compromising the accuracy of estimation. In
Keq 3066 10.592
2
Keq −2073 −2.029
this work, a similar approach is adopted based on the Thiele modu-
lus. In the work by Lommerts et al. (2000) rate equations based on
Graaf et al. (1990) kinetics are used and equilibrium composition of
as described by Bussche and Froment (1996) are based on reactions the components are required to calculate the pseudo-equilibrium
(2) and (3) only as compared to that of Graaf et al. (1990) which is constants and pseudo-rate constants. We extend that derivation to
based on all the three reactions. However, Graaf et al. (1990) did kinetics from Bussche and Froment (1996):
not account for the fact that some intermediates like the formyl pCH3 OH pH2 O
1
and methoxy species are predicted simultaneously in their model Keq = K1 (11)
pCO2 p3H
and feature in two different reactions at the same time. The kinetic 2

model of Bussche and Froment (1996) is formulated to overcome 2

pCO pH2 O
Keq = K2 (12)
this limitation and has been shown to accurately predict the exper- pCO2 pH2
imentally obtained conversions even for extrapolation outside the
3 and K 2 =  
where K1 = CH3 OH H2 O /CO2 H
originally used experimental window, and will be used in our work  CO H2 O /CO2 H2 are
2
reported here. the correction factors based on fugacities of the components calcu-
The kinetics are based on the Langmuir–Hinshelwood– lated using the SRK equation of state.
Hougen–Watson (LHHW) mechanism. The reverse water-gas shift The LHHW kinetics (Eqs. (9) and (10)) for methanol synthesis
reaction occurs through a redox mechanism and methanol is were then reduced to linear rates by defining pseudo constants as
formed through the hydrogenation reaction via formate. The exper- given below:
imental conditions for development of the original kinetic model  

pCH3 OH
were as follows: rMeOH = k1 pH2 − 1  (13)
Keq
Temperature: 180–280 ◦ C;  

pH2 O
Pressure: 15–51 bar; rRWGS = k2 pH2 − 2  (14)
Catalyst: ICI 51–2 Cu/ZnO/Al2 O3 . Keq

where
The rate expressions for the two rate-determining steps are 1 p2 p
kMeOH pCO2  Keq H CO2
given below and the values of the rate constants and the equilib- k1 = 1
= 2

rium constants are given in Tables 1 and 2.

 3
, Keq
pH2 O
(1 + K1 (pH2 O /pH2 ) + K2 pH2 + K3 pH2 O )
1 )(p
kMeOH pCO2 pH2 [1 − (1/Keq 3
H2 O pCH3 OH /pH pCO2 )]
rMeOH = 
2
(9) 2 p
3 kRWGS (pCO2 /pH2 )  Keq CO2
(1 + K1 (pH2 O /pH2 ) + K2 pH2 + K3 pH2 O ) k2 =  2
and Keq =
(1 + K1 (pH2 O /pH2 ) + K2 pH2 + K3 pH2 O ) pCO
2 (p
kRWGS pCO2 [1 − Keq H2 O pCO /pCO2 pH2 )]
rRWGS =  (10) The Thiele modulus for first order linear kinetic expressions in
(1 + K1 (pH2 O /pH2 ) + K2 pH2 + K3 pH2 O )
cylindrical pellet is given by (Levenspiel, 1993):


3.1.2. Intra-particle diffusion limitations rp ki (Keq
i + 1)

It was shown by Graaf et al. (1990) that pore diffusion effects are M,i =  (15)
2 i
Deff,H2 Keq
significant in low-pressure methanol synthesis using Cu/ZnO/Al2 O3
catalyst. On the other hand, in their experiments to develop a where the effective diffusivity, Deff,j is defined for H2 as the prin-
kinetic model, Bussche and Froment (1996) did not observe any cipal component in the gas-mixture. The effective diffusivity is
internal or external diffusional limitations. The effect of diffusion calculated by combining the effect of both Knudsen and molecular
can be incorporated in the form of an effectiveness factor which diffusion.
is the ratio of observed reaction rate to the theoretical reaction ⎛ ⎞
rate. The observed reaction rate can be increasingly less than the
⎜ ⎟
ideal rate if the reaction becomes more diffusion limited. Graaf ⎜ ⎟
et al. (1990) mention that at temperatures >227 ◦ C the activation 1  ⎜ 1 N
y ⎟
= ⎜ i ⎟
energy of the methanol synthesis reaction falls by 50% owing to Deff,H2 εs ⎜ Dk,H + D ⎟ (16)
⎜ 2 i,H2 ⎟
intra-particle diffusion. Also, as the catalyst deactivates the reac- ⎝ i=1 ⎠
tion becomes more diffusion limited as the activation energy of the i=
/ H2
reaction increases.
The effectiveness factor in the reactor varies with composition where yi is the mole fraction of component i, and  is the tortu-
and temperature (assuming constant pressure in the pellet) of the osity and εs is the porosity of the catalyst. The Knudsen diffusion
 S. Abrol, C.M. Hilton / Computers and Chemical Engineering 40 (2012) 117–131 121

Table 3 Due to the need for simplicity and accuracy in predicting k-

Diffusion volumes of components.
values for a wide range of temperature, use of an equation of state

Component i (EOS) was considered. A pressure–temperature (P–T) non-ideal
flash was modeled using the Soave–Redlich–Kwong (SRK) equation
CH3 OH 29.9
CO2 26.9 of state in this work. The form of the EOS is (Soave, 1972):
CO 18.9
RT a
H2 O 12.7 p= − (20)
H2 7.07 Vm + c − b (Vm + c)(Vm + c + b)
CH4 24.42
N2 17.9
where a = a0 + a1 and the quadratic mixing term, a0 and the asym-
metric polar term, a1 are given by:

coefficient for a cylindrical pellet is defined as (Reid, Prausnitz, & 

N 
N

a0 = yi yj ai aj (1 − kij ) (21)
Poling, 1997):
 i=1 j=1

Dk,H2 =
2
rp
8RT
(17) ⎛ ⎞3
3
MH2 
N

N
1/3
a1 = yi ⎝ yj ((ai aj ) 1/2
lji ) ⎠ (22)
where rp is the radius of the pellet and MH2 is the molecular weight
i=1 j=1
of diffusing component, H2 . The binary diffusion coefficients can
be estimated by the Fuller, Schettler and Giddings’ (FSG) equation where
(Fuller, Schettler & Giddings, 1996):

N 
N
−4 −1 1/2 b= yi bi , c = yi ci , ai = f (T, Tci , Pci , wi ),
9.9794 × 10 T 1.75 [Mi−1 + MH ]
Di,H2 =  
2
(18)
 2 i=1 i=1
P
1/3
i +
1/3
H  RT 
2 ci
bi = f (T, Tci , Pci ) and ci = 0.40768 (0.29441 − zi ).
 Pci
The atomic diffusion volume increments, i , are taken from
Fuller et al. (1996) and are used to calculate the diffusion volume The values of critical temperatures, pressures and the acentric fac-
of the components. The calculated diffusion volumes are given in tors from Reid et al. (1997) are given in Table 4.
Table 3. For a P–T flash the mass balance for each component is given by:
Using the estimated Thiele modulus, the effectiveness factor
Fyi = Fl li + Fv vi (23)
for a cylindrical pellet can be calculated from the modified Bessel
function (Levenspiel, 1993):  N

li = 1 (24)
1 I1 (2M,i )
i = (19) i=1
M,i I0 (2M,i )

The pore diffusion effect included as the effectiveness factor for the 
N

reactions takes into account any diffusional limitations that might

vi = 1 (25)
be observed in low pressure methanol synthesis, and is included in i=1

the reactor model defined by Eqs. (5) and (6). li l,i = vi v,i (26)

3.2. Separator model where yi is the mole fractions of component i in the feed to the
flash and l and v denote the liquid and vapor output streams from
The flash separator in Fig. 1 has to be modeled to predict the non- the flash. The fugacity coefficients, , are calculated using the SRK
ideal behavior of polar components in the system. Chang, Rousseau, EOS. The k-value is defined as the ratio:
and Kiipatrick (1986) mentioned that components CO, H2 and CO2 vi
Ki = (27)
exhibit nearly ideal behavior because the reaction temperature li
(200–400 ◦ C) is much higher than the critical temperatures of these
and the equilibrium separation, ˇ which is the fraction of vapors is
components, whereas, CH3 OH and H2 O vapors are highly non-ideal
given by the Rachford–Rice equation:
at these temperatures because of their polar molecular structure.
To estimate the k-values (as ratio of saturated vapor pressure 
N
yi (Ki − 1)
and total pressure) and subsequently the ideal vapor-liquid equilib- =0 (28)
1 + ˇ(Ki − 1)
rium composition, vapor pressures have to be accurately estimated i=1
at the given temperature. Use of an empirical relation like the
Antoine equation or the Wagner equation (which is known to be 3.3. Loop convergence
applicable for a wide temperature range) to calculate the saturated
vapor pressure of the components in the mixture does not work The solution of the flash mass balance, along with SRK EOS gives
for methanol synthesis because of the reason mentioned before, the flow rate and mole fractions of components leaving the flash in
resulting from non-ideal behavior of CH3 OH and H2 O. the liquid and vapor streams. As shown in Fig. 1, the liquid stream

Table 4
Thermodynamic parameters for the components.

Parameter CH3 OH CO2 CO H2 O H2 CH4 N2

Tc (K) 512.6 304.1 132.9 647.3 33 190.4 126.2

Pc (bar) 80.9 73.8 35 221.2 12.9 46 33.9
wi 0.556 0.239 0.066 0.344 −0.216 0.011 0.039
122 S. Abrol, C.M. Hilton / Computers and Chemical Engineering 40 (2012) 117–131

from the flash is the crude methanol product, and the vapors are Table 5
List of process parameters.
recycled back to the reactor (after allowing a small purge).
The loop is torn on the recycle stream and Wegstein algo- Parameter Value Units
rithm is used to update the value of the recycle stream variables Length of reactor 7 m
instead of successive substitution for faster simulation convergence Number of tubes 6560 –
(Wegstein, 1958). Details of the algorithm are given in Appendix Heat transfer coefficient 631 W/(m2 K)
B. The iterations were stopped when either maximum number of Catalyst bulk density 1100 kg/m3
Catalyst activity 1 –
allowed iterations were exceeded or the tolerance criterion for all
Internal diameter of tubes 1.3 in.
the recycle stream variables (recycle: total flow rate and mole frac- Void fraction 0.38
tions of all components) was met. This criterion is defined as: ◦
Shell temperature 238 C
 recycle − recycle
 Flash Pressure 730 psi
cal ◦
abs ≤ εtol (29) Flash Temperature 65 C
recycle Ratio of purge to vapor flow from flash 0.01

4. Solving the model

Hertzberg, 1986). Each module has its own integrator which solves
4.1. Reactor dynamic model a set of PDEs/ODEs/AEs for that unit. In the past, researchers have
proposed different algorithms for independent modular integra-
Eqs. (5) and (6) describing the open-loop dynamics of the reactor tion (Fagley & Carnahan, 1990; Liu & Brosilow, 1983), which may
are hyperbolic PDEs and are solved by discretizing them on the or may not exploit the flowsheet structure.
nodes of a one-dimensional mesh in the axial z-direction. Backward The advantage of simultaneous equation-based approach is that
finite difference approximation was used to reduce the equations by integrating the equation set as a whole, no inaccuracies are intro-
to a set of ODEs which were then solved using a stiff solver, ODE15s duced by the recycle streams. A sequential-modular approach on
in MATLAB. the other hand is more intuitive given the flowsheet structure and
Eqs. (5) and (6) can be represented with discretized terms along material flow in the process, and may require fewer calculations
the z-direction as follows: per time step. However, the time steps taken at every iteration
   needs to be small enough to generate accurate results depending on
∂xik ṁkt xik − xik−1 ki rik cat a
=− + MWi (30) the slope of the variable trajectory generated (Fagley & Carnahan,
∂t k
Aεcat Cm z k
εcat Cm 1990). Whereas, an equation-oriented approach is more rigorous
   and gives accurate solution even for large time steps. Variable step
∂T k ḟtk T k − T k−1
Di integrators have been suggested for sequential-modular techniques
=− + Ushell (Tshell − T k ) which try to automatically control the error resulting from a large
∂t Aεcat Ctk z Aεcat Ctk Cpk
time step (Kulikov et al., 2005).
cat a 
N The simulator programmed here integrates the equations (for a
+ ki rik (−Hf,i
k
) (31) small time step at every iteration) defining the reactor and the flash
εcat Ctk Cpk system separately in a sequential manner, that is, the output from
i=1
solving the reactor PDEs are the inputs to the flash for integration
The set of ODEs are solved for each node in the reactor starting from
over any particular time interval. Also, at each time step the loop is
the first node and solving the equations using the inlet conditions
converged by tearing on the recycle stream and using the Wegstein
to the reactor and the solution of the steady-state equations solved
algorithm. Fig. 2 shows the flowchart for sequentially solving the
separately. The procedure is repeated for the next node, k using the
dynamic flowsheet. A time step of 1 s was used for the runs. Smaller
solution of the previous node, k − 1 as the inlet condition and the
time intervals of 0.1 and 0.01 s were also tested and about the same
steady-state results as initial conditions. For each time iteration,
accuracy was observed for the response generated.
the procedure is repeated until the equations for the last node in
the reactor have been solved.
5. Steady-state analysis
4.2. Loop dynamic model
To solve the steady-state loop model the differential terms with
To run a dynamic simulation on the loop with a recycle respect to time in Eqs. (5) and (6) were set to zero. Sensitivity stud-
tear stream, different approaches can be used. An equation-based ies were made to see the effect of various process variables on total
approach solves all the modules simultaneously by a single integra- conversion of carbon (Eq. (33)) in the loop.
tion routine (Fagley & Carnahan, 1990). The reactor-flash process molar flowrate of crude MeOH
with recycle studied here can be represented as a system of differ- % carbon-conversion = ×100
molar flowrate of(CO2 + CO) in MUG
ential algebraic equations (DAEs), after the reactor PDEs have been
(33)
discretized on the reactor mesh. Solving a stiff system of DAEs such
as these can be an issue in MATLAB, which can solve only DAEs of The process parameters used for the studies are listed in
‘index 1’. The index refers to the sum of the ‘mass’ matrix (M) in the Tables 5 and 6.
DAE (Eq. (32)) and the Jacobian (multiplied by a constant) (MATLAB For an initial steady-state run using the process parameters
users’ guide, 2009). listed in Tables 5 and 6, the total carbon-conversion obtained was
96.76%. Because the purge is fixed at 1% of the flash vapors, the
M(t, y)y = F(t, y) (32)
recycle ratio is allowed to float and comes to ∼4.5. Fig. 3 shows
Another approach called sequential-modular solves independent the composition and temperature profiles inside the reactor for the
modules representing a unit or a ‘cluster’ of units sequentially. steady-state run.
In the sequential-modular approach modules are integrated over The temperature of the boiling water which is used as a coolant
a common time interval in a sequential manner as dictated by the for the reactor is controlled by changing the pressure of the steam
material flow in the process. A coordination algorithm maintains drum supplying the boiling water. The coolant carries away the
integration over the same time step at every iteration (Hillestad & exothermic heat of reaction for methanol synthesis and ensures
 S. Abrol, C.M. Hilton / Computers and Chemical Engineering 40 (2012) 117–131 123

Initial guess for recycle

stream variables
t=0

Solve steady-state
loop (obtain reactor
initial conditions)

t = t+dt

Save flowsheet solution

Integrate reactor and solve

flash for the current time
interval

No

No Check for Yes

recycle Is t>=tmax
convergence
Yes

Update recycle stream End

Fig. 2. Modular dynamic simulation algorithm with recycle convergence.

0.4 0.12
[CH3OH]

[CO2]

0.2 0.1
0 0.08
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
Reactor length (m) Reactor length (m)
0.4 0.02
[H2O]
[CO]

0.2 0.01

0 0
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
Reactor length (m) Reactor length (m)
0.25 260
T (C)
[H2]

0.2 240
220
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
Reactor length (m) Reactor length (m)

Fig. 3. Steady-state composition and temperature profiles in methanol reactor. [i] is the mole fraction of component i in the gas mixture.

that temperature inside the reactor never exceeds the maximum 97.2
temperature at which catalyst sintering occurs. Fig. 4 shows the
variation in total carbon-conversion with changing boiling water 97

temperature in the shell side.

% carbon-conversion

96.8
It is observed that for the given feed conditions (assuming fresh
96.6
catalyst), maximum conversion is obtained at a coolant tempera-
ture of 230 ◦ C. Lovik (2001) has cited experimental work by various 96.4
authors on catalyst sintering at elevated temperatures. At typi-
96.2
cal temperatures of methanol synthesis (200–240 ◦ C), Cu which
is an active element in the catalyst shows considerable sinter- 96
ing and at higher temperatures (>280 ◦ C), sintering of ZnO sites is
also observed. Sintering occurs by migration of atoms/crystals to 220 225 230 235 240 245 250 255 260
Tshell (C)
form large agglomerates, which decreases the active surface area
for reaction. Typically during the initial operation of the plant a Fig. 4. Total conversion of carbon with varying coolant temperature in the shell
lower temperature for reaction is used, and as the catalyst ages and side.
124 S. Abrol, C.M. Hilton / Computers and Chemical Engineering 40 (2012) 117–131

Table 6
MUG variables in the process.

Parameters Value Units

Total flow rate 4685 lbmol/h

Pressure 750 psi
◦
Temperature 250 C
Composition
[CH3 OH] 0 mol. fr.
[CO2 ] 0.032 mol. fr.
[CO] 0.285 mol. fr.
[H2 O] 0 mol. fr.
[H2 ] 0.675 mol. fr.
[CH4 ] 0.006 mol. fr.
[N2 ] 0.003 mol. fr.
Fig. 7. Contour plot with % carbon-conversion with varying MUG temperature and
pressure.
100

95
arises for reducing compression cost or the heat exchanger duty
% carbon-conversion

90
on the MUG, a lower temperature and pressure can be used with a
85
modest reduction (∼1.0–1.5%) in carbon-conversion.
80 The next study involved the flash separator in the synthesis loop.
75 The pressure (Pflash) and temperature (Tflash) of the flash unit were
changed to see their effect on separation of the reactor effluent into
70
liquid and vapor streams at equilibrium. Fig. 8 shows the isolines
65
for carbon-conversion with changing Pflash and Tflash. The temper-
60 ature of the stream entering the flash maybe varied by changing the
0.2 0.25 0.3 0.35 0.4 0.45
[CO] in MUG (mol. fr.) flow of cooling water stream in the heat exchanger (Fig. 1) used to
further cool down the reactor effluent. The pressure may also be
Fig. 5. Total carbon-conversion with varying CO in MUG. varied within reasonable limits by installing a control valve on the
inlet stream to the flash or the vapor exit stream from the flash.
deactivates, a higher temperature would be required to ensure One would like to operate the flash at a high pressure, so that
desired conversion rates. a smaller compressor is required in the recycle loop to compress
Another study was made by varying the composition of CO and the vapor effluents from the flash back to the reactor pressure.
CO2 in the make-up gas (MUG). Figs. 5 and 6 show the changing Operating at a higher temperature with a reduction penalty in
conversion versus concentration of these components in the feed carbon-conversion of ∼1% may be desirable depending on the heat
gas. duty requirements of the heat exchanger or the design and control
The plot for % carbon-conversion with varying CO2 concentra- strategy employed.
tion in MUG shows a similar trend as originally observed in the Another study was made by varying the amount of purge in the
experimental work by Klier, Chatikavanij, Herman, and Simmons recycle stream as a percent of the effluent vapors from the flash unit.
(1982). Skrzypek, Słoczyński, and Ledakowicz (1994) explained the A small amount of purge ensures removal of inerts (CH4 and N2 )
mechanism, with a maximum conversion caused by a reversible and prevent their build-up in the loop, which otherwise negatively
redox reaction resulting in the optimal oxidation state of Cu. A syn- impacts the conversion in the reactor. Table 7 shows the results
gas rich in CO2 also leads to higher deactivation rates of the catalyst. from runs made for different amount of purge (0.1–10% of flash
The deactivation is caused by the H2 O which is produced from CO2 vapors).
through the reverse water-gas shift reaction in the reactor. As expected, increasing the flow on the purge lowers the con-
If the pressure and temperature of the MUG are varied to see the version of carbon, and on the other hand a lower recycle ratio is
effect on carbon conversion, a plot with contours for % conversion observed as more ‘carbon’ is thrown away in the leaving purge
obtained is shown in Fig. 7. The pressure of the stream entering stream. More importantly, as the purge flow is increased the water
the flash unit in the loop was assumed to be 20 psi less than the in crude methanol is lowered. Depending on the design goals and
MUG pressure for these runs. As expected, a higher temperature the grade of crude methanol desired, excess water may have to
and pressure of the MUG favors more conversion. But if a need be removed in a downstream distillation column, which otherwise

100

95
% carbon-conversion

90

85

80

75

70
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
[CO2] in MUG (mol. fr.)
Fig. 8. Contour plot showing effect of flash pressure and temperature on % carbon-
Fig. 6. Total carbon-conversion with varying CO2 in MUG. conversion.
 S. Abrol, C.M. Hilton / Computers and Chemical Engineering 40 (2012) 117–131 125

Table 7
Simulated demonstration of purge effect on composition of crude methanol.

Purge flow %Carbon-conv. Recycle Crude flow rate MeOH in crude CO2 in crude CO in crude H2 O in crude H2 in crude CH4 in crude N2 in crude
(% of flash ratio (lbmol/h) (mol. fr.) (mol. fr.) (mol. fr.) (mol. fr.) (mol. fr.) (mol. fr.) (mol. fr.)
vapors)

0.1 98.71 19.79 1631.024 0.8980 0.0069 0.00097 0.0845 1.2E−06 0.008 0.0012
1 96.76 4.51 1595.394 0.8998 0.0227 0.00076 0.0691 1.1E−06 0.007 0.0008
2 95.13 3.02 1566.351 0.9011 0.0339 0.00094 0.0578 9.8E−07 0.006 0.0006
3 93.87 2.40 1543.783 0.9021 0.0418 0.00113 0.0494 9.1E−07 0.005 0.0005
4 92.82 2.03 1525.001 0.9030 0.0477 0.00131 0.0430 8.6E−07 0.005 0.0004
5 91.91 1.78 1508.708 0.9039 0.0522 0.00149 0.0379 8.2E−07 0.004 0.0004
6 91.11 1.59 1494.183 0.9047 0.0557 0.00167 0.0337 7.9E−07 0.004 0.0004
7 90.38 1.45 1480.952 0.9054 0.0585 0.00184 0.0302 7.7E−07 0.004 0.0003
8 89.70 1.33 1468.669 0.9061 0.0608 0.002 0.0273 7.5E−07 0.003 0.0003
9 89.06 1.23 1457.134 0.9068 0.0625 0.00217 0.0249 7.3E−07 0.003 0.0003
10 88.45 1.15 1446.148 0.9074 0.0640 0.00233 0.0228 7.2E−07 0.003 0.0003

could be avoided by allowing a higher purge, and utilizing the purge 96.8

as a fuel gas. Although a lower purge ensures higher conversion, it

also results in more inerts going in the liquid flash effluent, which 96.75

% carbon-conversion
may have to meet certain environmental regulations and limits. The
same is true for higher concentration of inerts for a larger purge, 96.7

and more dissolved CO2 in crude methanol.

A higher recycle ratio is also undesirable as it increases the oper- 96.65

ating cost of the recycle compressor, the size of the reactor, and the
overall catalyst cost. From a control’s perspective, the amount of 96.6

purge can be manipulated to control the recycle ratio and/or the

composition of the liquid stream leaving the flash. Control of inerts 0 10 20 30 40 50 60 70 80 90 100
in the loop is also necessary to prevent undesired effect of pro- Time (sec)

cess disturbances like varying MUG composition from affecting the

Fig. 9. Dynamics at peak flow without any step change.
reactor yield, which otherwise has been known to have a ‘snowball
effect’ in recycle loops (Luyben, 1994).
it from half to maximum were simulated. Figs. 11 and 12 show the
6. Dynamic simulation dynamic responses for these step changes.
Looking at the step responses in Figs. 11 and 12, the dynamics
Next, case studies were performed using the dynamic model for look like that of a second order process with a ‘zero’ in the numera-
the loop to analyze process behavior before doing the controller tor. Numerator dynamics might also arise due to the recycle of the
design. Such analyses are particularly relevant in solar syngas gen- product stream in the process, which adds a ‘zero’ to the overall
eration, where the flow rate sometimes cannot be guaranteed to process transfer function even if the units involved in the process
be constant. Fig. 9 shows the open-loop dynamics when the sys- have first-order dynamics (Seborg, Edgar, & Mellichamp, 2004).
tem starts at steady-state at t = 0 and is allowed to run (without
any disturbances) to a new steady-state in about 45 s. Also, Fig. 10 6.1. Isolated methanol synthesis process
shows the dynamic response of the composition and temperature
of the gas mixture at the reactor exit with time for the same run. For a process driven by solar energy there is a diurnal down-
A step change was also made in the flow rate of MUG. The total time period when there is no production of syngas. The methanol
flow rate value in Table 6 (4685 lbmol/h) is the peak flow of synthe- synthesis loop in that case can be isolated with the recycle com-
sis gas that is expected from upstream solar reactor, therefore step pressor running with a flow large enough so that the reactor can
response by cutting the flow to half the maximum and by stepping run at the minimum design capacity or gas hourly space velocity

0.32 0.086
[CH3OH]

[CO2]

0.3 0.084
0.082
0 20 40 60 80 100 0 20 40 60 80 100
Time (sec) Time (sec)
0.085 0.013
[H2O]
[CO]

0.08 0.012
0.075 0.011
0 20 40 60 80 100 0 20 40 60 80 100
Tme (sec) Time (sec)
0.19 240
T (C)
[H2]

239
0.185 238
0 20 40 60 80 100 0 20 40 60 80 100
Time (sec) Time (sec)

Fig. 10. Dynamic composition and temperature response at reactor exit. y-axis is the mole fraction of the components in the gas mixture.
126 S. Abrol, C.M. Hilton / Computers and Chemical Engineering 40 (2012) 117–131

97 260

Temperature (C)
250
96.95
% carbon-conversion

240

96.9 230
0 5 10 15 20 25 30 35 40 45 50
Time (sec)
96.85 750

Pressure (psi)
700
96.8
650

96.75 600
0 10 20 30 40 50 60 70 80 90 100 0 5 10 15 20 25 30 35 40 45 50
Time (sec) Time (sec)

Fig. 11. Dynamic response to a step change in MUG flow from FMUG to FMUG /2 at t = 0. Fig. 14. Temperature and pressure inside the batch reactor.

97 inside, it essentially becomes a batch reactor, and the dynamics of

such a reactor can be described using the following equations:
96.9
dCi  r cat a
% carbon-conversion

= i i (34)
dt εcat
96.8
 
∂T ḟt ∂T
Di
96.7 =− + U (T − T)
∂t Aεcat Ct ∂z Aεcat Ct Cp shell shell
96.6
cat a 
N
+ i ri (−Hf,i ) (35)
96.5
0 10 20 30 40 50 60 70 80 90 100
εcat Ct Cp
i=1
Time (sec)
Using the batch reactor model for the isolated system, the tem-
Fig. 12. Dynamic response to a step change in MUG flow from FMUG /2 to FMUG at t = 0.
perature and pressure goes to a new steady-state in less than a
minute as shown in Fig. 14. At the new steady-state value of pres-
(GHSV). Running an isolated loop ensures that there is no conden- sure and temperature (equal to boiling-water temperature) and
sation in the pipelines which would otherwise be an issue during given composition, the gas inside the reactor does not condense.
startup procedures for the next operation cycle. Depending on the In the next section a multivariable controller is designed to allow
composition of the gas and the pressure and temperature main- for non-continuous operation of the methanol loop.
tained inside the reactor, condensation can be avoided. In case the
recycle stream needs to run during the downtime, the bypass across 7. Model predictive control
the separator (Fig. 1) lets all the gas pass through it to the recycle
and there is no purge withdrawn from the loop. For this case, the Model predictive control or MPC refers to the advanced con-
dynamic response of methanol composition at the reactor exit and trol technique that uses an explicit process model, first-principles
the temperature of the reactor effluent stream are shown in Fig. 13. or empirical, to predict control moves of a plant driving it to
Methanol concentration and temperature of the reactor effluent go the desired set point while rejecting any process disturbances.
to simulated steady-state values in less than 15 s. In this example, The approach is well suited for multi-input multi-output (MIMO)
the loop is isolated when the flow rate of MUG at steady-state (t = 0) systems with strong interactions between the variables while sat-
is at peak value. isfying certain constraints or bounds on the input and output
In another scenario only the reactor may be idled with gas variables.
trapped inside, which provides a highly reducing environment. A model predictive controller utilizes the process model to solve
The boiling-water temperature is maintained on the shell-side to an online optimization problem to determine a sequence of control
ensure no condensation takes place inside the reactor. Condensa- moves, so that the predicted response reaches the set point in an
tion can also be prevented by having the pipelines in the recycle optimal manner. To solve the optimization problem a linear model
loop heavily insulated. When the reactor is isolated with the gas is often used to do the MPC calculations in real-time. Even today,
after so many decades when MPC was first introduced, practition-
ers are nervous about implementing MPC directly in the plant to
0.31
[CH3OH] (mol. fr.)

manipulate control valves. More often the classic PID controller is

0.3 used at the regulatory level and MPC may be used at a supervisory
0.29 level to provide the controller with set points based on the solution
of an economic optimization problem (Qin and Badgwell, 2003).
0 5 10 15 20 25 30 35 40 45 50 In this work, a regulatory (linear) MPC is designed for online
Time (sec)
implementation with the methanol synthesis loop. The objective
239
of the controller is to reject any (measured) disturbances observed
Temperature (C)

in the MUG flow rate, resulting from upstream solar variations

238
over short intervals, and also to track the changes in set points
corresponding to these disturbances. Override loops are not dis-
237
0 5 10 15 20 25 30 35 40 45 50 cussed here, which would be important in scenarios where there
Time (sec)
are long periods of zero MUG flow, and would require operation
Fig. 13. Composition of methanol and temperature at reactor exit for the isolated of the methanol plant as discussed in Section 6.1. Different types
loop. of linear models can be fitted to the input–output data like the
 S. Abrol, C.M. Hilton / Computers and Chemical Engineering 40 (2012) 117–131 127

Table 8 x 10
2.6
DVs, MVs and CVs for methanol synthesis loop. Transfer function
2.4 First-principles
a
Variables
2.2

Fin,reactor (lbmol/hr)
Disturbance variables (DVs) FMUG 2
Manipulated variables (MVs) Fbypass , Fpurge , Tshell
1.8
Controlled variables (CVs) Fin,reactor , [H2 O]liquid,flash , Tout,reactor
1.6
a
F, flow rate (lbmol/h); P, pressure (psi); T, temperature (◦ C); [·], mole fraction.
1.4

1.2
finite-impulse response (FIR) model, transfer function (TF) model, 1
state-space (SS) model or the auto-regressive with exogenous input 0.8
0 200 400 600 800 1000 1200 1400 1600 1800 2000
(ARX) model. Qin and Badgwell (2003) provide a review of differ- Time (sec)
ent types of models typically identified and used by industrial MPC
packages. Fig. 15. Comparison of the mathematical first-principles model and the fitted trans-
The first step in MPC design is the selection of controlled, manip- fer function model for output #1 (Fin,reactor ); Fit = 87.07%.
ulated and (measured) disturbance variables. For the synthesis loop
the variables chosen for control design are listed in Table 8. The
choice of the CVs was based on the known physical limitations and
design specifications for the process. The flow rate going into the
reactor, Fin,reactor determines the GHSV for the methanol reactor,
which has a lower bound as a design specification, below which the
diffusional limitations for the catalyst become significant, or there
are problems with the distribution of the gas inside the reactor
resulting in side reactions or development of hot spots. The amount
of water in crude methanol ([H2 O]liquid,flash ) obtained from the high
pressure separator may have a desired upper limit for certain grade
of methanol and/or to avoid the cost of having a distillation col-
umn downstream to remove the excess water. The temperature at
the reactor exit (Tout,reactor ) has to be controlled to avoid high reac-
Fig. 16. Comparison of the mathematical first-principles model and the fitted trans-
tion temperatures inside the reactor, which otherwise can cause fer function model for output #2 ([H2 O]liquid,flash ); Fit = 74.63%.
sintering of catalyst sites.
Selection of MVs, DVs and CVs is generally followed by a period
were deterministic and have an autocorrelation function similar
of open-loop pretests in which the steady-state gains for the pro-
to white noise (like in PRBS). Since the size (or amplitude) of the
cess and the settling time for each input–output pair is determined
excitation signals does not deviate too much from the nominal
(Seborg et al., 2004). Steady-state gains are estimated to ensure
value, such a signal also reduces the effect of nonlinearities on the
that during the actual plant test the inputs are not changed beyond
identified linear model.
a certain limit which might result in unstable behavior for the pro-
In this work, each of the MVs and DVs were perturbed one-at-a-
cess or for the outputs to exceed specified constraints. Since we
time in a sequential manner using PRBS. PRBS is a deterministic
will be using the first-principles loop model developed earlier in
sequence which has spectral characteristic of a random signal,
this work to make the ‘plant’ tests, only settling time was deter-
and is essentially a two-level periodic signal of a specified length.
mined by running the pretests on the dynamic model. The settling
Seborg et al. (2004) have mentioned various practical issues asso-
time is then used to determine the duration or frequency of the
ciated with moving several inputs simultaneously, and why the
excitation signals used in the subsequent ‘plant’ tests done to fit a
traditional practice of changing each variable individually is still
linear empirical model.
preferred.
Several authors have worked on identification of multivariable
In the case study shown here the maximum duration for a step
processes (Choi, Lee, Jung, Lee, & Han, 2000; Sotomayor, Park, &
change was 50 s based on the maximum settling time obtained dur-
Garcia, 2003; Wibowo, Saad, & Karsiti, 2009) including empirical
ing pretests. The changes in each variable take place for 500 s and
modeling of the large-scale Tennessee Eastman benchmark process
the other variables are fixed at their steady-state values during that
(Juricek, Seborg, & Larimore, 2001; Sriniwas & Arkun, 1997). Most of
time, after which the other variables are changed one-at-a-time
these identification techniques use subspace methods with a state-
in a sequential manner. So the total number of time steps in the
space model to fit experimental or simulation data. An advantage
of using the state-space structure is that it allows inclusion of noise
in the model, and more importantly any linear model structure
(ARX, FIR or TF) can be easily represented by a state-space model
(Ljung, 1999). Different types of excitation input signals have been
used by these authors for model identification. Juricek et al. (2001)
used a three-level sequence similar to the pseudo random binary
sequence (PRBS) as it has been shown that such signals reduce the
effect of nonlinearities on the resulting linear model. In the three-
level sequence, the middle level is the nominal values of the inputs
and the other two are ±ı from the nominal values. Wibowo et al.
(2009) developed three different models based on different types
of input signals used, with sequential step changes and simultane-
ous PRBS signals in the inputs. Sotomayor et al. (2003) used inputs
sequences that correspond to multi-level uniformly distributed Fig. 17. Comparison of the mathematical first-principles model and the fitted trans-
(and not cross-correlated) random sequences. The sequences used fer function model for output #3 (Tout,reactor ); Fit = 83.63%.
128 S. Abrol, C.M. Hilton / Computers and Chemical Engineering 40 (2012) 117–131

Table 9 x 10
3
Constraints on the variables in MPC.

Fin,reactor
2
Lower bound Upper bound Units 1
0 50 100 150 200 250 300 350 400
MVs Time (sec)

[H2O]liq,flash
Fbypass 0 23,000 lbmol/h 0.08
Fpurge 0 21,000 lbmol/h 0.06
◦
Tshell 200 260 C 0.04
CVs 0 50 100 150 200 250 300 350 400
Time (sec)
Fin,reactor 10,000 50,000 lbmol/h 240

Tout,reactor
[H2 O]liquid, flash 0 0.08 mol. fr.
◦ 230
Tout,reactor 200 260 C
220
0 50 100 150 200 250 300 350 400
Time (sec)
dynamic test run were 2000 (=500 time steps each x (3 MVs + 1 DV)).
The size of each time step was 1, and with the initial steady-state Fig. 18. Closed-loop response for the CVs with the designed MPC.
data point the total number of points were 2001.
A total of 12 (=4 inputs × 3 outputs) TF models were fitted to
the data. Different orders of TF models were tried, with time delays feasible regions as defined by the constraints. The rate of change of
and/or zeros in the numerator, and a second order TF with numera- MVs for the flow rates (Fbypass and Fpurge ) were given equal weights
tor dynamics was found out to be the best fit to represent dynamics (of 0.1) and the temperature of the boiling water (Tshell ) was given
of each input–output pair. As a single run of 2001 time steps was a much higher weight of 10, as the temperature of this huge boiling
used to fit the models, multi-input single-output (MISO) models mass would require much longer time to be manipulated. Higher
were used for each output. These models were identified using the values for the weights on the rate of change of MVs were also
prediction-error minimization (PEM) approach in the system iden- tried ([10 10 1000] instead of [0.1 0.1 10]), which resulted in large
tification toolbox of MATLAB. The transfer function models fitted
to each input–output pair are as shown in Eq. (36).
⎡ ⎤
−5.63(−226.58 s + 1) −606.7(77.37 s + 1) 401(111.82 s + 1)
  ⎢ (1188.6 s + 1)(0.44 s + 1) (4018.2 s + 1)(11.93 s + 1) (172.06 s + 1)(0.45 s + 1) ⎥F 
Fin,reactor (s) ⎢ −2.14e−9 (127.85 s + 1) −1.4e−4 (220.47 s + 1) 3.56e−4 (1061.7 s + 1) ⎥ bypass (s)
[H2 O]liquid,flash (s) =⎢
⎢
⎥ Fpurge (s)
⎥
(18.26 s + 1)(18.71 s + 1) (581.3 s + 1)(23.95 s + 1) (1644.4 s + 1)(0.134 s + 1)
Tout,reactor (s) ⎣ −7.2e−4 (−75.39 s + 1) −3.65e−3 (10.58 s + 1) 0.944(97.49 s + 1)
⎦ Tshell (s)
(849.08 s + 1)(2.25 s + 1) (187.2 s + 1)(0.001 s + 1) (94.33 s + 1)(0.64 s + 1)
⎡ ⎤
4.68(50.92 s + 1)
⎢ (56.55 s + 1)(0.26 s + 1) ⎥
⎢ 2.86e−6 (11.34 s + 1) ⎥
+⎢
⎢
⎥ [FMUG (s)]
⎥ (36)
(19.39 s + 1)(19.42 s + 1)
⎣ −1.11e−4 (−1446.1 s + 1)
⎦
(6182.9 s + 1)(3.56 s + 1)
As a measure of the accuracy for the fitted models, the following
criterion was used steady-state offsets for the CVs and increased the settling time for
 
||Y − Ŷ || the response. The CVs were given equal weights of 1. ‘Tuning advi-
Fit = 1 − × 100% (37) sor’ in the MPC toolbox was later used to fine tune these weights
||Y − Ȳ ||
based on the controller’s performance.
where Y is the output from the first-principles model, Ŷ is the output Wojsznis, Gudaz, Blevins, and Mehta (2003) have outlined a
for the fitted transfer function model and Ȳ is the mean of first- heuristic approach for practical tuning of an industrial MPC, and
principles model outputs for the entire run. Figs. 15–17 compare the Garriga and Soroush (2010) provide a comprehensive review of the
response of the original mathematical model and the fitted transfer available tuning methods for MPC both from theoretical and from
function models for the three outputs, and also give the fit error as practical perspectives. For this case study, a prediction horizon of
defined by Eq. (37). 50 and a control horizon of 20 were chosen based on the thumb
The fitted TF models are then merged and converted to a single rules given by Seborg et al. (2004) as typically used in the indus-
state-space model assuming that the process can be approxi- try, and based on the settling time obtained for the methanol loop
mated by these linear models and that superposition works (Qin during the pretests.
& Badgwell, 2003). There are tests that can be performed to check Next, a test run was made using the designed controller for a step
if the nonlinear model data can be well approximated by a lin- change in MUG flow rate (maximum design rate changed to half the
ear model (Billings & Voon, 1983). The linear state-space model maximum) at t = 10 s. The set points of Fin,reactor and [H2 O]liquid,flash
obtained is then used to solve the optimization problem in the MPC were changed along with the step change in MUG flow rate. The
toolbox of MATLAB (Morari & Ricker, 1994). result of the closed-loop simulation run with the MPC controller
The set points for the controller are usually based on economic is shown in Fig. 18. The red dotted lines are the set points for the
optimization (maximizing profit or production, or minimizing a CVs. The corresponding changes in the input variables are shown
cost function) on a steady-state model, and in the case study shown in Fig. 19.
here, it is calculated using the first-principles steady-state model As seen from Fig. 18, the controller is able to track the set point
for the disturbance in the MUG flow rate considered. The lower and changes corresponding to the change in the MUG flow rate with no
upper bounds (Table 9) are based on the design specification on the steady-state error or offset. Also, the controller tuning used here
CVs considered here. does not result in oscillation of the outputs, and as expected is
The weights in the MPC are tuned such that the MVs have fast for flow control (Fin,reactor ) as compared to the other two CVs.
zero weights, implying that they can vary within their respective The MVs change accordingly within the allowed limits to track the
 S. Abrol, C.M. Hilton / Computers and Chemical Engineering 40 (2012) 117–131 129

100 analysis if strong interactions are observed between the variables a

Fbypass
50
0
0 50 100 150 200 250 300 350 400
decoupling strategy using decouplers or a detuning strategy based
Time (sec) on Lambda control can also be adopted to tune the PID loops. A
600
Fpurge

400 plantwide control strategy may be developed to allow communi-

200
0 50 100 150 200 250 300 350 400 cation between the MPC for the solar-driven gasification process
Time (sec)
240 on the front-end and the methanol loop at the back-end. As sug-
Tshell

230
220
gested by Venkat, Rawlings, and Wright (2007) a centralized control
0 50 100 150 200 250 300 350 400
Time (sec) strategy with a distributed or hierarchical MPC technique (either
6000 communication-based or cooperation-based) can also be consid-
FMUG

4000
2000
0 50 100 150 200 250 300 350 400
ered for such a control structure.
Time (sec)

Fig. 19. MVs and DVs for measured disturbance (in FMUG ) and set point tracking 8. Conclusion
using designed MPC.
In this work, development of the methanol synthesis (reactor-
separator) loop model is described in detail. Approximate
set points, and because of interactions between the variables the
pore-diffusion effects for the Bussche and Froment (1996) kinetics
outputs change simultaneously with changing inputs.
are included in the model to account for any diffusional limita-
Ricker and Lee (1995) demonstrated application of a nonlinear
tions. Dynamic homogeneous model of a fixed-bed catalytic reactor
MPC (NMPC) to the Tennessee Eastman process for an uncon-
is developed and combined with an SRK-based separator using
strained 8 × 8 system. For the controller designed by the authors
a sequential-modular approach, and the recycle stream is torn to
it was seen that even if only one of the MVs was constrained it
simulate the loop. Variable-space simulation runs were carried
was impossible to maintain all the 8 CVs at their set points and
out using the steady-state loop model to study different operat-
steady-state offsets were observed in all the CVs. Also, the authors
ing regimes resulting in similar loop performance, and to provide
found that no single set of weights and constraints could provide
guidance in making design and economic decisions. Case studies
desired control performance in all scenarios of set point tracking or
using the dynamic model were done to provide insight into the
disturbance rejection, and SISO override loops were added to the
process behavior including settling times, process interactions dur-
designed MPC to achieve this. It is emphasized by Ricker (1996) that
ing transients and design of a model predictive controller. A bypass
using PI loops for the benchmark process, the performance is nearly
stream was included in the process to allow cycling and opera-
as good as NMPC but the control design is much more tedious.
tion of the methanol loop for varying syngas flow rates. Methanol
The theoretical MPC designed here aims at achieving robust
loop performance was also studied when the loop or the reactor is
control for a typical scenario during operation of the methanol
isolated during downtime of the upstream syngas process. Finally,
loop which is driven by an upstream solar syngas production pro-
design of a preliminary model predictive controller for the control
cess. Inclusion of the bypass stream in the methanol loop allows
of the methanol synthesis process is described. The controller aims
for an extra degree of freedom to control the flow going into the
at rejecting disturbances in syngas flow rate due to variability in
reactor, especially during low flow rates of MUG. As mentioned
the solar-driven upstream process, and also track the correspond-
before, the control design does not include any override loops,
ing changes in the set points at a regulatory level. A linear MPC is
which would play an important role in scenarios when there is zero
designed based on a linear process model identified using the data
flow of MUG. Level control of the reactor or the separator was not
generated from running the first-principles model. A more com-
included, which would override the flow loops in certain scenar-
prehensive approach as an extension of the work done here can be
ios. Ricker (1996) has outlined use of overrides for the Tennessee
adopted to design a plantwide control structure for a solar-driven
Eastman process which can be included in a decentralized control
process, and cycling of the downstream non-continuous methanol
scheme. The author mentions that inventory control (accumulation
synthesis process.
or depletion) of chemical species is also crucial for processes with
recycle.
Another important issue that needs to be addressed is that of Acknowledgments
time-delay in the process, mainly because of transport of material.
Time-lag has to be included in the dynamic model to get a better The authors would like to thank Dr. Babatunde A. Ogunnaike at
understanding of the response time to disturbances, and for robust University of Delaware for his useful comments and suggestions on
control design. The time-delay can be estimated knowing the pipe this work, and Sundrop Fuels Inc. in Louisville, CO for providing the
sizing and the changing flow rates in the loop. A control structure financial support on the project.
which takes into account the varying time-delay effect needs to be
studied. Appendix A.
The compressor speed (not included in the model) in the loop
can also be included in the control structure to accommodate for A.1. Heat of formation
changing recycle flow rates. In the base control of Tennessee East-
man process, McAvoy and Ye (1994) have included compressor Heat of formation for the reacting components CH3 OH, CO2 , CO,
power and compressor exit flow rate as controlled variables. H2 O and H2 from the following equations
As mentioned earlier, in practice MPC might be used as a
supervisory controller to perform optimization based set-point cal- 1
C(s) + O2 (g) + 2H2 (g) → CH3 OH(g), Hf,298 ◦ = −200, 660 J
culations, which will provide changing set-points to the lower-level 2
regulatory PID control loops through the distributed control sys- (A1)
tem or DCS. In such a design, set-point calculation based on an
economic cost function can be included in the objective of the C(s) + O2 (g) → CO2 (g), Hf,298 ◦ = −393, 509 J (A2)
MPC optimization routine (Rawlings & Amrit, 2009). To tune multi-
loop PID controllers, sequential loop tuning methods can be used 1
C(s) + O2 (g) → CO(g), Hf,298 ◦ = −110, 525 J (A3)
(Hovd & Skogestad, 1994). Based on the relative gain array (RGA) 2
130 S. Abrol, C.M. Hilton / Computers and Chemical Engineering 40 (2012) 117–131

1 References
H2 (g) + O2 (g) → H2 O(g), Hf,298 ◦ = −241, 818 J (A4)
2
Apt, J., Newcomer, L., Lave, L. B., Douglas, S., & Dunn L. M. (2008). An Engineering-
can be calculated as a function of temperature, T with respect Economic Analysis of Syngas Storage (Final Report DOE/NETL-2008/1331).
to standard temperature, T0 (=298 K) using the equation given in Billings, S. A., & Voon, W. S. F. (1983). Structure detection and model validity tests
in the identification of nonlinear systems. Control Theory and Applications, IEE
Smith et al. (2003):
Proceedings D, 130–193.
Bussche, K. M. V., & Froment, G. F. (1996). A steady-state kinetic model for methanol
HT ◦ = H298 ◦ + Cp ◦
(T − T0 ) (A5) synthesis and the water gas shift reaction on a commercial Cu/ZnO/Al2 O3 cata-
lyst. Journal of Catalysis, 161, 1–10.

Cp ◦
H B
T  Camacho, E. F., Berenguel, M., & Rubio, F. R. (1997). Advanced control of solar plants.
London: Springer.
= A + T0 +1 Chang, T., Rousseau, R. W., & Kiipatrick, P. K. (1986). Methanol synthesis reactions:
R 2 T0
 2 T  Calculations of equilibrium conversions using equations of state. Industrial and
Engineering Chemistry Process Design and Development, 25, 477–481.
C 2 T D
+ T + +1 + (A6) Choi, J. Y., Lee, J., Jung, J. H., Lee, M., & Han, C. (2000). Sequential loop closing iden-
3 0 T0 T0 TT0 tification of multivariable process models. Computers and Chemical Engineering,
24, 809–814.
The values of the constants A, B, C and D for the species in reactions Fagley, J. C., & Carnahan, B. (1990). The sequential-clustered method for
dynamic chemical plant simulation. Computers and Chemical Engineering, 14(2),
(A1)–(A4) are given in Smith et al. (2003). 161–177.
Fuller, E. N., Schettler, P. D., & Giddings, J. C. (1996). A new method for prediction of
binary gas-phase diffusion coefficients. Industrial and Engineering and Chemistry,
A.2. Specific heat at constant pressure 58(5), 19–27.
Garriga, J. L., & Soroush, M. (2010). Model predictive control tuning methods: A
Specific heat for all the components CH3 OH, CO2 , CO, H2 O, H2 , review. Industrial and Engineering Chemistry Research, 49, 3505–3515.
Graaf, G. H., Scholtens, H., Stamhuis, E. J., & Beenackers, A. A. C. M. (1990). Intra-
CH4 and N2 can be calculated based on the equation given in Smith particle diffusion limitations in low-pressure methanol synthesis. Chemical
et al. (2003). Engineering Science, 45(4), 773–783.
Hillestad, M., & Hertzberg, T. (1986). Dynamic simulation of chemical engineer-
CP,i D ing systems by the sequential modular approach. Modeling, Identification and
= A + BT + CT 2 + 2 (A7) Control, 7(3), 107–127.
R T Hovd, M., & Skogestad, S. (1994). Sequential design of decentralized controllers.
Automatica, 30.
The specific heat at constant pressure for a gas mixture with a mole
Juricek, B. C., Seborg, D. E., & Larimore, W. E. (2001). Identification of the Tennessee
fraction yi for component i can be defined by a simple mixing law: Eastman challenge process with subspace methods. Control Engineering Practice,
9, 1337–1351.

N Klier, K., Chatikavanij, V., Herman, R. G., & Simmons, G. W. (1982). Catalytic synthesis
Cp = yi Cp,i (A8) of methanol from CO/H2 : IV. The effects of carbon dioxide. Journal of Catalysis,
74, 343.
i=1 Kulikov, V., Briesen, H., Grosch, R., Yang, A., Wedel, L., & Marquardt, W. (2005). Mod-
ular dynamic simulation for integrated particulate processes by means of tool
integration. Chemical Engineering Science, 60, 2069–2083.
Appendix B. Levenspiel, O. (1993). Chemical reactor omnibook. OSU Publishing.
Liu, Y. C., & Brosilow, C. B. (1983). Modular integration methods for large scale dynamic
B.1. Wegstein convergence block systems. Washington, DC: AIChE Meeting.
Ljung, L. (1999). System identification (2nd Ed.). Upper Saddle River, NJ: Prentice Hall.
Lovik, I. (2001). Modelling, estimation and optimization of the methanol synthesis with
Following are the steps for Wegstein update to the recycle catalyst deactivation (Ph.D. Dissertation, Norwegian University of Science and
stream: Technology).
Lommerts, B. J., Graaf, G. H., & Beenackers, A. A. C. M. (2000). Mathematical mod-
eling of internal mass transport limitations in methanol synthesis. Chemical
1. Initial guess of recycle flow rate, Frecycle0, and composition, yre- Engineering Science, 55, 5589–5598.
cycle0; Luyben, W. L. (1994). Snowball effects in reactor/separator processes with recycle.
Industrial and Engineering Chemistry Research, 33, 299–305.
2. Calculate recycle flow rate and composition by solving the loop, MATLAB users’ guide, version 2009b (2009).
Frecycle cal0 and yrecycle cal0; McAvoy, T. J., & Ye, N. (1994). Base control for the Tennessee Eastman problem.
3. Frecycle1 = Frecycle cal0 (=Frecycle cal prev) and yrecy- Computers and Chemical Engineering, 18(5), 383–413.
Meyer, R. A. (1984). Handbook of synfuels technology (1st ed.). McGraw Hill Higher
cle1 = yrecycle cal0 (=yrecycle cal prev); Education.
4. Calculate Frecycle cal1 (=Frecycle cal) and yrecycle cal1 (=yre- Morari, M., & Ricker, N. L. (1994). Model predictive control toolbox. MA: The Math-
cycle cal) by solving the loop; works Inc.
Frecycle cal − Frecycle cal prev Mueller, R. W., Dagestad, K. F., Ineichen, P., Schroedter-Homscheidt, M., Cros, S.,
5. wF = , Dumortier, D., et al. (2004). Rethinking satellite-based solar irradiance mod-
Frecycle − Frecycle prev elling: The SOLIS clear-sky module. Remote Sensing of Environment, 91(2),
160–174.
Parvasi, P., Rahimpour, M. R., & Jahanmiri, A. (2008). Incorporation of dynamic flex-
yrecycle cali − yrecycle cal previ
wy,i = , ibility in the design of a methanol synthesis loop in the presence of catalyst
yrecyclei − yrecycle previ deactivation. Chemical Engineering Technology, 31(1), 116–132.
Petrasch, J., Osch, P., & Steinfeld, A. (2009). Dynamics and control of solar thermo-
chemical reactors. Chemical Engineering Journal, 145, 362–370.
Reid, R. C., Prausnitz, J. M., & Poling, B. E. (1997). The properties of gases and liquids
wF wy,i (4th ed.). McGraw Hill.
qF = , qy,i = , i = 1, 2, . . . , N Qin, S. J., & Badgwell, T. A. (2003). A survey of industrial model predictive control
wF − 1 wy,i − 1 technology. Control Engineering Practice, 11, 733–764.
Rawlings, J. B., & Amrit, R. (2009). Optimizing process economic performance
The lower and upper bounds on the accelerating parameter, q using model predictive control, nonlinear model predictive control. LNCIS, 384,
119–138.
were −5 and 0. Rezaie, N., Jahanmiri, A., Moghtaderi, B., & Rahimpour, M. R. (2005). A comparison of
6. Frecycle = qF Frecycle + (1 − qF )Frecycle cal, homogenous and heterogenous dynamic models for industrial methanol reac-
tors in the presence of catalyst deactivation. Chemical Engineering and Processing,
44, 911–921.
yrecyclei = qy,i yrecyclei + (1 − qy,i )yrecycle cali ; Ricker, N. L. (1996). Decentralized control of the Tennessee Eastman challenge pro-
cess. Journal of Process Control, 6(4), 205–221.
Ricker, N. L., & Lee, J. H. (1995). Nonlinear model predictive control of the Tennessee
7. Check for convergence, if ‘no’, go to step 4, if ‘yes’, exit. Eastman challenge process. Computers and Chemical Engineering, 19(9), 961–981.
 S. Abrol, C.M. Hilton / Computers and Chemical Engineering 40 (2012) 117–131 131

Seborg, D. E., Edgar, T. F., & Mellichamp, D. A. (2004). Process dynamics and control Sriniwas, G. R., & Arkun, Y. (1997). Control of the Tennessee Eastman
(2nd ed.). John Wiley & Sons. process using input–output models. Journal of Process Control, 7(5),
Shahrokhi, M., & Baghmisheh, G. R. (2005). Modeling, simulation and control 387–400.
of a methanol synthesis fixed-bed reactor. Chemical Engineering Science, 60, Tindall, B. M., & Crews, M. A. (1995). Alternative technologies to steam-methane
4275–4286. reforming. Hydrocarbon Processing, 74(11), 75–82.
Silva, R. N., Lemos, J. M., & Rato, L. M. (2003). Variable sampling adaptive control of a Venkat, A. N., Rawlings, J. B., & Wright, S. J. (2007). Distributed model predictive
distributed collector solar field. IEEE Transactions on Control Systems Technology, control of large-scale systems, assessment and future directions. LNCIS, 358,
11(5), 765–772. 591–605.
Skrzypek, J., Słoczyński, J., & Ledakowicz, S. (1994). Methanol synthesis. PWN Wegstein, J. H. (1958). Accelerating convergence of iterative processes. Washington,
Warszawa. DC: National Bureau of Standards.
Smith, J. M., Van Ness, H. C., & Abott, M. M. (2003). Chemical engineering thermody- Wibowo, T. C. S., Saad, N., & Karsiti, M. N. (2009). System identification of an inter-
namics (6th ed.). McGraw Hill. acting series process for real-time model predictive control. In American Control
Soave, G. (1972). Equilibrium constants for modified Redlich–Kwong equation-of- Conference St. Louis, MO, June 10–12.
state. Chemical Engineering Science, 27, 1196–1203. Wojsznis, W., Gudaz, J., Blevins, T., & Mehta, A. (2003). Practical approach to tuning
Sotomayor, O. A. Z., Park, S. W., & Garcia, C. (2003). Multivariable identification of an MPC. ISA Transactions, 42, 149–162.
activated sludge process with subspace-based algorithms. Control Engineering
Practice, 11, 961–969.