A Superstructure-Based Optimal Synthesis of

PSA Cycles for Post-Combustion CO2 Capture

Anshul Agarwal and Lorenz T. Biegler
Collaboratory for Process and Dynamic Systems Research, National Energy Technology Laboratory,
Morgantown, WV 26507
Chemical Engineering Dept., Carnegie Mellon University, Pittsburgh, PA 15213

Stephen E. Zitney
Collaboratory for Process and Dynamic Systems Research, National Energy Technology Laboratory,
Morgantown, WV 26507

DOI 10.1002/aic.12107
Published online November 9, 2009 in Wiley InterScience (www.interscience.wiley.com).

Recent developments have shown pressure/vacuum swing adsorption (PSA/VSA) to

be a promising option to effectively capture CO2 from flue gas streams. In most com-
mercial PSA cycles, the weakly adsorbed component in the mixture is the desired prod-
uct, and enriching the strongly adsorbed CO2 is not a concern. On the other hand, it
is necessary to concentrate CO2 to high purity to reduce CO2 sequestration costs and
minimize safety and environmental risks. Thus, it is necessary to develop PSA proc-
esses specifically targeted to obtain pure strongly adsorbed component. A multitude of
PSA/VSA cycles have been developed in the literature for CO2 capture from feedstocks
low in CO2 concentration. However, no systematic methodology has been suggested to
develop, evaluate, and optimize PSA cycles for high purity CO2 capture. This study
presents a systematic optimization-based formulation to synthesize novel PSA cycles
for a given application. In particular, a novel PSA superstructure is presented to
design optimal PSA cycle configurations and evaluate CO2 capture strategies. The
superstructure is rich enough to predict a number of different PSA operating steps.
The bed connections in the superstructure are governed by time-dependent control var-
iables, which can be varied to realize most PSA operating steps. An optimal sequence
of operating steps is achieved through the formulation of an optimal control problem
with the partial differential and algebraic equations of the PSA system and the cyclic
steady state condition. Large-scale optimization capabilities have enabled us to adopt
a complete discretization methodology to solve the optimal control problem as a large-
scale nonlinear program, using the nonlinear optimization solver IPOPT. The super-
structure approach is demonstrated for case studies related to post-combustion CO2
capture. In particular, optimal PSA cycles were synthesized, which maximize CO2 re-
covery for a given purity, and minimize overall power consumption. The results show
the potential of the superstructure to predict PSA cycles with up to 98% purity and
recovery of CO2. Moreover, for recovery of around 85% and purity of over 90%, these
cycles can recover CO2 from atmospheric flue gas with a low power consumption of

Correspondence concerning this article should be addressed to L. T. Biegler at

lb01@andrew.cmu.edu.

This is a U.S. Government work and, as such, is in the public domain in

the United States of America.

AIChE Journal July 2010 Vol. 56, No. 7 1813

 465 k Wh tonne
1 CO2. The approach presented is, therefore, very promising and quite
useful for evaluating the suitability of different adsorbents, feedstocks, and operating
strategies for PSA, and assessing its usefulness for CO2 capture. Published 2009 Ameri-
can Institute of Chemical Engineers AIChE J, 56: 1813–1828, 2010
Keywords: adsorption/gas, optimization, design

Introduction ter the adsorption step in the PSA cycle. Hence, to obtain a
Today, fossil fuels provide about 85% of the global energy pure heavy product gas, a heavy product pressurization step
demand and the outlook is that they will remain the domi- or a heavy reflux step is necessary in the cycle, similar to a
nant source of energy for decades to come. Consequently, light product pressurization step or a light reflux step in the
global energy-related CO2 emissions, especially from power conventional PSA cycles. This idea was first suggested in a
plants that burn fossil fuels, have increased, thereby increas- patent by Tamura,6 and has been incorporated in most of the
ing CO2 concentration levels in the atmosphere.1 One option PSA cycles that have been suggested in the literature for
to mitigate the emission of CO2 is to capture it from emis- high purity CO2 separation from flue gas.
sion sources, store it in the ocean or underground, or use it A fairly comprehensive review of the previous studies on
for enhanced oil and coal bed methane recovery. Before CO2 PSA cycles for concentrating CO2 from flue gas is presented
can be sequestered, it must be separated and concentrated in the next section. This review highlights the difficulties
from a flue gas with a low CO2 concentration. There are a associated with choosing one PSA cycle over another for a
variety of approaches to CO2 separation from other flue gas given application. From these studies, it is not clear why a
components, such as gas absorption, membranes, cryogenic particular cycle was chosen or one performed better than
distillation, gas adsorption and others, each with their own other configurations. More importantly, the review shows
pros and cons.2 Currently, absorption based technologies are that so far no systematic algorithm or method has been
commercially utilized for CO2 capture, in which different developed in the literature to design and evaluate a PSA
kinds of amines are used as solvents for absorbing CO2 from cycle configuration to obtain a pure light or a heavy product
flue gas. Typical energy requirements for the leading absorp- or both simultaneously. Therefore, our objective is to present
tion technologies are quite high, ranging from 765 to 950 a systematic approach to develop, evaluate, and optimize
k Wh tonne
1 CO2 captured (excluding energy requirement PSA cycles for CO2 capture.
for CO2 compression).3 Recent developments have shown In this article, we present an optimization-based frame-
pressure/vacuum swing adsorption to be a promising option work to generate optimal PSA cycles from a 2-bed PSA
for separating CO2 at high purity and with low power superstructure, and apply the approach to develop optimal
requirements. PSA processes have been widely applied for configurations to separate CO2 from a flue gas mixture. The
the removal of CO2 from various feed mixtures, such as CO2 cycles are optimal because they are realized after solving an
in the steam reformer off-gas, natural gas, and flue gas.4 optimal control problem. It is worth noting that in this work,
They are also commercially used to remove trace amounts of we consider a detailed partial differential algebraic equation
CO2 from air.5 In these commercial PSA cycles, the weakly (PDAE) based mathematical model, with the cyclic steady
adsorbed (or light) component in the mixture is the desired state condition, for the optimization problem and the model
product and enriching the strongly adsorbed (or heavy) com- accommodates detailed adsorption physics. The PDAE-based
ponent (in this case, CO2) is not a concern. On the other model is completely discretized in both spatial and time
hand, for CO2 sequestration, it is necessary to concentrate domains, and the resulting large-scale nonlinear program-
CO2 to a high purity to reduce the compression and the ming problem is solved using an interior point solver. ‘‘PSA
transportation cost. Moreover, safety and environmental Superstructure’’ section describes the optimal control frame-
issues are the additional reasons for concentrating CO2 to a work formulated to realize different operating steps in a PSA
high purity. cycle. ‘‘PSA Model and Solution Methodology’’ section
Typically adsorbents preferentially adsorb CO2 from a flue describes the PSA model equations and the solution strategy
gas mixture, consequently making it a heavy product. The while the subsequent section describes the case studies. Con-
conventional PSA cycles are inappropriate for concentrating cluding remarks are outlined in the final section.
heavy product because the light product purge step (or the
light reflux step) in these cycles uses a portion of the light
product gas, which necessarily dilutes the heavy component
in the heavy product stream. As a result, a pure light compo- Literature Review
nent is easy to attain from such cycles, but not a pure heavy Since 1992, when the Japanese power industry started
component. Thus, it is necessary to develop PSA processes investigating flue gas CO2 removal using gas adsorption,7–11
specifically targeted to obtain pure strongly adsorbed compo- a multitude of PSA/VSA cycles have been developed in the
nent, CO2 in this case. literature to produce pure CO2 from a flue gas mixture. We
Because the product purity of the heavy component is lim- provide a summary of these studies in Table 1. In this table,
ited by the gas mixture occupying the void spaces in the yf is the CO2 % in feed, while pCO2 and rCO2 are CO2 purity
bed, its purity can be increased by displacing the gas mixture and recovery in the heavy product stream, respectively, and
in the void spaces with a pure heavy product gas. For Pl is the vacuum/low pressure used to extract CO2 at high
instance, for the separation of N2-CO2 mixture, the displace- purity. The terminology for various operating steps in a
ment can be accomplished by purging the bed with CO2 af- cycle is adopted from Reynolds et al.12 Most of these studies

1814 DOI 10.1002/aic Published on behalf of the AIChE July 2010 Vol. 56, No. 7 AIChE Journal
 Table 1. PSA Cycles Suggested in the Literature for Post-Combustion CO2 Separation
PSA Cycle Operating Step Feed throughput
Configuration Sequence* Ads† yf (%) pCO2 (%) rCO2 (%) Pl (kPa) (kg-mol h
1) References
5-bed 5-step F,HR,CnD,LR,LPP HTlc 15 72 82 11.49 0.001 12‡
5-bed 5-step F,HR,CnD,LR,LPP HTlc 15 76 49 11.49 0.003 12‡
4-bed 4-step F,HR,CnD,LPP HTlc 15 83 17 11.49 0.001 12‡
5-bed 5-step F,HR,CnD,LR,LPP HTlc 15 98.7 98.7 11.64 0.00052 13‡
5-bed 5-step FþR,HR,CnD,LR,LPP HTlc 15 98.6 91.8 11.64 0.00052 13‡
4-bed 4-step F,HR,CnD,LPP HTlc 15 99.2 15.2 11.64 0.006 13‡
4-bed 4-step FþR,HR,CnD,LPP HTlc 15 99.2 15.2 11.64 0.006 13‡
4-bed 4-step LPP,FþR,HR,CnD AC 17 99.9 68 10.13 16.19 14‡
3-bed 8-step FP,F,CoD,R,N,HR,CnD,N 13X 16 99 45 6.67 0.049 15
4-bed 8-step FP,F,HR,LEE,CnD,LR,LEE,N NaX 13 95 50 10 1.116 16‡,§
2-bed 4-step FP,F,CnD,LR 13X 10 70 68 4 0.331 17‡
2-bed 6-step LEE,FP,F,LEE,CnD,LR 13X 10 82 57 6.67 0.331 17‡
3-bed 5-step FP,F,HR,CnD,LR 13X 10 83 54 6.67 0.331 17‡
2-bed 4-step FP,F,CnD,LR 13X 8.3 78 50 101.3 0.004 18§
3-bed 8-step FP,F,CoD,LEE,HPP,HR,CnD,LEE AC 17 99.8 34 10.13 0.027 19‡,§
3-bed 7-step FP,F,LEE,HR,N,CnD,LEE AC 13 99 55 10.13 0.204 20‡,§
3-bed 8-step FP,F,CoD,LEE,HPP,HR,CnD,LEE 13X 13 99.5 69 5.07 0.025 21‡,§
2-bed 4-step HPP,FP,CoD,CnD 13X 20 48 94 5.07 – 22
2-bed 5-step HPP,FP,F,CoD,CnD 13X 20 43 88 5.07 – 22
3-bed 4-step LPP,F,CnD,LR 13X 20 58 75 5.07 – 22
3-bed 6-step LPP,FP,F,HR,CoD,CnD 13X 20 63 70 5.07 – 22
2-bed 4-step FP,F,CnD,LR 13X 15 72 94 90 30.35 23
1-bed 4-step FP,F,CoD,CnD 13X 15 90 94 70 1.741 23
2-bed 4-step LPP,F,CnD,LR 13X 15 52 66 10 0.007 24§
3-bed 5-step LPP,F,HR,CnD,LR 13X 15 83 66 10 48.57 24§
3-bed 6-step F,LEE,CnD,LEE 13X 12 83 60 4 0.193 25‡,§
3-bed 9-step F,LEE,HR,CnD,LEE 13X 12 95 60 5 0.193 26‡,§
3-bed 9-step F,LEE,I,LEE,CnD,LEE,FP 13X 12 92.5 75 3 0.327 27‡

*Cycle-step legend: CnD, counter-current depressurization; CoD, co-current depressurization; FP, feed pressurization; F, feed or adsorption; HPP, heavy product
pressurization; HR, heavy reflux; LEE, light end equalization; LPP, light product pressurization; LR, light reflux; N, null or idle; R, recycle.
†
Adsorbent legend: HTlc, K-promoted Hydrotalcite; NaX, 13X, molecular sieve zeolites; AC, activated carbon.
‡
Studies with experimental results.
§
Multicomponent study.

are bench-scale and deal with extremely small feed through- rinse step and the equalization step in the 3-bed 5-step
put. cycle improved their CO2 purity and recovery, it didn’t
Ritter and co-workers have studied numerous PSA cycles decrease their power requirements, which were 106.91
for CO2 capture from a feed at high temperature using kWh tonne
1 CO2 captured for the 2-bed 6-step cycle, and
K-promoted Hydrotalcite as the adsorbent.12,28,29 They have 147.64 kWh tonne
1 CO2 captured for the 3-bed 5-step
emphasized the importance of including heavy reflux step cycle. Though the power consumption was quite low, the
to obtain heavy product at a high purity. They compared feed throughput of 0.331 kg-mol h
1 was also on the
seven different 4-bed 4-step, 4-bed 5-step, and 5-bed 5- lower side. The conventional 2-bed 4-step Skarstrom cycle
step cycle configurations with and without heavy reflux was also studied by Gomes et al.,18 in which they didn’t
step. In another work,13 they analyzed nine different PSA apply vacuum to recover CO2. Their work also showed
configurations and achieved better purities and recoveries that the light reflux step itself is not sufficient to obtain
for CO2, although, at an extremely small feed throughput. pure heavy component.
Kikkinides et al.14 studied a 4-bed 4-step vacuum swing Na et al.19,20 and Choi et al.21 studied 3-bed 8-step and
process and improved CO2 purity and recovery by allowing 3-bed 7-step VSA configurations experimentally as well as
significant breakthrough of CO2 from the light end of the numerically. Light reflux step was not used for any of
column undergoing heavy reflux, and then, recycling the these configurations, whereas heavy reflux was used in all
effluent from this light end back to the column with the of them. The 2-bed cycles of Chou and Chen22 did not
feed. Chue et al.15 compared activated carbon and zeolite use any kind of reflux steps whereas the 3-bed cycles used
13X using a 3-bed 9-step VSA process. They suggested both light and heavy reflux steps. The 2-bed cycles were
that despite a high heat of adsorption of CO2, zeolite 13X unconventional as flow reversal was implemented in
is better because of its higher working capacity, lower between the pressurization and depressurization steps. Simi-
purge requirement, and higher equilibrium selectivity. PSA larly, the 3-bed 6-step cycle incorporated an unusual co-
cycle sequences that took advantage of both light and current light product pressurization step. They couldn’t go
heavy reflux steps were explored by Takamura et al.16 and beyond 63% CO2 purity, which was achieved using the 3-
Park et al.17 Takamura et al. studied a 4-bed 8-step VSA bed 6-step cycle. Ko et al.23 optimized a 2-bed 4-step PSA
process whereas Park et al. analyzed three different cycle process to minimize power consumption, and a 1-bed 4-
configurations for VSA processes. Although the pure CO2 step fractionated VPSA process to increase CO2 purity to

AIChE Journal July 2010 Vol. 56, No. 7 Published on behalf of the AIChE DOI 10.1002/aic 1815
 PSA Superstructure
With the development of optimization strategies for pro-
cess synthesis, it is natural to consider a superstructure
based approach to design PSA processes. Figure 1 shows
the 2-bed PSA superstructure developed in this work. It has
a co-current bed (CoB) and a counter-current bed (CnB)
that determine co-current and counter-current operating
steps in the cycle, respectively. We consider only two beds
to ensure that the direction of the flow, and thus, the super-
ficial velocity, remains co-current for CoB and counter-cur-
rent for CnB. This strategy avoids flow reversals in the
bed, as well as a possible network of bed connections with
embedded logical conditions to realize different operating
steps. This superstructure is consistent with the concept of
unibed models,35 where no more than two beds interact at
the same time, and the steps can be grouped into adsorbing
steps and desorbing steps. Consequently, it can accomplish
a wide variety of these operating steps with just a single
Figure 1. A 2-bed PSA superstructure. bed connection, as shown in Figure 1. Furthermore, this
helps to avoid discrete variables and uses only continuous
variables for the optimization problem.
90% and recovery to 94%. Grande et al.24 studied a classi-
The superstructure is designed to get the light product
cal Skarstrom cycle with light product pressurization and a
from the upper end (light end) of CoB and heavy product
3-bed 5-step process which included a pure CO2 rinse step
after the adsorption step. Their scale-up study showed that from the lower end (heavy end) of CnB. The time depend-
a purity of 83% and a recovery of 66% is possible with ent variables b(t) and a(t) determine the fraction of the
the 3-bed 5-step process at a much higher feed throughput light product and the heavy product streams that go into
of 48.57 kg-mol h
1. the top and the bottom reflux, respectively. The flue gas,
Webley and co-workers25–27,30 have done an extensive coming at atmospheric pressure, is compressed to Pfeed to
research in the field of CO2 separation by adsorption. Chaffee increase the feed throughput. The time dependent feed frac-
et al.25 and Zhang et al.26 studied two different VSA processes. tion /(t) determines the feeding strategy. For CoB, pressure
For a low feed throughput of 0.193 kg-mol h
1 for both the is specified at the light end by Pads, whereas the pressure
cycles, they achieved a low power consumption of 192 kWh at the other end Pa is determined from the pressure drop in
tonne
1 CO2 captured for the 3-bed 6-step and 240 kWh this bed. The velocity va, concentration for ith component
tonne
1 CO2 captured for the 3-bed 9-step cycle. Xiao et al.27
Ca,i, and temperature Ta at the light end are determined
studied a similar 3-bed 9-step cycle and were able to increase
from the outlet flux. Similarly, for CnB, pressure is speci-
CO2 recovery to 75%. In another study, Zhang and Webley30
compared numerous VSA cycle configurations, and showed fied at the heavy end by Pdes, while Cd,i, Td and vd are
that CO2 purity can be increased by particularly incorporating obtained from the output flux, and Pd is obtained from the
heavy reflux and equalization steps. pressure drop. The superstructure also incorporates com-
Although this review offers some trends and guidelines, pressors and valves to account for different pressure levels
a fully systematic methodology is still required to design in the beds, and a vacuum generator to extract the
PSA cycle configurations. Zhang and Webley30 outlined an strongly-adsorbed component.
approach for cycle development by understanding the roles It is possible to accomplish a wide variety of different
of individual operating steps and adsorption fronts. How- operating steps of a PSA process by varying the control vari-
ever, they identified optimal configurations with the help of ables a(t), b(t), /(t), Pads(t), and Pdes(t), as shown in Figure
a pre-decided set of operating steps and a simplified mathe- 2. As a consequence, the temporal profiles of a(t), b(t), /(t),
matical model. Chiang31 and Smith et al.32 described heu- Pads(t) and Pdes(t) result into a sequence of operating steps,
ristic-based and mixed integer nonlinear programming thus generating a PSA cycle. For instance, the profiles of
(MINLP) based formulations, respectively, to determine a(t), b(t), and /(t), shown in Figure 3, translate into the clas-
minimum number of beds required to design a PSA process sical 2-bed 4-step Skarstrom cycle (FP,F,CnD,LR).36 CoB
around a given sequence of operating steps, but didn’t dis- generates pressurization (FP) and feed (F) steps, while CnB
cuss how the steps should be chosen. Smith and Wester- simultaneously generates depressurization (CnD) and light
berg33,34 also gave a 3-step scenario to design an industrial reflux (LR) steps. Thus, the overall cycle includes these four
PSA system, but again with a known sequence of operating steps (FP,F,CnD,LR). In an actual 2-bed PSA unit, after per-
steps. In contrast, this work presents a novel superstructure- forming its steps, CoB will follow the steps of CnB and
based approach to obtain optimal sequence of operating vice-versa. However, in the mathematical framework, this is
steps in a PSA cycle without any assumption on the kinds realized by giving final conditions of CoB as the initial con-
of steps that should be included within the cycle. As dis- ditions for CnB and vice-versa, thus modeling the true 2-bed
cussed in the next section, this approach relies on the for- behavior. This multibed approach is described in detail in
mulation of an optimal control problem. Jiang et al.37 An optimal sequence of operating steps, along

1816 DOI 10.1002/aic Published on behalf of the AIChE July 2010 Vol. 56, No. 7 AIChE Journal
 Figure 2. Realization of different operating steps by varying control variables.

with other decision variables, such as cycle time, individual (1) All of the gases follow the ideal gas law.
step times, and bed dimensions, is obtained by solving the (2) There are no radial variations in temperature, pressure,
following optimization problem. and concentrations of the gases in the solid and the gas
phase.
min (3) The gas and the solid phases are in thermal equilib-
Uðzðx; tf Þ; yðx; tf Þ; aðtf Þ; bðtf Þ; /ðtf Þ; zðx; 0Þ; pÞ rium and bulk density of the solid phase remains constant.
(4) Pressure drop along the bed is calculated by the Ergun
s:t: equation.


@z @z (5) The adsorption behaviors are described by the dual-
f ; ; zðx;tÞ;yðx;tÞ; aðtÞ; bðtÞ; /ðtÞ; Pads ðtÞ; Pdes ðtÞ; zðx; 0Þ; p
@t @x site Langmuir isotherm. This monolayer-based isotherm has
¼0 the ability to accurately capture linear behavior at low pres-
sures, a requirement for thermodynamic consistency in any
zCoB ðx; 0Þ ¼ zCnB ðx; tf Þ; zCnB ðx; 0Þ ¼ zCoB ðx; tf Þ physical adsorption system.38
sðzðx; tÞ; yðx; tÞ; aðtÞ; bðtÞ; /ðtÞ; pÞ ¼ 0 (6) The adsorption rate is approximated by the linear driv-
gðzðx; tÞ; yðx; tÞ; aðtÞ; bðtÞ; /ðtÞ; pÞ  0 ing force (LDF) expression. Sircar and Hufton39 demon-
strated that the LDF model is sufficient to capture the
0  aðtÞ; bðtÞ; /ðtÞ  1 kinetics of adsorption because the estimation of the separa-
bL  ðPads ; Pdes ; pÞ  bU ð1Þ tion performance of an adsorptive process requires several
sets of averaging of kinetic properties and the effect of local
Here, U is the objective function related to overall power characteristics are lumped during integration.
consumption, component purity or recovery. It can depend On the basis of the above assumptions, the mathematical
upon differential variables z(x, t), algebraic variables y(x, t), model for the PSA process is listed in Table 2. Here, we
control variables a(t), b(t), and /(t), initial conditions z(x, 0) consider a lumped mass transfer coefficient for the LDF
and other decision variables p. The first equation represents equation. Since a smaller magnitude of UA makes energy
the PDAE-based model for the PSA system, while the second balance a weak function of the ambient temperature, Tw is
equation is the cyclic steady state (CSS) condition (see assumed constant. As a convention, flow in the counter-
Table 2). As mentioned before, the CSS condition is
implemented by giving final conditions of CoB as the initial
condition for the CnB and vice versa. Additional constraints
for the optimization problem are given by the algebraic
equations s and the inequalities g. The control variables a(t),
b(t), and /(t) are fractions bounded between 0 and 1. Other
control variables, Pads(t) and Pdes(t), and decision variables p
are bounded between their respective bounds bL and bU.
It is important to note that although optimal 2-bed PSA
configurations are construed from the optimal profiles of
a(t), b(t), /(t), Pads(t), and Pdes(t), multibed cycles (with
more than two beds) follow immediately from these solu-
tions. These are generated by staggering the steps over mul-
tiple beds and ensuring that a bed with a product flow step
occurs at all points in time.

PSA Model and Solution Methodology

Model equations
We consider a detailed PDAE-based mathematical model
for the optimal control problem. The model is fairly general Figure 3. Realization of 4-step Skarstrom cycle from
and can also be extended beyond the following assumptions. the superstructure.

AIChE Journal July 2010 Vol. 56, No. 7 Published on behalf of the AIChE DOI 10.1002/aic 1817
 Table 2. PSA Model Equations

Component mass balance

@Ci @qi @ðvCi Þ

eb þ ð1
eb Þqs þ ¼0 i 2 fL, Hg (2)
@t @t @x
LDF equation

@qi
¼ ki ðqi
qi Þ i 2 fL, HgÞ (3)
@t
Energy balance
!
X @T X @qi @ðvhÞ
eb Ci ðCipg
RÞ þ qb Cps
qb DHiads þ þ UA ðT
Tw Þ ¼ 0 (4)
i
@t i
@t @x

Cipg ¼ aic þ bic T þ cic T 2 þ dci T 3 i 2 fL, Hg

X
Z 
h¼ Ci Cipg dT
i
Dual-site Langmuir Isotherm

qs b C RT qs b C RT
qi ¼ P1i 1i i þ P2i 2i i i 2 fL, Hg (5)
1þ b C
j2fL;Hg 1j j RT 1 þ j2fL;Hg b2j Cj RT

4
kmi
qsmi ¼ kmi
1
þ kmi
2
T bmi ¼ kmi
3
exp i 2 fL, Hg m ¼ 1; 2
T
Ergun equation

 X !
@P 150lð1
eb Þ2 1:75 1
eb

¼ v þ M i
C
w i vjvj (6)
@x dp2 e3b dp e3b i

Ideal gas equation

X
P ¼ RT Ci (7)
i

Bed Connection equation (See Figure 1)

Fi ðtÞ ¼ /ðtÞvfeed Cfeed;i þ aðtÞð
vd ðtÞÞCi;d ðtÞ i 2 fL, Hg (8)

TRi ðtÞ ¼ bðtÞva ðtÞCa;i ðtÞ i 2 fL, Hg (9)

LPi ðtÞ ¼ ð1
bðtÞÞva ðtÞCa;i ðtÞ i 2 fL, Hg (10)

BRi ðtÞ ¼ aðtÞð
vd ðtÞÞCd;i ðtÞ i 2 fL, Hg (11)

HPi ðtÞ ¼ ð1
aðtÞÞð
vd ðtÞÞCd;i ðtÞ i 2 fL, Hg (12)

Cyclic Steady State (CSS) condition for beds CoB and CnB

zCoB ðx; 0Þ ¼ zCnB ðx; tf Þ; zCnB ðx; 0Þ ¼ zCoB ðx; tf Þ;

(13)
zðx; tÞT ¼ ½CL ðx; tÞ; CH ðx; tÞ; qL ðx; tÞ; qH ðx; tÞ; Tðx; tÞ

current bed is considered negative and a minus sign is used the specified feed flux is achieved and the model assump-
for vd. Since the bed model is based on fluxes, the feed tions are not violated. The following equations are used to
throughput and the bed diameter can be adjusted as long as evaluate the performance variables.

1818 DOI 10.1002/aic Published on behalf of the AIChE July 2010 Vol. 56, No. 7 AIChE Journal
 R
ð1
bðtÞÞva ðtÞCa;L ðtÞdt optimization problem. The model equations are solved only
purityL ¼ R P (14a) once at the optimum and the excessive computational effort
ð1
bðtÞÞva ðtÞ i Ca;i ðtÞdt
R of getting intermediate solutions is avoided.41 However, the
ð1
aðtÞÞð
vd ðtÞÞCd;H ðtÞdt performance of this approach substantially depends upon the
purityH ¼ R P (14b) optimization solver, and therefore, it is crucial to choose an
ð1
aðtÞÞð
vd ðtÞÞ i Cd;i ðtÞdt
efficient NLP solver. Hence, we use the state-of-the-art NLP
R
ð1
bðtÞÞva ðtÞCa;L ðtÞdt solver IPOPT 3.4 for our case studies. This interior point
recoveryL ¼ (14c) solver uses a barrier method to handle inequalities and exact
Qfeed;L
second derivative information for faster convergence to the
R
ð1
aðtÞÞð
vd ðtÞÞCd;H ðtÞdt optimum.42
recoveryH ¼ (14d) Hyperbolic PDAEs in Table 2 are characterized by steep
Qfeed;H
adsorption fronts in spatial domain for concentrations and
Z
temperature. To capture such steep profiles, it is important to
Qfeed;i ¼ /ðtÞvfeed Cfeed;i dt i 2 fL, Hg (14e) use a spatial discretization method which not only avoids
physically unrealistic oscillations, but also is suited to model
Here, Qfeed is the feed flux. The total power consumption, conservative properties.43,44 Therefore, a first-order finite
given by the following equations, is the sum of the work done volume method is applied for spatial discretization. The dif-
by the compressors and the vacuum generator. ferential algebraic equations (DAEs), resulting from spatial
discretization, are then converted into a set of algebraic
Wtotal equations by applying orthogonal collocation on finite ele-
Z "P  
c
1 ! ments in temporal domain. A Radau collocation scheme is
cRTd i /ðtÞvfeed Cfeed;i þaðtÞð
vd ÞCd;i
c
Pa
¼
1 considered because it allows us to set constraints at the ends
c
1 gc Pfeed of the finite elements.45 A 3-point collocation scheme is
P (
c
1
c
1 ) ! used for state variables while control variables are consid-
aðtÞð
v d ÞC d;i P feed
c
P feed
c
þ i min ;
1 ered to be piecewise constant. We also consider a moving
gh Patm Pdes finite element strategy in which the length of each temporal
P (
c
1 !)# finite element is considered a decision variable in the optimi-
i ð
vd ÞCd;i Patm c
þ max 0;
1 zation problem. With moving finite elements, it is possible
gv Pdes to locate optimal breakpoints of the control variables with

c
1 ! variable element lengths. Appropriate bounds are imposed
c /ðtÞvfeed Pfeed Pfeed c on the variable element lengths of each finite element to
þ
1 dt ð15aÞ
c
1 gfg Patm guarantee the accuracy of the discretization.
Since a large number of spatially discretized nodes are
Wtotal
Power ¼ R (15b) required to capture steep adsorption fronts, such fine spatial
ð1
aðtÞÞvd ðtÞCd;H ðtÞdt discretization leads to a large set of DAEs which after tem-
poral discretization leads to a very large set of algebraic
Here, the max function ensures that the work done by equations, that becomes expensive to solve. Moreover, in the
the vacuum generator is zero when Pdes is greater than absence of any error checking mechanism for temporal inte-
atmospheric pressure Patm. Similarly, since the vacuum gration, pre-determined temporal discretization may lead to
generator discharges heavy reflux at Patm, the min operator inaccurate profiles due to insufficient number of finite ele-
ensures a proper upstream pressure for the heavy product ments. Although a large number of elements will improve
compressor. Since min or max operators introduce non- the accuracy, it will again make the problem computationally
differentiability, the following smoothing approximations are challenging to solve. Hence, to get the solution in a reasona-
adopted.40 A value of 0.01 is used for e in the following ble amount of time, we consider only 20 spatial finite vol-
equations. umes and around 25 temporal finite elements for the optimi-
zation problem. Because such a small number of nodes can
minðf1 ðxÞ; f2 ðxÞÞ ¼ f1 ðxÞ
maxð0; f1 ðxÞ
f2 ðxÞÞ (16a) cause inaccuracies to creep in the NLP solution obtained

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi from IPOPT, verification of this solution with an accurate
dynamic simulation is essential. Thus, we carry out dynamic
maxð0; f ðxÞÞ ¼ 0:5 f ðxÞ þ f ðxÞ2 þ e2 (16b)
simulations in MATLAB46 at the optimal values obtained
from IPOPT. The DAE system obtained after applying
method of lines is integrated in MATLAB at the optimal
Solution strategy values of the decision variables. The profiles and perform-
We adopt a complete discretization approach to solve the ance variables obtained from MATLAB are then compared
system of PDAEs in Table 2. The PDAEs are converted into with those obtained from IPOPT.
a set of algebraic equations by discretizing the state and the Finally, because all of the control variables appear line-
control variables both in space and time. As a result, the arly, problem (1) is a singular optimal control problem. Such
PDAE-constrained optimal control problem (Eq. 1) gets con- problems are characterized by shallow response surfaces.
verted into a large-scale nonlinear programming (NLP) prob- Moreover, the Euler-Lagrange equations obtained after
lem. One of the advantages of this approach is that it applying the maximum principle to (1) are high index in na-
directly couples the solution of the PDAE system with the ture and ill-conditioned. Consequently, a fine discretization

AIChE Journal July 2010 Vol. 56, No. 7 Published on behalf of the AIChE DOI 10.1002/aic 1819
of the control variable profiles also leads to an ill-condi- Table 3. Zeolite 13X Properties and Model Parameters23
tioned problem with solutions often characterized by oscilla- Parameter Value
tions that do not subside with increasing mesh refinement.47
Bulk porosity (eb) 0.34
As a result, we adopt a relatively coarse discretization of the Particle diameter (dp) 0.002 m
control profiles. Although this will lead to suboptimal (but Adsorbent density (qs) 1870 kg m
3
generally quite good) solutions, it also helps to regularize Bulk density (qb) 1234.2 kg m
3
the solution profiles and ameliorate the singular nature of the Heat capacity of solid (Cps) 450.54 J kg
1 K
1
Heat transfer coefficient (UA) 926.7 J m
3 s
1 K
1
optimal control problem. Gas viscosity (l) 1.7857  10
5 kg m
1 s
1
Gas constant (R) 8.314 J g-mol
1 K
1
Mass transfer coefficient (k) CO2 ¼ 0.1631 s
1
Case Studies and Computational Results N2 ¼ 0.2044 s
1
Heat of adsorption (DHads) CO2 ¼ 23011.14 J g-mol
1
In this section, we consider a post-combustion CO2 sep- N2 ¼ 14452.72 J g-mol
1
aration system that illustrates the superstructure methodol- Ambient temperature (Tw) 298 K
ogy to obtain optimal PSA cycles. We consider an 85%–
15% N2 -CO2 feed mixture which is a typical composition Isotherm parameters CO2 N2
of a post-combustion flue gas stream. As an initial study, k11 2.817269 1.889581
the focus is on a binary feed mixture. A multicomponent k21
3.51  10
4
2.25  10
4
feed mixture also having water, oxygen and other trace k31 2.83  10
9 1.16  10
9
components will be considered in the future extensions of k41 2598.203 1944.606
k12 3.970888 1.889581
this work. We assume that the flue gas enters at a temper- k22
4.95  10
3
2.25  10
4
ature of 310 K, and a maximum velocity (vfeed) of 50 cm k32 4.41  10
9 1.16  10
9
s
1. The inlet pressure Pfeed varies with the case studies. k42 3594.071 1944.606
Zeolite 13X is chosen as the adsorbent to separate CO2;
Chue et al.15 suggested it to be a preferable adsorbent
over others for this separation system. The adsorbent prop-
erties for 13X and other model parameters are listed in max
Table 3.23 We consider three different cases to explore
CO2 recovery ðfrom Eq. 14dÞ
different facets of the superstructure approach. The first
case study optimizes the 2-bed 4-step Skarstrom configura- s:t:
tion, obtained after fixing the control variables in the cðwÞ ¼ 0 ðfully discretized Eqs. 2--13Þ ð17aÞ
superstructure, and shows the ineffectiveness of such tradi-
tional cycles for high-purity CO2 separation. The second CO2 purity  0:4 ðfrom Eq. 14bÞ (17b)
case then finds an optimal PSA configuration which sepa-
rates CO2 at a high purity and recovery. Finally, in the Pads  Pd (17c)
third case, we find an optimal configuration which Pdes  Pfeed (17d)
achieves high-purity separation with minimal power
requirements. Pa  Pfeed (17e)
1 m  BLen  6 m (17f)
Case 1: Optimization with a conventional configuration
First, we explore the potential of the conventional 2-bed 20 s  Tc  2400 s (17g)
4-step Skarstrom cycle for post-combustion CO2 capture. For
101:32 kPa  Pads  1000 kPa (17h)
this, we fix the profiles of a(t), b(t), and /(t) over time, as
shown in the Figure 3. While fb is chosen as 0.3 , f/ before Pdes  10 kPa (17i)
tswitch is fixed to 0.35. This ensures that the superficial veloc-
ity is close to zero towards the light end of CoB during the Here w and c(w) ¼ 0 represent the set of completely
FP step. In this case, Pads and Pdes remain constant for the discretized variables and model equations, respectively.
entire cycle. The inlet pressure Pfeed is fixed to 300 kPa, as Constraint (17c) ensures that the pressure always drops around
considered by Gomes et al.18 as well. the pressure reducing valve in the superstructure. Similarly,
With this configuration we maximize CO2 recovery. Since constraints (17d) and (17e) ensure that the gas is never
the lack of any heavy reflux step in the configuration may expanded by the heavy gas and the feed compressors,
not enable a high purity separation, a relatively low value of respectively. The rest of the inequalities are bounds on the
40% is chosen for the lower bound on CO2 purity. Besides decision variables.
Pads and Pdes, we also consider bed length BLen, and cycle With 24 temporal finite elements and 20 spatial finite vol-
time Tc as decision variables. Since a moving finite element umes, we solved the NLP in AMPL48 using IPOPT. Table 4
strategy is adopted, the length of each finite element is also includes a summary of the optimization results. With 35,000
considered as an optimization variable. Because none of the variables and 29 degrees of freedom, we were able to solve
decision variables are functions of time, the optimal control it to optimality in around 3 CPU h. Optimal moving finite
problem (1) becomes a dynamic optimization problem, element lengths and cycle time of 2140 s yield an optimal
which becomes the following NLP after discretizing PDAEs step time of 760 for pressurization (and depressurization)
in both space and time. step, and 410 for feed (and light reflux) step. Such a long

1820 DOI 10.1002/aic Published on behalf of the AIChE July 2010 Vol. 56, No. 7 AIChE Journal
 Table 4. Optimization Results for Case 1 The control variables a(t), b(t), and /(t) are freed to let them
No. of variables 35022 achieve an optimal sequence of operating steps. The pres-
No. of equations 34993 sures Pads and Pdes are converted back to time dependent con-
CPU time 176.94 min trol variables. To keep this case comparable to the previous
Optimal step times one, we fix the bed length to 5 m. A desired CO2 purity of at
Step 1 (and 3) 760 s
Step 2 (and 4) 410 s least 95% is chosen. Besides this, we impose a lower bound
Optimal parameters BLen ¼ 5.51 m Tc ¼ 2140 s on feed flux Qfeed, in the absence of which the optimizer may
Pads ¼ 276.43 kPa Pdes ¼ 21.75 kPa force the feed fraction /(t) to zero to maximize CO2 recov-
ery. Finally, we add the equation for power consumption
Accuracy check Full discretization MATLAB verification (15b) to the optimal control problem. A 72% efficiency is
N2 purity 91.25% 90.74% assumed for all the compressors and the vacuum generator
N2 recovery 85.88% 85.94% unit. As in the previous case, we fix the flue gas inlet pres-
CO2 purity 40% 38.65% sure Pfeed to 300 kPa to achieve a reasonable Qfeed. The fol-
CO2 recovery 53.36% 50.22%
lowing large-scale NLP results after complete discretization
of state and control variables in the optimal control problem.

pressurization step is due to a small amount of feed during max CO2 recovery ðfrom Eq. 14dÞ
that step, which requires longer time for bed to get pressur- s:t: cðwÞ ¼ 0 ðfully discretized Eqs. 2--13Þ
ized and CO2 to adsorb. At the optimum, the cycle handles a CO2 purity  0:95 ðfrom Eq. 14bÞ
feed flux of 96.4 kg-mol m
2 h
1, which is higher than the Qfeed;L þ Qfeed;H  80 kg-mol m
2 h
1
corresponding 2-bed 4-step case studies in Table 1. At a pu- Pads  Pd
rity of 40%, a maximum CO2 recovery of only 53.4% was Pdes  Pfeed (18)
achieved. Such poor performance proves the point made in Pa  Pfeed
the introduction; classical cycles without heavy reflux cannot 0  aðti Þ; bðti Þ; /ðti Þ  1 8ti
produce heavy product at high purity since a light reflux step 20 s  Tc  2400 s
dilutes the heavy product and decreases its purity. Table 4 101:32 kPa  Pads ðti Þ  600 kPa 8ti
also lists the MATLAB verification of the AMPL results. Pdes ðti Þ  50 kPa 8ti
We considered 20 spatial finite volumes for MATLAB as
well. A comparison of the purities and the recoveries indi-
cates reasonable accuracy for the complete discretization As in the previous case study, c(w) is the set of com-
approach. pletely discretized PDAEs with CSS condition. We choose a
lower bound of 50 kPa for the vacuum generated, which is
not a substantially high vacuum. Similarly, the chosen upper
Case 2: Cycle synthesis to maximize CO2 recovery bound of 600 kPa for Pads is also not substantially high. No
Since a high-purity CO2 separation wasn’t achieved by the bounds are specified for the purity and the recovery of nitro-
Skarstrom cycle, in this case we solve the optimal control gen. We impose a lower bound of 80 kg-mol m
2 h
1 on
problem (1) to obtain an optimal configuration which yields the total feed flux. Because the bound is not on the feed
better performance. For this, a few modifications are essential throughput, a bigger diameter PSA bed will be able to han-
to the optimization problem presented in the previous case. dle much higher feed throughput and the optimal

Figure 4. Optimal control profiles for Case 2.

[Color figure can be viewed in the online issue, which is available at www.interscience.wiley.com.]

AIChE Journal July 2010 Vol. 56, No. 7 Published on behalf of the AIChE DOI 10.1002/aic 1821
 configuration need not change. For instance, for a 3 m bed
diameter, one PSA column will be able to handle a signifi-
cantly high feed throughput of 565 kg-mol h
1 for the same
optimal configuration. Also note that the value of 80 kg-mol
m
2 h
1 is significantly higher than feed flux chosen in the
literature studies in Table 1, as the focus here is to synthe-
size industrial scale PSA systems.
The NLP was solved in AMPL with 26 temporal finite
elements and 20 spatial finite volumes. The optimal control
profiles are shown in Figure 4. The profiles are drawn
against cycle time normalized between 0 and 1. These pro-
files suggest an optimal 2-bed 6-step VSA process, illus-
trated in Figure 5, which can be deciphered in the following
Figure 5. Optimal VSA configuration for Case 2.
manner. The cycle starts with a(t) ¼ 1, b(t) ¼ 0, and /(t) ¼

Figure 6. Gas-phase CO2 concentration profiles for Case 2.

[Color figure can be viewed in the online issue, which is available at www.interscience.wiley.com.]

1822 DOI 10.1002/aic Published on behalf of the AIChE July 2010 Vol. 56, No. 7 AIChE Journal
 Table 5. Optimization Results for Case 2

No. of variables 50162

No. of equations 49956
CPU time 756.22 min.
Optimal step times
Step 1 (and 4) 50 s
Step 2 (and 5) 405 s
Step 3 (and 6) 685 s
Optimal cycle time 2280 s
Feed flux 80 kg-mol m
2 h
1
Power consumption 637.25 kWh tonne
1 CO2
CO2 purity 95%
CO2 recovery 80.09%

1. This suggests a bottom reflux from CnB to CoB and feed

being fed to CoB. From the profiles of Pads(t) and Pdes(t),
CoB operates at around 450 kPa while CnB operates at Figure 7. Purity-recovery trade-off curve for Case 2.
around 85 kPa during this step. Thus, we have a low-pres-
sure adsorption step with a heavy reflux for CoB (step 1)
and a low-vacuum desorption step for CnB (step 4) with de- CoB purges nitrogen out of CoB and enriches itself towards
sorbed CO2 being sent as a heavy reflux from CnB to CoB. the heavy end of CoB and pushes its front towards the light
This heavy reflux joins the feed low in CO2 concentration end of CoB. Step 3 in Figure 6 confirms the movement of
and enriches the adsorbed CO2 concentration towards the this CO2 front. A decrease in Pads(t) and an increase in Pdes(t)
heavy end of CoB. This is evident from the gas-phase CO2 towards the end of this step halts this recirculation. After the
concentration profile for step 1 in Figure 6. After this step, total reflux step, the co-current bed follows the steps of
both top and bottom reflux disappear, while /(t) indicates the counter-current bed and vice-versa. This completes the
continuation of the feed to CoB. CoB gets pressurized to the cycle.
upper bound of 600 kPa and N2 is withdrawn at a high pres- The optimization results for this case are summarized in
sure, while CO2 is extracted from CnB at a vacuum of 50 Table 5. With 50,162 variables and 206 degrees of freedom,
kPa. This suggests a pressurization and high pressure adsorp- it was solved to optimality in 12.5 CPU hours on an Intel
tion step for CoB (step 2) and a high vacuum desorption Quad core 2.4 GHz system with 8 GB RAM. At the optimum,
step for CnB (step 5). The feed fraction /(t) drops at the be- the feed flux attained its lower bound of 80 kg-mol m
2 h
1.
ginning of this step to facilitate CoB pressurization. We For such a high feed flux, we obtained a reasonable power
observe a drop in the CO2 concentration in CoB (see step 2 consumption of 637.25 kWh tonne
1 CO2 captured. Also, an
in Figure 6) because of its low concentration in feed. Also, optimum CO2 recovery of 80% at a purity of 95%, for such a
because of the application of vacuum, the gas-phase CO2 high feed flux, is substantially better than the literature stud-
concentration decreases sharply for both step 4 and 5, as is ies for post-combustion capture that deal with high feed
evident Figure 6. throughput. These results confirm our assertion that steps like
Further, the pressures in the beds are held at their same re- heavy reflux are essential for high-purity CO2 separation.
spective levels, while a(t) becomes 1, b(t) approaches 1, and In Table 6, we provide a validation of the optimal results
the feed is stopped completely. Because there is no feed nor obtained from AMPL with method of lines simulations in
product at this time, we have a total reflux step (step 3 and MATLAB for varying numbers of spatial finite volumes. The
6), in which both the beds are connected to each other and a results from AMPL are in good agreement with those from
recirculation of the components occurs within the system. A MATLAB, and the accuracy doesn’t suffer as we consider a
small amount of N2 is withdrawn at the beginning of this large number of finite volumes in MATLAB. This indicates that
step, and is shown as a dotted line in Figure 5. During this the complete discretization approach is reasonably accurate.
step, nitrogen from CoB to CnB purges CO2 out of CnB from Figure 7 shows a trade-off curve between CO2 purity and
its heavy end and enriches itself towards the light end of recovery. We construct this curve by varying the lower
CnB while pushing its front. Similarly, CO2 from CnB to bound on CO2 purity in the NLP (18), and then maximizing

Table 6. Accuracy Validation for AMPL Results of Case 2

MATLAB Verification
AMPL Results
Spatial finite volumes 20 20 40 80
N2 purity 96.58% 96.28% 96.19% 96.16%
N2 recovery 99.26% 99.26% 99.28% 99.29%
CO2 purity 95% 94.95% 95.04% 95.07%
CO2 recovery 80.09% 78.26% 77.70% 77.51%

AIChE Journal July 2010 Vol. 56, No. 7 Published on behalf of the AIChE DOI 10.1002/aic 1823
Figure 8. Optimal control profiles for Case 3.
[Color figure can be viewed in the online issue, which is available at www.interscience.wiley.com.]

CO2 recovery for each lower bound. As a result, it is possi- generator. The rest of the optimization problem remains
ble that a different optimal cycle configuration is achieved at same, and is as below.
each point plotted on the curve. However, each configuration
is the best possible cycle for a particular CO2 purity. Conse- min
quently, this yields an optimal purity-recovery trade-off Power (from Eq. 15bÞ
curve. The curve shows that if a very high purity CO2 sepa- s:t:
ration is desired then the recovery falls drastically. A similar cðwÞ ¼ 0 (fully discretized Eqs. 2–13)
trend is observed with the purity when a very high CO2 re- CO2 purity  0:9 (from Eq. 14b)
covery is sought. The intermediate section of the curve is a CO2 recovery  0:85 (from Eq. 14d)
preferable region to operate. Qfeed;L þ Qfeed;H  65 kg-mol m
2 h
1
Pads  Pd (19)
Case 3: Cycle synthesis to minimize power consumption Pdes  Pfeed
Although we achieved a high purity separation in the pre- Pa  Pfeed
vious case, the power consumption was also quite high. 0  aðti Þ; bðti Þ; /ðti Þ  1 8ti
Therefore, the objective of this case is to obtain an optimal 20 s  Tc  2400 s
configuration which yields a high-purity separation at mini- 101:32 kPa  Pfeed  600 kPa
mal power requirements. To achieve this, a few small modi- 101:32 kPa  Pads ðti Þ  600 kPa 8ti
fications are done to the NLP (18). Although the lower Pdes ðti Þ  50 kPa 8ti
bound on CO2 recovery is relaxed to 85%, the lower bounds
on CO2 purity and feed flux are relaxed to 90% and 65 kg-
mol m
2 h
1, respectively. To minimize the work done in The NLP was solved in AMPL with 24 temporal finite
compressing flue gas from Patm to Pfeed (in Eq. 15a), we elements and 20 spatial finite volumes. The optimal control
consider Pfeed a decision variable instead of fixing it to 300 profiles are shown in Figure 8. Although the profiles for the
kPa. Appropriate bounds are imposed on Pfeed. The effi- top and the bottom reflux are quite similar to the previous
ciency is kept same as 72% for all compressors and vacuum case, we observe a stark difference in the profiles of the
operating pressures and the feeding strategy. As a result, we
obtain an entirely different 2-bed 8-step VSA configuration,
illustrated in Figure 9.
The cycle begins with a(t) ¼ 1, b(t) ¼ 0 and /(t) close to
one. This suggests a heavy reflux from CnB to CoB and
feed being fed to CoB. From the profiles of Pads(t) and
Pdes(t), the pressure rises in CoB and falls in CnB during
this step. Thus, we have a pressurization step for CoB (step
1) and a depressurization step for CnB (step 5). The heavy
reflux increases the adsorbed-phase CO2 concentration
towards the heavy end of CoB, which is evident from the
gas-phase CO2 concentration profile for step 1 in Figure 10.
Next, both a(t) and b(t) go to zero, while Pads(t) and Pdes(t)
Figure 9. Optimal VSA configuration for Case 3. attain their maximum and minimum allowed values,

1824 DOI 10.1002/aic Published on behalf of the AIChE July 2010 Vol. 56, No. 7 AIChE Journal
Figure 10. Gas-phase CO2 concentration profiles for Case 3.
[Color figure can be viewed in the online issue, which is available at www.interscience.wiley.com.]

respectively. This suggests an adsorption step with the re- CO2 is collected. After this step, we observe that both a(t)
moval of light product for CoB (step 2), and a high vacuum and b(t) go to 1. However, unlike the total reflux step in pre-
desorption step for CnB (step 6), during which high purity vious case study, b(t) doesn’t go to 1 at once and nitrogen is

AIChE Journal July 2010 Vol. 56, No. 7 Published on behalf of the AIChE DOI 10.1002/aic 1825
 Table 7. Optimization Results for Case 3 Figure 11 shows the trade-off curve between power con-
sumption and CO2 recovery. As in the previous case, we
No. of variables 46313
construct the curve by varying the lower bound on CO2 re-
No. of equations 46122
CPU time 273.08 min covery, while keeping purity at 90%, and optimizing NLP
Optimal step times (19) multiple times. Thus, we obtain an optimal trade-off
Step 1 (and 5) 56.77 s curve, although it is possible to obtain a different optimal
Step 2 (and 6) 500.03 s cycle configuration at each point plotted on the curve. As
Step 3 (and 7) 614.79 s
Step 4 (and 8) 28.41 s expected, the curve shows that the power requirements
Optimal cycle time 2400 s increase if a high-recovery separation is desired. However,
Optimal Pfeed 182.31 kPa the increase is almost linear up to a recovery level of 84%.
Feed flux 65 kg-mol m
2 h
1 The power requirements then start growing exponentially if
Power consumption 464.76 kWh tonne
1 CO2
more than 84% CO2 recovery is sought.
CO2 purity 90%
CO2 recovery 85%
Conclusions
still constantly removed from the system. Feed is also fed to A fairly extensive review of the previous work on post-
CoB for a considerable amount of time at the beginning of combustion CO2 capture reveals that a systematic methodol-
this step. Therefore, this translates into a heavy reflux step ogy is still required for the design of PSA cycles. To address
for CoB (step 3) and a light reflux step for CnB (step 7). this, we introduce an optimal control based framework to de-
Nevertheless, the intent of this step is similar to that of the velop optimal PSA cycle configurations. A novel 2-bed PSA
total reflux step: enrich the N2 front towards the light end of superstructure is presented which is able to predict a number
CnB, and CO2 front towards the heavy end of CoB. The gas of different PSA operating steps. The bed connections in the
phase CO2 concentration profiles for both these steps in Fig- superstructure are governed by time-dependent control varia-
ure 10 validate this behavior. bles, such as the fractions of light and heavy product recycle
In the previous case, the total reflux step occurred with between the beds, or the amount of feed to the beds. Differ-
Pads(t) and Pdes(t) operating at their upper and lower bounds, ent operating steps are accomplished by the superstructure
respectively. In contrast, CoB in this case operates at a lower by varying these control variables. To achieve an optimal
Pads(t) instead, which leads to savings in power consumption. sequence of operating steps, an optimal control problem,
Another power-saving aspect of this 2-bed 8-step VSA cycle with a PDAE-based model of the PSA system and the cyclic
is the pressure equalization step. After the third step, a(t) and steady state condition, is formulated.
b(t) remain at 1, while Pads(t) starts to drop, Pdes(t) starts to The superstructure approach is illustrated for three case
jump sharply, and the two pressures come very close to each studies of post-combustion CO2 capture. The first case study
other. In fact, Pads(t) and Pd(t) are approximately equal during optimizes the standard 2-bed 4-step Skarstrom cycle, and shows
this step. This translates into a short pressure equalization step that such conventional cycles, which focus on separating light
(step 4 and 8). Since the heavy reflux from CnB is negligible, product at a high purity, fail to produce heavy product at a high
we show it as a dotted line for this step in Figure 9. Clearly, purity because of the absence of a heavy reflux step. To obtain
from Figure 10, CO2 concentration drops substantially in CoB high-purity separation, the superstructure is optimized in the
and rises steadily in CnB during the equalization step. After second case study. A 2-bed 6-step VSA cycle is derived from
this, CoB follows the steps of CnB and vice-versa. the solution of the optimal control problem. With this configu-
Table 7 summarizes the optimization results. With 46,313 ration, we are able to recover about 80% of CO2 at a substan-
variables and 191 degrees of freedom in the NLP, the opti- tially high purity of 95%, and at a significantly high feed flux of
mal solution was obtained in 4.5 CPU hours. At the opti- 80 kg-mol m
2 h
1, but with a power consumption of 637 kWh
mum, the feed flux, CO2 purity, and CO2 recovery were at tonne
1 CO2 captured. Thus, in the third case study, we focus
their respective lower bounds of 65 kg-mol m
2 h
1, 90% on developing optimal configuration which yields high-purity
and 85%. Under these conditions, and at an optimum Pfeed separation with minimal power requirements. We construe a 2-
of 182.3 kPa, we achieved a power consumption of 464.76 bed 8-step VSA configuration from the optimal profiles, with
kWh tonne
1 CO2 captured, which is over 27% lower than which, at 90% purity and 85% recovery, CO2 is extracted with
Case 2. Table 8 lists a validation of the optimal results a substantially low power consumption of 465 k Wh tonne
1
obtained from AMPL with accurate simulations in MATLAB CO2 captured. Hence, with the proposed superstructure
for varying number of spatial finite volumes. As observed in approach, we are able to design optimal configurations that
the previous case, the purities and recoveries are in reasona- make pressure swing adsorption a promising option for high pu-
ble agreement, even for large number of finite volumes. rity CO2 capture from flue gas streams.

Table 8. Accuracy Validation for AMPL Results of Case 3

MATLAB Verification
Spatial Finite Volumes AMPL Results 20 20 40 80
N2 purity 97.38% 97.16% 97.03% 96.99%
N2 recovery 98.33% 98.36% 98.40% 98.42%
CO2 purity 90% 90.01% 90.16% 90.22%
CO2 recovery 85% 83.72% 82.93% 82.68%

1826 DOI 10.1002/aic Published on behalf of the AIChE July 2010 Vol. 56, No. 7 AIChE Journal
 Notation
BRi ¼ flux of ith component in the bottom reflux stream (g-mol m
2
s
1)
Ci ¼ gas-phase concentration of ith component (g-mol m
3)
Cipg ¼ heat capacity of ith component (J g-mol
1 K
1)
Cps ¼ heat capacity of the adsorbent (J kg
1 K
1)
dp ¼ particle diameter (m)
Fi ¼ input flux of ith component to the co-current bed (g-mol m
2
s
1)
h ¼ total gas-phase enthalpy (J m
3)
DHads
i ¼ isosteric heat of adsorption (J g-mol
1)
HPi ¼ flux of ith component in the heavy product stream (g-mol m
2
s
1)
ki ¼ lumped mass transfer coefficient for ith component (s
1)
LPi ¼ flux of ith component in the light product stream (g-mol m
2
s
1)
Miw ¼ molecular weight of ith component (kg g-mol
1)
P ¼ total bed pressure (kPa)
Pads ¼ pressure at the light end of the co-current bed (kPa)
Pdes ¼ pressure at the heavy end of the counter-current bed (kPa)
Figure 11. Power-recovery trade-off curve, at 90% CO2 qi ¼ solid-phase concentration of ith component (g-mol kg
1)
purity, for Case 3. qi ¼ equilibrium solid-phase concentration of ith component (g-
mol
1 kg
1)
[Color figure can be viewed in the online issue, which is T ¼ gas-phase temperature in the bed (K)
available at www.interscience.wiley.com.] Tw ¼ wall/ambient temperature (K)
TRi ¼ flux of ith component in the top reflux stream (g-mol m
2 s
1)
UA ¼ effective heat transfer coefficient (J m
3 s
1 K
1)
A complete discretization approach is used to solve the v ¼ gas superficial velocity (m s
1)
optimal control problem as a large-scale nonlinear program, Wtotal ¼ total work done by the compressors and the vacuum generator
using the nonlinear optimization solver IPOPT. Verifications (J m
2 s
1)
of the accuracy of the discretization scheme show this Greek letters
approach is reasonably accurate in capturing the dynamics of a ¼ fraction of the heavy product going as a bottom reflux
PSA systems governed by hyperbolic PDAEs and steep b ¼ fraction of the light product going as a top reflux
adsorption fronts, and can be used for PSA systems with ef- c ¼ heat capacity ratio (¼1.4)
ficient NLP solvers like IPOPT. In future, a sensitivity-based eb ¼ bulk void fraction
/ ¼ fraction of the feed going to the co-current bed
sequential approach, similar to Jiang et al.37 will also be l ¼ gas viscosity (kg m
1 s
1)
developed to solve the optimal control problem for the gc ¼ efficiency of the feed compressor
superstructure without a separate verification step. Instead, gfg ¼ efficiency of the flue gas compressor
the PDAEs for the PSA system will be decoupled from the gh ¼ efficiency of the heavy product compressor
optimization problem, and the partially discretized PDAEs, gv ¼ efficiency of the vacuum generator
qb ¼ bulk density (kg m
3)
together with the sensitivities of the state variables with qs ¼ adsorbent density (kg m
3)
respect to decision variables, will be integrated outside the
optimization problem using a sophisticated dynamic simula-
tor which is able to capture the state variable profiles with
high accuracy. The optimization problem will then be solved
for the decisions using these sensitivities. Literature Cited
Finally, our superstructure based methodology, demon- 1. IEA/WEO. World Energy Outlook 2006 tech. rep., Paris, France:
strated for CO2 capture in this work, is quite generic and International Energy Agency; 2006.
can be extended to many other PSA applications; no 2. Aaron D, Tsouris C. Separation of CO2 from flue gas: a review.
Separ Sci Technol. 2005;40:321–348.
assumptions are made on the adsorbent or feedstock, the 3. IPCC.Carbon Dioxide Capture and Storage tech. rep. Geneva, Swit-
operating steps that can be predicted, or details of the bed zerland: Intergovernmental Panel on Climate Change; 2005.
models. This makes the approach fairly general. Moreover, 4. Sircar S, Kratz WC. Simultaneous production of hydrogen and car-
the superstructure can also be used to evaluate different bon dioxide from steam reformer off-gas by pressure swing adsorp-
kinds of adsorbents for the same feedstock and process con- tion. Separ Sci and Technol. 1988;23:2397–2415.
5. Kumar R. Removal of water and carbon dioxide from atmospheric
ditions. Although the current superstructure involves only air. US Patent 4,711,645, 1987.
two beds, we also plan to extend the formulation to a super- 6. Tamura T. Absorption process for gas separation. US Patent
structure incorporating more beds, more complex flow pat- 3,797,201, 1974.
terns and more challenging multi-component mixtures. 7. Hirose M, Omori I, Oba M, Kawai T. Carbon dioxide separation
and recovery system. Jpn Patent 2005262001, 2005.
8. Ishibashi M, Ota H, Akutsu N, Umeda S, Tajika M, Izumi J, Yasu-
Acknowledgments take A, Kabata T, Kageyama Y. Technology for removing carbon
dioxide from power plant flue gas by the physical adsorption
This technical effort was performed in support of the National Energy method. Energ Convers Manage. 1996;37:929–933.
Technology Laboratory’s on-going research in Process and Dynamic 9. Ito K, Otake K, Itoi M. Carbon dioxide desorption method. Jpn Pat-
Systems Research under the RDS contract DE-AC26-04NT41817. ent 2004202393, 2004.

AIChE Journal July 2010 Vol. 56, No. 7 Published on behalf of the AIChE DOI 10.1002/aic 1827
10. Sasaki A, Matsumoto S, Fujitsuka M, Shinoki T, Tanaka T, Ohtsuki 28. Reynolds SP, Ebner AD, Ritter JA. New pressure swing adsorption
J. CO2 recovery in molten carbonate fuel cell system by pressure cycles for carbon dioxide sequestration. Adsorption. 2005; 11:531–536.
swing adsorption. IEEE Trans Energ Convers. 1993;8:26–32. 29. Reynolds SP, Ebner AD, Ritter JA. Carbon dioxide capture from
11. Yokoyama T. Japanese R&D on Large-Scale CO2 Capture in ECI flue gas by pressure swing adsorption at high temperature using a
Symposium Series on Separations Technology VI: New Perspectives K-promoted HTlc: effects of mass transfer on the process perform-
on Very Large-Scale Operations; RP3 2004. ance. Environ Prog. 2006;25:334–342.
12. Reynolds SP, Ebner AD, Ritter JA. Stripping PSA cycles for CO2 30. Zhang J, Webley PA. Cycle development and design for CO2 cap-
recovery from flue gas at high temperature using a hydrotalcite-like ture from Flue gas by vacuum swing adsorption. Environ Sci Tech-
adsorbent. Ind Eng Chem Res. 2006;45:4278–4294. nol. 2008;42:563–569.
13. Reynolds S, Mehrotra A, Ebner A, Ritter J. Heavy reflux PSA 31. Chiang AST. Arithmetic of PSA process scheduling. AIChE J.
cycles for CO2 recovery from flue gas: Part I. Performance Evalua- 1988;34:1910–1912.
tion. Adsorption. 2008;14:399–413. 32. Smith OJ, Westerberg AW. Mixed-integer programming for pressure
14. Kikkinides ES, Yang RT, Cho SH. Concentration and recovery of swing adsorption cycle scheduling. Chemi Eng Sci. 1990;45: 2833–
carbon dioxide from flue gas by pressure swing adsorption. Ind Eng 2842.
Chem Res. 1993;32:2714–2720. 33. Smith OJ, Westerberg AW. The optimal design of pressure swing
15. Chue KT, Kim JN, Yoo YJ, Cho SH, Yang RT. Comparison of adsorption systems. Chem Eng Sci. 1991;46:2967–2976.
activated carbon and zeolite 13X for CO2 recovery from flue gas 34. Smith OJ, Westerberg AW. The optimal design of pressure swing
by pressure swing adsorption. Ind Eng Chem Res. 1995;34:591– adsorption systems–II. Chem Eng Sci. 1992;47:4213–4217.
598. 35. Kumar R, Fox VG, Hartzog D, et al. A versatile process simu-
16. Takamura Y, Narita S, Aoki J, Hironaka S, Uchida S. Evaluation of lator for adsorptive separations. Chem Eng Sci. 1994;49:3115–
dual-bed pressure swing adsorption for CO2 recovery from boiler 3125.
exhaust Gas. Separ Purif Technol. 2001;24:519–528. 36. Skarstrom CW. Method and apparatus for fractionating gaseous mix-
17. Park JH, Beum HT, Kim JN, Cho SH. Numerical analysis on the tures by adsorption US Patent 2,944,627, 1960.
power consumption of the PSA process for recovering CO2 from 37. Jiang L, Fox VG, Biegler LT. Simulation and optimal design of
flue gas. Ind Eng Chem Res. 2002;41:4122–4131. multiple-bed pressure swing adsorption systems. AIChE. 2004;
18. Gomes VG, Yee KWK. Pressure swing adsorption for carbon diox- 50:2904–2917.
ide sequestration from exhaust gases. Sep Purifi Technol. 2002; 38. Ruthven DM, Farooq S, Knaebel KS. Pressure Swing Adsorption.
VCH Publishers, Inc, New York, 1994.
28:161–171.
39. Sircar S, Hufton JR. Why does the linear driving force model for
19. Na BK, Koo KK, Eum HM, Lee H, Song H. CO2 recovery from
adsorption kinetics work? Adsorption. 2000;6:137–147.
flue gas by PSA process using activated carbon. korean J Chem
40. Balakrishna S, Biegler LT. Targeting strategies for the synthesis and
Eng. 2001;18:220–227.
energy integration of nonisothermal reactor networks. Ind Eng Chem
20. Na BK, Lee H, Koo KK, Song HK. Effect of rinse and recycle Res. 1992;31:2152–2164.
methods on the pressure swing adsorption process to recover CO2
41. Biegler LT, Cervantes AM, Wächter A. Advances in simultaneous
from power plant flue gas using activated carbon. Ind Eng Chem
strategies for dynamic process optimization. chem Eng Sci.
Res. 2002;41:5498–5503. 2002;57:575–593.
21. Choi WK, Kwon TI, Yeo YK, Lee H, Song H, Na BK. Optimal 42. Wächter A, Biegler LT. On the implementation of an Interior-point
operation of the pressure swing adsorption (PSA) process for CO2 filter line-search algorithm for large-scale nonlinear programming.
recovery. Korean J Chem Eng. 2003;20:617–623. Math Program. 2006;106:25–57.
22. Chou CT, Chen CY. Carbon dioxide recovery by vacuum swing 43. LeVeque RJ. Numerical Methods for Conservation Laws. Bir-
adsorption. Separ Purif Technol. 2004;39:51–65. khäuser, Basel, 1992.
23. Ko D, Siriwardane R, Biegler LT. Optimization of pressure swing 44. Webley PA, He J. Fast solution-adaptive finite volume method for
adsorption and fractionated vacuum pressure swing adsorption proc- PSA/VSA cycle simulation; 1 Single step simulation. Comput Chem
esses for CO2 capture. Ind Eng Chem Res. 2005;44:8084–8094. Eng. 2000;23:1701–1712.
24. Grande CA, Cavenati S, Rodrigues AE. Pressure swing adsorption 45. Kameswaran S, Biegler LT. Convergence rates for direct transcrip-
for carbon dioxide sequestration presented at ENPROMER 2005, tion of optimal control problems using collocation at radau points.
4th Mercosur Conference on Process Systems Engineering, Rio de Comput Optimi Appli. 2008;41:81–126.
Janeiro, Brasil, 2005. 46. MATLAB User’s Guide. The Mathworks, Inc., Natick, MA, 1994–
25. Chaffee AL, Knowles GP, Liang Z, Zhang J, Xiao P, Webley PA. 2005.
CO2 capture by adsorption: materials and process development. Int J 47. Kameswaran S, Biegler LT. Simultaneous dynamic optimization
Greenhouse Gas Control. 2007;1:11–18. strategies: recent advances and challenges. Comput Chem Eng.
26. Zhang J, Webley PA, Xiao P. Effect of process parameters on power 2006;30:1560–1575.
requirements of vacuum swing adsorption technology for CO2 cap- 48. Fourer R, Gay DM, Kernighan BW. A modeling language for math-
ture from flue gas. Energ Conver Manage. 2008;49:346–356. ematical programming. Manage Sci. 1990;36:519–554.
27. Xiao P, Zhang J, Webley PA, Li G, Singh R, Todd R. Capture of
CO2 from flue gas streams with zeolite 13X by avacuum-pressure Manuscript received Mar. 25, 2009, revision received July 31, 2009, and final
swing adsorption. Adsorption. 2008;14:575–582. revision received Oct. 1, 2009.

1828 DOI 10.1002/aic Published on behalf of the AIChE July 2010 Vol. 56, No. 7 AIChE Journal