Journal of Cleaner Production 170 (2018) 704e714

Contents lists available at ScienceDirect

Journal of Cleaner Production

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

Optimal design and control of pressure swing adsorption process for

N2/CH4 separation
Qiang Fu, Haiyu Yan, Yuanhui Shen, Yingjie Qin, Donghui Zhang*, Yan Zhou
Collaborative Innovation Center of Chemical Science and Engineering, School of Chemical Engineering and Technology, Tianjin University, Tianjin 300072,
China

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

Article history: Pressure swing adsorption (PSA) is a very versatile, albeit complex gas separation and purification
Received 9 December 2016 technology. Due to its complexity and periodic operation, calculation of the optimal PSA system that
Received in revised form simultaneously obtains process design and control decision variables is a complicated task. This work
5 September 2017
presents detailed design and control optimization study of a two-bed, six-step PSA system aimed at
Accepted 17 September 2017
Available online 19 September 2017
heavy component CH4 upgrading. The key optimization objective is to enhance product CH4 recovery
while achieving a closed loop product CH4 purity of 75% for separating 68%N2/32%CH4 feed under
external disturbances. Traditional sequential and simultaneously design and control approach are
Keywords:
Simultaneous optimization and control
employed and compared based on this purpose. The benefits of simultaneous methodology over con-
VPSA optimization ventional sequential approach are successfully demonstrated by closed-loop performance results and
Process control simulation profiles. The simultaneously design and control methodology has succeeded in synthesizing
Methane upgrading the optimal PSA cycle which can generate CH4 recovery as high as 97.30%.
Coal bed methane © 2017 Elsevier Ltd. All rights reserved.

1. Introduction engineering processes like cryogenic distillation, membrane and

Coal bed methane (CBM) is a kind of gas mixture that discharged
in series of underground coal mining activities. Inside of this
mixture, CH4 is a kind of major component with certain added
economic value. CBM usually contains 20e35% of CH4 as well as
other contaminants like N2, O2, CO2, H2O (Bhadra and Farooq, 2011)
Such specific gas constitution made it hard to be utilized directly.
Emitting it to atmosphere directly would not only be a waste of
clean energy but also lead to severe environmental destruction
because the greenhouse warming potential of CH4 is 25times
higher than CO2 (Krzysztof, 2008; Flores, 1998; Wu et al., 2015;
Rosaria et al., 2017). Although low concentration of CH4 cannot be
utilized, secondary-quality methane (30e95%) can be used for
electricity generation or fuels (USA EPA, 2005). Thus, recovering
CH4 from CBM is a kind of activity that could converts waste into
wealth and capture greenhouse gas to avoid potential global
warming (Karapidakis et al., 2010; Bo€rjesson and Berglund, 2007).
To separate and enrich CH4 from CBM, many chemical
* Corresponding author. The Research Center of Chemical Engineering, School of
Chemical Engineering and Technology, Tianjin University, Tianjin 300072, China.
Tel.: þ086 022 27892097.
E-mail address: donghuizhang@tju.edu.cn (D. Zhang).
pressure swing adsorption (PSA) or vacuum pressure swing
adsorption (VPSA) have already proved to be feasible (Zhong and
Daraboina, 2013; Huang and Paul, 2007; Cavenati et al., 2006).
Although cryogenic distillation is nearly suitable for all kinds of
separation process, its enormous energy consumption could be the
biggest drawback considering the specific CBM constitution and
physical properties. For method of membrane separation, it is
difficult to make a proper trade-off between an ideal selectivity and
lower industrial-scale construction cost considering the present
technical state. PSA is much more suitable for this gas mixture
comparing with the former two methods because of its low cost in
energy consumption and highly automatic operation (Liu et al.,
2011).
Since its first practical application in later 1950s, PSA and VPSA
have been used in many mixtures separation and there is no
exception in CBM separation and concentration (Hiro et al., 2005).
Gomes and his coworkers simulated the VPSA process to recover
CH4 from coal-seam methane then verified their validation by ex-
periments. Results show that CH4 could be concentrated from 50%
to no less than 90% with recovery of 60% by 2-bed 4-step operations
(Gomes and Hassan, 2001). Besides, Olajossy uses both simulation
and experimentation to study processes of CH4 recovering from
CBM. Their pilot plant research was carried out based on their

https://doi.org/10.1016/j.jclepro.2017.09.169
0959-6526/© 2017 Elsevier Ltd. All rights reserved.
 Q. Fu et al. / Journal of Cleaner Production 170 (2018) 704e714
simulation prediction and reached a recovery of no less than 96%
(Olajossy et al., 2003).
From above work, it can be easily known that researchers have
used both simulation and experiment to study the PSA system.
However, as a result of the fact that the performance of a PSA
process correlates, it is difficult to get the optimal condition and
control it under outside disturbances just by experiments and pilot
plants. Luckily, in the last three decades, the systematical mathe-
matical models of PSA system have been established and simulated
successfully. Based on this, it is now often used the sequential-
based approach to get the best simulation results while control
the system under external disturbances. This means the model-
based system would be optimized to achieve certain goals while
meet some constraints firstly. Then, the optimized condition is
simulated within some unknown scenarios and use controllers to
make it satisfied some production standards (Sircar, 2006;
Jayaraman et al., 2004; Khajuria and Pistikopoulos, 2013).
For the optimization part, Smith and Westerberg presented a
mixed-integer non-linear programming (MINLP) to optimize the
operation cost (Smith and Westerberg, 1992). Then Nilchan pro-
posed a simultaneous discretization in both spatial and temporal
and converts the partial differential algebraic equations (PADEs)
into nonlinear algebraic equations and made it possible to use the
NLP solver to solve it (Nilchan and Pantelides, 1998). Soon after-
wards, Ko et al. configured the r-SQP to decide the optimal operate
variables like feed flowrates, step pressure and time duration (Ko
and Moon, 2002). Huang et al. built a rigorous PSA model and
solved it in software gPROMs. Also in Sun's work, she built a
fundamental PID control strategy and testified it with several kind
of external disturbance (Huang et al., 2008; Sun et al., 2015).
Then in the control part, a controller is needed for the chemical
system. A few numbers of studies have focused on control system in
chemical unit operation like distillation and crystallization. For
example, Singh et al. designed a proportional-integral-derivate
(PID) controller and compared its performance with open loop
simulations (Singh et al., 2014). Torre et al. established a model
predictive controller for periodic adsorption process (Torre et al.,
2005). Khajuria et al. proposed a model predictive control (MPC)
strategy for a H2/CH4 PSA separation system and tested the per-
formance of the MPC controller and compared it with the PID
controller (Khajuria and Pistikopoulos, 2011).
In industrial PSA separation processes, there are inevitable un-
known disturbances that may induce performance variables
changes. This means that the desired quality of product needs to be
controlled and enhanced through incorporating proper controller
during the design stage. In traditional sequential approach, opti-
mization is carried out firstly to get the best operate condition.
External disturbances are then added into the newly got optimal
condition and applied certain control strategy to make product
meet necessary constraints. Such working order may possibly meet
product's quality demand but loses its best operation condition
after the second step. Thus, developing a systematic integrated
design and control strategy that satisfied the design target during
both steady state and transient in the presences of various un-
known disturbances and process has become urgent to PSA systems
(Sakizlis et al., 2004). Simultaneous design and control PSA system
employing the optimization-based approach has significant
meaning for its practical operation in industries. Focusing on this
point, we have built a dynamic optimization approach and used it
in the PSA system.
This paper presented a systematic model-based simulation,
optimization and control of the PSA process for methane produc-
tion. A PDAEs-based model of the vacuum pressure swing
adsorption process for separation of CH4 from CBM mixture is
studied with feed composition of 68%N2/32%CH4. The accuracy of
705
mathematical models was verified by a series of experiments that
differ in feed and replacement feed flowrates from 32 m3/h to
50 m3/h. Based on the cyclic steady state that reached from the
simulation work, the sequential approach that optimizing recovery
of CH4 to maximum on the basis of product purity not less than 75%
with certain design constraints (pressures, valve constant,
replacement gas flowrate and evacuation gas flowrate) was firstly
carried out. Using the optimization results, disturbance like tran-
sient impulse in feed flow rate was added into the system while
employing the embedded PID controller to manipulate the
adsorption step time duration to maintain product purity as no less
than 75%. Afterwards, the newly built simultaneous optimization
and control framework was presented and added the same external
disturbance onto it with same optimization objective and con-
straints. Finally, we compared the simultaneous optimization-
based simulation performances with sequential-based approach.
Results show that comparing with traditional sequential approach,
the simultaneous approach has shorter response time, better inside
operation condition and prior performance facing external
disturbances.
2. PSA model validation
2.1. PSA model
A systematical PSA model should be able to describe the
inherent dynamic changes of the process as time goes on while
other constraints like definition of cycle steady state and boundary
conditions within it. In this paper, such a PSA model as well as its
optimization and control framework is built in gPROMS.
The following assumptions are used for the PSA bed model:
(1) The gas phase behaves as an ideal gas.
(2) No radial variation in gas concentration, temperature and
pressure.
(3) Pressure drop along the bed is calculated by the Ergun
equation.
(4) Thermal equilibrium between the gas and solid phases.
(5) The porosity of bed and adsorbent particle is uniform along
the bed.
(6) Adsorption rate is approximated by linear driving force (LDF)
approximation model with a single lumped mass transfer
coefficient.
(7) Adsorption behaviors are described by Langmuir isotherm
model.
In this work, the competitive adsorption behavior is described
by the extended Langmuir isotherm based on pure component data
that presented in Eq. (5). Activated carbon (AC) is selected as the
adsorbent in this paper just as most N2/CH4 research did. The ki-
netic adsorption rate is represented by the linear driving force (LDF)
approximation model, which is described in Eq. (6). Other mathe-
matical equations and sub-models like valves could be seen in
Tables 1e3 list key parameters for N2/CH4 system. Corresponding
boundary conditions for individual steps are listed in Tables 4 and 5.
As most papers did, we also used purity and recovery to assess the
performance of a PSA process and the definition of these two pa-
rameters are listed in Table 6.
2.2. Cycle scheme design
In general, a PSA system usually contains two or more adsorp-
tion beds interacting with each other via a net of solenoid valves.
Different steps are achieved by a system that controls opening of
different valves in different steps. The cycle scheme for a heavy
706 Q. Fu et al. / Journal of Cleaner Production 170 (2018) 704e714

Table 1
Mathematical model for VPSA systems simulations used for CH4-N2 separation.

Parameter Expression

Component mass balance


v2 ci v vg ci
 vci vq
εb Dax þ þ εb þ ð1  εb Þεp þ rp ð1  εb Þ i ¼ 0 (1)
vz2 vz vt vt

Overall mass balance


v2c v vg c
 vc vq
εb Dax þ þ εb þ ð1  εb Þεp þ rp ð1  εb Þ ¼0 (2)
vz2 vz vt vt

Energy balance
!

XN
 XN
 vT XN
vT
εb þ ð1  εb Þεp ci cpg;i  R þ ð1  εb Þrp cps þ ð1  εb Þrp qi cpg;i  R þ vg rg cpg;i
i¼1 i¼1
vt i¼1
vz
(3)
XN
vqi T  Tw
 vP v2 T
þ ð1  εb Þrp DHi þ 2h  εb þ ð1  εb Þεp  kg 2 ¼ 0
i¼1
vt Rb vt vz

Gas phase momentum balance

vP 150mð1  εb Þ2 ð1  εb Þrg  
 ¼
2 vg þ 1:75 vg vg (4)
vz ε3 2Rp 2Rp ε3
b b

Langmuir isotherm
qm;i bi Pi
q*i ¼ Pn ; (5a)
1þ bi Pi
i¼1
DHi 
bi ¼ b0 exp (5b)
RT

Linear driving force model

vqi 15Dc;i
* 
¼
2 qi  qi (6)
vt Rp

Diffusion coefficients
vg Rp
Dax ¼ 0:73Dm þ  ; (7a)
εb Dm
εb 1 þ 9:49 2v g Rp
εp Dk;i Dm
Dc;i ¼ (7b)
t Dk;i s
þ ffiffiffiffiffiffiffiffiffiffiffiffiffi
Dm

T
Dk;i ¼ 48:5Dp ; (7c)
M
ri ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

0:1013T 1:75 1
MCH4 þ MN2
1

Dm ¼  2 (7d)
1 = 1 =
P Dv;CH4 3 þ Dv;N2 3

Valve equation
.
unidirection : if Pin > Pout ; then F ¼ CV ðPin  Pout Þ 106
else; then F ¼ 0 . (8)
bidirection : F ¼ CV ðPin  Pout Þ 106

component upgrading process is different from processes that remain the same during the whole optimization and control part.
collecting light products. There usually contains a replacement (RP) The configuration of the two-bed six-step PSA process is shown in
step, also called heavy component or product purge, in schedule Fig. 1 and A framework for simultaneous design, operational, and
which uses high concentration of CH4 in product gas buffer to raise control process under disturbance is shown in Fig. 2. The schedule
the product purity following adsorption (AD) step. This step can not of PSA system is presented in Table 7.
only raise adsorption quantity of CH4 on adsorbents but also In the first step AD, feed gas passed into adsorption bed with
exclude light product like N2 in dead spaces. Normally after various certain pressure after compressor (Compressor 1). Most of N2
pre-capture operations, the CBM system would be stored in raw
material buffer tank under atmosphere pressure and normal tem-
perature. Therefore, the feed flue gas is assumed to be 0.1 MPa with Table 3
temperature of 303.15 K and contains 68% of N2 and 32% of CH4. Modeling parameters for the gas-solid system.
Both beds are 1.0 m high and 0.15 m in radius and buffer tanks in Parameters Value
Fig. 1 are all 0.1256 m3 in volume. All these geometric parameters
Cpg(J/mol K) 0.735
Cps(J/mol K) 0.887
dp(m) 1  109
Table 2 Rp(m) 6.0  104
Langmuir isothermal model parameters of N2 and CH4 on AC. rp(kg m3) 525
DvCH4 (cm3/mol) 24.5
Parameters N2 CH4
DvN2 (cm3/mol) 18.5
1 6
b0(Pa ) 1.73✕10 1.17✕106 kg (J/m K s) 0.242
qm(mol/kg) 2.97 3.47 mðPa s) 1.68✕105
-DH(KJ/mol) 17.082 21.267 εp 0.33
 Q. Fu et al. / Journal of Cleaner Production 170 (2018) 704e714 707

Table 4 two stainless steel columns packed with AC. Pressure gauges (PGs)
Boundary conditions for the adsorption and cocurrent blowdown steps for the six- used in experiment are produced by Shanghai Sighto Instrument
step PSA cycle.
Co., China and there uncertainty is 1% in full scale. The PGs are fixed
Z¼0 Z¼L at the top of bed to monitor the internal bed pressure. These two
Adsorption beds are connected with series of solenoid valves. These valves in
vg Abed ¼ Qfeed vg ¼ fvalve ðP; Poffgas ; CVVW ; CVVW0 Þ our research could achieve different steps under the control of S7-
yi ¼ yfeed vyi
vz ¼ 0 200 Micro PLC scheme. Mass flow controllers (MFC) and mass flow
T ¼ Tfeed vT ¼ 0 meter (MFM) are manufactured by Chengdu Laifeng Co., China and
vz
Cocurrent blowdown uncertainties of both are 0.2%. Raw gases were made by pure N2 and
vg ¼ 0 vg ¼ fvalve ðP; PCoD ; CVCoD Þ CH4 mixed up in proportion of 68%N2/32%CH4. These gases were
vyi vyi
vz ¼ 0 vz ¼ 0 pumped into the bed through the Fujian Juba BX2025 type, China
vT ¼ 0 vT ¼ 0
vz vz produced compressor. The CH4 concentration in the product gas
stream under different flowrates had been measured by the gas
Chromatograph (Beijing North sub-Rayleigh Analytical Instruments
Table 5 SP2100A). While these products were extracted from beds via a
Boundary conditions for the replacement (cocurrent purge with product),
diaphragm pump (ULVAC DAT-500, Japan). Figs. 3 and 4 show re-
evacuation and cocurrent repressurization with feed steps for the six-step
PSA cycle. sults of both simulation and experimentation under different raw
gas feed flowrates and replacement feed flowrates. It can be easily
Z¼0 Z¼L
seen from Figs. 3 and 4 that both results of purity and recovery of
Replacement(Cocurrent purge with product) simulation and experiment show same changing tendency as feed
vg Abed ¼ QRP vg ¼ fvalve ðP; Poffgas ; CVRP Þ
flowrate changes and have little derivation which means the
yi ¼ yproduct vyi
¼0
vz models and parameters are accurate enough to describe the process
T ¼ Tproduct vT ¼0
vz and suitable for further model simulation and computation.
Evacuation
vg Abed ¼ QVU vg ¼ 0
vyi vyi
vZ ¼ 0 vz ¼ 0 3. Sequential process design
vT ¼ 0 vT ¼ 0
vZ vz
Cocurrent repressurization with feed The traditional sequential process contains both optimization
vg Abed ¼ Qfeed vg ¼ 0
step and control step while the latter one is based on the results of
yi ¼ yfeed vyi
vz ¼ 0 the former. In this paper, we have employed the SRPQD method in
T ¼ Tfeed vT ¼ 0
vz gPROMS to solve the optimization problems and use the self-built
PID controller to control the external disturbances. Both steps are
done in software gPROMS. PSA is a dynamic process, which means
Table 6
it would changes regularly in cycles once it reaches CSS. The
Process performance indicators.
judging criteria is adopted to define the state of CSS and it can be
Performance indicators Expression used expressed as follows:
Purity CH4 Purity ¼ moles of CH4 in extract  
moles of extract  
Recovery CH4 Recovery ¼ moles of CH4 in net extract ecss ¼ yt¼Ntcycle  yt¼ðNþ1Þtcycle  < εcss lim
moles of CH in feed
4
x/∞

In this equation, y could represent any significant variables in

would be exhausted from top of the bed while CH4 is captured in
bed. In the following co-current blowdown step (CoD), valve VV_D is
open so that N2 was excluded into the gas phase to enrich purity of
CH4 without any energy consumption. RP step then uses high
concentration of CH4 that stored in product buffer-tank passed
through compressor 2 to further increase product concentration
and expel N2 from both adsorbent and dead spaces. The vacuum
(VU1 and VU2) step extracted high purity of CH4 product through
the Vacuum and regenerated the adsorbent fully. In fact, VU1 and
VU2 are the same vacuum step while writing them separately could
make it convenient for our control part analyzation. Finally, feed
pressurization (FR) step is conducted to raise pressure in bed and
make it ready for the next cycle.
Numerical simulations of the above PSA process are solved using
gPROMS. The modeling parameters for the gas-solid system are
adopted from our previous work and listed in Table 8. Firstly, the
central finite difference method (CFDM) is employed to convert the
PDAEs system to a differential algebraic equation (DAEs) system.
The corresponding DAEs system was solved and intergrading over
time dynamically by method of lines (MOL) in a DAEs integrator to
cyclic steady state (CSS).
2.3. Experimental verification
The bench scale experimental unit that used in experiments has
model like purity or recovery. Once the value difference between
any variable at one cycle and the next is less than the critical value
εCSS , we would define this condition has reached to its CSS. In this
work the value of εCSS is 105. In this paper, 50 cycles is chosen as
the time horizon for the PSA dynamic optimization studies to reach
CSS. Framework for optimization and control part is given below.
3.1. Optimization step in sequential process
Recovery, which is its direct relation to the plant operating cost,
is selected as the optimal objective function for PSA optimization
process. CVV-W and CVV-W0 affect the adsorption step by controlling
the flow rates at the outlet. The parameter of CVV_D is crucial for the
co-current blowdown step. The system would attempt to reduce
the impurity intake during this step, as well enrich purity of pru-
duct without any energy consumption. Switch valve CVV_RP is a
crucial parameter of the replacement step. The replacement step,
also called the product purge step, aims to increase the concen-
tration of CH4 in the column and exclude N2 from the column. The
optimal CVV_RP which is sensitive to flow rate of replacement step
(QRP) can push the adsorption front forward and widen the mass
transfer zone, largely enhancing the amount of replaced N2 and the
purity of CH4. The value of CVV_P is sensitive to flow rate of evac-
uation step (QVU). QRP and QVU will seriously impact on the product
purity and recovery. The key process design variables switch valves
708 Q. Fu et al. / Journal of Cleaner Production 170 (2018) 704e714

Fig. 1. Graphical overview of the PSA process and process control scheme.

Fig. 2. A framework for simultaneous design, operational, and control process under disturbance.

(ON-OFF operation) that CVV-W, CVV-W0, CVV_D, CVV_RP, CVV_P are showed in Table A1 of Supplementary material. Time duration is
considered. It is assumed that the structure of the PSA cycle and fixed for each step in optimization process to simplify the calcula-
sequence of processing steps remains fixed for the optimization tion process and all PDAEs are discrete in spatial alone. The opti-
studies. The switching schedule of valves and pumps over a cycle is mization formulation can be expressed as follows:
 Q. Fu et al. / Journal of Cleaner Production 170 (2018) 704e714 709

Table 7
Schedule of the VPSA process. PurityCH4  75:00%
Time/s 200 25 120 200 25 120

Bed1 AD[ CoD[ RP[ VU1Y VU2Y FR[ PVU  0:1bar

Bed2 VU1Y VU2Y FR[ AD[ CoD[ RP[

FR: concurrent repressurization with feed; AD: adsorption; CoD: concurrent PAD  2:5bar
blowdown; RP: replacement (concurrent purge with product); VU: evacuation step.

PRP  1:5bar
Table 8
Optimal decision variables, upper and lower boundaries and optimal performance
for PSA sequential approach optimization. PCoD  1:3bar
Decision variable Initial Value LB UB Optimal Value

Process Design VarLB  Var  VarUB

CVV_W 6.5 0.2 8 7.14
CVV_W0 0.5 0.2 4.0 0.62 In this optimization framework, F represents the DAEs equations
CVV_RP 1.8 0.1 10 1.92 in the VPSA model. Also in this equation, y is the state variable, q
CVV_D 2.0 1.0 20 3.23 stands for decision variables. Initial values used in optimization
CVV_P 4.0 0.2 10 0.94 process come from the simulation state that set replacement
Operational Variables
QRP 32 3.0 40 35.78
flowrate as 32Nm3/h and vacuum flowrate as 45 Nm3/h. Table 8
QVU 45 30 80 41.65 shows the calculation results of the optimization process while
Performance indicators Table 9 gives the corresponding computational statistics.
PurityCH4 75.03% 75.01% It can be seen from Tables 8 and 9 that with a fixed feed flowrate
RecoveryCH4 80.91% 98.01% of 45 Nm3/h, the process used 42.81 h to maximize recovery from
80.91% to 98.01%. Although purity dropped a little as recovery
increasing, no less than 75% concentration of methane is get to
further utilization. Pressure and temperature changes of a cycle
before and after optimization are shown in supplementary infor-
mation. Detailed changes inside the bed are shown in the 5th part
comparing with that of simultaneous approach.

3.2. Control step in sequential process

Once got the best operate condition, the sequential approach

Fig. 3. Methane purity and recovery comparison of experimental and simulation re-
sults for different feed gas flowrate.
would add the external disturbances into the system and use the
embedded controller to maintain certain parameters beyond some
standard (Khajuria and Pistikopoulos, 2013; Santos et al., 2013;
Lopes et al., 2011). In this research, product purity is set as the
control objective that is to maintain it as no less than 75% as purity
standard (Ps) under feed flowrate changes. The design parameters
are all based on the optimization values that have just get and fixed
during the PID controller. Here is the framework for the PID
controller in this paper:

Disturbance: Q feed ¼ Q 0feed  4:5;

 
2pt t
Tfeed ¼ 303:15 þ 5sin exp
5 10

verror

Controller: ¼ 100 purityCH4  Ps
vt


eðkÞ ¼ 100 purityCH4  Ps

Ts
PID ¼ Kc ðeðkÞ  eðk  1ÞÞ þ Kc eðkÞ þ Kd error
tI

tmin  t  tmax
Fig. 4. Methane purity and recovery comparison of experimental and simulation re-
sults for different replacement gas flowrate.
Dtmin  Dt  Dtmax
Objective: Max e Recovery of CH4
Constraints: F(y, y', q, t) ¼ 0; if jeðk  1Þj > jeðkÞj then
710 Q. Fu et al. / Journal of Cleaner Production 170 (2018) 704e714
tk ¼ tk1  0:5PID
else if P > 0 then
tk ¼ tk1  PID
else tk ¼ tk1  10PID
tAD ¼ tVU1 ¼ t
Time duration of the AD step is selected as the objective that
been controlled and adjust itself judging from the controller feed-
back. This is because this time duration is closely co-related with
product purity and can be easily achieved in real industrial process.
It is well known that time duration of each step in one PSA process
is always coupling with other steps so that we have wrote the
equation that time duration of AD step is the same as VU1 step even
though it changes in the control process. Besides a feedback from
controller, constraints of largest time change (Dt) on time adjust-
ment also added to avoid drastic time duration which may lead to
over saturation (Dtmin) or sudden surge of inflow (Dtmax). Other
key parameters in the controller are listed in Table 10. Time
schedule as well as pressure and temperature changes of a cycle
after control are shown in supplementary information. Based on
the above structure, performances of the control system is obtained
and shown in 5th part.
4. Simultaneous optimization and control of PSA process
system
It is the same for the simultaneous approach that an intelligent
objective function of PSA dynamic optimization is to maximize
product recovery with strict constraint for purity, which is set as
the interior constraint in the optimization framework of gPROMS
dynamic optimizer gOPT in this paper. On the other hand, the
controller performance is usually defined as the ability to stabilize
the product purity to its set-point under random process distur-
bances in closed loop. So the simultaneous optimization and con-
trol methodology based on optimization framework is the ideal
strategy to incorporate these conflicting performance indictors on
real-time PSA process operations. In the following paragraph, the
configuration for the optimization based framework is briefly
explained and the final optimization problem formed is listed
below. To provide the control actions for a PSA process under dis-
turbances, a computationally efficient controller is desired. The
MPC control scheme, as an advanced control methodology, is
perfectly suitable for highly complex, interconnected dynamic
systems according to related papers had published (Seferlis et al.,
2012; Papadopoulos and Seferlis, 2017; Bollas et al., 2007). But a
PID control as a simplicity and practicality methodology was
Table 9
Solver and convergence statistic for optimal system.
parameter value
Number of axial nodes 20
Discretization scheme CFDM
DAE solver absolute and relative convergence tolerance 10e5
DAE solver relative convergence tolerance 10e5
NLP solver convergence tolerance 10e5
Number of NLP iterations 27
Number of NLP line searches 39
CPU times 42.81 h
Fraction of CPU spend on sensitivity integration 67.61%
Machine details Intel Core
i5-2100@ 3.10 GHz
Table 10
Parameter setting for the PID controller.
Ps Kc Kd tmax (s) tmin (s) Dtmax (s) Dtmin (s)
0.7500 0.2 0.001 220 180 20 20
extensively used in most practical industrial PSA processes control
systems. The design of our PSA simultaneous optimization and
control process has considered PID controller to make our study
more authentic with the industrial PSA design (Mulgundmath et al.,
2012; Khajuria and Pistikopoulos, 2010). Just the same as sequential
approach, product purity is selected as the control variable and the
center for PID controller design. All the process variables are
interacted with each other over PSA cycle which makes the deter-
mination of manipulative variable is of crucial importance. The
potential manipulative variable should able to fast track purity to
its set-point without dramatic deterioration. Hence, the following
PID controller that uses tAD as regulated variable is designated as
controller for PSA system under consideration. The key process
design variables switch valves that CVV-W, CVV-W0, CVV_D, CVV_RP,
CVV_P are considered, while QRP and QVU are included as operational
decision variable.
To ensure all the physical behaviors of optimal designed PSA
process are closely realistic under the whole time horizon of
regulation process, some constraints for the process variables are
necessary and essential. Besides, simultaneous approach needs
more constraints because it faces both control and optimization at
the same time, which may lead to non-convergence if computation
goes on without those constraints, and this situation won't happen
in sequential approach that the two compute goes separately. Gas
velocity, bed temperature and adsorption step duration are chosen
for this purpose which can represent the process operation con-
dition, adsorbent activity and controller extent, respectively. Inte-
rior constraints are set for those variables at every cycle step of
dynamic optimization and control in gPROMS optimizer. Further-
more, the constrained PID controller performance is weighed by
integral square error (ISE) as defined in the following framework.
Manipulated variable bias value, controller gain, and integral time
are considered as decision variables. The above objective function,
operational constraints plus the decision variables and constraints
for simple optimization of PSA process make up the ultimate
simultaneous optimization framework:
Objective: Max e Recovery of CH4
Constraints: hd1(xd, xa, dp, v, t) ¼ 0
hc1(xd, xa, dc, v, t) ¼ 0
Zt

2
ISE ¼ purityCH4  75:00 dt
0
0  ISE  εc
purityCH4  75:00%
PVU  0:1bar
PAD  2:5bar
PRP  1:5bar
 Q. Fu et al. / Journal of Cleaner Production 170 (2018) 704e714 711

Table 11 Table 13
Optimal decision variables, upper and lower boundaries and optimal performance Comparison of optimal performance indicators obtained from sequential and
for PSA simultaneous approach optimization. simultaneous approach.

Decision variable Initial Value LB UB Optimal Value Performance variables Purity/% Recovery/% ISE

Process Design initial cycle steady state 75.03 80.91 e

CVV_W 6.5 0.2 8 6.58 sequential approach optimization step 75.01 98.01 e
CVV_W0 0.5 0.2 4.0 0.49 sequential approach control step 75.01 92.15 1.002
CVV_RP 1.8 0.1 10 1.83 simultaneous approach 75.00 97.29 0.45
CVV_D 2.0 1.0 20 1.92
CVV_P 4.0 0.2 10 1.21
Operational Variables
QRP 32 3.0 40 37.30 longer than the sequential approach. But the simultaneous
QVU 45 30 80 40.50 approach is still promising because control part is also been done in
Controller Design this time duration with better performance indicators. Recovery of
Kd 0.001 0.0001 10 0.0026
the process has been maximized to 92.15% and maintains its purity
Kc 0.2 0.01 100 0.65
tI 190 180 220 185.28 at an acceptable state. Time schedule as well as pressure and
Performance indicators temperature changes of a cycle after simultaneous approach
PurityCH4 75.03% 75.00% computation are shown in supplementary information. To make an
RecoveryCH4 80.91% 97.30% insight observation of both approaches as well as making a com-
parison, the following part is given.

5. Results comparison and analyzation

PCoD  1:3bar
Performance indicators are the main standard to evaluate
180  tAD  200 approach performances. Table 13 shows the results of three states
after the same out disturbances, they are: original operate state
0:01  vgAD  0:1; 0:0001  vgCoD  0:001 without any control and optimization, sequential approach out-
comes at the end of control step, as well as simultaneous approach.

0:01  vgRP  0:1; 0:8  vgVU  0; 0:1  vgFR  20

VarLB  Var  VarUB

where hd1 and hc1 stand for PSA model and controller, respec-
tively, while the decision variables are represented by dp (process
design), dc (controller design), and with v being the operational
variables, xa and xd refer to algebraic variables and differential
variables respectively. The εc is a tuning parameter that can be
changed to obtain a stable and fast controller response. In this
paper, the value of εc is fixed at 1.0. VarLB and VarUB represent the
lower and upper bounds of decision variables which are visible in
Table 11. The control part framework is the same as it is in
sequential part thus we do not list its equations again. In our
research, essentially, simultaneous approach is a kind of optimi-
zation approach in which certain step time duration is variable and
modified itself using the feedback from a PID controller. Such
control process is achieved by adding the control equations into the
Task modules in gPROMS software. Results of the simultaneous
approach are shown in Tables 11 and 12.
It can be seen in the two tables that simultaneous approach
takes 61.1 h to reach it's optimize operate conditions, which is

Table 12
Solver and convergence statistic for optimal system.

Parameter Value

Number of axial nodes 20

Discretization scheme CFDM
DAE solver absolute and relative convergence tolerance 10e5
DAE solver relative convergence tolerance 10e5
NLP solver convergence tolerance 10e5
Number of NLP iterations 48
Number of NLP line searches 62
CPU times 61.10 h
Fraction of CPU spend on sensitivity integration 77.58%
Machine details Intel Core
Fig. 5. Comparison of purity changes (A) and response time (B) in sequential and
i5-2100@ 3.10 GHz
simultaneous approach.
712 Q. Fu et al. / Journal of Cleaner Production 170 (2018) 704e714
Fig. 6. Comparison of axial distribution of gas-phase methane concentration at the end of (A) AD step (B) CoD step (C) RP step (D) VU step.
Results of both approach show that they all have a good per-
formance in enhance product recovery with certain constraints.
Although optimization results in sequential step show the best
recovery in value, this advantage soon lost once external distur-
bance added into the process. Recovery resulted from the simul-
taneous approach is nearly 5% higher than the final result of
sequential optimization approach. Besides, the simultaneous
approach also has a smaller ISE in value which means that this
approach has relatively smaller errors for the PID controller and
could adjust itself much more rapidly. To highlight this feature,
Fig. 5 is made to show the corresponding response time with per-
formance variables.
From the picture we could see that both approaches have good
performance facing external disturbance and could finally reach to
a satisfied purity with the PID controller. The simultaneous
approach reached to its optimized state much more rapidly,
smoothly and has little errors comparing with the traditional
sequential approach. These can be a good explanation to its lower
ISE value. The simultaneous approach could maintain a stable pu-
rity while doing its optimization and this is why it can reach its
optimal condition more rapidly. Although the sequential approach
has a better recovery after its optimization step, it uses more time
to convergence and has a larger vibration in its control step. Be-
sides, the final results of the sequential step are not that ideal as the
recovery dropped nearly 6% comparing with the optimization re-
sults. All in all, comparing with the traditional sequential approach,
the simultaneous approach is much more time saving, better in
performance indicators and suitable for industrial application. To
deeply analyze the insight changes in bed, gas phase distribution at
the end of some key steps are shown in Fig. 6.
Sharper adsorption front is visible from all steps under
simultaneous approach and this especially conspicuous in AD and
RP step. The optimization computation uses a better feed flowrate
and reasonable valve opening to make the adsorption front shaper
and make further influence on performance indicators. In AD step,
all three states have the same adsorption front at 0.25 in axial
position, but the sharper shape after optimization leaves more
clean adsorbents for following steps. While the flat adsorption front
under initial step make it difficult to get a better recovery. The in-
fluence of the adsorption front becomes much more conspicuous in
CoD step. In this step, CH4 purity enhanced without any energy
consumption, but the higher purity at the top of the bed (axial
position of 1.0 m) means more CH4 is lost. Thus, it can be conclude
that simultaneous approach has a better result than sequential
approach and the initial state has the worst recovery. Then in the RP
step, front of simultaneous approach moves 0.1 m further in axial
position than the rest two because of its larger input flowrate.
Finally in VU step, decrease in raw gas feed flowrate is accompanied
with decrease in vacuum pump flowrate while VU step time
duration is coupling with a longer AD step time duration after
control. The sequential approach shows the best desorption effect
in Fig. 6. This could desorb the adsorbents more clearly for the next
cycle and contributes to a better adsorption front. All in all, these
observations suggest that the simultaneous based optimization
approach predicts a more efficient and robust PSA system, while
also maintaining the closed loop purity in the presence of specified
disturbances. Solid-phase distribution at the end of each step under
these three states is shown in supplementary information.
6. Conclusions
In this paper, an optimization-based simultaneous optimal and
 Q. Fu et al. / Journal of Cleaner Production 170 (2018) 704e714 713

control strategy for PSA process design is presented. A PDAEs-based qi adsorbed phase concentration in equilibrium with gas
model of VPSA process for CH4 separation and enrichment from phase, mol/Kg
CBM mixture is considered with the feed composition of 32% CH4 Q 0feed initial volumetric feed flow rate, Nm3/h
and 68% N2. Afterwards, the traditional two-stage, sequential-based Q volumetric feed flow rate, Nm3/h
approach is employed to obtain a relatively good PSA cycle process R universal gas constant, J/mol K
on purpose of comparison with the optimization-based optimal Rb bed radius, m
PSA process system. Critical PSA system controller decision vari- Rp particle radius, m
ables are designed so efficient controller system is incorporated in t time, s
the simultaneous optimization-based framework. The above two T bed temperature, K
different design methodologies had been implemented success- Ts sampling interval, s
fully in gPROMS to maximize the product recovery with some Tw wall temperature, K
necessary constraints. The simultaneous optimization-based vg superficial velocity, m/s
methodology was applied to the two-bed six-step PSA process y gas phase molar fraction
system in the presence of various disturbances. Comparison of the z axial position in the adsorption bed
simultaneous optimization-based methodology closed-loop simu-
lation performances with sequential-based approach show that the Greek symbols
simultaneous approach has a better performance facing external εb bed porosity
disturbance, shorter response time and better inside operation εc control tolerance
conditions. εp particle phase porosity
rp adsorbent particle density, kg/m3
Acknowledgments t tortuosity factor
tI integral time constant for PI controller
This research is financially supported by the State Key Labora- m gas viscosity, Pa s
tory of Chemical Engineering of Tianjin University (No.SKL-ChE- DHi isosteric heat of adsorption, kJ/mol
16B05) and the Key Laboratory of Ethnic Affairs Commission and Dts sampling interval, s
Ministry of Education of China (No.KF2015003).
Subscripts
Appendix A. Supplementary data i gaseous species
in model entrance
Supplementary data related to this article can be found at out model exit
https://doi.org/10.1016/j.jclepro.2017.09.169. offgas waste gas from PSA cycle
product waste gas from PSA cycle
feed feed step of PSA cycle or feed conditions
Notation CoD cocurrent blowdown step of PSA cycle
RP replacement (cocurrent purge with product) step of PSA
cycle
English Symbols FR cocurrent repressurization with feed step of PSA cycle
b adsorption affinity parameter, 1/Pa VU evacuation step of PSA cycle
b0 adsorption affinity constant, 1/Pa
c gas phase molar concentration, mol/m3 References
Cpg gas phase heat capacity at constant pressure, J/Kg K
Bhadra, S.J., Farooq, S., 2011. Separation of methane-nitrogen mixture by pressure
Cps adsorbent heat capacity, J/Kg K
swing adsorption for natural gas. Ind. Eng. Chem. Res. 50, 14030e14045.
Cv value constant, mol/bar s Bollas, G.M., Anastasiou, I., Papadopoulou, S.A., Voutetakis, S.S., Seferlis, P., 2007.
Dax effective axial dispersion coefficient, m2/s Feed conversion targeting in an FCC pilot plant using a non-linear MPC strategy.
Dc effective diffusivity coefficient, m2/s In: Proceedings of the 2007 American Control Conference, pp. 632e638.
Bo€rjesson, P., Berglund, M., 2007. Environmental systems analysis of biogas sys-
Dm molecular diffusion coefficient, m2/s tems: the environmental impact of replacing various reference systems.
Dk Knudsen diffusivity coefficient, m2/s Biomass & Bioenergy 31, 326e344.
dp adsorbent particle pore diameter, m Cavenati, S., Grande, C.A., Rodrigues, A.E., 2006. Separation of CH4/CO2/N2 mixtures
by layered pressure swing adsorption for upgrade of natural gas. Chem. Eng. Sci.
DvCH4 molecular diffusion volume of component CH4, cm3/mol 61, 3893e3906.
DvN2 molecular diffusion volume of component N2, cm3/mol Flores, R.M., 1998. Coal-bed methane: from hazard to resource. Int. J. Coal Geol. 35,
3e26.
e purity deviation from its set-point Gomes, V.G., Hassan, M.M., 2001. Coalseam methane recovery by vacuum swing
h overall heat transfer coefficient, J/m2 K s adsorption. Sep. Purif. Technol. 24, 189e196.
ISE integral square error Hiro, M, Omori I, Oba M, Kawai T. 2005. Carbon dioxide separation and recovery
system. JP Patent 2,005,262,001.
Kc controller gain Huang, Y., Paul, D.R., 2007. Effect of film thickness on the gas-permeation charac-
Kd derivative gain teristics of glassy polymer membranes. Ind. Eng. Chem. Res. 46, 2342e2347.
kg effective axial thermal conductivity, J/m K s Huang, Q., Malekian, A., Eic, M., 2008. Optimization of PSA process for producing
enriched hydrogen from plasma reactor gas. Sep. Purif. Technol. 62, 22e31.
kLDF;CH4 mass transfer coefficient of component CH4 Jayaraman, A., Hernandez-Maldonado, A.J., Yang, R.T., Chinn, D., Munson, C.L.,
kLDF;CH4 mass transfer coefficient of component N2 Mohr, D.H., 2004. Clinoptilolites for nitrogen/methane separation. Chem. Eng.
Sci. 59, 2407e2417.
Hb bed length, m
Karapidakis, E.S., Tsave, A.A., Soupios, P.M., Katsigiannis, Y.A., 2010. Energy effi-
M gas mole weight, g/mol ciency and environmental impact of biogas utilization in landfills. Int. J. Environ.
N total number of component Sci. Technol. 7, 599e608.
p pressure, Pa Khajuria, H., Pistikopoulos, E.N., 2010. An explicit/multi-parametric controller
design for pressure swing adsorption system. IFAC Proc. 43, 206e211.
q adsorbed phase concentration, mol/Kg Khajuria, H., Pistikopoulos, E.N., 2011. Dynamic modeling and explicit/multi-
qm maximum adsorption capacity, mol/Kg parametric MPC control of pressure swing adsorption systems. J. Process
714 Q. Fu et al. / Journal of Cleaner Production 170 (2018) 704e714
Control 21, 151e163.
Khajuria, Harish, Pistikopoulos, Efstratios N., 2013. Optimization and control of
pressure swing adsorption processes under uncertainty. AIChE J. 59, 120e131.
Ko, D., Moon, I., 2002. Multiobjective optimization of cyclic adsorption process. Ind.
Eng. Chem. Res. 41, 93e104.
Krzysztof, W., 2008. Harnessing methane emissions from coal mining. Process Saf.
Environ. 86, 315e320.
Liu, C., Zhou, Y., Sun, Y., 2011. Enrichment of coal-bed methane by PSA com-
plemented with CO2 displacement. AIChE J. 57, 645e654.
Lopes, Filipe V.S., Grande, Carlos A., Rodrigues, Alırio E., 2011. Activated carbon for
hydrogen purification by pressure swing adsorption: multicomponent break-
through curves and PSA performance. Chem. Eng. Sci. 66, 303e317.
Mulgundmath, V.P., Jones, R.A., Tezel, F.H., Thibault, J., 2012. Fixed bed adsorption
for the removal of carbon dioxide from nitrogen: breakthrough behaviour and
modelling for heat and mass transfer. Sep. Purif. Technol. 85, 17e27.
Nilchan, S., Pantelides, C., 1998. On the optimisation of periodic adsorption pro-
cesses. Adsorption 4, 113e147.
Olajossy, A., Gawdzik, A., Budner, Z., Dula, J., 2003. Methane separation from coal
mine methane gas by vacuum pressure swing adsorption. Chem. Eng. Res. Des.
81, 474e482.
Papadopoulos, A.I., Seferlis, P., 2017. Process Systems and Materials for CO2 Capture:
Modelling, Design, Control and Integration. John Wiley and Sons Ltd., Chi-
chester, pp. 501e521.
Rosaria, A., Maria, C., Maria, C.A., 2017. Pressure swing adsorption for biogas
upgrading. A new process configuration for the separation of biomethane and
carbon dioxide. J. Clean. Prod. 140, 1390e1398.
Sakizlis, V., Perkins, J.D., Pistikopoulos, E.N., 2004. Recent advances in optimization-
based simultaneous process and control design. Comput. Chem. Eng. 28,
2069e2086.
Santos, M.P.S., Grande, C.A., Rodrigues, A.E., 2013. Dynamic study of the pressure
swing adsorption process for biogas upgrading and its responses to feed dis-
turbances. Ind. Eng. Chem. Res. 52, 5445e5454.
Seferlis, P., Stavrakis, I., Papadopoulos, A.I., 2012. Homotopy continuation solution
method in nonlinear model predictive control applications. Comput. Aided
Chem. Eng. 30, 1327e1331.
Singh, R., Barrasso, D., Chaudhury, A., Sen, M., Ierapetritou, M., Ramachandran, R.,
2014. Closed-loop feedback control of a continuous pharmaceutical tablet
manufacturing process via wet granulation. J. Pharm. Innov. 9, 16e37.
Sircar, S., 2006. Basic research needs for design of adsorptive gas separation pro-
cesses. Ind. Eng. Chem. Res. 45, 5435e5448.
Smith, O.J., Westerberg, A.W., 1992. The optimal design of pressure swing adsorp-
tion systems. Chem. Eng. Sci. 47, 4213e4217.
Sun, W., Shen, Y., Zhang, D., Yang, H., Ma, H., 2015. A systematic simulation and
proposed optimization of the pressure swing adsorption process for N2/CH4
separation under external disturbances. Ind. Eng. Chem. Res. 54, 7489e7501.
Torre, V.D.L., Fox, V.G., Bandoni, J.A., Biegler, L.T., 2005. Model predictive controller
for periodic adsorption processes. In: ENPROMER, Proceedings of 2nd Mercosur
Congress on Chemical Engineering, Brazil, pp. 403e418.
USA EPA 430-K-04e003, 2005. Identifying Opportunities for Methane Recovery at
U.S. Coal Mines: Profiles of Selected Gassy Underground Coal Mines
19992003.
Wu, B., Zhang, X., Xu, Y., Bao, D., Zhang, S., 2015. Assessment of the energy con-
sumption of the biogas upgrading process with pressure swing adsorption
using novel adsorbents. J. Clean. Prod. 101, 251e261.
Zhong, D.L., Daraboina, N., 2013. Recovery of CH4 from coal mine model gas mixture
(CH4/N2) by hydrate crystallization in the presence of cyclopentane. Fuel 106,
425e430.