Sie sind auf Seite 1von 13

Computers and Chemical Engineering 35 (2011) 177–189

Contents lists available at ScienceDirect

Computers and Chemical Engineering


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

Derivation of optimal operating policies under safety and technological


constraints for the acetoacetylation of pyrrole in a semi-batch catalytic reactor
Gheorghe Maria ∗ , Anca Dan
Department of Chemical Engineering, University Politehnica of Bucharest, P.O. 35-107, Bucharest, Romania

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

Article history: A large number of optimization strategies have been developed for semi-batch reactors (SBR) with a fixed
Received 29 June 2009 or free terminal time, with or without considering various sources of uncertainty. Such strategies account
Received in revised form 25 April 2010 for safety constraints rather empirically determined in the form of parameter thresholds, while safety
Accepted 5 May 2010
indices are seldom integrated in the optimization objective function. The present work illustrates how
Available online 13 May 2010
the runaway boundaries, and their confidence region associated to the parameter uncertainty, can be
evaluated using the process model and a generalized sensitivity criterion, and how they can be included
Keywords:
in the SBR optimization. A concrete example is provided for the SBR used for the acetoacetylation of
Semi-batch reactor
Optimization
pyrrole with diketene in homogeneous catalysis, a process known to be of high risk due to the very
Safety boundaries exothermic side-reactions. While previous studies approached the isothermal SBR, the present work is
Pyrrole focused on optimizing the non-isothermal SBR operation.
Diketene © 2010 Elsevier Ltd. All rights reserved.

1. Introduction operating policy and of technological and safety margins for the
operating variables/parameters.
Optimal operation of a batch or a semi-batch reactor (SBR) is In spite of being intensively investigated, derivation of the opti-
a classical, but still of high interest problem due to several rea- mal operating policy of a SBR remains a complex problem to be
sons. On the one hand, it is the economic interest to increase the solved for every particular case. As reviewed by Bonvin (1998),
plant productivity, the reactor being the main equipment on which there are several important aspects to be considered in formula-
the optimization efforts are focused due to the high value of raw- tion of an optimization objective function: maximization of the SBR
materials and products related to the production cost, but also due productivity and product quality with using minimum necessary
to its high sensitivity to operating conditions affecting the prod- raw-materials and utilities; fulfilment of safety requirements in the
uct quality and raising possible safety problems (Bonvin, 1998). On case of highly exothermic primary (desirable) or secondary (unde-
the other hand, it is the current trend to move production of a lot sirable) reactions, possibly developing chain exothermic reactions
of chemicals from the stable continuous plants to multi-product with gas production; fulfilment of the technological and ecological
(semi-)batch reactors, as being more flexible and easily adaptable constraints; scale-up the laboratory information to get a process
to market requirements. If the chemical reaction exhibits a high model and SBR optimal operating conditions; ensure a certain flex-
thermal effect, it is preferable to be conducted in a semi-batch ibility of the SBR optimal operation, to become readily adaptable
mode: one of the reactants is slowly added to the second com- to the available equipment in a multipurpose-plant and multi-
ponent, which has already been fully loaded to the reactor. Thus, product development strategy.
the evolution of the reactor temperature can be kept under con- Consequently, optimization of the SBR operation involves
trol, while a certain optimal feeding policy can ensure production multiple challenging issues to be considered: time-varying char-
maximization. An advanced control is usually implemented to keep acteristics of the process, due to the SBR way of operation, but
the reaction synthesis within defined safety limits and in an eco- also from batch-to-batch due to adjacent unit operations, resulting
nomic operating region (Loeblein, Perkins, Srinivasan, & Bonvin, in a variability of raw- and recycled-material purity; nonlin-
1999). However, frequent perturbations in the operating parame- ear characteristics of the process and heat transfer, difficult to
ters, raw-materials and recycling conditions, catalyst replacement be precisely modelled; model inaccuracies coming from a large
or reactivity modifications, all require periodical updates of the SBR variety of sources (experimental data quality, adopted simplifi-
catory hypotheses, material variability, identification procedure);
few available measurements, especially when frequent changes
∗ Corresponding author. Tel.: +40 744 830 308. in production are required; the presence of a large number of
E-mail address: gmaria99m@hotmail.com (G. Maria). technological, economical, ecological, and safety constraints to

0098-1354/$ – see front matter © 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compchemeng.2010.05.003
178 G. Maria, A. Dan / Computers and Chemical Engineering 35 (2011) 177–189

Nomenclature
j stoichiometric coefficient of species j
Notations  liquid phase density
Ar , Ae , Am heat exchange surface of the reactor: Ar is mea-
standard deviation
sured inside the reactor; Ae is measured cooling fluid
2 variance
side; Am is the logarithmic mean between Ar and Ae  time constant
A Arrhenius frequency factor  ad induction time to explosion under adiabatic condi-
aj correlation coefficients in Eq. (2) tions
B = Tad E/(RTo2 ) reaction violence index D time of addition of co-reactant D
cj component j concentration
cp molar specific heat Superscripts
d distance ∧ estimated value
dr reactor inner diameter - average value
ds stirrer diameter * reference value
Da = (−A )(rA  D )/cA,o Damköhler number for SBR ∼ distribution
E activation energy
F feed flow-rate (liquid) Index
f model function vector a cooling agent
g path constraint function vector ad adiabatic
G terminal constraint function vector c critical
Hfo standard enthalpy of species formation f final
−H reaction enthalpy in inlet
he reactor convective heat transfer coefficient, on the l liquid phase
cooling agent side max maximum
hi heat transfer coefficient, on the reactor side min minimum
IHi the increments of structural groups i for enthalpy of o initial
formation in the Joback formula w referring to the reactor wall
k rate constants
nj species j number of moles Abbreviations
N normal distribution, or stirrer speed BR batch reactor
Ndiv number of equal divisions of the batch time tf D diketene
Pr Prandtl number DHA dehydroacetic acid
P probability FDM finite difference method
PN normal cumulative distribution function MINLP mixed-integer NLP
p overall pressure MV Morbidelli–Varma criterion
Qevap vaporization heat NC nominal conditions
R universal gas constant NLP nonlinear programming
Re Reynolds number P pyrrole
r chemical reaction rate PAA 2-acetoacetyl pyrrole
si the number of structural groups of type i (in the Py pyridine
chemical formula) SBR semi-batch reactor
s(y;) absolute sensitivity, ∂y(t)/∂ STD standard deviation
S(y;) normalized sensitivity, (∗ /y∗ )s(y; )
St = (U Ar D )/(V ¯ c̄p ) Stanton number for SBR
t time
be considered with a certain weight in order to obtain the best
tf final batch time
operating solution (compromise); presence of disturbances in the
T temperature
operating parameters making accurate implementation of the opti-
Tad temperature rise under adiabatic conditions
mal policy questionable; limited on-line corrective actions due to
TMRad time-to-maximum-rate under adiabatic conditions
the finite duration of a (semi-)batch.
U overall heat transfer coefficient
Optimal operation of a SBR is usually focused on deriving an
u manipulated variable vector
optimal feeding policy for the co-reactant and, sometimes, an
V liquid (reactor) volume
optimal temperature dynamic profile to be realized in the reac-
x, y state variable vector, or dependent variables
tor, by usually manipulating one or two input variables. In all
w penalization weights in the Lagrange function
alternatives, the optimal policy for a fixed or free terminal batch
time results from the minimization of an objective function in the
Greek symbols
presence of problem constraints. If known, the sources of uncer-
 finite difference
tainty and its characteristics can be included during optimization,
ı random variation
leading to a certain degree of the problem solution uncertainty.
ıt reactor wall thickness
As a consequence, various procedures have been developed for
˚ optimization objective function
the model-based SBR optimization (Abel & Marquardt, 2003;
 operating parameter vector
Kadam, Schlegel, Srinivasan, Bonvin, & Marquardt, 2007; Rahman &
t ,  thermal conductivity of the reactor material; ther-
Palanki, 1996; Srinivasan, Palanki, & Bonvin, 2002, 2003; Terwiesch,
mal conductivity
Agarwal, & Rippin, 1994): robust (off-line) one-time optimization
dynamic viscosity
without (nominal) or with considering the parameter uncertainty,
and using the off-line derived process model; optimization of the
G. Maria, A. Dan / Computers and Chemical Engineering 35 (2011) 177–189 179

expected value of the objective function under uncertainty; batch- considering any systematically derived safety constraint. In the
to-batch optimization using a model updating step and additional present study, the optimal feeding policy with diketene solution
information from past batches (the so-called ‘tendency modelling’); will be obtained under non-isothermal conditions, including safety
on-line SBR optimization, the solution being adapted based on requirements explicitly, in the probabilistic sense. Besides, both the
the acquired on-line information from the process and used to solution-feeding rate, and its concentration will be considered as
improve the model adequacy and the operating policy for the rest optimization variables, leading to an increased SBR flexibility.
of the batch. All these formulations are solved using various nonlin- The developed computing methodology can be easily extended
ear programming (NLP) or mixed-integer NLP (MINLP) numerical to optimization of other SBRs of industrial interest, by realizing a
solvers due to the presence of a large number of constraints. satisfactory compromise between the productivity/low-cost goal
The safety requirements, especially for highly thermal sensitive and safety requirements.
BRs or SBRs, are usually considered during the reactor optimiza-
tion as explicit constraints (many times empirically determined
from observations and past experience with the process), by tak- 2. Process characteristics, reaction hazard, and semi-batch
ing certain levels of state sensitivity functions as constraints (Chen, reactor model
Erickson, & Fan, 1970; Seinfeld & McBride, 1970), or by keeping a
certain distance from the runaway boundaries due to the uncer- The acetoacetylation of pyrrole with diketene in homogeneous
tainty in the operating parameters (Mönnigmann & Marquardt, liquid phase (toluene), at around 50 ◦ C and normal pressure, hav-
2003; Takamatsu, Hashimoto, & Ohno, 1970; Watanabe, Nishimura, ing pyridine as catalyst, is a common method for producing pyrrole
& Matsubara, 1973; Wen & Chang, 1968). Recent developments derivates, such as 2-acetoacetyl pyrrole (PAA) used in the drug
check safety constraints in the form of scenario-integrated path industry. The process is of high thermal risk due to the tendency
constraints, via repeated (time-consuming) simulations over a cer- of very reactive diketene to polymerize at temperatures higher
tain time horizon (Abel & Marquardt, 2003). All such strategies than 60–70 ◦ C, or in the presence of impurities that can initiate
suffer from the lack of a systematic and precise model-based deriva- highly exothermic side-reactions difficult to be controlled (Nakano,
tion of the safety limits of the operating region to be considered Kosaka, Hiyama, & Nozaki, 2003). A few number of investigations
during SBR optimization. on the process kinetics have been reported, the most relevant being
In the homogeneous reaction case, the accumulation of the co- the study of Ruppen et al. (1997) in a bench-scale isothermal SBR
reactant is possible at low temperatures due to the reduced reaction operated at 50 ◦ C using a high excess of toluene as solvent.
rates. As a consequence, a sudden increase in the reaction tem- Due to such reasons, the reaction at industrial scale is con-
perature leads to an exponential rise in the reaction rates (when ducted in a semi-batch mode: the solution of diketene in toluene is
the co-reactant is in large amounts), which in turn will lead to a slowly added to the reactor, initially loaded with solution of pyrrole
continuous rise in the reactor temperature up to conditions when in toluene (recovered from previous batches and including some
the cooling system becomes ineffective, eventually determining impurities). In principle, the evolution of the reactor temperature
the reactor explosion. The situation is worsened by the occurrence can be kept under control by adjusting the feeding rate and cool-
of exothermic chain/polymerization reactions that lead to a quick ing regime, while a certain feeding policy can ensure production
increase in temperature or pressure with eventually the same effect maximization. Although optimal operating conditions can be deter-
(Grewer, 1994; Stoessel, 2008; Varma, Morbidelli, & Wu, 1999). mined from lab-scale experiments, operation of the industrial SBR
Conditions for runaway in the heterogeneous liquid–liquid systems is however difficult. Small variations in the catalyst characteristics,
are also given in the literature (Westerterp & Molga, 2004). raw-material or solvent quality, a sudden increase in temperature
The scope of this paper is to investigate the advantages of using a due to fluctuations in the cooling system performance, or an accu-
more elaborated and general strategy to evaluate the SBR runaway mulation of reactive diketene favour secondary reactions of risk.
boundaries for a known process, to be considered when deriving Such variations significantly change the selectivity between the
the optimal operating policy in the presence of parameter uncer- desired and undesired products, and therefore lead to also signifi-
tainty. Past researches suggested keeping the optimization solution cant variations in the product quality from batch-to-batch.
inside the random variation region of parameters that never inter- The reaction scheme advanced by Ruppen et al. (1997) includes
sects the safety boundaries (Mönnigmann & Marquardt, 2003). four main steps presented in Table 1. The synthesis of the desired
Our investigations suggest also considering the runaway bound- PAA product is accompanied by several side-reactions of diketene,
ary uncertainty when optimizing the SBR, by keeping the optimal leading to the dimmer DHA, oligomers Dn , or PAA derivatives
solution outside the confidence region of the runaway boundaries. denoted by G (the intermediate reaction of diketene with DHA has
Such a solution corresponds to a minimum probability to inter- been neglected from the model). The co-reactant diketene presents
sect the safety limits for an assumed confidence level, and for an an extreme reactivity and hazardous properties, as revealed by
adopted weight of the safety constraint in the objective function. To existing databases. Because of that, the temperature regime must
derive the SBR thermal runaway conditions, the generalized sen- be strictly controlled and diketene concentration in the reactor
sitivity criterion of Morbidelli and Varma (MV; Varma et al., 1999) kept lower than a certain critical threshold (ca. 0.025 mol L−1 ,
has been used, as being generally applicable irrespectively to the empirically determined; Ruppen et al., 1997; Martínez, 2005).
kinetics or reactor complexity. As another element of novelty, an The kinetic model considers only first or second-order reactions,
adequate algebraic correlation of the critical conditions with the with the rate constants evaluated by Ruppen et al. (1997) at
operating parameters has been derived, valid over the considered nominal operating conditions of 50 ◦ C and [PAA] > 0.1 M. Due to
operating region. Such a simple correlation allows simplifying the incomplete experiments, an activation energy of E/R = 10,242 K was
SBR optimization steps including probabilistic evaluations. adopted in this paper for all reactions of diketene, by analogy
The examined SBR is the bench-scale jacketed reactor for the with the diketene derivate polymerization (Sato, Morita, Tanaka,
acetoacetylation of pyrrole (P) with diketene (D) in homogeneous & Ota, 2003), and with the initiation energy of olefin polymer-
liquid phase, used by Ruppen, Bonvin, & Rippin (1997) to identify ization (Lenz, 1967). The resulted Arrhenius constants (Ai , Ei ) are
the optimal isothermal feeding policies that maintain the reactor displayed in Table 1. All reactions are moderately exothermic,
within technological limits and lead to an acceptable PAA prod- excepting the diketene oligomerization of standard heat around
uct yield. Several other authors investigated this SBR to derive −340 kcal mol−1 (for an n = 10–12 degree of oligomerization, see
an optimal feeding policy under isothermal conditions, without footnote of Table 1).
180 G. Maria, A. Dan / Computers and Chemical Engineering 35 (2011) 177–189

Table 1
Reaction schema, kinetic model, and some hazard indices of the process (Notations: D = diketene; P = pyrrole; PAA = 2-acetoacetyl pyrrole; DHA = dehydroacetic acid;
Py = pyridine).

Reaction, rate expression Arrhenius constants A Reaction enthalpy Tad (K)c B indexd TMRad (min)e Daf Da/Stg
(L mol−1 min−1 ; min−1 ) E/R (K)a (kcal mol−1 )b
ka ,Py
P + D −→PAA Aa = 3.1324 × 1012 Ha = −31.70 61.8 6.07 3.57 7.1 × 107 4.6 × 105
ra = ka [P][D] Ea /R = 10,242.4
b k ,Py
D + D −→DHA Ab = 7.5651 × 1012 Hb = −21.96 5.3 0.52 3.95 1.7 0.01
2
rb = kb [D] Eb /R = 10,242.4
kc
nD−→(D)n Ac = 1.6549 × 1012 Hc = −340.69 83.0 8.15 4.75 4.2 0.03
rc = kc [D] Ec /R = 10,242.4
dk ,Py
PAA + D −→G Ad = 1.7731 × 1012 Hd = −31.70 7.7 0.76 4.21 4.5 × 106 2.9 × 104
rd = kd [PAA][D] Ea /R = 10,242.4
a
k = A × exp(−E/RT ). Lacking of kinetic data, the activation energy has been adopted the same for all reactions, corresponding to that of diketene derivate polymerization
(Sato et al., 2003), being in the same range with the free-radical polymerization of olefins (Lenz, 1967). Rate constants at 50 ◦ C are (Lee et al., 2001): ka = 0.053; kb = 0.128;
kc = 0.028; kd = 0.003 (all in L mol−1 min−1 , excepting kc in min−1 ). Rate constants increase with the medium polarity (not accounted here), the reported values being valid for
[PAA] > 0.1 mol L−1 (Ruppen et al., 1997). Other secondary reactions (such as DHA + D → G ) have been neglected.
b
Ha was derived from species Hfo (Jolkkonen, 2000). Hb was taken from Lopatin, Popov, Epshtein, Mikhaleva, & Makarov (1992). Hc was derived using the standard
-1
Hfo = −897.5 kcal mol of diketene oligomer Dn , corresponding to a n = 10 cyclic oligomer or a n = 12 acyclic (Nakano et al., 2003), and evaluated by means of Joback’s

formula (Reid, Prausnitz, & Poling, 1987): Hfo = 68.29 + si IHi (kJ mol−1 ), where si = the number of structural groups of type i; IHi = the increments of structural groups i
for enthalpy of formation. Hd was adopted at the same value with Ha .
c
Tad = (−H)cj,o /(l cp,l ) = temperature rise of the reaction mixture under adiabatic conditions (Froment & Bischoff, 1990), where cj,o = initial concentration of key
species; l = liquid mixture density; cp,l = specific heat of the liquid mixture.
d
B = Tad E/(RTo2 ) = reaction violence index (Grewer, 1994; Maria, 2007), where To = initial temperature of the reaction. Values of B > 5 indicate hazardous reactions, of
high risk.
e
TMRad = time-to-maximum-rate under adiabatic conditions, obtained by reactor simulations.
f
Da = D /1 = (−A )(rA D )/cA,o = Damköhler number; St = (U Ar D )/(V ¯ c̄p ) = Stanton number for SBR. Slow reactions correspond to Da < 50–100 and Da/St < 1
(Grewer, 1994). These numbers are evaluated at initial SBR conditions.

To quickly assess the hazard of reactions (kinetic and ther- the so-called exothermicity, reactivity and cooling numbers, that
modynamic), some probability, severity and criticality indices for combine influences of various parameters leading to operation
runaway scenario in SBR can be used, critical operating condi- within a critical region (Westerterp & Molga, 2004).
tions being defined by simple inequalities (Grewer, 1994; Hugo Application of such a quick hazard assessment in the present
& Steinbach, 1986; Maria & Heinzle, 1998; Shah, Fischer, & case study reveals that reactions #1–3 (and especially #3,
Hungerbühler, 2003, 2005): the oligomerization of diketene) from Table 1 are potentially
−1
TMRad = t|r=rmax < 8h; ad = (ko B) (probability indices)
(−H)cj,o Tad E
Tad = > 50 K, B= >5 (severity indices)
¯ c̄p RTo2
D (−A )rA D (1)
Da = = < 50 − 100 (criticality index for slow reactions vs. addition time)
1 cA,o
Da U Ar D
< 1, St = (combined criticality index)
St V ¯ c̄p

where TMRad = time-to-maximum-rate under adiabatic dangerous even under the nominal initial operating condi-
conditions;  ad = adiabatic induction time to explosion; tions mentioned in Table 2. Thus, Tad = 62 K (reaction #1) and
B = reaction violence index; Tad = adiabatic temperature rise; Tad = 83 K (reaction #3) are larger than the threshold 50 K; B = 6
Da = Damköhler number for the key reactant A (start material); (reaction #1) and B = 8.1 (reaction #3) are larger than the thresh-
St = Stanton number; cj ,o = initial concentration of key species; old 5; Da = 1.7 (reaction #2) and Da = 4.2 (reaction #3) are smaller
¯ = reacting mixture density; c̄p = average specific heat; than the threshold 50–100 (slow reactions); avg. TMRad ≈ 4 min
To = initial temperature of the reaction; R = universal gas con- is much smaller than the 8 h threshold; Da/St = 0.01 (reaction #2)
stant; t = reaction time;  D = time of addition of the co-reactant; and Da/St = 0.03 (reaction #3) are much smaller than the thresh-
 l = reaction time constant, e.g. 1/ko for the first-order reaction, old 1. Consequently, reactions #1–3 display a high thermodynamic
and 1/(ko cBo ) for the second-order reaction between A and B at and kinetic hazard, being susceptible to runaway at higher tem-
small conversions; ko = reaction rate constant at the cooling tem- peratures if diketene is present in large amounts. The adiabatic
perature T˛ ; U = heat transfer coefficient; Ar = heat transfer area; induction time to explosion of  ad = (ko B)−1 ≈ 4 min for reaction
V = reactor volume; −A = stoichiometric coefficient of reactant #3 (Grewer, 1994) confirms the conclusion that oligomerization
A. Indices’ values overstepping the recommended thresholds of diketene is the most dangerous reaction for this process.
indicate potential dangerous reactions, presenting a fast evolution To make quick estimations of the critical conditions and optimal
and significant exothermicity. It should be noted that SBR safety operating policy, a simple SBR model was adopted, correspond-
conditions are related not only to how fast and exothermic the ing to a perfectly mixed vessel, with no mass and heat transfer
reactions are, but also to their dynamics in relationship with the resistances in the liquid (Froment & Bischoff, 1990). The solution
rate of addition of the co-reactant. The mentioned approximate of diketene is continuously added with a constant/variable fed
methods based on simple engineering numbers can only give a flow-rate F(t) over the continuously stirred pyrrole solution (of
rough idea about unsafe operating conditions. For heterogeneous known initial composition, including impurities coming from pre-
SBR, similar runaway conditions can be derived by accounting for vious batches). The cooling agent circulating in the reactor jacket
G. Maria, A. Dan / Computers and Chemical Engineering 35 (2011) 177–189 181

Table 2
Semi-batch reactor model, nominal operating conditions, technological and safety constraints (D = diketene; P = pyrrole; PAA = 2-acetoacetyl pyrrole; DHA = dehydroacetic
acid).

Observations

Differential balance equations


Species mass balance: At t = 0: cj = cj,o ;T = To
dcD
dt
= (−r̃a − 2r̃b − r̃c − r̃d ) + (cD,in − cD ) VF(t)
(t)
;
dcP dcPAA
dt
= −r̃a − cP VF(t)
(t)
; dt = (r̃a − r̃d ) − cPAA VF(t)
(t)
;
dcDHA F(t) dV
dt
= r̃ b − cDHA V (t)
; volume variation: dt
= F(t);
Reaction rate correction with the catalyst dillution: rc is not corrected with the catalyst dillution, the reaction
r̃j = rj VV(t)
o
; j = a, b,d displaying another mechanism (the stoichiometric
 coefficient was included in the rate constant)
(− Hj ) r̃j V (t)−UAr (T −Ta )
dT F(t)(Tin −T ) j
Heat balance: dt
= V (t)
+ ¯ c̄p V (t)
,
j = a, b,c, d
Model hypotheses
Semi-batch reactor model with perfect mixing and
uniform concentration and temperature field
dr2 4
Cyllindrical reactor of variable liquid volume and heat Ar = 4
+ dr
V (t)
transfer area.
Heat transfer coefficient computed with the formula: ReN = (ds2 N )/
¯ ¯
ıt Ar hi dr

0.14
1
= 1
+ + 1 Ar
; = 0.36 Re0.66 Pr 0.33
Pr = (c̄p )/
¯  ¯
U hi t Am he Ae ¯ N w
(approximate value for nominal conditions is
U = 500 kcal m−2 h−1 K−1 )
Heat of solvent vaporization in the reactor is neglected Qevap ≈ 0
Heat capacity and density of fed solution are the same with ¯ in c̄p,in ≈ ¯ c̄p
those of reactor content.
Nominal operating conditions and range of variationa
Initial liquid volume (Vo , L) 1 Initial [PAA] (mol L−1 ) 0.08 ≤ 0.10 ≤ 0.20
Reactor inner diameter (dr , m) 0.1 Initial [DHA] (mol L−1 ) 0.01 ≤ 0.02 ≤ 0.04
Stirrer speed (N, rot. min−1 ) 640 Fed D-solution flow-rate (F ×100, L min−1 ) 0.05 ≤ 0.15 ≤ 0.20
Liquid physical properties: toluene solvent Batch time (tf , min): 120 ≤ 145 ≤ 378
−1
Inlet [D] (mol L ) 4 ≤ 5.82 ≤ 6 Initial temperature (To , ◦ C) 40 ≤ 50 ≤ 60
−1
Initial [P] (mol L ) 0.4 ≤ 0.72 ≤ 0.8 Cooling agent temperature (Ta , ◦ C) 50
Initial [D] (mol L−1 ) 0.005 ≤ 0.09 ≤ 0.14 Feeding solution (Tin , ◦ C) 50

Process constraint expression Significance


G1 = cDHA,f − 0.15 ≤ 0 (mol L−1 ) Prevent precipitation of DHA at room temperature
(solubility at 50 ◦ C is 0.20 mol L−1 )b
G2 = cD,f − 0.025 ≤ 0 (mol L−1 ) Avoid high concentrations of toxic D in productb ; empirical
critical runaway conditionc
g1 = Max(T (t)) − 70 ≤ 0 (T, ◦ C) Prevent toluene solvent excessive vaporization, pressure
  increase, and dangereous exothermic side-reactionsd
¯
2 )
g2 = g2,j , g2,j = Min P(Fj (t) > Fc ), j = 1, ..., Ndiv ; Fc ∼N(F̄c (), Prevent rapproachemets of all arc input variables Fj on the
F c
Fc runaway boundary under random variations of process
parameters j e
g3 = Da − 100 > 0 Thermal safety critical index in SBR for slow reactionsf
g4 = Da/St − 1 < 0 Thermal safety critical index in SBR for slow reactions and
improper cooling systemg
a
Brendel, Mhamdi, Bonvin, & Marquardt (2003); Martínez (2005); Lee et al. (2001). Operating conditions are varied for ensuring the maximum productivity of the reactor,
i.e. Max PAA yield (nPAA,f /np,o ), or Max P production (nPAA,f = cPAA,f Vf ) (Lee et al., 2001; Loeblein et al., 1997; Martínez, 2005; Ruppen et al., 1997; Srinivasan & Bonvin, 2003).
b
Ruppen et al. (1997).
c
Martínez (2005).
d
Empirically predicted by adding 2 × Tc = 20 K to the nominal temperature, where Tc = RTo2 /E corresponds to the critical conditions of Semenov for zero-order reactions
(Varma et al., 1999; Maria, 2007).
e
For normal distributed random variations of the (uncorrelated) operating parameters around nominal conditions j ∼N( ¯ j ,
 ), the approximate variance of critical
 2 j

Fc can be evaluated using the error propagation formula:


F2 = ∂Fc /∂j
2 (Park & Himmelblau, 1980). Thus, for every batch time arc the constraint becomes:
c j

  j
 
g2,j = Min 1 − PN (Fc − dj < Fc < Fc + dj ) , j = 1, ..., Ndiv , for a given distance from the running to the critical conditions, dj = Fj () − Fc (), and for Fc ∼N(F̄c (),
¯
2 ).
F c
D (−D )rD D
f
The Damköhler number Da = 1
= cD,o
is evaluated for the co-reactant D, where (−D ) = stoichiometric coefficient of D reactant;  D = time of addition of the
co-reactant;  1 = reaction time constant (Grewer, 1994).
UA 
g
The Stanton number St = V ¯r c̄pD is evaluated for the co-reactant D (Grewer, 1994).

continuously removes the reaction heat. The mass and heat bal- been approximated to those of the toluene solvent, and a constant
ance equations, presented in Table 2, explicitly account for the overall heat transfer coefficient has been adopted (evaluated at
liquid volume (and heat transfer area) variation due to the added nominal conditions with the formula from Table 2). The reactor
diketene solution. The continuous catalyst dilution is accounted for model has been solved with a common low order stiff differential
when correcting the reaction rates, excepting for reaction #3 pre- equation integrator.
sumed to be promoted not by pyridine, but by some impurities The characteristics of the reactor, the nominal operating condi-
(of quasi-constant total concentration). To speed-up the computa- tions, and operating parameter ranges reported by various authors
tional steps, the physical properties of the reaction mixture have under isothermal conditions are presented in Table 2. Parameter
182 G. Maria, A. Dan / Computers and Chemical Engineering 35 (2011) 177–189

Fig. 1. (Left, centre) Evolution of the reactor temperature, P-conversion, and PAA yield over the batch time under nominal conditions (NC, Ta = To = 323 K, see Table 2). The
curve indices correspond to constant feed flow-rates of: F = 0.0015 L min−1 (1); F = 0.0050 L min−1 (2); F = 0.0085 L min−1 (3); F = 0.0090 L min−1 (4). (Right) Dynamics of the
temperature and its sensitivity function S(T ; F)/2000 during the batch time, for two inlet flow-rates (F = 0.0010 L min−1 , and F = 0.0084 L min−1 ).

ranges are determined experimentally or by simulation to fulfil sensitivity S(Tmax ; ) under critical conditions, independently of the
the technological constraints G1 and G2 derived from empirical considered operating parameter (e.g. To ,cj,o , Fin , Ta , B, Da, St,. . ., or
observations on safe operating conditions. Consequently, an opti- a combination of them). Thus, critical conditions
 correspond
  to the
mal feeding policy and a tight control of the reactor temperature maximum of the absolute value of s(Tmax ; ), or S(Tmax ; ). Eval-
are necessary to simultaneously fulfil all the productivity and safety uations of s(y; )t = ∂y(t)/∂ derivatives have been performed in
requirements. the present study by means of finite difference numerical method
(FDM), even if its accuracy is computationally costly, because it
3. Derivation of SBR runaway boundaries and their requires small discretization steps near critical conditions.
confidence region To precisely determine the runaway boundaries of the operating
parameters  with the MV-method, it is necessary to investigate
Runaway and eventual explosion of a SBR may occur when first the regions in which the state variables, and especially the
the rate of heat generation becomes higher than the rate of heat reactor temperature, exhibit high sensitivities to the operating con-
removal by the cooling system. Model-based accurate evaluation ditions, even if such set points will not be later used as nominal
of the safe operating region that exhibits low thermal sensitivities, operating conditions. By successively perturbing the main operat-
by setting the runaway boundaries in the parametric space, is not an ing parameters around the nominal conditions, the most influential
easy task. That is because of the computative numerical algorithms variables of the process can be roughly established. Then, moving
to be applied, but also due to fluctuations in the characteristics of the set point in a more sensitive region and performing repeated
the process, catalyst, impurities and raw-materials, thus requiring simulations of the reactor dynamics, critical conditions for the oper-
periodic process model updating (Fotopoulos, Georgakis, & Stenger, ating variables in the parametric space can be obtained (Maria &
1994; Maria, Terwiesch, & Rippin, 1996; Ruppen et al., 1997), and Stefan, 2010).
consequently an update of the runaway boundaries themselves. To determine where the sensitive operating region is located, it
Among alternatives, a sensitivity method has been applied is enough to check by simulation the SBR state variable dynamics
in this paper to evaluate the runaway boundaries in the oper- for various parameter values. Simulations for nominal conditions
ating parametric space. Such methods are based on evaluation (cD,in = 5.82 M) and constant F during the batch in the range of
of the sensitivity function of a state variable y with respect F ∈ [0.0015–0.0090] L min−1 are presented in Fig. 1 (left and cen-
to a parameter , i.e. s(y; )t = ∂y(t)/∂ (in absolute terms), or tre). The plotted dynamics of the reactor temperature, P-conversion
S(y; )t = (∗ /y∗ )s(y; )t (in relative terms; * = nominal operat- and PAA yield, together with the dynamics of reaction rates, reac-
ing point in the parametric space). Unsafe conditions correspond tion Damköhler (Da) or Damköhler/Stanton (Da/St) numbers (not
to a sensitive operation when “the reactor performance becomes plotted here) offer a first guess of the operating region of high ther-
unreliable and changes sharply with small variations in param- mal sensitivity. For instance, starting with F = 0.0085 L min−1 , the
eters” (Varma et al., 1999). The time-dependent state sensitivity temperature T(t) profile not only exhibits values higher than the
functions s(y; )t = ∂y(t)/∂ (i.e. of temperature, conversion, selec- critical threshold of 70 ◦ C (g1 from Table 2), but the curve tends
tivity, species concentrations, pressure) vs. process parameters (i.e. to oscillate. Higher fed flow-rates produce amplification of oscil-
j = [To , Ta , cjo , cj,in , po , B, U, . . .]) are computed by means of a sen- lations, toluene vaporization, a dangerous pressure increase, and
sitivity differential equation solved simultaneously with the reactor eventually the reactor runaway.
dynamic model (Varma et al., 1999). Of most interest in SBR hazard Such an effect can be explained by the slow secondary reac-
assessment is the sensitivity of the reactor temperature hot spot tions #2–3 which become dangerous when the co-reactant D is
Tpeak = (Tmax − To ) or Tmax during the batch operation, i.e. S(Tmax ; accumulating at low temperatures. For small F-levels, the main
). Detection of the operating regions, where certain combinations reaction consumes the co-reactant, the side-reactions are negli-
of parameters lead to a high thermal sensitivity of the SBR, offers gible, and the temperature raise is small. For high F-levels, the
the possibility to accurately set the safety limits of the operating slow side-reactions lead to the co-reactant D accumulation. Conse-
region. Thus, the generalized sensitivity criterion of thermal run- quently, the increased rates of side-reactions will start to generate
away of Morbidelli and Varma (MV; Varma et al., 1999) has been more energy that will increase the reactor temperature. In turn,
used in the present study, as being applicable to complex reactions a high temperature will lead to an increase of all reaction rates,
irrespectively to the reactor type. According to the MV criterion, and to a rapid consumption of the D co-reactant. Depletion of D
the hot spot (Tmax − To ) induces a sharp increase of the normalized will slow-down the reaction rates and will diminish the generated
G. Maria, A. Dan / Computers and Chemical Engineering 35 (2011) 177–189 183

Fig. 2. Runaway boundary curve (—) in the [F− vs.−Ta ] parametric plane depends on the chosen operating conditions: (Left) evaluation for cP,o = 1 mol L−1 , cD,in = 5.82 mol L−1 ;
(Right) evaluation for cP,o = 0.72 mol L−1 , cD,in = 7.5 mol L−1 (other parameters are set to the nominal values). The confidence band (—) of risk curves corresponds to the random
deviations of Ta ± ıTa with ıTa = 3 K.

heat leading to a temperature decrease. At low temperatures, the presented here). It should be observed that critical values of feed-
accumulation of the co-reactant is again possible due to the slow ing rates Fc () depend on the operating parameter vector  = [F,
reaction rates (Fig. 1) and the temperature will start to rise again. Ta , cD,in , cP ,o, To ] around which evaluation is made. For instance, in
The result is a continuous oscillation of the reaction temperature Fig. 2 the runaway boundaries in the F−Ta plane have been obtained
and D-concentration in the reactor, with larger amplitudes as the for two different operating conditions corresponding to high cD,in
F-level is higher. When exceeding a certain critical Fc value, the and cP,o levels. As expected, Fc decreases as the operating severity
temperature oscillations are higher than a tolerable limit, leading increases, corresponding to high Ta temperatures, large cD,in , and
to process escalation and eventually explosion, since the cooling small cP,o concentrations. As the Fc (To ) is roughly parallel to the To
system cannot remove the heat at the same rate. abscissa, it results that initial batch temperature has little influence
The same analysis has been conducted by successively vary- on the critical Fc over the investigated parameter domain.
ing the cooling agent temperature Ta ∈ [40–60] ◦ C, the initial The manipulated inlet variable is the molar feeding rate with
temperature To ∈ [20–60] ◦ C, the diketene solution concentration D co-reactant, that is the product FD,in,c = (FcD,in )c . Keeping in mind
cD,in ∈ [4–7.5] mol L−1 , and the initial concentration of pyrrole that all calculations of critical Fc have been done for a constant
cP,o ∈ [0.4–1] mol L−1 . By comparing the results, one can conclude cD,in = 5.82 mol L−1 (nominal value), extension of the validity of the
that the most influential variables for the process safety are  = [F, runaway boundary plots can be easily performed by replacing the
Ta , cD,in , cP ,o ]. The batch initial temperature presents a small influ- Fc ordinate with the product (5.82 Fc ) moles D min−1 . Thus, the esti-
ence on the reactor dynamics, while small variations of the initial mated boundary curve diagrams can be used for checking operation
concentrations for other species than pyrrole do not present a high at various feeding solution concentrations.
interest. Bringing together all these runaway limit curves in various
To precisely determine the runaway boundaries in every F vs. parameter planes determined over the whole parametric space
j plane (denoted as F−j ), by keeping nominal states for all other Fc (j ), it is possible to obtain the global sensitivity of the reactor
parameters, the time-dependent temperature sensitivity functions by accounting for concomitant variations of several input/process
have been evaluated by replacing the derivatives with finite differ- parameters. An empirical algebraic correlation Fc (), of logarithmic
ences s(T ; j )t = T (t)/ j , and performing repeated simulations type, has been found to adequately represent the critical reactor
under various operating conditions. The time discretization step conditions vs. the most influential operating parameters, of the
has been set very small (tf /5000) to consider all temperature peaks form:
in the critical operating region, and a small discretization step

3
has been used in evaluating the sensitivity functions, i.e. (max − Ln(F̂c ) = ao + aj Ln(j ), with  = [Ta , cD,in , cP,o ], (2)
min )/n, with n = 200–500. The dynamics of such a relative sensi-
j=1
tivity function is displayed in Fig. 1 (right) for two extreme values
of the feeding flow-rate F-levels. Of interest is only the sensitiv- Estimated regression coefficients a = [26.6766425, −5.1050358,
ity of the temperature maximum vs. the control variable, that is −1.05356372, 0.186367982] from a large number of preliminary
S(Tmax ; F), or S(Tmax ; FD,in ). evaluations of critical conditions correspond to a satisfactory model
Let us consider repeated simulations of the SBR dynamics for low standard deviation (8.9 × 10−5 L min−1 for the predicted F̂c ), the
various F-levels, for a chosen operating parameter j (e.g. Ta ), by relative residuals being small (avg. 4.4%) and alternatively positive
retaining the obtained relative  S(Tmax ;F) curve. It is to observe and negative (Maria, 2004).
that a sharp increase in the S(Tmax ; F) is recorded for a certain Uncertainty of type  ¯ j ± ıj can be included in evaluation of
F exceeding a critical value F > Fc . Calculations are then repeated for the confidence region of the runaway boundaries Fc ± ıFc . The
another j value, and retaining every time the corresponding Fc . confidence band in the parametric plane corresponds to a 100% con-
Critical values Fc (j ) determined in this way are then represented fidence level if random fluctuations of parameters are uniformly
in the [Fc vs. j ] plane. The runaway boundary divides the plane into distributed, or to a lower confidence level for normal distributed
two regions, corresponding to a safe or an unsafe operation of the parameters depending on the distribution characteristics (i.e. a
SBR (see Fig. 2). The procedure is repeated by choosing every time 68% confidence level for ıj =
j , a 95% confidence level for ıj =
another operating parameter j (e.g. cD,in , cP,o , or To ), by keeping 2
j , etc.). For instance, the resulted confidence band of Fc (Ta ) for
nominal values for all other parameters, and deriving the corre- ıTa = 3 K is presented in Fig. 2 (with dotted lines). Using the regres-
sponding runaway boundaries Fc (j ) in separate planes (graphs not sion model F̂c (), approximate estimation of Fc variance (denoted
184 G. Maria, A. Dan / Computers and Chemical Engineering 35 (2011) 177–189

by
F2c ) can be obtained using the error propagation formula for the (a) propose Ndiv = 3–8, without including model updating (Lee,
assumed uncorrelated parameters (Park & Himmelblau, 1980): Lee, Lee, & Park, 2001; Loeblein, Perkins, Srinivasan, & Bonvin,

 ∂Fc
2
1997); results indicate Ndiv = 8, tf = 138.62 min, nPAA,f = 0.4409
a21 a22 a23 moles, G1 being slightly violated;

F2c =
2 = Fc2
2 +
2 Ta

T2a +
T2a . (3)
∂j ¯
j Ta 2
cD,in 2
cP,o (b) propose Ndiv = 4–8, with including rate constant updating steps
j 
from on-line measured data (Maria et al., 1996; Srinivasan
Critical conditions and their associated uncertainty must be & Bonvin, 2003); results indicate Ndiv = 7, tf = 378 min,
considered in all studies dealing with determining the optimal nPAA,f = 0.4305 moles, with constraint G1 violated (Ruppen
operating policy of the SBR. Since in many cases the most severe et al., 1997);
running conditions correspond to a higher productivity, such a (c) maximization of PAA production, using Ndiv = 1, with find-
running solution is expected to offer a higher degree of safety in ing the optimal tf and F(t) (Martínez, 2005); results indicate
operation, with the expense of supplementary computational steps tf = 227.50 min, nPAA,f = 0.5045 moles, F(t) = 0.0012 L min−1 ;
with MV-method to derive the confidence of the safety limits. (d) maximization of PAA production, using Ndiv = 2, by impos-
ing the supplementary constraint cDHA,f + cG,f < cmax ,
cmax = 0.15 mol L−1 (Srinivasan & Bonvin, 2003); results
4. Optimization problem formulation and safety
indicate tf = 120.00 min, reaction #3 being neglected (only
constraints
constraint G1 is considered).
4.1. Problem formulation and past results
In solving the associated NLP problem, the constraints have
Derivation of an optimal running policy for a SBR consists in been integrated in the objective function as additive penalty func-
determining the manipulated variable vector u, and sometimes of tions multiplied by certain weights, or as simple penalty barriers
the final batch time tf , from minimization of a suitable objective added to ˚ for avoiding production of the undesired by-products.
function: The reported results given in literature (plots not presented here)
(û, t̂f ) = arg Min ˚(x, u, , tf ) display zigzagging time-profiles for input F(t), usually declining
towards the batch end time. As expected, the SBR productivity
s.t. ẋ = f (x, u, , t), x(to ) = xo (process dynamic model) increases with the reaction time, but also the production of by-
(4)
g(x, u, , t) ≤ 0 (path constraints) products and the risk of violating the technological constraints. An
increased number of time-arcs leads to an improvement of the reac-
G(x(tf ), ) ≤ 0 (terminal constraints)
tor productivity, allowing a more flexible reactor feeding policy, but
where  = operating parameter vector; x = state variable vector. The requires a considerable additional computational effort.
uncertainty of the solution comes from various sources: (i) model Reported optimization solutions suffer from several major
uncertainty in both structure and parameters due to the avail- drawbacks: (i) they do not consider the temperature dynamics dur-
able data uncertainty; (ii) operating parameter uncertainty due to ing the batch (i.e. the energy differential balance of the reactor); (ii)
random disturbances ı␾; (iii) constraint uncertainty ıg,ıG due to they do not consider a preliminary safety assessment step to pre-
their simplified formulation and uncertainty of other variables; (iv) cisely define the runaway boundaries in the parametric space, to be
objective function uncertainty ı˚ due to uncertainty considered for considered during SBR optimization; (iii) they do not account for
the variables and parameters. the parameter or other source of uncertainty which can affect the
For the approached case study, the manipulated variable con- problem solution. Simple setting of safety constraints by means of
sidered for controlling the SBR temperature is FD,in (t) = F(t)cD,in (t). semi-empirical observations, or a rough evaluation using approxi-
However, practical implementation requires a separate treatment mate explicit criteria is not enough to correctly set the search region
of the feed flow-rate F(t) and inlet solution concentration cD,in (t), in the parameter space.
offering a larger degree of freedom by manipulating two physi-
cal input variables. Besides, operation at low F(t) and high cD,in (t),
or at high F(t) and low cD,in (t), even if it might present certain 4.2. Derivation of optimal operating policies with/without
equivalence, can lead to different thermal regimes and production considering parameter uncertainty
flexibility as below proved.
Published trials to find the optimal feeding policy F(t) for this SBR To solve the SBR optimization problem in a more precise and
by several authors, reported various results for a constant inlet con- systematic way, several elements of novelty have been considered.
centration of diketene solution in toluene (cD,in = 5.82 mol L−1 ) and First, the non-isothermal thermal regime is explicitly considered
under isothermal operating conditions. In fact, the derived feeding by means of the heat balance of the reactor, by also imposing to the
policy is given on time-intervals, obtained by dividing the batch temperature to not overstep a threshold suggested by the Semenov
time tf in Ndiv equal parts (the so-called ‘arcs’; Kadam et al., 2007), BR adiabatic runaway theory (the technological constraint g1 from
of length t = tf /Ndiv , where the reactor input is continuous and Table 2).
differentiable. By considering constant feeding characteristics over The compromise between the productivity/low-cost goal and
every such interval, that is constant Fi (t) and cD,in ,i (t), over ti−1 ≤ t < safety requirements, in the presence of parameter uncertainty, is
ti , i = 1, . . . , Ndiv , ti = it (switching points), then an optimal pol- generally solved in literature either by minimizing a cost func-
icy [F1 , F2 , . . . , FNdiv ] (for constant cD,in ), or [FD,1 , FD,2 , . . . , FD,Ndiv ] tion with imposing a threshold for the failure probability, or by
(for both flow-rate and cD,in adjustable) can be derived by maxi- minimizing the failure probability with imposing a threshold for
mizing the reactor performance index in the presence of various the cost/productivity requirements. In the present study, the first
technological and safety constraints. alternative is adopted, by keeping the optimal solution outside the
The reported results have been obtained by minimizing tf confidence region of the runaway boundaries. Such a solution cor-
under constant cD,in = 5.82 mol L−1 , and constraints G1 and G2 responds to a minimum probability of intersecting the safety limits
of Table 2, by imposing a certain productivity threshold, i.e. for an assumed confidence level.
nPAA ,f = cPAA ,f Vf ≥ 0.42 mol (equivalent to 45% PAA net yield), as fol- For the fixed terminal time case, the SBR optimization criterion
lows: has been formulated as an augmented objective Lagrange function,
G. Maria, A. Dan / Computers and Chemical Engineering 35 (2011) 177–189 185

of the form: (i.e. a 68% confidence level for ıFc =


Fc , or a 95% confidence level
[F1 , F2 , ..., FNdiv , cDin,1 , cDin,2 , ..., cDin,Ndiv ] = arg Max ˚(Fj , cD,in,j , ), for ıFc = 2
Fc ). Estimations of Fc () and
F2c are made in every iter-
ation of the solution search, using the correlation models (2–3).

Ndiv
Normal fluctuations have been considered for the operating param-
˚ = wPAA nPAA,f − wG,1 G1 − wG,2 G2 − wg,1 g1 − wg,2 g2,j
eters, with the standard deviations of
Ta = 3 K,
[D]in = 0.1 M,
j=1

[P]o = 0.05 M. It should also be mentioned that a more elabo-
s.t. Fj ≥ 0, cD,in,j ≥ 0, j = 1, ..., Ndiv .
rated formulation of the optimization objective function in the
(5)
presence of multiple sources of uncertainty will require comput-
ing the expected value of ˚, that is: ˚ = P()˚(FD,in (t, ), )d
By considering both input variables during the search, a higher 
SBR flexibility is expected. The constraint functions are displayed (Terwiesch et al., 1994).
in Table 2. The chosen penalization weights in the objective function
Only parameter uncertainty is considered here, other uncer- depend on the relative importance given to each constraint. For
tainty sources being neglected. This is accounted for by including instance, by adopting the weighting factors wPAA = 1, wG,1 = 10,
the probability of violating safety limits by the running profile wG,2 = 100, wg,1 = 0.1 for the technological constraints g1 , G1 –G2
[Fj ,cD,in,j ]. Thus, the constraint g2,j penalizes any rapprochement of one scales the variations of all terms of ˚ in the [0,1] range. As
the tried arc flow-rate Fj () on the (known) runaway boundary the sum terms of constraints g2 represent failure probabilities for
Fc (), in the probabilistic sense. By assuming a normal distribu- every time-interval of the batch, a weighting factor wg,2 = 1 assigns
tion for the parameter random fluctuations, ∼N(, ¯
2 ) (where  the same importance to the safety requirements and productivity

denotes the nominal point), a normal distribution of the runaway goal (wPAA = 1). In contrast, a small safety weighting factor, e.g.
¯
2 ). The
boundary Fc () can be assumed, of the form Fc ∼N(F̄c (), wg,2 = 0.1, can be adopted when the productivity goal is preva-
Fc
constraint g2,j requires maximizing the distance dj from the run- lent, and the safety problems are not very stringent. The last choise
ning point to the critical conditions, which means minimizing the can be an option when the distance from the operating profile to
probability that Fj ≥ Fc , i.e.: the runaway boundaries is large and an effective process control is
  implemented on the SBR.
g2,j = 1 − PN (Fc − dj < Fc < Fc + dj ) , (Minimum); By solving Eq. (5), optimal molar feed flow-rates
 
dj = Fj () − Fc () , j = 1, ..., Ndiv
[FD,1 , FD,2 , . . . , FD,Ndiv ] with D-solution are obtained. However,
(6)
practical implementation allows separate treatment of the feed
where PN denotes the Normal cumulative distribution function of flow-rate F(t) and inlet solution concentration cD,in (t), offering a
the random variable. As small values for g2,j ≥ 0 are expected (usu- larger degree of freedom. In fact, only a finite number of D feeding
ally in the range of 0–5%), such a policy tries to keep the operating solutions of known concentrations can be prepared before starting
trajectory outside the confidence region of the critical conditions the batch, such that the SBR feed might be switched during

Fig. 3. Optimal feed flow-rate policies [F(t), cD,in (t)], temperature profile, and PAA yield dynamics, for the semi-batch reactor for acetoacetylation of pyrrole. Optimal policies
ensure Max nPAA ,f using Ndiv = 8 feed flow-rate arcs, and fixed tf = 145 min, under the technological constraints g1 , G1 –G2 (Table 2), and without any safety constraint. The
checked operating conditions are the followings: (a) isothermal operation with constant cD,in = 5.82 mol L−1 , realizing nPAA ,f = 0.4358 moles (net PAA yield 0.4664); (b) non-
isothermal operation with constant cD,in = 5.82 mol L−1 , realizing nPAA ,f = 0.4708 moles (net PAA yield 0.5150); (c) non-isothermal operation with optimal found cD,in = 3 mol L−1
(derived from cD,in ∈ [3,4,5,6] mol L−1 possibilities), realizing nPAA ,f = 0.4820 moles (net PAA yield 0.5305).
186 G. Maria, A. Dan / Computers and Chemical Engineering 35 (2011) 177–189

operation according to a certain policy. In the present study, literature for the same isothermal SBR and tf . The non-isothermal
feeding solutions of concentrations cD,in,j ∈ [3,4,5,6] mol L−1 have operation solution has been derived also by keeping a constant low
been considered. An effective MINLP procedure was used to derive cD,in = 3 mol L−1 level over the batch (solution (c) in Fig. 3). Even if
the feeding policy, searching for Fj continuous variables, and cD,in,j lower cD,in will lead to higher optimal F(t) levels, and vice-versa,
discrete variables (j = 1,. . .,Ndiv ). Constraints g3 and g4 from Table 2 policies (b) and (c) are not fully equivalent, neither as temperature
have not been included in the objective function, but are only used profile (sharper for (c) case), nor as realized PAA yield (better for
for checking the severity of the operating conditions. As recom- (c) policy). Operation al lower cD,in clearly induces more flexibil-
mended in literature, the number of time-arcs was considered ity in managing the feed flow-rates without escalating the process.
Ndiv = 8, larger values offering an improved solution flexibility, but Isothermal operation, even if desirable for safety reasons, is how-
with higher computational and implementation costs. Other fixed ever impossible to be implemented due to parameter variation
operating parameters concern the final batch time tf = 145 min, random character and very sensitive exothermic reactions.
the initial and feeding temperature To = Tin = 50 ◦ C, and initial
composition of Table 2. As MINLP solver, the adaptive random
search MMAMI of Maria (2003) has been used, together with the 4.2.2. Feeding policies (d1–d2)
random differential evolution algorithm of Storn (2008) with our In a second approach, one determines the optimal feeding policy
adaptation for MINLP problems. Fj , j = 1,. . .,8, with considering technological and safety constraints
SBR optimization has been performed in several alternatives, g1 –g2 , G1 –G2 (Eq. (5)), and keeping cD,in = 5.82 mol L−1 constant over
and the derived feeding policies are compared to derive general the batch. In the alternative (d1), a small weight wg,2 = 0.1 has
conclusions. been assigned to the safety requirements, the productivity goal
being prevalent. In the alternative (d2), the safety requirements
have been included in the SBR optimization with the same weight
4.2.1. Feeding policies (a–c) wg,2 = 1.0 as for the productivity goal. The results, comparatively
In a first approach, one determines the optimal feeding policy presented in Fig. 4, reveal a continuous decline in the PAA yield as
Fj , j = 1,. . .,8, with considering only technological constraints g1 , the safety requirements are stronger in the optimization criteria.
G1 –G2 , without including the safety constraints g2 (Eq. (5)), and This decline might be acceptable for a small weight of constraints
keeping constant cD,in = 5.82 mol L−1 over the batch. Optimization g2 , the realized yield of 0.5079 being close to the maximum of
solution was derived under isothermal (a) or non-isothermal con- 0.5150. However, the probability of the risk that Fj (t) > Fc is in the
ditions (b). The results, presented in Fig. 3, reveal the superiority range of 1.3–9%, intersecting during one-time-arc the 95% confi-
of non-isothermal approach, with a realized PAA yield of 0.5150 dence region of the runaway boundaries (corresponding to random
vs. 0.4664, being superior vs. all other feeding policies reported in fluctuations of ±2
Fc , and ±2
 ). In a more prudent sub-optimal

Fig. 4. Optimal feed flow-rate policies F(t), with cD,in = 5.82 mol L−1 , compared to the critical values Fc (t) , temperature profile, and PAA yield dynamics, for the non-isothermal
SBR (STD(Fc ) =
Fc = standard deviation of Fc ). Optimal policies ensure Max nPAA ,f using Ndiv = 8 feed flow-rate arcs, and fixed tf = 145 min, under the technological constraints
g1 , G1 –G2 , and safety constraints g2 (Table 2). The imposed safety constraints are the followings: (b) non-isothermal operation with constant cD,in = 5.82 mol L−1 , under
constraints g1 , G1 –G2 (wg,2 = 0, g2 -neglected), realizing nPAA ,f = 0.4708 moles (net PAA yield 0.5150); (d1) non-isothermal operation with constant cD,in = 5.82 mol L−1 , under
constraints g1 , G1 –G2 , with safety constraints activated of weight wg,2 = 0.1, realizing nPAA ,f = 0.4657 moles (net PAA yield 0.5079); constraint g2 is violated with a time-
interval probability of g2 = [0.013, 0.018, 0.025, 0.032, 0.059, 0.020, 0.090, 0.033] (see Table 2 footnote); (d2) non-isothermal operation with constant cD,in = 5.82 mol L−1 , under
constraints g1 , G1 –G2 with safety constraint activated of weight wg,2 = 1, realizing nPAA ,f = 0.3587 mol (net PAA yield 0.3593); constraint g2 is violated with a time-interval
probability of g2 = [0.003, 0.005, 0.001, 0.009, 0.012, 0.019, 0.001, 0.008] (see Table 2 footnote).
G. Maria, A. Dan / Computers and Chemical Engineering 35 (2011) 177–189 187

Fig. 5. Optimal feed flow-rate policies F(t), with optimal cD,in = 3 mol L−1 (derived from cD,in ∈ [3,4,5,6] mol L−1 possibilities), compared to the critical values Fc (t), temperature
profile, and PAA yield dynamics, for the non-isothermal SBR (STD(Fc ) =
Fc = standard deviation of Fc ). Optimal policies ensure Max nPAA ,f using Ndiv = 8 feed flow-rate arcs,
and fixed tf = 145 min, under the technological constraints g1 , G1 –G2 , and safety constraints g2 (Table 2). The imposed safety constraints are the followings: (c) non-isothermal
operation with constant cD,in = 3 mol L−1 , under constraints g1 , G1 –G2 (wg,2 = 0, g2 -neglected), realizing nPAA ,f = 0.4820 moles (net PAA yield 0.5305); (e1) non-isothermal
operation with constant cD,in = 3 mol L−1 , under constraints g1 , G1 –G2 , with safety constraints activated of weight wg,2 = 0.1, realizing nPAA ,f = 0.4745 moles (net PAA yield
0.5201); constraint g2 is violated with a time-interval probability of g2 = [0.001, 0.007, 0.001, 0.016, 0.012, 0.014, 0.019, 0.007] (see Table 2 footnote); (e2) non-isothermal
operation with constant cD,in = 3 mol L−1 , under constraints g1 , G1 –G2 , with safety constraint activated of weight wg,2 = 1, realizing nPAA ,f = 0.4520 moles (net PAA yield 0.4889);
constraint g2 is violated with a time-interval probability of g2 = [0.001, 0.001, 0.002, 0.003, 0.006, 0.008, 0.010, 0.005] (see Table 2 footnote).

operation, the realized SBR productivity is lower (i.e. 0.3593 yield), runaway happens with a certain frequency. More complicated sim-
but the probability of violating the safety limits for Fj is very small, ulations of accident consequence scenarios (AIChE, 2000; Maria,
i.e. in the range of 0.1–1.9%. 2007), together with a detailed economic analysis of the batch in
connection to the market requirements might help in taking the
appropriate operating policy decision.
4.2.3. Feeding policies (e1–e2)
For instance, the optimal policy (d2) that severely met the safety
In a third approach, one determines the optimal feeding pol-
constraints for a cD,in = 5.82 mol L−1 level, leads to a decrease with
icy [Fj , cD,in,j ], j = 1,. . .,8, with including all technological and safety
24% of the PAA productivity but consuming 47% less D reactant
constraints g1 –g2 , G1 –G2 (Eq. (5)). In the alternative (e1), a small
than the policy (b) that ignores the safety constraints g2 [the con-
weight wg,2 = 0.1 is assigned to the safety constraints, in contrast
sumed
NdiD number of moles over the batch is given by the formula
to a high weight wg,2 = 1 used in alternative (e2). The results, com- v
(tf Fc
j=1 j D,in,j
)]. Similar conclusions can be derived for operation
paratively presented in Fig. 5, reveal two interesting conclusions.
Even if a continuous decrease in the PAA yield is obtained as more at low cD,in = 3 mol L−1 level, the decline in the PAA productivity for
importance is paid to the safety constraint (from 0.5305, to 0.5201, the operating policy (e2) being of 6% vs. the policy (c), but realizing
and 0.4889 for policies c-e1–e2), running at low cD,in,j = 3 mol L−1 is an economy of 16% of raw-material D (not mentioning that policies
more advantageous, offering a higher yield comparatively to oper- e1–e2 fully fulfil the safety and flexibility requirements).
ation at a higher cD,in = 5.82 mol L−1 (policies b-d1–d2). The paid
price consists in manipulation of larger feeding flow-rates, and in an 5. Conclusions
insignificant increase of the thermal peak (with ca. 1 K). As another
conclusion, the probability of violating the safety limit confidence Optimal operating policy of a non-isothermal SBR can be sub-
region is much smaller at lower cD,in (in the range of 0.1–1%). Con- stantially improved in terms of safety requirements if accurate
sequently, operation at lower co-reactant D feeding concentrations safety limits of the feasible parametric space can be determined.
is more economic and safer. It should also be observed that, for this Safety constraints can be included in the optimization objective
case study, the operating severity declines for all policies towards function in a probabilistic sense, by accounting for the parameter
the end of the batch, due to requirements to limit the amount of uncertainty and penalizing risky rapprochements on the runaway
formed by-products. boundary confidence region in the parametric space.
In fact, a decision on choosing a suitable SBR feeding policy The present work illustrates how the safety operating limits can
that realizes a satisfactory productivity vs. safety compromise is be model-based derived using a generalized sensitivity criterion
not easy to be taken, depending not only on raw-materials, prod- (MV), and how to be included during SBR optimization together
ucts and other batch costs, but also on consequence costs when a with the associated uncertainty. A concrete example is provided
188 G. Maria, A. Dan / Computers and Chemical Engineering 35 (2011) 177–189

for the SBR used for the acetoacetylation of pyrrole with diketene Fotopoulos, J., Georgakis, C., & Stenger, H. G., Jr. (1994). Uncertainty issues in the
in homogeneous liquid phase. While previous studies tried to maxi- modeling and optimization of batch reactors with tendency models. Chemical
Engineering Science, 49, 5533–5547.
mize the isothermal reactor performances by model-based deriving Froment, G. F., & Bischoff, K. B. (1990). Chemical reactor analysis and design. New
various optimal feeding policies of the co-reactant under semi- York: Wiley.
empirically derived technological constraints, the present work Grewer, T. (1994). Thermal hazards of chemical reactions. Amsterdam: Elsevier.
Hugo, P., & Steinbach, J. (1986). A comparison of the limits of safe operation of a SBR
approaches the non-isothermal SBR with two manipulated input and a CSTR. Chemical Engineering Science, 41, 1081–1087.
variables (feeding solution concentration and flow-rate), by also Jolkkonen, M. (2000). Standard thermodynamic properties of chemical substances.
considering the safety requirements. Results suggest that a reason- In The Royal Institute of Technology in Stockholm. Boca Raton: CRC Press.
Kadam, J. V., Schlegel, M., Srinivasan, B., Bonvin, D., & Marquardt, W. (2007). Dynamic
able compromise between reactor productivity vs. safety can be optimization in the presence of uncertainty: From off-line nominal solution to
realized, avoiding risky rapprochements of the running point on measurement-based implementation. Journal of Process Control, 17, 389–398.
the runaway boundary confidence band in the parametric planes. Lee, J., Lee, K. S., Lee, J. H., & Park, S. (2001). An on-line batch span minimization and
quality control strategy for batch and semi-batch processes. Control Engineering
The study also suggests that operation at low concentration levels
Practice, 9, 901–909.
of the inlet solution is preferred, even if higher feed flow-rates have Lenz, R. W. (1967). Organic chemistry of synthetic high polymers. New York: Inter-
to be manipulated. Thus, a considerable increase in both SBR pro- science Publishers.
ductivity and safety is obtained, without a considerable variation Loeblein, C., Perkins, I. D., Srinivasan, B., & Bonvin, D. (1997). Performance analysis
of on-line batch optimization systems. Computers & Chemical Engineering, 21,
of the thermal regime. S867–S872.
Even though it is computationally more intensive, the model- Loeblein, C., Perkins, J. D., Srinivasan, B., & Bonvin, D. (1999). Economic performance
based MV-method to derive the process safety limits is generally analysis in the design of on-line batch optimization systems. Journal of Process
Control, 9, 61–78.
applicable, offering less conservative predictions. Regular determi- Lopatin, E. B., Popov, V. V., Epshtein, N. A., Mikhaleva, L. M., & Makarov, Y. N. (1992).
nation of the safety operating margins, exhibiting a high thermal Kinetic and thermochemical characteristics of diketene based reactions. Khimiko
sensitivity, is necessary when changes in the systems’ geometry, Farmatsevticheskii Zhurnal, 26, 76–78.
Maria, G., Terwiesch, P., & Rippin, D. W. T. (1996). Reliable on-line kinetics iden-
input conditions, raw-materials or catalyst’s characteristics occur. tification for a semi-batch process by parallel use of shortcut transfer of
The interest in accounting for precise safety limits during the SBR information rule and recursive estimators. Chemical Engineering Communica-
optimization is justified by the economic implications of getting a tions, 143, 133–147.
Maria, G., & Heinzle, E. (1998). Kinetic system identification by using shortcut
higher productivity by setting the nominal operating policy in the techniques in early safety assessment of chemical processes. Journal of Loss
vicinity of the safety limits. Prevention in the Process Industries, 11, 187–206.
The study also points-out that parameter random fluctuations Maria, G. (2003). ARS combination with an evolutionary algorithm for solving MINLP
optimization problems. In M. H. Hamza (Ed.), Modelling, Identification and Control
have to be considered not only in determining the new operation
(pp. 112–118). Canada: IASTED/ACTA Press.
set point of the reactor, but also when evaluating the confidence of Maria, G. (2004). A review of algorithms and trends in kinetic model identification
safety region according to the parameter distribution. This double for chemical and biochemical systems. Chemical and Biochemical Engineering
uncertainty in the operating parameters and critical conditions of Quarterly, 18, 195–222.
Maria, G. (2007). Chemical process quantitative risk analysis and modelling of accident
reactor operation has to be considered when determining the opti- consequences. Bucharest: Printech Publication. (in Romanian).
mal operating policy of the SBR in respect to a certain performance Maria, G., & Stefan, D. N. (2010). Variability of the risk operating limits with the cat-
criterion. The optimal operating solution has to be kept inside the alyst properties in a fixed-bed vapour-phase catalytic reactor for nitrobenzene
hydrogenation. Journal of Loss Prevention in the Process Industries, 23, 112–126.
safety region, without intersecting the confidence region of the Martínez, E. (2005). Batch-to-batch optimization of batch processes using the STAT-
runaway boundaries, instead preventing intersection of runaway SIMPLEX search method. In 2nd Mercosur Congress on Chemical Engineering Costa
boundaries themselves. Such a solution corresponds to a minimum Verde, Rio de Janeiro, Brasil, paper #20.
Mönnigmann, M., & Marquardt, W. (2003). Steady-state process optimization with
probability of intersecting the safety limits for an assumed confi- guaranteed robust stability and feasibility. AIChE Journal, 49, 3110–3126.
dence level and for an adopted weight of the safety constraint in Nakano, K., Kosaka, N., Hiyama, T., & Nozaki, K. (2003). Metal-catalyzed synthesis of
the objective function. stereoregular polyketones, polyesters, and polycarbonates. Dalton Transactions,
4039–4050.
Since in many cases the most severe running conditions corre-
Park, S. W., & Himmelblau, D. M. (1980). Error in the propagation of error formula.
spond to a higher productivity, such a modified solution is expected AIChE Journal, 26, 168–172.
to offer a higher degree of safety in operation with the expense of a Rahman, A. K. M. S., & Palanki, S. (1996). On-line optimization of batch reactors with
two manipulated inputs. Computers and Chemical Engineering, 20, S1023–S1028.
certain decline in reactor productivity and of some supplementary
Reid, R. C., Prausnitz, J. M., & Poling, B. E. (1987). The properties of gases and liquids.
computational steps to derive the confidence of the safety limits. New York: McGraw-Hill.
Ruppen, D., Bonvin, D., & Rippin, D. W. T. (1997). Implementation of adaptive optimal
operation for a semi-batch reaction system. Computers & Chemical Engineering,
Acknowledgement 22, 185–199.
Sato, T., Morita, N., Tanaka, H., & Ota, T. (2003). Synthesis and radical polymerization
of ethyl alpha-acetoacetoxymethylacrylate. Die Makromolekulare Chemie, 191,
This work was partially supported by CNCSIS-UEFISCSU, project 2599–2607.
number PNII-IDEI 1543/2008–2011 “A nonlinear approach to con- Seinfeld, J., & McBride, W. L. (1970). Optimization with multiple performance crite-
ria. Application to minimization of parameter sensitivities in a refinery model.
ceptual design and safe operation of chemical processes”. Industrial and Engineering Chemistry Process Design and Development, 9, 53–57.
Shah, S., Fischer, U., & Hungerbühler, K. (2003). A hierarchical approach for
the evaluation of chemical process aspects from the perspective of inher-
References ent safety. Transactions of the Institution of Chemical Engineers (London), 81(B),
430–443.
Abel, O., & Marquardt, W. (2003). Scenario-integrated on-line optimisation of batch Shah, S., Fischer, U., & Hungerbühler, K. (2005). Assessment of chemical process
reactors. Journal of Process Control, 13, 703–715. hazard in early design stages. Journal of Loss Prevention in the Process Industries,
AIChE. (2000). Guidelines for chemical process quantitative risk analysis. New York: 18, 335–352.
Center for Chemical Process safety of the American Institute of Chemical Engi- Srinivasan, B., Palanki, S., & Bonvin, D. (2002). Dynamic optimization of batch pro-
neers. cesses II. Role of measurements in handling uncertainty. Computers and Chemical
Bonvin, D. (1998). Optimal operation of batch reactors—A personal view. Journal of Engineering, 27, 27–44.
Process Control, 8, 355–368. Srinivasan, B., & Bonvin, D. (2003). Convergence analysis of iterative identification
Brendel, M., Mhamdi, A., Bonvin, D., & Marquardt, W. (2003). An incremental and optimization schemes. Proceedings of the American Control Conference, 3,
approach for the identification of reaction kinetics. In Preprints of the ADCHEM 1956–1960.
2003, International Symposium on Advanced Control of Chemical Processes Hong Srinivasan, B., Palanki, S., & Bonvin, D. (2003). Dynamic optimization of batch
Kong. processes I. Characterization of the nominal solution. Computers and Chemical
Chen, M. S. K., Erickson, L. E., & Fan, L. (1970). Consideration of sensitivity and param- Engineering, 27, 1–26.
eter uncertainty in optimal process design. Industrial and Engineering Chemistry Stoessel, F. (2008). Thermal safety of chemical processes. Risk assessment and process
Process Design and Development, 9, 514–521. design. Weinheim: Wiley-VCH.
G. Maria, A. Dan / Computers and Chemical Engineering 35 (2011) 177–189 189

Storn, R. (2008). Differential Evolution—Scientific Report. Berkeley Univer- Varma, A., Morbidelli, M., & Wu, H. (1999). Parametric sensitivity in chemical systems.
sity, CA: International Computer Science Institute. http://www.icsi.berkeley. Cambridge, MS: Cambridge University Press.
edu/∼storn/index.html. Watanabe, N., Nishimura, Y., & Matsubara, M. (1973). Optimal design of chemi-
Takamatsu, T., Hashimoto, I., & Ohno, H. (1970). Optimal design of a large com- cal processes involving parameter uncertainty. Chemical Engineering Science, 28,
plex system from the viewpoint of sensitivity analysis. Industrial and Engineering 905–913.
Chemistry Process Design and Development, 9, 368–379. Wen, C. Y., & Chang, M. T. (1968). Optimal design of systems involving parameter
Terwiesch, P., Agarwal, M., & Rippin, D. W. T. (1994). Batch optimization uncertainty. I & EC Process Design and Development, 7, 49–53.
with imperfect modelling: A survey. Journal of Process Control, 4, 238– Westerterp, K. R., & Molga, E. J. (2004). No more runaways in fine chemical reactors.
258. Industrial and Engineering Chemistry Research, 43, 4585–4594.

Das könnte Ihnen auch gefallen