You are on page 1of 15

Chemical Engineering Journal 97 (2004) 87–101

Optimal control of product quality for batch nylon-6,6 autoclaves
Marcel Joly a , José M. Pinto a,b,∗
a Department of Chemical Engineering, University of São Paulo, São Paulo SP 05508-900, Brazil
b Department of Chemical Engineering and Chemistry, Polytechnic University, Brooklyn, NY 11201, USA
Received 29 November 2002; accepted 29 April 2003

Dynamic optimization problems are usually solved by transforming them into nonlinear programming (NLP) models with either sequen-
tial or simultaneous approaches. In this paper, the potential and limitations of both procedures in solving a general batch nylon-6,6 autoclave
optimization model are evaluated and discussed. Nylon-6,6 is a highly value-added good produced in large-scale under many different spec-
ifications. Highly nonlinear behavior, lack of on-line measurements of polymer qualities and several disturbances inherent to this process
motivate its optimal operation. In addition, the operation during the finishing stage of the batch polycondensation is crucial since it largely
determines the final product molecular weight and quality. Results show that both strategies can be successfully applied to solve the dy-
namic optimization problem only after an expressive level of investments in terms of modeling implementation. However, the simultaneous
strategy has the advantage that specification constraints can be directly enforced in the model, thus generating better and more robust results.
© 2003 Elsevier B.V. All rights reserved.
Keywords: Optimal control; Nylon-6,6; Nonlinear programming; Optimization

1. Introduction On the other hand, limited up to the middle of the 1980s
not only by computational restrictions but decisively due
1.1. Preliminaries to little availability of published information, detailed sim-
ulation and optimization of nylon-6,6 reactors had to be
Due to the ever-increasing global competition environ- postponed to the 1990s. Till then, kinetic information was
ment, several factors have encouraged better monitoring and practically reduced to the pioneers, nevertheless limited re-
control strategies of end-use quality variables in nylon-6,6 sults from Ogata [2,3]. Additionally, there was no published
polymerization processes. Nylon-6,6 is a highly value-added data on vapor–liquid equilibrium to model water vaporiza-
good produced in large-scale under many different specifica- tion. This information is very crucial since the final degree
tions. Moreover, it is well-known that process operation play of polymerization depends on the water concentration in
an important role in both reactor design and control. In par- the melt, and this in turn depends on the pressure and
ticular, the highly nonlinear behavior of the process, the fre- temperature, as well as the composition of the liquid phase.
quent obstacles in on-line reactor measurements of nylon-6,6 As a consequence of subsequent studies about kinetic
end-use qualities as well as the several disturbances inher- modeling produced ever since [4,5], the recent years have
ent to this process (e.g., water content in the feed) have em- been characterized by an increasing number of modeling
phasized the importance of reactor modeling. Furthermore, developments and refinements reported in the open liter-
this problem is often embedded in higher-level problems, as ature that addresses nylon-6,6 production in a number of
the plant-wide dynamic optimization aiming to produce op- different reactors. Thin or continuous wiped film reactors
timally integrated planning and scheduling policies, whose modeling have been studied by Steppan et al. [6] and Choi
benefits are well documented in other fields of chemical en- and Lee [7]. Giudici et al. [8] addressed the catalytic and
gineering [1], given the important ease that relatively sophis- homogeneous process conduced in twin-screw extruder
ticated plant-wide dynamic models involving thousands of reactors. Batch autoclave models and control strategies
states can now be handled through the existing technology. have been investigated by Russell et al. [9], whereas prac-
tical approaches to the control of final product quality in
semi-batch reactors are discussed by Yabuki et al. [10] and
∗ Corresponding author. Tel.: +1-718-260-3569; fax: +1-718-260-3125. Clarke-Pringle and MacGregor [11]. Tubular reactors un-
E-mail address: (J.M. Pinto). der two-phase flow conditions without external catalyst are

1385-8947/$ – see front matter © 2003 Elsevier B.V. All rights reserved.

[12]. nonlinear programming (NLP) under Mii molecular weight of a i–i segment constraints remains as basis of major solution techniques within the polymer molecule currently used by schedulers [1]. (dimensionless) Following this trend. Pinto / Chemical Engineering Journal 97 (2004) 87–101 Nomenclature xi mole fraction of i in the liquid phase Ci concentration of i (mol/l) yi mole fraction of i in the vapor phase Cp specific heat capacity of the liquid phase (cal/(g ◦ C)) Greek letters CT total concentration (mol/l) ε extent of reaction Eapp activation energy for degradation µi mass fraction of i in the liquid phase reaction j (cal/mol) ρ liquid phase density (g/l) Ekapp apparent activation energy (cal/mol) ωie mass fraction of i in the vapor phase Happ apparent enthalpy of reaction (cal/mol) Subscripts Hvi heat of vaporization of i (cal/mol) A amine end groups kapp apparent forward reaction constant C carboxylic end groups (mol/total mol h) HMD hexamethylene diamine monomer kj kinetic constant for degradation i components A.14]. the solution of state equations specially with un- Pvap solution vapor pressure (psia) stable components and the definition of the finite element Pisat pure component vapor pressure (psia) lengths for discretization [16]. so they R gas constant may offer a practical approach given the current state of Rj rate of reaction j (mol/(l h)) technology. the main deficiency of direct search T reactor temperature (K) comes from the fact that the nonlinear model. these mod- (g/(h psia)) els. Qheat heat input (cal/h) On the other hand.88 M. unit (l) conventional use of decomposition schemes for dynamic n1 fraction of polymer molecules systems implies in additional difficulties associated with that are HMD the resulting NLP. L. the existence of local so- reaction j (K) lutions in optimization problems is frequent. that may be v vent flow rate (g/h) partially circumvented by running the optimization algo- vMi molar rates of vaporization of i (mol/h) rithm several times with different starting points for decision V volume of liquid phase (l) variables. which in most part rely on (strongly) nonlinear systems Kapp apparent equilibrium constant for of differential-algebraic equations (DAEs). However. as the presence of algebraic equations NH2 amine end concentration (mol/g) (the index problem). As pointed out in [13]. must be solved at each T0j reference temperature for degradation algorithm iteration.M. which is T0 reference temperature (K) normally a time-consuming task. Firstly. Furthermore. the high dimensionality of discretized P reactor pressure (psia) problems. Vˆ specific volume (l/kg) The contribution of this research is twofold. Joly. HMD. Moreover. MW average molecular weight of the NLP gradient-based approaches usually have to cope with polymer problems such as numerical evaluations of derivatives (Ja- nw moles of water/mole of fundamental cobians and Hessians) and feasibility issues [15]. while alternative optimization ap- K0 reference apparent equilibrium proaches. derivative-free (or direct search) meth- r stoichiometric ratio of monomers ods usually overcome some of these drawbacks. Concomitantly. SE and W reaction j (mol/total mol h) j reactions k0 reference apparent forward rate constant L polymer link (mol/total mol h) SE stabilized end groups k0j reference forward rate constant for W water molecule degradation reaction j (mol/total mol h) k0j kinetic constant rate for degradation reaction j (h−1 ) K overall mass transfer coefficient reported by Giudici et al. have motivated polycondensation reaction the study and development of solution techniques. J. Nevertheless. as neural networks (NN) or multi-way projection constant (dimensionless) to latent structures (MPLS) are living important devel- Mi molecular weight of i opments [13. two general and comprehensive NLP-based methodologies for . C. NLP opti- M0 molecular weight of one unit of the mization technology still requires ability to deal with some polymer chain well-known algorithmic limitations.

In this process. amine end groups on either HMD or polymer batch nylon-6. obtained from thermal decomposition of nylon-6. the chemistry involved is essentially the same. Overview of the process sections take a more in-depth view at the inherent features of the nylon-6. Nylon-6. since it captures (–CONH–) and SE refers to a cyclic or stabilized end group.6 salt (hex. C is a carboxyl end ment is satisfied by an autoclave reactor (Fig. the loss of HMD through vaporization (and groups for the sake of simplicity. In order to achieve This paper is concerned with the production of nylon-6. optimization results and evaluation of the solu- since they affect the temperature dependence of the result- tion techniques are presented in Section 4. the industrial objective is to maximize synthesized in a multistage process in which the product is productivity regarding the required end-use product quality. The fundamental model chain react with carboxylic end groups on either adipic acid is introduced in Section 2.6 polymerization chemistry as well as at the In short. 1. maximizing nylon-6. tions for constrained batch run times.6 or poly(hexamethylene adipamide) is usually In most cases. degra- differential-algebraic optimization problem (DAOP). the extent of reaction by direct amidation of adipic acid and hexamethylene di. (3)). W is a water molecule. Nylon-6.6 production may be carried on concentration (NH2 ). the equilibrium is shifted to- amine in a batch process carried out in autoclave reac. The chemistry of nylon-6.6 reaction. The polyamidation and C. L is a polymer link damental process model presented in [9]. sensitive to the stoichiometric feed ratio (r) of monomers A proximately stoichiometric proportions. which can perform as set-point programs Degradation C → SE + W (1) for control systems. Although nylon-6. Therefore. In this paper.6 autoclave process. Vaporization require- A is an amine end group (–NH2 ). 1).6 autoclave model Table 1 are proposed in order to provide useful tools for defining Nylon-6. M. the dominant features of the process behavior. the next sub- 1. according to Eq. we are concerned with quite a number of different ways from an engineering stand. Some cludes the paper and summarizes future directions. The end-use prod- an exact 1:1 stoichiometric ratio between –NH2 and –COOH uct quality is defined by the MW level and the amine end groups.3.6 the desired end-use product quality. other objectives should be considered due to the amethylene diammonium adipate). Another complicating factor is that MW is very by molten adipic acid and hexamethylene diamine in ap. which must be kept near unity. to form cal background of the approaches used to solve the previous a polymer chain link and a molecule of water. In order to provide a better understanding of the problem. Joly. the reactor feed is composed Eq. Section 3 outlines the mathemati- monomer or polymer chain. we have selected the fun- group (–COOH). Second.6 production under product specifica- point. wards completion by continuous vaporization of water (see tors (Fig. . Pinto / Chemical Engineering Journal 97 (2004) 87–101 89 optimal control of a generic batch nylon-6. in benefit of a simpler formulation. which a–a/b–b type commonly described in terms of functional is also volatile. Eventually. (1) and (2)) must be considered [6] simulation. J. Also.M.2.6 reaction model [9] operating profiles. It is worth to note can be treated as a second-order reversible reaction of the that since A is supplied by HMD (two A per HMD). As basis for the proposed work. As will be discussed in need of blending with some off-spec product or even when the next subsection. The overall reaction model the subsequent deviation of r from unity) represents an im- for the nylon-6. However. Model dation reactions (Eqs. Section 5 con- ing average molecular weight of the polymer (MW).6 polymerization model in Table 1. (3). must be above 99%. as in [9]. 2) where there Fig. where portant drawback to this approach [9]. the computational performance L → SE + A (2) of the two approaches are compared and discussed in terms Polyamidation A+C↔L+W (3) of algorithmic structures and solution methods. this salt has the advantage of presenting batch time does not need to be minimized. we neglect that reaction.6 batch autoclave is shown in Table 1. authors [12] also consider a more complete model by adding the degradation of SE (into CO2 and a Schiff base) to the 1.

Also. Ultimately.17] since it sets the quality on molecular size). .6 autoclave. but also by reactor fouling. increase of ings of the reactor pressure (P). the initial heating phase. [9] is se- perature and complementary heat is supplied to the system lected since it incorporates a comprehensive and well-known by the steam jacket to force the polymerization reaction to phenomenological structure supported by first-principles begin. is continuously vented off the reactor with the objective (c) spatial variations are neglected.90 M. in which water (b) thermal effects of the vapor phase are neglected. of the final product. the fundamental portant amount of the necessary heat to raise the batch tem. the reactants are not yet well-mixed and along with some empirical relations that are tailored for the the pressure above the reacting solution is maintained higher system.e. nylon-6.6 batch autoclave model from Russell et al.M. Therefore. 2. of allowing high MW polymer levels. the model relies on the following major since the vent valve is kept closed during this initial phase. Joly. In the first one. are neglected. these measurements include read. that also supplies heat to the reaction medium. Once charging is concluded. the reaction rate does not depend has been extensively studied [6.6 process (g) equal reactivity (i. boiling and a well-mixed medium is obtained. since MW and NH2 can be solely measured after some measures are frequently taken to facilitate removal of the batch run. At this stage. Typically. the model is complete with respect to than the vapor pressure of the liquid phase to avoid prema. as inert gas bubbling through the melt. When a sufficient rate of poly- merization in terms of reacted HMD is achieved. The valve opening starts the boiling phase. In short. the vent (a) thermal effects of the side reactions (Eqs. assumptions [9]: the ratio r remains constant. an aqueous effects. The reactor the desired end-use product quality should normally be operation can be suitably divided into three distinct stages met under time limitations in order to reduce degradation [9]. heat addition is compli. An extra amount of HMD can also be added in order to compen. Therefore. (h) saturated steam is fed into the reactor jacket and heat cated not only by the increased viscosity in the medium due losses are negligible. flow rate and temperature of the liquid phase (T). all complementary information (kinetic and physical prop- ture loss of volatile amine feed monomer HMD. equimolar mixture of HMD and adipic acid is charged to the autoclave that has been previously purged of oxygen. steam jacket pressure. Furthermore. water.. J. As a basis for the proposed study. transport in the liquid phase becomes more effective due to (e) P (reactor pressure) is a manipulated variable. because of the high polymer melt viscosity during the latter During the batch operation. the rate of water removal becomes monitored by on-line measurements of secondary process a critical factor for the polymerization extent. the exothermic salt formation provides an im. In his phase. product quality is usually stages of polymerization. is a valve for venting vaporized water and a steam jacket to the high polymer MW. erties). Moreover. 2. vent agitation rate or application of partial vacuum. mass (d) the vapor phase molar balance is neglected. longer occurs. Pinto / Chemical Engineering Journal 97 (2004) 87–101 Fig. The finishing (f) water and HMD leave the liquid phase according to their phase starts when water has been removed and boiling no vapor–liquid equilibrium relations. Schematic representation of the batch nylon-6. variables. (1) and (2)) valve is then opened to allow the escape of vaporized water. This important stage of the nylon-6. Fundamental model sate losses of this reactant by vaporization.

provided that condensate flow tion.2  x  W Happ = 7650 tanh[6. K = 25 000 g/(h psia) dt v=0 otherwise dV CC ωie = −V(R3 + R1 ) vM i = v. where T0 = 473. HMD dt Mi dV CL = V(R3 − R2 ) vM A = 2vHMD M dt dV CW = V(R3 + R1 ) − vM w Vapor–liquid equilibrium (VLE) relations dt dV CSE dρ V yHMD xHMD P sat = V(R1 + R2 ) and = −v =α . k2c . as a consequence.6 autoclave model—mass and energy model [9] Mass and energy balances Mass transfer model dV CA v = K(P vap − P) if P vap > P = VR2 − (VR3 + vM A) .99 cal/mol K R T T0   √   xw K0 = exp 1 − 0. where Eapp = 30 kcal/mol. xHMD = CHMD /CT a For rate constant—k1 . the manipulation of the jacket rates are much lower than that of the polymerization reac.55)] + 8. Ekapp = 21. R = 1. R = 1.15 K.5n1 (CA + CC + CSE ). T in K dt ρCp ρCp V ρCp V ρCp V (T − 45. xC = CC /CT . xL = CL /CT .065     Ekapp 1 1 kapp = k0 exp − − . Due to the small mass of the vapor phase and its fast rate is measurable.6 autoclave model—kinetics [9] Reaction Kinetics Degradation C → SE + W R 1 = k1 CC    CA L → SE + A R2 = CL k2 + k2c CT Polymerization     kapp CL C W A+C↔L+W R3 = CA C C − CT Kapp Kinetic relations a       Eapp 1 1 ki = k0i exp − − .58(tanh[50(xW − 0.05xC )}. pressure. k0i : 0.99 cal/mol K R T T0i        Happ 1 1 Kapp = K0 exp − − . k2 .4 kcal/mol. Original references on the kinetics and sure control loop is fast (i. affects the liquid phase dy.15 K. T0i : 566 K. While P2 relies on Table 3 The batch nylon-6.1)] − 1)(1 − 30. Since only the overall reactor The model is characterized by a set of DAEs and nonlin- pressure in the vapor phase. size distribution in terms of the Flory distribution [18]. (Table 3) involves the vapor pressure term (Pvap ) given as vap vap vap Qheat . k0i : 0.M.15)2 . P.4 kcal/mol R T T0 k0 = exp{2. CHMD = 0. T0i : 578 K. k0i : 0. HvHMD = 12. ear expressions for constitutive equations.32 h−1 . dynamics.52)] + 6500 exp − − 800 0. where CT = CA + CC + CL + CW + CSE .2xW ) 0. Finally. In order to describe the molecular the modeling may be found in [9].005 h−1 . from (h) the heat input to the reactor. is fairly simple.06 h−1 . rized in Tables 2–6. has negligible dynamics) and physical property calculations as well as specific features of is sufficiently accurate. M. where α = HMD dt dt yW xW PWsat dT Hp R3 HvHMD vM HvW vM Qheat 7. xW = CW /CT . (g) As part of the overall formulation.45 − 4. i = W. becomes simply the heat given up by the condensing the maximum between P1 and P2 . (d) holds and (e) emerges from the fact that the pres.724T 2 =− − HMD − W + with HvW = . R = 1.47 exp − (8. where T0 = 473. Joly. the mass transfer model is postulated. Assumption (a) comes from the fact that side reaction steam and.99 cal/mol K.5(xW − 0.e. J. Pinto / Chemical Engineering Journal 97 (2004) 87–101 91 Table 2 The batch nylon-6.45 tanh[25(xW − 0. and are summa- namics.. T0i : 578 K.55 − 0. (b) takes place. Pj .

Pinto / Chemical Engineering Journal 97 (2004) 87–101 Table 4 The batch nylon-6. J. Joly.M.6 autoclave model—quality variables [9] t .92 M.

are obtained being removed through vaporization whereas C is subject through VLE relations. In this sense.6 autoclave model—physical properties [9] Physical properties Specific volumes (V’s in l/kg and T in K) Vˆ W = (1. This ap. P2 = xW PW sat + xHMD PHMD sat 760P1 /14. Indeed.15])−1 Vˆ sol = Vˆ W [µW + 0. P1 corresponds to the Henry’s tribution model for a–a/b–b polymerization [18].13 − 0. tion reaction may change from C towards A during the pro- notes the total number of polymer molecules and parameter cess as a consequence of the HMD vaporization.925(1 − µW − µL )] + Vˆ poly µL Vapor pressure relations (P’s in psia and T in K)  sat    760PW 1668. [9] consider the amount of HMD that has left the those at the extreme of polymer molecules [9].7 T Specific heat capacities (Cp in cal/(g ◦ C) and T in K) Cpw = 0. The mass fraction of water and HMD in the modified approach accounts for the fact that HMD is the vapor phase. (1) and (2)).4726 × 10−7 (T − 273.1 wt. Table 6 Modified molecular weight distribution model [9] Limiting group C (r < 1) Limiting group A (r > 1) (1 − rε)2 r(1 − ε)2 n1 = n1 = r + 1 − 2rε r + 1 − 2rε t (VCC )0 − VCC − VCSE (VCA )0 − VCA − 0 (2ωHMD e v/MHMD ) dt ε= ε= t (VCC )0 (VCA )0 − 0 (2ωHMD v/MHMD ) dt e t (VCC )0 (VCA )0 − 0 (2ωHMD e v/MHMD ) dt r= t r= (VCA )0 − 0 (2ωHMD v/MHMD ) dt e (VCC )0 .00052[T − 273. respectively. which ward/backward reactions occur fast enough to allow the dis- combined with a suitable modification in Flory’s molecu.96106 × 10−5 (T − 273. as shown in Table 6. where log10 vap = − 10. in which the term in parenthesis de. M0 (1 + r) 106 CA MAA (VCA )0 − 0 (2ωHMD e v/MHMD ) dt + MCC (VCC )0 MW = and NH2 = with M0 = t 1 + r − 2ε ρ (VCA )0 − 0 (2ωHMD e v/MHMD ) dt + (VCC )0 Table 5 The batch nylon-6. equations must be properly considered. is given by CW /CL .7 T − 45. To overcome these effects in such mode that it is not possible to distinguish between that are unaccounted for in Flory’s original derivation. the mass balances were to degradation by forming stabilized end groups that can no postulated in terms of total number of A amine end groups longer react (Eqs. two sets of n1 denotes the fraction of polymer molecules that are HMD. result in an expression for the HMD concentration.5(1 − ε2 )] hypothetical ideality condition for the liquid phase during Also based on the modified Flory’s molecular weight dis- vap most part of the process.21 PHMD sat = exp[−1. HMD.4ε2 + 0.09. P2 ).15)2 + 1. In short.99824 × 10−4 [T − 273.164316 + 6237. the quality law behavior of the solution for water content not greater variables are expressed as functions of the modeled process vap than 0. lar weight distribution relations (outlined below) proposed the integral term accounts for the total amount of vaporized in [9].44(0.6321 × 10−6 [T − 273.15] − 2.0046715 − 1.002807 − T −1 )] log10 =− + 7. Rus- amine end groups at the extreme of HMD molecules and sell et al. ωW e and ωe HMD . since the limiting group for the polymeriza- CHMD (see Table 2).% [2].15)3 Cp = µW Cpw + (1 − µW )[0.96681 14.15) − 5. reactor as if it has never been there by assuming that the for- parent difficulty is circumvented by assumption (g).2037 × 10−8 (T − 273.15    vap vap nW 3050 vap P vap = max(P1 . tribution to reach its new value instantaneously. However.997092 + 3.15]2 )−1 Vˆ poly = (1. The calculation of P1 involves nw that states and their respective initial feed conditions.

Z(t)] accomplish the optimization by applying variable parame- x. this class. veloped and refined [15. x the time-invariant decision variable vector. the first attempt is then applied to solve a master NLP in an outer loop to to solve the optimal control problem employed concept of update the control actions.1. 3. the key idea is to discretize the control that address such class of problems. M.Z(t) terization approaches. ZU (UL .25–28]. c the design equality con. U(t). namely sequential or UL ≤ U(t) ≤ U U . the calculus of variations introduced in the 1960s [22]. profiles into piecewise polynomials on finite elements. For conve. This paper is concerned with the third class. methods aim to transform the infinite-dimensional dynamic ˙ Z(t)=F [x. Indeed. U(t). t the time.2. it is possible to Min φ[x. nience. Since the cally posed as an initial value ODE model as in Eq. In this paper. An optimization routine and control variables by Luus [21]. have been continuously de- where φ is the objective function. the optimization problem is trans- formed into a two-point boundary value problem (TPBVP) 3. (4): decision variables of the nylon-6. c[x. which fall on three cat. ZL . In short. the dy. we have straint vector. J. Z(t)] = 0. indirect inner loop.16. ZL ≤ Z(t) ≤ ZU control parameterization strategy [23] and simultaneous or collocation strategy [24]. U(t). Sequential strategy (control) profile bounds. implemented both strategies as discussed next. Joly. See [19] for review.6 process model can be identified as continuous time-invariant. the process model previously introduced is generi. two general approaches. (4) optimization problem into a finite-dimensional NLP.M. problem [16]. xL . Z(t) (U(t)) the state (control) profile vector. Scheme of the control parameterization approach to the model (4).U(t). Fig. or direct methods. Numerical solution of dynamic optimization problems with DAEs. Z(t). xL ≤ x ≤ xU . t] ∀t ∈ [0. tf ]. This approach constitutes the basic idea of the indirect methods and performs very well for unconstrained A brief overview of general-purpose methods for the nu.t. and this yields function and [20] and was extended to include constraints on the state gradient values for the NLP solver. the control U(t) is approximated as piecewise Fig. Pinto / Chemical Engineering Journal 97 (2004) 87–101 93 3. There is an extensive literature on numerical strategies In this approach. . Solution strategies this variational approach. In an egories: dynamic programming based approaches. in an initial value problem (IVP) for fixed control profiles namic programming method was first described by Bellman using standard DAE solvers. In summary. the state and sensitivity equations are integrated and direct methods. direct s. merical solution of the DAOPs that include a system of especially with the addition of profile inequalities in the ordinary differential equations (ODEs) follows. UU ) the state 3. but the solution of the TPBVP remains difficult. xU the decision variable bounds. In In this figure. problems. 3 summarizes the method. Within Z(0) = Z0 .

The stability issues for the simultaneous . where. the optimization procedure. This step yields z(y). This approach leads to a decomposition of the original DAOP into two subproblems. which is the numerical integration based on derivative evaluation. for example. t)] deter- T T T T SQP algorithm structured as shown in Fig. or gradient-free ap. t).t. t). Therefore. which constitute the ma. the major deficiency of sive quadratic programming (SQP) [31]. In principle. u2 (v. uTN (v. Therefore. be stable [37]. The ODEs are discretized v target using. but a justifies in part the extensive work that has been produced in number of equality path constraints. Based on results from the integration step. . intercepts and exten- sions of each finite element can be taken as optimization variables in the master NLP. the A-stable orthogonal collocation on s. This becomes an important drawback to the use of sensitivity approaches (a) An IVP subproblem. y] + ω err MW err NH2 of both state and control profiles. we Fig. in such mode that analytical derivation quential procedure consists of the following steps: of the gradients would be a hard task.94 M. 4. sensitivities are of the DAE system formed by Z(t) ˙ and c for the fixed not derived and the optimization is carried out through an control y = [x . features of the main approaches for computing gradients in jor state constraints of the problem are imposed as equality DAE systems may be found in [35]. However. Functional structure of the implemented sequential strategy. Neither this proaches such as direct search or stochastic search [32. Here. in the general case. intermediate solutions. Reviews on both theoretical and computational targets for the product qualities.3. Section 2 involves a complex mathematical dependence sured violations are taken into account through penalization between the objective (as well as part of the constraints) of objective in the master NLP. Joly. . inequality constraints and/or boundary conditions on state variables. the solution of the IVP might be infeasible or too costly to converge at intermediate trial points [16]. . which can be included order to evaluate the sensitivities in mathematical program- into the DAE model to be solved simultaneously [29].1) is ensured to Nevertheless. In this paper. u1 (v. original objective of (4) to account for deviations from the Consequently.6 process model presented in these constraints are evaluated at the outer loop and the mea. the master approach lies on the treatment of profile constraints as well NLP can be solved with either gradient-based approaches as bounds and elimination of costly and possibly infeasible such as generalized reduced gradient (GRG) [30] or succes. the fundamental process model synthesized in with respect to the optimization parameters [19]. . the slopes. (5) finite elements [25] and the resulting algebraic equations are yL ≤ y ≤ yU added to the NLP formulation having the polynomial co- efficients as optimization variables. Zq − Zq = errq q = MW. for which the In the simultaneous strategy. The ming [34]. The main advantages of this tor and γ and η are real exponents. with the requirement for continuity of the con- trol from one finite element to another. the nylon-6. NH2 . efficient control parameterization usually re.6 quality specifications. Conversely. Simultaneous strategy (4). 4.36]. ω is a weighting fac. Discretization using control parameterization allows ob- jective function and point constraints of (4) to be expressed as functions of a finite number of time-invariant variables v (or parameters). Excluding bounds quires the gradients of the objective function and constraints on profiles. since the number of control variables u and parameters remains small even for very complex models. use piecewise linear control policy under N equal sampling intervals. strategy nor the preceding one (Section 3. This in turn (4) does not contain any inequality path constraint. cannot be directly enforced [25]. In addition. Pinto / Chemical Engineering Journal 97 (2004) 87–101 linear functions (hence the term control parameterization). the ODE model is solved only once over product qualities targets in (5).M. as the nylon-6. we have included the term ω errMW err NH2 to the such as B-spline collocation can also be applied [24. However. this strategy lies in its explosive dimension. The major advantage of this strategy lies in the reduced dimensionality of the NLP. J. and the control y. the resulting se. Alternative approaches γ η As seen. in which the constraints may in prin- ciple be handled in a number of ways [29]. the DAE solution and opti- bounds are determined directly from (4): mization converge simultaneously through the discretization γ η Min φ[z(y). (b) A master NLP subproblem in terms of y. point constraints. mined by the optimization step (or by the initial guess at the first iteration) using the initial state condition of 3. which characterize the basis functions.33].

K (12) coefficient at the beginning of each element. .19]. NE. l = 1. . NE. the relationship between t and t becomes as fol. strategy are depicted in [37. .j tij − tik j=1 k=1. t ∈ [0. State and con- trol variable profiles. tf ] is R(til . . . . quantities. . The end point of its element and provide an initial condition for term ziK+1 (t) denotes a (K + 1)th order (degree < K + 1) the next polynomial ziK+1 (t). . . where three finite elements in each element at the K shifted roots (til ) of a Kth degree (NE = 3) and two inner collocation points (K = 2) are rep. . K (7) through the following recursive formula:   where tj is the relative. . . . J. . αi j=0 and the locations of the orthogonal Legendre roots in each i = 1. ziK+1 (t) at the interior knots αi . . K. αi and αi+1 . NE. . i = 1. 5. . The expression for ϕ˙ j (tl ) can be easily calculated off-line i = 1. . t0 = 0 is imposed to ensure the existence of a j = 0. . . j = 0. normalized location of the jth (j > 2tl − K j  =0. i = 1. The calculation of ϕ˙ ij (til ) can be simplified [25] and lows: Eq.j tj ϕ˙ j (tl ) = 2  K . such as temperature or concentration. . j denotes k = 0. i = 2. . NE. which is as follows. . where j=0 element i is bounded by two knots. Eqs. . The difference in the orders is due to the initial is characterized by a high degree of nonlinearity associ. . l = 1. where clearly α0 = 0 and αNE+1 = i = 1. ini- defined in (4) into a nonlinear program we first discretize tializing a profile or interpreting solution profiles. normalized and partitioned into NE finite elements. . l = 1. til ] = 0. K (11) element i are mapped to the following points: with z10 = Z0 . . condition for Z(t). . . re. Also. . As pointed out in [25]. . Note that ziK+1 (tij ) = zij and uiK (tij ) = ated to the element placement constraints. . M. . . . this property of the Lagrange procedure can be very sensitive to starting points and must polynomials is very desirable. NE. zil . uil . . . ψij (t) = (9) k=0. Scheme of finite element collocation for K = 2 and NE = 3. the uij . . . Pinto / Chemical Engineering Journal 97 (2004) 87–101 95 Fig. The time domain t ∈ [0. Joly. By mapping each finite element to normalized time. 1]. . As a result. The location of the collocation points are chosen residual) equations: to correspond to the shifted roots of an orthogonal Legen- K  dre polynomial of degree K. Legendre polynomial leads to the following collocation (or resented. tij = αi + tj (αi+1 − αi ). 0) shifted root of an orthogonal Legendre polynomial of de. This in turn location method.j K  t − tk K  K  ϕj (t) = (13) t − tik tj − tk uiK (t) = uij ψij (t). .M. . . say with z10 = Z0 . . . . j + 1.j (tj ) + j  =0. K. ing problems states and controls often represent physical In this paper. the formulation polynomial. j–1. zil . . tj − tj Kj  =0. . . as follows: spectively: K  K  K  t − tik zi0 = zi−1. 5. Consider Fig. (10) is rewritten as: t = αi + t(αi+1 − αi ). (8) and (9) into the ODE model and enforcing the ODE ements [25]. . NE K  ϕ˙ j (tl ) for t ∈ [αi .j These equations extrapolate the polynomial zi−1 K+1 (t) to the where k = 0.j ϕj (t = 1). NE. ziK+1 (t) = zij ϕij (t).38]. Z(t) and U(t) are approximated over An additional set of equations enforces continuity for each element i by Lagrange interpolation polynomials. To convert the DAOP becomes useful when providing bounds on variables. . . .j (tj ) gree K. with αi = αi+1 –αi . Replacing the ODEs by applying orthogonal collocation on finite el. uil . αi ) = zij ϕ˙ ij (til ) − F [x. . since for chemical engineer- require a careful initialization [16. K (10) 1. In addition. . ϕij (t) = (8) tij − tik j=0 j=0 k=0. i = 1. we are concerned with the orthogonal col. til ] = 0. . Although differentiable and polynomial whereas uiK (t) denotes a Kth order (degree < K) piecewise polynomial approximations for the state profile . αi+1 ] (6) R(til ) = zij − F [x.

. Cuthrell and Biegler [25] develop a nonfully rig. zi0 = zi−1j ϕj (t = 1). K  (14) the resulting optimization problem is as follows [37]: z10 =Z0 . NE. τ). . K. R2 = l k2 + k3 . m can be calculated through Eq. θ) Mass transfer model: v. l = 1. xL ≤ x ≤ xU . . uil . an appropriate placement j=0 of these elements has a severe impact on this number and i = 2. zil ] i = 2. this enforcement is only effective equations can be regarded as major deficiencies. hmd. (8) and (9). . . which is based on the en. as saturation of state variable pro. This work for index 1 problems. . c[x. The fundamental nylon-6. a different error veloping a more reliable strategy where the approximation control strategy on higher index variables (usually the con- error can be controlled directly in parameter optimization. Therefore. τ) Happ = f(w. . Numerous strategies exist in the approximation theory lit. i = 1. T. k0 = f(w. Hvw = f(T). se. .zil s. c. four-point collocation. 4) e . developed by Brenan and Petzold [40]. zi0 = zi−1j ϕj (t = 1). K0 = f(w. J. the accuracy of the approximation however has not been discussed. index 3 problems. high index systems don and Biegler [26] where many of the previous concepts (i. 3. Incorporating knot placement in optimization i = 1.t. vM i = f(P. Joly.1. two-point collocation. . . . R(ti. j=0 Min φ[x. . three-point collocation. . R3 = ac − τ τ Kapp ki = f(T). l. l. NE. particularly since the process model may have nondif. states and controls at tnc with Eqs. se) and hmd = f(a. w. w. . . τ).e. . i = 1. (R(tnc )m αi ) ≤ ε. trolling the integration error. . αi = tf . uil .uil . K. a. τ). τ). 2 m=1 3. files. xL ≤ x ≤ xU . the continuous and piecewise approxi. .uil . zil ] = 0. i = 1.  NE erature for selecting a suitable distribution of knots when fi. c. since for higher index systems the is extended by Cuthrell and Biegler [38] for discontinuous accuracy of the higher index variables is affected due to control and refined by Vasantharajan and Biegler [27] by de. . Kapp = f(T. NE. ZL ≤ zil ≤ ZU i = 1. the approximation error. methods are [37]: index 1 problems.zil ε is a user-specified error tolerance. NE. . .l ) = 0.t. the following NLP 2 m=1 formulation is obtained: where R(tnc i ) denotes the residual at the noncollocation m Min ϕ[x. . Indeed. . ωHMD = f(T. index 2 or higher) can be considered not only by con- can be applied. c. in which the introduction of nondifferentiabilities and highly nonlinear As pointed out in [37]. reduction of the error order. w. NE (15) ferentiable solution profiles. ential equation. The residuals R(tnci ) s. . c[x. Therefore. M 1 in fact is the procedure to be followed. zil ] = 0. w. which is an index 1 DAE Table 7 Functions of (17)   a      kapp lw Kinetics: R1 = k1 c. which is not index 2 problems. Pinto / Chemical Engineering Journal 97 (2004) 87–101 can also be used. zil ] point tnc of the mth state profile at the finite element i and x. . kapp = f(T). Cp = f(T. as discussed in the i 2 next section. Russell and Christiansen [39] present i=1 an excellent review where numerical comparisons are in- cluded. . . W. but by additionally choosing ditions for optimal control problems are systems of DAEs a suitable method. 4. HMD.6 process model forcement of the following constraints in the optimization is rewritten as in (17) and Table 7. se. nite elements are used. αi ) = 0. R(ti. . i = A. Choosing a number of elements so that their K  lengths are “sufficiently small” is the first idea to minimize z10 = Z0 . In order to control the overall error within every element The integral terms in Tables 4 and 6 may be properly re- for all of the differential equations in the simultaneous moved by adding an extra state and a corresponding differ- approach. ri = i 2 (R(tnc )m αi ) ≤ ε.96 M. In these cases.M. . formulation: mation described in this section is sufficient in the present M 1 work. . NE. x. . hmd. U L ≤ uil ≤ U U . c. uil . 2. τ). Furthermore. . from theoretical properties Extension to optimal control problems is explored by Logs. l = 1. U L ≤ uil ≤ U U .3. So far. encountered with parameter optimization problems [27]. where τ = f(a. trols) is required. . . ZL ≤ zil ≤ ZU (16) orous method based on error equidistribution. we consider [37]. (11) by extrapolating the i = 1. . uil . . The major distinction is that optimality con. The minimum requirements for these and for certain conditions. αi ≥ 0. w. . NE. stability and accuracy problems can occur.l .

θ) where Vˆ sol = f(T.971 d se dθ V0 (l) 1802 = R1 + R2 = −v. Joly. 2. for instance. 4). τ). This allows two-point collocation to find the so-   a  lution because the higher index portion does not propagate R1 = k1 c. = −(R3 + R1 ). Note that constraints on nylon-6. w. τ). hmd. 4). T0 (K) 425 dt dt (17) dT Happ R3 HvHMD vM =− − HMD In (17). w.t. the lower case letters denote the product of vol- dt θCp θCp ume and concentration. se. =R2 −(R3 +vM A ). the volume of liquid phase can be com- = . Table 8 pression for P and Qheat from active bounds: Initial conditions for the nylon-6. M. a. w. τ). c. w. a = CA V and so on. c. Pinto / Chemical Engineering Journal 97 (2004) 87–101 97 because one differentiation step is needed to obtain an ex. se) and hmd = f(a. c.6 problem State variable Initial value max l(tf ) P. i = 1. HvW vM Qheat − W + . Model (17) has as initial condition 4(0) = 0. the process model (Tables 2–6) was firstly i = A. . R2 = l k2 + k3 . se. l. Kapp = f(T.1. 3. c. kapp = f(T). as Tables 2–6: 4. 6 and 7 Fig. since the θCp θCp amount of vaporized HMD is zero at the beginning of the d4 e 2ωHMD v batch. w. a.M. θ).6 qualities must only be active at one point of the trajectory where Ri (i = 1. v. 6. T. Typical simulation results for concentration of functional groups during a batch run. NH2 = f(a. MW = f(c.Qheat CA0 (mol/l) 6 da dc CC0 (mol/l) 6 s. J. 3) are defined as: (t = tf ). ωHMD e = f(T. 2. Results and. vM In order to provide an accurate starting point for the simul- i = f(P. = R3 + R1 − vM w. Model simulation ki = f(T). HMD. l. W.2 dl dw = R3 − R2 . taneous strategy.     τ kapp lw R3 = ac − τ Kapp 4. se. CL0 (mol/l) 0 dt dt CW0 (mol/l) 11. θ)e. l. τ) simulation for a horizon of 3 h for typical initial conditions summarized in Table 8 was accomplished by MATLAB® on where τ = f(a. simulated by using typical operational values as estimates k0 = f(w. The model Cp = f(T. 4. w. Happ = f(w. a Pentium III 700 MHz platform in 0. Hvw = f(T). τ). hmd. τ). K0 = f(w. Figs. dt MHMD puted by incorporating the algebraic constraint V = θ Vˆ sol .77 CPUs. into the index 1 portion of the system. If necessary. CSE0 (mol/l) 0 dt dt ρ0 (kg/l) 0. for the input heat and reactor pressure profiles.

Based on the previous results.2. plishes the simultaneous based optimization. the opti. The choice for cretize the horizon. the number of finite elements used to discretize (16) was mization of Eq.7) for a nominal batch run time of Table 9 where computational details. In the first one.5. For two collocation points (K = 2). mension and CPU times are also tabulated. On the reactor pressure profiles are shown in Fig. Joly.33 7. expensive computation and frequent failures were ments (NE = 10) have proven to be sufficient for obtaining encountered during the optimization. Typical simulation results for the autoclave model parameters during a batch run and initial estimate for the heat input. The sequential strategy employs a piecewise linear con. sired polymer satisfying product specifications (MWtarget = The results of the simultaneous optimization are given in target 13 479 and NH2 = 49. variables. satisfactory profile approximations (ε = εr = 10−3 ) within The objective is to maximize the production of the de. Model optimization approach is used to implement the feasible path optimiza- tion in MATLAB® . finite element to another. display typical simulated trajectories for the key process trol policy under 30 sampling intervals of the same length. the time-invariant modeling system GAMS 2.50 [42] was used in order to im- model. As mentioned other hand. (16) is then accomplished by relaxing the supported by comparisons between simulated and optimized aforementioned knot placement constraint. the GRG-based solver CONOPT2 [41] accom- in the preceding section.722 . disconti- lies on the simulated response as initial point for the NLP nuities introduced by max operators (Table 5) may be easily solver. two models are sequentially solved within eliminated through smooth approximation techniques.5 and ω = 7 × 10−5 . The the simultaneous approach. Table 10 reports Table 9 Simultaneous approach performance and related information (ε = εr = 10−3 ) Number of constraints Number of variables Number of iterations CPU time (s) Objective = max [CL ] (mol/l) Evenly spaced knots based model 1327 1379 451 20. acceptable computational effort. the resulting quadratic program In order to start the optimization procedure. 7. 7.M. The initial guesses for the input heat and reactor with the requirement for continuity of the control from one pressure profiles are also presented in Fig. Here. 10 finite ele- cedure. an evenly spaced set of fixed knots are used to dis. Moreover. J. determines the search direction whereas a line search pro- gies make use of the same initial estimate for heat input and cedure along the direction defines the next point. the simultaneous strategy also re. The coefficients of (5) were cho- sen as γ = 1. both strate. Pinto / Chemical Engineering Journal 97 (2004) 87–101 Fig. including the model di- 3 h. Here. η = 0.16 7. reactor (solid line) and vapor (dashed line) pressures. The SQP 4. profiles. Without this pro. 7.722 Direct error enforcement based model 1760 1639 430 38. plement the model and its solution method.98 M.

enforced in a straightforward way. . this feasible approach is of the horizon (product depuration). i = 1. J.20 58. although substantially larger.3 0.7 3.13 133 Simultaneous 7. note that how the heat was solved in a smaller number of iterations.9 1. Optimal profiles for key variables obtained with sequential (solid line) and simultaneous (dashed line) approaches.1 2. The reliability of the approximation may putational performance of the proposed sequential approach be checked by comparing the final collocation profiles with can be acceptable for industrial purposes. However. Despite appreciable changes in the ments in the objective value and the sequential procedure knot distribution (Table 10). this results from its starting point (note that the existence of multiple (local) optimal solutions. the com- pared in Fig.17 1. verified for the objective value. tional performance of the simultaneous strategy. 8. which is the solution of After 133 iterations there were no significant improve. those obtained by numerical integration. 9.323 0. computational cost is still less constrained than the first model (fixed knot location) Fig. Furthermore. 8.71 1.992 13479 46. Pinto / Chemical Engineering Journal 97 (2004) 87–101 99 Table 10 Initial and optimal knot distribution (simultaneous approach) αi . objective value remains the same).6 0.839 2. since the state variable bounds cannot be in Table 11.50 0. Table 9 summarizes model features and related computa- Finally. As can be mal profiles produced by the two strategies as well as the seen.699 1. Objec.4 2. 10 the heat input and reactor pressure opti.03 1. Here. M.0 Final 0 0. particularly in the final instants has stopped at this point. which were input peak provided by the simultaneous solution shifts in clearly more time-consuming tasks than the previous ones. However.8 Initial estimate 7. The resulting profiles for key variables are com. .57 0.72 2. . as shown in Fig. Nevertheless. the fixed knot model. NE + 1 Initial 0 0. this is not surpris- quires repeated and expensive solution of the overall DAE ing since the second model (variable-knot location).722 13490 50. . the variable-knot model initial estimate are compared.687 10918 64.2 1. albeit system (Table 11).M. no improvement was more time-consuming than the simultaneous one since it re. size and position from the sequential one as result of the Apparently.8 2. high due to the need of parallel convergence of the quality tive values and resulting product qualities are summarized parameters in (4).77 the initial (uniform) and optimal knot distribution. .86 3.7) Total CPU time (s) Sequential 7.5 1. Joly. in Fig.0 Table 11 Summary of the results Solution strategy Objective = max [CL] (mol/l) MW quality (target: 13479) NH2 quality (target: 49.

strategies can be successfully applied to solve the nylon-6. the im- namic optimization that takes into account operation condi. it is mandatory to batch nylon-6. therefore generates better and more robust solutions. ous approach. regarding the well-known diffi- lated results provided by the initial estimate for the control culties in finding global solutions for large-scale nonconvex profiles. the process model from ficients as well as the scaling of the controls is imperative Russell et al. proved performance of both strategies is usually achieved tions. Table 11 demonstrates the was also outlined and implemented within the simultane- effectiveness of both optimization strategies over the simu. In particular. product quality and only after an expressive level of investment in terms of energy consumption is then accomplished through two dis. As In counterpart. through the sequential strategy. the necessary background for the proposed study. As a consequence. Beyond the aforementioned dimension and nonconvexity 5. [9] was chosen due to the its suitable and to accomplish the optimization in competitive CPU times well-known phenomenological structure. 9. from the knot placement viewpoint is more restricted in tures of the sequential and simultaneous methods were then what concerns the approximation error. Concluding remarks issues.M. modeling implementation. Optimal profiles obtained through collocation (dashed) and the correspondent ones produced by numerical integration (solid line) of the solution. Pinto / Chemical Engineering Journal 97 (2004) 87–101 Fig. capabilities for determining optimal operating profiles.100 M. NLP models. . J. Nevertheless. this study is concerned with local optima de- termination.6 batch autoclave model is characterized by its large diversity in size order of the variable set (values In this contribution the dynamic optimization problem of ranging from ∼10−5 to 109 ). 10. Fig. Joly. Inherent fea. wear and tear of the equipment. discussed and formalized. A qualitative dy.6 autoclave process state variables are imposed in a straightforward manner and and introduced the fundamental process model. A simpler and reliable strategy comparative worst objective were not rare when running from [37] where the approximation error can be controlled the second formulation. the nylon-6. a suitable selection of the error term coef- a basis for the proposed study. Results have shown that both tinct NLP-based solution strategies. As a consequence.6 autoclave process has been presented and perform good variable and equation scaling or meaningless solved through NLP techniques in order to evaluate their solutions can be reported with the simultaneous approach. Therefore. Finally. Optimal profiles obtained through the sequential (long dashed) and simultaneous (short dashed) approaches and the initial estimate (solid line) for the controls. First. we took simultaneous strategy enables that quality constraints on an in-depth view of the batch nylon-6.6 Several concepts were discussed in order to introduce the batch autoclave optimization problem.

CA. Int. Polym.W. Dynamic optimization with state variable [42] A. C.A. Eng. W.F. The Scientific Press. Polym. Pierucci (Ed. R. Chem. L. Wolfe. Phys. [39] R.R. [41] A.F. Cuthrell.M. stochastic algorithms. Part 2. AIChE J.S.W. Res.H. Vasantharajan. Raman. Barton. Neuman. Eng. Eng.I. Sci. Feehery. Banga. 36 (1997) product quality control in semi-batch processes. 54 (1999) 3243–3249. Eng. 33 (1994) 2111–2122. film reactors for reversible condensation polymerization. Nascimento. Studies on polycondensation reactions of nylon salt. MacGregor. A. I.F. Dynamic optimization of batch reactors [10] T. T. M. [17] M.H. M. Eng. Sci. in: S. Molecular size distribution in linear condensation Manual for CONOPT Subroutine Library.F.T.. Macromol. Kendrick. ESCAPE 10. Phenomenological model and [32] J.-Rev. Chem. Comput. Marquardt.6 polymerization. R. Redwood City. dynamic optimization problems. Res.6 polymerization. A. Barton. Studies on polycondensation reactions of nylon salt. [35] V. NJ. Russell. Dahmen. Logsdon.S. Chem. Appl. Luus. [13] C. A. Fundam. Chem.D. Lab. Eng. Duchesne. M. Feehery. 58 (1936) 1877–1885. Macromol. Am. R. 22 (9) (1998) 1241–1256. J.S. R. D. Jacobsen. Eng. Doherty. Application of dynamic programming to high-dimensional This work was financially supported by FAPESP grants nonlinear optimal control problems. P. Flory. path constraints. Drud. [16] P. (1980) 415–421. 13 of process variable trajectories for batch processes. Comput. Global optimization of chemical processes using parameter adjusting. Russell. Adaptive mesh selection strategies for [15] W. Problems [12] R.H. 2000. [34] O. 21 (12) (1997) 1395–1409. kinetics. Sen. L. Clarke-Pringle. Vassiliadis. Multivariate analysis and optimization methods for batch reactor control profiles. [31] S. Transient and steady-state behavior of wiped. Amsterdam. L. Moro. Biegler.G. [22] L. Sullivan.E. 2252–2261. 1995. 35 (5) (1996) 1550–1555. Steppan. References Interscience. Part II. Lee. Ogata. multiscale dynamic optimization. Res. J.F. C.6 polycondensation process in 263–282. Nascimento. in: Proceedings of the Eighth IFIP for refinery operations. Banga. Comput.6 autoclaves. SIAM J. C32 (3–4) [28] W. . class of multistage dynamic optimization problems.O. ARKI Consulting and polymers. 20 (1996) S707– [6] D. 19 Livermore. New York. 33 (8) (1987) 1257–1270. Seider.O. in: State of the Art in Global Optimization: [9] S.R. Simultaneous strategies for optimi- 2333–2344. 59–80. Sci. P.C. Control Computational Methods and Applications. Chem. Res. Ogata. Blank. 67 (1998) 1573–1587. semi-batch polymerization reactors. 1962. differential/algebraic equations by implicit Runge–Kutta methods. Prog.G. Neural network based [37] J. Reference [18] P. of multistage dynamic optimization problems. Brooke. Eng. Giudici. A suboptimal control algorithm for constrained The equilibrium in the system of polyhexamethylene adipamine and problems using cubic splines. Simplified analysis of the performance of UCRL Preprint 95905. Eng. A. GAMS—A User’s path constraints. Res. 33 (1994) 2123–2133.). Pellicciotta. Chem. Luus. water. I.D. Steppan. Ogunnaike. Beiler. [33] E. Pinto / Chemical Engineering Journal 97 (2004) 87–101 101 Acknowledgements [20] R. Sargent. Accurate solution of differential-algebraic approach for optimization of industrial chemical processes. Solution of a Comput. Princeton. Biegler. Chem. 1977. A kinetic and equilibrium 862. Tech. Num. Industrial experiences with using adaptive stochastic algorithms. Biegler. Malone. Giudici. J. B.I. Nonlinear adaptive temperature optimal control. Tresmondi.F.6 polymerization in a two-phase flow tubular reactor. 33 (21) (1998) 3685–3702. Grid refinement in nylon-6. 51 (2000) 125–137.P. Problems with [7] B. H. Comput.A. 42 (1960) 52–67. Stable decomposition for dynamic [40] K. Res. optimization problems. J. J. Math. Yabuki. Comput. R. Chem. [38] J.F. Bellman. Ind. Dynamic optimization with equality path solving boundary-value problems. Sci. Eng.. Guide.L. [24] C. Sargent. Barton. W. J. 34 (1995) 1253–1266. Robertson.A. Planning and scheduling models optimal control package. M. Logsdon. J. Ind.I. Princeton University of product quality for batch nylon-6.H.S. Unified modeling for polycondensation criteria. The numerical solution of higher index optimization. Ind.R.T. Amon. 15 (1) (1978) constraints. Sargent. Domingues. Decomposition strategies for large scale of water. 29 (1990) 2012– [29] V. Chem. Tanartkit. Solution of a class 2020. Choi. J. Denmark. 43 (1961) 117–131. Princeton University Press.F.F. Gamkrelidze. Vassiliadis. Petzold. A system for large scale nonlinear optimization. P. 1986. Meeraus. Eng. 1996. Mathematical modeling of an industrial process of [36] T. 33 (1987) [27] S. Simultaneous optimization and solution [14] C. Ind. Control 52 (1990) 239– 98/14384-3 and 99/09897-4. Res. [30] P. Automatica 9 (1973) 601–603. J.D.6 salt in the presence [26] J. Eng. 24 (2000) 2259–2276. Denson.F. On the optimization of differential- [3] N. 28 (1989) 1628–1639. Pontryagin. The development of an efficient [1] J. Princeton. N. Joly. Chem. Appl. Eng. J. Macromol. Comput. G. L. MacGregor.H. Eng. [21] R.T. Carrasco. Eng. R. Malone.O. Han.A. Cuthrell. Macromol. 1998.E. The Mathematical Theory of Optimal Processes. Chem. Dynamic Programming. R. Ind. Chem.A. Feehery. Soc.S. Eng.M. Chem. Lee.T. Chem. without path constraints. rate of polycondensation reaction of nylon-6. [19] W. Boltyanskii. Chem. Eng. Nagasawa.D. L. Doherty. Brenan. RAND Co. [2] N. Ind. Pantelides. [23] R. Int. zation of differential-algebraic systems with enforcement of error [5] L. Conference on Optim. 11 (1976) Modeling of industrial nylon-6.F. D. Chem. a twin-screw extruder reactor. Res. 250. L. wiped-film polycondensation reactors. J. 24 (2000) 2303–2314. Eng. Biegler. L. L. C. Ind. M. Chem. Chem. control of multi-product. Giudici. Syst. CA.V. I. W. 47 (4) (1992) 851– [4] D. Anal. Lawrence Livermore National Laboratories. 15 (4) (1991) 273–281. Joly.E. general nonlinear programming problems. C.F. Ray. Press. Chem. M. Biegler. Guardani. MacGregor. Christiansen. 14 (10) (1990) 1083–1100. C.C. The Reduced Gradient Method.T. 1957. Development A/S.H. Eng. Elsevier.J. (1989) 49–62. L. for nylon-6. Pantelides. Scherbakoff. Eng. Binder. Ind. Rosen. V. J. II. 38 (1999) 2350–2363. Chem. NJ.T. Mishchenko. Dynamic simulation and optimization with (1992) 407–519. Evaluation of gradients for piecewise constant [11] T.F. Sci. Comput. The algebraic process systems. Sci. 31–36.C. Bagsvaerd.P. Pinto. Eng. Wiped film reactor model S712. Chem. 1962.W. inequality path constraints. Chem. Chem. model for nylon-6. Ind. [25] J. Chem. E. Superlinearly convergent variable metric algorithms for [8] R. Comput. W.S. Biegler. Chemomet. pp. Part I. Chem. R. Nascimento.R. 26 (2002) 205–212.