Vous êtes sur la page 1sur 9

chemical engineering research and design 1 1 7 ( 2 0 1 7 ) 168–176

Contents lists available at ScienceDirect

Chemical Engineering Research and Design

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

Improved prediction of alkyd reactors via


infrequent-delayed observations

F.C. Uzoh ∗ , O.D. Onukwuli


Department of Chemical Engineering, Faculty of Engineering, Nnamdi Azikiwe University, Awka, Nigeria

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

Article history: The problem of alkyd reactor polycondensation kinetics has been addressed. This was
Received 10 June 2016 enhanced through a novel mechanistic approach and discrete-delayed measurements. The
Received in revised form 24 August predictor provides adequate empirical-analytical basis for evaluating and operating alkyd
2016 batch reactor based on the present performance in order to monitor, control, correct the
Accepted 8 October 2016 reactor and ensure prompt termination of the batch such that the final product quality
Available online 20 October 2016 index is maintained within specification. The current proposal considers modification of the
existing models to include certain adjustable parameters whose values must be computed
Keywords: numerically such that the error measured between the actual and the predicted results are
Mechanistic model adequately minimized. The new approach was tested and validated in a reasonable number
On-line monitoring of reactions. The results were found to be reliable than the conventional approach. Further
Alkyd reactor detailed comparison of the results achieved through the error dynamics and RMSE revealed
Polyesterification a significant overall improvement in the studied reactor calibration problems.
State estimation © 2016 Institution of Chemical Engineers. Published by Elsevier B.V. All rights reserved.

1. Introduction literatures and no satisfactory solution has been reported. However,


previous works intended to address this problem only achieved cer-
Due to the stringent industrial specification that today’s alkyd resin tain levels of compromise between the observable outcome and the
must meet, the need for fine tuning their properties through a more desired result. The earlier alkyd kinetics studies report kinetics models
effective reactor operation is more acute than ever. For instance, in whose prediction accuracies degrade either at high or low conversions
order to produce alkyd resin offering a combination of both desired dry- (Lin and Hsieh, 1977; Aigbodion and Okeimen, 1996). In a consistent
ing time, chemical and mechanical properties, the molecular weight, effort to address this problem, works abound in literature where the
viscosity, conversion and some instance, undesirable side product state estimation methods (Alvarez and Lopez, 1999), system identifi-
concentration must be controlled within limits that would not dam- cation/recursive parameter estimation (Araromi and Adegbola, 2014)
age the industrial or end-use requirements (Lin and Hsieh, 1977). and nonlinear model predictive control (Valeel and Georgakis, 2002;
These alkyd resin property variables, in most instance, are difficult Eaton and Rawlings, 1992; Castellanos and Zaldo, 2015; Bachir et al.,
or impossible to measure on-line, specifically, viscosity. Thus, it is 2005–2008) were recommended for the treatment of the alkyd kinet-
very difficult to control the alkyd resin product quality when all ics prediction–estimation problem. The nonlinear state estimation
that is available is infrequent and delayed-offline laboratory measure- method due to its compatibility with discrete-time system seems to
ments. One major problem facing the study of alkyd polycondensation be a preferred choice for handling monitoring and control problems in
processes up to the present time is the difficulty in modeling the polycondensation processes. However, reports (Bachir et al., 2005–2008;
reaction kinetics (Lin and Hsieh, 1977; Aigbodion and Okeimen, 1996). Buruaga et al., 2003; Patton, 1962) show that the lack of reliable kinet-
Due to the complexity of alkyd polymerization mechanism, and ics models preclude direct application of the standard state estimation
other reasons related to the uncertainty of the monomer structure, methods such as extended Kalman filter (Bequette, 1991; Wang et al.,
detailed kinetics model capable of yielding accurate prediction of 1997) and Luenberguer-type model-based observer (Deza et al., 1992;
reaction motion and product quality in alkyd reactor is basically lack- Ciccarella et al., 1993) to alkyd reactors. Further attempts to overcome
ing. This problem has been well-emphasized in reaction engineering this problem led to the development of geometric estimation method


Corresponding author.
E-mail addresses: cf.uzoh@unizik.edu.ng, epio4real@yahoo.com (F.C. Uzoh).
http://dx.doi.org/10.1016/j.cherd.2016.10.016
0263-8762/© 2016 Institution of Chemical Engineers. Published by Elsevier B.V. All rights reserved.
chemical engineering research and design 1 1 7 ( 2 0 1 7 ) 168–176 169

(López et al., 2000; Hernández and Alvarez, 2003; Henerndez-Escoto alkyd resin from oxy-polymerizable gmelina seed oil, palm stearin and
et al., 2010) which attempts to reconstruct the kinetics uncertainties a published data (Nanvaee et al., 2013).
via a dynamic observer driven by discrete-delayed measurements. The
method basically involves designing a compensator network (or matrix 2. Mathematical modeling of the alkyd kinetics
of gain functions) that could be linked to the available kinetics models
to allow them perform within specification. The level of success attain- The following relationships are already known in coating industry for
able with this method remains a major breakthrough in the history of modeling the kinetics of polycondensation processes:
alkyd kinetics studies.
Optimal estimation is a process of extracting information con-
i The empirical third-order reaction rate model for conversion of acid
cerning a parameter or function from noise corrupted observation.
functional group (Lin and Hsieh, 1977; Henerndez-Escoto et al., 2010)
Considering model based and estimator based kinetic study of the
polycondensation process, it is seen that model based kinetic study
2
utilizes analytical models which often comes with unknown param- r =  (c, k, ce) :=k (1 − c) (ce − c) (1a)
eters. These unknown parameters, usually, are difficult to estimate
but can be estimated using experimental data. However, prediction of ii Viscosity-conversion expression in a form suggested by free-volume
the reactor parameter based on empirical-analytical model often intro- theory (Odetoye et al., 2011)
duces some unavoidable error in the results due to observed delay from
measurement time. The present value of the measured quantity actu- 
 =  (c, p ) :=a eb /(c −c) , p = [a , b , c ] (1b)
ally reflects value at certain time in the past since reasonable time is
devoted to experimentation. To accurately predict the reaction motion
iii Average molecular weight conversion—relationship (Patton, 1962)
over the entire course of the reaction requires an appropriate online
sensor capable of time ahead prediction of the reaction based on pre-
vious or current data. This online sensors are equally referred to as 1 + a c
M =  (c, p ) :=
dynamic data processor (DDP) (Henerndez-Escoto et al., 2010). This is b + c c + d c2
usually achieved using a well-constructed estimator which acts as a
process filter. In order to apply the technique of estimation theory to

a particular engineering problem, an appropriate mathematical model p = [a , b , c, d, ] (1c)
must be found for the system of interest. Most chemical engineering  2

processes, due to their highly non-linear nature, cannot be readily mod- The third-order rate equation k (1 − c) (ce − c) in its basic form
eled and thus defining control strategies become cumbersome. Since is nonlinear with respect to conversion. Thus, it is less possible to
the implementation of estimation algorithm is based primarily on sys- integrate the equation analytically. To overcome this problem, it was
tem state observation or function thereof, the use of well-constructed conceived in this study, that, instead of searching for an appropriate
filter based on third-order rate model have assisted in parameter esti- numerical integration scheme (which has also been proven difficult
mation and model development; given that a measure of the accuracy from earlier study (López et al., 2000)), a possible linearization scheme
of the model is directly obtainable (Hernández and Alvarez, 2003). could be imposed on the empirical third-order rate model. This was
Although the possibility of resolving the alkyd reactor fully developed in this paper via the following steps.
prediction–estimation problem using the geometric estimation
technique is well established in literature, yet, there are indications (i) Substitution of relevant parameter of the third-order rate model
that the accuracy of such a method fluctuates considerably in the reac- with an arbitrary parameter ␹.
tion path (especially in the low conversion range <90%). Other major (ii) Approximating ␹ with a linear integrable function with some use-
issues associated with the implementation of geometric estimation ful tuning parameters. This involves trying some basic linear,
method as inferred from the referenced articles include; trigonometric, polynomial, exponential and logarithmic function
for adequacy bearing in mind that:
i. Conditional stability of the iteration process: setting of the tuning (a) the most appropriate fitting function, must be able to predict
parameters required to achieve adequate convergence to the exper- ␹ = 1 − c/ce adequately, over the entire reaction course,
imental data is based on trial and error, and the stability of the (b) the approximation function must have tuning parameters that
computation process is not always guaranteed at all trial settings would allow for improvement of the model performance in any
of the tuning parameters. This makes the task involved in realizing given data set,
the convergence criteria a bit difficult. (c) the tuning scheme need to be systematic for easy implemen-
ii. Low dimensional integrity: the task involved in describing the non- tation.
linear gain functions which determine the performance of the
estimator progressively becomes cumbersome as the dimension of Considering these three points, an approximating polynomial of the
the problem increases. form q (t) = ˛tn + ˇt(n−1) + . . . + t + was considered most appropriate
for the following reasons;

In a decisive effort to simplify the alkyd reactor


prediction–estimation problem, a pure mechanistic approach to (1) The q (t) = ˛tn + ˇt(n−1) + . . . + t +
function is linear integrable,
the solution of the problem was introduced in this paper. An integral (2) Adequate fit to the value of ␹ resulting from the experimental data
conversion model was developed based on the empirical third-order can be obtained by mere selection of the parameters of the function,
rate model. The development of the integral conversion model follows (3) Adjustment could be made easily in any new data set.
from the knowledge that with proper adjustment of parameters, the
third-order rate model Eq. (1a) could be integrated analytically to The application of Eqs. (1a)–(1c) in modeling reaction kinetics for a
obtain a generalized integral conversion model of the form Eq. (7); given batch process requires that the unknown set of model parame-
which could be linked to existing kinetics laws to yield adequate ters k, p and p are known. Basically, p and p are evaluated by fitting
prediction of the entire reactor motion and the emerging product the calibration models to appropriate empirical data, while k are deter-
quality. The performance of the proposed method in yielding the mined following standard linearization scheme. This study noted that
present-time estimate of the key reactor variables that are related to unadjusted model(s) resulting from direct application of the third-order
product quality including; conversion viscosity, molecular weight, was reaction rate model (1a) lead to poor prediction of the reactor variables
compared with that of the standard geometric estimation method at certain conversion range. Hence, in advanced kinetics modeling, the
using data taken from the results of a laboratory scale synthesis of reaction rate model (1a) is usually forced to adequately describe the
170 chemical engineering research and design 1 1 7 ( 2 0 1 7 ) 168–176

entire course of reaction by assuming the lumped reaction rate con-


stant k, a time-varying parameter, and incorporating the dynamics of
k in the kinetics state set.
⎡ ⎤−1 ⎡ ⎤
J11 (xc , pe ) 0 0 1 D 1/2D2
⎢ ⎥ ⎢ ⎥
2.1. Overview of geometric estimation method
= ⎣ J21 (xc , pe ) J22 (xc , pe ) 0 ⎦ ⎣0 1 D ⎦
J31 (xc , pe ) J32 (xc , pe ) J33 (xc , pe ) 0 0 1
The geometric estimation method (GEM) is a known reliable means to
⎧ ⎫
handle the alkyd kinetics prediction–estimation problem. ⎪ g (D, , ω) ⎪
⎨ 1 ⎬
For typical reactor estimation problem the reaction rate parameter
g2 (D, , ω) (5)
k is assumed a time-varying parameter given by; ⎪
⎩ ⎪

g3 (D, , ω
k (t) =  (c (t) , w (t)) (2)

which depends on conversion of acid functional group c and other and the terms of the observability matrix J (xc , pe ) are
un-modeled variables w in an unknown possibly nonlinear form  obtained from the partial derivative ∂xc  (xc , pe ) described in
(Henerndez-Escoto et al., 2010). Considering k and its time derivative v Eq. (6).

J11 (xc , pe ) =∂c c (xc , pe ) ; J21 (xc , pe ) =∂c k (xc , pe ) ; J31 (xc , pe ) =∂c v (xc , pe ) ; J22 (xc , pe ) =∂k k (xc , pe ) J32 (xc , pe ) =∂k v (xc , pe ) J33 (xc , pe ) =∂v v (xc , pe )

(6)

as dynamical variables, the kinetics state set is given in terms of the


Given the maps r and  as described in Eqs. (1a) and (1b),
observable motion as,
respectively, and the kinetics state set xc (3), the observability
xc = [c, k, v] ;

pe = [ce , p ] (3)
map  (xc , pe ) is defined as Eq. (7)

Assuming the key interest variables including; conversion c, viscos-  (xc , pe ) = [c (xc , pe ) , k (xc , pe ) , v (xc , pe )] (7)
ity v and molecular weight M to only depend on the kinetics state set,
dropping the unobservable dynamics, the standard estimator structure where;
is recalled in the form suggested by (Henerndez-Escoto et al., 2010) for
the present-time estimate of the interest variables based on discrete- c (xc , pe ) =  (c, p ) (8a)
delayed measurement of viscosity.

  k (xc , pe ) = ∂c  (c, p ) r (c, k, ce ) (8b)


ĉ (ti ) = c ti , ti−1 , ĉ (ti−1 ) , k̂ (.) , p̂e + Gc (ĉ (ti−1 ) , p̂e , D, , ω)

(y (ti ) −  (ĉ (ti−1 ) , p̂e )) (4a)


v (xc , pe ) = ∂2 c2  (c, p ) r(c, k, ce )2
  +∂c  (c, p ) ∂c r (c, k, ce ) r (c, k, ce) + ∂k r (c, k, ce ) v (8c)
k̂ (ti ) = k̂ (ti−1 ) + Dv̂ (ti−1 ) + Gk ĉ (ti−1 ) , k̂ (.) , p̂e , D, , ω

(y (ti ) −  (ĉ (ti−1 ) , p̂e )) (4b)


⎧ ⎫
⎪ 3 − 2e  cos () − e ⎪
  ⎪
⎪ ⎪


⎨ ⎪
v̂ (ti ) = v̂ (ti−1 ) + Gv ĉ (ti−1 ) , k̂ (ti−1 ) v̂ (ti−1 ) , p̂e , D, , ω
1  ⎬
 3g1 − 4 + e2  + 2e( +1) cos () + e(2 +1)
[g1 , g2 , g3 ] = (9)
(y (ti ) −  (ĉ (ti−1 ) , p̂e )) (4c) ⎪

2D ⎪


⎪  ⎪

⎩ 1
g − 2 + e2  + 2e( +1) cos () − e(2+1)

1
D2
v (t) =  (ĉ (t) , p̂e ) (4d)
and

M (t) =  (ĉ (t) , p̂e ) (4e)  


[, ] = 1 − 2 , −Dω (10)

ĉ (t0 ) ≈ c (t0 ) ; k̂ (t0 ) ≈ k (t0 ) ; v̂ (t0 ) = 0 A comprehensive review of the geometric estimation
where ẑ is the estimate of the variable z, c is the transition map of the method noted that proper monitoring of the reaction progress
differential equation ċ = r (c, k, ce ) , c (t0 ) = c0 , yielding c (t, t0 , c0 , k (.) , ce ); and control of product quality could be achieved for a typical
the maps  and  are defined in Eqs. (1b) and (1c), Gc , Gk and Gv are alkyd reactor. This can be achieved via a detailed numeri-
nonlinear gains constructed on the basis of the observability matrix cal estimation method which guarantees convergence to the
−1
J(xc , pe ) , which depend on the state, a tuning parameter pair ( , ω) cho- actual experimental data at both low and high conversion up
sen in such a way as to have stable estimation error dynamics (ε = y − ŷ) to the batch end. However, a decisive attempt to test the per-
and the sampling-delay time (D). formance of such a method in predicting the major reactor
 
Eqs.  (4a)–(4e)
    yields the 
estimate sequence
variables over the entire course of reaction using appropriate
ĉ (ti ) , k̂ (ti ) , v̂ (ti ) , ˆ (ti ) and M̂ (ti ) convergent to the
          empirical data (Table 1), reveals that the prediction accu-
current sequence c (ti ) , k (ti ) , v (ti ) , v (ti ) and M (ti ) ,
racy obtainable via the geometric estimation method degrades
respectively.
considerably at certain useful conversion range. A further
The nonlinear gain G (xc , pe , D, , ω) is described by the expression
treatment of the alkyd kinetics prediction–estimation problem
 −1

G (xc , pe , D, , ω) = [Gc , Gk , Gv ] = J(xc , pe ) [ (D)] [g (D, , ω)] is thus proposed in this paper. The proposal targets devel-
opment of a mechanistic model with improved capacity to
chemical engineering research and design 1 1 7 ( 2 0 1 7 ) 168–176 171

Table 1 – On-line measured batch reactor data for alkyd resin synthesis.
Time (min) Acid value Conversion Degree of polymerization MWA (g/mol) Viscosity (cP)

0 248.17 0.1530 1.1806 199.50 85.70


20 196.89 0.3280 1.4881 251.50 102.80
40 152.06 0.4810 1.9268 325.60 110.00
60 112.51 0.6160 2.6042 440.10 135.00
80 79.11 0.7300 3.7037 625.90 148.00
100 60.35 0.7940 4.8544 820.40 160.00
120 43.65 0.8510 6.7114 1134.20 188.70
140 34.57 0.8820 8.4746 1432.20 207.40
160 27.83 0.9050 10.5263 1778.90 225.00
180 21.09 0.9280 13.8889 2347.20 253.00
200 16.99 0.9420 17.2414 2913.80 273.00
220 15.23 0.9480 19.2308 3250.00 295.00
240 10.25 0.9650 28.5714 4828.60 317.00
260 8.79 0.9700 33.3333 5633.30 335.00
280 6.44 0.9780 45.4545 7681.80 350.40
300 5.86 0.9800 50.0000 8450.00 370.00

Initial Acid value of reacting mixture (AVo ) = 293.00.

predict accurately the major reactor variables at any useful by measuring the evolved water of reaction and periodically
conversion range and over the entire course of reaction. checking the acid number by titration at time step of 20 min.
Palm kernel stearin modified alkyd resin (experiment 2) was
equally synthesized from a blend of 35 g palm kernel stearin
3. Material and methods
(PKS) and 15 g rubber seed oil (RSO). This combination which
represent 70:30 wt% ratio of PKS to RSO was done in other
3.1. Materials
to introduce unsaturation to the PKS molecules. The reac-
tion was carried out following the method of experiment 1,
The gmelina fruits were collected locally from a forest in min-
however, the duration was about 7 h. Data taken from partial
istry of forest reserve, Anambra State in Nigeria. It was soaked
replacement of palm oil stearin with linseed oil (PSLO2080)
in water for 8 days so as to easily separate the fruit pulp from
for alkyd resin synthesis reported by Nanvaee et al. (2013) was
the seed (de-pulp). The seeds were sun-dried and crushed
considered as experiment 3.
mechanically using corona blender. The oil was extracted
using soxhlet extractor. The palm kernel stearin was col-
lected from the Upglobal Agro oil mill, Isiakpu village Nise, 3.2.2. Formulation of approximating model for the
Anambra State. The refined rubber seed oil was purchased proposed alkyd kinetics study
from the Rubber Research Institute of Nigeria, Iyanomon, For the purpose at hand, Eq. (1a) is transform into a state dif-
Benin city and was used without further purification. Ana- ferential equation and re-expressed in parabolic form as Eq.
lytical grade glacial acetic acid, H2 SO4 , phthalic anhydride, (11).
maleic anhydride, acetic acid, sodium bicarbonate, lithium
oxide and glycerol were obtained from commercial sources
(Merck Chemicals, BDH Chemical Limited, Poole, England and ċ = r (c, k, ce) = kce 2 (1 − c) 2 (11)
May and Baker Limited, Dangeham, England) and used with-
out further purification.
where  = (1 − c/ce ).
Taking similar steps from Lopez et al, 2000, knowing that
3.2. Methods
the kinetics parameter ce typically takes a constant real num-
ber slightly greater than or equal to 1 and the value of c lies
3.2.1. Experimental
typically in the range of (0 − 1). In order to force the trans-
The raw gmelina seed oil (GSO) was chemically modified
formed rate model Eq. (11) to adequately describe the entire
(desaturation) according to the method described in Odetoye
course of reaction, the product of the lumped reaction rate
et al. (2011). Medium oil lenghth alkyds (experiment 1) were
constant k and  could be considered a time-varying parame-
prepared with the modified GSO using lithium oxide as cata-
ter such that the two identified parameters of the transformed
lyst, phthalic anhydride and maleic anhydride. The reactions
reaction rate model,  (t) and k (t) whose values depend on c are
were carried out in a three necked round bottom flask titled
lumped into a nonlinear function of time f (t) according to Eq.
with a mechanized stirrer, a dean-stark trap titled with water-
(12), keeping in mind that this methodical assumption will be
cooled condenser and nitrogen in let tube at a temperature
removed in the later stage of the analysis.
of 220–260 ◦ C. Xylene was employed as an azeotropic solvent.
During alcoholysis, the triglycerides of GSO were converted
to monoglycerides by reaction with glycerol at 180–200 ◦ C. PA, f (t) = k (t) 2 (t) (12)
glycerol and xylene were then added, and polymerization was
carried out at 220–260 ◦ C for about 5 h. The required MA was
then added to complete the reaction. The concentrations of The present proposal hence departs from López et al.
alkyd solutions were determined and finally diluted to 70–75% (2000), Hernández and Alvarez (2003), Henerndez-Escoto et al.
by addition of xylene. The progress of reaction was followed (2010) by considering the substitution of Eq. (12) in Eq. (11) to
172 chemical engineering research and design 1 1 7 ( 2 0 1 7 ) 168–176

obtain a variable-separable differential conversion model Eq. operation data. It is quite interesting that satisfactory results
(13). could be achieved using low order polynomials where n ≤ 2.

dc = ce 2 (1 − c) f (t) dt (13) 4. Implementation

Eq. (13) is a linearized expression for conversion rate. It then The predictions of viscosity (v) and molecular weight average
follows that the integration of Eq. (13) could be simply com- (M) via the proposed adjusted mechanistic model follow from
pleted analytically to obtain a generalized integral conversion the application of Eqs. (1b) and (1c), respectively. From nonlin-
model Eq. (14). ear regression fitting over the experimental data (Table 1), the
following viscosity and average molecular weight parameters
c (t) = 1 − e−F(t) (14) were obtained.

p = [a , b , c ] = [70.2752, 0.3157, 1.1725]


 Although f (t) is typically nonlinear, however, its integral
f (t) dt could be well approximated in time domain using a
linear polynomial function q (t) of the general form Eq. (15) p = [a , b , c, d, ] = [0.7563, 0.0059, −0.0014, −0.0045]

q (t) = ˛tn + ˇt(n−1) + . . . + t + (15) The substitution of p , p resulting from nonlinear regres-
sion, and the adjusted c (t) obtained from Eq. (16) into the
such that the expressions for the kinetics state set [c, ċ, k, v] viscosity-conversion relationship Eq. (1b) and the molecular
are written inthe following steps: weight-conversion expression Eq. (1c) gives the final form of
Replacing f (t) dt with the approximating polynomial q (t), the adjusted models applied in the solution step.
the adjusted integral conversion model is inferred from Eq. (14) Evidently, the proposed method for alkyd kinetics study fol-
as, lows simple implementation procedure and ultimately leads
to overall improved prediction accuracy compared to the stan-
2 n (n−1) +...+t+
c (t) = 1 − e−ce (˛t +ˇt ) (16) dard geometric estimation method and the unadjusted model.

while the conversion rate ċ (t) corresponding to Eqs. (1a) and 4.2. Estimator functioning
(11) is re-expressed as the first order derivative Eq. (17),
  For the estimator, the initial conversion was set to the actual
ce 2 ˛nt(n−1) + ˇ (n − 1) t(n−2) + . . . +  f (t)
ċ (t) =   =   conversion c (t0 ) = 0.153 measured at the initial monitoring
e
− ˛tn +ˇt(n−1) +...+t+ − ˛tn +ˇt(n−1) +...+t+
e period t0 . A typical value for k0 = 0.0367 min−1 has been
reported for a relatively high initial conversion c0 = 0.794 in
c (t0 ) = c0 , pc = [˛, ˇ, , ] (17)
Henerndez-Escoto et al. (2010) following the method described
in Lin and Hsieh (1977). By interpolation on the k0 − c0 plane
The kinetics parameter k (t) is obtained by re-writing Eq.
assuming k0 ∝ c0 in the low conversion range, the initial value
(12) as in Eq. (18).
of the reaction rate parameter for the current case where
ce 2 q̇ ce 4 q̇ c0 = 0.153 was found to be 0.0071 min−1 . The time deriva-
k (t) = = 2 (18) tive of the reaction rate parameter was set to v0 = 0 in line
 2 c − 2cce + ce 2
with Henerndez-Escoto et al. (2010). The values of the tuning
and its time derivative v is expressed as Eq. (19). parameters ω = 4.5/210 and = 0.71 was found to be the best
   estimator performance which was compared with the best
ce 4 c2 − 2cce + ce 2 q̈ − 2q̇ (cċ − ċce )
v = k̇ (t) = performance of the adjusted mechanistic model in Fig. 3.
2
(c2 − 2cce + ce 2 )
4.3. Performance evaluation of the proposed model to
k (t0 ) = k0 (19) alkyd kinetics studies

where To test the performance of the proposed alkyd reactor pre-


 diction method, the data obtained from the kinetics study
 
F (t) = ce 2 f (t) dt ∼
= ce 2 q (t) = ce 2 ˛tn + ˇt(n−1) + . . . + t + of oxy-polymerization of fatty acid of gmelina oil in pres-
ence of phthalic anhydride, maleic anhydride and glycerol was
The model adjustment procedure suggested above simply considered as given in Table 1. The case study presents an
describes the construction of a standard integral model for operation carried out in an experimental batch reactor from
conversion of acid functional group Eq. (16) which is nonlin- the beginning of the reaction when (c = 0.153) up to the batch
ear in terms of the unknown parameter set pc = [˛, ˇ, , ]. The end. The proposed method requires that a useful order n of
reaction motion in a typical batch alkyd polycondensation the approximating polynomial q (t) which guarantees good
process could be reconstructed using the proposed integral performance of the integral conversion model Eq. (16) is first
conversion model if the unknown parameter set pc is prop- determined. This was accomplished for the study by fitting
erly estimated. The detailed kinetics modeling then involves Eq. (16) to the experimental data while the kinetics parameter
collective application of the Eqs. (16)–(19) and the other exist- ce was fixed at its nominal value of 1. The results presented
ing mechanistic models including Eqs. (1b) and (1c). Achieving in Fig. 1 show that although high prediction accuracies could
high prediction accuracy through either the proposed integral be achieved assuming q (t) to be either linear (n (q) = 1) or
conversion model or the adjusted rate model Eq. (17) requires quadratic(n (q) = 2), yet, a better performance of the proposed
proper selection of the order n of the  polynomial q (t) that integral conversion model is identified for the present case-
accurately approximates the integral f (t) dt for a given batch study with the linear approximation of the integral. The
chemical engineering research and design 1 1 7 ( 2 0 1 7 ) 168–176 173

Table 2 – RMSE for DDP and adjusted model.


DDP Adjusted model

Experiment 1 (GSO) RMSE


Conversion 0.02936 0.01214
Viscosity 282.0286 8.7234
MWA 503.4955 963.0724

Experiment 2 (PSLO2080)
Conversion 0.0206 0.00996
Viscosity 51.4461 49.5514
MWA 0.9846 1.0343

Fig. 1 – The performance of the proposed integral Experiment 3 (PKS)


conversion model at n = 1 (–) and n = 2(–). Conversion 0.01443 0.0043
Viscosity 30.9463 56.4792
MWA 944.1482 1734.0024
optimal values of the parameter set pc = [0.0136, 0.1644] cor-
responding to the linear case were found numerically using variables suggest that the reaction was relatively fast in the
unconstrained nonlinear optimization method ‘fminsearch’ early period of batch operation approaching 80% conversion
found in MATLAB. Fminsearch algorithm was adapted to con- within the first 100 min of the reaction, but gradually slows
duct a global search for the optimal pc that minimizes the down as the reaction progresses towards gelation.
sum of squared error between the measured conversions, and
those predicted by the adjusted model. 4.4. Validation of the proposed pure mechanistic
The outcome of Fig. 1 in which considering the order of the approach to alkyd reactor prediction
approximating polynomial n (q) = 1 gives a better convergence 
to the experimental data suggests that the integral f (t) dt In the coating industry, liquid resins with correct viscosity
based on the present experimental data show uniform gradi- (in the range 249–300 cP) and molecular weight (in the range
ent within the monitoring period. Earlier related studies show 5000–10,000) are basically desired. Hence, the greatest con-
that this trend is, however, not generic to all cases of alkyd cerns to product analyst are the risks of obtaining either an
polymerization processes. Hence, useful order of the approxi- inviscid product (usually caused by shortage of reactants or
mating polynomial n (q) must be recalibrated for any different premature termination of the batch) or unwanted gelatinous
set of reactor prediction data. product if the reaction is allowed to progress up to and beyond
In pursuing the usual production goal of achieving a prod- the critical gel-point. Therefore, the need to correct the reac-
uct quality within specification at the end of the batch, tor motion where necessary by addition of reactants or to
operators are consigned with monitoring on-line the reaction stop the batch before the critical gel point so as to obtain
progress and product quality. Often, the dynamic states of the a product within specification implies that the key variables
reacting mixture are estimated in terms of conversion of acid directly related to product quality including; conversion of
functional group c, the reaction rate parameter k and their time acid functional group, viscosity and molecular weight of the
derivatives ċ, v, respectively. reacting mixture must be monitored collectively over the reac-
With the proposed method for studying alkyd polymeriza- tion period. To validate the proposed mechanistic approach
tion kinetics, it becomes indeed easier to predict accurately to alkyd reactor prediction, efforts were made to compare
these variables that are related to product production rate. The the performances of standard geometric estimation method
kinetics state set [c, ċ, k, v] are described by Eqs. (16)–(19). The also known as discrete data processor (DDP). The proposed
results presented in Fig. 2a show exponential rise and fall of adjusted model and the unadjusted model in predicting the
conversion and conversion rate, respectively, witnessed dur- reactor variables in terms of the level of convergence to the
ing the reaction period. Fig. 2b describes the dynamics of the experimental data achievable with such methods are shown
reaction rate parameter k and indicates that both k and its time in Fig. 3a.
derivative v were apparently constant in the low conversion From the results collected in Fig. 3(a) it could be inferred
range but increased geometrically after the first 120 min of the that the existing GEM and the proposed adjusted (mechanistic)
reaction. The overall results in terms of the predicted state model have comparable capacities to yield adequate predic-

Fig. 2 – Predicted kinetics states of the studied alkyd reactor based on the proposed method.
174 chemical engineering research and design 1 1 7 ( 2 0 1 7 ) 168–176

Table 3 – Calculated values of the existing reactor calibration model parameters.


Model parameters a b c a b c d ce

Exp.1 (GSO) 70.2752 0.3157 1.1725 0.7563 0.0059 −0.0014 −0.0045 1


Exp.2 (PKS) 8.7587 0.1694 1.0306 248.8433 0.0046 1.4692 −1.4738 1
Exp.3 (PSLO2080) 5.4479 0.2517 1.0416 39.7694 0.0060 0.2292 −0.2352 1

Table 4 – Calculated values for the parameters of the proposed adjusted model.
Adj. model parameters ˛ ˇ  n ce

Exp.1 (GSO) 0 0 0.0136 0.1644 1 1


Exp.2 (PKS) 0 0 0.0091 1.2931 1 1
Exp.3 (PSLO2080) 0 0 0.0084 1.1610 1 1

Fig. 3 – The performances of the geometric estimation method, adjusted (mechanistic) model, and unadjusted model in
yielding the present-time estimate of the major variables of gmelina oil alkyd reactor.

tion of the studied reactor variables which include; conversion validity of the proposed reactor prediction method were con-
(c), viscosity () and molecular weight average (MWA). More ducted using data obtained from experiments conducted at
detailed comparison of the results achieved through the error different reaction conditions. Two additional reactor calibra-
dynamics presented in Fig. 3(b) and root mean square error tion problems both of which involve the synthesis of alkyd
(RMSE) shown in Table 2, reveals that the proposed adjusted resin from palm kernel stearin (experiment 2) and palm oil
model guarantees a significant overall improvement in the stearin (experiment 3) (Nanvaee et al., 2013) were considered.
studied gmelina alkyd estimation–prediction problem. The analysis and the collection of the results follow the same
To formalize the new method for general-purpose alkyd procedure as the one described for gmelina oil. The results
reaction prediction–estimation studies, further tests on the were presented in Figs. 4 and 5, respectively. The values of
chemical engineering research and design 1 1 7 ( 2 0 1 7 ) 168–176 175

Fig. 4 – The performances of the geometric estimation method, adjusted (mechanistic) model, and unadjusted model in
yielding the present-time estimate of the major variables of palm kernel stearin alkyd reactor.

Fig. 5 – The performances of the geometric estimation method, adjusted (mechanistic) model, and unadjusted model in
yielding the present-time estimate of the major variables of palm oil stearin (PSLO2080) (Nanvaee et al., 2013) alkyd reactor.

the existing reactor calibration model parameters and the 5. Conclusion


proposed mechanistic model used in the overall study are
presented in Tables 3 and 4, respectively. A mechanistic model for alkyd polymerization kinetics which
guarantees high-order accuracy in wide-range of applications
176 chemical engineering research and design 1 1 7 ( 2 0 1 7 ) 168–176

is developed in this work. The study found that with proper Ciccarella, G., Dala Mora, M., Germani, A., 1993. A
adjustment of the proposed mechanistic models, prediction Leunberguer-type observer for nonlinear systems. Int. J.
accuracy significantly higher than that of the popular geo- Control 57 (3), 537.
Deza, F., Busvelle, E., Gauthier, J.P., Rakotopara, D., 1992. High-gain
metric estimation method could be obtained for the studied
estimation for nonlinear system. Syst. Control Lett. 18, 295.
alkyd kinetics problems. With the success of this study, model- Eaton, J.W., Rawlings, J.B., 1992. Model predictive control of
based calibration of alkyd reactors which essentially reduces chemical processes. Chem. Eng. Sci. 47, 705–720.
on-line measurement efforts (required to achieve products Hernández, H., Alvarez, J., 2003. Robust estimation of continuous
quality index within specification), and avoids the rigorous nonlinear plants with discrete measurements. J. Process
tuning procedure that characterizes the existing numerical Control 13, 69–89.
estimation methods, becomes much more reliable. Henerndez-Escoto, H., Lopez, T., Alvarez, J., 2010. Estimation of
alkyd reactors with discrete-delayed measurements. Chem.
Eng. J. 160, 698–707.
References Lin, C.C., Hsieh, K.H., 1977. The kinetics of polyesterification,
adipic acid and ethylene glycol. J. Appl. Polym. Sci. 21,
Aigbodion, A.I., Okeimen, F.E., 1996. Kineticts of the preparation 2711–2719.
of rubber-seed oil alkyds. Eur. Polym. J. 32, 1105–1108. López, T., Hernández, H., Alvarez, J., 2000. Robust nonlinear
Alvarez, J., Lopez, T., 1999. Robust dynamic state estimation of estimation of alkyd reactors with discrete-delayed
nonlinear plants. AIChE J. 45 (1). measurements. In: ADCHEM2000 Congress, Pisa, Italy.
Araromi, D.O., Adegbola, B., 2014. System identification of Nanvaee, Azimi A.R., Yahya, Rosiyah, Gan, Seng-Neon, 2013.
nylon-6 caprolactam polymerization process. ARPN J. Sci. Cleaner production through using by-product palm stearin to
Technol. 4 (2). synthesis alkyd resin for coating applications. J. Clean. Prod.
Bachir, D., Mansouri, A., Bouhamida, M., Chenafa, M., 2005–2008. 54, 307–314.
State estimation in free radical polymerization reactor using a Odetoye, T.E., Ogunniyi, D.S., Olatunji, G.A., 2011. Improving
nonlinear interconnected high-gain observer. J. Theor. Appl. Jatropha Curcas Linnaeus oil alkyd drying properties. Elservier
Inf. Technol. J. Prog. Org. Coat. 73, 374–381.
https://www.researchgate.net/publication/242177807. Patton, T.C., 1962. Alkyd Resin Technology. John Wiley & Sons, NY.
Bequette, B.W., 1991. Nonlinear control of chemical processes: a Valeel, J., Georgakis, C., 2002. Nonlinear model predictive control
review. Ind. Eng. Chem. Res. 30 (7), 1391. of end-use properties in batch reactors. AIChE J. 48 (9).
Buruaga, I.S., López, T., Pérez, S., Alvarez, J., 2003. Calorimetric Wang, G.B., Peng, S.S., Huang, H.P., 1997. A sliding observer for
Estimation of Viscosity and Acid Number in Alkyd Reactors. nonlinear process control. Chem. Eng. Sci. 52 (5), 787.
IFAC.
Castellanos, E., Zaldo, F., 2015. Joint control and operation design
for alkyd resin polyesterification, polymer reaction
engineering IX. ECI Symposium Series.