European Symposium on Computer-Aided Process Engineering - 14 901

A. Barbosa-P6voa and H. Matos (Editors)

9 2004 Elsevier B.V. All rights reserved.

Periodic Scheduling of Multiproduct Continuous Plants

using a RTN Continuous-time Formulation
Pedro Castro *t', Ana Barbosa-P6voa ~ and Augusto Novais t
tDMS, INETI, 1649-038 Lisboa, Portugal
~CEGIST, Instituto Superior T6cnico, 1049-001 Lisboa, Portugal

Abstract
This paper considers the problem of deriving an optimal periodic schedule for an M-
stage, P-product multiproduct plant with sequence dependent changeovers. A Mixed
Integer Non-Linear Programming (MINLP) continuous-time mathematical formulation
is proposed, which allows for different product sequences in different stages and
variable processing rates. The performance of the Resource-Task Network (RTN) based
formulation is illustrated through the solution of a 2-stage, 3-product example problem,
with the results showing it to be capable of achieving significant better solutions than
the MINLP periodic scheduling formulation of Pinto and Grossmann (1994). However,
the complexity of the proposed formulation is apparent and increases steeply with an
increase in the number of products/stages.

Keywords:Periodic Scheduling, Resource-Task Network, Continuous-time Formulation

1. Introduction
Most manufacturing environments are currently changing very rapidly due to an
increasingly competitive global market. The trend is to move towards more flexible
process systems that can respond to market requirements quickly. In such systems the
procedure of allocating resources and equipment over time so as to execute the chemical
and physical processing tasks required to manufacture one or more products in a
profitable manner can be very complex. Thus, there is a real need for efficient
scheduling tools that can help this decision-making process.
Process plants are usually operated in either a short-term or a campaign mode. The
campaign mode of operation is appropriate where product demands are stable over
extended periods of time. In such cases it is often profitable to establish a regular
periodic operating schedule (typically of the order of days) in which the same sequence
of operations is carried out repeatedly. The goal of periodic scheduling is to optimize
the utilization of resources over a cycle by ignoring the start-up and shutdown phases of
the production and to determine the optimal cycle time.

*Corresponding author. Tel. :+351-217162712. E-mail: pedro.castro @ineti.pt

902

2. Problem Description
This work addresses a specific type of problem that arises frequently in the chemical
processing industries, namely that of continuous multiproduct plants with sequence
dependent changeovers, and was first addressed by Pinto and Grossmann (1994) with a
Mixed Integer Non-Linear Programming formulation. It is assumed that the
manufacturing process consists of a sequence of M-stages, each with its own dedicated
equipment, interconnected by intermediate inventory tanks for each of the P-products.
The objective is to maximize the profit realized by the process per unit time, where the
profit function is composed of several terms. These include the revenue from product
sales, as well as changeover and inventory costs. For the latter, it is assumed that the
inventory cost for the intermediates is a linear function of the maximum amounts that
need to be stored, and that the inventory cost for the final products depends on the
amounts stored per unit time.

3. Problem Formulation
The proposed mathematical formulation can be viewed as a generalization of the
periodic scheduling formulation of Castro et al. (2003) with the inclusion of the more
recent developments of the short-term formulation of Castro et al. (2004). However, the
nomenclature has been adapted to the unique characteristics of the problem being
addressed. The formulation is based on the Resource Task Network (RTN) process
representation of Pantelides (1994) and employs a continuous representation of time,
where the cycle time is divided into intervals (set T) of unknown duration (see figure 1).
Notice that a single time grid is used for all process resources.

!: Cycle N !' Cycle N+I

,=I] Slot1 Slot 2 Slot T-1 Slot T J Slot 1
"l
i
2
i
3
% i
T-1
s
T
i e
2
1 T+1-1
0 H

Figure 1- Time grid adopted in this work

| .I Make_A_1
"l Rate(kg/h)=0-800 J
I ~/ J
"l
Make_A_2
Rate(kg/h)-0-900 J
I

.! .I I
(~ "lRateMake B_I
(kg/h)-=0-1200 ] =~
Make_B_2
"I Rate(kg/h)=O'600 l
.I Make_C_2 I
(~ J Make_C_1 I
"1 Rate(kg/h)=0-1000 [
=~j
"i Rate(k~/h)=0"1100 I 0

Figure 2- RTN representation of the process (part I- processing tasks)

 903

....... ,4 Clean_A_C_l t ; ....... .~ Clea n_A_C_2

Duration (h)=8 J ........ ' Duration (h)=l t ........ '

i .......................... ~ .
......L i il ....Q ...., :: Clean A B2 " " ..... t CI.... C_A_2
Duration (h)=6 ~'~.... ]
',',
', l
,~ t ..... i l
._J ca . . . . a_A_l t . . . . . i__i
i l Duration(h)=10 | [ ......... 1 Cl.... C_B1 "" i Ou.,on,~,=~ t ....... . ........ t C' .... ,on,~,--,
o .... C_._= i" .... if=
i.......................... | Durati~ (h)=7 ~-" J
........ _.... i ............. i
' ....... -t ~ t,~,:~ ........ ! " ....... "1 ~tou=o,~,:~ ........

Figure 3- RTN representation of the process (part H- cleaning tasks)

The RTN representation of the 2-stage, 3-product process considered as the example
problem is given in figure 2 and figure 3. It uses two different types of resources" i)
three material states (set S) per product (set P); ii) one equipment state per product and
per stage (set M); and two different types of task: a) continuous processing tasks; b)
batch cleaning tasks. Notice that there are two types of interaction with resources. On
one hand there is a discrete interaction (double-arrow dashed lines) with equipment
states, which can be viewed as a temporary decrease of the total resource availability (a
task captives the resource at its start only to release it back at its end). On the other
hand, there is a continuous interaction (single-arrow solid lines) with material resources,
which results on a permanent decrease of the total resource availability.
Three sets of variables are used to characterize the execution of tasks in a particular
time interval. The occurrence of an instance of the processing task of product p in stage
m starting at event point t and ending at event point t' (equal to t+l for the continuous
tasks), is identified through the binary variable Nm,p,t,t+l, while the nonnegative
continuous value ~m,p,t,t+l gives the total amount of material being processed. For the
cleaning tasks only one set of variables is required since no material resources are
involved. These are the binary variables Nm,p,p,,t,t,. The other required variables, which
are all nonnegative, are the excess resource variables Rr,p, t (r~M) and S s,p,t, the
timing variables Tt and the cycle time (H), the total amount of material consumed over
the cycle As,p (only for raw-materials, s~ S ~ ) , and the maximum amount of material
stored over the cycle STs,p (only for intermediate products, s~ sRM^s~ sFr').
The model constraints are presented next, together with the objective function. Due to
the lack of space, and since these have been thoroughly discussed (in their general form)
in recent papers (Castro and co-workers, 2003, 2004), no description is here provided.
In some constraints we have used the following wrap-around time operator.

- I T l i f t >lTI
ff2(t) =
fi if 1 < t <i T I
t + l T l i f t <l
(1)
904
m m

Rr,p, t - Rr,p,~(t_l) - Nr,p,t,~(t+l) + Nr,p,~(t_l),t +

q-Z( ZNr,p,,p,t,,t - ZNr,p,p',t,t') VFE M,PE P, tE T (2)

p '6P t '~T t'~T
p',p t-At<t'<tvt'>t-At+]T I t<t'<kt+tvt'<At+t-lTI

Zer,p,T = n 0 +ZZ(N--r,p,t,~(t+l)-N--r,p,t,t+l)+
p~ P I~ P t~T
(3)
-]- Z Z Z ( Z gr'p''p't't' -- Zgr,p,p,,t,t,) Vr~ M
p~ P p '~P t6 T t '~T t '~T
p'~p t<t'<At+t t<t'<At+tvt'<At+t-IT
I

Ss,p,t = A s,p t=l,sESRM 4r Ss,p,~(t_l) q- Z/]'m,s~m,p,~(t-1),t --

m~M
(4)
I Zs,~SRMAs"p'(Ttlt~:l+Hlt=l-Tf2(t-1))] s~sFe Vs ~ S, p ~ P, t ~ T

as,p, t <_STs, p ~sq~ sRM,s(~ S FP, p 6 P, t6 T (5)

As,p > dpH Vs6 SRM,p6 P (6)

T t ' - T t >- Z (~r'p't't'lt':t+l+max Z~Zr,p,p'gr,p,p',t, t ' ) V F E M , t s T , t'~T,t<t'<at+t_ (7)

p~ P iOr, p p '~e
p ':~p

Tt - Tc<_ H - Z ( ~ r ,-p , t , t ' max

It'=t+a-lTI Z
"t- ~Zr,p,p,Nr,p,p,,t,t, ) Vr, t,t', t' < At + t - I T I (8)
peP iOr ,p p'~ e
p'~p

Tc-Tt <Hmax " ( 1 - Z Z gr'p'p',t,t') -[-Z Z ~r,p,p'Nr,p,P',t,t' Vr, t,t',t < t'< At + t (9)
p~_Pp'~P peP p'~P
p'~p p'~p

Tt - T t, > n - n max .(1- Z ZNr,p,p',t,t')-Z Z~r,p,p'gr,p,p',t,t ' Vt'<At+t-ITI(IO)

p~.Pp'~P pe.P p'~P
p'~p p'~p
max _ max ~,
~m,p,t,~2(t+l) <~H Pm,p1u m,p,t,~2(t+l) Vm ~ M, p ~ P,t ~ T (11)
Tt < H max '7'tE T (12)
H < H max (13)
CSs,p
Z ZCpAs,p - Z Z Z 2 (Ss'p't+Ss'p'f~(t+I))(Tt+IIt*ITI+HIt=ITI-Tt)
s~S RM peP s~S Fe peP t~T
m a x

H
(14)
Z
s~S RM,s~S FPpeP
Z ZEZ
m~M p~P p'~Pt~T t'~T
t <t'<At+tvt'<At+t-ITI
H
 905

4. Problem Data
The performance of the formulation is illustrated with a 2-stage, 3-product problem
(example 2 of Pinto and Grossmann, 1994). The data related to the maximum
processing rates of the continuous tasks ( lOm,p
max ), to the continuous rate of
generation/depletion of state s in stage m (A,m,s ) and to the duration of the cleaning
tasks ( ~'r,p,p' ) can be found in figure 2 and figure 3. The other required parameters are

now given: R ~ = 1Vr~ M ; the minimum demands (dp) are equal to 50, 100 and 250
kg/h of A, B and C, respectively; the final products are worth (cp) 150, 400 and 650
S/ton; the cost of finite intermediate storage is the same for all materials
c--ds,p = 140.6 S/ton Vs ~. S I~M, s ~ S FP, p ~ p., final storage costs cS~,p= 4.06 S/ton/h;
and finally, the transition costs (Cp,p,) are sequence dependent and are the same in both
stages (760 A-B or A-C, 750 B-A/C and 770 $/h C-A/B).

5. Results
The resulting non-convex MINLP was solved by GAMS/DICOPT (build 21.1), which
then used CONOPT for the NLPs and CPLEX for the MILPs. We assumed that the
cleaning tasks could only last one time interval (At=l) and placed an upper bound
(Hm~) of 250 h on the cycle time. In order to reduce solution degeneracy arising from
cyclic schedule permutations, there is need to anchor the schedule. We have chosen to
start the processing task of product A in stage 1 in the first event point ( Nstgl,A,1,2 = 1 ).
The best solution to the problem was found for 10 event points and corresponds to an
objective of 347.87 $/h and a cycle time of 177.92 h. The problem consists of 124
integer variables, 342 continuous variables and 275 constraints and was solved in 9.6 s
on a Pentium IV-2.53 GHz machine. The corresponding schedule is given in figure 4,
while the inventory profiles are shown in figure 5.

Figure 4- Best solution found

906

r
35000 -----Int A ~ Int B -- IntC '="='--FPA ---'FPB -- FPC ~

30000 t
l
r
A
oD 25000

r
l r
~ 20000 -r
r
r
n-
r

= ~5ooo r
8
10000

5000

0 20 40 60 80 100 120 140 160

T i m e (h)

Figure 5- Inventory profiles for the intermediates and final products

The best solution features a A-B-C sequence in stage 1, and a A-C-B sequence in stage
2. Three factors indicate that it is a very good solution: i) the production rates for the
less valuable products (A-B) are at the minimum demand rates; ii) only two continuous
tasks are not executed at their maximum processing rates; iii) almost all IntC is being
produced and consumed at the same time, which is particularly important for the most
required intermediate product. The best solution found by Pinto and Grossmann (1994)
has a profit of only 297 $/h. This can be explained by two factors. The first, the most
important, is that their formulation assumes that the product sequence is the same in all
stages. The second is that they assume that all processing equipments always work at
their maximum processing rates, a condition that is not observed in the current solution.

6. C o n c l u s i o n s
This paper addresses the periodic scheduling problem of continuous multiproduct plants
with sequence dependent changeovers, with a RTN based continuous-time formulation.
The proposed MINLP formulation is more general than the MINLP formulation of
Pinto and Grossmann (1994) and was shown to produce a significantly better solution
(corresponding to a 17 % increase in profit) for a well-known 2-stage, 3-product
problem. The drawback of the higher generality of the proposed formulation is that the
resulting MINLPs becomes very difficult to solve even for a single increase in the
number of products, mainly due to the need of considering at least two more event
points per added product (one for processing and another for cleaning). On the other
hand, problems resulting from the formulation of Pinto and Grossmann (1994) still
remain tractable for examples larger than the one considered here.

References
Castro, P., A. Barbosa-P6voa and H. Matos, 2003, Ind. Eng. Chem. Res., 42, 3346.
Castro, P.,A. Barbosa-P6voa,H. Matos, A. Novais, 2004, Ind. Eng. Chem. Res. 43, 105.
Pantelides, C., 1994, Proc. 2 nd Conf. on Foundations Computer Aided Operations, 253.
Pinto, J. and I. Grossmann, 1994, Comp. Chem. Eng. 18,797.