Professional Documents
Culture Documents
00
# Institution of Chemical Engineers
www.catchword.com=titles=02638762.htm Trans IChemE, Vol 80, Part A, September 2002
M
anagement of cleaning of heat exchanger networks (HENs) subject to signi cant
fouling plays an important role in oil re nery heat recovery network economics.
Recent studies have shown that where this can be applied1;2, signi cant bene ts can
be achieved in, for instance, reducing red-heater duties over the operating horizon considered.
These previous approaches have focused on using outer approximation (OA) based formula-
tions and solution techniques to solve the associated mixed integer-nonlinear programming
problem (MINLP) involving binary decision variables describing when and which unit to clean
in a multi-period formulation. In parallel to this avenue, this work explores the options offered
by stochastic optimization techniques for combinatorial problems applied to the HEN cleaning
problem. A particular method, the backtracking threshold accepting algorithm (BTA) is
developed, derived from the standard threshold accepting algorithm (TA), which is related to
the class of methods typi ed by simulated annealing (SA). Two large-scale case studies are
considered: one involving 14 heat exchanger units, and one involving 25 units. Both are
tackled by the OA and BTA approaches: the latter proved able to handle larger problems, for
example, with more time intervals or exchangers, than the commercial OA solver used.
561
562 SMAÏLI et al.
absence of ‘classical’ optimization techniques, these other order to maximize performance, for example, by adjust-
heuristic approaches are potentially equally valid for gene- ing ow rates between parallel sets of units with
rating cleaning schedules to aid engineers make decisions different extents of fouling. Constant values of Cp are
on the operation and cleaning of HENs subject to fouling. used within each unit.
Secondly, we note at this point that the multi-period (2) Fouling rate models used. Heat exchanger performance
formulation used to generate the MINLP problem represents
and fouling effects are modelled using a lumped para-
the application of a heuristic to the scheduling problem. By
subdividing the time horizon into regular periods in which meter approach. Equation (1) shows how the overall
cleaning actions may (or may not) be performed, the heat transfer coef cient, U , is related to the coef cient
exibility of the schedule is reduced. The truly global when the exchanger is clean, U0 , and the fouling
optimum schedule may require non-regular spacing of resistance, Rf .
events in time, such that the MINLP formulation may only Rf thus combines tube-side, shell-side, and ow pattern
be able to approximate, for example, from several similar effects.
solutions. Flexibility could be maintained by using a large
number of periods ( ne discretization), at a cost associated 1 1
with increased problem size. Rf …t† ˆ ¡ …1†
U …t† Uclean
The remainder of this work focuses on presenting the
HEN cleaning problem and scheduling of cleaning actions
The fouling resistance changes over time, with several
in a multi-period setting as a mathematical programming
formulation. The backtracking threshold accepting algorithm different behaviours being observed in practice7 . The
(BTA) is presented, suitably modi ed for the HEN cleaning work presented here uses a linear fouling model with
problem, along with a comparison to results obtained over constant parameters, based on plant data reconciliation.
the same problems using the MINLP formulations utilizing
the outer approximation method (OA)5 . The BTA algorithm Rf …t† ˆ Rf 0 ‡ at …2†
is implemented in standard ANSI C directly, while the OA
method is accessed through the solver DICOPT 5 within the where a is a constant, and Rf 0 is the initial fouling
GAMS mathematical programming language6 . resistance. The use of other fouling models in the
scheduling problem has been investigated by Smaõ¨li
THE HEN MINLP FORMULATION et al.8 The use of constant fouling rate parameters is
UNDER FOULING justi ed on the grounds that the temperature range in
each unit does not vary signi cantly over the total time
A heat exchanger network consists of a number of hot and
period, and constant ow rates maintain similar bulk
cold input streams, which pass through a set of heat transfer
velocity, wall shear stress, and residence time condi-
units in a xed con guration, as shown in Figure 1. Each
unit may be taken off line for cleaning without interrupting tions. Moreover, the variation in feedstock composition
the process; the con guration is temporarily changed while in a typical HEN operation is likely to render the use of
the relevant hot and cold streams either bypass the unit or more sensitive models inappropriate.
are diverted to another exchanger. The formulation (3) Pressure drop. The effect of fouling on the pressure
presented here is based on the following assumptions: drop across an exchanger is often as important as its
effect on the unit’s thermal performance. The reduction
(1) Constant inputs—heat capacity ow rates. We consider in ow area resulting from accumulated deposits will
here the situation where the network is working under result in higher velocities through the unit and conse-
constant input ow and temperature conditions, that is, quently increased pressure drop. Pressure drop effects
corresponding to the maximum throughput scenario. are therefore acknowledged to be important, but are not
Seasonal variations in ow rate could be considered in considered in detail here. Such effects have been
the multi-period formulation if necessary. This assump- discussed elsewhere8 . Under the constant ow rate
tion precludes the use of inter-period optimization in assumption, there will exist a maximum tolerable pres-
sure drop over a unit or group of units, effectively
imposing a particular operating constraint on the
model. One of the dif culties posed by pressure drop
considerations is obtaining a reliable mapping between
the effects of fouling on the thermal and hydraulic
performances of an exchanger. The relationship is
usually nonlinear, and depends strongly on whether
fouling occurs on both sides of the heat transfer surface,
and whether uneven ow leads to tube blockage.
The aim of the mathematical model is to determine which
unit should be cleaned and when. Each unit n over each
period p can be represented by a binary variable y where
»
1; if unit n is on-line over period p
Figure 1. Schematic diagram of a heat exchanger network (circle, heat yn;p ˆ
exchanger; C, cold utility; H, hot utility). 0; otherwise
and where the following are sets of indices used to describe The multi-period approach features discretizing the time
the network: horizon into NP equal periods (time slots) where each is
further split into a ‘cleaning interval’ (length Dtccl ) and a
p is the period number …1; . . . ; NP†; ‘processing interval’ (length Dtppr ), as shown in Figure 2.
n is the unit number …1; . . . ; NE† For unit n within period p, the heat transfer coef cients in
the cleaning and in the processing intervals are related to the
where NP is the number of periods and NE is the number of corresponding heat transfer coef cients from the previous
units. period:
Each heat exchanger unit is modelled here as a single pass pr
counter-current unit. Other models could be used as cl
Un;p¡1
required, without loss of generality in the approach. The Un;p ˆ …8†
1 ‡ U pr ¢ R_ f n ¢ Dtpcl
n;p¡1
following equations are written for one unit and are repeated
for each exchanger in the network. cl
Un;p
The heat balance is given by pr
Un;p ˆ ¢ y ‡ Uclean ¢ …1 ¡ yn;p †
cl ¢ R
1 ‡ Un;p _ f n ¢ Dtppr n;p
Q ˆ Fh ¢ Cph ¢ …Tinh ¡ Tout
h
† …3†
…9†
on the hot side and by
where Rfn is the fouling rate of each unit in each interval.
c
Q ˆ Fc ¢ Cpc ¢ …Tout ¡ Tinc † …4† Dtppr and Dtccl are xed. Dtccl represents the time that a unit
must be taken out of service in order to effect cleaning: Dtppr
on the cold side. The ow rates Fh and Fc may vary if is an operating window—such as weekends—when cleaning
cleaning actions cause the local con guration to change, and is not allowed, and may have zero length.
(typically) the hot stream from one unit is diverted to The fouling rate for each unit is constant and is given by
another unit operating in parallel.
The heat exchanger performance is given by R_ f n ˆ an …10†
Q ˆ U ¢ A ¢ DTLM …5† Unit temperatures in each interval are then calculated
by simultaneously solving the equations describing the
where DTLM is the logarithmic mean temperature difference. network: the above exchanger performance relationships,
Simulation of the network involves rating equations and and con guration equations describing ow splitting, ow
the performance of each unit is calculated explicitly using mixing, and recycle.
the NTU-effectiveness method; see, for example, reference
9. The outlet temperatures are calculated from the inlet
temperatures via the rearranged equations (3) to (5): Objective
Á !
Fc ¢ Cpc The objective function to be minimized is the total cost
h h c
Tout ˆ Tin ¡ …Tout ¡ Tinc † …6† incurred due to the loss of performance of the network:
Fh ¢ Cph … tF NE X
X NP
same for all units. A CE value of 0:34 £ kW¡1 day¡1 is used constraints both reduce the number of network simulations
in both case studies. and the size of the combinatorial problem.
Furthermore, it is noteworthy that equation (11) does not
consider the state of the network after the nal time tF ,
which has signi cant impact on the schedules generated. STOCHASTIC SEARCH METHODS:
The schedules generated under this objective function could SA, TA, AND BTA
be inoperable later on, incurring signi cant cleaning costs. The use of stochastic solution techniques arises from cases
Consideration of extended operation (post tF ) would require of practical combinatorial optimization problems that either
an amended objective function, such as by adding penalty because of their size, or because of their inherent complex-
terms related to the network ef ciency at tF . ity, are very dif cult to obtain good solutions with determi-
nistic methods.
In particular, the introduction of the simulated annealing
Constraints
method (SA) by Kirkpatrick et al.10 has inspired a multitude
The scheduling problem seeks the set of cleaning deci- of either direct variants or of related methods, all of which
sions that minimizes the objective function (11). Each case have found successful application in niche combinatorial
to be considered will feature its own constraint set, which problems providing stochastically good solutions. SA has
can involve binary and=or continuous variables (e.g., become an extremely popular method in practically every
Appendix A and Appendix B). The constraints frequently eld that in one way or another touches upon combinatorial
take the following forms, which may vary from period to (integer-variable-related) decision making in operations
period and=or unit to unit as the problem requires. research.
The basic SA algorithm operates using the analogy of
Simple performance targets cooling down a melt to obtain a defect-free crystal (the
The values of individual continuous variables in the globally minimal energy con guration at any temperature T ).
network, particularly temperatures, may be subject to It relies on getting an initial feasible system con guration s
bounds or limits, for example, and an initial ‘temperature’ T , which are modi ed according
to the generic algorithm given in Figure 3. The theoretical
T L < Tout
h
< TU …12† literature on SA and related algorithms contains proofs of
asymptotic convergence to the globally optimal solution. A
Particular examples on oil re nery heat preheat trains are the thorough review and comparison of SA and other stochastic
inlet crude temperature to a ash tank or desalter, and ‘run- methods can be found in Nurmela11 and Schoen12 .
down’ temperatures of hot streams. Cleaning decisions that However, such proofs are only of theoretical value since
would cause violation of these constraints are therefore they would imply an in nite sample of ‘moves’ (new points
excluded. generated from the old con gurations, as in step 2.1.1 in
Figure 3).
Complex performance targets Furthermore, the SA algorithm has many variants added
Several performance targets involve combinations of to it, depending on how the pseudo-temperature parameter T
continuous and=or binary variables, which introduce signif- is reduced from one outer iteration to the next. Different
icant complexity or bilinearities into the problem formula- schemes obviously affect the percentage of uphill moves
tion. An example on an oil re nery is a furnace ring limit, accepted at each inner iteration loop and the number of such
where there exists a maximum heating duty in the furnace. iterations required to reach ‘thermal equilibrium’ (or satisfy
The heat required will depend on both the ow rate and the stochastic termination criteria). Finally, the use of the
temperature of the incoming stream(s). In the work reported probabilistic acceptance of an uphill move (particularly
here, the constant ow rate assumption reduces such a limit frequent at lower pseudo-temperatures) follows a Boltzmann
into a simple constraint, based on inlet temperatures. Our probability distribution function according to the basic SA
experience is that such constraints can often render the algorithm presented.
scheduling problem intractable or very dif cult to solve, What distinguishes SA algorithms from random genera-
and thus we do not consider local control actions such as tion of con gurations s0 in the neighbourhood of a previous
partial bypassing of exchangers in order to avoid pressure
drop limits.
1. Get an initial system con guration s and an initial
Selection constraints temperature T
Operational protocols and management strategy will 2. While outer loop stop criterion not satis ed do:
frequently result in restrictions on the allowed combination 2.1 While inner loop stop criterion not satis ed do:
of cleaning actions. For instance, a unit will rarely be 2.1.1 Select a trial solution s0 derived from s.
allowed to be cleaned immediately after it has been returned 2.1.2 Let D ˆ c…s0 † ¡ c…s†
to service; thus can be expressed as: 2.1.3 If D < 0 set s ˆ s0
2.1.4 If D > 0 set s ˆ s0 with probability
yn;p ‡ yn;p‡1 ¶ 1 …13† exp…¡D=T †.
2.2 Reduce ‘‘temperature’’ T.
Selection constraints can also be used to preclude combina- 3. Report best solution found, along with other near
tions of cleaning decisions that would automatically lead to solutions for comparison.
violations of performance targets. These combinations may
be known from experience, or identi ed by simulation. Such Figure 3. The simulated annealing algorithm.
con guration s, is the ability gradually to restrict costly parameters. It is thus best to adopt the most simple meaning
con gurations and introduce a type of descent property by and ease of estimation of these parameters.
gradually requiring more and more accepted con gurations A signi cant issue concerning SA and TA algorithms is
to have a smaller cost function. Performing uphill moves for when to stop the outer and inner loops. For the outer loop
some of the time allows the routine to escape from local we select the following criteria:
minima, at least at the beginning of the run. It can be seen
(and indeed as SA is practised) that an initial sample is (1) The number of consecutive outer iterations without
necessary to evaluate the rst temperature parameter so that accepting any new con gurations reaches n1 (typically
a certain percentage of moves is accepted at the beginning. 5–10 have been used in our BTA implementation). A
Thus, depending on the function values, a suitable range of counter is incremented while no outer iterations yield
initial temperature is identi ed. new accepted con gurations, and is reset to zero when
The use of an arti cial temperature parameter and an one is accepted.
arbitrary reduction scheme have little relation to the actual (2) The number of consecutive outer iterations satisfying:
optimization problems attempted to be solved by SA. In our sin
opinion a more practical and simple algorithm is that of µ e1 …14†
threshold accepting (TA) and our variant, backtracking min
threshold accepting (BTA), which is presented directly in reaches the value of n2 iterations. The parameters min
Figure 4. and sin are the average and standard deviations of all
The TA algorithm is not new; indeed it can be found in the samples considered (total, not only the accepted
Dueck and Scheuer13 and Nurmela11. The TA algorithm is
ones) in the inner iteration; e1 is a user accuracy
effectively a very close variant of the basic SA algorithm.
Notably, acceptance is performed without a Boltzmann parameter for the coef cient of variation (typically set
probability distribution function, and in our scheme the to a small enough statistical signi cance value, e.g.,
threshold is simply reduced by a constant parameter a1, 10¡3 or 10¡4 ). Again, a counter is incremented while
where a1 < 1 and is used as a linear scaling factor. consecutive outer iterations satisfy the criterion in
Our modi cation is the addition of a backtracking step in Equation (14), and is reset to zero on the rst outer
the case where a given number of new con gurations do not iteration that does not satisfy it.
gain acceptance. This can also be applied to the SA basic (3) The total number of function=system evaluations. This is
algorithm. Since our intention was not to work with SA, but usually set to a suf ciently large value (in our case we used
with a closely related method requiring fewer arti cial typically 100 000 function evaluations, although in the
parameters, we provide no comparison of the two. In runs presented this bound was hardly reached). Alterna-
Figure 4, the parameters a1 , a2 are both less than unity. tively, CPU time may be used to terminate lengthy runs
Parameter a2 is a scale factor used to increase the threshold and present the best solution found to that point.
close to its previous value, following a failure to accept a
new move in the inner iteration loop. The BTA algorithm, as presented in Figure 4 with the
In general, comparisons of highly stochastic methods are termination=stopping criteria enumerated above, was imple-
not very useful, especially in generic methods that do not mented in standard ANSI C yielding a portable software
attempt to exploit speci c properties of a given problem. package. The particular implementation of a HEN cleaning
These methods are termed ‘meta-heuristics,’ implying that formulation abstraction within the framework of BTA is
the optimization is achieved by suitably modifying (optimiz- considered in the next section.
ing) the parameters of an algorithm to solve a particular
IMPLEMENTATION OF THE HEN
problem, such as in SA and TA. This displacement of the
SCHEDULING PROBLEM
optimization to the tuning of the algorithm to improve the
problem’s cost function implies that the performance of each Solution by the Outer Approximation Method
algorithm can be maximized by suitable modi cation of its
The MINLP scheduling problem consists of the above
performance equations, together with the input variables
and constraints set. Solutions are generated by simulta-
neously solving the MINLP problem for all periods under
1. Get an initial system con guration s and an initial consideration.
threshold Dnew ˆ Dold The case studies presented here were implemented in the
2. While outer loop stop criterion not satis ed do: GAMS modelling language6 and solved using DICOPT‡‡5
2.1 While inner loop stop criterion not satis ed do: on a Sun SPARC 10 workstation.
2.1.1 Select a trial solution s0 derived from s.
2.1.2 If c…s0† < c…s† ‡ Dnew set s ˆ s0 Implementation of the Backtracking
2.2 If s has changed reduce threshold: Threshold Algorithm
Dold ˆ Dnew ,
Dnew ˆ Dold £ a1 . Stochastic search methods such as TA and SA are
2.3 If s did not change increase threshold: applicable in general where system con gurations (in parti-
Dnew ˆ Dold ¡ …Dold ¡ Dnew † £ a2 cular, combinations of binary decision variables in the HEN
3. Report best solution found, along with other near scheduling) can be performed very cheaply. In other words,
solutions for comparison. as several thousands of system con gurations will be
evaluated, the associated objective (and constraint) function
Figure 4. The backtracking threshold acceptance algorithm. evaluations must require very small CPU times.
while the second is a 25-unit network, as summarized in search. The BTA searches were performed using two differ-
Figures 6 and 9, and Tables 2 and 4, respectively. Both ent types of ‘moves’, that is, search patterns. The rst type
networks are based on a crude oil preheat train. Linear of search (move-1) is based on globally changing, at
fouling behaviour is considered with all units starting from random, any binary variable indicating cleaning by ipping
an initially clean condition. The length of each interval (i.e., its value between 1 or 0. In this way the number of cleanings
cleaning and processing) is two weeks. Therefore, the length is not constrained, and it varies during the search procedure.
of each period is one month. When a unit is taken off line The second type of search (move-2) is carried out in an
for cleaning, the cold stream ow rate is bypassed to the independent run from the rst, using xed numbers of
next unit. In the case of the hot streams, the ow rate is cleanings for each unit in the operating horizon, equal to
either bypassed or diverted to parallel units. For instance, if the number found during the move-1 type search pattern of
unit 1A (Figure 6) is being cleaned then the hot stream ow the previous run. In effect, the cleanings of each unit are
rate is diverted to unit 1B. moved at random time slots and evaluated with the BTA
The rst case study introduces the complexities in re n- algorithm.
ery networks caused by interconnecting hot streams. The The number of inner iterations performed by the algo-
second, which resembles a real network more closely, rithm was xed initially at 50 until the 50th outer iteration
contains a signi cant amount of duplication, where several was reached. From then on, each outer iteration added two
units (and cold streams) are used in order to reduce the extra function evaluations to the inner cycle. This was done
impact of cleaning one exchanger. Duplication not only so that when the outer cycles appeared to reach equilibrium,
increases the size of the problem, but introduces issues of more samples were tested in the binary variable space to
degeneracy, where notionally different schedules are in fact allow the possibility of nding better solutions. The
identical, with interchangeable exchanger labels. complete settings of other parameters of the BTA algorithm
The results presented here are the best from a number of are given in Table 1; the settings were kept strictly the same
possible solutions reached using different starting points. for all reported case studies. The values in Table 1 were
The results are represented in terms of cost saving, when obtained by trial and error, for example, varying a1 from
cleanings occur, in comparison with the cost of the ‘no 0.99 to 0.8 and noting the effect on performance. The
cleaning’ scenario: threshold criterion was found by testing values, in increasing
size (if too small), to give 80–90% acceptances of the new
Cost…uncleaned† ¡ Cost…cleaned†
Saving ˆ ¢ 100 points generated around the starting point. This tuning
Cost…uncleaned† process took 1–2 hours. All runs reported are in CPU
…16† seconds on a SunSparc 10 workstation used in the
GAMS runs.
In order to obtain solutions with the BTA algorithm, ve
independent runs were performed, each of which was started
with a different initial seed provided by a random number
Case Study 1
generator (uniform distribution). It will be seen that, as
expected with stochastic search methods such as TA and Figure 6 is a schematic of the 14-exchanger network,
SA, each run produces different solutions at the end of the which features three cold streams and seven hot streams.
Figure 6. HEN for case study 1. Values of heat load, in MW, are given at initial, clean, condition: Solid lines, cold (crude) streams; dashed lines, hot streams;
SS, sidestream; PA, pumparound; VAC, vacuum tower; Atm, atmospheric tower.
Table 2. Stream data for Case Study 1 (variable values given at initial, clean, condition).
HEX Tinh ( C)
¯
Tinc ( C)
¯
Fh (kg s¡1 ) Fc (kg s¡1 ) Cph (kJ kg¡1 K¡1 ) Cpc (kJ kg¡1 K ¡1 ) U0 a (kW m¡2 k¡1 ) A (m2 ) R_ f ¢ 1011 (m2 K J¡1 )
1A 334b 210 17.4 46.0 2.8 2.4 0.5 55.6 1.9
2A 286b 191 22.8 46.0 2.9 2.4 0.5 61.3 1.8
3A 249b 178 9.5 46.0 2.8 2.4 0.5 91.0 1.6
1B 334b 210 17.4 46.0 2.8 2.4 0.5 55.6 1.9
2B 286b 191 22.8 46.0 2.9 2.4 0.5 61.3 1.8
3B 249b 178 9.5 46.0 2.8 2.4 0.5 91.0 1.6
4 254 167 45.5 95.0 2.9 2.3 0.5 67.1 1.5
5 205b 161 55.8 95.0 2.6 2.3 0.5 67.2 1.1
6 285 135 34.8 95.0 2.8 2.3 0.5 110.1 1.5
7 237b 116 49.7 95.0 2.6 1.92 0.5 121.6 0.8
8 170b 101 49.7 95.0 2.6 1.92 0.5 112.9 0.8
9 197 50 55.8 95.0 2.6 1.92 0.5 208.3 0.6
10 296b 45 3.3 95.0 2.9 1.92 0.5 8.9 0.9
11 194 26b 19.1 95.0 2.8 1.92 0.5 56.6 0.6
a
U0 ˆ Uclean .
b
Fixed temperature inputs.
Figure 7. Optimal cleaning schedule (a) and variation of CIT with time (b) for Case Study 1, obtained by GAMS with Cc ˆ £4000.
to add a number of periods after the NPth period in which periodicity in the schedule, with units 7–9 and 11 being
no cleaning is allowed. These ‘end effects’ illustrate the cleaned near the middle of the operating period, but the
need to (i) manage problems with a large number of periods, distribution obtained is not immediately obvious. The gure
such that the schedules obtained are less sensitive to the also shows that no more than two exchangers are scheduled
horizon boundaries, and (ii) consider the form of the to be cleaned in any one period; this is a realistic result
objective function carefully. which is not forced by the constraint set. The latter could,
The schedule presented in Figure 7(a) includes some moreover, be constructed such that only one cleaning action
intuitive and non-intuitive results. The exchangers that is allowed per period. A more sparse schedule would be
foul most seriously (units 1A=B, 2A=B) are cleaned most obtained by increasing the cleaning cost, limiting the total
often, and the distribution of numbers of cleaning actions in number of cleaning actions allowed or deciding that units
the hot section (1A=B etc.) is symmetrical, as would be with less severe fouling are not cleaned at all.
expected. Units 3A=B, however, are cleaned less often than The BTA schedule in Figure 8 exhibits the same features
unit 4, despite similar fouling rates, illustrating the impor- discussed in the GAMS results (Figure 7). Two more clean-
tance of unit 4 and the effect of coupling in the network. ing actions are performed in this case, although cleaning
Likewise, unit 10 is not cleaned at all. There is a quasi- starts at the fth period and ends at the 30th period
Figure 8. Optimal cleaning schedule (a) and variation of CIT with time (b) for Case Study 1, obtained by BTA algorithm with Cc ˆ £4000.
(cf. starting at the sixth period in the GAMS results). Similarly, MINLP problem, as solved by GAMS, was relatively large,
unit 10 is not cleaned and ‘end effects’ are very evident. with 518 binary variables, 7808 continuous variables, and
The decay in CIT whenever unit 1A is being cleaned is 10 1707 constraints. The CPU times for the OA method
greater than the decay in the GAMS case. This is due to the (GAMS) are reasonably low. However, a feasible starting
fact that units 1B and 2B are being cleaned simultaneously, point is necessary to achieve convergence of the problem in
with a commensurate effect on CIT. The difference in the these CPU times and some further work was required for
uncleaned objective values …k£1320 cf. k£1331) arises from this initialization. A set of 50 starting points has been used
the different numerical methods used to solve the equation and the best solution was achieved at the 11th 1oop after
sets, and represents a º l% difference. This difference is 15 887 iterations.
also evident in the results for Case Study 2. Table 3 shows that the two techniques yield schedules
Table 3 shows a summary of the cost performance and with effectively the same performance. The OA method
computational results for the two methods. The underlying proved to be faster in this case, as it was able to con-
Table 3. Total cost saving in comparison with the ‘no cleaning’scenario for
Case Study 1.
most appropriate schedule for a plant will frequently involve
further operating considerations, that is, extra selection
Best objective CPU time constraints, which serve to reduce the solution space. The
Model C …£†c † in k£ (Saving in %) (CPU s) NC range of move-2 results …£k755¡769† for this case study
Uncleaned — 1320 441 — was consistently lower than those obtained with move-1
OA 4000 755 (42.8) 199 25 types …£k771¡788†: it will be seen in Case Study 2 that the
BTA 4000 758 (43) 220 27 ranges for the two move types overlapped (Table 6) to a
Move 2 761 216 27 great extent.
Move 2 762 194 27
Move 2 766 283 26 Figures 7(b) and 8(b) show the variation of CIT with time
Move 2 769 195 27 for the optimal schedule and when no cleaning occurs. ‘End
Move 1 771 210 26 effects,’ as described above, are evident. Large drops in CIT
Move 1 774 265 27 could be managed by the use of constraints of the type
Move 1 784 290 27
Move 1 785 195 27
CIT > CITL . It is noteworthy that the CIT for the cleaned
Move 1 788 199 27 networks in both gures does not drop below the nal value
of CIT (uncleaned).
Figure 9. HEN for Case Study 2. Values of heat load, in MW, are given at initial, clean, condition. Solid lines, cold (crude) streams; dashed lines, hot streams;
SS, sidestream; PA, pumparound; VAC, vacuum tower; Atm, atmospheric tower.
HEX Tinh (¯ C) Tinc (¯ C) Fh (kg s¡1 ) Fc (kg s¡1 ) Cph (kJ kg¡1 K ¡1 ) Cpc (kJ kg¡1 K¡1 ) U0 a (kW m¡2 k¡1 ) A (m2 ) R_ f ¢ 1011 (m2 K J¡1 )
1A 334b 210 8.7 23.0 2.8 2.4 0.5 21.3 1.9
2A 286b 191 11.4 23.0 2.9 2.4 0.5 29.7 1.8
3A 249b 180 4.8 23.0 2.8 2.4 0.5 31.4 1.6
1B 334b 210 8.7 23.0 2.8 2.4 0.5 21.3 1.9
2B 286b 191 11.4 23.0 2.9 2.4 0.5 29.7 1.8
3B 249b 180 4.8 23.0 2.8 2.4 0.5 31.4 1.6
1C 334b 210 8.7 23.0 2.8 2.4 0.5 21.3 1.9
2C 286b 191 11.4 23.0 2.9 2.4 0.5 29.7 1.8
3C 249b 180 4.8 23.0 2.8 2.4 0.5 31.4 1.6
1D 334b 210 8.7 23.0 2.8 2.4 0.5 21.3 1.9
2D 286b 191 11.4 23.0 2.9 2.4 0.5 29.7 1.8
3D 249b 180 4.8 23.0 2.8 2.4 0.5 31.4 1.6
4A 293 132 23.0 47.4 2.8 2.3 0.5 26.7 1.5
4B 205 151 28.0 47.4 2.6 2.3 0.5 35.4 1.1
5A 254b 158 17.4 47.4 2.9 2.3 0.5 79.1 1.5
5B 293b 132 23.0 47.4 2.8 2.3 0.5 29.2 1.6
6A 205 151 28.0 47.4 2.6 2.3 0.5 35.4 1.1
6B 254 158 17.4 47.4 2.9 2.3 0.5 79.1 1.5
7A 237b 116 25.0 47.4 2.6 1.92 0.5 60.8 0.8
7B 196 116 25.0 47.4 2.6 1.92 0.5 80.3 0.8
8 164 100 49.6 95.0 2.6 1.92 0.5 129.2 0.8
9A 194 73 55.8 95.0 2.6 1.92 0.5 109.6 0.9
9B 180 46 55.8 95.0 2.6 1.92 0.5 96.6 0.9
10 296b 69 3.3 95.0 2.9 1.92 0.5 8.5 0.6
11 203 25.6b 19.1 95.0 2.8 1.92 0.5 56.6 0.6
a
U0 ˆ Uclean .
b
Fixed temperature inputs.
The best schedules obtained for a two-year horizon for the Study 1. Some of the units are not cleaned at all, duplicated
OA and BTA methods are presented in Figures 10 and 11, units (1A–D) exhibit symmetry, and the number of cleaning
respectively. The two-year OA problem featured 625 binary actions per month is µ 2. Units 4A and 4B, corresponding
variables, 9401 continuous variables, and 12 850 constraints. to unit 4 in Case Study 1, are again cleaned more often than
As in Case Study 1, 50 different starting points were used and those adjacent, con rming the importance of these units in
the best locally optimal schedule is presented here; the network. The ‘end effects’ discussed previously are very
6575 CPU s were necessary to obtain a solution with objec- evident, owing to the shorter operating period and give rise
tive function parameters Cm ˆ Cx ˆ 0 and Cc ˆ £5000 per to some congestion, with no cleaning occurring after 19
clean. months. This horizon is undesirably short compared to a
OA results have not been achieved for a three-year typical re nery operating horizon of 3–7 years.
horizon with 36 monthly periods …Dt cl ˆ Dt pr ˆ 15 days† The BTA results for the 24 month scenario in Figure 11
to date, owing to the size of the problem. Solutions for such show an even more centrally congested schedule, indicating
a scenario could be generated using a number of modi ca- that the boundary conditions are dominating this scenario. It
tions, including (i) increasing the length of the sub-periods, is very apparent that different schedules would be obtained
further constraining the schedule to the discretization if the objective function included a dependency of the state
pattern; (ii) exploiting the observed end effects by not of the network at the time tF , as outlined in Section 2. The
permitting any cleaning in the rst six or last six months; schedule presented in Figure 11 is likely to result in a
and=or (iii) starting from an initial, dirty condition. Table 5 furnace ring limitation: the low CIT value in month 13
includes results for the ‘free’ problem for a two-and-a-half- would be avoided by the use of temperature performance
year horizon, which is the longest horizon we have found to limits, which were not included in the constraint set used
be tractable for this problem using the OA method. The here. The number of cleaning actions is similar to that in
solution for the two-and-a-half-year scenario showed the Figure 10 (15 for BTA versus 17 for OA), although the
same features observed elsewhere. There is evidently a objective function values for the two techniques do not agree
limit to the size of the problem that can be solved using as well as in Case Study 1.
the MINLP formulation with the solver employed here. The BTA algorithm performance in the 24 month case
Alternatively, linearization of equations such as (5) could shows the trends observed in Case Study 1, where 10
be used to reduce the complexity of the simulation model. different schedules were identi ed within a narrow range
Inspection of the solver records indicated that the appli- of objective function values (here, k£493¡509. Type 2
cation found the simultaneous solution of the large nonlinear moves again proved more effective that type 1 moves in
problems (NLP) dif cult. This observation suggests that a three of the ve seeds, requiring comparable or noticeably
tailored NLP solver, which exploited the sequential nature of smaller CPU times.
the multi-period simulation as in the BTA model, would The results for the 24 month scenarios in Table 5 also
allow the OA method to be applied to larger problems. indicate the trends identi ed in Case Study 1. Each solu-
The optimal OA schedule and CIT–time pro le shown in tion technique identi ed optimal schedules with similar
Figure 10 illustrates many of the features discussed in Case objective function values, yet different cleaning patterns
Figure 10. (a) Optimal cleaning schedule and (b) variation of CIT with time for Case Study 2, obtained by GAMS with Cc ˆ £5000.
as evidenced by the NC values. The computational times 2695 CPU s, was nearly twice that of the type 1 moves
were again similar, indicating that where the OA method …3965 CPU s†.
was able to achieve a solution, it was not inferior to the The schedule in Figure 12 shows a ‘more acceptable’
BTA method. distribution of cleaning actions compared to the congested
It is noteworthy that, whereas the MINLP formulation patterns in Figures 10 and 11, although it does feature an
was unable to solve the 25 heat exchanger problem over the undesirably large drop in CIT in period 12. Such deviations
36-month horizon, the BTA algorithm was able to solve could readily be avoided by a simple performance target on
such large problems without any additional complications or CIT. Again, units 4A=B are cleaned more often than the
tuning. Figure 12 shows the schedule and CIT–time pro le adjacent units and identical units are cleaned a similar
obtained for this scenario, and Table 6 summarizes the number of times. Unit 10, the smallest, with a small duty,
algorithm performance measures. The results for 10 trials is not cleaned at all, as well as units 3A and 3B.
lay again within a narrow range …k£819¡840† and the type 2 Case Study 2 is related to the con guration in Case
moves were superior in speed without compromising Study 1 and several points can be drawn from comparing
performance. The speed of the type 2 moves, totalling the schedules for a 36-month horizon (Case Study
Figure 11. (a) Optimal cleaning schedule and (b) variation of CIT with time for Case Study 2, obtained by BTA algorithm with Cc ˆ £5000.
1, 14 units, Figures 7 and 8: Case Study 2, 25 units, Figure savings achieved in Case Study 1 are higher (¹ 43% cf.
12, BTA only). The results cannot be compared directly as 35%) because the units in the former network are mostly
the exchanger areas in the second case study are not larger, while Cc remains similar, so that the energy
directly proportional to those in the rst case study. The recovered on cleaning an individual exchanger will be
Table 5. Total cost saving in comparison with the ‘no cleaning’ scenario for Case Study 2.
Model Cc (£) Horizon length (month) Best objective in k£ (Saving in %) CPU time (CPU s) NC
Uncleaned OA — 24 622 717 —
Uncleaned BTA — 24 589 — —
OA 5000 24 525 (15.6) 6575 17
BTA 5000 24 493 (16.3) 184 15
OA 5000 30 730 (28) 544 31
BTA 5000 36 819 (35) 490 34
Figure 12. (a) Optimal cleaning schedule (three-year scenario) and (b) variation of CIT with time for Case Study 2, obtained by BTA algorithm with
Cc ˆ £5000.
greater. Comparison of the total computational time Discussion and Evaluation of OA and BTA Results
required to solve the BTA models (14 units – 2267 s
The results for both case studies indicate that both these
cf. 25 units – 6624 s) indicates a quadratic rather that a
techniques can be used to solve the multi-period MINLP
linear relationship with problem size. The schedules iden-
scheduling problem in HENs subject to long-term fouling.
tify the same groups of units as having most impact on
Similar objective value ranges and numbers of cleaning
network performance, so that the network is sensitive to
actions are obtained, although the optimal schedules differ.
fouling in these units and will thus clean them more
The BTA schedules tended to give more ‘bunched’distribu-
regularly. The converse also holds. This identi cation of
tions, which could be proscribed by a suitably designed
network sensitivities, discussed in other contexts previously
constraint set.
by Kotjabasakis and Linnhoff18, could be exploited when
Direct comparison of the CPU times for OA and BTA
considering retro tting options in order to improve
techniques can best be interpreted in terms of comparable
network performance.
speeds being achieved in these case studies. The difference of cleaning actions was often non-intuitive. The case studies
in performance for the largest problem, where the OA presented here were freely posed, with relatively few
technique failed to converge, is interpreted here as a constraints: faster solution speeds and more ‘acceptable’
dif culty in the optimization solver used. The tailored solutions are expected when constraints based on more
NLP solver in the BTA algorithm, which exploited the detailed operating criteria are included.
sequential nature of the multi-period problem, rendered Both techniques are able to generate optimal cleaning
this technique more robust and suggests that even larger schedules, which can now be compared with other potential
horizons could be addressed by this approach. strategies for mitigating fouling in heat exchanger networks.
The combinatorial nature of the problem, added to a Furthermore, the schedules identify those key exchangers in
highly non-convex solution space with many local optima, the network that are sensitive to the effects of fouling and
clearly presents a challenge to both solution techniques. whose performance may need to be augmented via retro-
tting strategy.
Future applications of the BTA methodology will focus
CONCLUSIONS on expanding the spectrum of case studies of cleaning in
This paper presents an alternative approach to the solution HENs, including varying ow rates, pressure drop limita-
of the MINLP problem generated by considering the HEN tions, and pumparound streams. Another possible activity is
scheduling problem in a multi-period formulation. The the consideration of dynamic effects caused by the insertion
approach is based on a meta-heuristic search method, the or removal of units within a network.
threshold accepting algorithm (TA), with backtracking
variation (BTA) that potentially allows it to escape from
premature termination in a local region of the search space. APPENDIX A
The BTA algorithm has proved reasonably straightforward Operational Constraints for 14-Unit Case Study
to implement and has been compared with a classical
optimization method (OA) in solving two case studies A set of operational constraints is provided for Case
based on oil re nery crude oil preheat trains. Study 1. These relationships could be set by consideration
The method is compared with a mixed integer-nonlinear of performance targets, or acceptable operating practice. In
programming problem formulation that has been described practice, a combination of operational constraints and
in previous work (see for example, ref. 2). The TA algorithm bounds on performance are likely to be employed.
as implemented is shown to be very competitive with the For all periods p ˆ 1; 2; . . . ; NP the following opera-
classical method. CPU times are signi cantly small tional constraints must hold true:
compared to the size of the problems solved with the BTA (I) Only one ‘hot end’ exchanger cleaned at a time:
algorithm, and the solutions are in some instances better
than those obtained with the classical method. y1A;p ‡ y2A;p ‡ y3A;p ¶ 2 …A1†
Overall, the performance of the BTA algorithm appears
stable and applicable to realistic size case study sizes with- y1B;p ‡ y2B;p ‡ y3B;p ¶ 2 …A2†
out complications. The drawback of such stochastic search
methods, namely their inability to guarantee global optima- (II) Maintaining ash and desalter temperatures
lity, is less problematic in these heat exchanger cleaning
problems as the formulation is multi-modal and local y4;p ‡ y5;p ‡ y6;p ¶ 2 …A3†
methods such as OA in themselves cannot be treated as y7;p ‡ y8;p ‡ y9;p ‡ y10;p ‡ y11;p ¶ 4 …A4†
more than heuristic descent algorithms as well.
The schedules obtained for the two case studies highlight (III) Vacuum residue rundown target
the importance of the form of the objective function. The
objective function used here did not consider the state of the y1A;p ‡ y1B;p ‡ y6;p ¶ 2 …A5†
network beyond the immediate horizon, resulting in notice-
able boundary effects for shorter horizon scenarios. The (IV) Atmospheric middle pumparound target
schedules showed several expected features, with key
exchangers being cleaned more often, but the distribution y2A;p ‡ y2B;p ‡ y4;p ¶ 2 …A6†
(V) Sidestream rundown temperature target Cf opportunity cost ¡ lost production, £ day¡1
Cm extra maintenance costs caused by fouling, £ day¡1
y3A;p ‡ y3B;p ‡ y11;p ¶ 2 …A7† Cp speci c heat capacity, kJ kg¡1 K¡1
CIT coil inlet temperature, ¯ C
(VI) Atmospheric top pumparound target CT (opt) total cost, optimal schedule, £
CT (uncl) total cost, no cleaning actions, £
y5;p ‡ y9;p ¶ 1 …A8† F mass ow rate, kg s¡1
MINLP mixed-integer nonlinear programming
For the rst period all units must be present: NC number of cleanings
NE number of heat exchangers
yn;1 ˆ 1; n ˆ 1; 2; . . . ; NE …A9† NP number of time periods
Q heat load, kW
where NE ˆ 14 and NP ˆ 36 for this case study. Rf fouling resistance, m2 K kW¡1
Rf 0 initial fouling resistance, m2 K kW¡1
R_ f n fouling rate, m2 K kW¡1
APPENDIX B t time, day
tF nal time, day
Operational Constraints for 25-Unit Case Study Dt time interval length, day
For all periods p ˆ 1; 2; . . . ; NP the following opera- T temperature, ¯ C
TL temperature lower bound, ¯ C
tional constraints must hold true: TU temperature upper bound, ¯ C
DTLM logarithmic mean temperature difference, K
(I) Vac residue rundown temperature target U overall heat transfer coef cient, kW m¡2 K ¡1
y1A;p ‡ y1B;p ‡ y1C;p ‡ y1D;p ‡ y6A;p ‡ y6B;p ¶ 5 U0 ; Uclean overall heat transfer coef cient: initial, clean status,
kW m¡2 K¡1
…B1† x fraction of ow rate
y binary variable to denote cleaning status
(II) Atmospheric middle pumparound target
Greek symbols
y2A;p ‡ y2B;p ‡ y2C;p ‡ y2D;p ‡ y4A;p ‡ y4B;p ¶ 5 a BTA effectiveness parameter
…B2† e user accuracy parameter for the coef cient of variation
m average deviation
(III) Sidestream rundown temperature target s standard deviation
Subscripts
y3A;p ‡ y3B;p ‡ y3C;p ‡ y3D;p ‡ y11;p ¶ 4 …B3† clean clean
c cold
(IV) Atmospheric top pumparound target h hot
in inlet
y5A;p ‡ y5B;p ‡ y9A;p ‡ y9B;p ¶ 3 …B4† l linear
n exchange unit, n 2 ‰1; NEŠ
(V) Vacuum pumparound target out outlet
p time period, p 2 ‰1; NPŠ
y7A;p ‡ y7B;p ‡ y8;p ¶ 2 …B5†
(VI) Only one ‘hot end’ exchanger cleaned at a time Superscripts
c cold
y1A;p ‡ y2A;p ‡ y3A;p ¶ 2 …B6† cl cleaning
h hot
y1B;p ‡ y2B;p ‡ y3B;p ¶ 2 …B7† i inlet
o outlet
y1C;p ‡ y2C;p ‡ y3C;p ¶ 2 …B8† pr processing
y1D;p ‡ y2D;p ‡ y3D;p ¶ 2 …B9†
(VII) Maintaining ash temperature
REFERENCES
y4A;p ‡ y5A;p ‡ y6A;p ¶ 2 …B10†
1. Wilson, D. I. and Vassiliadis, V. S., 1997, Mitigation of re nery fouling
:y4B;p ‡ y5B;p ‡ y6B;p ¶ 2 …B11† by management of cleaning, in Bott, T. R., Melo, L. F., Panchal, C. B.
and Somercales, E. F. C. (eds). Understanding Heat Exchanger Fouling
(VIII) Maintaining desalter temperature and its Mitigation, 299–307, (Begell House, NY).
2. Smaõ¨li, F., Angadi, D. K., Hatch, C. M., Herbert, O., Vassiliadis, V. S.
y7A;p ‡ y7B;p ‡ y8;p ‡ y9A;p ‡ y10;p ‡ y9B;p ‡ y11;p ¶ 6 and Wilson, D. I., 1999, Trans IChemE Part C, 77: 159–164.
3. Georgiadis, M. C., Papageorgiou, L. G. and Macchietto, S., 2000, Ind
…B12† Eng Chem Res, 39: 441–454.
4. Georgiadis, M. C. and Papageorgiou, L. G., 2000, Trans IChemE, 78:
For the rst period all units must be present: l68–179.
yn;1 ˆ 1; n ˆ 1; 2; . . . ; NE …B13† 5. Viswanathan, J. and Grossmann, I. E., 1990, Comput Chem Eng,
14: 769.
where NE ˆ 25 and NP ˆ f24; 36g for this case study. 6. Brooke, A., Kendrick, D. and Meeraus, A., 1996, GAMS Release
2.25: A User’s Guide (GAMS Development Corporation,
Washington DC, USA).
NOMENCLATURE 7. Bott, T. R., 1994, Fouling of Heat Exchangers (Elsevier, Amsterdam).
8. Smaõ¨li, F., Vassiliadis, V. S. and Wilson, D. I., 1999, Scheduling
A heat transfer area, m2 of heat exchanger cleaning in re nery heat exchanger networks,
Cc cleaning costs incurred for cleaning operations, £ per Proc. Engineering Foundation Conference ‘‘Mitigation of Heat
cleaning operation Exchanger Fouling and Its Economic and Environmental Implications’’
CE extra energy cost required due to fouling, £ kW¡1 day¡1 July 11–16, Banff, Canada.
9. Hewitt, G. F., Shires, G. L. and Bott, T. R., 1994, Process Heat Transfer ACKNOWLEDGEMENTS
(Begell House, NY).
10. Kirkpatrick, S., Gellatt, Jr., C. D. and Vecchi, M. P., 1983, Science, 220: The authors wish to acknowledge the support of Esso Petroleum (UK)
671. and Nalco-Exxon Energy Chemicals in association with the EPSRC under
11. Nurmela, K. J., 1993, Constructing Combinatorial Designs by Local grant GR=L 33658. The assistance of J. MeGillivary, A. Cracknell and
Search, Digital System Laboratory, Helsinki University of Technology D. Marnell from Esso is also gratefully acknowledged.
Finland, Series A: Research Report No. 27.
12. Schoen, F., 1991, Global Optimization, 1: 207.
13. Dueck, G. and Scheuer, T, 1990, J Comput Phys, 90: 161.
14. Moré, J. J., Garbow, B. S. and Hillstrom, K. E., 1980, User Guide
for MINPACK-1. Tech. Rep. Report ANL-80-74, Argonne National ADDRESS
Laboratory, Argonne, Ill.
15. Moré, J. J., Garbow, B. S. and Hillstrom, K. E., 1984, User Guide for Correspondence concerning this paper should be addressed to
MINPACK-1. (Prentice-Hall, New Jersey, USA). Dr V. S. Vassiliadis, Department of Chemical Engineering, University
16. Dongarra, J. and Grosse, E. (eds), 1999, Netlib Repository. World Wide of Cambridge, Pembroke Street, Cambridge CB2 3RA, UK.
Web Database http:==www.netlib.org= E-mail: vsv20@cheng.cam.ac.uk
17. Wilson, D. I., Condron, P. I. and Garrett, S., 1997, AIChE Symp Ser,
93: 305. This manuscript was received 26 February 2002 and accepted for
18. Kotjabasakis, E. and Linnhoff, B., 1986, Chem Eng Res Des, 64: 197. publication after revision 29 May 2002.