Professional Documents
Culture Documents
a r t i c l e in fo abstract
Article history: Hot rolling scheduling is a difficult problem in the steel processing industry. It involves many objectives
Received 28 April 2007 and constraints in both technical and practical respects. A two-stage scheduling method is proposed in
Accepted 21 October 2008 this paper. Batch planning of staple material is formulated as a VRPTW, which is solved with a modified
Available online 6 December 2008
PGA. Then, batches of the established units are optimized by adjusting rolling sequences using
Keywords: intelligent search algorithms to reach higher hot charge ratios. This method has been applied to a hot
Hot rolling production scheduling strip mill belonging to Baosteel in China. Our results demonstrate that the proposed technique can
Vehicle routing problem with time improve production efficiency and offer significant economic benefits.
windows & 2008 Elsevier Ltd. All rights reserved.
Heuristics
Intelligent search algorithms
0967-0661/$ - see front matter & 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.conengprac.2008.10.014
ARTICLE IN PRESS
substantially different from the parallel multi-furnace reheating heuristics based on expert rules, etc. Finally, we conclude in
mode, which is widely applied in most hot rolling plants today. Section 5.
Furthermore, a parallel multiple traveling salesman problem
(MTSP) strategy was proposed by Tang, Liu, Rong, and Yang
(2000) to model and solve this problem using a modified 2. Rolling batch planning models and algorithms
genetic algorithm (MGA). Although the method was effective,
some process constraints mentioned in that paper no longer exist 2.1. Problem description
today, with advances in manufacturing equipment. For example,
simultaneous changes of width, gauge and hardness of adjacent Most hot strip rolling production processes in large-scale steel
slabs are considered allowable in modern hot strip mills. plants have three different logistical modes as shown in Fig. 1.
Therefore, scheduling principles should be redefined consistent
with the new rolling constraints. Additionally, Tang, Luh, Liu, and
(1) After continuous casting, slabs are first temporarily placed
Fang (2002) presented an integer programming formulation with
into a heat preservation pit or a slab yard. Then, the slabs
a separable structure for steel-making processes. A heuristic
are taken out and loaded into reheat furnaces to be prepared
method based on Lagrangian relaxation and dynamic program-
for hot rolling. Consistent with initial slab temperatures
ming was developed to reduce the complexity of this scheduling
(300–800 1C), this process consists of the hot charge rolling
model. A commercial decision support system was developed
(HCR) mode and the cold charge rolling (CCR) mode. The ratio
by Cowling (2003) for hot rolling scheduling, and this provided
of these two modes is relatively high in most hot rolling
semi-automatic schedules that had some human involvement to
plants.
eliminate algorithmic instability and to control unforeseen
(2) After continuous casting, the higher temperature slabs are
production events.
directly loaded into reheat furnaces using conveyors. Then the
In this paper, we consider the production scheduling problem
hot rolling process begins. This production manner is called
of a hot strip mill operated by Shanghai Baosteel Co., Ltd, China.
direct hot charge rolling (DHCR), with temperatures exceeding
We approach the underlying problem with a two-stage hybrid
800 1C. It is an energy-saving production manner that allows
scheduling method that involves planning a batch of slabs and
for shorter wait times before reheating.
optimizing batches of the scheduled formed units. At the batch
(3) After being cut by continuous caster, the slabs are heated by
planning stage, a set of ‘‘coffin’’ shape rolling units are built up.
border heaters and rolled immediately. This is called direct
The staple material with single rolling type is modeled as a vehicle
rolling (DR), and the rolling temperature is close to 1200 1C.
routing problem with time windows (VRPTW), and the warm-up
This is the most effective production manner because of its
material section uses constraint satisfaction programming. At the
energy-saving qualities. However, the process is difficult to
optimization stage, the global scheduling solution is obtained
implement because it requires high-tech hardware and good
from a hybrid heuristic algorithm based on the TS and simulated
integration of casting and rolling.
annealing (SA). We establish principles of unit swapping and slab
swapping in TS and SA to generate new neighboring solutions.
Moreover, certain important practical factors that have not been Today, the logistic modes of the Baosteel hot strip mill are
mentioned in other literature to date are also taken into account mainly HCR and CCR, together with DHCR. The heat preservation
in this paper. For instance, we consider the influence of rolling pit and the slab yard act as buffers for connecting the continuous
sections on a single unit of a slab’s rolling surface grade, besides casting and the hot rolling processes. Compared with HCR and
formulating width, gauge and hardness jumps in the context of CCR, the proportion of DHCR scheduling in this plant is not high
adjacent slabs. In addition, jumps in three key temperatures in because of the complex production processes and specific rolling
the rolling process—namely discharging temperature, finishing restrictions associated with their hot strip mill. This paper mainly
rolling temperature and coiling temperature—are also taken into discusses the former two logistic modes. If DHCR is required for a
account in our rolling planning model. On the other hand, to manufacturing plant, scheduling systems can still be used to
guarantee continuity between continuous casting and hot rolling, generate feasible solutions using appropriate human–machine
and to enhance the hot charge ratio, our proposed method takes coordination and production time adjustments.
both real slabs and virtual slabs into consideration. Real slabs are Hot strip rolling planning can be described as follows. A large
those that have been produced and placed in a slab yard or heat number of slabs are first arranged into a series of rolling units. The
preservation pit, while virtual slabs are those that have not been number of units and number of slabs in each unit are initially
produced but have already been scheduled into the casting unknown. Fig. 2 depicts a series of ‘‘coffin’’ shaped units
process. Finally, the application software system for this two- consistent with a certain rolling width profile. There are warm-
stage method is showcased for a hot rolling plant operated by up material stages and staple material stages. In the warm-up
Shanghai Baosteel Co., Ltd, China. To date, the system has been stage, slabs are scheduled from narrow to wide. The key step is to
successfully operating for over two years. The operating results heat the rollers after they have been changed. In terms of
from the system demonstrate that our proposed scheduling technical approach, the slabs require a lower hardness value here.
solutions are more feasible and effective than existing methods. The majority of slabs in a unit are classified as staple material.
This paper is organized as follows. The logistical description Width profiles for this group are selected to avoid rolling marks on
of hot rolling processes are first described in Section 2, and we the adjacent steel coils. On the other hand, every manufactured
then propose planning models that include both single-type and
multi-type rolling units and their associated algorithms. In
Heat preservation pit/ Reheat furnace
Section 3, we establish an optimizing model that comprises both Slab yard
HCR/CCR
unit optimization sequences and slab optimization sequences. We Reheat furnace
Continuous Hot
also present hybrid intelligent search algorithms. The systemic DHCR Reheat furnace
rolling
casting
application of our scheduling method to the plant is demonstrated Reheat furnace
in Section 4, which illustrates that the proposed two-stage DR
approach is comprehensively superior than other methods. We
compare several aspects using standard evolutionary algorithms, Fig. 1. Logistic mode in hot strip reheating and rolling process.
ARTICLE IN PRESS
a rolling unit
rolling
width
warm-up material section staple material section
(1) the rolling width profile of each rolling unit is coffin shaped
(see Fig. 2); d1 d2
(2) continuous rolling length for an identical width is limited; Days
(3) jumps in rolling width, gauge and hardness must be smooth,
and their maximum values must be less than a certain Fig. 3. Calculation of the priorities of inner due date and working procedure date.
quantity;
(4) jumps in temperatures associated with discharging, finishing
rolling and coiling must be smooth and the maximum jump
levels are also restricted; In the above example, may be 300 slabs should be separated into a
(5) slabs with different surface grades must be rolled in different single batch.
rolling parts of a unit; In this paper, we describe a priority scheme that takes into
(6) total weights of single-type staple material must be within a account internal due dates and working procedure dates to select
certain range; and the available slabs before scheduling. Those high priority slabs are
(7) continuously rolling different types of slabs in a multi-type preferentially chosen to meet target supply quantities for down-
unit must be compatible with certain specific restrictions. stream processes. Eq. (1) shows how we calculate this priority.
PRI ¼ mo Po þ mp Pp (1)
We first consider the higher priority single-type unit in this
paper, and subsequently explore the multi-type staple material where Po and Pp denote the priorities of internal due date and
unit. working procedure date, respectively; the parameters m0 and mp
denote their corresponding weighted coefficients. Fig. 3 shows a
similar calculation of the two priorities. The x-axis represents the
2.2. Gathering the available slabs difference between the dates (the internal due date and the
working procedure date) and the planned production date, and
There are always a number of slabs awaiting scheduling; new the y-axis shows priority levels. If the difference is between d1 and
slabs are continually provided from the upper logistic layer, d2, the priority values satisfy the linear relationship. The rest of
known as the enterprise resource planning (ERP) system. In the axis is the maximum (less than d1) or the minimum (larger
general, the ERP system regulates internal due dates and working than d2). This also means that the closer the planned production
procedure dates for every slab according to customer order date to the two dates, the higher the priority that will be assigned
delivery times and the production capacity of the hot mill. It to the slab. In practical production scenarios, setting these
allocates target supply quantities for subsequent processes based parameters is based on the steel grade and product category.
on the order category, logistic situation, etc. The internal due date, The available slabs with higher priority rankings will be selected
which differs from the order delivery time, represents production to build the rolling units. The parameters will be adjusted using
deadlines inside the plant. The working procedure date represents any remaining materials. Since the priority ranking of each slab is
the prospective production date. Since each slab corresponds to a updated over time, it is often necessary to process particularly
specified customer order, it is impossible to schedule all available urgent slabs immediately.
slabs in the material set at one time. This might lead to slippage in
the due dates. For example, if there are 1000 available slabs in the
material sets, some of those listed as urgent order will likely be 2.3. Planning model and algorithms for the single-type staple
scheduled for later rolling times. This leads to slippage. In material scenario
addition, scheduling all the slabs might also lead to torpid
rescheduling when manufacturing failures occur. For these 2.3.1. Planning model
practical reasons, the available slabs should be selected and The production of single-type staple material involves the
handled batch by batch to maintain manufacturing performance. majority of rolling assignments. This problem is formulated as a
ARTICLE IN PRESS
VRPTW model. A VRPTW model is a generalization of the vehicle slab i from the continuous caster; Ttrans is the required transfer
routing problem (VRP) that includes allowable delivery time or time for the slab to move from the continuous caster to the rolling
time windows. Customer service starts within a time window mill; tir is the time that we start rolling slab i; tiR is the deadline for
defined by the earliest and the latest times. In the rolling rolling slab i based on its prospective charging mode; Mmin is the
scheduling problem, slabs and rolling units are viewed as possible minimum number of rolling units and Mmax is the
customers and vehicles, respectively. The penalty value that possible maximum number. The symbol int[a] allows us to
ðTypeÞ
includes jumps in width, gauge, hardness and so on represents calculate the maximum integer that is no larger than a. Wgtmin
ðTypeÞ
the costs of transportation. The due times in the casting process and Wgt max are the minimum and the maximum allowable
and in prospective charging modes for each slab are regarded as weights of each staple material, for which the rolling slab type is
time windows. Consistent with VRPTW, each slab must be Type. Wgt kðTypeÞ is the weight of a staple material of unit k, whose
arranged into one unit with a capacity that satisfies the chosen rolling type is Type. Additionally, the other variables are as
limits. The objective of our model is to load the slabs with follows:
minimum penalty and using the fewest vehicles given various
1 if rolling slab j right after slab i in unit k
constraints. The model is defined as follows: xijk ¼
0 1 0 otherwise
XM X
1 A
min @ cij xijk þ a ; M 2 ½M min ; M max (2)
KM 1 if slab i is arranged into unit k
k¼1 i;j yik ¼
0 otherwise
X
M
yik ¼ 1; i ¼ 1; 2; . . . ; n (3) 1 if the surface grade of slab i is p
sip ¼
k¼1 0 otherwise
X
m
1 if slab j is rolled after slab i in unit k; and their rolling widths are equal:
sip ¼ 1; i ¼ 1; 2; . . . ; n (4) zijk ¼
0 otherwise
p¼1
Eq. (2) is the objective function that minimizes the penalty costs
X
n
xijk ¼ 1; j ¼ 1; 2; . . . ; n (5) and maximizes the average rolling length of units. Eq. (3) is used
i¼1 to guarantee that any one slab is allocated to only one unit. Eq. (4)
is to guarantee that each slab has only one rolling surface grade.
X
n
Eqs. (5) and (6) denote the uniqueness in allocating slabs. Eq. (7)
xijk ¼ 1; i ¼ 1; 2; . . . ; n (6)
j¼1
ensures that the total rolling length in front of slab i with surface
grade p in unit k must be less than Rp. Eq. (8) illustrates the
X
m constraint of continuously rolling with the same width. Eq. (9)
Lik p sip Rp ¼ 1; i ¼ 1; 2; . . . ; n; k ¼ 1; 2; . . . ; M (7) shows the time window of slab i. Eq. (10) is the calculation of the
p¼1
possible minimum units, based on the restriction of rolling
surface grade. Since R1 indicates the strictest surface demand, it
X
n
zijk lj pW same ; k ¼ 1; 2; . . . ; M; i ¼ 1; 2; . . . ; n (8) is used to obtain the possible maximum number of units that are
j¼1 given in Eq. (11). Eq. (12) shows the weight constraint for a certain
type of staple material.
t i;cut þ T trans pt ir pt iR ; i ¼ 1; 2; . . . ; n (9) VRPTW is an NP-complete problem (Savelsbergh, 1985) and
" "Pp # # the solutions to this problem have been studied in recent years. To
q¼1 T q
approach most typical examples, Desrochers, Desrosiers, and
M min ¼ max int þ1 (10) Solomon (1992) formulated VRPTW in terms of a relaxation of set
p¼1;...;m Rp
partitioning. A dynamic programming technique was adopted to
" Pm # solve a 100-customer problem. However, the scale of that problem
q¼1 T q
M max ¼ int þ1 (11) is relatively small compared with our hot rolling batch planning
Rp
model. Homberger and Gehring (1999) compared a genetic
algorithm approach with evolutionary programming using several
ðTypeÞ
Wgt min pWgt kðTypeÞ pWgt max
ðTypeÞ
; k ¼ 1; 2; . . . ; M (12) benchmarks. The method showcased in that paper achieved
approximate solutions to routing problems that involve more
where n is the number of a certain type of slabs that have been
than 400 customers. Although the parameters in Homberger and
gathered; M is the number of rolling units; m is the categories
Gehring (1999) were of comparable size to our proposed planning
number of surface grade; cij represents the penalties to roll slab j
g model, a significant number of practical constraints were not
immediately after slab i, where cij ¼ pti þ pw h ht
ij þ pij þ pij þ pij þ
completely considered by these researchers.
pat
ij þ p ct
ij ; p t
i represents the penalty due to rolling time, and
g
pw h ht at ct
ij ; pij ; pij ; pij ; pij ; pij , respectively, represent the penalties due
to width jump, gauge jump, hardness jump, discharging tempera- 2.3.2. Parthenogenetic algorithm
ture jump, finishing rolling temperature jump and coiling Since there are a vast number of slabs in the hot rolling
temperature jump; K M is the average rolling length of these M scheduling problem at any one time, and because there are a
units; a is the penalty coefficient for average rolling length; Lik multitude of complex constraints, it is very difficult to obtain a
represents the total rolling length before slab i in unit k only if slab satisfactory solution—or even one that is feasible. Accordingly,
i is assigned to unit k; Rp is the constraint of largest allowable this paper proposes a modified parthenogenetic algorithm (PGA)
rolling length before the slab whose surface grade is p, usually associated with heuristic rules to solve this model. PGA is a special
R1oR2oyoRm; lj is the rolling length of slab j; Wsame is the genetic algorithm that resolves the crossover infeasibility be-
constraint of largest allowable continuously rolling length of the tween two chromosomes consistent with ordinal strings. The
slabs that are of same width. Tq is the total rolling length of method has been applied to certain fields (Katayama & Narihisa,
gathered slabs whose surface grade is q; ti,cut is the cutting time of 2001). By reproducing new individuals with the parthenogenetic
ARTICLE IN PRESS
operators, a new child in PGA will always come from a single The structure of these operations is specified in advance.
parent, but never from a pair of parents. As a result, the genetic During processing, if the memory swarm is filled, we then replace
operation tends to be simple and offers excellent computational individuals in the swarm with a number of new superior
efficiency. Furthermore, without crossover operations, PGA elim- individuals by means of an iterative procedure. Meanwhile, we
inates immature convergences and improves results. Taking into choose individuals to create new generations both from the
account these factors, our paper uses an n-bit integer string to memory swarm and from the colony. In terms of genetic
represent each individual in a colony. It defines gene exchange operations, the single-gene swap is designated the main parthe-
rules as parthenogenetic operators, which aims to enhance nogenetic operator for our purposes. We combine this operation
convergence speed and solution accuracy. with the principles of multi-gene swaps and gene-string moves.
Since the multi-gene swap operator is similar to the single-gene
one, we provide illustrative examples of single-gene swap and
2.3.2.1. Transformation from code to the feasible solution space. We
gene-string moves here.
assume that ni is the serial number of the ith-bit slab in an n-bit
string that is uniquely assigned to a certain n slabs of a given type.
The unit into which this slab is distributed will be defined as Single-gene swap:
ni(mod)M+1 (the operator mod is used to generate the re- The sequence number of the genes: 1 2 3 4 5 6 7 8
mainder). The value of M is assumed known. The n slabs will be The parents: 5 8 7 4 1 2 6 3
distributed according to sequences of the n-bit string into M The children: 5 2 7 4 1 8 6 3
rolling units, in order to generate the ultimate solution. Certainly, Gene-string move:
given a multitude of constraints that range from rolling length to The gene-string: 3 4 5
surface grade, such transformations may well generate infeasible The parents: 1 2 3 4 5 6 7 8
solutions. Several possible approaches can be used to process The children: 1 2 6 7 3 4 5 8
infeasible solutions in the context of evolutionary algorithms.
Options include deploying more effective encoding techniques, 2.3.3. Local search strategy
strengthening penalty forces consistent with the fitness value of As it approaches the global optimum during the solving
infeasible solutions, and making infeasible solutions feasible process, the entire colony may incline to some various local
using repair operators. In this paper, our infeasible solution space optima. A representative method that avoids this tendency
is large. To address this and to increase application reliability, we requires integrating multiple intelligent algorithms to strengthen
chose a heuristic operator to convert an infeasible solution into a the searching operation, or to provide an adaptive algorithm
feasible one. The transformation steps are as follows: (Shin, Kim, Choo, & Singh, 2004; Srinivas, & Patnaik, 1994).
Step 1. Initialize the current rolling length of each unit Lk However, all such methods are associated with long computational
(k ¼ 1,y,M) to zero; i ¼ 1. times. This necessarily increases the implementation complexity
Step 2. Check whether i4n holds true. If yes, move to step 6; for the scheduling system. In our research, we propose a local
otherwise, select the ith-bit slab and calculate k ¼ ni(mod)M+1 to search strategy that is suitable for engineering practice. Our
assign the slab to unit k. method enlarges the scope of search and guarantees algorithmic
Step 3. Compute Lk, and Lengthik to represent the current rolling efficiency. Given an n-bit integer string, its local operation can be
length of the slabs whose surface grade value is no higher than defined as a permutation of three slabs with equal surface grade
that of the ith-bit slab in unit k. If Lk satisfies Eq. (7), slab i can be that are randomly selected from the integer string. In other word,
arranged into unit k. Assign i ¼ i+1, and return to step 2; five new and different strings are produced as local sets from the
otherwise, if Lengthik satisfies Eq. (7), this means that this slab original string. The local search strategy then checks all the local
can be assigned to unit k but that rectification is needed. In this feasible solutions in order to obtain a better one. Our hot strip mill
case, go to step 4. If neither of these two conditions is met, slab i application showed the effectiveness of this strategy.
needs to be assigned repeatedly, in which case we go to step 5.
Step 4. Select the slabs whose surface grade values are higher
2.3.4. Controlling violations
than slab i in unit k. Exchange these with slab i consistent with
It is essential in practice to restrict jumps in width, gauge and
rolling length constraints until all the slabs in unit k satisfy the
hardness. In theory, this is implemented by constraining the
constraint. Set i ¼ i+1 and go back to step 2.
corresponding coefficients in Eq. (2). Jumps in temperature are
Step 5. Arrange the ith-bit slab repeatedly, by attempting to
assigned higher penalties, all of which may lead to violations
assign it to another unit using a sequence of current values of Lk
because of PGA randomness. Consequently, we use an expert-
where this parameter varies from minimum to maximum. Repeat
based revision method to guarantee the reliability and feasibility
step 3 until the ith-bit slab has been processed. Set i ¼ i+1 and go
of our scheduling processes. This includes the various operations
back to step 2.
of replacing, inserting and deleting. In addition, we propose that
Step 6. The transformation is completed.
once the most urgent slabs have been arranged into units before
When using this transformation, a phenomenon may occur
revision, they cannot be deleted. The arrangement operators are
whereby several integer strings map into an identical feasible
applied to the most urgent slabs before any others. In fact, the
solution. This would decrease the diversity of individuals. An
violations mentioned below emerge only very rarely. In manu-
approach that converts the mapped solutions into standard
facturing practice, these constraints are sufficiently weak in the
strings can avoid such phenomenon. This measure can be under-
process control layer—which appears below the planning and
stood as the reverse of the above-mentioned transformation,
scheduling layer in the plant—can always be used to make further
which forms an n-bit string from a series of M rolling units. The
adjustments. Accordingly, we consider that this violation restric-
detail of this process is similar to the aforementioned list.
tion process should be feasible.
Rule 1. Eliminating violations in respect of jumps in dischar-
2.3.2.2. Memory swarm structure and creating new individuals. To ging temperatures: select a suitable slab on the basis of priority
maintain a sufficient quantity of dominant individuals with high level from all available slabs. Ensure the slab satisfies the
fitness values, we used a memory swarm structure that preserves constraints of width, gauge, hardness and surface grade. Perform
a designated number of individuals who are superior to others. replace or insert operations accordingly.
ARTICLE IN PRESS
Step 8. Check whether the most recent slab type may be a term experiments in the rolling plant, production practices have
rolling segment. If yes, create a new staple material section and shown that a dependency-directed backtracking approach is
finish the search. Otherwise, delete the last rolling slab type and feasible given computational processing constraints in the context
record its type in the tabu list. Reset the search procedure to of normal amounts of warm-up material and normal numbers of
examine the final slab in the formed rolling segment. Return to rolling units.
step 6. To date, we have deployed our rolling batch algorithms and
Step 9. If the tabu list is full, output the current staple material have built a series of coffin-shaped rolling units.
information. If the staple material section has not been com-
pleted, set i ¼ i+1, and return to step 2 to commence a new search
procedure. 3. Optimization of the rolling plan
Step 10. Finally, check whether all the most urgent slabs have
been processed. If yes, the heuristic process finishes. Otherwise, 3.1. Problem description
rearrange the remaining slabs by ignoring all temperature
constraints. In other words, temperature violations can be One way to save energy and reduce setup costs is to use
subsequently removed in the process control layer. continuous casting procedures to improve the hot charge ratio.
Using both single-type and multi-type sections, almost all The temperatures of slabs that come out of the continuous caster
staple material slabs can be processed including all urgent slabs. A drop so rapidly that energy is required to reheat them before
few slabs may not undergo processing. Those can be assigned to rolling. In other words, increasing the charging temperature
the next batch after their priorities have been updated. The can save on both reheating time and energy. Accordingly, our
scheduling system in Baosteel demonstrates that one-off planning optimization objective is to improve rolling continuity and
ratios for the staple material can be above 97%. This figure meets increase charging temperatures. Although Section 2 of this paper
more than the relevant production requirements. proposes a rolling batch planning approach, this can only be
deployed from the current a batch of available slabs. To increase
the global hot charge ratio, we must optimize rolling planning
2.5. A warm-up material planning model based on CSP across several batches. The objective of this model is described by
Pn
C
Although the number of warm-up materials tends to be small Max i¼1 i (13)
in a rolling unit, processing such materials can significantly n
influence the performance of the warming rollers and the X
n
resulting roll quality. Compared with processing staple materials, Min ðt ir t i;cut Þ (14)
the warm-up material has its own rolling constraints and i¼1
production modes. The main constraints for processing warm- where n is the number of the all finished batches of slabs; Ci is the
up materials are: (1) steel grade and size of the planned slabs are hot charging sign of slab i, such that Ci ¼ 1 means that slab i is hot
subject to specific requirements; (2) jumps in width, gauge, charged and Ci ¼ 0 means that slab i is cold charged. The rest of
hardness and temperature must satisfy particular constraints; (3) parameters and constraints are similar to those listed in Section 2.
the first and the last slab in a section of warm-up material must Eq. (13) helps us maximize the global hot charge ratio. Eq. (14)
be within a given size range; (4) the capacity for processing minimizes the time for slabs to move from the continuous caster
warm-up material is usually limited by the number of slabs or the to the hot mill in order to reduce cooling. These two equations are
total rolling length. For example, in the hot strip mill, the complementary.
constraint was that each rolling unit must have more than four
slabs and that total rolling length must be less than 20 km. Lopez
3.2. Optimizing model and algorithms
et al. (1998) considered warm-up material and staple material in
the same way, the approach that may be appropriate for some
manufacturing plants. However, warm-up material slabs are Given the complexity of the hot rolling scheduling problem,
treated differently in the Baosteel plant, and as a result additional the entire optimization process in this paper is divided into a unit-
rolling constraints are necessary. This paper treats warm-up based optimization sequence and a slab-based sequence. The
material processing using the constraints satisfactory program- input to the optimization process is a few batches from the
ming (CSP) approach. planning stage. Their rolling sequence serves as the baseline for
The CSP has been widely studied across many application obtaining a new scheduling.
areas, including scheduling problems. Certain researchers have
proposed that it be solved using inconsistent assignments and 3.2.1. Optimizing the sequence for rolling units
rectification of constraint violations (Minton, Johnston, Philips, & This optimization process aims to improve the hot charge ratio
Laird, 1992). However, such a solution is never appropriate for hot by adjusting the sequence of rolling units without breaking the
strip rolling because of the presence of various complicated internal structure of each unit. A TS-based heuristic algorithm is
constraints, any of which make search times unpredictable. We applied. This is appropriate because this function only occurs
adopt a two-step approach to solve the CSP in our application occasionally within the broader scheduling problem. This paper
scenario. Firstly, we employ a dependency-directed backtracking defines two permutations of formed input units as the foundation
approach (Baker, 1995) to determine the last warm-up material of our operating heuristics. Both of these reflect the greedy nature
slab that connects with a section of staple material. A heuristic of our manufacturing sequences in two respects. List 1 is the
method that is similar to that used to process multi-type staple permutation sorted by average cutting time for the slabs in each
material is then applied to process a section of warm-up material. unit, and List 2 is sorted according to the latest cutting times for
Admittedly, certain scholars have argued strongly that the back- slabs from each unit. In addition to these heuristic definitions, the
tracking method has its drawbacks. It is too computationally concepts of neighbor unit and sub-neighbor unit are designated as
expensive and it is difficult to apply this approach in large-scale exchange elements by which a new feasible solution can be
computational scenarios such as for production planning and generated. The neighbors are the closest adjacent elements to a
scheduling (Kelleher & Cavichiollo, 2001). However, with long given unit in the above two lists, and the sub-neighbors are the
ARTICLE IN PRESS
sub-closest units in the lists. Based on the two arrays, eight charge ratios of the new state and current state. The new solution
different new unit permutations can be generated when a given can be selectively accepted based on the Metropolis Rule. The
unit does not appear at the head or at the tail of the arrays. Thus, slabs then enter the tabu list.
including the original, the best of these nine rolling sequences can Step 3. Check whether the solution is convergent (a condition
be identified by comparing their hot charging ratios. The selected of which is no change in the current best solution after 15
original unit can then be inserted into a tabu list. Consistent with iterations). If yes, update the annealing temperature and go to step
our experience of long-term simulation in the plant, this paper 4. Otherwise, repeat step 2.
empirically assumes that the maximum iteration number is Step 4. Check whether the finishing temperature reaches. If yes,
[N 0.9], where the operator [.] represents obtaining the integer, output the current optimal solution. Otherwise, repeat step 2.
and N is the number of rolling units.
3.2.2. Optimizing slab sequence 4. Application of the hot rolling scheduling method
Optimization that is achieved by adjusting the sequence of
slabs can help achieve a better hot charge ratio. Because so many The proposed scheduling method for hot strip rolling was
constraints exist in the planning stage, random modifications to successfully implemented in a hot rolling plant operated by
slab sequence may destroy the structure of planned units, and this Baosteel China. The production capacity of this plant is about
in turn may generate infeasible solutions. Fortunately, with 3.8 million tons of strip steel per year. This plant has a hot strip
longer-term simulation research we have discovered that an mill, whose maximum rolling width is 1580 mm, and three
approximate relationship between surface grade and rolling width walking beam reheat furnaces, which supply the heated slabs to
exists. In other words, the higher surface grade, the wider the this mill. The entire structure of our production scheduling
rolling width. We propose a heuristic method based on SA and TS, software system is illustrated in Fig. 4. This system is based on
in which a slab-swapping rule produces the new physical state. our proposed method and was developed using Pro*C tools for
Similarly, the swapping rule is defined as follows. Firstly, a background services, which involve scheduling algorithms, and
number of rolling units are selected, and from each, one slab is was operated on an IBM P_SERIES server. The graphical user
selected with the same surface grade and rolling type. Then, we interface (GUI) was written using Visual C# and TUXDEO
exchange these slabs to generate new solutions, and we identify middleware that connects with an ORACLE real-time database to
the best feasible solution from these as a result. Although we do balance the data load. Slab data for production materials is
not doubt the ability of SA to optimize, the optimal result depends collected from the Baosteel computer system. To date our hot
on a rigorous annealing process with a high initial temperature rolling production scheduling system has been in operation for
and a low annealing temperature. Therefore, a tabu structure is more than two years and has proved very helpful for plant
employed in this paper to assist the SA to achieve a high speed of logistics. Fig. 5 shows a typical operating interface for our system,
convergence. The detailed process is as follows: in which rolling width profiles and the jump trends of rolling
Step 1. Regard the rolling sequence as the initial solution after gauges are highlighted. By using the GUI, the schedulers can easily
optimization, and set the number of selected slabs consistent with change priority values or technical rules, and adjust the schedul-
the swapping rule. This paper uses a maximum value of 4. ing solution for consistency with an integrated production
Initialize the length of the tabu list and the parameters of SA, environment.
including initial temperature T0, finishing temperature Tm and The following are constraints that relate to the number of
cooling coefficient a. rolling units, certain penalty coefficient settings, and certain
Step 2. Randomly choose slabs that are not in the tabu list from algorithm coefficient settings. Table 3 lists constraints between
different units. Perform the swapping operation for the new state. the rolling surface grade and its corresponding rolling length for
Calculate the value of exp ((S-S0)/T), where S and S0 are the hot this hot strip mill. From this table, it is clear that the smaller the
ERP System
Manufacturing process
Fig. 4. The structure chart of hot rolling scheduling system.
ARTICLE IN PRESS
Table 3 Table 5
Relationship between surface grade and rolling length. Hardness penalty coefficients.
01 70 1 5
02 90 2 15
03 120 3 35
04 160 4 60
Table 6
Table 4 Gauge penalty coefficients.
Width penalty coefficients.
Gauge jump (mm) Penalty for up/down
Width jump (mm) Penalty
0.00–0.30 3 6
0–5 1 0.31–0.60 6 12
6–10 2 0.61–0.90 15 30
11–20 5 0.91–1.20 50 100
21–30 10 1.21–2.00 200 400
31–50 15
51–70 20
71–90 30 Table 7
91–115 50 Discharging temperature penalty coefficients.
116–140 70
141–165 90 Discharging temperature jump (1C) Penalty
166–190 120
191–215 150 0–5 1
215–250 200 6–10 3
11–15 6
16–20 12
20–30 30
430 1000
value of surface grade, the more rigorous the constraint of rolling
length, as previously explained. The penalty coefficients of width
jump, hardness jump and gauge jump are shown in Tables 4–6,
respectively. The penalty coefficients associated with jumps in meters of the PGA are as follows. The single-gene swap probability
discharging temperature, finishing rolling temperature and coiling is 0.9; the multi-gene swap probability is 0.15; the gene-string
temperature are listed in Tables 7–9. Table 10 describes the move probability is 0.15; the scale of the memory swarm structure
penalty coefficient associated with slab rolling time. The para- is designated as 20; the population of the colony is 50. In the
ARTICLE IN PRESS
Table 11
Comparison of planning results with four methods.
Method Batch Rolling unit Slabs Rolling Weight (t) Default of hot Average hot Time Batch
ID number length (km) charged slabs charge ratio (%) spending penalty
number
Table 12
Results of optimizing rolling sequence.
Rolling sequence Rolling unit ID Slabs number Rolling length (km) Weight (t) Hot charged slabs number after optimized
5. Conclusions
Fig. 7. The gauge jump trend in randomly selected 2 rolling units of Table 12. In this paper, we developed a hot rolling scheduling method for
a certain hot strip mill in Shanghai operated by Baosteel Co. Ltd.
Our approach formulates the scheduling problem using two
ninth slab. In terms of gauge jumps, given the width jump trend, stages, namely planning rolling batches and optimizing the rolling
these change relatively smoothly and meet all production sequence. A variety of intelligent search algorithms are used
requirements. to solve the problem. In a practical manufacturing scenario,
In order to demonstrate the effectiveness of this production production software based on a hybrid method that features
scheduling system in the Baosteel plant, Table 13 shows actual human–machine coordination improves scheduling flexibility.
manufacturing production statistics for this system during The results in Baosteel demonstrate the viability and effectiveness
March–August 2006. Production data includes average rolling of our proposed scheduling method.
ARTICLE IN PRESS
Table 13
Comparison of the production statistical data in two years.
Month Average weight of a rolling unit (t) Average rolling length of a unit (km) Average hot charge ratio Energy saving (standard coal equivalent) (t)
Acknowledgments Kelleher, G., & Cavichiollo, P. (2001). Supporting rescheduling using CSP, RMS
and POB—An example application. Journal of Intelligent Manufacturing, 12(4),
343–357.
This work was partially supported by the National Natural Kosiba, E. D., Wright, J. R., & Cobbs, A. E. (1992). Discrete event sequencing as a
Science Foundation of China (60534010, 60604026), and by the traveling salesman problem. Computers in Industry, 19(3), 317–327.
National High-Tech R&D Program of China (2007AA04Z156). The Lopez, L., Carter, M. W., & Gendreau, M. (1998). The hot strip mill production
cooperation of Shanghai Baosteel Co. Ltd., China, in this work is scheduling problem: A tabu search approach. European Journal of Operational
Research, 106(2-3), 317–335.
greatly appreciated. We would also like to thank the associate
Minton, S., Johnston, M. D., Philips, A. B., & Laird, P. (1992). Minimizing conflicts: A
editor and the anonymous reviewers for their valuable comments heuristic repair method for constraint-satisfaction and scheduling problems.
and constructive suggestions. Artificial Intelligence, 58(1), 161–205.
Peterson, C. M., Sorensen, K. L., & Vidal, R. V. V. (1992). Inter-process
synchronization in steel production. International Journal of Production
References
Research, 30(1), 1415–1425.
Savelsbergh, M. W. P. (1985). Local search in routing problems with time windows.
Assaf, I., Chen, M., & Katzberg, J. (1997). Steel production schedule generation. Annals of Operations Research, 33(4), 285–305.
International Journal of Production Research, 35(2), 467–477. Shin, D., Kim, J., Choo, J., & Singh, C. (2004). Optimal service restoration and
Baker, A. B. (1995). Intelligent backtracking on the hardest constraint problems. reconfiguration of network using genetic-tabu algorithm. Electric Power
Journal of Artificial Intelligent Research. Systems Research, 71(2), 145–152.
Cowling, P. (2003). A flexible decision support system for steel hot rolling mill Srinivas, M., & Patnaik, L. M. (1994). Adaptive probabilities of crossover and
scheduling. Computers and Industrial Engineering, 45(2), 307–321.
mutation in genetic algorithms. IEEE Transactions on System, Man and
Desrochers, M., Desrosiers, J., & Solomon, M. (1992). A new optimization algorithm
Cybernetics, 24(4), 656–667.
for the vehicle routing problem with time windows. Operations Research, 40(2),
Tang, L., Liu, J., Rong, A., & Yang, Z. (2000). A multiple traveling salesman problem
342–354.
Homberger, J., & Gehring, H. (1999). Two evolutionary metaheuristics for the model for hot rolling scheduling in Shanghai Baoshan Iron & Steel Complex.
vehicle routing problem with time windows. INFOR Journal, 37(3), 297–318. European Journal of Operational Research, 124(2), 267–282.
Katayama, K., & Narihisa, H. (2001). On fundamental design of parthenogenetic Tang, L., Luh, P. B., Liu, J., & Fang, L. (2002). Steel-making process scheduling using
algorithm for the binary quadratic programming problem. In: Proceedings of Lagrangian relaxation. International Journal of Production Research, 40(1),
the congress on evolutionary computation, Seoul. 55–70.