NIPPON STEEL & SUMITOMO METAL TECHNICAL REPORT No.

121 March 2019

UDC 658 . 51 : 669 . 1

Technical Report

Production Planning and Scheduling Technology

for Steel Manufacturing Process
Kuniharu ITO* Tetsuaki KUROKAWA
Masanori SHIOYA Hirokazu KOBAYASHI
Masatoshi AGO Junichi MORI

Abstract
The steel manufacturing process is a V-type production process that separates products
from natural raw materials so as to satisfy product quality and delivery times of each order
while orienting large lot variety of small lot orders, making large lots from each process.
Because of the large-scale and complicated production process, the burden of production
planning and scheduling work is high, and support needs by system technology are strong.
In this paper, we describe the development status and future prospects of optimization algo-
rithms to support decision making in our production planning and scheduling tasks.

1. Introduction production has been targeted. However, as explained above, condi-

In steel manufacturing processes, various types of steel products tions for manufacturing products vary from process to process and
are manufactured from iron ore, coal, and other raw materials im- delivery dates are also different. Therefore, it is difficult to plan a
ported from overseas based on requests by customers in various in- single production schedule throughout all manufacturing processes
dustries (e.g., automobiles, shipbuilding, bridges, and home appli- while balancing the quality, costs, and delivery dates, so such work
ances) through processes from blast furnace to converter, continuous depends on the expertise of experienced workers.
casting, rolling, annealing, and surface treatment. The production Meanwhile, technologies for supporting production planning
pattern is V-type (branching) flow shop. Product specifications in- and scheduling from the aspect of systems are being increasingly
clude various requirements based on the application of products demanded because one generation of experienced planners has been
such as quality of materials (e.g., strength and toughness), grades of giving way to another recently, manufacturing conditions are becom-
the inside and surface of slabs, and size (e.g., thickness and width). ing more difficult due to a shift to high-grade steel, and workloads
Specifications range in number from several thousands to tens of need to be reduced toward work-style reform. In addition, regarding
thousands although this varies depending on the product type. In ad- computer technologies, in addition to the enhanced performance of
dition, their manufacturing conditions consist of a combination of computers themselves, various algorithms represented by mathe-
various factors such as molten steel components, rolling size, an- matical optimization solvers have been advanced and they have en-
nealing temperature, and plating type. The variety of manufacturing abled support by systems in domains for which practical use used to
conditions is equivalent to that of product specifications. be difficult.
Production planning and scheduling determine processing tim- Against such a background, Nippon Steel & Sumitomo Metal
ing and sequence in each manufacturing process for each order Corporation has been developing technologies for supporting pro-
while manufacturing conditions and delivery date in each process duction planning and scheduling in various manufacturing processes
are satisfied. Comprehensive judgment in consideration of quality, such as raw materials, steelmaking, hot rolling, and logistics. Chap-
costs, delivery dates, and other various performance indexes is re- ter 2 introduces cases in which Nippon Steel & Sumitomo Metal
quired. In the steel manufacturing processes, in particular, large lot worked to optimize production and logistics simultaneously from
production in which products with the same manufacturing condi- raw material transportation to their physical logistics schedules at
tions are continuously manufactured is advantageous from the as- steelworks. Chapter 3 introduces the development of a scheduler
pects of quality and costs (including profitability) and thereby such considering the balance between productivity and cost (temperature)

* General Manager, Head of Laboratories, Intelligent Algorithm Research Center, Process Research Laboratories
20-1 Shintomi, Futtsu City, Chiba Pref. 293-8511

- 31 -
NIPPON STEEL & SUMITOMO METAL TECHNICAL REPORT No. 121 March 2019

targeted at steel mills. Chapter 4 shows a case in which an explicit million tons. Coal baked in a large furnace called a coke oven turns
solution technique in mathematical programming was applied to a to coke. Coke is put into a blast furnace with iron ore; they are melt-
problem involving complicated manufacturing conditions (con- ed at high temperatures and they chemically react. Through such
straints) targeted at hot rolling mills. processes, iron oxide contained in the iron ore is reduced by coke
Recently, support for planning production schedules throughout and it turns to iron containing carbon of slightly less than 5% (pig
multiple manufacturing processes is highly desirable in addition to iron). This pig iron is processed by adjusting the components in var-
schedules of individual manufacturing processes. As explained ious ways to make end products, such as sheets, plates, and billets. 1)
above, production lot conditions vary from manufacturing process As explained above, the steel industry handles large quantities of
to process, so increasing the size of individual lots in each manufac- raw materials, pig iron, half-finished goods, and products, so an
turing process hinders the synchronization of the production timing enormous quantity of goods is distributed every day and the logis-
between upstream and downstream processes, which increases tics cost is very high. 2) Therefore, in addition to regulating the logis-
workpieces and varies production periods. Chapter 5 introduces a tics of goods to stabilize operation and reduce costs, adjusting the
technology for estimating standard production periods throughout components of slabs is essential to satisfy customer requirements
multiple manufacturing processes highly accurately by machine and maintain the high quality of products. To achieve these, plan-
learning targeted at steel plate mills. Chapter 6 introduces a technol- ning schedules for transporting raw materials and planning produc-
ogy in which such standard production periods are used to support tion and logistics schedules in manufacturing processes are impor-
planning weekly schedules for manufacturing steel plates in consid- tant.
eration of the balance between larger lot size at steel mills and level- Against such a background, Nippon Steel & Sumitomo Metal
ling of loads in the refining process that follows the rolling process has been working to optimize production and logistics for raw mate-
while aiming at starting just-in-time production according to deliv- rials simultaneously aiming at streamlining and advancing the pro-
ery dates. duction and logistics related to raw materials throughout the compa-
To regulate production and logistics throughout manufacturing ny. Specific optimization targets are assigned to the Head Office and
processes, logistics between manufacturing processes needs to be steelworks at each region (Fig. 1). The Head Office determines the
regulated in addition to that in mills. Chapter 7 introduces the devel- outline of annual and termly schedules for multiple steelworks in
opment of a discrete event simulator that is a base for estimating consideration of advantages for the entire company. Steelworks plan
and controlling logistics. Chapter 8 introduces a case in which a daily schedules based on such schedules determined by the Head
mathematical optimization technique was applied to the instruction Office so that they can carry out daily operations within the outline.
of logistics of slabs (half-finished goods) between steel mills and Specifically, the Head Office plans the following schedules: (1) Ship
hot rolling mills. chartering schedules to secure ships to transport raw materials con-
Meanwhile, there are cases, depending on application targets, in sidering the tapping quantity (outputs of pig iron) and outputs of
which all conditions cannot be modelled on a computer and in coke at the entire company, (2) ship load-schedules to determine
which the presentation of grounds for planned schedules that is dif- which ships are assigned to which loading place (mine and coal
ficult for computers to achieve is required. Section 2.2 and Chapter mine) considering purchased quantity, (3) ship unloaded- schedules
9 introduce the development of a human cooperative scheduler as a to determine which ship that took in raw materials at a loading place
solution for such problems that are often seen in steel production is assigned to which steelworks, and (4) blending schedules of raw
planning and scheduling involving complicated work adjustment. materials to determine the ratio of raw materials to be used consid-
ering the transported raw materials and tapping quantity and coke
2. Simultaneous Optimization of Production and outputs.
Distribution for Raw Materials On the other hand, steelworks determine the following schedules
The quantity of iron ore and coal (raw materials) consumed by based on the schedules for ships and blending raw materials deter-
the entire Japanese steel industry per year is several hundreds of mined by the Head Office such that they can carry out daily opera-

Fig. 1 Target of planning, scheduling and logistics related to raw material

- 32 -
 NIPPON STEEL & SUMITOMO METAL TECHNICAL REPORT No. 121 March 2019

Fig. 2 Overview of the hierarchical dividing term and moving horizon algorithm 4)

tions: (5) Storage yard allocation schedules to determine places to

receive raw materials at yards and discharge them, (6) stacker
schedules to determine the operation time of transportation systems
to receive raw materials at the receiving places determined in the
storage yard allocation schedules, (7) reclaimer schedules to deter-
mine the operation time of transportation systems to deliver raw
materials to the discharging places determined in the storage yard
allocation schedules and determine storage tanks such as silos that
store discharged raw materials temporarily, and (8) blending sched-
Fig. 3 Example Gantt chart of a scheduling result
ules of raw materials to adjust the termly and monthly blend sched-
ules determined by the Head Office such that they match actual dai-
ly operations. Nippon Steel & Sumitomo Metal has developed a vided into a time line and hierarchical time sharing where mathe-
system that allows the Head Office and steelworks to manage infor- matical programming is linked to the simulator in each divided peri-
mation on the supply and demand of raw materials in an integrated od is repeated. This has made it possible to obtain solutions that can
way for the targets listed above and has introduced an optimization be used for actual operation for problems that used to be impossible
technology for each target. 3) to solve (Fig. 2).
This chapter describes, as one example, the raw material trans- 2.1.3 Scheduling result
portation and ship allocation optimization system as work at the Figure 3 is a Gantt chart as an example schedule. The schedule
Head Office 4) and raw material yard allocation optimization system shows that the overstays of the ships at unloading places are reduced
as work at steelworks. 5) and the entry to ports becomes smooth (demurrage at the loading
2.1 Raw material transportation and ship allocation optimiza- places is a given condition). In addition, increase in the ratio of un-
tion system loading at a single port by Cape size or larger ships reduces extra
2.1.1 Details of work charges for unloading at multiple ports, which can reduce the total
In this work, schedules for chartering ships and allocating each transportation cost.
ship to loading and unloading places are determined based on in- 2.2 Raw material yard allocation optimization system
formation on the allocation of signed ships (ship types, quantity, 2.2.1 Details of work
and current locations), available ships, purchased quantity under This work determines charge positions (piles) required to stack
contracts, and raw materials to be used at each steelworks while multiple types (several tens) of raw materials transported by ships at
constraints to secure the stock at each steelworks are met. The trans- empty spaces in yards and determines discharge positions to supply
portation fee varies depending on the type of ships to be signed, raw materials in accordance with the usage schedules for blast fur-
contract form, and port call patterns, so optimizing the configuration naces and coke ovens along with their amounts. In these operations,
of ships to be chartered and transportation routes can reduce the 1) stacking raw materials in many small piles creates dead spaces,
costs. which reduces the yard efficiency, so the number of small piles
However, ships’ transportation conditions vary, for example, be- needs to be reduced and 2) securing a large empty space in a yard in
tween China from which the transportation time to Japan is short advance can reduce the time from when ships enter a port to when
and Brazil from which it is long. Therefore, it is difficult to optimize they depart, which can reduce the cost of the ships’ waiting at an-
ship allocation schedules half a year in advance considering the chor offshore, thus enabling efficient yard use.
movements of each ship at an unloading berth in the accuracy of However, the relationship between the length of a yard and ton-
one-hour units while avoiding out-of-stock at each steelworks. 6, 7) nage capacity is not linear and thus it requires complicated calcula-
2.1.2 Configuration of the scheduling function tion. In addition, many factors need to be considered: e.g., signifi-
Considering factors to be combined in long-term schedules and cant changes in the real yard capacity as a result of stacking the
quantity determining factors simultaneously was difficult from the same brand by overlapping and securing of a space between a pile
aspects of calculation scale and time. In the developed system, to and an adjacent pile in a different brand; and constraints on storage
consider such extensive conditions, two layers are provided: One to spaces vary from brand to brand. Therefore, staff used to plan such
determine a long-term broad outline and the other to optimize arriv- schedules based on their experience and intuition, so planning a
al and departure timing and other factors in detail using such deter- long-term schedule in a short period was impossible, which posed
mined outline as fixed information. In addition, we have developed problems, for example, planning schedules looking ahead to future
an original technology in which the total period for each layer is di- delivery of goods was impossible.

- 33 -
NIPPON STEEL & SUMITOMO METAL TECHNICAL REPORT No. 121 March 2019

Fig. 4 Human cooperative optimization algorithm

2.2.2 Configuration of the scheduling function

When optimization technologies were applied to production and
logistics schedules, planning was regarded as a mathematical pro-
gramming problem and Nippon Steel & Sumitomo Metal focused
on mathematical optimization technologies and technologies to use
rule-based techniques, etc. for optimization, which allowed Nippon
Steel & Sumitomo Metal to significantly advance such technologies.
Meanwhile, it was revealed that it needed to address cases that were
difficult to model into formulas and thereby that required high-level
judgment by humans, and cases where know-how that had not Fig. 5 Result example of ship scheduling
turned to explicit knowledge was secured by experienced workers
and to address gradually changing operation conditions to allow dressing changes in operation and requirements is essential for the
systems to be continuously used in actual operation. 8) system to be used for a long period of time. Therefore, we have de-
Therefore, this system was aimed at fully utilizing human veloped a new scheme that cooperates with humans to support plan-
knowledge and addressing operation flexibly by creating a scheme ning by providing functions for planning initial schedules, changing
that cooperates with humans. Specifically, this system has enabled 1) such schedules interactively, and storing revision history. This tech-
planning an initial schedule that satisfies constraints quickly by a nology has been contributing to reduction of the costs of ships’ wait-
simulator that simulates actual operations finely and the greedy al- ing at anchor offshore.
gorithm, 2) planning a schedule that satisfies constraints by allowing
humans to revise schedules freely and interactively and by using 3. Development of a Steelmaking Scheduling Tech-
such details to operate the simulator, 3) storing details that humans nology
planned or revised one by one and allowing humans to revise each In the steelmaking process, the components of molten iron sup-
detail freely, and 4) allowing humans to revise parameters (e.g., plied from a blast furnace are adjusted in a converter and secondary
constraints and performance) freely and storing changed details one refining equipment manufactures slabs (half-finished goods) by a
by one to allow humans to revise each detail freely (Fig. 4). continuous-casting machine. Molten steel for which the components
2.2.3 Scheduling result were adjusted in a converter is poured into a ladle (transportation
Figure 5 is a ship schedule planned using the developed raw container) and then transported to a continuous-casting machine
material yard allocation optimization system. In the schedule, the through the secondary refining process. Molten steel in a single ladle
ships can leave the ports earlier than the target arrival and departure is called a charge. Multiple charges to be continuously casted in a
time given by the Head Office, which can reduce the costs of the continuous-casting machine are called a cast. One characteristic in
ships’ waiting at anchor offshore. the steelmaking process is that the production target is a high-tem-
2.3 Summary perature liquid called molten steel. The temperature of molten steel
This chapter described our work aiming at simultaneous optimi- gradually decreases as time passes. When the waiting time between
zation of production and logistics for raw materials. As explained manufacturing processes is too long, the molten steel in the ladle so-
above, we have developed an optimization technology that can be lidifies, which hinders the productivity significantly. Therefore, it is
applied to large-scale planning problems by combining hierarchical impossible to manufacture extra molten steel in advance and store it
division consisting of a long-range decision layer and detailed oper- in a buffer. In the converter and secondary refining process, molten
ation decision layer with time sharing. This technology was used to steel for which the temperature was appropriately adjusted needs to
develop the raw material transportation and ship allocation optimi- be supplied to the continuous-casting machine according to the cast-
zation system that could plan schedules that were accurate enough ing timing.
to be used for actual operations and the results were favorable. In In addition recently, the processing flow in the refining process
addition, it was found through the development that flexibly ad- has become complicated to satisfy the needs for high-quality prod-

- 34 -
 NIPPON STEEL & SUMITOMO METAL TECHNICAL REPORT No. 121 March 2019

ucts. Several methods have been developed as molten iron pretreat- above. 11) The new algorithm has made it possible to model opera-
ment methods using converters; such as the LD converter-optimized tional constraints flexibly and plan schedules within a practical time
refining process (LD-ORP) method that achieves desiliconization (approximately one minute for a schedule for three days) by narrow-
and dephosphorization at the same time and a multi-refining con- ing down the range for searching for a solution during scheduling
verter (MURC) method that enables a single converter to continu- only to feasible solutions (Fig. 6).
ously process dephosphorization and decarbonization interrupted by Furthermore, we have developed an optimization technique in
the discharge of dregs in between. Such methods have been applied which the temperature of molten steel is considered to control its
to actual equipment. 9) The processing flow and time in the second- temperature appropriately during casting. As molten steel tempera-
ary refining process also differ for each steel type. Therefore, it is ture control, the target temperature in each manufacturing process
becoming difficult to plan a schedule in which various indexes, such used to be calculated separately with the planned operation schedule
as high productivity, minimum cost, and strict adherence to delivery as preconditions. However, the schedule needed to be readjusted be-
dates, are satisfied in the entire steelmaking process while the peak cause the temperature of molten steel decreased as time passed and
of the processing load in each manufacturing process is reduced. the processing time had to be changed to secure the time required to
3.1 Problem with steelmaking scheduling adjust the temperature in the schedule. We have developed a simul-
One problem with steelmaking scheduling is to determine opera- taneous optimization model in which variation in the temperature of
tion schedules of converters, secondary refining, and continuous- molten steel during transportation and processing is modelled and
casting machines such that the objective functions become optimum such variation is linked to a schedule (Fig. 7). 12) This technique al-
while constraints on manufacturing and logistics at steelworks are lows the temperature of molten steel when it arrives at a continuous-
met with charge configuration of casts for the continuous-casting casting machine to match the target temperature and an operation
machines and casting sequence given. As constraints, there is a con- schedule to be calculated such that the temperature of the molten
straint on the processing flow according to which each charge steel in each manufacturing process is within the upper and lower
moves on through the predetermined manufacturing processes suc- limits. In addition, we have been studying a robust scheduling tech-
cessively and another one on interference that prohibits the process-
ing time of one charge from overlapping with that of another one in
a manufacturing process.
In addition to those constraints, there is also a constraint regard-
ing the consecutive continuous casting of casts to allow multiple
charges to be casted without interruption in the same cast. As repre-
sentative performance indexes included in objective functions, resi-
dence time and casting completion time are used. The residence
time refers to the time from when molten steel is tapped from a con-
verter to when casting begins. The objective is to reduce the de-
crease in the temperature of molten steel to the extent possible to
minimize the cost generated as a result of the temperature increase
of molten steel. The casting completion time refers to an index indi-
cating the productivity that is used to maximize the operation rate of
a continuous-casting machine. We have developed algorithms for
which mathematical optimization techniques are applied to solve the
afore-mentioned problem with steelmaking scheduling.
3.2 Example of optimization of the problem with steelmaking
scheduling Fig. 6 Gantt chart of scheduling result 11)
Planning steelmaking schedules manually took time and labor
and the optimality and feasibility of operations of obtained sched-
ules were not always satisfactory, so we have developed an algo-
rithm based on a mathematical optimization technique. 10) All opera-
tional constraints were considered to be difficult as the calculation
would take time. Therefore, the main operational constraints and
performance indexes only were considered in the mathematical op-
timization technique and other specific constraints were described in
the logistics simulator. Combining the two has enabled planning of
a schedule in a few minutes that is optimum to a level equal to or
higher than those planned by experienced workers. In addition, be-
cause the converter operation procedures were changed and new
equipment was added, it needed to handle various factors: Increase
in the number of combinations in the sequence of tapping of molten
steel from a converter; the performance of cranes and other trans-
portation systems becoming an obstacle; and increase in the number
of charges to be included in schedules due to increase in the output.
Therefore, we have developed another new algorithm in which con-
straint logic programming is combined with the technique described Fig. 7 Model for molten steel temperature and schedule 12)

- 35 -
NIPPON STEEL & SUMITOMO METAL TECHNICAL REPORT No. 121 March 2019

nology that includes the appropriate temperature control time in the chapter introduces solutions using the explicit solution technique.
secondary refining process to reduce variation in the temperature of 4.2 Technology for optimizing charging and rolling sequences
molten steel caused by changes in operational conditions in order to (extraction sequence) simaltaneously 17)
make casting stable. 13) To maintain the temperature when slabs are charged high into a
3.3 Summary hot-rolling heating furnace, hot-rolling yards usually have equip-
Almost all molten iron is processed at steel mills following the ment for keeping slabs warm recently. Under such a circumstance,
blast furnace process and it is turned into sheets, plates, and other consideration on the charging schedule side is also required to avoid
various end products, so it is an important process from all aspects the rearrangement of piles at the time of charging into a heating fur-
of quality, cost, and delivery dates. To continue manufacturing high- nace. This demands consideration of the charging sequence in a hot-
quality products while keeping the productivity high by maximizing rolling schedule for which the original requirement is to make the
the operation rate of equipment owned, planning more accurate rolling sequence (sequence of extraction from a heating furnace) ap-
steelmaking schedules is required. To that end, processing and propriate.
transportation conditions in steel mills need to be more specific and 4.2.1 Relation between the rolling sequence (extraction sequence)
integrated optimization for the entire steelworks even considering and charging sequence
production and logistics schedules for upstream and downstream When the ratio of the numbers of slabs to be extracted from mul-
processes is important. To achieve this, development of modelling tiple furnaces is different, the time during which slabs stay in the
technology for large-scale optimization problems and technologies high-ratio furnace is shorter than that in the low-ratio furnace.
for speeding up calculation is expected through the development of Therefore, the charging sequence is relatively late comparting to the
algorithms and parallelization. extraction sequence. A charging event occurs when an empty space
is formed on the charging side after an extraction event. Therefore,
4. Development of a Hot-rolling Scheduling Tech- based on the relationship between the width of a slab extracted and
nology the width of a slab to be charged, there are three charging cases after
4.1 Problem with hot-rolling scheduling a single slab was extracted: (1) No slab can be charged (width of the
In the hot-rolling process (hot strip mill), usually slabs reheated extracted slab < width of the slab to be charged), (2) a single slab
in multiple heating furnaces (three or four units) are extracted one can be charged (width of the extracted slab ≥ width of the slab to be
by one in accordance with the predetermined sequence of extraction charged), and (3) two slabs can be charged (width of the extracted
furnaces and rolled by a roller. Therefore, scheduling needs to deter- slab ≥ total width of the two slabs to be charged). The charging se-
mine distribution to heating furnaces and the charging sequence for quence differs depending on which case occurs. That is to say, the
each furnace such that heating constraints are met and it needs to charging sequence depends on the rolling sequence (extraction se-
determine the rolling sequence appropriately such that rolling con- quence) and it can be uniquely determined, but its formularization is
straints are met. difficult and charging and extraction simulations are required for
As a heating constraint, there is a constraint on the heat change such determination.
that the charging temperature, target extraction temperature, and 4.2.2 Simultaneous optimization of charging and rolling sequences
heating characteristics should be at the same level between neigh- When considering re-stacking at yards, the schedule should be
boring slabs (in three to five meters) in the same furnace because the determined such that slabs can be charged from the top of piles for
furnace temperature is equally controlled for them due to the fur- both rolling and charging sequences to the extent possible. If the
nace’s thermal inertia. As typical constraints on rolling, there is a functional relation between the rolling and charging sequences can
coffin constraint that in a single rolling chance (one schedule), wider be formulated, simultaneous optimization of the rolling and charg-
slabs should be processed first because controlling such shape is ing sequences is possible using such relation as a constraint. How-
easier and then the width should be gradually reduced; and another ever, as explained above, the functional relation between the rolling
constraint on the changes in the width and thickness that demands and charging sequences can be identified only by charging and ex-
that differences in the thickness and width of the coils of two slabs traction simulations. Therefore, we have determined to solve such
to be rolled successively should be small and thereby the change problem by convergent calculation where the relation between the
should be smooth. rolling and charging sequences is calculated through simulations;
Meanwhile, for the problem with hot-rolling scheduling, several such relation is used as a constraint in the calculation; and these pro-
performance indexes are used to minimize the changes in tempera- cedures are repeated until the relation between the rolling and charg-
ture, width, and thickness to the extent possible, reduce the quantity ing sequences for the obtained solution matches the hypothesis. At
of fuel consumed by heating furnaces, and enhance the productivity this time, the determination of the rolling and charging sequences
by shortening the total extraction time from the first extraction to the was formulated as a 0/1 integer programming problem where slabs
last in a schedule although they overlap with the constraints. In ad- are assigned to the charging and rolling sequences as a double as-
dition, as described later, regarding the relation with yards, the signment problem.
charge sequence needs to be adjusted in some cases such that the 4.2.3 Procedures for formulating a double assignment problem of
number of re-staking becomes the smallest possible with stacking charging and extraction sequences
conditions in yards as preconditions. Decision variables:
Regarding methods to solve the problem with hot-rolling sched- xc[i][jc] (charging sequence assignment variable):
uling using mathematical programming, some researchers have 0/1 variable where when slab i is charged in charging sequence
studied heuristics methods such as methods using genetic algorithms jc, it is 1, and when not, it is 0
(GA) 14, 15) and another method in which assignment problem formu- xe[i][je] (rolling sequence assignment variable):
larization is combined with a local search 16) because the scale of the 0/1 variable where when slab i is rolled in rolling sequence je,
problems is large and too many constraints need to be met. This it is 1, and when not, it is 0

- 36 -
 NIPPON STEEL & SUMITOMO METAL TECHNICAL REPORT No. 121 March 2019
Table 1 Effect of optimization orders received, products with specifications demanded by custom-
Load of pile (ave.) ers should be delivered by designated delivery dates. Staff in charge
of production control always monitor the progress of the production
Manual 112
of each order so that the delivery dates can be met. The most impor-
Optimize 72 (−40)
tant thing in production control is to determine an appropriate date
on which the production of each order begins. The reason is that if
the date is too late, the delivery date cannot be met and if it is too
early, the yards and warehouses become full, which stops the manu-
facturing lines. However, production periods for steel products vary
significantly; the production periods of product types that pass
through many refining lines vary even more, in particular, so man-
aging their production is difficult.
A standard production period refers to the standard value of pro-
duction periods used to calculate the production start date. Staff in
charge of production control calculate a standard production period
based on the specifications of an order. They plan a rolling schedule
and cast formation such that the rolling can be performed on the
Fig. 8 Width change with rolling order 17) date and time that are calculated by deducting the standard produc-
tion period from the delivery date. When there is no pile of work-
Constraint formulas: pieces in each manufacturing process and manufacturing processes
Charging sequence constraints: through which products pass are known, the production period can
Constraint on the pile sequence at a yard and constraint on the be calculated by adding the processing time in each manufacturing
relation between the rolling and charging sequences process and transportation time. 22) However, in plate manufacturing
Rolling sequence constraints: processes (as is the case with steel sheets and pipes), there is a large
Constraint on the changes in the width, thickness, and temper- pile of workpieces before each manufacturing process and there are
ature, constraint on the input position, constraint on charging some stochastic processes (e.g., conditioning and leveler process)
regulation for each furnace, etc. for which whether products should pass through the processes is de-
Objective functions: termined in the middle of manufacturing (whether products pass
Objective function for the charging sequence: through such processes is determined only after the production be-
Minimizing the number of pile rearrangements gins), so estimating production periods accurately is difficult.
Objective function for the rolling sequence: This chapter describes the procedures for calculating standard
Minimizing the amounts of the changes in the width, thick- production periods of plates for which estimating production peri-
ness, and temperature, minimizing fuel for a heating furnace, ods is difficult. The next section introduces the plate production
etc. flow and conventional procedures for calculating standard produc-
4.2.4 Results of the application of the technology for optimizing tion periods first. Then the newly developed procedures for calcu-
charging and rolling sequences (extraction sequence) simulta- lating standard production periods using decision trees and the
neously maximum likelihood estimation are described. Lastly, effects of the
It has been found that simultaneous optimization of the charging application of such procedures to actual equipment are shown.
and rolling sequences can reduce the total number of pile rearrange- 5.1 Plate manufacturing process and conventional procedures
ments per schedule by 30% or more as shown in Table 1 and the for calculating standard production periods
changes in the width and thickness do not break as a result. Figure Figure 9 illustrates the plate production flow. Slabs heated in re-
8 shows example width changes. heating furnaces are rolled into a designated size at the roughing and
4.3 Future prospect finishing mills. Then they are water-cooled by an accelerated cool-
This chapter described a problem with scheduling when the ex- ing device such that designated crystal structure is obtained and they
traction ratio is fixed (the extraction furnace sequence is fixed). are cooled at room temperature at a cooling bed. The processes fol-
However, it is ideal to handle the extraction ratio as a variable that lowing the cooling bed are referred to as refining processes: They
relies on slabs to be charged into each furnace, so such solution is are divided into normal processes (e.g., heat treatment and coating)
expected. for which whether products pass through the processes is deter-
In addition, schedules for charging slabs into heating furnaces mined based on the production specifications of an order; and sto-
and extracting them are closely related to the combustion control in chastic processes (e.g., conditioning and leveler process) for which
the heating furnaces. Some researchers have been working on si- whether products pass through the process is determined based on
multaneous optimization of schedules and combustion control. 18–20) the quality in the middle of the production. In Fig. 9, the processes
Such technologies may be realized soon thanks to the future ad- enclosed with the solid lines are normal processes and those en-
vancement in computer technologies. Some researchers have also closed with the dashed lines are stochastic processes. The ultrasonic
reported expansion to integrated scheduling of steelmaking and roll- test equipment (UST) is a process having both characteristics. A
ing 21) and practical use of such technology in the future is expected. production period in this paper refers to the number of days from
rolling to when preparation for shipment (or certificate test) has be-
5. Development of Technology for Designing Stand­ come complete, that is to say, the period during which products re-
ard Plate Production Periods main in a plate mill.
In the steel industry where products are manufactured based on A conventional standard production period used to be managed

- 37 -
NIPPON STEEL & SUMITOMO METAL TECHNICAL REPORT No. 121 March 2019

Fig. 9 Manufacturing process of steel plate (solid line: normal processes, dash line: stochastic processes) 23, 27)

Fig. 10 Tables of old standard production period 27)

Fig. 11 Margin of old standard production period 27)

using tables as shown in Fig. 10. The period is broadly divided into
the base period, processing period, grade margin, inspection margin,
and customer margin: They are further divided into small categories.
The value in each table refers to the days based on the quality speci-
fications of the order and such values are added to obtain the stand­
ard production period. However, only normal processes for which
Fig. 12 Calculation algorithm of new standard production period 23)
whether products pass through can be determined based on the order
specifications are included in the processing period and stochastic
process periods are included in the grade margin and other margins. 5.2 New procedures for calculating standard production periods
Therefore, the margins that are differences between the standard Analyzing actual production periods, the production periods
production periods and actual periods spread horizontally in a histo- highly correlated with processes through which products pass (here-
gram as shown in Fig. 11. The proper production completion rate inafter, “transit processes”). Therefore, we devised the procedures
defined by the percentage of production completion within the for calculating standard production periods after estimating transit
stand­ard production period (ratio of “actual period ≤ standard pro- processes. Figure 12 shows a new algorithm for calculating stand­
duction period”) remained at 91.5% (the periods have been normal- ard production periods. 23) Details of the processing in each step in
ized with reference to a certain value). the algorithm are explained below.
A. Determination of a production class to be manufactured

- 38 -
 NIPPON STEEL & SUMITOMO METAL TECHNICAL REPORT No. 121 March 2019
N
If the transit processes and production class of orders are both J= ∑  ln p ( t | f ) ~
  n     n 
the same, it is considered that the probability distribution of the pro- n =1
(5)
duction periods must be the same. Decision trees 24) are used to esti-
[ (t  − μ )
]
N ~ 2

mate the transit processes through which the product would pass
1  
 = − —
2
∑ ln (2π) + ln (v ) + —
n =1
v
~
n 
n
~
n
n 
→ Max
based on the production specifications of the order as shown in Fig.
13 and 0 or 1 that indicated whether the product passes through each N in Formula (5) is the number of plates in learning data. It is
process is arranged in a line to form a process flow. The order class desirable to use the number of plates for half a year to one year dur-
was added before the process flow as shown in Formula (1) to deter- ing which all production classes are manufactured for the time. The
mine a code. This code is called production class ci (i is the index of value of N becomes several hundreds of thousands. Therefore, when
the production class). 25) Formula (5) is used as it is, the scale of the optimization problem
def
Production class (ci ) = order class _ process flow (1) becomes very large. However, fortunately, Formula (5) can be
B. Calculation of the probability distribution of the transit processes grouped by the process flow type and the scale of the problem is re-
Plates in the same production class ci are extracted from past op- duced to the number of the types (L) (several hundreds), which
eration data, and the occurrence rates (empirical distribution) of ac- makes the calculation easy. 27)
tual process flow fj = (f1, f2, …, fM) are calculated as shown in For- D. Estimation of production periods for the production class
mula (2). This is an occurrence probability model of production The occurrence probability model (P(fj|ci)) of the process flow
class ci. for the production class calculated in step B is combined with the
Number of ci & fj plates probability density function (p(t|fj)) of the production periods for the
P (  fj | ci ) = — (2)
Number of ci plates process flow calculated in step C as shown in Formula (6) to calcu-
C. Calculation of the production periods for each process flow late the production period (p(t|ci)) of the production class.
L L
On the assumption that when an order is manufactured in certain
process flow fj, the probability distribution of the production periods
p ( t | ci ) = ∑  p ( t | f , c )  P ( f | c ) ≈ ∑  p ( t | f )  P ( f | c ) (6)
j =1
       j  i     j    i 
j =1
       j     j    i 

can be calculated by the summation of the processing periods of the E. Calculation of the standard production period
manufacturing processes through which the order has passed (in- When the cumulative distribution function of the probability
cluding waiting time to be processed). Specifically, if the processing density function (p(t|ci) in Formula (6) is determined as F(t|ci), stan-
periods are assumed to be a normal distribution, the probability den- dard production period tˆ95i can be calculated using Formula (7).
sity function of the production period (t) of actual process flow fj is Where, 0.95 is the designed value for the proper production comple-
expressed as Formulas (3) and (4). The average (μm) and variance tion rate.
(vm) are calculated such that the likelihood function of actual period tˆ95i = F −1 (0.95 | ci ) (7)
tn shown as Formula (5) becomes the maximum (maximum likeli-
~
hood estimation 26)). Where, f n is the actual process flow of the nth 5.3 Effects of new standard production periods
~ ~
plate, and μn and vn are the average and variance of the production Standard production periods were designed using actual data on
period, respectively. plates manufactured at a steelworks during a certain period. The

{ }
(t − μ~j )2 margins of the standard production periods in another period are
1
p ( t |  fj ) = N ( t |  μ~j , v~j ) = —  exp  − — (3) shown as a histogram in Fig. 14. Compared to Fig. 11, the new
~
√ 2 π vj 2v~j  
stand­ard production periods were steeply distributed with a higher
μ~j =  fj  μT, v~j =  fj  vT peak. The average of the new standard production periods was al-
(4) most the same as that of the conventional production periods, but
μ = (  μ1, μ2, …, μM ) ≥ 0, v = ( v1, v2, …, vM ) ≥ 0 the proper production completion rate can be improved by 3.2%
(91.5 → 94.7%).
Next, Fig. 15 shows changes in actual proper production com-
pletion rates at three steelworks before and after switching from
conventional standard production periods to new standard produc-
tion periods. These actual proper production completion rates are
not the margin of “standard production period ≥ actual period.”

Fig. 13 Decision trees to decide production class 23, 27) Fig. 14 Margin of new standard production period 27)

- 39 -
NIPPON STEEL & SUMITOMO METAL TECHNICAL REPORT No. 121 March 2019

6.1 Output schedule optimization problem

Orders to steelworks are grouped by steel type and processing
route pattern. The output weight for each grouped order is deter-
mined using mixed integer programming under various constraints
(e.g., production capacity) such that the evaluation value of Formula
(8) for the increase in the output lot, adherence to output deadlines,
and levelling of processing loads becomes optimal. An estimation
model that was obtained by statistical analysis of actual data was
used to estimate processing loads.
Fig. 15 Changes in proper production completion rate 30) Constants:
x̂ j, t: Order weight for grouped order j and output deadline t
They are the rates of production completion by shipping deadlines. ŷ k, t: Processing capacity for manufacturing processes k and date t
That is to say, they are management indexes used in actual produc- Decision variable:
tion control work including the differences between actual rolling xj, t: Output weight for grouped order j and planned output date t
dates and appropriate rolling dates. When the conventional standard Intermediate variables:
production periods were used, the actual proper production comple- δi, t: 0/1 variable. When steel type i is output on planned output date
tion rates largely varied from steelworks to steelworks and dropped t, it is 1. When not, it is 0.
lower than 80% in some cases. After the use of the new standard yk, t =  ∑jfk(∙)xj, t: Weight of the processing load for manufacturing
production periods began, the actual proper production completion process k and planned output date t
rates remained at a high level. The new standard production periods Note: fk(∙) is a function that estimates the occurrence rate of the
designed using this algorithm were shorter than conventional ones processing load in manufacturing process k from the produc-
by one to three days on average, so this algorithm has made it possi- tion specifications.
ble to shorten production periods and improved the actual proper Performance indexes:
production completion rate at the same time. In addition, the new Minimization J = W1 J1 + W2 J2 + W3 J3 (8)
standard production periods designed using the proposed technique J1 = ∑ i ∑ t δi, t: Number of seams of different steel types
have been used for some scheduling systems at steelworks, 28, 29) con-  p  p
J2 = ∑ j ∑ p | ∑ t = 0 x̂ j, t − ∑ t = 0 xj, t |: Delivery schedule delay
tributing to improving the accuracy of production control at steel-
works. J3 = ∑ k ∑ t | ŷ k, t − yk, t |: Amount exceeding the processing capacity
5.4 Summary Where, W1, W2, and W3 are weighting factors.
This chapter described the procedures for designing the standard Constraint formulas:
production periods for plates. In the proposed technique, decision Steelmaking constraint, processing and storage space constraint,
trees are used to estimate transit processes because production peri- etc.
ods relate to the transit processes; then, the maximum likelihood es- 6.2 Optimization algorithm
timation is used to estimate the probability density function of the A mathematical optimization technique is used to calculate the
production periods; and the production period that satisfies the des- output weight (output schedule) for each production class such that
ignated proper production completion rate is determined as the performance indexes (e.g., increase in the productivity, levelling of
stand­ard production period. The new technique for designing stand­ processing loads, and adherence to delivery dates) become optimal
ard production periods is used in the order entry system (OES) that under various constraints (e.g., production capacity). As the optimi-
issues production instructions for received orders, 30) contributing to zation technique, basically mixed integer programming is used.
maintaining the high production completion rates at each steelworks However, some cast formation constraints (e.g., constraints on
and reducing the stock. charge locations in casts) to be considered when output schedules
are planned are complicated and thereby modelling is difficult by
6. Development of a Technology for Optimizing mixed integer programming alone. Therefore, we have developed a
Weekly Plate Output Schedules 28, 29) multistep solution in which a tentative output schedule for which
Appropriate weekly plate output schedules need to be quickly only linear constraints are taken into account is planned by mixed
planned in consideration of various performance indexes, such as integer programming; the charge quotas are formulated for the ob-
improvement in the productivity (maximization of the sequential tained output schedule for each continuous-casting machine; a
number of continuous casting), levelling of processing loads, and search technique is used to rearrange the charge quotas in consider-
adherence to delivery dates, while operational constraints (produc- ation of all constraints including those indicated in the non-linear or
tion capacity and storage capacity) are met. The parameters for If-then rule; and linear programming is used to assign a production
planning schedules, such as processing loads and production peri- class to each of all the obtained casts and charge quotas (Fig. 16).
ods, vary depending on the order specifications and operation 6.3 Example schedule
change. In addition, it is difficult to solve large-scale and complicat- Table 2 compares a schedule planned using this technique and
ed optimization problems within a practical time if each order is di- another schedule manually planned for actual manufacturing data
rectly modelled. Therefore, we have developed an optimization al- for eight days. The table shows that this new technique can increase
gorithm that can consider production class models that can be used the output lot size and reduce variation in delivery dates while the
throughout processes from parameter estimation to schedule plan- constraints are met. Where, the seam length refers to the perfor-
ning by learning of actual results and that can consider complicated mance index value for the length of waste of seams and the variation
constraints and performance indexes. of delivery dates refers to the standard deviation of differences be-
tween the output deadlines and planned output dates.

- 40 -
 NIPPON STEEL & SUMITOMO METAL TECHNICAL REPORT No. 121 March 2019

Fig. 17 Example of colored Petri Net 31)

put arc expression is a rule required for firing for a single token. Fir-
ing conditions in which multiple tokens’ attributes are combined can
be described in a dialog shown by double-clicking a transition.
In this simulator, a Petri Net can be input using a mouse and
keyboard, and a simulation can be executed immediately to see the
Fig. 16 Optimization algorithm 29) network behavior. At the beginning of a simulation, compilation is
performed in the background and the behavior of the tokens can be
seen immediately on the graphical user interface (GUI), so debug-
Table 2 Effect of optimization
ging the simulator is inherently easy.
Proposal Manual In this colored Petri Net simulator, tokens’ attributes can be de-
Number of the violation of the restrictions [-] 0 0 fined. List 1 shows examples of attribute definition. In List 1, a
Amount of the waste of the steel [-] 16 26 Product token and Automated guided vehicle (AGV) token are de-
Standard deviation of the due date [day] 14.17 15.04 fined: The weight (product weight) and deadline (delivery time) are
defined as attributes of the Product token, and the capacity (maxi-
mum authorized payload), battery (battery capacity), and product
6.4 Summary (product loaded) are defined as attributes of the AGV token.
This chapter described the optimization algorithm that supports
planning output schedules that balance the production in larger out- List 1 Examples of token definition
put lot size, levelling of loads in the processing process, and reduc- Product(double weight, double deadline);
tion of variation of delivery dates in the plate production. Trade-off AGV(double capacity, double battery, Product product);
adjustment between the production in larger output lot size and re-
duction of variation of delivery dates is a task common in steel pro- The locations of tokens at the beginning of a simulation (initial
duction. Application of the algorithm to product types other than markings) are described in a dialog shown by double-clicking a
plates is expected. place as shown in List 2. Setting specific value(s) as attribute(s) cre-
ates a token instance.
7. Development of General-purpose Colored Petri
Net Simulator List 2 Examples of initial marking
Many heavy objects are transported at steelworks, so reducing Product(20, 800);
the logistics costs is an important task. Therefore, logistics control AGV(50, 80, Product(0,0));
that allocates transportation equipment to objects to be transported
properly is important to transport more objects with fewer numbers An input arc expression that is a token condition that enables a
of transportation equipment to deliver products to destinations as transition to fire can be given to the input arc of the transition. List 3
scheduled. Nippon Steel & Sumitomo Metal has developed a real- shows such examples. The upper row in List 3 means that when a
time advanced simulator tool for colored Petri Net (TrasCPN) as a Product token is present in the place, the transition can fire. The
logistics simulation tool that can handle entire processes from analy- term “product” is the name of the token instance variable and is re-
sis of logistics to control in order to improve the efficiency of logis- ferred to in the transition’s firing conditions and output arc expres-
tics control. The simulator can cover complicated logistics condi- sions to be described later. Meanwhile, the lower row in List 3
tions, perform fast calculations, and can make simulations linked to means that when an AGV token for which the battery capacity is 20
the optimization algorithm. Nippon Steel & Sumitomo Metal has % or more is present in the place, the transition can fire.
been using the simulator to solve various problems with logistics at
steelworks. 31) This chapter describes the functions of TrasCPN. List 3 Examples of input arc expression
7.1 Formats in the colored Petri Net simulator Product product
Figure 17 illustrates an example of colored Petri Net created us- AGV agv(agv.battery>20)
ing this simulator. The thin rectangle in Fig. 17 indicates a transi-
tion, the square containing a circle indicates a place, and the small As an input arc expression, conditions in which multiple tokens’
circle in a circle indicates a token. A firing condition can be given to attributes are combined cannot be described. To define this firing
the arc connecting a place to a transition. Token conditions required condition, a transition setting dialog is used. List 4 shows an exam-
for firing can be given to the input arc to a transition (input arc ex- ple of firing conditions. As shown in List 4, the C++ language is
pression) and conditions of a token generated after firing can be giv- used to describe firing conditions. Once the processing completes,
en to the output arc of a transition (output arc expression). This in- firing is enabled and when zero is returned in the middle of process-

- 41 -
NIPPON STEEL & SUMITOMO METAL TECHNICAL REPORT No. 121 March 2019

ing, firing is disabled. In List 4, when the maximum authorized pay-

load of the AGV is smaller than the product weight, firing is dis-
abled. Together with multiple token combinations that enable firing
are provided, the keyword “fire_priority” can be used to designate
the priority of firing. In List 4, the priority of a token with closer de-
livery time is set to higher. In addition, the keyword “fire_delay”
can be used to specify a delay time in which a token created after a
transition fired becomes ready to fire next.

List 4 Example of transition firing conditions 1

if (agv.capacity < product.weight) return 0;

fire_priority = 1.0/product.deadline;
fire_delay = 60;
Fig. 18 Example of hierarchical colored Petri Net
In addition, a new variable can be created in a firing condition as
shown in List 5 and it can be used in an output arc expression to be
described later. Complicated logistics events can be easily described, The lower-layer Petri Net of the module created by the reference
for example, the calculation result of a designated function can be copy is the same as the original before the simulation begins. At the
set as the attribute of a transition that is created after another transi- beginning of the simulation, the lower-layer module is copied and it
tion fired. becomes an instance different from the original, turning into another
lower-layer Petri Net for which the locations of the tokens are dif-
List 5 Example of transition firing condition 2
ferent from those in the original.
int N = agv.product.weight<50? 1: 0; In addition, regarding places, complete copy and reference copy
are available. Complete copy is to create instances different from
List 6 shows example output arc expressions that are the condi- those in the original. Reference copy means creating and displaying
tion of a token to be created after a transition has fired. In the upper a place the same as that in the original in a different location. This
row in List 6, the AGV and product token instance variables in the function eliminates the use of a long arc to connect a place to a far
input arc expressions in List 3 are used; the maximum authorized transition, allowing complicated Petri Nets with good readability to
payload is the same and the battery capacity is reduced by 10% to be created.
create an AGV with product. Meanwhile, in the lower row in List 6, 7.3 Summary
variable N described in List 5 is used to create no (0) or one AGV This chapter described the colored Petri Net tool developed to
token. Such case is useful to select an output place depending on the solve the problem with logistics control. In this simulator, multiple
token attribute. In addition, the term “@60” means the delay time tokens with different attributes can be defined to make it possible to
from when a token is created to when the next firing is enabled as is simulate complicated logistics flow freely and such attributes can be
the case with the keyword “fire_delay.” used to describe complicated firing logic. In addition, the hierarchi-
cal structure makes it possible to create Petri Nets with good inher-
List 6 Examples of output arc expressions ent reusability even for large-scale logistics.
AGV(agv.capacity, agv.battery-10, product) This colored Petri Net tool has been used to solve many prob-
N*AGV(agv)@60 lems with logistics control in the steelmaking process, such as AGV
allocation control, 33) molten iron control system, 34) molten steel out-
7.2 Colored Petri Net GUI put schedule planning support system, 10, 11) and yard control sched-
Providing easy-to-use GUI is important when creating a large- uler. 35)
scale complicated simulator. In this simulator, Petri Nets can be eas-
ily created with a mouse and keyboard. Part of a Petri Net can be di- 8. Development of a Technology for Optimizing Lo-
vided into some modules to improve the readability and reusability gistics at Slab Yards
of complicated Petri Nets as is the case with other tools 32), which al- 8.1 Problems with slab yard control optimization
lows hierarchical Petri Nets to be created. In this tool, two methods When steel materials (slabs) are supplied from the steelmaking
are provided to reuse modules and users can select whichever they process to the rolling process, slabs are once placed in temporary
require. Figure 18 is an example of a simple hierarchical Petri Net. storage spaces called slab yards (hereinafter “yards”) and then trans-
The double-lined blocks are modules having lower-layer Petri Nets. ported in accordance with the processing time at heating furnaces.
Two methods are available to copy and reuse modules: Com- Yards are intermediate storage spaces working as buffers between
plete copy and reference copy. In complete copy, the original mod- different manufacturing processes. That is to say, they work to ad-
ule and copied module become different objects and all the lower- just the manufacturing sequences between different manufacturing
layer Petri Nets are copied, so a change to a lower-layer Petri Net in processes of steelmaking and rolling and to supply materials to the
one side does not affect the other one. On the other hand, in refer- rolling process without hindrance. Lately, with the increasing de-
ence copy, lower-layer Petri Nets are not copied and only the infor- mands for CO2 emissions reduction, needs for control optimization
mation relationship with the original is copied. Therefore, when one at yards are increasing because preventing decrease in the tempera-
side is modified, the other side is also changed. The middle module ture of slabs at yards is expected to reduce fuel consumed by heating
in Fig. 18 is a module created by copying the relationship with the furnaces. In yard control, items ① to ④ in Fig. 19 need to be deter-
top module and the bottom module is one created by complete copy. mined.

- 42 -
 NIPPON STEEL & SUMITOMO METAL TECHNICAL REPORT No. 121 March 2019

S1 S2 … Sj … Sm = N, Sj1∩Sj2 = φ ∀j1, j2 (9)

∩ ∩ ∩ ∩ ∩
φ in Formula (9) is an empty set. In SPP, 0/1 decision variable
x[j] is provided for arbitrary subsets Sj. The variable indicates
whether each subset Sj is adopted as an optimum subset for dividing
universal set N. It can be formulated with Formulas (10) to (12) list-
ed below as a 0/1 planning problem below. At this time, subset
group M for universal set N needs to be listed in advance. For the
slab stacking problem, feasible pile group M that meets the pile
form constraint needs to be predetermined.
Fig. 19 Slab yard control decision flow SPP: Min. ∑j∈M cj ∙ x [  j] (10)
M: Feasible pile set
Items ① to ④ are related, so they cannot be determined sepa- Subject to ∑j∈S(i) x [  j] = 1 (i∈N) (11)
rately. However, determining all the factors at the same time is also S(i): Set of subsets j including slab i
difficult because the scale is large. Therefore, control staff usually x [  j]∈{0, 1} (∀j∈M)
determine them in the following sequence: Target slabs are allocated cj = k1 ∙ 1 + k2 ∙ cj2 = k1 + k2 ∙ cj2 (∀j∈M) (12)
to one of multiple buildings (①); the slabs allocated to each build- cj in Formula (10) is the evaluation value of feasible pile j and it
ing are further divided into piles (unit) because approximately 10 is evaluated based on the number of piles and the number of rear-
slabs are stacked in each building for storage (②); the transportation rangements as shown in Formula (11). The number of piles is one
sequence is determined to stack the slabs efficiently by overhead for all the piles. The number of rearrangements at the time of stack-
cranes and other transportation equipment (③); and the transporta- ing is cj2, the weighting factor for the number of piles is k1, and that
tion tasks are allocated to transportation equipment (④). for the number of rearrangements is k2.
Technologies for using mathematical optimization techniques In SPP, the evaluation of the pile form constraint that is a con-
and simulation technologies for such determination are under devel- straint unique to the target problem and the evaluation of the number
opment. Regarding item ①, some researchers have reported tech- of rearrangements that becomes an objective function are included
nologies for formulating such determination as flow optimization in the processing that lists feasible piles. Therefore, a constraint as
problems and calculating appropriate flow routes. 36, 37) Regarding the 0/1 planning problem is only Formula (11) requesting that the
items ③ and ④, some researchers have reported techniques in target slabs belong to any pile without omission and overlapping. As
which GA and simulation technologies are combined. 38) In addition, described above, constraint and evaluation formulas for which the
determination of piling groups in ② and determination of the trans- formularization is difficult can be handled in the listing processing.
portation sequences for piling in ③ (slab stacking problem) have It can be said that this is as an advantage in applying a set partition-
been studied most intensely. This is because slabs are stored in piles ing problem to an actual problem. Where, how quickly the pile form
at yards to save space and they are rearranged at the time of the re- constraint and the number of temporary placed slabs are evaluated
ception such that slabs to be discharged early are positioned in the in the listing processing are important. (For the details, refer to the
upper part of a pile to allow them to be discharged quickly. Improv- document 44).
ing the efficiency of such rearrangement is a target in work at yards. 8.3 Results of the application of the set partitioning problem
There are various preceding researches on such problem including method to the slab stacking problem
similar problems of reducing the rearrangements of containers at Table 3 compares the results of slab stacking planned by this
ports. 39) method and those of stacking that humans planned. The table shows
This slab stacking problem assumes that when slabs arrive at a that the rates of temporary placed slabs are at the same level, but the
yard, the heating furnace schedule in the next manufacturing process number of piles can be reduced by approximately 25%. This method
has been determined. If such schedule has not been determined at is expected to increase the charge rate to hot pits, which may in-
the time of the arrival, different approaches are required after the crease the charge temperature.
schedule is determined, for example, slab rearrangement (pile rear- 8.4 Summary
rangement problem 40, 41)) or adjustment in the charge schedule. This chapter introduced the method to solve the slab stacking
The slab stacking problem is a combination optimization prob- problem by SPP when the discharge schedule to the following man-
lem to form a pile in which slabs are stacked in the discharge se- ufacturing process has been determined when slabs arrive at a yard.
quence at small workloads to the extent possible under a pile form However, this method consumes a lot of memory and when N ex-
constraint. Some researchers regard this problem as a grouping ceeds 50, the calculation becomes impossible in some cases. In ad-
problem and they have proposed formularization procedures using dition, the conditions of yards change hourly, so real-time resched-
the vertex coloring method 42, 43) and those based on the set partition uling procedures according to such condition changes are required.
method. 35, 44) This chapter introduces the set partition method briefly. To this end, the pile form rearrangement problem described above
8.2 Formularization of the slab stacking problem as a set parti- needs to be addressed.
tioning problem
The slab stacking problem can be regarded as a grouping prob-
lem in which target slab set N is divided into subsets (piles). This Table 3 Effect of optimize (n: 3 115)
problem can be formulated using a set partitioning problem (SPP) 45) Number of Rate of
that is a type of combination optimization problem. SPP is a prob-
Slab per one pile (piles) temporary placed slabs
lem in which the elements in N are divided into subsets S1, S2, …, Sj,
…, S without overlapping and omission so as to minimize the total Manual 6.9 (451) 36%
m
cost on the assumption that any subset Sj has its cost cj. Optimize 9.0 (346) 33%

- 43 -
NIPPON STEEL & SUMITOMO METAL TECHNICAL REPORT No. 121 March 2019

9. Development of Human Cooperative Scheduler

Production planning and scheduling systems assist humans in
making decisions. Normal production schedulers have no function
for explaining the reason why final schedules were planned. Users
may be dissatisfied with schedules planned by such systems depend-
ing on the application targets.
Many researchers have reported research outcomes for decision
support systems. For example, Qian, C. et al. 46) have reported that
rearranging items displayed on screens can reduce the time required
to solve problems and can improve the quality of the solutions in ar-
rangement problems including production schedules.
This chapter describes the study results of functions required as
scheduling systems and of a human cooperative algorithm along
with its application to cast formation when schedulers are regarded
as decision-making systems.
9.1 Guidelines for designing cooperative type schedulers
To identify specific design guidelines for scheduling problems, a
simple scheduling problem was used to research how means to pres-
ent information and function allocation would affect decision mak-
ing. 47) Results are shown below.
(1) When some attributes that a system did not use for optimiza- Fig. 20 Scheduling algorithm for interactive cast scheduler 48)
tion calculation were hidden from view, the speed at which
people judged the system’s proposals increased in some cases. for which the cast position constraints were severe and thereby their
However, users who referenced such hidden attributes to make production was difficult first and then planned the positions of other
judgments felt more dissatisfaction with the system. charges. In addition, after they planned a single cast formation
(2) Providing functions that allowed users to select the cooperative schedule, they revised it repeatedly while checking the balance of
ratio by humans and the system in a form that users could rec- the entire schedule until there were no more improvement ideas to
ognize increased the degree of acceptance of schedules pro- determine it as the final schedule.
posed by the system: Such functions include one for displaying Figure 20 shows the scheduling algorithm of the cooperative
results in the course of optimization calculation; one for calcu- cast formation system. The system loads target charges first and pro-
lating the schedule by system in half and calculating by hu- poses a rough schedule to a user only showing key charges for
mans after that; one in which humans’ calculation results are which the position constraints are severe. Next, the user checks the
used as initial values and the following processes are calculat- proposed rough schedule and revises the cast positions of the key
ed by the system; and one for manually revising schedules charges. The key charge positions that the user is more concerned
finely. about are agreed at this stage, so the system determines the positions
The item (1) above is common to all decision-making systems: of the remaining charges and proposes the schedule to the user. This
Presenting too much information to humans decreases their cogni- is the detailed primary schedule including all the targets’ charges.
tive capacity, so it is desired to display the minimum necessary in- The user checks the balance of the entire detailed schedule and re-
formation on decision-making systems. In steel production sched- vises it manually, if necessary. The user also uses the system’s auto-
ules, the amount of information required to judge whether schedules matic scheduling function, if necessary, to improve the production
can be accepted or rejected is huge and functions for displaying as sequence in a designated range. These manual revisions and auto-
much information as possible need to be provided. However, we matic improvements in a designated range by the system are repeat-
think that the user interface where only main items are always dis- ed until the user is satisfied with the results and the final results are
played and other necessary information is displayed based on user output as the final schedule.
requests should be provided. In this way, providing two stages in which a rough schedule is
Regarding the item (2), providing a function that leaves a margin agreed first and then a detailed schedule is proposed allows the sys-
for humans to make judgments even in the course of calculation, in tem to propose a detailed schedule that the user can easily accept.
addition to judging calculation results, possibly allows humans to To allow the afore-mentioned scheduling algorithm to work as
feel more that the final schedules are not forced by the system and expected, an easy-to-use user interface is important. Main user in-
they planned the schedules in cooperation with the system, which terfaces realized in the cooperative cast formation system are listed
possibly makes the planned results more acceptable. We consider below.
that in production schedulers, it is important to present a schedule (1) Function for displaying evaluation values for a schedule and
gradually in multiple stages to give a margin for humans to adjust it their details (showing grounds for the proposed schedule)
at each stage, rather than presenting the final schedule one time. (2) Function for coloring violation of constraints, etc. (to make it
9.2 Human cooperative cast formation system easier to judge whether the schedule is acceptable or not)
We have developed a cast formation system that plans produc- (3) Charge drag-and-drop function (for improving schedules coop-
tion schedules for continuous-casting machines based on the design eratively)
guidelines determined in the previous section. 48) It was found that (4) Function for fixing the cast positions of charges (for improving
persons in charge of cast formation did not consider all target charg- schedules cooperatively)
es to be formed equally but they determined the positions of charges (5) Function for grouping multiple charges (for improving sched-

- 44 -
 NIPPON STEEL & SUMITOMO METAL TECHNICAL REPORT No. 121 March 2019

per described the development of production planning and schedul-

ing algorithms for solving problems with raw materials, steelmak-
ing, hot-rolling, and logistics as actual efforts for supporting such
planning work. Another of our works for plates was introduced from
the aspect of production planning throughout the manufacturing
processes. In addition, our work for linking humans and systems at
a higher level to make it possible to plan production schedules that
could not be made only by humans or systems was described.
Improving the efficiency of the implementation and maintenance
of the developed algorithms is also an important task. As an exam-
ple of such effort, Nippon Steel & Sumitomo Metal has been devel-
oping a tool that can create mathematical optimization models auto-
matically just by entering some data on screens without a high level
of mathematical technical knowledge. 51)
As the labor population is expected to further decrease in the fu-
ture, there will possibly be a need to develop technologies for sup-
Fig. 21 User interface of interactive cast scheduler 48)
porting production planning and scheduling throughout multi-stage
manufacturing processes from raw materials to production and ship-
ules cooperatively) ment and for supporting planning schedules for multiple steelworks
(6) Function for designating an improvement range (for improving from the aspect of the entire company. To that end, in addition to the
schedules cooperatively) development of high-speed algorithms for larger-scale mathematical
(7) Functions for enabling and disabling constraints, changing optimization problems, next-generation architectures (e.g., quantum
thresholds, and revising performance index weights (for adjust- computers) may need to be used.
ing the planning performance) In addition, re-planning and rescheduling performance needs to
(8) Function for saving revision history, comparing multiple be enhanced so as to quickly cope with changes in the order config-
schedules, and showing their differences (for comparing multi- uration and operation changes as a result of changes in market
ple schedules in detail) needs. To detect changes promptly and cope with them quickly,
(9) Function for recommending positions into which targets should technologies for analyzing order data and actual operation data will
be put (for supporting manual planning and training to new become increasingly important. Furthermore, with the remarkable
staff) advancement of AI technologies, it is important to appropriately
Figure 21 shows example scheduler screens with the functions judge how humans and systems share manufacturing process man-
listed above. agement tasks including production planning and how scheduling is
9.3 Summary desirable on a long-term basis. Designing advanced user interfaces
Normal production schedulers have no function for explaining that give awareness to humans and that can be intuitively operated
planned results, so users may not be satisfied with the results de- is also important. Nippon Steel & Sumitomo Metal will provide
pending on the application target. Therefore, in this chapter, a sched- these technologies as comprehensive system solutions and will work
uler was regarded as a decision support system and guidelines for to establish general production management technologies that are
designing human cooperative schedulers were proposed. These linked to quality management and operation management in addi-
guidelines were applied to the cast formation system and the effects tion to manufacturing process management.
were verified. The developed cast formation system has been used
at actual manufacturing sites and it has been confirmed that users References
are satisfied with the proposed schedules and excellent cast forma- 1) Nippon Steel Corporation: What are Steel and Steelmaking. Nippon
Jitsugyo Publishing, 2004
tion schedules are obtained. The system has been useful to reduce
2) Iwatani, T.: Problems with Planning Receipt and Shipping of Goods at
planning workloads and manufacturing costs and improve the cast- Steelworks and Optimization Technologies. Communications of the Op-
ing quality. In addition, requirements that cooperative schedulers erations Research Society of Japan. 51 (3), 143–148 (2006)
should have are very versatile, so they have been applied to other 3) Kobayashi, H., Yaji, Y., Saitoh, G., Suzuki, Y.: Optimization of Plans for
scheduling problems in the steelmaking process, achieving suc- Blending Raw Materials for Steel. Communications of the Operations
Research Society of Japan. 56 (11), 633–639 (2011)
cess. 49, 50) 4) Kobayashi, H., Suzuki, Y., Sano, T., Ushioda, Y., Kanazawa, N., Yaji, Y.:
Optimum Scheduling System for Ship Charge and Discharge Schedul-
10. Conclusions ing. CAMP-ISIJ. 25 (2), 1029 (2012)
The steelmaking process is a V-type job shop production process 5) Kobayashi, H., Yaji, Y., Yamada, H., Iwami, N.: Optimum Scheduling
System for Coal Pile Deployment in Coal Stock Yard. CAMP-ISIJ. 22
in which various types of steel products are manufactured in small
(1), 345 (2009)
lots using natural raw materials so as to satisfy the product specifi- 6) Christiansen, M., Fagerholt, K., Nygreen, B., Ronen, D.: Ship Routing
cations and delivery date for each order while aiming at increasing and Scheduling in a New Millennium. European Journal of Operational
the size of lots in each manufacturing process (maximizing the pro- Research. 228, 467–483 (2013)
ductivity). In addition, production schedules need to be planned for 7) Kobayashi, K., Kubo, M.: Optimization of Oil Tanker Schedules by De-
composition, Column Generation, and Time-space Network Techniques.
each manufacturing process in consideration of the various produc- Japan Journal of Industrial and Applied Mathematics. 27 (1), 161–173
tion constraints and performance indexes from the aspects of prod- (June 2010)
uct quality and costs. Experienced skills and know-how are required 8) Iwatani, T.: Field Force at Street Making Process-Scheduling. Tetsu-to-
even when the scheduling unit is a manufacturing process. This pa- Hagané. 97 (6), 316–319 (2011)

- 45 -
NIPPON STEEL & SUMITOMO METAL TECHNICAL REPORT No. 121 March 2019
9) Sasaki, N., Ogawa, Y., Mukawa, S., Miyamoto, K.: Improvement of Hot trol Problems. Transactions of the Society of Instrument and Control En-
Metal Dephosphorization Technique. Shinnittetsu Giho. (394), 26–32 gineers. 53 (8), 437–447 (2017)
(2012) 32) Jensen, K., Kristensen, L. M., Wells, L.: Coloured Petri Nets and CPN
10) Ito, K., Yaji, Y., Komiya, N., Matsumura, T., Oshina, M.: Operational Tools for Modelling and Validation of Concurrent Systems. International
Scheduling System for Steelmaking Process. CAMP-ISIJ. 20 (2), 299 Journal on Software Tools for Technology Transfer. 9 (3), 213–254
(2007) (2007)
11) Ito, K., Umemura, J., Yaji, Y., Urakami, T., Matsumura, T.: Operational 33) Shioya, M., Sadaki, A., Sugiyama, K., Fujii, A., Nishimura, S., Enami, R.,
Scheduling System for Steelmaking Process-2. CAMP-ISIJ. 21 (2), 1145 Fukuya, S.: Automatic Allocation System for AGV by Petri Net Simula-
(2008) tor. CAMP-ISIJ. 21 (1), 285 (2008)
12) Ago, M., Kitada, H., Ito, K.: Development of Models for Optimizing 34) Shioya, M., Yaji, Y., Nagata, S., Uragami, T., Iguchi, H.: Simulator and
Temperature and Schedules at Steelworks Simultaneously. CAMP-ISIJ. Scheduler for Hot-metal Transportation. CAMP-ISIJ. 22 (14), 346
28, 725 (2015) (2009)
13) Ago, M., Kitada, H., Ito, K.: Development of Models for Optimizing 35) Kurokawa, T., Nakajima, H., Suzuki, Y.: Development of Technologies
Schedules Considering Variation in Temperature of Molten Steel at for Solving Slab Stacking Problems at Slab Yards by the Set Partition
Steelworks. CAMP-ISIJ. 30, 783 (2017) Method. CAMP-ISIJ. 29 (2), 688 (2016)
14) Kanisawa, Moriwaki, Umeda et al.: Hot Rolling Order Scheduling by 36) Kurokawa, T., Yaji, Y., Nakajima, H.: Yard Operation Plan Making
Using Genetic Algorithm. CAMP-ISIJ. 14 (2), 16 (2001) Method, Device and Program. Japan Patent Office, Publication of Japa-
15) Sannomiya, Kita, Tamaki, Iwamoto: Genetic Algorithm and Optimiza- nese Patent No. 4987602
tion. Asakura Shoten, 1998, p. 166–178 37) Kurokawa, T., Nakajima, H.: Transportation Control Procedures, Equip-
16) Hama, Yoshizumi: Scheduling Techniques for Sheet Hot-Rolling Proc­ ment, and Computer Programs. Japan Patent Office, Publication of Japa-
ess. Proceedings of Scheduling Symposium. 2006 nese Patent No. 5332872
17) Kurokawa, Adachi, Nagasaka: Development of Hot-Rolling Schedulers 38) Kuyama, S., Tomiyama, S.: Optimization Technology for Crane Han-
Based on Technologies for Optimizing Heating Furnace Charging and dling Scheduling in a Steel Manufacturing Process. JFE Technical Re-
Rolling Sequences Simultaneously. CAMP-ISIJ. 27, 338 (2014) port. 35 (2), 43–47 (2015)
18) Fujii, S., Urayama, K., Kashima, K., Imura, J. et al.: Simultaneous Opti- 39) Tanaka, S.: Recent Studies on Reduction of Item Relocation, Reshuf-
mization of Charging Scheduling and Heating Control in Reheating Fur- fling, and Rehandling. Systems, Control and Information. 61 (3), 88–94
nace. Tetsu-to-Hagané. 96 (7), 434–442 (2010) (2017)
19) Suzuki, M., Katsuki, K., Imura, J., Nakagawa, J., Kurokawa, T., Aihara, 40) Kurokawa, T.: Yard Management Equipment, Procedures, and Computer
K.: Modeling and Real-time Heating Control of a Reheating Furnace Programs. Japan Patent Office, Publication of Japanese Patent No.
Using an Advection Equation. IFAC. 3M, 2012 5365759
20) Suzuki, M., Katsuki, K., Imura, J., Nakagawa, J., Kurokawa, T., Aihara, 41) Kurokawa, T.: Transportation Control Procedures, Equipment, and Com-
K.: Simultaneous Optimization of Slab Permutation Scheduling and puter Programs. Japan Patent Office, Publication of Japanese Patent No.
Heating Control for a Reheating Furnace. J. Process Control. (2013) 5434267
21) Hama, T., Yoshizumi, T.: Procedures for Creating DHCR Schedules 42) Kurokawa, Matsui, Takahashi: Equipment for Planning Slab Stacking
through Linkage of Steelmaking and Hot-Rolling Schedulers. Commu- Schedules, Procedures, and Computer Programs. Japan Patent Office,
nications of the Operations Research Society of Japan. 56 (11), 646–653 Japanese Unexamined Patent Application Publication No. 2017-39556
(2011) 43) Kurokawa, Matsui: Equipment for Planning Slab Stacking Schedules,
22) Ashayeri, J., Heuts, R. J. M., Lansdaal, H. G. L., Strijbosch, L. W. G.: Cy- Procedures, and Computer Programs. Japan Patent Office, Japanese Un-
clic Production - Inventory Planning and Control in Pre-Deco Industry. examined Patent Application Publication No. 2017-40985
Int. J. Production Economics. 103, 715–725 (2006) 44) Kurokawa, T., Oogai, H.: Development of Solving Slab Stacking Prob-
23) Shioya, M., Uchida, K.: Prediction Model to Design Standard Produc- lem by Set Partitioning Approach. Transactions of the Society of Instru-
tion Period for Steel Plate Mills. 3rd International Conference on Con- ment and Control Engineers. 54 (2), 298–306 (2018)
trol, Automation and Robotics (ICCAR 2017). 2017, p. 346–351 45) Suzuki, H., Iwamura, K.: Knapsack Problem and Set Covering (Partition)
24) Kohavi, R., Quinlan, J. R.: Handbook of Data Mining and Knowledge Problem. Communications of the Operations Research Society of Japan.
Discovery. Oxford University Press, 2002, p. 267 24 (6), 359–368 (1979)
25) Shioya, M., Mori, J., Ito, K., Mizutani, Y., Torikai, K.: Development of 46) Qian, C., Itoh, K., Enkawa, T., Akiba, M.: Display factors on Human
Estimation Technology for Standard Production Time of Plate Mill. Problem Solving Performance for Human-computer Cooperative Prob-
CAMP-ISIJ. 26 (1), 238 (2013) lem. The Japanese Journal of Ergonomics. 26 (5), 233–242 (1990)
26) Author; Bishop, C. M., Translation; Motoda, H., Kurita, T., Higuchi, T., 47) Shiose, T., Yoshino, H., Yanagihara, M., Motoyoshi, T., Kawakami, H.,
Matsumoto, Y., Murata, N.: Pattern Recognition and Machine Learning. Katai, O.: On Credible Information Presentation Enhancing the Persua-
Tokyo, Springer Japan, 2007, p. 138 siveness of the Decision Support Systems. Proceedings of the Human
27) Shioya, M., Mori, J., Ito, K., Mizutani, Y., Torikai, K.: Development of Interface Symposium. 2008, p. 835–840
Stochastic Model of Production Time and Estimation Technology for 48) Shioya, M., Shiose, T., Aono, M.: Development of Interactive Cast
Standard Production Time in a Plate Mill. Tetsu-to-Hagané. 101 (11), Scheduling System. Transactions of the Institute of Systems, Control and
574–583 (2015) Information Engineers. 29 (9), 391–400 (2016)
28) Mori, J., Ito, K., Mizutani, Y., Torikai, K., Senzaki, M., Yaji, Y.: Devel- 49) Kobayashi, H., Yaji, Y., Yamada, H., Iwami, N.: Optimum Scheduling
opment of Optimization Method for Continuous Casting Schedules for a System for Coal Discharge from Coal Stock Yard. CAMP-ISIJ. 21 (1),
Plate Mill. CAMP-ISIJ. 25 (1), 355 (2012) 127 (2008)
29) Ito, K., Mori, J., Shioya, M., Nishimura, R., Torikai, K., Mizutani, Y.: 50) Kobayashi, H., Yaji, Y., Yamada, H., Iwami, N.: Optimum Scheduling
Development of Plate Output Schedule Optimization Technologies (2). System for Coal Pile Deployment in Coal Stock Yard. CAMP-ISIJ. 22
CAMP-ISIJ. 30 (1), 288 (2017) (1), 138 (2009)
30) Shioya, M., Mori, J., Ito, K., Mizutani, Y., Torikai, K.: Development of 51) Furukawa, A., Inatomi, M., Umeda, K.: Development of Production and
New Order Entry System for Plate Mills. Asia Steel 2015. 2015, p. 546– Logistics Optimization Technologies and Future Prospects. TEXENG
547 Report, on the Website of Nippon Steel & Sumikin Texeng. Co., Ltd.
31) Shioya, M.: Colored Petri Net Tool Suitable for Solving Logistics Con- (http://www.tex.nssmc.com/pdf/report/ texeng_te1_01.pdf)

- 46 -
 NIPPON STEEL & SUMITOMO METAL TECHNICAL REPORT No. 121 March 2019
Kuniharu ITO Hirokazu KOBAYASHI
General Manager, Head of Laboratories Senior Researcher
Intelligent Algorithm Research Center Intelligent Algorithm Research Center
Process Research Laboratories Process Research Laboratories
20-1 Shintomi, Futtsu City, Chiba Pref. 293-8511

Tetsuaki KUROKAWA Masatoshi AGO

Senior Researcher Senior Researcher
Intelligent Algorithm Research Center Intelligent Algorithm Research Center
Process Research Laboratories Process Research Laboratories

Masanori SHIOYA Junichi MORI

Chief Researcher, Dr.Eng. Senior Manager, Ph.D
Intelligent Algorithm Research Center Computer System Dept.
Process Research Laboratories Production & Technical Control Div.
Kimitsu Works

- 47 -