Professional Documents
Culture Documents
D Ronen
To cite this article: D Ronen (2002) Marine inventory routing: shipments planning, Journal of the
Operational Research Society, 53:1, 108-114, DOI: 10.1057/palgrave.jors.2601264
Article views: 37
This paper addresses a shipments-planning problem faced by producers of large volume liquid bulk products. Producing
origins with limited tank storage capacity supply multiple products by ships (or barges) to consuming destinations that
also have limited storage capacity. Timing, origin, destination, and product quantities of shipments have to be determined
in a manner that minimizes costs and does not violate storage capacity constraints at both ends (neither stopping
production at the origins, nor running out of stock at the destinations). A mixed integer-programming model is used to
derive cost effective solutions within a few minutes. A cost-based heuristic algorithm is used to assure that acceptable
solutions are obtained quickly.
Journal of the Operational Research Society (2002) 53, 108–114. DOI: 10.1057=palgrave=jors=2601264
Keywords: distribution planning; marine transport; integer programming; supply chain; inventory
Marine inventory routing problems often involve inven- tactical ones which focus on the efficient routing and
tory considerations at the sources (in addition to the scheduling of vessels over a relatively short planning hori-
destinations)—exceeding tank capacity at the sources zon. Strategic decisions determine the required storage
should be avoided because it halts production. capacities at the sources and at the destinations and the
Due to the relatively small number of standardized fleet size and mix (especially when the fleet consists of
products involved, shipments may often be sourced at special purpose vessels or when no chartering of vessels is
several (alternate) facilities. Thus, shipment sourcing possible). The tactical decisions are shipment sourcing,
decisions may be required. sizing and timing, as well as vessel chartering, routing and
A voyage of a vessel usually has one or very few scheduling. Larson7 dealt with a strategic problem: the
unloading locations. design of a system to transport sludge by barges from
Cost of inventory at both ends (at the sources and at the sewage treatment plants to an ocean dumping site. Miller8
destinations) is usually borne by the shipping organiza- dealt with a tactical inventory routing problem: scheduling
tion. In the short run the major component of the of chemical tankers to deliver products to warehouses
inventory holding cost is the marginal cost of capital worldwide while maintaining minimal inventories (safety
which is usually the same corporate-wide. In the short run stocks) at the destinations. Similarly, this work addresses a
the cost of storage facilities cannot be changed. Therefore tactical marine inventory routing problem but also takes into
inventory holding cost may be ignored in this short run account storage volume limitations at the sources. The major
planning problem. However, when the ends are in differ- contribution of this work is in the determination of the
ent countries import taxes may differentiate the inventory timing, source and product composition of each shipment
holding costs. (multiple, non-mixable products), while accounting for the
Information about inventory levels at the destinations is product-specific storage limitations at the origins and the
usually available, but marine transit times are usually destinations. All that is achieved at minimal set-up and
much longer (days instead of hours). variable shipping costs. The next section provides a descrip-
(Optional) backhauls may be available for the vessels. tion of the problem. It is followed by the mathematical
Often vessels may be chartered-in or chartered-out, as model used to solve it, and results for ten problem instances
necessary (ie the fleet size and mix may be changed at a of different sizes. Finally, practical extensions are outlined.
fairly short notice).
Vessels come in a large variety of sizes with different The problem
compartment sizes. When a shipment consists of multiple
An organization ships bulk products from a set of origins to
products that have to be loaded on the same vessel the
a set of destinations by a fleet of ships or barges. In contrast
product quantities will have to be adjusted to fit the
to packaged goods, the products cannot be mixed and must
available compartments.
be stored and shipped in separate compartments. There is a
This paper focuses on marine inventory routing, and known limited storage capacity for each product in each
addresses the question of how much to ship of each product, origin and at each destination. The products are produced at
from which origin to which destination, when, and by which the origins according to a production plan, and consumed at
vessel. That has to be accomplished in a manner that the destinations according to a demand forecast. Due to
minimizes shipping costs and does not violate inventory uncertainties in demand and in production, prescribed safety
and storage limitations at both the origins and the destina- stocks of each product have to be maintained at the origins
tions. During the last decade the author has encountered and the destinations. Physical limitations at the loading
about a dozen companies that face such marine inventory facilities dictate the maximal vessel size that may load at
routing problems. any given origin, and similarly there is a maximal vessel size
Large volume product storage facilities are substantial which may discharge at any given destination. A vessel’s
capital investments that are not made easily. Therefore large voyage has a single loading port and a single discharging
volume shippers usually operate with limited storage capa- port, but a vessel may carry multiple products (in separate
city at their origins and destinations. This limited storage compartments). The problem is to determine when and how
capacity necessitates planning shipments to prevent storage much to ship of each product from which origin to which
overflow at the origins (which may cause stoppage of destination and by which vessel in a manner that will
production), and product shortage at the destinations. minimize the total shipping cost and not violate the safety
In contrast to the vast body of literature dedicated to stock and storage volume limitations. This type of problem
vehicle routing, relatively little attention has been directed to has been encountered in the transportation of crude oil to
ship routing and scheduling (see Ronen6). Only a couple of refineries, oil distillates to tank terminals (distribution
works address marine inventory routing problems. Larson7 centers), base lubes from refineries to lube plants, heavy
differentiated between strategic inventory routing problems oil from refineries to industrial customers, intermediate
which are focused on the distribution system design, and chemicals among plants, and other instances.
110 Journal of the Operational Research Society Vol. 53, No. 1
The approach used here separates the solution of the Fij fixed (setup) shipping cost from source i to
problem into two stages, first the determination of the slate destination j ($/shipment).
of shipments to be shipped, and second, the scheduling of INVDjk starting inventory at destination j of product k
vessels to ship them. This approach was chosen because (units).
solving the whole problem simultaneously for problems of INVSik starting inventory at source i of product k (units).
practical size is intractable, and the interaction between the MAXDj maximal shipment (vessel) size to destination j
two stages is limited. The same approach was taken by Equi (units).
et al 9 in solving a similar problem involving trucking of MAXSi maximal shipment (vessel) size from source i
lumber. An earlier work (Bausch et al10) dealt with schedul- (units).
ing of vessels for delivery of a specified slate of shipments, Pikt planned production at source i of product k on day
and therefore the determination of the slate of shipments is t (units).
addressed here. SSSik safety stock level at source i of product k (units).
The customary planning horizon for such problems SSDjk safety stock level at destination j of product k
usually covers several voyages of the vessels. When vessel (units).
voyages are relatively short (several days), the planning Tij transit time from source i to destination j (days).
horizon is one month with daily time resolution. When Ujkt planned usage at destination j of product k on day
vessel voyages are longer (weeks), the planning horizon may t (units).
span several months with weekly time resolution, or even
daily resolution for the first several weeks.
The cost of a shipment from a specific origin to a specific
destination depends on many factors. In this work that cost Decision variables
is represented by a linear function of the size of the
shipment with a large set-up component. This shipping Xijkt Quantity shipped from source i to destination j of
cost relationship has been verified by using linear regression product k on day t (units).
to analyse spot charter rates (regression of spot charter cost
vs shipment size) for such an operation in a market where
the rates are set on a cost-plus basis. R-squared values of (
0.98 and higher were received. These results suggest that the 1 if any xijkt > 0
spot charter rates may have been derived from such a linear Yijt ¼
relationship. 0 otherwise
Continuous Binary
Problem no. Origins Destinations Products Time periods variables variables Constraints
1 1 2 2 15 330 30 240
2 1 2 2 30 660 60 480
3 1 3 2 30 900 90 660
4 1 4 2 30 1140 120 840
5 1 5 2 30 1380 150 1020
6 1 5 3 30 2070 150 1380
7 2 5 2 30 1860 300 1440
8 2 5 3 30 2790 300 1860
9 2 5 4 30 3720 300 2280
10 2 5 5 30 4650 300 2700
Table 2 Solutions
MIP solution
Problem no. Cost* Seconds Nodes LP bound Heuristic solution MIP/bound HEUR=bound
1 57 600* 15 4686 47588 64 500 1.21 1.36
2 134 100* 4 271 132 900 147 900 1.01 1.11
3 296 000 10 460y 275 600 297 000 1.07 1.08
278 000* 237 18 711 1.01
4 278 000 20 680y 269 000 293 200 1.03 1.09
274 000 56 3600y 1.02
Same 7202 500 000
5 299 200 10 190y 295 400 332 600 1.01 1.13
Same 7786 500 000
6 358 300 20 190y 347 533 358 400 1.03 1.03
Same 13 703 250 000
7 222 150 7 130y 192 300 227 200 1.16 1.18
193 650* 164 5581 1.01
8 267 850 22 220y 244 975 270 000 1.09 1.10
261 400 102 1080y 1.07
251 350 461 8180y 1.03
Same 7851 250 000
9 332 750 12 0y 293 633 332 350 1.13 1.13
313 600 241 860y 1.07
302 600 57 708 221 600y 1.03
Same 64 986 250 000
10 369 350 15 0y 329 658 375 700 1.12 1.14
340 800 39 650 122 400y 1.03
Same 84 960 250 000
*Optimal within 1% tolerance.
y
Several nodes later.
Table 3 Analysis of values of binary variables in some MIP and LP bound solutions
Problem no. Binary variables Positive Value less than 0.5 Positive Positive in both LP and MIP
1 30 8 6 4 3
2 60 18 12 7 2
7 300 74 64 18 5
bound and the best known integer solution at which the SOLARIS 5.6 operating system. The number of branch-and-
search for better integer solutions stops). The processing bound nodes was limited to 500000 for the smaller
times required to identify the integer solutions were problems and 250000 for the larger problems. No additional
recorded. The reported times are in seconds on a 248 MHz parameter settings were used in CPLEX. For each problem
Sun Microsystems ULTRA Enterprise computer under the the best integer solutions identified by CPLEX within 30
D Ronen—Marine inventory routing 113
seconds and within 300 seconds are presented in Table 2. In of vessel loadings (discharging) or volume loaded
addition, the best integer solution found within the limit on (discharged) in a time period. Voyages with multiple loading
the number of nodes is listed thereafter. (or discharging) ports are harder to incorporate because
The heuristic identified solutions that are usually within transit times depend on the itinerary of the specific vessel,
15% of the LP bound, and within 10% of the best known but may be accommodated through higher safety stocks. If
integer solution. These solutions are derived in a negligible there are differences in inventory holding costs among the
amount of time (usually less than a second). However, in all various locations the inventory holding costs can be easily
cases, a better solution than the heuristic one was found by added to the objective function because the daily inventory
the MIP model within at most five minutes. Because such balances are provided as decision variables.
performance is not assured we retain use of the heuristic. Although the results are not as close to optimality as one
The MIP problems are very hard to solve (see Table 3). would like them to be, this is a hard real-life problem, and a
Comparing the best known MIP solutions to their LP bound first attempt to solve it optimally. While obtaining optimal
solution shows that a relatively large number of binary solutions is desirable, deriving high-quality solutions
variables (close to 30%) have a positive value in the LP quickly is essential for any practical application.
bound solution. The number of binary variables that are set
to 1 in the best known MIP solutions is much smaller. In
addition, the overlap between these two sets of binary
Appendix: outline of the heuristic
variables is small, and variables that have a value of 1 in
the LP bound end up with a value of 0 in the best known 1. Start and load data
integer solution. Also, most of the positive binary variables 2. Project inventories at all sources and destinations for
in the LP bound solutions are close to zero. Thus, the LP every product every day in the planning horizon
bound solution does not help to determine which binary 3. Pull moves:
variables should be set to 1 or 0. 3.1. Take the next day (if none left go to 4)
Comparison of the heuristic solutions to the correspond- 3.2. Take the next destination (if none left go to 3.1)
ing best known integer solutions indicates that the heuristic 3.3. Take the next product (if none left go to 3.2)
tends to ship as late as possible (in order to save money and 3.4. Is the product below its safety-stock level? If no—
retain shipping flexibility), whereas earlier shipments in the go to 3.3
best known integer solutions save shipments later on. The 3.4.1. Identify all products that are below their
heuristic takes a greedy approach while the optimization has safety-stock level
a wider perspective. In most cases, good solutions (within 3.4.2. Identify the cheapest source that has all these
10% of the bound) were found, either by the heuristic or by products available
solving the MIP model, within several minutes. 3.4.3. Construct the largest shipment possible from
that source by equalizing product run-out
times at this destination (integer number of
days) without violating storage constraints of
any product and shipment size limits
Conclusion
3.4.4. Update projected inventories at the source
A common shipments planning problem is presented and and the destination
solved using a heuristic and a mixed integer programming 3.4.5. Go to 3.2
model. Between these two methods good solutions of 4. Push moves:
known quality are provided quickly. Instead of looking for 4.1. Take the next day (if none left go to 5)
feasible solutions using rules of thumb, shipments plan- 4.2. Take the next source (if none left go to 4.1)
ning can be performed with the explicit objective of cost 4.3. Take the next product (if none left go to 4.2)
minimization. 4.4. Is product overflow projected? If no—go to 4.3
The model presented here may be extended to reflect 4.4.1. Identify all products that cause overflow
additional practical considerations. Imbalance between 4.4.2. Identify the cheapest destination that can take
production, demand, and storage capacity may cause these products
complications that are not recognized by the MIP model. 4.4.3. Construct the largest shipment possible to
Product tank overflow at a source may be prevented by early that destination by equalizing product run-
loading of shipments on vessels while paying the daily out times at the destination (integer number
demurrage of the vessels (ie using the vessels as floating of days) without violating storage constraints
tanks). Product stock-outs at the destinations can be incor- of any product and shipment size limits
porated by imposing penalties on them. In addition, limited 4.4.4. Update projected inventories at the source
loading capacity at a source (or discharging capacity at a and the destinations
destination) may be conveyed as a constraint on the number 4.4.5. Go to 4.2
114 Journal of the Operational Research Society Vol. 53, No. 1