Logistical Inventory Approach in Forecasting and Relocating Share-use Vehicles

Hao Wang
Ningbo Institute of Technology,
Zhejiang University, P.R. China
wanghao@nit.zju.edu.cn

Ruey (Kelvin) Cheu

Der-Horng Lee

Dept. of Civil Engineering,

University of Texas at El Paso,
United States
rcheu@utep.edu

Dept. of Civil Engineering, National

University of Singapore, Singapore
dhl@nus.edu.sg

from a model of logistical inventory management. The

proposed model consists of three main components, i.e.,
focus forecasting, inventory replenishing, and microscopic
traffic simulation.

AbstractThis research involves a study of the existing

carsharing services in Singapore. In this paper, a novel
method to forecast and relocate vehicle resources for
carsharing services was proposed. The idea is adapted from a
model of logistical inventory management. The proposed
model consists of three main components, i.e., focus
forecasting, inventory replenishing, and microscopic traffic
simulation. Based on the experimental results, the proposed
method has the potential to improve the carsharing service
currently operating in Singapore.

THE FOCUS-FORECASTING MODEL

Introduction
Over the past decade, carsharing has become more
common, especially in Europe and North America, and
increasing developments have also been achieved in Asia.
Auto manufacturers, such as Honda, Toyota, Ford, Nissan,
and DaimlerChrysler, have become more involved in the
deployment of shared-use vehicle fleets in the last few years.
Carsharing may be thought of as organized short-term
car rental. Individuals gain access to vehicles by joining
organizations that maintain a fleet of cars and light trucks
in a network of locations. Instead of owning one or more
vehicles, a household can access shared-use vehicles on an
as-needed basis. Hence, carsharing provides a way to
reduce the number of cars in an urban area and supports a
practical shift away from over-dependency on automobiles,
thus improving air quality and alleviating parking and
traffic problems.
Although most organizations still require two-way
rentals (i.e., vehicles must be returned to the same location
where the trip is originated), many organizations have finetuned their programs and businesses, which could facilitate
one-way rentals (i.e., members are not required to return
their vehicles to the same location). However, this will
inevitably induce unevenly distributed vehicles between
carports. In such situation share-use vehicles will have to
be relocated to ensure a constant, or at least sufficient
supply at every carport to meet the demand.
This paper presents an innovative method to relocate
vehicle resources for carsharing services, which is adapted
978-1-4244-5848-6/10/$26.00 2010 IEEE

314

All supply chain designs and planning decisions are

based on a forecast of customer demand. A focus
forecasting model based on an aggregate approach is
proposed for this ad hoc problem of car-sharing.
Aggregated forecasts are usually more accurate than the
disaggregate forecasts. Aggregate forecasts tend to have a
smaller standard deviation of error relative to the mean. For
example, it is easy to forecast the gross domestic product
(GDP) of Singapore for a given year with less than a 2
percent error. Howevex, it is much more difficult to forecast
yearly earnings for a company with the same degree of
accuracy. Likewise for car sharing system, it is easier to
forecast the total number of vehicles rented out and retuned
over time at each port, than to forecast the details of every
individual vehicle/customers routing behavior.
While it would be advantageous to be able to identify a
specific forecast technique that is appropriate for each
application, the development and evaluation of forecast
techniques are not that exact. The selection of the
appropriate technique is much more of an art than a science.
The concept of focus forecasting illustrates one approach
that is based on results, which incorporates a number of
techniques ranging from very simple to reasonably complex
(see Figure 1).

presented in this report will focus on the direct forecast of

net flow.

Figure 1. Conceptual Framework of Focus Forecasting

Model
In this study, several time series techniques such as
selective moving averages, Holts model, Winters model as
well as Tabu Search heuristics were deployed and
compared (see Figure 2). For each time period, multiple
forecasts are generated. Focus forecasting then applies the
forecast using the technique that would have been the most
accurate if it were used for the most recent historical time
period.

 The inventory-replenishing model

Shortage of inventories can disrupt planned
marketing/manufacturing operations and overstocks
increase cost and reduce profitability in a logistics supply
chain. Likewise, vehicle shortage in a carsharing service
will discourage customers future patronages and even
make them quitting memberships. While overstocks, i.e.
too many shared-use vehicles available in one port, will
induce parking problems.
For each inventory decision of vehicle stocks, major
considerations are reviewed and guidelines are developed.
The decisions include vehicle inventory tracking,
determining when to replenish inventory (determining
order point, when to order?), and determining the amount
to replenish (determining lot size, how much?).
The carport with available vehicles more than the
necessary inventory (including safety stock) is defined as an
overstocked carport. All the overstocked carports constitute
the candidate list of suppliers during the reallocating
operation.
Once the relocating decision is made (i.e. from which
port(s) to which port(s), when to reallocate and how many
vehicles involved), the under-stocked carport will be
replenished from the nearest overstocked carport in terms
of the lowest travel cost at that moment of reallocating (the
shortest lead time for the requested products), which was
determined by the microscopic traffic simulation model.

Figure 2. The Methods Adopted in Focus Forecasting

Model
The forecasted demands were then fed into the
following inventory-replenishing model.
Before we proceed further, it is important to first define
the meanings of supply and demand as used in this paper.
They are as follows:
Supply Number of vehicles being returned to port
Demand Number of vehicles being rented from port
The net flow of vehicles into the port may be defined as
follows:

The microscopic traffic simulation model

4.1 The study network
Instead of using centroids and constant travel costs
between carports, a customized microscopic simulation
model, PARAMICS[1], was adopted to model real-streetnetwork traffic conditions. A portion of the Central
Business District (CBD) area in Singapore, which is
bounded by the Electronic Road Pricing (ERP) gantries,
covering an area of approximately 3.0 km by 2.5 km, and
including several carsharing ports, was used for the
simulations conducted in this study.

Net Flow = Supply Demand

A positive net flow would thus indicate a net flow of
vehicles into the port while a negative net flow would
indicate a net flow of vehicles out of the port. Net flow may
be forecasted using two methods. Utilizing our forecasted
results of supply and demand, net flow may be derived from
the difference between forecasted supply and forecasted
demand. Alternatively, it may be forecasted directly from a
model developed with actual net flow figures in historical
data. Our initial results show that direct forecasting of net
flow produces better results as compared to deriving it from
forecasted demand and supply. Therefore, the results

Figure 3. CBD Network in Singapore

315

able to ensure that better paths will not be discovered before

the entire searching procedure is completed [16,17]. On
the other hand, the label-setting approach will stop the
shortest path search once the destination node becomes
permanently labeled. In view of the purpose of applying
the shortest path algorithm, in this research, adopting the
label-setting approach will save the computational time
significantly.
Typically, the origin and destination are considered as
nodes in the shortest path algorithms. The shortest paths
found by such algorithms are thus the node-to-node (or
node-based) shortest paths. However, for a real road
network, the intersections are usually represented as nodes
while the roads or streets are usually represented as links in
the shortest path algorithms. In the microscopic simulation
in this research, the vehicles are able to adjust the route
choices from their current positions, which are often
located on roads/streets rather than at intersections. Hence,
it could be more reasonable to consider the origin and
destination as links in the shortest path search.
Furthermore, the characteristics of the real road networks
such as the one-way links and turning restrictions imposed
by traffic signal control or regulations would also prefer the
shortest path algorithm based on links rather than the ones
based on nodes. For example, there are usually some
turning restrictions imposed at the intersections and very
few shortest path algorithms in the literature have thus far
incorporated the turn penalties at each node into
consideration. The traditional node-based shortest path
algorithms are not able to differentiate the physically
connected but operationally prohibited movements. Though
such a drawback can be avoided by expanding each turning
movement as separate links via the so-called expanded
network representation, such expansions unavoidably will
incur the increase in network scale. To this end, in order to
take the turning restrictions and penalties into proper
accounts, a link-to-link (or link-based) shortest path
algorithm is proposed and adopted. It is based on the labelsetting approach, Dijkstras algorithm [8].
The proposed algorithm maintains and adjusts a
candidate list V : ( d 1 , d 2 ,... d n ) but is different from
Dijkstras algorithm in that each dj, called the label of link
j (rather than the label of node j ) , is either a scalar or .

The CBD network was chosen because it has the highest

concentration of offices and commercial activities, making
it a densely-populated area within Singapore island.
Several commercial buildings serve as carports for
carsharing service in the traffic simulation model, which is
otherwise relatively uncommon in other road networks such
as expressways and suburban areas.
For network coding, the details of the geometry and
physical layout of the roads were collected via field surveys,
and included such information as the number of lanes (midblock and at intersections), turning restrictions, post speed
limit, etc. The data on signal timing, origin-destination
(OD) and information on the demarcation of zones in the
CBD area was collected from related transportation
authorities [2-4].
The CBD network consists of a total of 894 nodes and
2,558 links. The 100 traffic analysis zones in this network
were defined according to the traffic demands of each zone,
which were allocated according to the acquired OD data.
Figure 3 provides an overview of the CBD network.
A. 4.2

Application program
program for traffic simulation

interface

(API)

An API program was developed to collect the travel time

of each link along the CBD network through traffic
simulations. A link-to-link shortest path algorithm with
turn penalties was embedded into the API program to
search for the shortest time paths for each origindestination pair.
Five locations in CBD area were selected as carports at
major trip generators, e.g., shopping malls, hospitals and
convention centers. Initially, there were 10 cars at each
carport for carsharing service.
1000 sets of car renting demands been tested based on
hypothetical data as well as real-life data from local
carsharing operators. Each set of demand was set with
pickup time points, possession time during four weeks of
the same month, and pickup and drop off locations
distributed among the five carports.
B. 4.3 Link-to-link shortest path algorithm with

turn penalties
The shortest path algorithm has been intensively studied
in the literature. Since the 1950s, a number of shortest path
algorithms have been proposed [5-13]. In general, there are
two types of methods for solving the shortest path problems:
label-setting and label-correcting. The label-correcting
method is usually more efficient for sparse networks such
as an urban road network [14,15]. As a matter of fact, the
nature of the shortest path problem considered in this
research is a indeed a single origin/single destination
problem. As known, the label-correcting approach is not

316

( a 1 , a 2 ,... a n ) is the link cost of each link. In addition,

this algorithm exploits another cost, incurred by the turn

penalties from each link to its outgoing links: Tij , i.e. the
turning cost from link i to link j . In this study, Tij as well
as the link travel cost ( a 1 , a 2 ,... a n ) are calculated and
updated through simulation.
Initially, (link 1 is assumed as the origin)
V = {1},
d1 = 0,

di = , i

1.

The approach proceeds in iterations and

terminates when the destination link is removed
from list V. A typical iteration (assuming V is
nonempty) is as follows:

Do while V is not empty

{

Port 1
Port 2
Port 3
Port 4
Port 5
Sum
Service
level

1
3
2
3
7

2
5
7
1
2
69

2
5
8
1
6

93.1%

&RQFOXVLRQ
This research identifies and explores the potential
employment of a new approach modeled after a logistical
inventory management system. The proposed model
consists of three main components, i.e., focus forecasting,
inventory replenishing, and microscopic traffic simulation.
Experimental results show the efficiency of the proposed
system, and the novel approach has the potential to improve
the carsharing service currently operating in Singapore.
.
Acknowledgements

Removed from the candidate list V, link i

such that

d i = min d j
jV

If ( i is the destination link ) stop the

iteration;

Else
{
For each outgoing link j from link i, if
d j > d i + a j + Tij , set

The authors express their appreciation to the Land

Transport Authority (LTA) for providing the OD data of
the CBD area in Singapore. Special thanks are also due to
Honda Diracc for their assistance in verifying the detailed
information of the existing carsharing operations in
Singapore.

d j = d i + a j + Tij ;
And add j to V if it does not already
belong to V;
}

REFFERENCES
[1] Quadstone, Paramics User and Reference Manuals, Version 3.0,
Scotland, 2000.
[2] Ministry of Law, Singapore Street Directory, 20th Edition, SNP
Publishing, Singapore, 2000.
[3] Land Transport Authority, White Paper: A World Class Land Transport
System, 1996.
[4] Urban Redevelopment Authority, http://www.ura.gov.sg/, Accessed July
1, 2002.
[5] Ahuja, R. K., K. Mehlhorn, J. B. Orlin and R. E. Tarjan, Faster
algorithms for the shortest path problem, Journal ACM 37, 1990, pp.213-223.
[6] Bertsekas, D. P., An auction algorithm for shortest paths, SIAM Journal
on optimization 1, 1991, pp.435-447.
[7] Dial, R. B., F. Glover, D. Karney and D. Klingman, A computational
analysis of alternative algorithms and labeling techniques for finding shortest
path trees, Networks 9, 1979, pp.215-248.
[8] Dijkstra, E. W., A note on two problems in connection with graphs,
Numeriche Mathematik 1,1959, pp.269-271.
[9] Floyd, R. W., Algorithm 97, shortest path, Communication ACM 5,
1962, pp.345.
[10] Gallo, G. and S. Pallottino, Transportation planning models, Shortest
path methods in transportation models, M. Florian (Editor), North-Holland,
1984, pp.227-256.
[11] Glover, F. D., D. Klingman and N. Philips, A new polynomially
bounded shortest path algorithm, Operation Research 33, 1985, pp.65-73.
[12] Goldberg A. V. and T. Radzik, A heuristic improvement of the BellmanFord algorithm, Application mathematics letter 6, 1993, pp.3-6.
[13] Vliet, D. Van, Improved shortest path algorithms for transport networks,
Transportation research 12, 1978, pp.7-20.
[14] Bertsekas, D. P., F. Guerriero, and R. Musmanno, Parallel asynchronous
label correcting methods for shortest paths, Journal of optimal theory and
applications, 89(1), 1996, pp.297320.

}
Using this algorithm and the link travel time updated
through simulations, the shortest paths of all interested
vehicles, from their current positions to their destinations
can be obtained.
Experimental results
The computation works were carried out in a personal
computer with a Pentium IV 1.8 GHz CPU and 512 MB of
RAM.
The performance of the proposed approach has been
tested based on hypothetical data as well as real-life data
from local carsharing operators. The initial results indicate
that the proposed method has good potential in relocating
vehicle resources to balance the demand and supply of
share-use vehicles.
The results are shown in Table 1
Table 1. Experimental Results
No.
of
Shortage
of cars
(Week 1)

2
1
6
4
1

No.
of
Shortage
of
cars
(Week 2)

No.
of
Shortage
of
cars
(Week 3)

No.
of
Shortage
of
cars
(Week 4)

317

[15] Gallo, G. and S. Pallottino, Shortest path algorithms, Annals of

operation research 7, 1988, pp.3-79.
[16] Bertsekas, D .P., A simple and fast label correcting algorithm for
shortest paths, Networks 23, 1993, pp.703709.

[17] Bertsekas, D. P., Network optimization: continuous and discrete models,

Athena Scientific, Belmont, Massachusetts, 1998.

318