You are on page 1of 20

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/254609330

Keeping Passenger Surveys Up to Date: A Fuzzy Approach

Article  in  Transportation Research Record Journal of the Transportation Research Board · January 2000
DOI: 10.3141/1735-05

CITATIONS READS

13 199

3 authors, including:

Markus Friedrich Peter Mott


Universität Stuttgart PTV Group, Karlsruhe, Germany
75 PUBLICATIONS   645 CITATIONS    2 PUBLICATIONS   13 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Qualitätssicherung von Verkehrsmodellberechnungen (Forschungsprojekt SVI 2015/001) View project

Anforderungen an städtische Verkehrsnachfragemodelle View project

All content following this page was uploaded by Markus Friedrich on 08 February 2015.

The user has requested enhancement of the downloaded file.


Veröffentlichung / Publication

Keeping Passenger Surveys up-to-date


A Fuzzy Approach

Autoren / Authors:

Markus Friedrich
PTV AG, Karlsruhe

Klaus Nökel
PTV AG, Karlsruhe

Peter Mott
PTV AG, Karlsruhe

Veröffentlicht in / Published in:

Friedrich, M., Nökel, K., Mott, P. (2000): Keeping Passenger Surveys up-to-date: A
Fuzzy Approach, Transportation Research Records, No. 1735, p. 35-42.

Universität Stuttgart
Institut für Straßen- und Verkehrswesen
Lehrstuhl für Verkehrsplanung und Verkehrsleittechnik
www.uni-stuttgart.de/isv/vuv/
Keeping Passenger Surveys up-to-date
A Fuzzy Approach

recommended for presentation


th
at the 79 Annual Meeting of the Transport Research Board
Washington, January 2000

PTV AG
Stumpfstr. 1
D-76131 Karlsruhe
Tel.: +49-721-9651-0
Fax: +49-721-9651-299
Email: ptv@ptv.de
Friedrich, Mott, Noekel 2

Keeping Passenger Surveys up-to-date – A Fuzzy Approach

Authors:

Markus Friedrich, PTV AG


Stumpfstr. 1, D-76135 Karlsruhe, Germany
Phone ++49-721-9651316, fax ++49-721-9651299, email markus.friedrich@ptv.de

Peter Mott, PTV AG


Stumpfstr. 1, D-76135 Karlsruhe, Germany
Phone ++49-721-9651203, fax ++49-721-9651299, email peter.mott@ptv.de

Klaus Noekel, PTV AG


Stumpfstr. 1, D-76135 Karlsruhe, Germany
Phone ++49-721-9651328, fax ++49-721-9651299, email klaus.noekel@ptv.de

Abstract:

The knowledge of travel demand is an essential prerequisite for analyzing and planning the transport
supply. Obtaining travel demand data for a transit system requires passenger surveys which combine
counts and interviews. As a matter of fact passenger surveys have two unpleasant characteristics. They are
expensive and the results of such studies tend to lose their validity fairly rapidly. For this reason, the
development of techniques which reduce survey costs and keep demand matrices up to date are gaining
increasing interest. The paper will give details of a technique for computer-aided processing of passenger
surveys and present a method for continuous updating of demand matrices. Since traffic surveys only
represent a snapshot situation the proposed updating method employs a Fuzzy approach to consider that
traffic volumes vary within a certain bandwidth.

Keywords:
Passenger survey, matrix correction, Fuzzy theory
Friedrich, Mott, Noekel 3

1 INTRODUCTION

The knowledge of travel demand is an essential prerequisite for analyzing and planning the transport
supply. In transit networks with an integrated fare system precise information on travel demand is also
crucial, if travel demand serves as a key for distributing revenues onto lines of individual operators.
Obtaining travel demand data for a transit system requires passenger surveys which combine counts and
interviews. As a matter of fact passenger surveys have two unpleasant characteristics. They are expensive
and the results of such studies tend to lose their validity fairly rapidly. For this reason, the development of
techniques which reduce survey costs and keep demand matrices up to date are gaining increasing interest.
This paper will give details of a technique for computer-aided processing of passenger surveys and present
a method for continuous updating of demand matrices. Figure 1 outlines the data flow within the
procedure:

• Base year: Computer-aided processing of passenger survey data can help to reduce survey costs and
save time. In order to compute an o-d matrix and traffic volumes for the base year it is necessary to
check the survey data for plausibility, to determine projection factors and to assign the surveyed travel
demand onto the network.

• Current year: The updating process corrects the base matrix by employing current passenger count
data from manual or automatic counts. Since traffic surveys only represent a snapshot situation the
proposed updating method TFlowFuzzy applies a Fuzzy approach to consider that traffic volumes vary
within a certain bandwidth.
Friedrich, Mott, Noekel 4

Figure 1: Overview

passenger survey data supply data base year


Input • boarding counts • stops
• interviews • lineroutes + timetables

processing of survey data

base year
Program
plausibility check, projection, “direct” assignment

base matrix volumes for base year


1 2 3
Result 1 0 10 18
2 12 0 14
3 20 15 0

Input base matrix supply data current year passenger counts


as weighting • stops • link / stop counts
matrix • lineroutes + timetables • fluctuation ranges

current year
Program TFlowFuzzy
matrix correction with entropy maximation

updated matrix volumes for current year


Results 1 2 3
1 0 11 19
2 14 0 14
3 20 16 0
Friedrich, Mott, Noekel 5

2 COMPUTER-AIDED PROCESSING OF PASSENGER SURVEYS

A frequent approach for passenger surveys in urban transit networks suggests to interview the
passengers onboard of the transit vehicle. Surveyors travel on selected service trips of a transit line (=
survey line), where they perform two central tasks:
1. At each stop they count the number of boarding passengers. This results in one boarding-count record
for each stop of a service trip.

2. After each stop they attempt to interview as many of these passengers as possible. The surveyors are
requested to record four stops: (1) the boarding stop where the passenger entered the survey line, (2)
the stop where the passenger intends to alight the survey line, (3) the origin and (4) the destination stop
of the passenger’s trip. Depending on the survey design the passengers may also be asked to provide
additional information, e.g. their ticket-type or trip purpose. Each interview creates one interview
record.

Application of software can help to reduce costs and save time during all phases of a passenger survey:
• Preparation of survey: In the preparation phase software may be applied to obtain a sample of service
trips. Service trips are usually selected with a stratified random sampling method [1] which subdivides
all service trips into homogeneous strata, e.g. all service trips of one line during peak-hours. Software
can also improve duty schedules for the surveyors by minimizing the duty time with crew scheduling
algorithms.

• During interview: Computer-aided interviewing using a palmtop computer allows online validation by
comparing a passenger’s route data to the network and timetable data stored in the computer.
Unfortunately this method has two weak points limiting its value for passenger surveys: it is slower
than the conventional paper & pencil method and it is more expensive, if many surveyors need to be
equipped simultaneously.

• After interview: In case of paper & pencil surveys, software can speed up the process of transferring
paper data into digital databases by scanning the questionnaires. PTV has successfully applied this
approach in numerous surveys in Germany ranging from 10.000 to 500.000 interviews. One machine
can process approximately 3.600 questionnaires per hour.

• Data processing: Once the survey records exist in digital form, three main tasks need to be completed
before the survey data are ready for analysis: (1) plausibility check and completion of the routes
obtained from the interviews, (2) projection of passenger trips based on the boarding counts, (3)
assignment of the trips onto the network.

The following sections of this chapter present the processing method of survey data as it is
implemented in the comprehensive transportation model VISUM [2] (see Figure 2). In addition to
checking and projecting the survey data records, VISUM is capable of assigning the projected interview
Friedrich, Mott, Noekel 6

records as travel demand onto the network similar to o-d matrices during standard transit assignment. This
special assignment process is called “direct” assignment.

Figure 2: Processing of survey data

interview data
? ?
⇒ core information on
S S S S
passenger trips

supply data
plausibility check and
• stops
data completion
• line routes
• timetables ⇒ complete routes S S S S

count data projection


• boarding counts ⇒ routes with
projected volumes S S S S

supply data „direct“ assignment 50


• stops ⇒ traffic volumes S S
• line routes ⇒ service indicators 50
S S
• timetables ⇒ o-d matrix
Friedrich, Mott, Noekel 7

2.1 Plausibility Check and Completion of Routes

A data record obtained from a passenger interview according to the design described above will
contain the following attributes:
• name or identifier of survey line,
• number of boarding stop into survey line,
• departure time at the boarding stop,
• number of alighting stop from survey line,
• number of origin stop or origin zone,
• number of destination stop or destination zone.

These attributes hold the core information on a passenger’s trip. This data set needs to be checked for
plausibility and completed, as it does not include information on the lines and service trips the passenger
used prior to or after the survey line. It also does not contain information on the departure and arrival
times at origin stop, destination stop or transfer stops which is important for calculating travel and transfer
times.

A network model [2]containing detailed data describing the transit lines, i.e. the line routes and
timetables, serves a basis for a plausibility check. The plausibility check compares the stated transit stops,
lines and departure times with the data of the network model. After the plausibility check VISUM tries to
complete each interview record to a whole route which entirely describes a passenger’s trip with all transit
lines used and their related departure / arrival times. This is achieved by using the survey line as an anchor
from where to look back to the origin and ahead to the destination stop. The objective of this step is to
identify suitable preceding and succeeding service trips which were most likely used by the passenger (see
Figure 3). Identification of preceding service trips starts at the boarding stop (S2). Knowing the boarding
time it is possible to determine suitable service trips in two ways:
• direct connections: all service trips which provide a direct connection between origin stop (S1) and
boarding stop (S2) are suitable, as long as they arrive before the stated departure time at S2.
• connection search: all connections (i.e. a sequence of service trips) which can be calculated by a
connection search procedure are appropriate, if they arrive in time at S2.

From the set of possible connections the procedure selects the connection which assures the latest
departure time at the origin stop. Identification of succeeding service trips starts at the alighting stop (S3)
and works likewise.
Friedrich, Mott, Noekel 8

Figure 3: Completion of routes

preceding survey succeeding


lines line lines

S1 ?

S2 S3

point of
observation ? S4

destination stop
boarding stop

alighting stop
origin stop

2.2 Projection of Passenger Trips

Since a passenger survey will normally capture only a sample of all passengers it is necessary to
determine a projection factor for each interview record. This projection factor stands for the number of
travelers which are represented by one interview record. The projection process in VISUM consists of
three steps (Figure 4):

1. Projection of interviewed passengers to all boarding passengers: This projection considers the fact
that surveyors may fail to interview all boarding passengers or that passengers may refuse to answer.
The projection factor PFac1 is obtained by comparing the number of boarding passengers from the
boarding count with the number of interviewed passengers.

2. Projection of surveyed vehicle section to whole vehicle: If a passenger survey does not cover all
sections of the surveyed vehicle, i.e. not all seats of a vehicle or train, a projection factor PFac2 is
Friedrich, Mott, Noekel 9

needed for adjustment. This factor can be derived from the number of seats in the surveyed section and
the total number of seats on the survey line.

3. Projection of surveyed services to all services: As a passenger survey may only inspect selected transit
lines or selected service trips of a line, a projection factor PFac3 is required to consider the entire
transit supply. Projection factor PFac3 can be determined by finding surveyed service trips which
match the non-surveyed service trips, i.e. which show a comparable line route and operation time.

Figure 4: Projection of passenger trips for one interview record

1. Projection of interviewed passengers to all boarding passengers


Input:
• number of boarding passengers per stop BPass S
• number of interviewed passengers per stop IPass
Output:
• Projectionfactor1 = BPass / IPass BPass = 4, IPass = 2
e.g.
BPass / IPass = 2

2. Projection of surveyed vehicle section to whole vehicle


Input:
• total vehicle seats TSeat
• surveyed vehicle seats SSeat
Output:
e.g.
• Projectionfactor2 = TSeat / SSeat total seats / surveyed seats = 4

3. Projection of surveyed services to all services


Input:
6:00
• surveyed service trips
• non-surveyed service trips 7:00 service trip1 1,5
Output:
8:00 service trip2 0,0
• Projectionfactor3 = specific weighting factor
for each service trip service trip3 1,5

In a comprehensive passenger survey which includes transit lines of the entire network, a passenger
using several lines may be interviewed more than once during his trip. To take this into consideration, two
approaches are possible:
Friedrich, Mott, Noekel 10

• The surveyors may only interview passengers on the first transit line, i.e. after boarding at the origin
stop. Thus it can be avoided to interview the same passenger twice.
• The projection includes an additional factor taking into account the number of transit lines used for a
passenger’s trip. The number of transit lines used is equal to the number of transfers + 1. Thus the
following projection factor can be determined for each interview record:

PFac1 × PFac2 × PFac3


Projectionfactor PFac =
number of transfers + 1

Using the data from Figure 4 an interview record containing three transit lines, i.e. two transfers, for
example, would obtain a projection factor of 4 (= 2 × 4 × 1.5 / 3). Thus this interview record represents
four passenger trips which must be assigned onto the transit network.

2.3 Assigning Surveyed Passenger Trips onto the Network

Standard transit assignment distributes passenger trips stored in an o-d matrix onto the lines of a transit
network. “Direct” assignment of passenger trips from interview records differs from standard assignment
in a number of aspects:

• Direct assignment does not require a route or connection search procedure. The routes of the pass-
engers are already identified during they step “plausibility check and completion of routes” (see 2.1).

• Direct assignment has a lower degree of freedom, since interview records provide more information
than o-d matrices. In addition to the origin and destination zones, interview records hold information
on transfer stops and departure times. Traffic volumes calculated under consideration of these
constraints are generally very close to cross-sectional traffic counts.

• Whilst an o-d matrix contains exactly one demand value for any o-d pair, there may exist several inter-
view records for one o-d pair, each describing a specific route. The projection factor PFac of the
interview record corresponds to the demand value of the o-d matrix.

Despite these differences both, standard and direct assignment, produce the same types of result: traffic
volumes, service indicators (e.g. travel time), travel demand indicators (e.g. passenger-kilometers) and
routes for analysis of traffic flows on selected links and stops.

Within VISUM direct assignment is implemented as an alternative assignment method. This permits
users to analyze and display the results of standard and direct assignment in the same manner. It has been
found that the results of direct assignment are beneficial calibrating the assignment parameters of standard
assignment, e.g. transfer penalties and sensitivity to marginal changes in travel time.
Friedrich, Mott, Noekel 11

3 UPDATING DEMAND MATRICES

3.1 Basic Principles

For some 20 years now, primarily in English-speaking countries, so-called matrix correction (or matrix
update) techniques have been used to produce a current travel demand matrix from an earlier travel
demand matrix (base matrix) using current traffic count values. Based on research by VAN ZUYLEN /
WILLUMSEN [3, 4], BOSSERHOFF [5] and ROSINOWSKI [6] which focuses on matrices for private
transport, PTV has extended the application of these techniques to public transport.

Starting point of the classic approach is the travel demand tij for o-d pairs. Travel demand is usually
described as a matrix, but for our purposes a vector representation containing all non-zero o-d trips is
more suitable:
⎛ t12 ⎞ ⎛ t1 ⎞
⎜ ⎟ ⎜ ⎟
⎜ t13 ⎟ ⎜ t 2 ⎟
⎜ M ⎟ ⎜t ⎟
t1n ⎞ ⎜ ⎟ ⎜ ⎟
3
⎛0 t12 t13 L
⎜ ⎟ ⎜ t1n ⎟ ⎜ M ⎟
⎜ t 21 0 t 23 L t 2n ⎟ ⎜ ⎟ ⎜ ⎟ where:
⎜t t tk
t 32 0 L t 3n ⎟ = ⎜ 21 ⎟ = ⎜ ⎟ tij... travel demand between zone i and j
⎜ 31 ⎟ ⎜t ⎟ ⎜ M ⎟
⎜ M M M O M ⎟ ⎜ 23 ⎟ ⎜ ⎟ p ... number of non-zero o-d pairs
⎜t M M
⎝ n1 tn2 t n3 L 0 ⎟⎠ ⎜ ⎟ ⎜⎜ ⎟⎟
⎜ t 2n ⎟ M
⎜ ⎟ ⎜ ⎟
⎜ t 31 ⎟ ⎜ M ⎟
⎜ M ⎟ ⎜t ⎟
⎝ ⎠ ⎝ p⎠

This vector describes the travel demand of an earlier state. Vector value tk describes the number of trips
for the kth o-d pair with non-zero trips. Index p gives the total number of non-zero o-d pairs Considering
the current travel demand, it is assumed that no o-d specific information is available but only traffic
counts. For public transport such traffic counts may be available either as counts of boarding and alighting
passengers at transit stops or as link counts. In case of boarding/alighting counts it is important to note,
that only initial boardings at the origin stops and final alightings at the destination stop may be employed
to update a matrix, i.e. the counts must not include transferring passengers. The following vector v denotes
traffic counts at m locations:
v T = (v 1 v2 v3 L vl L vm ) where: m ... number of traffic count locations

The trips of any o-d pair contribute a certain share to each traffic count. In case of boarding and
alighting passengers the marginal sums of the demand matrix are known. In case of link counts the
Friedrich, Mott, Noekel 12

counted volumes correspond to the sum of all o-d trips travelling on this link. In general there is a linear
relation between the demand on the o-d pairs and the traffic counts:

⎛ a11 L a1 p ⎞
⎜ ⎟
⎜ M O M ⎟ ⋅ t= A ⋅ t= v
⎜a ⎟
⎝ m1 L a mp ⎠

A is called the share-matrix. The number of columns of this share-matrix refers to the number of non-
zero o-d pairs, the number of rows corresponds to the number of traffic counts. Each element alk of this
share-matrix expresses the share of trips of one o-d pair k which uses link l. The share-matrix A has an
exceptionally regular form in case of boarding and alighting counts. The following example shows the
share-matrix A for a network with 3 zones (n = 3) and 6 counts (m = 6), i.e. 3 boarding and 3 alighting
counts:

⎛1 1 0 0 0 0 ⎞⎛ t12 ⎞ ⎛ board1 ⎞
⎜ ⎟⎜ ⎟ ⎜ ⎟
⎜0 0 1 1 0 0 ⎟⎜ t13 ⎟ ⎜ board 2 ⎟
⎜0 0 0 0 1 1 ⎟⎜ t 21 ⎟ ⎜ board 3 ⎟
⎜ ⎟⎜ ⎟ = ⎜ ⎟
⎜0 0 1 0 1 0 ⎟⎜ t 23 ⎟ ⎜ alight1 ⎟
⎜1 ⎜ ⎟ ⎜ ⎟
⎜ 0 0 0 0 1 ⎟⎟⎜ t 31 ⎟ ⎜ alight 2 ⎟
⎜0 1 0 1 0 0 ⎟⎠⎜⎝ t 32 ⎟⎠ ⎜⎝ alight 3 ⎟⎠

For boarding and alighting counts the share-matrix A does not depend on the transit supply, i.e. the line
routes and timetables. In case of link counts, however, the share-matrix A needs to consider the route
choice of passengers which is affected by the transit supply. To create a share-matrix for link counts one
can assign any demand matrix, e.g. the outdated base matrix, onto the current network. It is also possible
to combine boarding/alighting counts and link counts.

The chief problem of matrix correction methods results from the fact that typically m << n 2 , so that
the new matrix is underconstrained by the counts. The consequence of this is an extremely large number
of combinations for the o-d values (tij) matching with the counts. In order to select the best of all possible
matrices an evaluation function q is applied as objective function:

maximize q(t )
so that A ⋅ t = v

Usually a combination of entropy and weighting with the structure of the outdated matrix serves as
evaluation function. Maximizing q in the following evaluation function favours matrices which differ least
from the base matrix:
Friedrich, Mott, Noekel 13

p
tk
q(t ) = − ∑ t k ⋅ ln − tk
k =1 tˆk

where:
tˆk ... travel demand on one o-d pair in the base matrix
t k ... travel demand on one o-d pair in the new matrix

3.2 The Fuzzy Approach

The formulation of the matrix-correction-problem, as it is described above, has one serious weakness:
the vector v of traffic counts is assumed to be a known input value without any uncertainty. As a result the
traffic counts obtain an inappropriate weight since any count only provides figures, that represent a
snapshot situation which are subject to considerable sampling error. For this reason, PTV decided to
employ an approach developed by ROSINOWSKI [6] who models the counts as imprecise values based
on Fuzzy Sets theory. If one knows, for example, that the number of alighting passengers in an area
fluctuates by up to 10 % on a day-to-day basis, but in other areas by up to 20 %, then this is represented by
appropriate bandwidths. To consider this in the matrix-correction-problem the exact count values are
replaced by Fuzzy Sets ~ v s with varying bandwidths (see Figure 5):

maximize q(t )
such that A ⋅ t = ~
v

Compared to simple intervals, representing counts as Fuzzy Sets allows to favour count values close to
the mean value. Values which are closer to the bandwidth border are also accepted, if the decrease in the
fuzzy membership function is more than offset by the increase in the entropy function. Unfortunately in
the general case fuzzy membership functions do not have the analytic properties (continuous,
differentiable) that are required for an efficient solution algorithm. The compromise taken in this approach
resembles a straightforward representation of counts as intervals with upper and lower bounds. However,
the entropy function is extended to cover also the slack variables introduced by the inequality constraints.
Just like the outdated matrix values are used as weights for the o-d relations, the fuzzy bandwidths are
used as weights for the slack variables, i.e. solutions are preferred in which the counts are matched
exactly. Hence the maxmization problem becomes:
Friedrich, Mott, Noekel 14

maximize q ( t ) + q ( s ) + q (s )
such that A ⋅ t + s = v
A⋅t −s = v
s≥0
s≥0

where:
m
sl
q(s ) = − ∑ sl ⋅ ln − sl
l =1 sˆl
m
sl
q(s) = − ∑ s l ⋅ ln − sl
l =1 sˆ l
v, v ... maximum/minimum of Fuzzy Set
s , s ... slack variables
sˆ = v - v ... upper bandwidth for traffic count
sˆ = v - v ... lower bandwidth for traffic count

Incorporating the slack variables s , s into the weighted entropy maximization gives preference to
matrices which achieve A ⋅ t= v within the bandwidth „as good as possible“. Thus the introduction of
Fuzzy Sets expands the solution space and produces generally better values for the evaluation function q
which maximizes entropy.

The new formulation of the problem is structurally identical to the original formulation of the problem
with exact constraints. The non-linear maximization problem with equality constraints is solved by the
Lagrange multiplier method, solving for zeros in the (non-linear) first partial derivatives using Newton’s
method. By exploiting the symmetry in the structure of the constraints the computational effort can be
reduced to little more than in the original (exact) problem.
Friedrich, Mott, Noekel 15

Figure 5: Fuzzy Sets

“approximately 40” as Fuzzy Set


1

Fuzzy function

0
30 40 50
v-s count v v +s

bandwidth

3.3 Procedure

The following gives an overview of the steps involved in the updating process for an existing matrix
based on current passenger counts with TFlowFuzzy:

• Collect traffic counts of boarding / alighting passengers or cross-sectional link counts and estimates of
the fluctuation ranges obtained from random samples.
• If link counts are applied, then build a network model [2] containing the transit supply, i.e. stops, line
routes and timetables.
• Determine share-matrix A by assigning an outdated o-d matrix onto the current transit network. The
share matrices may be calculated automatically by TFlowFuzzy as VISUM stores all routes during the
assignment procedure.
• Calculate the new demand matrix applying the TFlowFuzzy method.
• Display and check validity of the new matrix by assigning it to the current network and comparing the
resulting passenger volumes with the counted volumes.
Friedrich, Mott, Noekel 16

4 PRACTICAL EXPERIENCE

The method of processing passenger survey data, as it is described in chapter 2, is regularly applied in
Germany for urban surveys as well as interregional surveys. The recent development of the TFlowFuzzy
methods was first tested in 1998. In a pilot study for a public transport operator operating in the North
German city of Lübeck (Stadtwerke Lübeck) TFlowFuzzy was employed to update a demand matrix from
1996 with traffic counts from 1997. In 1997 the operator had fundamentally modified the transit network
so that out of 24 transit lines only nine were left without changes.

The validity of the new matrix for 1997, calculated with TFlowFuzzy, was checked as follows: first the
previous matrix (1996) and the new matrix (1997) were assigned to the new network (1997). Then the
assigned link volumes as well as the calculated number of boarding and alighting passengers were
compared to the count data. In both cases, it was evident that the recalculated matrix provided a far better
match than the original matrix. The link-related volumes of the ‘97 matrix deviated from the counted
volumes only by ±2% for almost 75% of the 950 links. For the ´96 matrix this could only be achieved for
13% of the links (see Figure 6). These results demonstrate clearly that the use of the updated matrix for
analysis and planning purposes produces significantly better results than could have been expected with
the 1996 matrix.

Figure 6: Comparison of 1996 and 1997 matrices based on deviations of assignment result and
traffic counts

300

250

200
number of links

150

100
-69%

-63%

-58%

-53%

-48%

-43%

-38%

-33%

-28%

-23%

-18%

50
-13%

-8%

-3%

2%

7%

12%

17%

22%

27%

32%

37%

0
43%

51%

63%

deviation outdated
matrix 96
updated
matrix 97
Friedrich, Mott, Noekel 17

In the meantime TFlowFuzzy is the established method used by German Railways (DB AG) for
updating ridership figures on metropolitan commuter trains (S-Bahn). It is currently employed in the
Cities of Hamburg and Munich with the purpose of obtaining matrices and performance indicators on a
monthly basis. Conducting regular counts allows to produce matrices which do not describe only a
snapshot situation but show the dynamic demand flow throughout the seasons.

Figure 7 shows a VISUM screenshot with the Munich network. In Munich 34 out of 290 cross-sections
are surveyed once a month on working days and on weekends. 76 % of the passenger trips traverse at least
one of these 34 census points.

Figure 7: Application of TFlowFuzzy for Munich’s S-Bahn system

census
point
Friedrich, Mott, Noekel 18

5 CONCLUSION

The presented methods for processing survey data and updating demand matrices, both implemented in
the transportation model VISUM which is part of the commercial software package PTV-VISION, supply
transit planners with travel demand data. Providing up-to-date demand data is vital for the planning
process and improves the quality of planning results. Although each method may be applied independently
the combination of both methods offers an efficient way to integrate the processes of surveying and
updating.

The success of several pilot projects shows that the technique developed by PTV for updating existing
demand matrices using simple passenger counts is a very practical tool and produces excellent results.
Using automatic passenger count systems would add three benefits:

• The need for census personnel would be eliminated or would be required only in areas where suitably
equipped vehicles cannot be used.

• By collecting count data over a number of days, the fluctuation range of the ridership counts can be
determined precisely, thus improving the results.

• Frequent count surveys, conducted not only on an annual basis, would allow to produce matrices for
different times of the year or for certain days.

6 REFERENCES

1. Ortúzar, J. D., Willumsen, L. G.: Basic sampling theory. In Modelling Transport, Wiley, Chichester,
1990.

2. Friedrich, M., Modelling Public Transport - A European Approach. Preprint CD-ROM of 78th Annual
Meeting, Transport Research Board, Washington 1999.

3. Zuylen, H. J. van, Willumsen, L. G.: The most likely trip matrix estimated from traffic counts, Trans-
portation Research, Vol. 14b, 1980.

4. Willumsen, L. G.: Estimation of an O-D matrix from traffic counts – a review, Working Paper No. 99,
Inst. for Transport Studies, Univ. of Leeds, 1978.

5. Bosserhoff, D.: Statische Verfahren zur Ermittlung von Quell-Ziel-Matrizen im öffentlichen Personen-
nahverkehr – ein Vergleich, Dissertation, University of Karlsruhe 1985.

6. Rosinowski, J.: Entwicklung und Implementierung eines ÖPNV-Matrixkorrekturverfahrens mit Hilfe


von Methoden der Theorie unscharfer Mengen (Fuzzy-Sets-Theorie), Master thesis, University of
Karlsruhe 1994.

View publication stats

You might also like