Sei sulla pagina 1di 25

Validating the passenger traffic model for Copenhagen

Goran Vuk
1,
* & Christian Overgaard Hansen
2


1
Danish Transport Research Institute, Knuth-Winterfeldts All, Bygning 116 Vest, 2800 Lyngby, Denmark (E-mail:
gv@dtf.dk);
2
Centre for Traffic and Transport, Technical University of Denmark, Bygning 115, 2800 Lyngby, Denmark (E-mail:
coh@ctt.dtu.dk)
* Author for correspondence

Abstract The paper presents a comprehensive validation procedure for the passenger traffic model for Copenhagen
based on external data from the Danish national travel survey and traffic counts. The model was validated for the
years 2000 to 2004, with 2004 being of particular interest because the Copenhagen Metro became operational in
autumn 2002.
We observed that forecasts from the demand sub-models agree well with the data from the 2000 national travel
survey, with the mode choice forecasts in particular being a good match with the observed modal split. The results
of the 2000 car assignment model matched the observed traffic better than those of the transit assignment model.
With respect to the metro forecasts, the model over-predicts metro passenger flows by 10% to 50%.
The wide range of findings from the project resulted in two actions. First, a project was started in January 2005 to
upgrade the models base trip matrices. Second, a dialog between researchers and the Ministry of Transport has been
initiated to discuss the need to upgrade the Copenhagen model, e.g. a switching to an activity-based paradigm and
improving assignment procedures.

Key words: demand traffic modelling, assignment modelling, model validation, metro system

1. INTRODUCTION
Model validation is a method of ensuring that the model replicates the observed conditions and produces reasonable
forecasting results. The literature in the field of model validation in general reveals that no well-established
framework exists while research into the validation of demand and assignment models for transit modes is
alarmingly poor (Barton-Aschman Associates and Cambridge Systematics, 1997).
The travel demand part of the traffic model for Copenhagen, the Orestad Traffic Model (OTM), is a state-of-practice
nested logit model in which the generation, distribution and mode choice sub-models are connected via logsums.
The OTM forecasts tours for an average workday undertaken by car, bus, train, metro, bicycle and walk. The tours
are converted into day trips before the assignment procedure. The OTM was estimated and calibrated for the year
2000, when the Copenhagen Metro was still under construction, and the models metro forecasts were therefore the
main preoccupation of city planners and local politicians at the time. On the other hand, the model accuracy for
personal car traffic was the main concern of the Danish Road Directorate. As a result, the OTMs objectives relate to
private cars and transit modes, while the slow modes are of secondary importance, mainly due to poor counts.
The OTM is an operative traffic model for Copenhagen often used for generating 10- to 20-year forecasts. The
prime objective of the paper is to list and apply a range of available tools for model validation, according equal
weight to the demand and assignment elements of the model. These tools enabled us to identify the accuracy and
weaknesses of the model, which should be included as input to model updates. The OTM is calibrated to its base
1992 year. Model validation has been designed to include all steps of the model in a dynamic approach not restricted
to the base year only. The paper reports primarily on the validation for 2000, but 2003 and 2004 are also used to
validate the forecasting capability and sensitivity of the model. We also try to point out that, data permitting, the
validation design should focus on model objectives and the potential uses of the model such as forecasting
passengers in transit, which is particularly important in a country like Denmark where car taxation is extremely high.
Our hypothesis is that acceptance of a model depends on factors such as model type, objectives and year of
validation (base year, forecast or backcast). The validation criteria can therefore not be exactly formulated and the
validation results should be presented and discussed in a framework of policy requirements and experiences.
External data from the Danish national traffic survey were used to validate the passenger demand model. The transit
assignment results for separate modes and geographical areas were validated on the basis of well-known existing
statistical tools such as %RMSE and GEH statistics. A similar validation was conducted for cars as well, where
efforts were made to validate the model forecasts for the peak traffic.
Section 2 contains a short review of the OTM. Sections 3 and 4 are dedicated to the validation of the demand and
assignment models, respectively, of the OTM for 2000, while Section 5 deals with the model validation for period
2000 to 2004. Conclusions are summarised in Section 6.
2. MODEL STRUCTURE
2.1 Model analysis area
The OTM covers the Greater Copenhagen area with 1.7 million inhabitants. Figure 1 illustrates the part of
Copenhagen in the catchment area of the metro. It shows the road network in white, the rail network in grey, and the
metro network in black. A motorway runs across southern Amager (thick black line in the bottom of the figure),
extending east towards Malm, Sweden, and south towards south Sjlland.
Metro line 1 (M1) runs to a new town, restad in west Amager, while metro line 2 (M2) runs to Lergravsparken in
east Amager. The two lines meet on the Amager side of the harbour corridor (a sea channel between the islands of
Sjlland and Amager delimited by the Knippelsbro and Langebro bridges), just before Knippelsbro bridge where
the metro goes underground towards the island of Sjlland.
The analysis area is split into 601 internal zones while the surrounding area is defined by 17 port zones. The OTM
network includes some 13,000 nodes, of which 10,000 are bus nodes, 150 are train nodes, while there are 2,850 road
nodes. Apart from that, the transit network includes 1,200 lines, while the road network includes 4,100 road links.
Figure 1. The road, rail and metro networks in the city area covered by the metros alignment

2.2 Passenger demand model
The OTM predicts traffic demand for an average workday, omitting weekend traffic. The models nested logit
structure uses revealed preference (RP) and stated preference (SP) data in which the sub-models are connected via
the measure of accessibility (logsums). The model includes four segments: the business (BS), commuter (HW),
education (HE) and leisure (PR) segments and four travel modes: car, public transport, bicycle and walk. The BS
segment is trip based while the other three are tour based. In the OTM, a tour is defined as the sequence of a simple
trip from home to destination and a simple return trip from destination to home.
Input data to the demand model are zone data and files of level of service (LOS). The zone data describe the
distribution of the population, jobs, education places, shopping places, car ownership and parking costs in the 601
zones. Zone-to-zone travel times and travel costs for each available mode are represented in the LOS files. These
files are produced in the car and transit assignment models, and hence the LOS variables are exogenous to the
demand model. The distribution model is conditional on the generation model, and the mode choice model is
conditional on the distribution model (the three models previously described as demand sub-models).
Based on a number of traffic surveys at the end of the 1980s and the beginning of the 1990s in the Greater
Copenhagen area, 1992 travel matrices for walk, bicycle, car and public transport modes were built. They are 24-
hour tour matrices, which have been continuously improved in a number of traffic planning projects. The OTM
consistently uses base 1992 observed matrices 1992 being the models base year in a pivot point procedure
throughout the demand model as the last step of the sub-models in order to usefully exploit the available data
sources. For instance, the procedure is used in the generation model as follows. The number of tours for both the
forecasting year and the base 1992 year for each zone are calculated first (so-called synthetic matrices). The
forecasted generated tours by zone are then calculated as the ratio of the synthetic matrices multiplied by the
observed zone tours from the base year tour matrix.
Sixteen matrices of one days travel (four journey purposes by four travel modes) are produced after the execution
of the mode choice model. They are then split into three time-of-day matrices, namely matrices for the morning peak
period (7 a.m.-9 a.m.), afternoon peak period (3 p.m.-5 p.m.) and out-of-peak period (rest of the day) based on the
observed time split existing in the base 1992 matrices. The HW, HE and PR tour matrices are simultaneously
converted into trip matrices. The midpoint of a trip defines time of day. A detailed description of the OTMs
passenger demand model structure is given in Jovicic and Hansen (2003).
2.3 Assignment models
The OTM includes assignment models for all four modes, where walk and bicycle trips are assigned by an all-or-
nothing procedure based on travel times. In larger planning projects and long- term forecasts, the scenario is built by
creating a pseudo equilibrium situation where the demand and assignment models are run in a number of iterations
by a manually operated feedback loop, typically in three iterations.
A probit-based stochastic user equilibrium model is used in the car assignment model, applying the principles
developed by Danganzo and Sheffi (1977) and Sheffi and Powell (1982). In the OTM, the travel resistance at a link
level, U
a
, is given by:
( )
a a d, 3 a o, 2 a 1 a
t t l U ) 1 ( c | | | + + + + =
where l
a
is travel costs of link a, t
o,a
is free travel time, t
d,a
is delay, and | is a parameter. represents variation of the
cost coefficient while c
a
represents variation at link level. Delay is calculated as the difference between travel time
and free flow time.
In the OTM, c
a
follows the gamma distribution since it is reproductive in mean and variance (just like the normal
distribution) but has the advantage of being non-negative, i.e. it results in non-negative link costs (see Nielsen et al.
(2002) for more details).
In the transit assignment model, trips are categorised by transit mode (bus, train and metro) and route. It is a simple
assignment procedure based on a frequency-aggregated network using the following route costs V
r
:
r t, t r v, v r w, w r a, a r
t t t t V ) 2 ( o o o o + + + =
where t
a,r
is access/egress time using route r, t
w,r
is waiting time at first boarding, t
v,r
is in-vehicle time, and t
t,r
is
transfer time. Waiting times are calculated as half the headway aggregated across different but similar routes for the
same origin-destination zone pairs. To allow for the time it takes for travellers to adapt to timetables, waiting time at
first boarding is maximised to 6.5 minutes in the model. The parameters in the car and transit assignment models are
estimated using SP data.
The transit assignment procedure starts by selecting the applicable routes between each pair of zones. Extensive
calibrations based on 1992 data suggested that only the two best routes according to minimum values of formula (2)
should be used, otherwise inferior and non-used routes would enter the assignment. In the second phase of the transit
assignment procedure, the probability of using route re{1, 2} is calculated by the Kirchhoffs formula where = 2
(as found in the model calibration):

=
=
2
1 i
-
i
-
r
r
V
V
P ) 3 (


3. DEMAND MODEL VALIDATION FOR THE YEAR 2000
3.1 Generation model
The generation model in the OTM was estimated on the basis of a sample of 12,821 respondents from the 1997,
1998 and 1999 national travel surveys, from the sub-sample for the Greater Copenhagen area. The number of HW,
HE, PR and BS tours/trips made by each respondent per day was observed, including if no trips were made. The
model was then calibrated against the observed 1992 tour/trip matrices.
The OTM forecasts 3.55 trips per person per workday for the year 2000, half a trip more than observed in the Danish
national travel survey 2000 (Danish Transport Research Institute, 2003) in the sub-sample for the Greater
Copenhagen area (GCA). Table 1 compares the trip rates observed in the 2000 national travel survey and those
forecasted by the model for an average workday, split between three travel purposes: home-work/education, home-
leisure and business. The models home-work and home-education travel purposes had to be combined into one
segment because the available travel behaviour data do not distinguish between these travel purposes. The table
shows that the disagreement between the totals is entirely explained by the difference in trip rates for home-
work/education trips. The later work has showed that the 2000 national travel survey underestimates the trip rate for
this travel purpose.
Table 1. Trip rates from the Danish national traffic survey and the model forecasts by travel purposes on an average
workday in 2000, in number of trips per person per workday
National traffic survey,
2000
OTM
Home-work/education trips 0.98 1.42
Home-leisure trips 1.52 1.59
Business trips 0.56 0.54
Total trips per person per day 3.06 3.55
3.2 Distribution model
In the OTM, the choice of destination depends on zone attractiveness, distance and logsums from the mode choice
model (Jovicic and Hansen, 2003). According to the national survey, residents of Copenhagen travelled 29 km per
workday in 2000. The model forecast for person-km by travel mode reveals a very good overlap with the observed
mode for person-km, as shown in Table 2.
Dividing the total person-km (presented in Table 2) by the person trip rate (presented in Table 1), we observe that
the average observed trip length in 2000 was 9.4 km while the modelled trip length was only 8.2 km. Using a pivot
point procedure, we see that the trip pattern in the OTM is strongly influenced by the 1992 trip matrices, which were
estimated by matrix estimation procedures based on traffic counts (e.g. Nielsen, 1997). Compared with the 2000
national survey, it seems that these procedures tend to split long trips into two or more sub-trips, thus resulting in
higher trip rates and shorter trips. It is also possible that the survey respondents tended to forget short trips. For
instance, in the survey, a work-shopping-home trip chain may be categorised as one long trip from work to home,
whereas in the 1992 matrices it is split into two trips.
Table 2. Person-km from the Danish national traffic survey and the model forecasts for four travel modes on an
average workday in 2000, in km
National traffic survey,
2000
OTM
Private car 19.43 20.02
Transit modes 7.20 7.01
Bicycle 1.82 1.69
Walking 0.42 0.46
Total person-km per day 28.87 29.18

In the AKTA
1
project, 280 car drivers reported that home-work trips by car are 13.6 km long on average. In the
OTM, the forecasted total average car driving distance for home-work and home-education purposes is 11.3 km.
Again, this supports the conclusion that the applied matrix estimation procedures may tend to produce too many
short trips.
3.3 Mode choice model
The model forecast for modal shares for the year 2000 harmonises well with the shares observed in the national
travel behaviour survey for the same year, as shown in Table 3. The predicted and observed shares are particularly
comparable for transit and car modes, for which the OTM includes numerous travel time components. In all, forty
different values of time (VOT) were estimated in the OTMs mode choice model (a full discussion of the obtained
values is provided in Jovicic and Hansen, 2003).
Table 3. Modal shares from the Danish national traffic survey and the model forecasts for four travel modes on an
average working day in 2000

1
AKTA (www.akta-kbh.dk) is a research study under the EU project PROGRESS (www.progress-project.org),
which is part of the EUs 5
th
framework programme The Growth Programme on Sustainable Mobility and
Intermodality (Nielsen and Jovicic, 2003).
National traffic survey,
2000
OTM
Private car 49% 48%
Transit modes 15% 14%
Bicycle 21% 19%
Walking 15% 19%
Total modal shares 100% 100%

It is probably reasonable to state that the OTM overestimates the modal share for walking at the expense of cycling.
The trade-off between these two modes has never been investigated in the across-mode SP experiments. Therefore,
in the model, many of the modal shares for walking and bicycling were explained by their alternative specific
constants. Also, the 1992 matrices used in the pivot point procedure for walking and bicycle trips were inaccurate
(particularly in suburban and rural areas) due to few existing counts for these two modes.
3.4 Demand elasticities
The model has been tested for changes in demand when travel costs and travel times for car and transit modes
change. Table 4 shows the calculated direct and cross-trip elasticities aggregated across the home-work, home-
education and leisure travel purpose segments. Note that the values below do not take the business segment into
account because this segment shows almost no change in demand for tested changes in travel supply. It should also
be noted that the elasticities in the table do not include the feedback effects from the assignment model caused by
network changes, e.g. an increase in car travel time is directly coded into the demand model utilities.
Changes in travel cost and travel time resulted in much higher direct elasticities for transit modes than for private
cars. This is most likely due to differences in income between these two groups of travellers, as transit travellers
tend to have lower incomes than car users. The sensitivity analysis also showed that the direct car elasticity for
driving time is higher than for driving costs, while the opposite is true for transit modes. Copenhagen car drivers are
probably not fully aware of the cost of driving, because they generally only relate the cost of driving to petrol prices.
Thus, the cost-of-driving factor is less important because car drivers have a rather vague idea of the actual cost of
driving, a conclusion also reached in the AKTA project. Driving time, on the other hand, was perceived more
realistically and thus considered more important (an increase in congested travel time was particularly unwelcome).
Transit travellers focus keenly on fare due to direct out-of-pocket costs, making their perception of travel costs fairly
accurate.
Table 4. Demand model trip elasticities, aggregated across travel purpose segments
Car Walking Bicycle Transit
Car driving costs increased -0.11 +0.06 +0.06 +0.07
Car driving time increased -0.16 +0.12 +0.09 +0.11
Transit fare increased +0.06 +0.05 +0.11 -0.35
Transit time increased +0.06 +0.04 +0.08 -0.30

The following discussion of each of the four sensitivity tests (the four rows in Table 4) refers to specific trip
purposes where appropriate.
A 10% increase in driving costs resulted in a 1.1% decrease in the number of car trips. The education segment was
strongly influenced by cost increase (elasticity of 0.63), which can be explained by the low income of travellers in
this group. The commuter segment had a numerically higher elasticity (0.096) than the leisure segment (0.070).
This is probably due to differences in car occupancy for the two travel purposes, since car occupancy for leisure
trips is higher than for home-work trips.
A 10% increase in driving time resulted in a 1.6% decrease in the number of car trips. Again, the education segment
was strongly influenced by increase in driving time (elasticity of 0.48). In this case, the commuter and leisure
segments had identical elasticities of 0.15.
Students proved extremely sensitive to changes in transit fare, with a 10% increase resulting in a 13% decrease in
transit trips in this segment. The sensitivities of commuters and persons travelling for leisure were 0.29 and 0.26,
respectively. The average direct transit elasticity for travel costs was 0.35.
A 10% increase in transit travel time led to a 5.4% drop in transit demand related to education travel. According to
the OTM demand model, a 10% increase in transit travel time would lead to a 3.6% switch in commuting transit
trips to other modes, while 2% of leisure transit trips would switch to other modes or not be done. The average direct
transit elasticity for travel time is 0.30.
An EU report on car travel elasticities, TRACE (Jong at al., 1998), showed some similarities with the OTMs car
elasticities. The project reported that in the Netherlands, France, the UK and Sweden, the direct car elasticity for
driving time was numerically higher than for driving costs. However, it seems that the OTMs elasticities in absolute
terms are generally lower than the other EU countries elasticities. For instance, Swedish car travel cost elasticity
was reported to be
0.14 while car travel time elasticity was 0.32. One explanation is that Danish car ownership is very low compared
with many other EU countries, owing to the high taxes and duties levied on cars.
4. ASSIGNMENT MODEL VALIDATION FOR YEAR 2000
4.1 Validation methods
Since the assignment model is the last element of the modelling process, its output can be considered to reflect the
performance of the model as a whole, and most model validation efforts have traditionally focused on generating
accurate link volumes.
Barton-Aschman Associates and Cambridge Systematics (1997) suggest three measurements for validating link
volumes:
- Comparing observed and estimated volumes by screen line where a reference of 5% divergence is targeted.
- Comparing observed versus estimated volumes for all links with counts. To compute aggregate statistics for
validating traffic assignment results, the Percent Root Mean Square of the Error (%RMSE) is proposed
according to (4):
( )

=
=

=
n
1 i
i
n
1 i
2
i i
Obs
n
1
1 n
Obs - X
%RMSE ) 4 (
where X
i
is assigned traffic and Obs
i
observed traffic at link i. Theoretically, the nominator in (4) consists of
the standard deviation and bias, since (X
i
- Obs
i
) = ((X
i
- X ) + ( X - Obs
i
)). The denominator is simply the
average of observed values.
- Comparing the R
2
(Coefficient of Determination) region-wide traffic counts and estimated values. Barton-
Aschman Associates and Cambridge Systematics (1997) suggest that R
2
should be greater than 0.88.
Additionally, a scattergram of the counts versus assigned volumes should be drawn and points lying outside a
reasonable boundary of a 45

degree line are recommended for review.
The UK Department of Transport (DoT) (1996) offers two methods for comparing modelled values against
observations:
- Calculating the GEH statistic according to (5):
( )
( )
( )
2
Obs X
Obs - X
2
Obs X
Obs - X
GEH ) 5 (
i i
i i
i i
2
i i
+
=
+
=
The GEH values can be calculated either for individual links or for a screen line. It is assumed that, being a
special case of the more general gamma distribution G(k, ) with k = and = 2, the GEH statistic follows a
chi-squared distribution
2
(1). For individual flows, the UK DoT requires that the GEH statistic be less than 5,
corresponding to
2
97.5%
(1), and for screen lines with more than five links, the GEH statistic should be less than
4, corresponding to
2
95%
(1) in 85% of cases.
- Plotting modelled values against observed values and carrying out linear regression analysis. The Coefficient
of Determination R
2
of the best-fit regression line through the origin should be above 0.90 and the slope
between 0.9 and 1.10.
The %RMSE (4) is dominated by discrepancies on high-volume links and, therefore, it seems to work best on
samples of similarly sized road stretches. Since (X
i
+ Obs
i
) Obs
i
, the estimated volume X
i
in the GEH statistic
(5) approximately follows a gamma distribution:
|
.
|

\
|
e
|
.
|

\
|
e
i i i
i
i i
Obs ,2
2
1
Obs - X
,2
2
1
Obs
Obs - X
) 6 (
G
G

However, a normal distribution of estimated link volumes is more usually assumed, and the weaknesses of GEH and
%RMSE will be discussed and assessed in the validations below. The methods are applied in the model validation of
transit assignment (Section 4.2), car assignment (Section 4.3), and slow mode assignment (Section 4.4).
4.2 Validation of transit assignment
The bus company in the GCA performs systematic, periodical, automatic, onboard bus counts. The Danish State
Railway counts train passengers once a year, on the Tuesday in the first week of November. However, no transit
data are available for passenger km and passenger hours.
For the model validation, 103 bus passenger link counts and 29 train passenger link counts are available for an
average workday in 2000. The distribution of bus links is as follows: 52 in the downtown area, 33 in suburban areas,
and 18 in rural areas. The corresponding figures for train links are 17, 7 and 3, respectively. The hourly volumes,
which were not available to this study, were approximated by dividing the daily volumes by 32 (16 daytime hours,
from 6 a.m. to 10 p.m., multiplied by two directions).
Table 5 shows the statistical validation of bus and train passenger assignment. For bus passengers, the poor %RMSE
range from 29% to 64% is due to rather low bus passenger volumes, particularly in rural areas. The %RMSE for
train passengers is much lower due to accurately estimated high traffic volumes.
Table 5. Statistical validation of assigned bus passengers in 2000
Area Bus Train
Observations %RMSE GEH < 5 Observations %RMSE GEH < 5
Downtown 52 38.2% 61.5% 19 13.5% 63.2%
Suburban 33 28.9% 87.9% 7 19.8% 71.4%
Rural 18 64.2% 83.3% 3 6.3% 100.0%
Total 103 43.2% 73.8% 29 14.9% 69.0%

The GEH statistic in the table is less than 5 in 74% of the 103 bus links and does not comply with the 85%
requirement, which is only achieved for links in the suburban area. The train assignment also falls short of the GEH
statistic threshold.
The figures in Table 5 show that the performance of the transit assignment model is unacceptable according to the
boundaries set by Barton-Aschman Associates and Cambridge Systematics (1997), and the UK Department of
Transport (1996). The validation procedures are primarily developed on the basis of car traffic. Since car assignment
performance is usually higher than the more complex transit assignment, it seems appropriate to reduce threshold
requirements when applied in validation of transit assignment. However, we conclude that the accuracy in assigning
downtown transit users is inadequate at the link level.
Figure 2 plots modelled bus passenger volumes against observed values, and Table 6 shows the best fit of a linear
regression through the origin. According to these sources, the bus passenger flows on links in the downtown area are
overestimated while those on links in rural areas are underestimated. The Coefficient of Determination R
2
is low in
rural areas, and the line of regression is therefore not a good estimate.
The same procedure as above is used for train passenger assignment in Figure 3 and Table 6. The accuracy of
assigned train passengers is better than for bus passengers, although there appears to be some underestimation in
downtown and suburban areas. The underestimation of train passengers in these two areas is probably due to an
incorrect split between bus and train caused by the simplicity of the transit assignment procedure, e.g. comfort and
timetables are not considered.
Figure 2. Assigned versus observed bus passengers

Figure 3. Assigned versus observed train passengers
0
200
400
600
800
1000
1200
1400
0 200 400 600 800 1000 1200 1400
Suburan
45 degree line
Rural
Downtown
Observed hourly bus
Estimated hourly bus

Table 6. Linear regression analysis through the origin based on estimated versus observed directional hourly bus
passenger volume
Area Bus Train

Observations Slope R
2
Observations Slope R
2
Downtown 52 1.142 0.889 19 0.931 0.970
Suburban 33 1.019 0.857 7 0.854 0.974
Rural 18 0.669 0.572 3 1.047 0.999
Total 103 1.131 0.920 29 0.922 0.974

4.3 Validation of car assignment
The Danish Road Directorate and the Copenhagen Municipality conduct annual vehicle traffic counts on predefined
road links. To validate the car traffic model on an average workday in 2000, 395 road link statistics were used, and
the estimated directional peak hour flows were compared with 178 (8-9 am) and 179 (4-5 pm) road links
automatically counted and averaged over 2000. Manual one-day counts were only used in the validation of average
workday flows due to considerable uncertainty in the observed peak volumes.
0
500
1000
1500
2000
2500
3000
3500
4000
0 500 1000 1500 2000 2500 3000 3500 4000
Suburban
45 degree line
Downtown
Rural
Estimated hourly rail passengers
Observed hourly rail
In Figure 4, the modelled daily volumes are plotted against observed values illustrated with reasonable boundaries.
Since the best fit of a linear regression through the origin results in a slope of 0.99 and R
2
= 0.95, the guidelines for
slope and Coefficient of Determination recommended in Section 4.1 are met, indicating no apparent over- or
underestimation.
In Figure 5, a similar procedure has been used to illustrate directional peak hour flows (8-9 a.m. bullet and 4-5 p.m.
cross).




Figure 4. Assigned versus observed daily car traffic volume

Figure 5. Assigned versus observed directional peak hour car traffic volume
0
10000
20000
30000
40000
50000
60000
70000
80000
90000
100000
0 10000 20000 30000 40000 50000 60000 70000 80000 90000 100000
Observed daily hour flow
Estimated daily flow

The model underestimates the large morning and afternoon peak hour volumes resulting in a best-fit regression
slope of 0.83 (R
2
= 0.87). The low peak shares in the observed pivot point matrices of 1992, indicating a
proportionately higher increase of commuters on motorways over the period 1992 to 2000, may explain why the
model does not meet the requirements of Section 4.1. Another likely explanation is a faster decline in car occupancy
in long-distance commuter traffic relative to other travel purposes.
In Table 7, the GEH statistic is less than 5 in 51% of the 178 road links and 40% of the 179 road links, which is
unacceptable according to the required 85% threshold. The %RMSE test also shows significant divergences.
Table 7. Statistical validation of assigned car traffic in 2000 by road link
Area Observations
8-9 a.m.
GEH < 5
8-9 a.m.
%RMSE
8-9 a.m.
Observations
4-5 p.m.
GEH < 5
4-5 p.m.
%RMSE
4-5 p.m.
Central
municipalities
13 54% 22% 13 39% 33%
Copenhagen County 61 49% 32% 60 45% 40%
Frederiksborg County 80 56% 38% 80 39% 42%
Roskilde County 24 38% 39% 26 31% 49%
Total 178 51% 38% 179 40% 47%

0
500
1000
1500
2000
2500
3000
3500
4000
4500
5000
5500
6000
6500
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 5500 6000 6500
8-9 am
45 degree line
4-5 pm
Observed directional peak
Estimated directional peak
Table 8 shows the GEH statistics and relative divergences for a set of 16 screen lines across the GCA. Only a small
share of screen lines meet the requirements of GEH <4 and show a divergence less than 5%. The relatively poor
results may primarily be explained by the poor basis of peak hour data as mentioned above. However, it should be
stressed that we would expect less accurate results in forecasting year 2000 than in base year 1992. The results also
reflect some of the weaknesses of %RMSE and GEH, since %RMSE is dominated by discrepancies on high-volume
links (Figure 5) and GEH boundaries seem most likely achievable at low volumes.
Finally, we used the GEH statistics by converting daily volumes into average directional hourly day-time flows
similar to Section 4.2. This statistic implies that 81% to 88% of the road links in the three counties of the GCA
achieve values below 5. Only the estimated flows in central municipalities seem not to meet the requirements.
At 21% of daily flows, the %RMSE reflects some accurately estimated high traffic volumes in suburban areas (13%)
and less accurate estimates in the central districts (25%).


Table 8. Statistical validation of assigned car traffic in 2000 by screen line
Screen line Observations
8-9 a.m.
GEH
8-9 a.m.
Divergence
8-9 a.m.
Observations
4-5 p.m.
GEH
4-5 p.m.
Divergence
4-5 p.m.
1 5 0.2 -0.2% 5 6.3 8.8%
2 6 5.7 8.9% 6 11.6 18.6%
3 21 4.6 -2.9% 21 4.8 3.0%
4 4 10.1 -21.0% 4 13.4 -24.9%
5 9 43.5 -28.7% 10 58.0 -33.4%
6 26 12.6 -8.9% 26 13.3 -9.3%
7 3 9.1 26.0% 3 5.8 -11.8%
8 13 13.8 -8.0% 10 2.5 -1.9%
9 9 19.2 -15.6% 9 16.8 17.6%
10 3 20.5 -23.7% 4 7.9 -8.8%
11 21 7.3 -8.0% 21 19.7 -18.5%
12 10 6.1 5.1% 10 11.9 -8.7%
13 8 3.2 4.4% 10 13.4 -14.5%
14 16 4.8 8.4% 16 7.5 -10.8%
15 4 4.6 -28.3% 4 10.7 -45.5%
16 20 11.4 -11.5% 20 12.8 -12.4%

4.4 Validation of bicycle and walk assignment results
The OTMs assignment model for slow modes (cycling and walking) is an all-or-nothing model based on the
shortest travel time. Little attention was paid to these two modes in the model estimations predominantly due to the
lack of observed data. Even in the SP experiments completed under the OTM, walking and cycling are presented
only in the context of across-mode experiments, i.e. no within-mode SP experiments were completed for the two
modes.
Table 9 shows available observed bicycle and walk trips across Knippelsbro and Langebro bridges (the harbour
corridor in the city area) in 2000 compared with estimated volumes.

Table 9. Observed and estimated daily flows by bicycle and walk across Knippelsbro and Langebro bridges, 2000
Observed Estimated Divergence (%)

Knippelsbro, bicycle 23,465 24,097 632 (2.7%)
Langebro, bicycle 21,912 27,040 5,128 (23.4%)
Bicycle, total 45,377 51,137 5,760 (12.6%)
Knippelsbro, walk 4,889 1,898 -2,001 (-61.2%)
Langebro, walk 2,164 2,034 -130 (-6.0%)
Walk, total 7,053 3,932 -3,121 (-44.2%)
Slow modes, total 52,430 55,069 2,639 (5.0%)

The figures in the table show that bicycle traffic over the harbour corridor is overestimated by some 13%, where the
vast majority of missing traffic is related to the Langebro bridge. The volume of observed walk traffic in the corridor
is about 15% of bicycle traffic. The corridor walk traffic is largely underestimated by the OTM. This time the
mismatch occurs on the Knippelsbro bridge. In total, the model overestimates the slow modes traffic in the Harbour
corridor by 5%.
The doubtful modelling of slow modes is a serious shortcoming of the OTM because bicycling, in particular, is an
important mode in Copenhagen and, up to certain distances, a popular alternative to public transport modes.
5. MODEL VALIDATION FOR 2000-2004
5.1 Non-metro modes
Table 10 shows the relative change for observed versus modelled traffic for car, train, bus and bicycle modes in the
GCA for the period 2000-2004. Note that train traffic includes trips by the Copenhagen city train (S-train) and
regional trains within the GCA.

Table 10. Modelled versus observed traffic changes in the period 2000-2004
Modes Observed OTM forecast
Car +4% +6%
Copenhagen city train (S-train) -2% +11%
Regional train +1% +10%
Bus -21% -6%
Bicycle +6% +2%

According to the table, the OTM forecasts are higher than the observed traffic changes for car, train and bus modes.
The countrys stagnant economy in this period provides a possible explanation, i.e. the GDP figures remained almost
unchanged after 2000, while GDP had grown by almost 2% annually prior to this period (Danish Statistical Bureau,
2004). The train and metro have experienced a long period of unreliable operations, which may also have negatively
affected ridership. Finally, the observed increase in transit fares is not taken into account in the model, where the
travel costs are kept constant by the scenario definitions specified by the client.
5.2 Assignment model for the metro
Phase 1 of the metro opened in October 2002. In 2003, screen line counts were conducted once a week, rotating
between weekdays. In most cases, counts were located on the stretch between Nrreport and Kongens Nytorv
stations (sample counts), and metro passenger estimates were computed from the sample counts. Total counts (all
metro stations) of boarding passengers have only been conducted a couple of times to date owing to the high costs
associated with conducting this type of count.
Following an initial running-in process, the metro became fully operational in mid-January 2003. As the summer
and autumn of 2003 were characterised by the opening of new metro phases and unstable operation, model
validation is only possible on the basis of counts from the first months of 2003. Table 11 shows available metro
passenger counts for February and March 2003. A total count was conducted on 6 February 2003 when 50,354
metro passengers were counted. On average, 55,000 metro passengers travelled on a workday in the counting period.
That is 5,900 passengers fewer, or 10.7%, than the traffic flow predicted by the OTM.
Table 11. Metro passenger counts
Number of passengers
February 2003 56,054
March 2003 53,795
Average workday 54,925

The general overprediction by the model was mainly caused by incorrect forecasts at Lergravsparken and Island
Brygge metro stations. Lergravsparken is the terminus of the line M2, and until the second week of January 2003,
headways were 12 minutes in peak periods and 16 minutes in out-of-peak periods, which is considerably worse than
the pre-existing bus service frequency. After the second week of January 2003, headway improved to 6 minutes in
peak periods and 8 minutes in out-of-peak periods. Thus, the disappointing metro service had a longer- term effect.
The Island Brygge metro station is the station closest to the University of Copenhagen, Amager, and the model
predicted that many students would shift to the metro here. This seems not to have been the case. The overprediction
of metro flows is also related to the fact that the model overpredicts bus and train traffic (see Table 10). Passengers
using these two modes were expected to transfer many of their trips to the metro (Vuk, 2005).
In 2003, the metro was extended to Vanlse station in Phase 2. In early 2004 a new interchange station was opened
at Flintholm, connecting the metro to the existing S-train network and a new S-train line around Copenhagen.
During the first half of 2004, the metro service continued to be rather unreliable but performance improved
significantly in fall 2004, and the number of passengers rose.
The metro boarding and alighting passenger counts are now automatically conducted at each station. Due to poor
service performance and the gradual extension of the metro system, it is difficult to compare the observed and
predicted metro passenger flows. In March 2005, the metro recorded 130,000 passengers in an average workday
compared with a forecast of 198,000 passengers. The high overprediction can largely be attributed to four main
factors. First, the negative transit trend in the period 2000 to 2004 (Table 10) had a substantial impact on the number
of metro passengers. Second, the unreliable metro service delays the process of adapting to the new system. Third, it
seems that overpredictions primarily relate to phase 2 of the metro (the connection of Vanlse station to Forum
station), implying an inaccurate data foundation. Fourth, the bus counts at certain links show that many bus
travellers still have not switched to the metro, as predicted by the model forecasts.
6. CONCLUSIONS
The paper validated the OTM, the passenger traffic model for Copenhagen, for the years 2000 to 2004. The
validation procedure for the year 2000 is two-pronged. First, the observed travel behaviour from the 2000 national
travel survey, which is external to the model, was compared with the results of the generation, distribution and
modal split models. Second, the car and transit assignment forecasts were tested against the traffic counted in 2000.
With respect to the validation of the demand sub-models, we conclude that a tour-based demand model structure of
the OTM, in which there is logsum feedback between the sub-models, produces satisfactory results for the following
two main reasons. First, the modelled 2000 trip rates, as forecast by the generation model, closely approximated
observed trip rates at both national and international levels. However, the forecast trip rate for home-work/education
trips was somewhat higher than the observed 2000 trip rate. Because the OTM applies a pivot point procedure in
each of the sub-models, the trip pattern is strongly influenced by the base 1992 trip matrices. In conclusion, it seems
that that the models pivot point procedure has a tendency to split long home-work/education trips into two or more
sub-trips, thus resulting in higher trip rates and shorter trips than observed.
Second, the differences between the forecast and observed modal split for the year 2000 were small. While the
modal shares for car and transit modes were almost identical in the national travel survey and the OTM, the
differences were slightly greater for bicycle and walk modes. The reason is that the trade-off between these two
modes has never been investigated in the walk-bicycle across-mode SP experiments. The modal shares for these two
travel modes were therefore largely explained in the OTM by their alternative specific constants.
The sensitivity analysis showed that car cost direct demand elasticity was considerably lower than transit fare direct
elasticity, which can be explained by the difference in income between the two groups of travellers.
With respect to the validation of the assignment forecasts for the year 2000, the overall conclusion was that the car
assignment model performed better than the transit assignment model. The car assignment model is fairly complex
in structure while the transit assignment model uses a simple simultaneous mode/route procedure.
The car assignment model met the accuracy threshold values of the applied statistics with respect to 24-hour traffic.
However, the analysis of directional peak hour car traffic shows greater discrepancies between the observed traffic
and the modelled traffic at 16 screen lines in the analysis area, where the model underestimated observed traffic. The
forecast car peak traffic in the OTM is calculated on the basis of the observed percentage shares of the peak traffic in
the 1992 base matrices, which seem inaccurate and outdated.
The overall performance of the transit assignment model was not entirely satisfactory, particularly for the
assignment of bus trips. The statistical test concerning the correspondence of the model results with observed values
showed large deviations in the downtown area.
The model seems to overpredict metro passenger flows by 11% in early 2003. In March 2005, the model
overpredicted metro passenger flows by 50% due to factors such as the negative transit trend in Copenhagen over
the period 2000 to 2004 and unreliable metro services.
The efforts to validate the OTM revealed a need to update the base 1992 matrices if the pivot point procedure is to
be maintained. The 1992 matrices include outdated flows and inappropriate peak shares. The validation conducted
has also taught us that future modelling work should pay more attention to the importance of regularity and the
period of adapting to the new travel mode in order to provide ideas for improving the transit assignment model (e.g.
a time-table based assignment) and establishing a model for choice of time of day. In this area, an activity-based
approach to modelling travel demand might be necessary. An update of the models trip matrices was initiated in a
project started in January 2005 while the model update itself is currently being discussed by researchers and the
Danish Ministry of Transport.
We wish to stress the importance of conducting model validation, often neglected in practical model development.
Validation should include three main tasks: an assessment of the coefficients used in model estimation, a test of the
model system and sensitivity, and a validation of results versus observed behaviour. We recommend that the design
of the validation procedure should already be addressed in the planning phase of model development and should
focus on:
- The importance of validating all model parts and not only the assignment models, as is usually the case.
- According equal weight when validating the model for the base year and forecasting/backcasting years in a
dynamic procedure.
- The importance of validating transit traffic and slow modes (bicycle and walk) relative to the objectives of
urban traffic models.
- Resources (budget) for assembling appropriate validation data.
Finally, procedures and a methodological framework for model validation need to be further developed. In the
paper, we applied the usual procedure for assignment validation, including GEH and %RMSE statistics. Those tests
miss some basic features such as statistical foundation and assessment of biases due to the numerical formation and
they cannot, therefore, be used without supplementary validations.
REFERENCES
Barton-Aschman Associates and Cambridge Systematics (1997). Model Validation and Reasonableness Checking
Manual. Travel Model Improvement Program, U.S. Department of Transportation.
Daganzo, C.F. Sheffi, Y. (1977). On Stochastic Models of Traffic Assignment. Transportation Science 11 (3), pp.
253-274.
Danish Statistical Bureau (2004). A databank available at the Danish Statistical Bureaus website
(http://www.statistikbanken.dk/statbank5a).
Danish Transport Research Institute (2003). Danish National Travel Survey 2000 Summary tables. www.dtf.dk.
Jovicic, G. and Hansen, C.O. (2003). A Passenger Travel Demand Model for Copenhagen. Transportation Research,
Part A 37, pp. 333-349.
Jong, G.C. and Tegge, O. (1998). TRACE, Cost of private road travel and their effect on demand, including short
and long term elasticities. European Commission Directorate-General for Transport.
(www.rand.org/randeurope/hcgprojects/trace/documents.htm).
Nielsen, O.A. (1997). On the Distribution of the Stochastic Component in SUE Traffic Assignment Models.
European Transport Forum (PTRC Annual Meeting). Seminar F, Transport Planning Methods vol. II, pp. 77-94,
Uxbridge, UK.
Nielsen, O.A., Frederiksen, R.D. and Daly, A. (2002). A Stochastic Route Choice Modsel for Car Travellers in the
Copenhagen Region. Networks and Spatial Economics no. 2 pp. 327-346, Kluwer.
Nielsen, O.A., and Jovicic, G. (2003). Sensitivity of variable definitions in SP analyses- An empirical study of car-
users evaluation of length cost and time components. 10
th
International Association for Travel Behaviour
Research Conference. Switzerland.
Sheffi, Y. and Powell, W.B (1982). An algorithm for the Equlibrium Assignment Problem with Random Link Times.
Networks vol. 122 pp. 191-207.
UK Department of Transport (1996), Traffic Appraisal in Urban Areas, volume 12 Section 1 of the Design Manual
for Roads and bridges (DMRB v12s1).
Vuk, G. (2005). Transport impacts of the Copenhagen Metro. Journal of Transport Geography, Volume 13, pp. 223-
233.

Potrebbero piacerti anche