Download as pdf or txt
Download as pdf or txt
You are on page 1of 19

T,,,,q,n. R;I.-B. Vol. 236. No. I. pp. 29-47. 1989 0191.2615189 13.00 + .

oO
Primed m Great Bnrain. 0 1989 Pergamon Press plc

MACROSCOPIC MODELLING OF TRAFFIC FLOW


ON THE BOULEVARD PfiRIPHgRIQUE IN PARIS

MARKOS PAPAGEORGIOU, JEAN-MARC BLOSSEVILLE


and HABIB HADJ-SALEM
Institut National de Recherche sur les Transports et leur S&wit6 (INRETS),
Dkpartement Analyse et Regulation du Traffic (DART), 2, Av. du GI. Maileret-Joinville,
94114 Arcueil, Cedex, France

(Received 14 December 1987)

Abstract--Three macroscopicmathematicalmodels are presented and compared on the basis


of real traffic data collected from the Boulevard Pkriphtrique in Paris. All models include the
conservation equation and a nonlinear, static, or dynamic volume-density relationship. Param-
eter estimation is effectuated for each model so as to reflect the characteristics of the particular
application case considered. Sensitivity investigations for changed application conditions are
included.

1. INTRODUCTION

Availability of adequate macroscopic mathematical models of freeway traffic is a pre-


requisite for the development of efficient control strategies (see Papageorgiou, 1983).
More precisely, mathematical models of sufficient accuracy are required for: traffic data
processing and estimation, design of control strategies, and testing of control strategies
via simulation.
For the Boulevard PCriphCrique de Paris (BP) it is intended to develop and compare
a number of on-ramp metering strategies with different degrees of sophistication. In a
first phase, implementation and comparison of several control strategies has been ef-
fectuated for a single on-ramp (Brancion) in real-life. Results of the first phase being
judged very encouraging, a second experimentation phase has been started that incor-
porates ramp metering at three on-ramps included in a southern part of BP with a total
length of 6 km. For this part of BP it is intended to develop traffic responsive overall
control strategies that coordinate the control actions in the individual ramps so as to
achieve overall satisfactory traffic conditions (see Papageorgiou, 1986). Performance of
different overall control strategies with different degrees of sophistication will be com-
pared to each other and to local control performance.
In this context, a macroscopic mathematical model has been developed for the
considered part of BP along with parameter estimation on the basis of real traffic data,
and performance investigation under changing application conditions. The present paper
is an extract of the corresponding detailed results reported in Papageorgiou (1987).

2. MACROSCOPIC MODELS OF FREEWAY TRAFFIC

This section presents three macroscopic models of freeway traffic following the
nomenclature of Papageorgiou (1983). Each model includes

(i) the conservation equation and


(ii) a (static or dynamic) nonlinear relationship between mean speed and traffic
density.

The models are formulated in a continuous space-time framework before being


discretized in space and time. In view of the approximate, empirical nature of (ii),
however, it does not appear necessary to approximate the continuous solutions with
highest accuracy using fine discretisation. In fact, we focus on the discrete versions of
the models, and it is for these versions that parameter estimation on the basis of real
29
30 M. PAPAGEORGIOU
et al.
traffic data is effectuated. Nevertheless, caution is necessary for one of the three models
(model B below) in order to avoid structural instability of the discrete version.
The desired space-time discretization intervals are in the order of 500 m resp. 10 s.
With these values, computational effort is low both for simulation and for control strategy
design. However, available empirical material did not permit time-space intervals to be
longer than 125 m resp. 2.5 s. Hence, some complementary investigations will be nec-
essary in the near future.

The fundamental diagram


Macroscopic description of traffic flow implies the definition of adequate flow vari-
ables expressing the average behaviour of the vehicles at a specific location and time
instant. We define the traffic density p(x, f) as the number of vehicles per length unit
(veh/km), the space mean speed u(x, t) (km/h) as the instantaneous average speed of
vehicles in a length increment (see Wardrop, 1952), and the traffic volume (17(x,r) as
the number of vehicles passing a specific location in a time unit (veh/h).
As a direct consequence of the continuum variables’ definition, we get the rela-
tionship

4 = P’fJ, (1)

which has a direct analogon in hydromechanics.


For reasons of safety, distances between vehicles are adjusted to the vehicles’ ve-
locities and vice versa. Hence, increasing density in a given freeway stretch. correspond-
ing to shorter headways, leads to reduced mean speed. Several mathematical formulae
have been proposed for the description of a speed-density relationship under homo-
geneous conditions (see, e.g. May and Keller, 1967). A fairly general formula V(p)
satisfying appropriate boundary conditions reads

V(P) = q[l - (P~P,am)‘l’n. (2)

where I > 0 and m > I are real valued parameters. Several V(p) formulae proposed in
the literature can be considered as special cases of (2) by appropriate choice of the
parameters m and 1. For example, the formula

V(p) = u,. exp


[
$ 1
(P/P2

results from (2) for 1 = a, pjam = per - (a - m)““, and m * =. A polynomial formula
sometimes utilized reads

V(p) = i &Pi, (4)


i=O

where a;, i = 0, . . . , p, are constant parameters.


Substituting V(p) into (l), we get a volume-density relationship Q(p) that is broadly
known as the fundamental diagram of traffic engineering. As it is well known, the Q(p)
relationship provides maximum volume qmaxat a critical density pCr.

The conservation equation


Let us consider traffic on a long multilane freeway with several off-ramps and on-
ramps. We are interested in developing mathematical models of traffic flow describing
the dynamic evolution of traffic variables along the freeway.
Traffic flow on the Boulevard Piripherique in Paris 31

Regarding traffic flow as a fluid with density p(x, t) and volume 4(x, t), we may
write the fundamental equation of conservation of matter

ap a4
dt+dx=r-S’
where r - s is the exogeneous on-ramp/off-ramp source term. For our purposes, it is
more convenient to consider difference equations. For this reason, we subdivide the
freeway into a number N of sections with lengths A,, i = 1, . . . N, each having at most
one on-ramp and one off-ramp. Furthermore, T being the sample time interval, we
introduce space-time-discretized traffic variables (Fig. 1) as follows:

p,(k) = number of vehicles in the freeway section i at time k . T divided by the length A, of the section
u,(k) = space mean speed of vehicles in the freeway section i at time k * T
q,(k) = number of vehicles leaving section i during the time period [kT, (k + l)T], divided by T
r,(k), s,(k) = on-ramp, off-ramp volumes of section i (if any).

With these variables, a space-time discretized form of eqn (5) can be given as
follows:

P,(k + 1) = P,(k) + $, [qi-l(k) - q,(k) + r,(k) - Si(k)l.


I

It is obvious that eqns (5) and (6) hold exactly. The conservation equation can be
expanded to become a complete model of traffic flow if traffic volumes are expressed
in terms of traffic densities in (5), resp._(6). This can be accomplished in different ways
as will be shown in the next three sections.

An algebraic volume-density relationship


Assuming the fundamental diagram to be valid even under nonhomogeneous con-
ditions, we have

dx, t> = Q[P(x, t)l. (7)


In view of (l), eqn (7) corresponds to the assumption

u(x,t> = V[P(X, t>l = Q[P(x> WP(G 0. (8)

Equations (5) and (7) constitute a complete traffic flow model for the nondiscretized
case. Lighthill and Whitman (1955a; 1955b) have shown that in order for this model to
provide unique solutions, traffic variables should be allowed to have discontinuities
(shock waves) in space and time. More concretely, discontinuities may be developed by
continuous wave forms due to overtaking of slower waves by faster ones. At points of
discontinuity, a shock wave moving condition

dx,_ Q(Pz) - Q(P,>


dt - P2 - PI

replaces the use of (5) and (7) which are valid elsewhere, X, being the position of the
shock wave and p, and p2 (pZ > p,) being the density values upstream resp. downstream
of the discontinuity.
In the discretized case, eqn (7) should be changed in a suitable way, i.e. possible
development of discontinuities as expressed in (9) should be taken into consideration.
At a first trial, traffic volume between two freeway sections might be expressed as a
Fig. 1. A freeway system subdivided into sections.

weighted sum of the traffic volumes corresponding to the densities of the sections

q;(k) = a - Q[pi(k)]+ (1 - a>*Q[~l+dk>l (10:


with 0 % a s 1. If a is constant, however, it can be shown that the resulting discretizec
model cannot be stable for all possible density values. In fact, in view of (9), it seem!
plausible to choose a = 1 whenever dx,ldt > 0 in (9), and a = 0 when dx,ldt < 0, where
pi,pi+, replace p,,p2 in (9). However, comparisons with real traffic data led to the folIowine
somehow different rule:

q;(k) = 5 * Q[minb;(k), P~JI + (1 -aI * Q[p;+l(k)l (11)


CL if pi+,(k) < PC~
E=
1-a if p,+,(k) Z per (12)

with a parameter a having a constant value close to one. This rule was proposed with
a = 1 by Lebacque (1983). In the case pi+1 < per < p{, the rule gives (if a = 1) qi =
Q(P,,) = qmax, which is in accordance with the results of the kinematic wave theory
exposed by Lighthill and Whitham (1955a; 1955b).
In case there is an on-ramp included in section i + 1, eqn (11) is modified as follows:

q;(k) = E * Q[minbi(k), Pi,)] + (1-3{Q[Pi+dk)l -r,+dkN. UW

In the presence of an off-ramp in section i, eqn (11) is replaced by

q,(k) = iji - Q[min{p,(k),


pCJl + (1 -a> * Qh+l(k>l - s;(k). (lib)

In both cases, cl calculation (12) remains unchanged.


Naturally, q;(k) obtained by application of (11) should be limited to the capacity
of the sections i and i + 1. In presence of an on-ramp, the limitation to the capacity of
section i + 1 is changed to q;(k) 5 q,,,.i+I - r,,,(k). In presence of an off-ramp, the
limitation to the capacity of section i is modified to q,(k) S q,,,i - s/(k).
Equations (6), (ll), and (12) constitute a complete traffic flow model that will be
referred to as model B.

Introduction of a density gradient term


Assume that mean speed adjusts to traffic density according to

u(x, t) = V[p(x + Ax, t)] (13)

instead of (8). For Ax > 0, eqn (13) suggests that drivers anticipate density changes
downstream when adjusting their speed. Assuming Ax is small, we obtain by expanding
(13) in a Taylor series and neglecting high-order terms

v(x, t) = V[p(x, t)] + [dV/dp - apla~]~,,,, - Ax. (14)


Traffic flow on the Boulevard PCriphkique in Paris 33

Payne (1971) proposed the space increment to be Ax = 0.5/p based on heuristic


microscopic considerations. Furthermore, one may assume aV/ap to be approximately
constant, in which case one obtains from (14)

u(x, t) = V[p(x, t)] - VlP(XY t) * dP(X, t)lax, (15)

where v = -0.5 - dV/dp > 0 is a constant parameter. Using (l), we get for the traffic
volume

q(x, t) = Q[p(x, t)] - v * dp(x, t)lax, 06)

which replaces (7).


Equations (5) and (16) constitute a complete traffic flow model for the nondiscretized
case. This model leads to unique solutions without the need of introducing discontinuities
of traffic variables, as was done in the preceding section. Hence, possible abrupt changes
of traffic density always occur in a finite length of space and time that corresponds better
to real traffic phenomena.
In view of (16) it becomes apparent that even under steady-state conditions, max-
imum traffic volume is not constant but depends on the density gradient. In particular,
it is interesting to note that eqn (16) allows for q to become greater than qmax, namely
if p is sufficiently high and, furthermore, aplax is negative.
In the discretized case, eqn (15) may be written

vi(k) = VIP,(~)] - v/Ai * [Pi+,(k) - P;(~)I/(PA~) + KIT (17)

where the constant parameter K has been added to limit the second term in case of very
low density values.
Traffic volume qi(k) between two sections may be obtained in a way similar to (10).
Recalling (1)) we get

qi(k) = ci - p,(k) * Vi(k) + (1 - a)Pi+,(k) * s+,(k). (18)

In case there is an on-ramp included in section i + 1, eqn (18) is modified as follows:

q,(k) = 0. ’ Pick) * ui(k) + (l - a)[Pi+l(k) * ui+l(k) - ri+I(k)l. 084


In presence of an off-ramp in section i, eqn (18) is replaced by

qi(k) = OL ’ p,(k) ’ U;(k) + (1 - CY)Pi+t(k) ’ ui+tCk)- si(k). (18b)


By proper choice of the constant parameter ct, no stability problems will occur in
this case due to the extention introduced in (15) compared to (8). Hence, eqns (6), (17),
and (18) constitute a complete traffic flow model that will be referred to as model E.
Model E is better applicable to control strategy design because it does not contain jumps
in the sense of the rule (12) of model B.

A dynamic mean speed-density relationship


Models B and E contain stationary mean speed-density relationships, i.e. mean
speed is supposed to adjust instantaneously to traffic density whenever a density change
occurs. A more realistic description of traffic flow is expected if this hypothesis is removed
by introducing a small time delay T in (13) as proposed by Payne (1971):

u(x, t + T) = V[p(x + Ax, t)]. (19)

Expanding the left hand side of (19) in a Taylor series with respect to T and the
right hand side with respect to Ax, and using the same assumptions as in the preceding
TR(Bl 23:1-c
34 hl.PAPAGEORGIOC et a/.

section, we get after rearranging terms

T - dddr = V(p) - v - VIP. aplax, (20)


where the arguments x, t are depressed for convenience. dvidr is the acceleration of an
observer moving with the traffic stream, i.e.

duldt = adat + v - adax. (21)


Combining (20) and (21), we finally get

addt = u + adax + [v(p) - v - VIP


. apIa,v]h, (22)
where T is an additional constant parameter. Equations (l), (5), and (22) constitute a
complete traffic flow model for the nondiscretized case which is expected to describe
dynamic phenomena (particularly in congested traffic) with higher accuracy than previous
models.
Discretization of (22) leads to the following nonlinear difference equation:

u,(k + 1) = u,(k) + T [V(p,(k))- u,(k)]+ y 4k)[4,(k) - v,(k)1


I
_ VW,+,(k) - P!(k)1
7 ' L$ [P,(k)+ K] ’ (23)

where a constant parameter K was added in order to keep the third term limited when
pi becomes small. Furthermore, a parameter c was introduced in the convection term of
(23). This enables investigation of the importance of this term in the model calculations.
In this way, we get traffic flow model D, which consists of equations (6), (18), and
(23).

Influence of on-ramp volumes


On-ramp and off-ramp volumes, which are included in the conservation eqn (6),
have a direct impact on the density of the corresponding freeway section. Density, in
its turn, influences mean speed by means of the speed-density relationship. In this
way, we obtain an indirect influence of on-ramp resp. off-ramp volumes on mean speed.
However, entering and exiting cars usually have a lower speed than cars on the main
street. Therefore, it is interesting to investigate whether direct consideration of on-ramp
or off-ramp volumes on mean speed calculation may lead to a more realistic description
of traffic phenomena. This seems particularly important in cases with high entering and
exiting rates, e.g. in urban freeways.
Let us first assume existence of an on-ramp in a freeway section. Entering cars
generally have a mean speed v, that is not higher than mean speed in the main street
and they will therefore have to accelerate accordingly. Since acceleration effects are
ignored in models B and E for mainstream speed, it does not appear necessary to
introduce acceleration terms for entering traffic.
On the contrary, model D can be modified so as to consider directly a possible
influence of entering traffic on mean speed. Recall eqn (22) and assume lateral traffic
enters at the point x with injection rate r(x,t) [in veh/(h * km)] at a speed v,. Since
generally u, $ u, the rate of change dvlat is influenced by an additional term, namely
r/p - (u, - u). Then, if we assume that the entering speed is given by v, = 6’ * v, 0 d
6’ ~5 1, the additional term becomes - Gurlp, 0 Z 6 5 1.
In this way, eqn (22) may be changed to

adat = u - au/ax + [v(p) - u - V/P


- aplax]l7 - Gurlp, (24)

where 6 is assumed to be a constant parameter depending upon the layout of the ramp.
After space discretization, one may add the constant parameter K in the denominator
Traffic flow on the Boulevard Ptriphtrique in Paris 35
of the new term as was done for (23). Hence, recalling r = r-,/h;,the discretized form
of the additional term becomes -6 * T/h, - u, * ri/(pl + K). A similar term has been
proposed by Cremer and May (1986).
At this point, it is interesting to note that eqns (l), (5), and (22) have a perfect
analogon in the description of unsteady water flow in open channels (see, e.g. Ligget,
1975). The analogon of eqn (22), which is called the momentum equation, is derived
from conservation of momentum in water flow. The additional term to be included in
(23) in case of lateral inflow has the same structure as in the momentum equation of
water flow.
In the case of off-ramps, exiting cars generally decelerate, leading to lower mean
speeds in the corresponding freeway sections. Hence, the impact of off-ramp volumes
may be described in the same way as it was done for the on-ramp volumes, r being
replaced by s.

Limitations of the models


Aggregate models presented in this section describe traffic flow in one dimension
(along the road axis). Although traffic flow along street width is generally nonhomo-
geneous (e.g. trucks and slow vehicles generally use the right lanes of the street), traffic
flow dynamics are described accurately enough by one-dimensional models under certain
conditions. It seems rather difficult to state these conditions precisely but it is possible
to name some traffic situations that probably cannot be described accurately by one-
dimensional models.
The first of these traffic situations is off-ramp congestion occurring when, for any
reason, actual off-ramp throughput is inferior to the portion of traffic wishing to exit.
In this case, one or more right lanes of the main street may be blocked whilst traffic on
the left lanes may be fluid.
A second traffic situation probably leading to model failure may occur when trucks
are not allowed to use the far left lane, as is the case on the Boulevard PCriphkrique.
If the number and the percentage of trucks exceeds certain limits, lanes used by trucks
may be blocked whilst traffic in the left lane may be fluid.
Finally, the same objection applies to the case of special lanes dedicated to buses,
taxis, etc.
Suitable extensions of the model equations may be necessary in these cases in order
to facilitate adequate description of traffic phenomena.
The speed-density relationships of the various models presented in this section do
not take into account the percentage of trucks in traffic. Although different truck per-
centages are not expected to give raise to a structural change of the models, results may
differ quantitatively for different truck portions. Cremer (1976) and Cohen (1981) have
proposed suitable extensions of the fundamental diagram so as to take into account the
actual percentage of trucks. Since a fundamental diagram is included in each of the
models, B, D, and E, an extension of this type can be easily accomplished in the models
of this section.

3. TRAFFIC DATA AVAILABLE

Parameter estimation and validation of the macroscopic models presented in Section


2 requires availability of adequate real traffic data. This Section presents the data utilized
in the present study along with application of the models to the corresponding street
geometry.

Description of street geometry


Traffic data utilized in the present study were collected from a three-lane stretch
of the BP with a lenght of 976 m including one on-ramp and one off-ramp, as shown in
Figure 2. There are five detector stations (denoted A, B, C, D, and E) installed along
the main street as well as two detector stations (denoted F and G) for the on-ramp and
off-ramp, respectively. Each station contains one to three detectors.
M. PAPAGEORGIOUet al

E D
dktector stations
direction off traffic flow

Fig. 2. The BP stretch considered.

Available measurement information


Collected data consist of measurements of car passages and occupancies for each
detector over 5-s intervals. Using estimated mean effective vehicle lengths for each lane,
mean speed estimates w have been derived for each detector station (see Papageorgiou,
1987, for details).
Measurement errors are caused by lane changes of vehicles activating either both
or neither of two adjacent detectors. In fact, balancing (conservation equation!) total
measured inflow vs. total measured outflow for the given BP stretch over sufficiently
long time periods was found to lead to unrealistically high resp. negative densities in
the stretch due to counting errors.
For technical reasons, synchronized measurement sets are available for either of
the following two station configurations: A, B, C, D, F, and G (configuration 1) or B,
C, D, E, and F (configuration 2). Three of four available measurment sets are based
on the first configuration and a fourth set is based on the second configuration. Each
of them has a duration of four hours. Since T = 5 s, this corresponds to a number K
= 2,280 of time intervals for each set.
Each measurement set includes at least a part of a rush-hour period. Hence, available
traffic data cover the whole range of possible traffic conditions (fluid, dense. and con-
gested traffic). Volume-occupancy diagrams (corresponding to the fundamental dia-
gram), obtained for one-minute intervals, differ for the individual detector stations. In
particular, station D has a relatively high critical occupancy value compared to the other
stations, probably due to the merging process. Stations D and E have similar maximum
volumes in the order of 6,500 veh/h, whereas 5,300 veh/h was the maximum volume
observed at stations B and C, due to underutilization of the right lane between the ramps.
Maximum on-ramp volumes during rush hours are in the order of 1,500 veh/h, the
corresponding off-ramp volumes being in the order of 600 veh/h. Congestion in the
considered stretch mostly mounts from downstream, only a few occupancy peaks of short
duration being due to high on-ramp volumes.

Subdivision into sections


Application of dynamic models B, D, and E implies subdivision of the BP part
under consideration into a number of sections. Referring to Fig. 2, the stretch AB is
subdivided into two sections of 110.5 m each, the second one including the off-ramp.
The stretches CB and DC constitute two further sections, the latter including the on-
ramp. Finally, stretch DE is subdivided into four sections of equal length A = 125 m.
In this way, the first configuration contains four sections with an off-ramp in the second
section and an on-ramp in the fourth section, whereas the second configuration contains
six sections with an on-ramp in the second section.
The same fundamental diagram (same parameters) will be used for all street sections
despite the remarks of the preceding section.
Results obtained by Cremer and Papageorgiou (1981) suggest that the time interval
T in the time discretized case should be chosen

T < A/v,
Traffic flow on the Boulevard Ptriphkrique in Paris 31

in order to guarantee model stability. In our case, the smallest sections are in the order
of 110 m and the free speed is expected in the order of 90 km/h, which gives T < 4.4
s. Thus, T is chosen to be 2.5 s. Since the measurement sets are discretized on the basis
of 5 s, the same measurement values are used for pairs of subsequent time intervals.

Definition of inputs and outputs for parameter estimation


Since the models are supposed to describe traffic dynamics for the whole density
range, measured traffic variables (i.e. traffic volumes and mean speeds) at both the entry
and the exit of the BP stretch are treated as input variables along with the on-ramp and
off-ramp volumes. Hence, the models being fitted with the input variables, mean speeds,
and volumes at the internal sites may be calculated. Deviations of the calculated out-
put variables from the corresponding real traffic data may then be used as a measure of
the models quality.

Utilisation of input data in the model equations is effectuated as follows:


(i) Upstream boundary volume replaces q,(k), appearing in (6) for i = 1 in all
models.
(ii) Upstream boundary means speed replaces u,(k), appearing in (23) for i = 1 in
model D.
(iii) The ratio of downstream boundary volume and mean speed gives downstream
boundary density pN+,(k) needed for the last freeway section in (12) (model B) resp.
(17) (model E) resp. (23) (model D).
(iv) Downstream boundary volume replaces Q(pN+,) appearing in (11) (model B)
resp. pN+, * IJN+I appearing in (18) (models E and D). It should be noted that downstream
boundary volume is not directly used in the conservation equation of the last section in
order to avoid accumulation of counting errors included in the measurements.
(v) On-ramp and off-ramp volumes replace the corresponding terms in the model
equations.

Mean speed w;(k) at the boundary of two sections is obtained by

w,(k) = a’ * Q(k) + (1 - cX’)&+r(k) (24)

with a’ = ol for model B and a’ = CYfor models D and E.

4. PARAMETER ESTIMATION FOR MODELS B, E, AND D

Choice of a Q(p) formula


Each one of the macroscopic models B, E, and D contains a Q(p) resp. V(p) formula.
In a preliminary investigation, five formulae were applied to the volume-occupancy data
of each detector station. These formulae are (2), (3), (3) with a = 1, (3) with a = 2,
and (4) with p = 3. As a best compromise of accuracy vs. computational effort, formula
(3) with a = 2 was selected for further use in the models B, E, and D. Application of
the selected formula to the two internal stations of each measurement set, with param-
eters optimized individually for each station and each set, leads to the average standard
deviations J, of traffic volume error shown in Table 1.

Nominal parameter values


The model parameters to be estimated are: Us, po, and cx for model B, ur, p_, CL,
and v for model E, and uf, per, 01, v, T, 5, 6,,, and &Offfor model D. The parameter K,
occurring in some denominators of (17) and (23), is chosen to K = 40 veh/km since any

Table 1. Average standard deviation for the two internal stations


applying formula (3) with a = 2

Measurement set I II III IV

Standard deviation JQ (veh/h) 492 486 499 463


38 hC PAPAGEORGIOL’
et al.
reasonable value of this parameter can be practically compensated by according adjust-
ment of the value of 1, resp. 8.
Parameter values for each model should be specified so as to minimize the deviation
of the models’ output from the corresponding measured values. Using appropriate op-
timization tools, a nominal parameter set was established for each model on the basis
of the four measurement sets available as follows:

Model B: uf = 88 km/h, per = 124 veh/km, a = 0.94


Model E: uf = 88 km/h, pci = 120 veh/km, IX = 0.6, v = 5 kmr/h
Model D: ur = 90 km/h, per = 120 veh/km, (Y = 0.8, v = 28 km’ih, T = 0.02
h = 72 s, 5 = 0.5, 6,” = 0.1, 6,tr = 0.
These parameters values give rise to the following remarks:

(i) Free speed uI and critical density per are similar for all models, hence the fun-
damental diagram used by all models is roughly the same. In fact, the maximum volumes
resulting from these values are 6.618, 6.405, and 6,551 veh/h for models B, E, and D,
respectively. Again it should be underlined that a unique fundamental diagram is used
for all sections in spite of the different shapes appearing for different sites. The value
of ur(pc,) in model D is low (high) compared to the free speed (critical density) specified
in Cremer and Papageorgiou (1981) for a German highway but is very similar to the
free speed (critical density) specified in Cremer and May (1986) for a Los Angeles,
California, freeway. This is probably due to the speed limitation that is present in BP
and the Los Angeles freeway but not in the German highway. Anyhow, the specified
values of free speed and critial density reflect the particular traffic conditions on Boulevar
Peripherique.
(ii) In view of the different mathematical structure of each model, discussed in
Section 2, it is not surprising that the parameters cr and v obtain different values for
each model. As far as model D is concerned. it is interesting to note that these values
are of the same order as in Cremer and Papageorgious (1981) (CY= 0.8, v = 21.6 km’/
h) and in Cremer and May (1986) (o = 0.95. v = 23.9 km’ih).
(iii) The value of the time constant T for model D is higher than in Cremer and
Papageorgiou (1981) (T = 36 s) and Cremer and May (1986) (7 = 20.4 s). In all cases,
however, the value of T is a “macroscopic” one, i.e. it is two orders higher than the
drivers’ reaction time in microscopic models from which T is supposed to originate (see
also Babcock, 1984).
(iv) The value of the convection parameter [ introduced in this report for model D
is seen to be different than one. The value of the on-ramp parameter 6,” is rather low
whilst no amelioration of results could be achieved for &,rf > 0. These values will be
further commented on in the next Section.

Performance criterion values


Table 2 gives the values of the one-minute standard deviations of traffic volume
error Jo and mean speed error JW obtained by application of each model with its nominal
parameters. The values of Table 2 are averages over the four available measurement
sets. The results give rise to the following comments:

(i) Model D is better than models B and E for all measurement sets considered.
Model E provides better results than model B for the sets I, II, and IV and is worse for
the set III. On the average, model E is better than model B.

Table 2. Results with nominal parameter


values

Model B E D

Jp (vehih) 486 427 410


IW (km/h) 17.5 14.4 8.3
Traffic flow on the Boulevard PiriphCrique in Paris 39

(ii) Performance differences are striking with respect to mean speeds. Model D gives
error standard deviations in the order of 8 km/h, whilst standard deviations for models
B and E vary between 10.2 and 25.8 resp. 10.6 and 22 km/h. In the average, standard
deviation of mean speed error increases by 112% for model B and by 75% for model
E as compared to model D.
(iii) Performance differences are less striking with respect to traffic volumes.
Model D gives throughout error standard deviations in the order of 400 veh/h, whilt
Jo for models B and E vary between 431 and 543 resp. 382 and 490 veh/h. On the
average, J, increases by 18% for model B and by 4% for model E as compared to
model D.
(iv) All dynamic models provide standard deviation of traffic volume that are equal
or lower than the ones of Table 1 obtained with a static fundamental diagram. This is
at first view surprising for model B because: Model B makes direct use of the fundamental
diagram with density values derived from the conservation equation, whereas the oc-
cupancy values used in Table 1 are measured-and hence more accurate-values, and
results of Table 1 have been obtained by individual optimization for each station and
each measurement set whilst results of model B have been obtained using a unique
parameter set.

These application conditions being in favour of the simple fundamental diagram, similar
quality of volume estimates obtained by application of model B can really be explained
by the following two facts: Results of Table 2 have been obtained using volume mea-
surements from both ends of each configuration. whereas results of Table 1 are based
on local occupancy measurements only, and model B is a dynamic model including shock
waves in the discretized form (11) and (!2), whereas the fundamental diagram is a purely
static, local relationship. Also. standard deviations of mean speed error provided by
models B and E for measurement set IV are particularly high and will be discussed in
the next section.

Output trajectories
The output trajectories, i.e. trajectories of the two internal stations for traffic
volumes, confirm that, generally, all models describe traffic volume evolution with
satisfactory accuracy. In fact, if model equations are numerically stable, each freeway
section acts as a nonlinear first-order system (for models B and E) or second-order
system (for model D) that sooner or later leads model traffic volume near the input
(boundary) traffic volumes. Thus, even if model mean speed is, say, unrealistically
high, traffic volume is described correctly by calculating accordingly low density
values. As an example, Fig. 3 depicts the traffic volume trajectories of station B for
measurement set II for model D. As for all figures of the present section, solid
lines depict measured trajectories, whilst broken lines depict the corresponding model
outputs.
Mean speed trajectories of station C for measurement set I are shown in Figures 4,
5, and 6 for models B, E, and D, respectively. Model B is seen to follow traffic congestion
in an acceptable way, but strong oscillations of mean speed are not accurately described.
On the other hand, model E and, particularly, model D describe speed variations with
high accuracy.
Figures 7 and 8 show the mean speed trajectories of station D for measurement set
IV and models B and D, respectively. Model D is again describing mean speed evolution
in a very accurate way. On the contrary, models B and E fail to reproduce congestion,
although results with respect to traffic volumes are still satisfactory. This is due to the
fact that the density values calculated by the two models are undecritical (around 85
veh/km) but produce the same traffic volume as the overcritical densities of real life.
Since measurement set IV is the only one to be based on configuration 2, it is not possible
to state whether the failure of models B and E is due to the geometrical configuration
or due to the particular measurement set.
40 M. PAPAGEORGIOUet al

80130 - traffic volume (veh/h)

6000

2000 I

1000
time (mn)
0, ,,,,, I ,,,..I ,*. # >.>..!. ‘..W
0 20 40 60 80 100 120 140 160 180 200 220 240

Fig. 3. Traffic volume for measurement set II, model D.

120 -traffic speed (km/h)

time (mn)
0. ’ ~
0 20 40 60 80 100 120 140 160 180 200 220 240

Fig. 4. Mean speed for measurement set I, model B.


Traffic flow on the Boulevard PCriphCrique in Paris

60

time (mn)
OI
0 20 40 60 60 100 120 140 160 160 200 220 240

Fig. 5. Mean speed for measurement set I, model E.

120 traffic speed (km/h)

100 -

time (mn)

0 20 40 60 80 100 120 140 160 180 200 290 240

Fig. 6. Mean speed for measurement set I, model D.


M. PAPAGEORGIOUer al.
speed (km/h)

60

I
time tmn)
0 .’ 1, 4 I....I....I....I....~._L.,_~..~.._.!_..I-.L.~--

0 20 40 60 60 100 120 140 160 160 200 220 240

Fig. 7. Mean speed for measurement set IV. model B.

Fig. 8. Mean speed for measurement set IV. model D.


Traffic flow on the Boulevard PCriphCrique in Paris 43

5. SENSITIVITY INVESTIGATIONS

Motivation and procedure description


Beside the results produced by the nominal parameter set in the last section, the
following investigations seem to be of particular interest:

(i) Do optimal parameter set, obtain for each measurement set individually, differ
substantially from each other? If yes, the model structure is suspect. If no, we may expect
to describe traffic conditions for any traffic situation in the given stretch.
(ii) D o o pt im a 1 parameter sets, obtained for each configuration, differ substantially
from each other? If yes, application of the model to other parts of the BP appears
questionable. If no, the model is very likely to reproduce traffic conditions at other BP
stretches with comparable accuracy.
(iii) What is the degree of model deterioration in case of simplications of the model
structure? This question appears particularly interesting in the context of computation
time, e.g. in the case of real-time optimal control.
(iv) How sensitive are the models with respect to parameter variations?

In order to investigate these questions let us define the following quantity

SJ,(P) - - JdP*)?
- = Jw(P> - (25)

where Jw is the standard deviation of mean speed error and B* is the nominal parameter
-
set of the corresponding traffic model:
Figure 9 contains sensitivity diagrams obtained for model D by changing one pa-
rameter at a time in the range

-0.6 < (pi - p*,)/B*I < 0.5,

whilst all other parameters are kept equal to their nominal values. The diagrams of Fig.
9 depict these parameter change ratios vs. the corresponding performance change ratios
SJ,/J; for each measurement set. It should be noted that mean speed sensitivity is much
higher than traffic volume sensitivity (see Papageorgiou, 1987, for detailed results).
Sensitivity results are commented on in the following sections for each parameter in-
dividually. Furthermore, sensitivity with respect to the sample time interval is investi-
gated in a final section.

Free speed and critical density


The parameters ur and perspecify the fundamental diagram to be used in each model.
These two parameters are the most sensitive ones for all models, a phenomenon that
was also observed in analogous investigations by Cremer and Papageorgiou (1981).
Hence, proper selection of the fundamental diagram to be utilized appears to be im-
portant for model application.
Sensitivity of model D with respect to these parameters is lower than sensitivity of
models B and E. In particular, standard deviations for mean speed error of model
D vary in the order of &30% if the free speed is changed in the order 80 s ur I 100
km/h. This is judged to be satisfactory in view of the wide range of ur variations.
On the other hand, the fact that the optimal values of v/ and per for the individual
measurement sets lie near each other confirms models’ robustness. Merely results ob-
tained for models B and E and measurement set IV are not satisfactory. This is due to
4-l M. PAPAGEORGIOU er al.

0.6 Pcr

L
;, :
:: :
:;:
0.4 -
;:
/
....
0.2 ,.t
,:
.:’
i ;
: :
0 i :
: .:: ....____...

.0.2
t, 8. 1. I. I 1. 08/B

-0.6 -0.4 -0.2 0 0.2 0.4

0.6

.0.2
08/S
-0.6 -0.4 -0.2 0 0.2 0.4

Fig. 9. Sensitivity results for model D, measurement sets I (_), II (-), III (_ -), and IV (-
II. )

the failure of these models to reproduce congestion for measurement set IV, as already
discussed in Section 4.

Weighting parameter cx
A dramatic performance deterioration, probably due to numerical instability, is
observed for all models if OLfalls below a critical minimum value. This critical value is
in the order of 0.7 for model D, 0.5 for model E, and 0.5 for model B.
Model D is fairly insensitive with respect to variations of OLabove the critical value.
In particular, for a = 1, leading to a simplification of the model structure [see eqn (IS)],
the maximum deterioration of the standard deviation for the mean speed error reads
only 10%.
On the other hand, model E and, particularly, model B are fairly sensitive with
respect to a variations. Furthermore, behaviour of model B is not uniform for the
individual measurement sets.

Anticipation parameter v
Model D is fairly insensitive with respect to v variations. Model E is more sensitive,
particularly for v smaller than the nominal value.
Traffic flow on the Boulevard PCriphCrique in Paris
a)“&
f

,00/S
t
-0.6 -0.4 -0.2 0 0.2 0.4

6 on 6 off

-0.2 -0.2
, mm LB/B
-0.6 -0.4 -0.2 0 0.2 0.4 -0.6 -0.4 -0.2 9 0.2 0.4

Fig. 9. (Continued).

Time constant 7 and convection parameter 5


Model D is rather insensitive with respect to variations of T and 5. In particular,
4 = 1 gives a maximum deterioration of standard deviation for the mean speed error
in the order of 10%.

On-ramp parameter a,,


As already mentioned in Section 4, the role of the on-ramp parameter 6,” was found
less crucial than initially conjectured.
A first observation extracted from the diagrams is that results for the measurement
sets I, II, and III of the first configuration are not better for 6, > 0. If we assume that
on-ramp volumes do exert a direct influence on mean speed dynamics, a possible inter-
pretation of the above observation is the following: The influence of on-ramp volumes
on mean speed of the last freeway section is soon visible at the exit of the system (station
D). Boundary mean speed wu being an input variable, the influence of on-ramp volumes
on mean speed is transmited to the model equation indirectly, via wu. This may be the
reason why direct consideration of the on-ramp volumes does not lead to any visible
amelioration,
This conjecture is strengthened by the fact that results in the second configuration
(measurement set IV) do ameliorate for a,,, > 0. Let us now focus our attention on the
nature of this amelioration.
46 M. PAPAGEORG[OUer al.
As already mentioned in Section 3, no significant breakdown of real traffic is caused
by high on-ramp volumes in the measurement sets available. Merely short negative speed
peaks were found to be created inside the considered street stretch, but no visible
correlation to positive on-ramp volume peaks could be established. It is therefore hy-
pothesized that, in the considered traffic situation, high on-ramp volumes merely aug-
ment the risk of short time breakdowns, a phenomenon that can hardly be reproduced
by a deterministic model.
In fact, for measurement set IV, an amelioration of more than 20% in mean speed
error standard deviation is achieved for 6,” = 0.2 as compared to 6,” = 0. However, a
comparison of the corresponding mean speed trajectories shows that the amelioration
is not due to the reproduction of the short negative speed peaks but due to a better
average description of the real-life speeds.
It is empirically known that high on-ramp volumes may cause severe congestion of
main-stream traffic. The question whether such congestions are better reproduced using
6,” > 0 and the degree of possible amelioration cannot be definitely answered on the
basis of the data available for this study. In any case, recalling the physical meaning of
6,,, it seems that a possible, more or less important influence of on-ramp volumes on
mean speed dynamics depends significantly on the on-ramp layout. Hence, estimation
of 6,” may be necessary for each type of ramp layout individually. However. results of
the present study indicate that the direct influence of on-ramp volumes on mean speed
dynamics may be of minor importance for a number of merging areas.

Off-ramp parameter
No amelioration of results was achieved for ?joff> 0. This is not really surprising
because, if the off-ramp is not blocked, no significant difference of speed between exiting
and main-stream vehicles occurs in real traffic on the Boulevard Peripherique.

Sensitivity with respect to the sample time interval T


In order to investigate sensitivity of the mode1 calculations with respect to the sample
time interval T, model D was run with the nominal parameter values for all measurement
sets with T = 1.25 s and T = 5 s.
For T = 5 s, results were extremely bad due to numerical instability as hypothesized
in Section 3. For T = 1.25 s, model trajectories were practically identical to the ones
produced with T = 2.5 s.
Due to the limited length of the BP part considered, no investigation of the model
quality in case of longer street sections have been effectuated. However, on the basis
of results reported by Cremer and Papageorgiou (1981), it is expected that the quality
of the model calculations with longer stree sections will be comparable to the ones
achieved here, provided the ratio A/T is kept constant. Hence, a section length of 500
m and a sample time interval of 10 s are expected to lead to good results.

6. CONCLUSIONS

As stated in Section 1, the present study aimed at developing a mathematical model


of traffic on Boulevard Peripherique de Paris to be used both for control strategy design
and for simulation purposes. Let us summarize the main conclusions:

(i) After thorough investigations with respect to mathematical structure, parameter


estimation, and sensitivity, model D was found to be clearly superior to the two simpler
alternative models B and E. More precisely, the mathematical structure of model D
allows for fairly accurate description of wave movement in traffic flow in both directions.
Model D, fitted with the estimated parameter values, provided excellent results for all
measurement sets used in this study. Furthermore, all sensitivity investigations effec-
tuated in this study demonstrated a considerable robustness of the model performance.
(ii) Models B and E were clearly inferior to model D with respect to modelling
Traffic flow on the Boulevard Peripherique in Paris 47

accuracy and sensitivity. Moreover, their performance was found to be subject to con-
siderable variations for different measurement sets. In particular, both models failed to
reproduce congestion in the case of measurement set IV. Although the reasons for this
failure are not exactly known, there is some insecurity in the application of these models
due to their unsteady performance.
(iii) Attempts to improve performance of model D by incorporating a direct influ-
ence of on-ramp volumes in the mean speed equation led to a limited observable success.
This may be due to the particular traffic situation reflected in the measurement data
available. In fact, no significant congestion of main-stream traffic originates from high
entering volumes at the on-ramp included in the considered BP part. Hence, the question
about utility of an additional term in the mean speed equation was only partially an-
swered. In any case, the parameter of the additional term is expected to depend upon
the layout of different on-ramp types and may hence have to be adjusted individually.
(iv) Application of model D with longer sample time intervals (e.g. 10 t 20 s) and
section lengths (e.g. 500 + 1.000 m) was not possible due to the limited measurement
data. However, experience from previous studies indicates robustness of model per-
formance also in this case.
(v) Some simplifications of the model structure (CX= 1, 6,” = 0, 6,e = 0) that re-
duce the computational effort were found to lead to only slight deterioration of perfor-
mance. This is particularly important for the design of control strategies on the basis of
model D.

In summary, model D appears to be an adequate tool both for control strategy


design and for simulation of traffic flow on Boulevard Peripherique. Minor modification
of some parameter values may further improve model performance in case of application
to a longer stretch with longer discretiiation intervals and different types of on-ramps
resp. off-ramps. The corresponding investigations will be effectuated in the near future
when additional data will be available.

Acknowledgemenr-The authors would like to thank the traffic engineering staff of Ville de Paris for the data
material they made available for the purposes of this study.

REFERENCES

Babcock P. S. (1984) Improved dynamics and performance for the FRECON freeway simulation model.
Technical Document UCB-ITS-TD-84-l. Institute of Transportation Studies, University of California,
Irvine, California, August 198-l.
Cohen S. (1981) Effet des poids lourds sur la capacite des autoroutes urbaines. Internal report, INRETS,
Arcueil, France.
Cremer M. (1976) A new scheme for traffic flow estimation and control with a two-component model. Proc.
3rd IFAC/IFIP/IFORS Symp. on Control in Transportation Systems, Columbus, Ohio, pp. 29-37.
Cremer M. and May A. D. (1986) An extended traffic flow model for inner urban freeways. Preprints 5th
IFACIIFIPIIFORS Intern. Conf. on Control in Transportation Systems, Vienna, Austria, pp. 383-388.
Cremer M. And Papageorgiou M. (1981) Parameter identification for a traffic flow model. Auromafica 17,
837-813.
Lebacque J. P. (1983) Simulation semi-macroscopique des reseaux urbains. Internal report, INRETS, Arcueil,
France.
Ligget J. A. (1975) Basic equations of unsteady flow. In Unsready Flow in Open Channels, 29-62 (Edited by
Mahmood K. and Yevjevich V.). Water Resources Publications, Fort Collins, Colorado.
Lighthill M. J. and Whitham G. B. (1955a) On kinematic waves I. Flood movement on long rivers. Proc.
Royal Society of London, Series A, 229, 281-316.
Lighthill M. J. and Whitham G. B. (1955b) On kinematic waves II. A theory of traffic flow on long crowded
roads. Proc. Royal Society of London, Series A, 229, 317-345.
May, Jr. A. D. and Keller H. E. M. (1967) Non-integer car-following models. Highway Res. Rec. 199,19-
32.
Papageorgiou M. (1983) Applications of automatic control concepts to traffic frow modeling and control.
Springer-Verlag, Berlin-Heidelberg-New York-Tokyo.
Papageorgiou M. (1986) Freeway on-ramp control strategies: Overview, discussion, and possible application
to Boulevard Peripherique de Paris. Internal report, INRETS, Arcueil, France.
Papageorgiou M. (1987) IMacroscopic modeling of traffic flow on the Boulevard Periphtrique de Paris. Internal
report, INRETS, Arcueil, France.
Payne H. J. (1971) Models of freeway traffic and control Simularion Council Proc. 1,51-61.
Wardrop J. G. (1952) Some theoretical aspects of road traffic research. Proc. In.stn. Civ. Engrs., pt. II, 1,
325-362.

You might also like