Anda di halaman 1dari 20

A Hybrid Model for Model Predictive Control of Urban Trafc

F. Basile, P. Chiacchio and D. Teta

Abstract In this preliminary paper the simulation of an efcient hybrid model is presented to predict the behavior of an urban area as controlled by a certain control strategy. Intersections and on-ramps are modelled by Timed Petri Nets while road links are modelled by a stochastic discrete time model. It is assumed that vehicle routings at each intersection are stochastic. An algorithm for the on-line identication of routing rates is presented. The control inputs are trafc lights switching times. The aim is to minimize the average number of vehicles in the urban area over a xed time horizon to avoid a trafc congestion in the area.

Index Terms Discrete event system, Hybrid model, Petri Nets, Urban trafc control.

Published as: F. Basile, P. Chiacchio, D. Teta, "A Hybrid Model for Model Predictive Control of Urban Trafc," 9th European Control Conference (ECC07), Kos, Greece, pp. 18231830, July 2007.

F. Basile, P. Chiacchio and D. Teta are with the Dip. Ing. dellInformazione e Ing. Elettrica, Universit di Salerno, Via Ponte don Melillo, 84084 Fisciano (Salerno), Italy. {fbasile,dteta}@unisa.it

I. I NTRODUCTION Models for the simulation of road trafc serve different purposes [12]. For instance, they may support off-line planning of road networks and trafc light settings, but the models may also be used for model predictive on-line control of complex trafc guidance systems [13]. Of course, even if it seems to be practical to increase the degree of abstraction in order to obtain a high computational performance (which is especially required in cases of on-line application), decisive effects of the road trafc behavior must be preserved.
state actions

Controller

Observer

Sensors

vehicle arrivals

Urban Area Plant

vehicle departures

Fig. 1.

A general scheme for urban trafc control.

In [1] a new hybrid model for urban trafc control has been proposed. This hybrid model is used in this paper since it is revealed to be enough accurate and efcient for on-line simulation purpose. An urban area, consisting of elementary components such as priority controlled intersections, simple road links, road links with on-ramps and off-ramps and intersections controlled by trafc lights, can be modelled by an hybrid model. Some components such as intersections require detailed DES models like interconnected (colored) timed Petri Nets (PNs) modules, while other components such as road links can better be represented as more coarsely discrete time systems like the one proposed in [4]. Token colors are used to model vehicle routings like in [7]. In other PN models the weights of net arcs have been used and usually constant vehicle routing are assumed [9]. The vehicle routing plays a key role in a detailed model like the one used in this paper. In this paper a modied version of the model in [1] is presented. Intersections are modelled by

Timed PNs (TPNs) where vehicle routings are modelled by stochastic routing rates associated to transitions in conict, i.e. having common input places. This choice is motivated by the possibility to identify on-line such routings, which can be time-varying or constant without loss of generality, and thus to use their actual values to make more accurate the model used to predict system behavior. Fig.1 shows a general scheme for urban trafc control used also in this paper. An urban area to be controlled (e.g. two signalized intersections connected to each other with a road link) together with sensor and actuator signals, is viewed as a plant, according to the classical closed loop scheme in the control theory. Sensor signals detect vehicle arrivals and departures at some locations and information on some internal events. These observations are inputs for the observer subsystem [17]. The observer contains a detailed hybrid model (modular and adaptive) of the controlled urban area. Actuator signals, such as switching time for trafc signals, are selected using the observer state. The modularity is realized by considering separate models for components of an urban area such as signalized intersections, road links, etc. The controller works in real-time: according to the model predictive control methodology it reads the estimated state generated by the observer and predicts the future evolution of the plant model for various controls (using what-if techniques). It is important to notice that controller actions are sent to the plant as well as to the observer, in order to update the estimated state using the true input values. The model presented in [1] - integrated in a simulation framework - may be used for urban trafc control by resorting to model predictive control (MPC) strategies. MPC, usually contains the following three ideas [6]:

Explicit use of a model to predict the process output in a future time horizon. Calculation of a control sequence to optimize a performance index. A receding horizon strategy, so that at each time instant the horizon is moved towards the future, which involves the application of the rst control signal of the sequence calculated at each step.

In literature, several authors have proposed the MPC approach to the problem of trafc control [5], [16], [14], [15], [10]. A critical point of such an approach is the optimization problem of minimizing a certain objective function over a future time horizon. Most optimization methods attempt to determine the values for the controllable variables of

Optimizer

kTc

x
Simulation Model

Algorithm

J ( , x )

Road Network State Traffic Light State

Urban Traffic Area Optimal Signal Settings

Fig. 2.

Simulation-based urban trafc control scheme.

a system so that a certain outcome of the system is optimized. Many of these systems can be formulated analytically and optimized by various techniques of mathematical programming also in trafc control applications [8]. However, there are more complicated systems, often stochastic in nature, for which the objective function cannot be expressed analytically. The problem formulated in this paper has the following characteristics:

There is no analytical expression of the objective function. This eliminates the possibility of differentiation or exact calculation of local gradients.

The objective function is a stochastic function of the deterministic decision variables.

Hence, such a problem can be classied as a simulation optimization problem [2]. Simply stated, a simulation optimization problem is an optimization problem where the objective function (objective functions, in case of a multi-criteria problem) and/or some constraints, are responses which can only be evaluated by computer simulation. These functions are stochastic and implicit functions of decision parameters of the system. As objective function to minimize, it is considered the average number of vehicles in the urban area over a time horizon of K trafc lights cycles having each one a duration of 90 seconds. This function depends on the overall system state x (road network state and trafc light state), is stochastic and can only be evaluated by simulating a urban trafc model; as for decision variables , it is possible to consider trafc light switching times for each controlled intersection of the urban area, over K cycles. The aim is to nd the optimal vector that minimize J (see Fig. 2). A variety of approaches have been proposed to deal with the optimization of systems

through simulation modelling (see [2] for a complete review). As in [3] a forward simulation of the whole urban area over K trafc lightss cycles is used to decide if a strategy is better than another one. In this work the key role of the choice of time horizon (i.e. the choice of K) is highlighted. II. T HE
URBAN AREA MODEL

In this section the main elements (TPN model of intersections and discrete time model of road links) of the urban trafc model used in the paper are presented. The model of intersections is described in detail, while road links model is briey recalled, further details can be found in [1]. We propose a hybrid approach in urban area modelling. A microscopic representation of urban trafc can be used for some parts of the network (intersections and on-ramps) while coarser, more aggregated models can be used for other parts (road links). PNs are very suitable in representing resource sharing, such as an intersection or arriving zone at an on-ramp. On the other hand, they are not suitable in modelling road links, where many vehicles concurrently move at different speeds. For a road link we prefer to adopt a hybrid stochastic model and to interface it to the PN model of an intersection or an on-ramp by means appropriate interfaces. Our model is modular: the urban area is modelled as the interconnection of timed PN modules, connected to each other by fusion places and stochastic hybrid systems. In this preliminary work the modelling of pedestrians or other car ows like the ones represented by vehicle leaving parking areas are not considered. A. The intersection model It is assumed that the reader knows the denition of Place Transition (P/T) nets, for further details see [11]. Fig.3-(b) shows the P/T net model of the simple intersection depicted in Fig.3(a). The idea behind this model is very easy: starting from a discretization of the crossing area in a nite number of zones (1 8 in Fig.3-(a)) and under the assumption that each zone can be occupied only by one vehicle at a time, places {pi }i=1..8 represent respectively the crossing area zones 1 8. To model the time that a vehicle needs to move from one zone to another (e.g., transition t1 models the move from the zone 5 to the zone 6), a time function f : T R is introduced, which associates to each transition t T , the deterministic delay from enabling to

ring. If f (t) = 0, t is said to be immediate. In Fig.3-(b) immediate transitions are represented by black bar while timed transitions are represented by empty box. Notice that our model can represent behavioral conicts between timed transitions. For instance, transitions t1 and t2 may be in behavioral conict when places p5 and p8 are both marked. This corresponds to two vehicles attempting to occupy the same zone 6: the fastest one wins. Places pc3 and pc6 guarantee that each zone can be occupied by one vehicle at a time only. A token in place p6 (p3 ) models a vehicle in zone 6 (3) that has to decide whether to move to zone 3 (2) or to zone 7 (1). To describe vehicle routings, in zone 6 and zone 3, routing rates associated to immediate transitions are used. Every time that a token is in p6 (p3 ) a conict between transition t63 (t31 ) and t67 (t32 ) occurs. The conicts between immediate transitions are solved by using routing rates. Let T = {t1 , . . . , tk } T be transitions in structural conict (i.e. having nonempty transition preset intersection: t1 t2 .. tk = ). The constants r1 , . . . , rk N+ (routing rates) are explicitly dened in the net interpretation in such a way that when t1 , . . . , tk are enabled, transition ti T res with probability ri /(
k j=1 rj ).

The value of the routing rates are obtained

by measures at the intersection during an appropriate time horizon.


p c1 p1 t6 p c2 p 31 t 32 p2 t7 p 32 t 31 p3 p c3 t4 p c5 t1 p5 p c6 t2 p c8 p8
(D

p c4 t3 p4

2 5

3 6 8

4 7

(a)

p 63 t 63 p6 t 67 p 67 t5

p c7

p G1

t G1

p Y1

t Y1

(D

G1 )

p R1 t R1

p7
t R2 p R2

t Y2 p Y2

(b)

G2 )

t G2 p G2

Fig. 3.

(a) Two-phase signalized intersection; (b) Timed PN model of the intersection (a).

p c in p in

N 1,k :=N 1,k-1 +1 v1,k :=v(N 1,k )


1 2 n
s-1

N ns,k :=N ns,k-1 -1 vns,k :=v(N ns,k )


n
s

p c5 p5

in

N 1,k ,v1,k S 1,k R 1,k+1 Q 1,k

N ns,k ,vns,k
... ..

pS t out

[N 1,k <N max 1,k ]

S ns-1,k R ns-1,k+1 Q ns-1,k

Fig. 4.

Hybrid stochastic model of a onelane road link.

Fig.3-(b) shows also the timed PN model of a twophase trafc light signal cycle. A token in place pG2 means that vehicles coming from zone 8 meet the green light. The duration DG2 of the green phase is modelled by the timed transition tG2 that can be changed by the controller. After DG2 time units, the token moves to place pY2 (yellow phase) and remains there for a xed time; afterwards, the token moves to place pR2 in which all trafc light meet for a brief duration (modelled by timed transition tR2 ) the red light. Subsequently the token moves in place pG1 and the green light is turned on for vehicles coming from zones 4 and 5 (transition t2 is disabled whereas transitions t1 and t3 are enabled). In addition, the model uses also self loop arcs graphically represented by dotted lines they have the function of test arcs, to test the presence of tokens in a place. By changing the duration of the green phases (DG1 and DG2 ) it is possible to control the signalized intersection. B. Physical model of a road link As for a road link we adopt a hybrid stochastic model, called physical model of a link [4]. Since we are going to follow a hybrid approach, it is necessary to interface PNs and link models in a suitable way, as will be explained afterwards. A road link consists of a sequence of i = 1, ..., ns sections (see Fig.4). Each section corresponds to a stretch of a road where the behavior is fairly homogeneous. It receives vehicles from the upstream section and sends vehicles to the downstream section. Let Ni,k [veh/section length] be the number of vehicles within section i (i = 1, ..., ns ) (length Li [km] with li,k number of lanes) at instant kt. The average speed [km/h] of these Ni,k vehicles in section i at instant kt is denoted by vi,k . If we dene Qi,k as the number of vehicles crossing the boundary between two sections [veh] (i.e. Qi,k represents the number of vehicles leaving section i and entering the

section i + 1) during the time interval [kt, (k + 1)t), it follows that Ni,k+1 = Ni,k + Qi1,k Qi,k . Qi,k may be computed as follows Qi,k = min{Si,k , Ri,k }, (2) (1)

where the Si,k is called sending function and Ri,k receiving function. The sending function Si,k is a random variable expressing how many among the Ni,k vehicles in section i at instant kt are at a distance less than vi,k t from the boundary between section i and section i + 1: Si,k = max Ni,k vi,k t v i out t + Si ,k , Ni,k min Li Li (3)

i out where Si ,k is a zero mean random noise and vmin is the minimum velocity of vehicles leaving

i-th section. Speed vi,k depends on Ni,k , with a certain noise added: vi,k = v(Ni,k ) + vi ,k .
2000

(4)

1500 flow [veh/h]

1000

500

20

40

60

80 100 density [veh/km]

120

140

160

180

50 40 speed [km/h] 30 20 10 0

20

40

60

80 100 density [veh/km]

120

140

160

180

Fig. 5.

Fundamental diagrams ow-density and speed-density used in road link models.

Since the ow qi,k in section i may be computed as qi,k = i,k vi,k li,k , if we assume the fundamental diagram ow-density1 as in Fig.5, the diagram speed-density can be easily deduced;
1

Actually, this fundamental diagram in which the ow initially grows up linearly until a maximum value, after that it remains

more or less constant and then it decreases, is obtained by real data about freeway trafc [4]. Since so far we do not have real data about urban trafc we assume the same behavior in urban trafc.

hence, being i,k = Ni,k /(Li li,k ), speed vi,k , as a function of Ni,k , is derived. The receiving function Ri,k is a random variable which gives the maximum number of vehicles allowed to enter section i + 1 during the time interval [kt, (k + 1)t). If Al is the average vehicle length and td is the safe time distance between vehicles, Ri,k may be expressed as follows Ri,k = Li+1 li+1,k Ni+1,k + Qi+1,k . Al + vi+1,k td (5)

Each vehicle must be detected at least once in each section, during the time interval [kt, (k +
max 1)t), which can be expressed mathematically as vi t < Li . This condition determines an max upper bound for t < Li /vi . An algorithm for the computation of link state variable can be

found in [1]. C. Interaction between TPN components and road links The inow into a road link at its upstream section 1, and the outow from a road link at its downstream section ns are described by discrete time variables (Q0,k , v0,k ) and (Qns ,k , vns ,k ). The inow Q0,k however may be obtained as the point process counting the number of times a transition (e.g. tin ) res in a timed PN representing the intersection upstream of the road link. On the other hand, the Qns ,k vehicles leaving section ns trigger Qns ,k ring of a transition (tout in Fig.4) of the timed PN representing the intersection downstream of the road link. We have implemented this idea in an algorithm simulating trafc in a large network as follows. At each state update time (k + 1)t of the discrete time model, we calculate N1,k+1 = N1,k Q1,k Nns ,k+1 = Nns ,k + Qns 1,k . (6) (7)

The incoming ow at the section 1 is given by the ring of transition tin (see Fig.4); each time transition tin res (a vehicle enters section one), the associated action2 (N1,k = N1,k1 + 1, v1,k = v(N1,k )) is executed. On the other hand, the outgoing ow from section ns is given by the ring of transition tout ; each time transition tout res (a vehicle leaves section ns ), the associated action (Nns ,k = Nns ,k1 1, vns ,k = v(Nns ,k )) is executed.
2

An action, associated to a transition, here is simply dened as a set of mathematical expressions used to update the internal

state executed each time the transition res.

y in 2

out y2

yiin

y out j

xi j 1 x23 x21
out y1 in y1

x2 j

xi N xi j +1

x12 x1 j x1N x N1

xN j xN N 1

out yN in yN

Fig. 6.

Generalized signalized intersection

III. E VALUATING

ROUTING RATES

In this section the evaluation of routing rates of urban transportation network (UTN) intersections is discussed. An UTN is a couple {I, R} where I is the set of the signalized intersections in the network and R is the set of the road links. A signalized intersection is an element connecting two or more road links, and including the physical area occupied by vehicles crossing the intersection. In words, a road link is an element connecting two adjacent signalized intersections. The ows of vehicles coming from incoming directions are controlled by trafc lights.

y1out y1in y1in b in y1 a

x1b 3 x1b 2 x1a1

Fig. 7.

Laned signalized intersection

In Fig.6 a generalized signalized intersection is represented. It consists of N incoming di-

rections, each one with a trafc light, and N outgoing directions. In general, the number of incoming directions and the number of outgoing directions are different, because some roads connected by the intersection can have only a unique specied direction (incoming or outgoing).
in out The incoming (outgoing) ows yi (yj ), for i = {1, . . . , N} (j = {1, . . . , N}) are obtained

by measuring at the edges of the intersection the number of vehicles arriving (departing) in each trafc light cycle. The ows xij , with i = j, represent the number of vehicles starting from the incoming direction i and arriving to outgoing direction j, in each trafc light cycle, like shown in Fig.6. They are related to the intentions of the driver to reach a specied outgoing direction. The routing rates (ij ) are dened by the following relation: ij = xij N j=1,i=j xij , (8)

by considering the ratio of the number of vehicles coming from the incoming direction i, and going towards the outgoing direction j, w.r.t. the total number of vehicles from the incoming direction i in each trafc light cycle. The ows xij are evaluated and then by the equation (8), the routing rates ij are obtained. To compute the ows xij , it is necessary to consider the balance between the total ow of the vehicles arriving into the intersection and the total ow of the vehicles leaving the intersection. In each trafc light cycle, consider the i-th incoming direction; an equation, matching the ow
in yi with the sum of the vehicles leaving the i-th direction, can be written.

Similarly, consider the j-th outgoing direction; an equation, matching the total ow of the
out vehicles yj with the sum of the vehicles which arrives from other directions, can be written.

composed by N equations for incoming directions and N equations for outgoing directions. It can be easily proved that system (9) has at most 2N 1 equations linearly independent, since the total number of vehicles arriving into the intersection is equal to the total number of vehicles leaving the intersection. Furthermore, system (9) has at most N(N 1) unknown variables (xij ), since for each incoming direction it is possible to arrive to each outgoing direction by excluding itself. In the

Hence, the following system of equations can be obtained: N x in j=1 ij,i=j = yi i = 1 . . . N N xij,i=j = y out j = 1 . . . N
i=1 j

(9)

real signalized intersections, the number of unknown variables is usually less than the N(N 1), since for each incoming direction not all the outgoing directions are allowed, but however their number is grater than the number of equations. This approach can be used also with laned signalized intersections. This situation is shown in
in in in Fig.7, where the incoming ow yi is split in many ows (yia ,yib ), as many as the number of out lanes, and the same assumption is considered for the outgoing ows yj .

The system of equations (9) can be written in matrix form Ax = y, in which the value of the generic element aij of the matrix A is 1 whether it exists a ow xij that contributes to the
in out total incoming/outgoing ow (yi /yj ), 0 otherwise. The column vector of unknown variables x

is (x12 , x13 , . . . , x1N , . . . , xN 1 , xN 2 , . . . , xN N 1 ). The column vector y of the incoming/outgoing


in in in out out out ows measured at the edges of the intersection is (y1 , y2 , . . . , yN , y1 , y2 , . . . , yN ).

As it has been discussed above, the matrix A is usually rectangular with size m n, with m = 2N and n N(N 1). Notice that an intersection is dened only when N 3. Since the number of ows xij to be evaluated is usually greater than the number of the equation, there are innite solutions for the ows vector x. A possible choice is to select the best solution in last square sense by the equation x = A y r where A is the right pseudo-inverse matrix of A, dened as follows: r A = AT (AAT ) . r
1

(10)

(11)

The n m degrees-of-freedom (DOFs) of the system can be used in order to consider an additional condition to evaluate the solution. In our approach, the routing rates are evaluated on-line every trafc light cycle. A possibility is to use the DOFs of the system to minimize the distance in mean square sense between the solution of the system (9) (at step k) and the values obtained at previous evaluation (at step k 1), since it can be assumed that between two consecutive algorithm steps no big differences can occur in routing rates values. In words, the routing rates dynamic is slower than the sampling time of the algorithm. Therefore to minimize the norm of the vector xk xk1 , where xk (xk1 ) is the column vector of ows xkij (xk1ij ) evaluated at the current step k (at

the previous step k 1), the Lagrange multiplier method can be used. Let us consider the cost function given by: 1 g(xk ) = (xk xk1 )T (xk xk1 ), 2

(12)

by adding the constrain given by the matrix form of the system (9), the cost function modied is obtained: 1 g (xk , ) = (xk xk1 )T (xk xk1 ) + T (y k Axk ) 2

(13)

where is the unknown vector of Lagrange multiplier that allows to include the constrain in the cost function. The solution must satisfy these conditions: g xk
T

=0

= 0,

(14)

those lead to the solution of the system, that respects the new condition about the minimum mean square error, given by: xk = A y k + (I A A)xk1 r r

(15)

where I is the identity matrix, (I A A) is the operator that projects the previous solution r xk1 in the null space of the matrix A and y k is the column vector of incomimg/outgoing ows measured at the step k. After evaluating the ows xij by (15), the routing rates are obtained by (8). Usually, the routing rates ij , which are used to determine the duration of the trafc light phases in an intersection, are evaluated on the basis of the ows of the vehicles for a long time, often on the annual mean ows. They result to be constant in a certain time of the day. Denote ij these constant routing rates. In the proposed algorithm such long-time routing rates are used are used at the initial step of routing rates computation. Let be ij for all i and j, the routing rates evaluated on the long time mean ows, the reference solution in terms ow of vehicles in a trafc light cycle (ij ) must be computed considering the x following relation:
in xij = yi ij .

(16)

in out Given all incoming/outgoing ows yi /yj measured at the edges of the intersection and the

reference solutions ij , algorithm 1 computes at every trafc light cycle the actual routing rates ij . It is assumed that x0ij = xij , i.e. the routing rates evaluation procedure is initialized to the long time mean ows. Algorithm 1: Routing rates computation.
(* Algorithm initialization *) x0 = x; k:=1; (* Computation of actual ows xkij *) while 1 do begin xk = A y k + (I A A)xk1 ; r r (* Computation actual routing rates kij *) for i=1 to N do for j=1 to N do begin kij = xkij / end end k:=k+1; end
N j=1

xkij

An arithmetic mean of the routing rates computed by previous algorithm is considered including an arbitrary number of previous cycles, in order to reduce the variability of the obtained solution. It is worth to take into account the relative percentage error between the generic routing rate ij computed in on-line mode by previous algorithm and the corresponding long time reference ij . The maximum relative percentage error is quite often obtained w.r.t. the smallest reference routing rates solution. To overcome this problem a weighted right pseudo-inverse matrix AW r = W 1 AT (AW 1 AT ) . can be considered.
1

(17)

Hence another solution is given by xk = A r y k + (I A r A)xk1 W W

(18)

with W = diag{w1 , . . . , w7 }, where wi represents the weight of the norm of i-th component of the distance vector xk xk1 . Therefore the coefcients wi have to be chosen in order to reduce the relative percentage errors w.r.t. the smallest components of the long time reference solution x. If the number of ows xij to be evaluated is minor than the number of the equation, the same approach can be used to minimize the error in last square sense w.r.t. the measured ows since there is not a solution for the ows vector x. A possible solution in last square sense is given by: x = A y l where A is the left pseudo-inverse matrix of A, dened as follows: l A = (AT A) AT . l Since this is a rare case, it is not considered in the following for the sake of brevity .
out 1 Zone 1 Link 1 Zone 2 in 5

(19)

(20)

out 2 in 1

Link 2 Link 3 Link 5

Link 6 Link 7
STOP

in 4 out 4

Intersection 1 Link 4 in 2 in 3

On-Ramp

Intersection 2 out 3

Fig. 8.

Example of an urban area to be controlled.

IV. P ROBLEM

FORMULATION IN URBAN AREA CONTROL

The main aim in trafc control is the improvement of the network performance, that can be interpreted in many ways, and for each interpretation a different objective function can be formulated. In this paper the objective is to minimize the average number of vehicles in the urban area over K trafc lights cycles, since it is directly related to the average travel time for all vehicles. Hence, the objective function is 1 J= K
K

n(k)
k=1

(21)

where n(k) is the total number of vehicles in the urban area at trafc light cycle k. Let us consider the following place partition in the net model P = P c P , P c P = . where P c represents the set of control places and P represents the set of places having a physical meaning (i.e. the presence of a token in place p P corresponds to the presence of a vehicle in a certain zone of the network). We denote as |mk (p)| the size (number of tokens at instant k) of the place p that represent a discretization intersection area.
j Moreover, we denote as Ni,k the number of vehicles present in the section i of the road link

j at the instant k. By assuming that nl is the number of road links in the considered urban area and nsj is the number of section of j-th links, n(k) may be expressed as follows
nl nsj j Ni,k . j=1 i=1

n(k) =
pP

|mk (p)| +

(22)

Finally, our objective function has the following form 1 J= ( K


nl nsj j Ni,k ). j=1 i=1

|mk (p)| +
pP

(23)

V. C ASE

STUDY

In this work we apply to the urban area in Fig.8 an algorithm to control trafc lights switching times. The algorithm is based on the actual hybrid model state and on the results obtained by forward simulations like in [3]. The urban area consists of two signalized intersections connected to each other with road links having also an on-ramp area. All the links consist

NORDout

OVESTout OVESTin

ESTin ESTout

SUDin

Fig. 9.

Detailed scheme of Intersection 1.

of two sections except link 6 which consists of four sections. All the sections have the same
max length, Lsection = 200 m. All the links have one lane and vi = 50 km/h. The picture

shows, just before the intersections and between link 5 and link 7, also some grey squares which represent the discretization urban area made for the connection between continuous road link model and discrete intersection model. All the parameters used are realistic but, so far, the model validation has been carried out only in simulation. The simulations are performed in Matlab/Simulink/Stateow integrated environment. To generate input vehicles at the inputs of the urban area an Arrival Poisson Process is used. The strategy of the algorithm used to control the urban area is described in the sequel. If the number of vehicles in the rst section of the link 7, just after the on-ramp, or the number of vehicles of the second section of the link 7, just before intersection 2, is greater than 75% of the respective maximum section capacity, two forward simulations are performed. Each forward simulation has a duration equal to K = 5 trafc light cycles. In the rst forward simulation the duration of the green light duration of the trafc light cycle at the end of the link 7 is 55 seconds while in the second one the green light duration of the same trafc light cycle is 45 seconds. The duration of green light of the trafc light cycle at the end of link 7 of the urban area is 55 seconds if the value of the objective function computed by the rst forward simulation plus a certain threshold is less than the value of the objective function computed by the second forward simulation, otherwise it is equal to 45 seconds. This control strategy depends on the

actual system state and on the expected system state given by the forward simulations. Many simulations are performed in order to investigate about the choice of the threshold between the two forward simulations and about the choice of the duration of the forward simulations. In this work all the simulations have been carried out by evaluating the routing rates for the forward simulations in adaptive way by using the methodology shown in Section III. In [3] the threshold between the objective functions of the forward simulations was xed to 5 veh and a mean value over twelve simulations of J tr=5 = 144, 98 veh is obtained. In this work, a rst series of twelve simulations has been carried out with threshold xed to 1 veh by obtaining J tr=1 = 137, 70 veh. Thus, it results an improvement of 5.02%. This improvement is related to the fact that the control strategy is applied to the urban area also when there is a little variation of the objective function between two forward simulations. A second series of twelve simulations has been carried out with the threshold xed to 5, but the routing rates are unknown and are evaluated on-line. The objective function with all parameters known is always J know = 144, 98 veh, while in the last case is J unknow = 148, 36 veh, the performances of the system are reduced of 2.33%. The loss of performance can be explained by the initial gap between the real value of the routing rates and their estimation. To investigate about the duration of the forward simulations, a third series of of twelve simulations has been carried out, all with threshold xed to 1 veh, but with a different duration of simulation. Such simulations have been performed with a forward simulations of 3, 5, 7, and 9 trafc light cycles, by obtaining respectively the following mean values of the objective functions: J K=3 = 158, 20 veh, J K=5 = 137, 70 veh, J K=7 = 139, 81 veh, and J K=9 = 133, 78 veh. It is worth to notice that the objective function is reduced of 12.95% if K is increased from 3 to 5, it is increased of 1.53% if K is increased from 5 to 7, and there is a reduction of the objective function of 4.31% if K is increased from 7 to 9. This non-monotonic behavior with respect to K is due to the stochastic and complex nature of the system. The choice of the duration of the forward simulations has to be done also with respect to to the computational time. It must be possible to execute the forward simulations during a real trafc light cycle, without introducing a delay in the control response of the system. Therefore the duration of the forward simulations is a trade-off between improvement of the performances of the system and computational time. The computational time of the forward simulation over K = 5 trafc light cycles (including the routing rates computation) is less then 25 sec on a 3.00 Ghz Pentium IV processor based PC

and then it could be executed during a real trafc light cycle, without introducing a delay in the control response of the system. The computational time of the forward simulation over K = 9 trafc light cycles is 80 sec on the same machine. By analyzing previous simulation results, it possible to conclude a good choice for the for the duration of forward simulation is 5 trafc light cycles. VI. F UTURE
WORKS

The research continue by a validation of the model, based on urban trafc real data. Also the development of more complex what-if algorithms to decide the optimal trafc signal plan is an important task to be carried out. R EFERENCES
[1] C. C. Avram, F. Basile, R. K. Boel, C. Carbone, and P. Chiacchio. A hybrid model for urban trafc control. IEEE International Conference on Systems, Man, and Cybernetics (SMC04), The Hague, Netherlands, pages 1795 1800, October 2004. [2] F. Azadivar. Simulation optimization methodologies. Proceedings of the 1999 Winter Simulation Conference, pages 93 100, 1999. [3] F. Basile, P. Chiacchio, and D. Teta. Simulation based control of urban trafc based on a hybrid model. ANIPLA 2006, Int. Conf. on Methodologies for Emerging Technologies in Automation, Rome, Italy, November 2006. [4] R.K. Boel and L. Mihaylova. A compositional stochastic model for real time freeway traic simulation. Transportation Research Part B, 40:319334, 2006. [5] F. Boillot, J.M. Blosseville, J.B. Lesort, V. Motyka, M. Papageorgiou, and S. Sellam. Optimal signal control of urban trafc networks. Proc. of the 6th IEE Int. Conf. Road Trafc Monitoring and Control, pages 7579, 1992. [6] E.F. Camacho and C. Bordons. Model predictive control in the process industry. Springer-Verlag, Berlin, Germany, 1995. [7] M. Dotoli and M.P. Fanti. An urban trafc network model via coloured timed Petri nets. Control Engineering Practice, 14(10):12131229, October 2006. [8] A. Di Febbraro, D. Giglio, and N. Sacco. Urban trafc control strcture based on hybrid petri nets. IEEE Trans. on Intelligent Transportation Systems, 5(4):224 237, December 2004. [9] A. Di Febbraro and N. Sacco. On modelling urban transportation networs via hybrid Petri nets. Control Engineering Practice, 12(4):1225 1239, 2004. [10] Andrtas Hegyi. Model predictive control for integrating trafc control measures. PhD thesis, Technische Universiteit Delft, Belgium, 2004. [11] T. Murata. Petri nets: Properties, analysis and applications. Proceedings of the IEEE, 77(4):541 580, April 1989. [12] Markos Papageorgiou, Christina Diakaki, Vaya Dinopoulou, Apostolos Kotsialos, and Yibing Wang. Review of road trafc control strategies. Proc. of the IEEE, 91(12):20432067, 2003. [13] I. Porche and S. Lafortune. Dynamic trafc control: decentralized and coordinated methods. IEEE Conference on Intelligent Transportation System (ITSC 97), pages 930 935, November 1997.

[14] I. Porche and S. Lafortune. Dynamic trafc control: decentralized and coordinated methods. Proc. of the 1997 IEEE Conference on Intelligent Transportation System (ITSC 97), pages 930935, 1997. [15] I. Porche and S. Lafortune. Coordination of local adaptive trafc signal controllers. Proc. of the 1998 American Control Conference (ACC98), pages 18331837, 1998. [16] I. Porche, M. Sampath, Y.-L. Chen, and S. Lafortune. A decentralized scheme for real-time optimization of trafc signals. Proc. of the 1996 IEEE Int. Conf. on Control Applications (CCA96), pages 582589, 1996. [17] Th. Riedel and U. Brunner. Trafc control using graph theory. Control Eng. Practice, 2(3):397404, 1994.

Anda mungkin juga menyukai