You are on page 1of 14

1 OPTIMAL DRIVING OF CONNECTED VEHICLES AT TRAFFIC LIGHTS

2
3
4
5 Marko Woelki, Corresponding Author
6 Institute of Transportation Systems, German Aerospace Center (DLR)
7 Rutherfordstr. 2, 12489 Berlin, Germany
8 Tel: +49/30/67055-710; Fax: +49/30/67055-289; Email: marko.woelki@dlr.de
9
10 Daniel Krajzewicz
11 Institute of Transport Research, German Aerospace Center (DLR)
12 Rutherfordstr. 2, 12489 Berlin, Germany
13 Tel: +49/30/67055-273; Fax: +49/30/67055-283; Email: daniel.krajzewicz@dlr.de
14
15 Andreas Leich
16 Institute of Transportation Systems, German Aerospace Center (DLR)
17 Rutherfordstr. 2, 12489 Berlin, Germany
18 Tel: +49/30/67055-409; Fax: +49/30/67055-289; Email: andreas.leich@dlr.de
19
20
21 Word count: 5,723 words text + 7 tables/figures x 250 words (each) = 7,473 words
22
23
24
25
26 Submission Date: 1st of August 2015

TRB 2016 Annual Meeting Paper revised from original submittal.


Woelki, Krajzewicz, Leich 2

1 ABSTRACT
2 This article considers different driving patterns at a traffic light (TLS). The motivation is to
3 determine and improve the behavior of traffic flow at controlled intersections, given the
4 availability of the so-called “EcoDriving” (USA) or “GLOSA” (Europe) application. This is an
5 advanced driver assistance system that tells the driver which speed to choose in order to pass the
6 oncoming traffic light at green. The speed is computed using information sent by the traffic light.
7 When entering the communication range of the TLS vehicles decide at what time they will pass the
8 halting line and follow that plan. The models of consideration either comprise speed or
9 acceleration advisory or no advice at all, the latter in case they are not equipped with a
10 communication device. Additionally, the cases of cooperative and non-cooperative driving across
11 the intersection are taken into account; that is either closing the gap to the car in front or simply
12 following a velocity that will lead to an arrival at green light. Finally, different extensions of the
13 situation are considered: smaller random demand and a finite penetration rate. Safety criteria
14 (brake gap) are fulfilled in any of these cases. This article presents simulation results and answers
15 the question which driving advice strategy is best with regard to different criteria as travel time and
16 emissions.
17
18
19
20 Keywords: emission reduction, GLOSA, V2X
21

TRB 2016 Annual Meeting Paper revised from original submittal.


Woelki, Krajzewicz, Leich 3

1 INTRODUCTION
2 Optimizing the traffic flow at signalized intersections with regard to criteria like minimal waiting
3 times, maximal throughput or minimal emission of pollutants, is a standard task in traffic
4 engineering (1). A reasonable approach to solve this task is to optimize the green, red and cycle
5 times according to the numbers of vehicles (traffic flows) passing the TLS (2). The inverse
6 approach is being investigated in this paper: constant green and cycle times with varying demand
7 and driving patterns. Special attention is paid to the vehicles’ trajectories. Driving behavior has a
8 high influence on energy consumption and emissions. In urban scenarios with speed limits below
9 80km/h, traffic lights have a high impact on pollutant emissions. Traffic lights can force a vehicle
10 to decelerate or even halt at the stop line, thus partly or even fully dissipating its kinetic energy.
11 Thereby, regarding the approach of vehicles towards a signalized intersection, it is of interest to
12 optimize certain vehicle parameters with regard to the perspectives of a) the individual driver, b)
13 the traffic management that has a global traffic system perspective, and c) the ecological point of
14 view. In the following, time efficiency of the drivers as well as emission reduction are considered.
15 One field of application for the results presented here is the optimization of a speed advice
16 as given to the driver by the application named “EcoDriving” in the U.S. or “GLOSA” (“Green
17 Light Optimal Speed Advisory”) in Europe. The concept of GLOSA is to send messages from the
18 TLS signal controller via a road side unit to the vehicles via wireless communication infrastructure
19 (I2V). The two I2V message types used for this purpose are standardized. The first, Signal Phase
20 and Timing (SPaT) (3) contains information about the current and future signal timings while the
21 road infrastructure ahead is described using INFRA messages. Given this information, GLOSA
22 can compute the speed to pass the next traffic light at green. This speed is given as an advice to the
23 driver.
24 GLOSA’s major scope is to increase traffic comfort by reducing accelerations and
25 decelerations but is as well assumed to increase traffic efficiency and to reduce emissions (4).
26 Whether or not it can be effective in doing so is discussed controversial in the literature. While in
27 (5) results of simulation and real world tests show a CO2 reduction in the order of magnitude of
28 10%, in (6) adverse effects in congested situations and even higher CO2 emissions for unequipped
29 vehicles were reported. In (4), simulation results are presented where GLOSA plus automatic
30 engine start-stop control yield up to 20% less emissions at city scale. In (7), it is shown that
31 emissions can rise in selected cases where communication range is beyond 500 m. In fact, this can
32 only happen if the advised speed is not chosen properly. (7) is thereby a direct motivation for the
33 work presented here.
34 This paper is a contribution to this discussion based on a new approach. Simulated vehicle
35 trajectories have been analyzed and benchmarked in (8), where three models from the literature
36 and two new ones were evaluated. In the setup used in (8), only one vehicle’s trajectory is
37 investigated at once. In principle, this corresponds to the extreme case of a vanishing deterministic
38 demand. In the following, a finite demand and especially the other extreme of a maximal
39 deterministic demand is considered: one vehicle every two seconds. Several driving strategies at a
40 signalized intersection are analyzed with regard to flow maximization and minimization of
41 emissions. In contrast to the majority of other investigations in the field, the key idea of the
42 approach presented here is that vehicles reserve a time-slot at which they cross the intersection.
43 This slot is computed using the current position and velocity, the planned arrival time of the
44 vehicle in front and the signaling state of the traffic signal.
45 The research presented in this article was conducted within the COLOMBO project (9, 10),
46 which was co-funded by the European Commission. The project’s major scope was to develop and
47 evaluate methods for traffic surveillance and traffic light control assuming low penetration rates of

TRB 2016 Annual Meeting Paper revised from original submittal.


Woelki, Krajzewicz, Leich 4

1 vehicles equipped with V2X technology. Environmental issues played a fundamental role in the
2 project. The research presented here is based on a new state-of-the-art microscopic vehicular
3 pollutant emission model PHEMlight (11). PHEMlight was derived from the instantaneous vehicle
4 emission model PHEM (Passenger Car and Heavy Duty Emission Model). PHEM itself is based
5 on an extensive European set of vehicle measurements and has been validated against the
6 emissions of these real-world vehicles. PHEMlight has been validated against PHEM, see (11).
7 This paper is structured as follows: the models, formulas and concepts for this
8 investigation are explained in the first section. Dedicated subsections deal with the traffic model
9 equations and the simulation scenario, the driving patterns covered by this study, the presentation
10 of simulation results illustrated with trajectory plots, and the presentation of emission results
11 corresponding with the different driving patterns. The second section contains consideration of the
12 vanishing demand. The performance at a finite penetration rate is discussed afterwards, followed
13 by an investigation of a stochastic driver behavior. The paper closes with conclusions.
14
15 COMPARISON OF DRIVING PATTERNS AT A TRAFFIC LIGHT
16 The scenario used in this section consists of a single-lane, unidirectional road with a traffic light at
17 position 𝑥𝑥 = 0𝑚𝑚. Vehicles are enumerated as 𝑖𝑖 = 0,1, … The minimal gap between vehicles is
18 2.5𝑚𝑚 and the car length is taken as 5𝑚𝑚 so that the minimal distance of the back ends of
19 consecutive vehicles is 𝐷𝐷 = 7.5𝑚𝑚. The maximum possible acceleration is 𝑎𝑎𝑚𝑚𝑚𝑚𝑥𝑥 = 1𝑚𝑚/𝑠𝑠² and the
20 maximum possible deceleration is 𝑏𝑏𝑚𝑚𝑚𝑚𝑥𝑥 = 4.5𝑚𝑚/𝑠𝑠². For vehicles being able to receive
21 information from the TLS, there is a certain interaction range, meaning that the vehicle starts to
22 interact at position 𝑥𝑥1 . Since it might happen that the desired speed for the approach to the halt line
23 cannot be reached instantaneously, a whole range for speed adjustment is given. This is of the
24 order of the braking distance from maximum speed which is 𝑣𝑣2𝑚𝑚𝑚𝑚𝑥𝑥 /(2𝑏𝑏𝑚𝑚𝑚𝑚𝑥𝑥 ) ≈ 21.5𝑚𝑚. The traffic
25 light has a cycle time of 𝐶𝐶 = 60𝑠𝑠, green light starts at 𝑡𝑡 = 𝑛𝑛𝑛𝑛 (𝑛𝑛 = 0,1, … ) and ends at 𝑡𝑡 = 𝑛𝑛𝑛𝑛 +
26 𝐺𝐺, with the duration 𝐺𝐺 = 25𝑠𝑠. This green time period is followed by a yellow time of 𝑌𝑌 = 3𝑠𝑠 so
27 that 𝑅𝑅 = 32𝑠𝑠 remain for the red time in each cycle.
28
29 Microscopic traffic model and scenario
30 The used car-following behavior is based on the microscopic traffic model introduced in (12). It is
31 discrete in time with a time step of one second. It is continuous in space. At each time step, the
32 velocity 𝑣𝑣𝑖𝑖 of a vehicle 𝑖𝑖 is adapted depending on the velocity 𝑣𝑣𝑖𝑖−1 of the vehicle in front and the
33 gap 𝑔𝑔𝑖𝑖 in between the two vehicles according to the safety condition:
34
35 𝑑𝑑(𝑣𝑣𝑖𝑖 ) + 𝑣𝑣𝑖𝑖 𝑇𝑇𝑖𝑖 ≤ 𝑑𝑑(𝑣𝑣𝑖𝑖−1 ) + 𝑔𝑔𝑖𝑖 (1)
36
37 Here, 𝑇𝑇𝑖𝑖 = 𝑇𝑇 = 2𝑠𝑠 is the preferred headway. Although in ref. (12) this time was considered
38 as reaction time it is more realistic to consider a preferred headway. An additional argument for the
39 two seconds is that in Germany it is a common rule to ensure a headway of two seconds. This
40 allows for comfortable braking and collision free dynamics. Further, 𝑑𝑑(𝑣𝑣) is the braking distance
41 𝑑𝑑(𝑣𝑣) = 𝑣𝑣𝑖𝑖2 /(2𝑏𝑏) with the comfortable deceleration 𝑏𝑏 = 0.8𝑚𝑚/𝑠𝑠 2 . This is assumed to be more
42 realistic than using the maximum deceleration. Solving this safety condition leads to the safe speed
43
44 𝑣𝑣𝑖𝑖,𝑠𝑠𝑠𝑠𝑠𝑠𝑒𝑒 (𝑡𝑡) = −𝑏𝑏𝑇𝑇𝑖𝑖 + �𝑏𝑏2 𝑇𝑇𝑖𝑖 2 + 𝑣𝑣𝑖𝑖−1 2 (𝑡𝑡) + 2𝑏𝑏𝑔𝑔𝑖𝑖 (𝑡𝑡) (2)
45
46 Vehicles start deterministically at maximum speed 𝑣𝑣𝑚𝑚𝑚𝑚𝑥𝑥 = 13.89𝑚𝑚/𝑠𝑠 from their initial

TRB 2016 Annual Meeting Paper revised from original submittal.


Woelki, Krajzewicz, Leich 5

1 position. Here, for vehicle 𝑖𝑖 = 0 the starting position 𝑥𝑥0 = −500𝑚𝑚 is assumed. The other vehicles
2 follow the first vehicle at safety distance:
3
4 𝑥𝑥𝑖𝑖 (0) = 𝑥𝑥0 − 𝑖𝑖(𝐷𝐷 + 𝑣𝑣𝑚𝑚𝑚𝑚𝑥𝑥 𝑇𝑇) (3)
5
6 Note that for reasons of computer-memory efficiency we cut all trajectories at position 𝑥𝑥0
7 (as if all vehicles would start from 𝑥𝑥0 in timewise succession). The position of vehicles is updated
8 until the final position 𝑥𝑥𝑓𝑓 is reached according to
9
10 𝑥𝑥𝑖𝑖 (𝑡𝑡 + 1) = 𝑥𝑥𝑖𝑖 (𝑡𝑡) + min( 𝑣𝑣𝑖𝑖,𝑑𝑑𝑑𝑑𝑠𝑠 (𝑡𝑡), 𝑣𝑣𝑖𝑖,𝑠𝑠𝑠𝑠𝑓𝑓𝑓𝑓 (𝑡𝑡)) (4)
11
12 Thus the driver decides on one of two velocities. 𝑣𝑣𝑖𝑖,𝑠𝑠𝑠𝑠𝑓𝑓𝑓𝑓 according to eq. (2) ensures a safe
13 speed with regard to the vehicle in front. 𝑣𝑣𝑖𝑖,𝑑𝑑𝑑𝑑𝑠𝑠 depends on the driving strategy. It ensures that the
14 traffic light is not passed at red and may additionally be such that the vehicle drives as time and
15 emission effective as possible:
16
𝑣𝑣𝑚𝑚𝑎𝑎𝑥𝑥 , 𝑥𝑥0 ≤ 𝑥𝑥(𝑡𝑡) < 𝑥𝑥1 ,
17 𝑣𝑣𝑑𝑑𝑑𝑑𝑠𝑠 𝑣𝑣
= 𝑣𝑣(𝑡𝑡) + max. acc./dec. towards � 𝑚𝑚𝑚𝑚𝑑𝑑𝑑𝑑𝑑𝑑 , 𝑥𝑥1 ≤ 𝑥𝑥(𝑡𝑡) < 0, (5)
𝑣𝑣𝑚𝑚𝑚𝑚𝑥𝑥 , 0 ≤ 𝑥𝑥(𝑡𝑡) < 𝑥𝑥𝑓𝑓 .
18
19 As one can see, all strategies have in common that outside of the interaction range of the
20 TLS drivers try to be as fast as possible. In the next section several strategies will be considered for
21 𝑣𝑣𝑚𝑚𝑚𝑚𝑑𝑑𝑑𝑑𝑑𝑑 = 𝑣𝑣𝑜𝑜𝑜𝑜𝑡𝑡 (𝑥𝑥, 𝑡𝑡) ∈ [𝑣𝑣𝑚𝑚𝑚𝑚𝑛𝑛 , 𝑣𝑣𝑚𝑚𝑚𝑚𝑥𝑥 ]. It depends on the distance 𝑥𝑥 to the traffic light as well as the
22 time relative to the signal timings. It is definitely bounded by the maximum velocity and optionally
23 by a suitably chosen minimal velocity. The acceleration in eq. (5) has to be in the physical range
24 𝑐𝑐 ∈ [−𝑏𝑏𝑚𝑚𝑚𝑚𝑥𝑥 , 𝑎𝑎𝑚𝑚𝑚𝑚𝑥𝑥 ]. Note that the drivers may use the maximum deceleration 𝑏𝑏𝑚𝑚𝑚𝑚𝑥𝑥 in order to
25 adapt to the infrastructure-motivated speed 𝑣𝑣𝑚𝑚𝑚𝑚𝑑𝑑𝑑𝑑𝑑𝑑 , however braking due to the vehicle in front is
26 assumed to be smoother (with the comfortable deceleration 𝑏𝑏. Consequently, vehicles are able to
27 adapt to 𝑣𝑣𝑚𝑚𝑚𝑚𝑑𝑑𝑑𝑑𝑙𝑙 rather quickly within a certain range Δ𝑥𝑥 and without affecting following vehicles
28 too much. The follower always keeps enough distance and drives at the (comfortable) safe speed
29 (eq. 2). For other decelerations the traffic flow can quickly break down when the adjustment range
30 is not small enough compared to the distance to the TLS.
31 The preferred time headway of two seconds assumed in (eq. 3) implies that the maximum
32 possible demand (or saturation flow) in this model is 𝑞𝑞𝑚𝑚𝑚𝑚𝑥𝑥 = 1(2𝑠𝑠). For the maximum
33 demand, 𝑖𝑖 takes all values 1, … , 𝑁𝑁 − 1, where 𝑁𝑁 is the total number of cars. Thus all vehicles start
34 at safety distance for maximum speed in succession. However, if the demand is smaller than the
35 saturation flow, one can think of vehicles being removed at random from this initial situation.
36 Formally, this initial configuration of vehicles driving at maximum speed for a finite demand is
37 achieved as follows: A random binary sequence (𝑠𝑠0 , 𝑠𝑠1 , … , 𝑠𝑠𝑁𝑁−1 ) is drawn in which each
38 component is represented by 𝑠𝑠𝑖𝑖 = 1 if 𝑟𝑟𝑟𝑟𝑛𝑛𝑛𝑛(𝑖𝑖) < 𝑞𝑞 and by 𝑠𝑠𝑖𝑖 = 0 otherwise. Then only those
39 values of 𝑖𝑖 for which 𝑠𝑠𝑖𝑖 = 1 represent a vehicle (which starts at 𝑥𝑥𝑖𝑖 (0) as given by the formula
40 above). Thus, for large 𝑁𝑁, there are 𝑞𝑞𝑞𝑞 vehicles on the road on average. To be able to compare
41 results from different driving models directly, this vector 𝑠𝑠 is taken for all of these models. Note
42 that this choice of initial condition implies that each vehicle 𝑖𝑖 > 0 has an initial distance to the car
43 in front that is an integer multiple of the safety distance at maximum speed.
44

TRB 2016 Annual Meeting Paper revised from original submittal.


Woelki, Krajzewicz, Leich 6

1 Different driving patterns


2 Four different driving patterns are presented in the following. Their respective behavior is depicted
3 in Figure 1. Note that all driving patterns ensure collision free dynamics.
4 • WITHOUT ADVICE, SIMPLE (WS) MODEL. This is the simplest realization of
5 driving behavior without any communication with the traffic signal or other vehicles. Vehicles just
6 drive as fast as allowed and accelerate as strong as possible until this velocity is reached if needed
7 and possible. When approaching a red signal they decelerate as strongly as possible if the distance
8 is equal or larger than the brake gap. Otherwise they just drive through what implies that they may
9 pass over yellow light. However, as in the rest of the paper, the braking due to a vehicle in front is
10 considered as a comfortable one of 𝑏𝑏 = 0.8𝑚𝑚/𝑠𝑠². One sees that the throughput of the TLS at any
11 green time period is close to the maximum one: Since the green time is 𝐺𝐺 = 25𝑠𝑠 and the preferred
12 headway is 𝑇𝑇 = 2𝑠𝑠 , this number is 12-13 vehicles.
13 • NON-COOPERATIVE SPEED ADVICE (NS) MODEL. In the interaction range
14 (between the dotted lines in figure 1), vehicles try to achieve a constant velocity that will carry
15 them at green light over the intersection. It is assumed that the vehicle typically adapts to the
16 velocity of its lead vehicle as quickly as possible if this leads to an arrival at green light. One sees
17 that the velocity as well as the number of vehicles that can pass through green light decreases with
18 time. Thus the traffic flow quickly breaks down. At one stage there will be a completely immobile
19 jam, when the target velocity falls below the formal lower critical velocity bound of the algorithm.
20 • COOPERATIVE SPEED ADVICE (CS) MODEL combines the previous NS approach
21 with a cooperative behavior of vehicles. Vehicles always aim to catch up with the vehicle in front
22 in order to allow as many vehicles as possible to pass through the green time. This implies that they
23 do not keep their constant velocity until the halt line but accelerate before, as explained later. This
24 leads to maximum throughput of vehicles at the stop line. Formally this is achieved in the
25 following way: The first vehicle checks after entrance into the interaction range of the TLS
26 whether it arrives at green when keeping its velocity (or accelerating to the maximum speed, if not
27 yet reached). Otherwise it targets the time tR when the next switch from red to green takes place
28 and computes its velocity accordingly. After having passed the adjustment range, the vehicle has
29 reached its desired speed 𝑣𝑣𝑖𝑖,𝑑𝑑𝑑𝑑𝑠𝑠 . It simply follows this speed until reaching the halt line. Then it
30 accelerates again. All following vehicles have the same strategy but have to consider the vehicle in
31 front as well. In the adjustment range they plan their arrival time as follows: Due to the car in front,
32 the earliest safe time to arrive at the halt line is 𝑡𝑡𝑖𝑖,𝑎𝑎𝑎𝑎𝑟𝑟,𝑠𝑠𝑠𝑠𝑓𝑓𝑓𝑓 = 𝑡𝑡𝑖𝑖−1,𝑎𝑎𝑎𝑎𝑟𝑟,𝑠𝑠𝑠𝑠𝑓𝑓𝑓𝑓 + 𝑇𝑇. The earliest possible
33 time (bounded by maximum velocity and acceleration) where it can reach the halt line is
34
𝑣𝑣𝑚𝑚𝑚𝑚𝑥𝑥 −𝑣𝑣𝑖𝑖 (𝑡𝑡) |𝑥𝑥|−0.5𝑎𝑎𝑚𝑚𝑚𝑚𝑥𝑥 (𝑣𝑣𝑚𝑚𝑚𝑚𝑥𝑥 −𝑣𝑣𝑖𝑖 (𝑡𝑡))2
35 𝑡𝑡𝑖𝑖,𝑎𝑎𝑎𝑎𝑟𝑟,𝑚𝑚𝑚𝑚𝑛𝑛 = 𝑡𝑡 + + (5)
𝑏𝑏𝑚𝑚𝑚𝑚𝑥𝑥 𝑣𝑣𝑚𝑚𝑚𝑚𝑥𝑥
36
37 Hence, the candidate arrival time at the halt line is 𝑡𝑡𝑖𝑖,𝑎𝑎𝑎𝑎𝑟𝑟 = 𝑚𝑚𝑚𝑚𝑥𝑥(𝑡𝑡𝑖𝑖,𝑎𝑎𝑎𝑎𝑟𝑟,𝑠𝑠𝑠𝑠𝑓𝑓𝑓𝑓 , 𝑡𝑡𝑖𝑖,𝑎𝑎𝑎𝑎𝑟𝑟,𝑚𝑚𝑚𝑚𝑛𝑛 ). If
38 this means arriving at green, the vehicle reserves this time. Otherwise, the arrival time is shifted to
39 the next green time: 𝑡𝑡𝑖𝑖,𝑎𝑎𝑎𝑎𝑟𝑟 = 𝑡𝑡𝑅𝑅 . Then its desired speed 𝑣𝑣𝑖𝑖,𝑑𝑑𝑑𝑑𝑠𝑠 (𝑡𝑡) is set accordingly. After having
40 passed the interaction range two situations can occur: If the vehicle is unaffected by the vehicle in
41 front because the distance is sufficiently large, the vehicle follows constantly its desired speed.
42 However, if its trajectory is affected by the car in front, it typically follows the safe speed
43 𝑣𝑣𝑖𝑖,𝑠𝑠𝑠𝑠𝑓𝑓𝑓𝑓 according to the car-following model described in the previous section. In such cases, the
44 vehicle arrives at the earliest possible time 𝑡𝑡𝑖𝑖,𝑎𝑎𝑎𝑎𝑟𝑟 usually by accelerating between adjustment range
45 and the halt line when the vehicle in front does so as well. If a sequence of vehicles follows each
46 other at the safety distance this has the effect that vehicles start to accelerate increasingly further

TRB 2016 Annual Meeting Paper revised from original submittal.


Woelki, Krajzewicz, Leich 7

1 ahead of the halt line (as a consequence of the acceleration of the first car in this green-time period
2 when passing the halt line). This can be seen in Figure 1 for the case of maximal demand where all
3 following vehicles drive at safety distance before deciding onto their arrival time at the halt line.
4 • COOPERATIVE ACCELERATION ADVICE (CA) MODEL. Here, the vehicle brakes
5 slowly when entering the interaction range until it reaches its velocity minimum at the halt line
6 where it accelerates again. The mechanism of time reservation is the same as in the cooperative
7 speed advice model. Thereby, when the planned arrival time is located during yellow or red, this
8 time is raised to the next green switch at time𝑡𝑡𝑅𝑅 . In this case, the vehicle calculates the optimum
9 deceleration 𝑎𝑎∗ to arrive at this time at the halt line, which is
10
𝑥𝑥𝑖𝑖 +𝑣𝑣𝑖𝑖 𝑡𝑡𝑅𝑅
11 𝑎𝑎∗ = −2 𝑡𝑡𝑅𝑅 2
(6)
12
13 For further details, see (8).
14

15

16
17 FIGURE 1 Trajectories for 27 vehicles passing over an intersection. Top left: Without any
18 advice (WA). Top right: Non-cooperative speed advice (NS). Bottom left: Cooperative speed
19 advice (CS). Bottom right: Cooperative acceleration advisory (CA).
20
21 Emissions computing
22 As in (8), the different driving modes a) Cruising (at constant speed), b) Acceleration, c)
23 Deceleration, and d) Stop time are considered. For each of these modes (despite stopping), the
24 respectively optimal velocities and accelerations have been computed using the instantaneous

TRB 2016 Annual Meeting Paper revised from original submittal.


Woelki, Krajzewicz, Leich 8

1 emission model PHEM (13-15). Deceleration has to be decomposed into the driving states
2 coasting and braking. When coasting, a vehicle decelerates due to the air drag and the drag losses
3 of the engine and of the drivetrain system while the fuel injection is switched off yielding in zero
4 emissions. When braking actively, fuel is continued to be injected. Figure 2 shows how the driving
5 mode is determined using the vehicle's velocity and acceleration in the current simulation time
6 step.
7
8 # get the driving mode
9 # v: velocity of the vehicle(at time t)
10 # a: acceleration of the vehicle (at time t)
11 def getMode(v, a):
12 mode = UNKNOWN
13 # the vehicle has a constant speed or halts
14 if a<=0 and a>-0.01:
15 # halting if velocity is smaller than .1 m/s
16 if v<0.1: mode = HALTING
17 # driving at constant speed otherwise
18 else: mode = CONSTANT
19 # the vehicle is braking
20 elif a<0:
21 # coasting if acceleration is small and speed is above idling speed
22 if a<-.3 and v>2.78: mode = COASTING
23 # braking otherwise
24 else: mode = BREAKING
25 # the vehicle accelerates
26 else: mode = ACCELERATING
27 return mode
28 FIGURE 2 Identification of driving modes according to velocity v and acceleration a.
29
30 The following results are best displayed when encoding the different driving modes with
31 color values. For the rest of the paper the convention given in Figure 3 holds.
32

33
34 FIGURE 3 Association of driving modes with colors for the rest of the paper.
35
36 Increasing the number of vehicles gives a more and more statistically significant view at
37 the reliability of the different approaches. It was determined that taking 1800 vehicles running
38 successively into the intersection already represents the asymptotic behavior very well. This leads
39 to a minimum number of 150 green phases if the maximum number of vehicles per green time
40 period is 12. Figure 4 shows the results for two different regions of measurement, see also the
41 Figure’s caption. The second approach of non-cooperative speed advice is not considered any
42 further because it leads to growing jams. Thereby, only the three remaining models are shown in
43 the following.
44 It turns out that both, in the vicinity of the halt line (±100 m), and when starting to measure
45 from the front end of the initial stream of vehicles to the same end position +100 m, favors the
46 cooperative speed advice CS with regard to travel time and CO2 emissions. Obviously, the WA

TRB 2016 Annual Meeting Paper revised from original submittal.


Woelki, Krajzewicz, Leich 9

1 model produces more emissions than CA in the vicinity of the halt line due to the heavy braking
2 and accelerating in that region. The reason why in total CA is worse than WA is the following: In
3 CA vehicles are braking very often but within a range of deceleration that is not in the coasting
4 regime. Therefore in (8) an amended model was considered that applies coasting deceleration
5 whenever possible. As a conclusion one can state that continuous deceleration towards the halt line
6 is not necessarily better than driving at constant speed.
7

9
10 FIGURE 4 Comparisons of the different driving patterns at maximum demand (1 vehicle /
11 2s), that is 1800 vehicles inserted during one hour. Measurements took place in the vicinity of
12 the halt line between -100m and +100m (top) and from the initial position -500m to +100m
13 (bottom).
14
15 VANISHING DEMAND
16 While in the previous section a demand of one vehicle every two seconds over a time period of one
17 hour has been considered, this section deals with smaller demands. Figure 5 shows the results for a
18 demand of 0.01 vehicles per second over a period of four hours. One may think of 7200 initially
19 consecutive placeholders each of which is occupied by a vehicle with a probability of 1 percent.
20 As can be seen in Figure 5, CO2 and PMx emissions as well as fuel consumption are lowest
21 for the CA model. The finding that CA is best for very small demand of 0.01 vehicles per second is
22 in line with the results from (8) where only single-vehicle trajectories were considered.
23
24
25

TRB 2016 Annual Meeting Paper revised from original submittal.


Woelki, Krajzewicz, Leich 10

2
3 FIGURE 5 Number of modes, CO2 and PMx emissions, as well as fuel consumption for the
4 different driving patterns at a demand of 0.01 vehicles per second being inserted randomly
5 during 7200 trials. Emissions were measured from the initial position -500 m to +100 m.
6
7 FINITE PENETRATION RATE
8 In this section, a finite equipment rate of pequip is considered. That means that a fraction 𝑝𝑝𝑒𝑒𝑒𝑒𝑢𝑢𝑢𝑢𝑢𝑢 of
9 the vehicles reserve their arrival time at the halt line beforehand while the remaining vehicles drive
10 according to the WA model. In order to keep things as formal and optimal as possible, an advised
11 vehicle is allowed to estimate its arrival time correctly even if its front car is not advised.
12 Otherwise an advised vehicle would have to guess the arrival time of the car in front when it
13 reaches the interaction range. Without going into detail, we found in our simulations that this
14 rapidly lead to complications since even a linear estimate can lead to crossing trajectories while
15 approaching towards the halt line.
16 Figure 6 shows how advised and unadvised vehicles interplay at a penetration rate of 20%.
17 If the latter follow an advised vehicle they practically behave as being advised themselves. The
18 trajectories in the top row are from a simulation run at demand of 0.05 vehicles per second. In the
19 second and third row one sees the results for demands of 0.15 and 0.5 vehicles per second,
20 respectively.
21 In the second and third row of Figure 6 one observes the failure of CA advice at larger
22 demand. One sees that for the CS advice (left) the equipped vehicle adjusts its speed at t = 650s to
23 a constant value in order to reach the TLS at the next green switch. The same vehicle in the right
24 diagram drives with the CA advice and first plans to follow its conventional lead vehicle through
25 the next green slot. Then around 180𝑚𝑚 before the traffic light it decides that it is too risky and
26 aims for the next green switch at t = 720s. As a consequence of this retardation, the same happens
27 at the end of this green slot: Another vehicle later decides that it better targets the green switch
28 at t = 780s. However, note that this vehicle under CS advice achieves to pass at t = 745s.
29 The reason for the failure of the CA advice in contrast to the CS advice is the following: In
30 the CS advice, the algorithm checks if the target speed is reached. If so, the driver can trust that it
31 arrives at green. In the CA advice, however, the speed typically varies more, since vehicles
32 approach the halt line at an optimal acceleration. In this case, the uncertainty in the arrival time at
33 the TLS is much higher than for constant speed. Therefore, sudden decisions to switch to the next

TRB 2016 Annual Meeting Paper revised from original submittal.


Woelki, Krajzewicz, Leich 11

1 green phase can occur.

4
5
6 FIGURE 6 Examples of trajectories for varying demand for CS (left) and CA (right).
7 Dashed lines represent unadvised vehicles. The penetration rate always is 20 %. Measured
𝒗𝒗𝒗𝒗𝒉𝒉
8 between −𝟓𝟓𝟓𝟓𝟎𝟎𝒎𝒎 and +𝟏𝟏𝟏𝟏𝟎𝟎𝒎𝒎. The demands are (from top to bottom): 𝟎𝟎. 𝟎𝟎𝟎𝟎, 𝟎𝟎. 𝟏𝟏𝟏𝟏, 𝟎𝟎. 𝟓𝟓 𝒔𝒔 .
9
10 The bottom diagrams of Figure 7 show an even stronger effect at maximum demand.
11 Consider first CS (left): During the green slot between t = 208s and t = 250s traffic jam spreads
12 backwards in space but all vehicles that attempted to pass through the green slot between
13 𝑡𝑡 = 240𝑠𝑠 and t = 265s succeed. The vehicle that wisely decides at t = 180s to slow down and

TRB 2016 Annual Meeting Paper revised from original submittal.


Woelki, Krajzewicz, Leich 12

1 target the green switch at t = 300s can assure himself that this was the right decision by counting
2 the vehicles in front starting from the halt line – those are 12 what is the maximum number of
3 vehicles that can pass through the slot (when accelerating from 0). So he will be the first to pass at
4 the next slot. Considering all the vehicles behind him one sees that the shape of their trajectories is
5 very linear (and possibly increasing towards the end). For each follower this makes the uncertainty
6 in planning the arrival time at the TLS very small. In contrast, in the bottom right diagram of
7 Figure 6 one sees that for the CA advice one recognizes a curvature towards zero velocity and
8 maximum uncertainty in the arrival time at the TLS. One driver at around t = 252s suddenly
9 comes to a halt when he feels that he cannot pass through the current slot. He does not take a
10 constant velocity to reach the next green switch because it falls short of the minimal allowed
11 velocity.
12
13 FINITE PENETRATION RATE AND DRIVER HETEROGENEITY
14 So far we have considered idealized situations where drivers always adapt to the optimum velocity
15 or acceleration. Now, WA driver heterogeneity is taken into account so that instead of the optimal
16 velocity a randomly reduced velocity is chosen. Following ref. (12), the randomness is described
17 by the parameter 𝜖𝜖, where 𝜖𝜖 = 0 corresponds to the deterministic case studied above. Since in the
18 previous section it turned out that CS is more stable than CA in case of interplay between equipped
19 and non-equipped vehicles, only CS is investigated in the following. For varying 𝜖𝜖, Figure 7 shows
20 a linear increase in travel time and CO2 production for a small simulation run of eight minutes.
21

22
23 FIGURE 7 Results for CS at maximum demand and varying strength of stochastic driver 𝝐𝝐.
24 The penetration rate always is 20%. Measured between −𝟓𝟓𝟎𝟎𝟎𝟎𝒎𝒎 and +𝟏𝟏𝟏𝟏𝟎𝟎𝒎𝒎.
25
26 CONCLUSIONS
27 Different driving patterns at a signalized intersection were considered in this paper, with special
28 attention to the trajectories of vehicles. The first important result is that following a constant speed
29 advice without catching up with the vehicle in front rapidly leads to a traffic breakdown in a high
30 traffic demand situation. Therefore a cooperative constant speed advice (CS) has been introduced
31 that reaches maximum throughput during each green phase of duration 𝐺𝐺 under idealized
32 conditions. Note that this maximum throughput is given by 𝐺𝐺/𝑇𝑇 with 𝑇𝑇 = 2𝑠𝑠 being the preferred
33 headway of vehicles. Two models serve for comparison. First: A cooperative acceleration advice
34 (CA) where vehicles slowly decelerate towards the halt line instead of driving at constant velocity.
35 Second: a real-world model without advice (WA) in which vehicles brake in the very last moment
36 in front of a red signal. It should be mentioned that this simple WA model is able to quite
37 effectively transport the vehicles over the traffic light position without communication with the
38 signal controller. This is achieved just by driving and accelerating as fast as possible and braking
39 whenever necessary. It also reaches the maximum throughput of vehicles through every green

TRB 2016 Annual Meeting Paper revised from original submittal.


Woelki, Krajzewicz, Leich 13

1 phase as the other two models of consideration.


2 In this paper special attention has been spent to the cases of maximum and vanishing
3 demand in order to analyze the limiting effects of different driving patterns. It turned out that a
4 deterministic maximal demand of automated connected vehicles (one vehicle every two seconds)
5 that drive at safety distance can most effectively be transported over the halt line when following
6 the CS advice. This strategy implies that the vehicles always catch up with the car in front. Thus
7 they should not simply drive at a constant velocity until the halt line. This also holds for the CO2
8 production. One reason why CS is more effective than CA at maximum demand is that
9 decelerations induce a larger safety gap than driving at a constant speed. However at very small
10 demand (0.01 vehicles per second) CA becomes better than CS since vehicles typically have a
11 headway of 100 seconds and therewith are usually not affecting each other. This is in accordance
12 with the results found in (8) where single trajectories were considered which corresponds to the
13 case of minimal demand.
14 The investigation of the concept of cooperative speed advice presented in this paper shows
15 that there is at least a technologically feasible approach to GLOSA for high demand situations.
16 Cooperative speed advice maintains the throughput of the signalized intersection as high as
17 without speed advice and performs a pollutant emissions reduction in the order of magnitude of
18 20% in the 100m surrounding of the intersection. The solution would require high penetration rates
19 of equipped vehicles and needs a mechanism that realizes a cooperative Electronic Tow-bar
20 between vehicles besides a road side unit that transmits SPaT and INFRA messages.
21 The present paper as well considered the interplay between equipped and non-equipped
22 vehicles. At small demand, the headway between vehicles is typically large enough so that
23 vehicles rarely interact. At maximal demand it turns out that CS is more effective than CA. The
24 main reason is that an approach at constant speed allows for a better planning of the arrival time at
25 the TLS. Driving at varying speed as in CA increases the uncertainty in the arrival time
26 considerably. This difference between CS and CA mainly affects vehicles that plan to arrive at the
27 end of a green phase. In case of the CA advice, the driver of an equipped vehicle targeting the end
28 of a green phase may change his mind to reach the next green switch instead when following a
29 conventional vehicle. Further, one can state that the WA model of the present paper (where
30 sometimes vehicles drive at maximum speed towards a red light and brake in the very last moment
31 to come to a halt) is too extreme for an examination of realistic interplay between equipped and
32 conventional vehicles apart from the regime of a small demand. In reality, drivers often anticipate
33 the red light and take their foot off the gas pedal. When further approaching towards the red signal
34 they additionally tend to brake comfortably. Such a behavior of unequipped vehicles should be
35 considered in future work.
36 Finally, both, mixed environment and stochasticity were considered. Travel time as well as
37 CO2 emissions scale linearly with the stochasticity parameter so that no traffic breakdown is
38 expected at least within the investigated range. Beyond that, it would be interesting to observe the
39 course of the quantities that were of interest in this paper (e.g. number of mode occurrences and
40 CO2 emissions) as a function of the delay or even the value of the interaction range. Additionally
41 one should compare the results with those obtained by the COLOMBO#2 model of reference (8)
42 which tries to be in coasting mode whenever possible rather than braking at a fuel consuming
43 deceleration.
44
45 ACKNOWLEDGEMENTS
46 The authors want to thank the European Commission for co-funding the work in the scope of the
47 COLOMBO project.

TRB 2016 Annual Meeting Paper revised from original submittal.


Woelki, Krajzewicz, Leich 14

1 REFERENCES
2 1. Miller, Alan J. (1963) "Settings for fixed-cycle traffic signals." OR: 373-386.
3 2. F. V. Webster (1958) Traffic signal settings. Road Res. Tech. Paper No. 39, H.M.S.O.
4 3. Misener, James, S. Shladover, and S. Dickey. "Investigating the Potential Benefits of
5 Broadcasted Signal Phase and Timing (SPAT) Data under IntelliDriveSM." ITS America
6 Annual Meeting, Houston, Texas. 2010.
7 4. Wegener, A., Hellbrück, H., Wewetzer, C., Lübke, A. (2008) VANET Simulation
8 Environment with Feedback Loop and its Application to Traffic Light Assistance,
9 GLOBECOM Workshops, 2008 IEEE , vol., no., pp.1,7, 2008, doi:
10 10.1109/GLOCOMW.2008.ECP.67.
11 5. Lebre, Marie-Ange, et al. "Real scenario and simulations on GLOSA traffic light system
12 for reduced CO2 emissions, waiting time and travel time." arXiv preprint
13 arXiv:1506.01965 (2015).
14 6. Eckhoff, David, Bastian Halmos, and Reinhard German. "Potentials and limitations of
15 green light optimal speed advisory systems." Vehicular Networking Conference (VNC),
16 IEEE, 2013.
17 7. Krajzewicz, Daniel und Bieker, Laura and Erdmann, Jakob (2012) Preparing Simulative
18 Evaluation of the GLOSA Application. In: Proceedings CD ROM 19th ITS World
19 Congress 2012 Vienna, Austria
20 8. Krajzewicz, D., Wagner, P., Woelki, M., Dippold, M. (2015) Driving patterns reducing
21 pollutant emission at traffic lights. Proceedings of the 22nd ITS World Congress,
22 Bordeaux, France, 5–9 October 2015,
23 https://proceedings.itsworldcongress.com/files/papers/754.
24 9. Krajzewicz, D., Heinrich, M., Milano, M., Bellavista, P., Stützle, T., Härri, J.,
25 Spyropoulos, T., Blokpoel, R., Hausberger, S., Fellendorf, M. (2013) COLOMBO:
26 Investigating the Potential of V2X for Traffic Management Purposes assuming low
27 penetration Rates. ITS Europe 2013, 04. - 07. Jun. 2013, Dublin, Irland.
28 10. COLOMBO consortium (2014) COLOMBO web site, http://www.colombo-fp7.eu.
29 11. Krajzewicz, D., Hausberger, S., Wagner, P., Behrisch, M., Krumnow, M. (2014) Second
30 Generation of Pollutant Emission Models for SUMO. SUMO2014 - Second SUMO User
31 Conference, 15.-16. Mai 2014, Germany. ISSN 1866-721X.
32 12. Krauss, S., Wagner, P., and Gawron, C.: Metastable states in a microscopic model of
33 traffic flow, Phys. Rev. E 55, 5597 (1997)
34 13. Hausberger S., “Simulation of Real World Vehicle Exhaust Emissions“, VKM-THD
35 Mitteilungen, Heft/Volume 82, Verlag der Technischen Universität Graz, ISBN
36 3-901351-74-4, Graz, Austria, 2003
37 14. Hausberger, S., Rexeis, M., Zallinger, M., Luz, R. (2009) Emission Factors from the
38 Model PHEM for the HBEFA Version 3, Report Nr. I-20/2009 Haus-Em 33/08/679.
39 15. Technical University of Graz (2014) pages of the Institute for Internal Combustion
40 Engines and Thermodynamics (IVT), last visited on 10.04.2014.
41

TRB 2016 Annual Meeting Paper revised from original submittal.

You might also like