optimizing traffic control to reduce fuel consumption and vehicular emissions

tions and frequent stops at intersections. However, low traffic andcontinuous progression along streets do not guarantee the lowest fuelconsumption and emissions. Excessive speeding, which may occur onroads with low traffic, may cause increased emissions for several pol-lutants. The best flow of traffic on arterial streets, in terms of fuel con-sumption and emissions, is the one with the fewest stops, shortestdelays, and moderate speeds maintained throughout the commute (1).

One of the ways to reduce excessive stop-and-go driving on urbanstreets is to optimize signal timings. Historically, signal timing opti-mization tools were developed to reduce delays and stops experiencedby urban drivers. The concept of optimizing signal timings to reducefuel consumption and emissions was first addressed by Robertsonet al. (2). However, at that time traffic was simulated by macroscopicand analytical tools, and individual driving behavior was not con-sidered. Similarly, the relationship between traffic activity, fuel con-sumption, and vehicular emissions, which was applied to all vehicles,was a simplistic and linear relationship (2).

In recent years powerful tools for traffic modeling, fuel consump-tion, and emissions modeling have been developed. Microscopicsimulation tools, such as VISSIM, have been used for more than adecade to model individual traffic behavior (3). Similarly, emissionsmodels, such as the comprehensive modal emission model (CMEM),were developed to estimate second-by-second emissions of individ-ual vehicles based on modes of a common driving cycle (4). Thesetwo types of microscopic models were coupled to estimate instanta-neous emissions based on second-by-second activities of individuallybehaved vehicles (5–7).

However, signal timing optimization models have been developedthat now use microscopic traffic models to evaluate and improve thequality of signal timings (8, 9). Researchers have reported that thesenew signal optimization tools generate signal timings that reducedelays and stops when compared with the ones generated by macro-scopic optimization tools (10). However, no research has been per-formed that integrates all these new microscopic tools in order to findthe best signal timings that would minimize fuel consumption andemissions. The research reported here aims to fill that gap in existingpractice by integrating a microscopic traffic simulator, a comprehen-sive microscopic emission estimation model, and a stochastic sig-nal optimization tool to provide signal timings that minimize fuelconsumption and vehicular emissions.

BACKGROUND

In previous decades, many researchers have evaluated the effects oftraffic signal timings on the environment (11–18). Effects are eval-uated through an investigation of the amount of fuel consumption

Optimizing Traffic Control to Reduce FuelConsumption and Vehicular EmissionsIntegrated Approach with VISSIM, CMEM, and VISGAOST

Aleksandar Stevanovic, Jelka Stevanovic, Kai Zhang, and Stuart Batterman

105

One way to reduce excessive fuel consumption and vehicular emissions onurban streets is to optimize signal timings. Historically, signal timing opti-mization tools were used to reduce traffic delay and stops. The concept ofoptimizing signal timings to reduce fuel consumption and emissions wasaddressed decades ago with tools that are now considered outdated. Thisstudy advocates a fresh approach to integrating existing state-of-the-arttools for reassessing and ultimately minimizing fuel consumption andemissions. VISSIM, CMEM, and VISGAOST were linked to optimizesignal timings and minimize fuel consumption and CO2 emissions. Asa case study, a 14-intersection network in Park City, Utah, was used.Signal timings were optimized for seven optimization objective functionsto find the lowest fuel consumption and CO2 emissions. Findings showthat a formula commonly used to estimate fuel consumption in trafficsimulation tools inadequately estimates fuel consumption and cannot beused as a reliable objective function in signal timing optimizations. Someof the performance measures used as objective functions in the opti-mization process were proved to be ineffective. When CMEM-estimatedfuel consumption is used as an objective function, estimated fuel savingsare around 1.5%, a statistically significant decrease. Further researchis needed to find an effective way to minimize fuel consumption andemissions by using the proposed approach.

Both continuous transportation growth in the Western world and therecent economic boom in India, China, and many third-world coun-tries have had a tremendous impact on the use of fossil fuels. Theincrease in fuel consumption affects the environment (the greenhouseeffect), health (air pollutants), and the economy (increased fuel prices).Increased fuel consumption is mainly caused by two factors. First,millions of new drivers start using private cars as a main mode oftransportation every year. Second, when these new travelers joinexisting traffic demand, traffic congestion increases because highwaycapacity does not increase commensurately with the new demand.

The highest fuel consumption on urban arterials is associated withdriving in congested traffic, characterized by higher speed fluctua-

A. Stevanovic and J. Stevanovic, Department of Civil and Environmental Engineer-ing, University of Utah, 122 South Central Campus Drive, Room 104, Salt LakeCity, Utah 84112-0561. Current affiliation for A. Stevanovic, Department of CivilEngineering, Florida Atlantic University, 777 Glades Road, Building 36, Room231, Boca Raton, FL 33431. Current affiliation for J. Stevanovic: 2145 NorthwestThird Court, Boca Raton, FL 33431. K. Zhang and S. Batterman, EnvironmentalHealth Sciences, School of Public Health, University of Michigan, 1420 WashingtonHeights, Room 6037, Ann Arbor, MI 48109-2029. Corresponding author: A. Stevanovic, [email protected].

Transportation Research Record: Journal of the Transportation Research Board,No. 2128, Transportation Research Board of the National Academies, Washington,D.C., 2009, pp. 105–113.DOI: 10.3141/2128-11

and vehicle emissions (both air pollutants and greenhouse gases) forvarious traffic conditions by using various methods and tools. Bothfield and theoretical tests have shown that optimized signal timingsdecrease fuel consumption and vehicle emissions compared withnonoptimized timings.

Robertson et al. optimized signal timings by using TRANSYT 8 tominimize fuel consumption (2). They found that when signal timingsare not optimized to reduce delays but to reduce total fuel consump-tion, the benefits of such signal timings may decrease fuel consump-tion by up to 3%. The fuel consumption was estimated from its linearrelationship with traffic performance measures (delay, stops, andaverage speed) (2). The research set an industry standard in optimiza-tion of signal timings by defining a performance index (PI) as a linearcombination of delay and stops that should be minimized to get min-imal fuel consumption. Experiments showed that each stop should beassociated with a penalty delay of 20 s if fuel consumption is going tobe minimized. This PI became a standard objective function for opti-mizing signal timings, and the defined weights for delay and stopshave not changed significantly since then.

To estimate air pollutant concentrations, Park et al. coupled theVISSIM microsimulator model with MODEM, an emissions inven-tory database (19). Concentrations estimated by using a Gaussian dis-persion model were comparable with those estimated from anothermacroscopic model but slightly different from levels measured inthe field.

Instead of using an emissions inventory database, Nam et al.coupled VISSIM with CMEM to estimate emissions from a singlevehicle (5). The comparison with the field measurements found thatCMEM is acceptable when capturing aggregated hydrocarbon (HC)and carbon monoxide (CO) trends but less accurate for carbon diox-ide (CO2) and nitrogen oxides (NOx). An integrated VISSIM–CMEMmodel was also used to show that the signal timings, optimized forprogression in TRANSYT 9, significantly reduced pollutant emis-sions and fuel consumption on an arterial road (6). Oda et al.developed a simulator to estimate CO2 emissions (20). They used amacroscopic traffic flow model to input traffic activities into the CO2

simulator. The authors wanted to optimize traffic control settings toreduce CO2 emissions. However, because of the huge computationalburden needed to estimate CO2 for all vehicles in the network, theauthors simplified the experiments. Instead of minimizing CO2 theyminimized the number of stops, which they had shown was highlycorrelated with CO2 (20). Another integrated VISSIM–CMEM modelwas used to show that a scenario with optimal traffic control reducedvarious pollutant emissions (CO, HC, NOx) from 3% to 15%. Theresearch was done for a road network in Beijing by Chen and Yu (7).

Qu et al. investigated impacts of reduced freeway speed limits ontraffic emissions in Houston, Texas (21). The authors used TRANSIMto model traffic. The traffic activities were imported into three emis-sion models: TRANSIMS (CMEM), MOBILE 5, and MOBILE 6.Emissions of three major pollutants [volatile organic compounds(VOC), NOx, and CO] were modeled in each of the three emissionsmodels to investigate the effectiveness of freeway speed limit reduc-tions as a way to decrease emissions. The results were mixed, show-ing that some models justify the reduction of speed limits while othersdo not. The study also showed TRANSIMS’s inability to modelchanges in speed limits accurately because of its discrete approach inmodeling vehicular speeds.

Another attempt to determine signal timings that minimize fuelconsumption and vehicular emissions was reported by Smith et al.(22), who briefly addressed SCOOT operations that minimize vehicleemissions. Traditionally, SCOOT has been used to minimize delaysand stops in traffic by adjusting signal timings based on traffic demand

106 Transportation Research Record 2128

measured in real time. The authors tested a new version of SCOOTthat can minimize any of the five emission pollutants—CO, CO2,VOC, NOx, and PM10—instead of the traditional PI. The pollutantswere estimated on the basis of the SCOOT traffic model. The authorsused a new SCOOT feature to minimize emissions by adjusting traf-fic control settings for the U.K. region of Leicester. The results showedthat emissions for any of the pollutants could be reduced by up to2% if an emission-related objective function is used during SCOOToptimizations. Unfortunately, these reductions were not statisticallysignificant at the 95% confidence level. A major limitation of theapproach was the fact that SCOOT’s mesoscopic traffic model wasnot capable of modeling second-by-second modular operations (accel-eration, cruising, idling) of individual vehicles. Rather, SCOOT basesits emission estimates on average emission rates for each vehicleclass (four classes are available), and traffic flow and speed estimatesare averaged over each link (13, 23).

In summary, researchers have used various traffic simulation toolsand various methods to estimate fuel consumption and vehicle emis-sions. Most applications have shown that optimized signal timingsdecrease fuel consumption or vehicle emissions or both but are basedon macroscopic or mesoscopic models and unreliable objective func-tions. However, no research has addressed the optimization of sig-nal timings based on evaluations of single-vehicle emissions anddriving behavior. Further, without an objective function related toaccurate fuel consumption and emission estimates, signal timings can-not be optimized to minimize these environmental impacts. Researchpresented here optimizes signal timings on the basis of CMEM emis-sions estimates for a population of vehicles whose individual drivingbehaviors were modeled in VISSIM. Optimization was used tominimize fuel consumption and CO2 emissions.

VISSIM–CMEM–VISGAOST CONCEPT

VISSIM Model

VISSIM is a microscopic, time-step and behavior-based model devel-oped to simulate urban traffic and public transport operations. Theprogram can analyze vehicle operations under different lane config-urations, traffic composition, traffic signals, and public transport stops.This ability makes it a useful tool to evaluate traffic in alternativenetworks and to develop transportation engineering and planningmeasures of effectiveness (3).

The accuracy of a traffic simulation model is mainly dependent onthe quality of the vehicle modeling, such as the methodology ofmoving vehicles through the network. In contrast to less complexmodels that use constant speeds and deterministic car-following logic,VISSIM uses the psychophysical driver behavior model developedby Wiedemann (3).

VISSIM has several ways of modeling traffic control. One of themost popular ways is the emulation of the industry standards in traf-fic control established by the National Electrical Manufacturers Asso-ciation (NEMA). Recent experiments showed that signal timingsgenerated by VISSIM’s NEMA emulator do not differ practicallyfrom those generated by real-world controllers.

CMEM Model

CMEM is a physically based, power-demand model developed bythe University of California at Riverside, the University of Michigan,and Lawrence Berkeley National Laboratory (4). After a variety of

enhancements, the latest version (3.0) includes submodels for light-duty vehicles (LDVs) and heavy-duty diesel (HDD) vehicles. Thesesubmodels estimate vehicle tailpipe emissions (CO, HC, NOx, andCO2) in different modes of vehicle operation, such as idling, cruising,acceleration, and deceleration. Scora and Barth suggested that tem-poral and vehicular aggregations were necessary in practice becauseCMEM was developed to predict emissions for vehicle categories (4).The temporal scale ranges from second-by-second, several seconds(mode) to driving cycle or scenario, and the vehicular scale rangesfrom a specific vehicle, vehicle technology category, to generalvehicle mix or fleet.

CMEM model inputs include traffic composition, vehicle and oper-ation variables (e.g., speed, acceleration, and road grade), and model-calibrated parameters (e.g., cold start coefficients and an enginefriction factor) (4). Outputs are tailpipe emissions and fuel consump-tion. Emissions (in grams per second) are predicted as the product offuel rate (FR, in grams per second), engine-out emission indices(grams of emission per grams of fuel), and time-dependent cata-lyst pass fraction (CPF), defined as the ratio of tailpipe to engine-out emissions. CPF is mainly affected by the fuel-to-air ratio andengine-out emissions.

LDV and HDD models have similar structures (4). Both are com-posed of six modules: engine power demand, engine speed, fuel-to-air ratio for the LDV model or engine control unit for the HDD model,fuel rate, engine-out emissions, and CPF for the LDV model or after-treatment pass fraction for the HDD model. Key parameters (e.g.,vehicle mass, engine size, fuel type) depend on vehicle technology,fuel delivery system, emission control technology, vehicle age, andother factors. CMEM has been calibrated by using data from theNational Cooperative Highway Research Program, which includesboth engine-out and tailpipe emissions of CO, HC, NOx, and CO2

for over 400 vehicles in 36 vehicle technology categories.

VISGAOST Program

VISGAOST is an optimization program for signal timings of trafficcontrollers based on their performance in VISSIM microscopic sim-ulation. The program bases its optimization on the stochastic natureof genetic algorithms (GAs). The general structure of VISGAOSTGA optimization is well documented (10). The basic version ofVISGAOST is written in C++ and relies on VISSIM’s input and out-put files (3). The key part of the program is a simple GA similar toother GAs used for signal timing optimization (24).

The first version of VISGAOST enabled the optimization of allfour basic signal settings: cycle, offset, split, and phase sequence.The program was tested and evaluated for the network in Park City,Utah, consisting of three groups of coordinated intersections and twoactuated intersections. Results confirmed that VISGAOST can findtiming plans that work better in VISSIM than the initial timing plansfrom the field (9). Further, the results showed that the GA-optimizedplan was better than the timing plan generated by the traditionaloptimization tool SYNCHRO.

VISGAOST application was extended to enable optimization oftransit signal priority (TSP) settings. The two most common TSPsettings—green extension and early green—were optimized for acorridor of seven signalized intersections in Albany, New York.Results showed that the optimized timing plan improved overalltraffic performance and reduced person delay (10).

The extended version of VISGAOST, presented in this paper,enables optimization of signal settings to minimize fuel consumption

Stevanovic, Stevanovic, Zhang, and Batterman 107

and vehicular emissions estimated by CMEM. The program has beenmodified to accommodate new linkage to CMEM and some new esti-mates from VISSIM. The steps below describe the basic operationsin the VISGAOST optimization process.

Step 0: InitializingG, total number of generations;T, total number of timing plans per generation;�, convergence threshold;i, current number of population;i = 0.Generation of initial population pi of timing plans tpk ∀ k ∈

[1, . . . , T]• Read field timing plan tp1 from database,• Generate tpk ∀ k ∈ [2, . . . , T].

Step 1: Evaluating populationEvaluation of tpk ∈ pi ∀ k ∈ [1, . . . , T]

• Write tpk to database,• Simulate tpk,• Estimate emissions for tpk,• Calculate fitnessk.

Step 2: Testing termination criteria• Find b, fitnessb for which fitnessb = max( fitness1, . . . , fitnessT);• Find fitnessa for which fitnessa = average( fitness1, . . . ,

fitnessT );• Test rule.

IF ((i = G) OR (( fitnessb − fitnessa) < �))Stop and RETURN tpb ∈ pi

ELSEGO TO Step 3

Step 3: Generating new populationi = i + 1Generation of new population pi

• Select best-ranking timing plans from pi−1,• Generate pi through GA operations.

GO TO Step 1

VISSIM–CMEM–VISGAOST Integration

Figure 1 shows the integration of VISSIM, CMEM, and VISGAOSTto find signal timings that reduce fuel consumption and vehicularemissions. The optimization process starts with the VISGAOST gen-eration of the initial population of signal timings, which is seeded bythe existing set of signal timings from the field. Each generated sig-nal timing plan is evaluated in VISSIM. As a result of the evalua-tion process, VISSIM outputs a vehicle record file with relevantsecond-by-second data for each vehicle in the network for the entiresimulation period.

The vehicle record file is processed by the VISSIM–CMEM inter-face and sent to CMEM. CMEM estimates emissions and fuel con-sumed during the evaluation of that particular signal timing plan.The CMEM estimates are then summed for all vehicles in the networkduring the entire simulation period.

VISGAOST receives the summed fuel consumption (or vehicularemissions) for each signal timing plan from the current population.A signal timing plan with the lowest fuel consumption (emissions)will be selected as the best one and saved to be compared with thebest one from the next generation. Then the GA procedure withinVISGAOST uses four basic GA operators to create a new popula-tion of signal timings. The whole process is repeated until one of twopredefined termination criteria is met.

VISSIM–CMEM–VISGAOST Interface

Connections between traffic microsimulation tools, such as VISSIM,and instantaneous emission models, such as CMEM, have beendescribed elsewhere (5–7), and other studies provide more informa-tion about the VISGAOST and VISSIM interface (9, 10). Here thefocus is on specific modifications of these two interfaces that enablefunctional communication among the three tools.

The major limitations of several previous attempts to integrate traf-fic simulation and emissions estimation models were that they werenot applicable to U.S. traffic conditions because the emissions werebased on European vehicles (15, 25, 26). In other studies, heavy vehi-cles were not modeled directly because there was no HDD model inCMEM at that time (5, 27–29). Finally, in several studies vehicleemissions were overestimated because low-emitting vehicles wereimproperly represented in the old CMEM version (4). This studyimproves the emissions modeling approach by using a recent versionof the CMEM software (3.0) and a representative sample of vehiclesused in the United States.


A program built in Java connects VISSIM with the LDV and HDDmodels in CMEM (Figure 1). The program’s logic is similar to thatdescribed previously (6, 28). The program improves on the previousdevelopments by modeling diesel trucks directly, calling either LDVor HDD core models for each individual vehicle (instead of the LDVbatch model, which limits the number of records and vehicles that canbe handled) (4), and using Java, a platform-independent language.

For each vehicle VISSIM provides simulation time, a vehicle iden-tifier, a vehicle type (LDV or truck), speed, and acceleration or decel-eration on a second-by-second basis. The Java interface programimports the VISSIM output file to CMEM, which uses individualvehicle data to estimate instantaneous emissions for each vehicle.Each VISSIM vehicle type is assigned (by the Java program) to aCMEM vehicle category.

The assignment of vehicle categories follows the mapping processdescribed in Table 1, which maps the vehicle types from MOBILE 6.2to the CMEM vehicle categories. It was assumed that the simulatedvehicle fleet is composed of the light-duty gasoline vehicles (LDGVs)and heavy-duty diesel vehicles (HDDVs) defined in MOBILE 6.2.

VISSIM VISGAOST

Simulation time: 600 to4200Parameter ValueTotal travel time[h] 835.8Total delay time[h] 159.2Number of stops 21828Stopped delay[h] 84.2

NetworkPerformance

VISSIM Output

Split[1,8]=[[10.0,23.0,10.0,23.0,10.0,23.0,10.0,23.0]];LeadPhase[1,8]=[[1,0,0,1,1,0,1,0]];CycleLength[1]=[66.0];Offset[1]=[30.0];

VISSIM InputSignalGroups[8]=[1.0,2.0,3.0,4.0,5.0,6.0,7.0,8.0];

Optimized Signal Settings

Measures of Effectiveness

t; Veh; Type; v; a;1.0; 2; 1001; 23.18; 0.86;1.0; 1; 1001; 25.75; 0.69;1.0; 3; 1001; 24.55; 0.82;2.0; 5; 1001; 23.80; 0.59;2.0; 4; 1001; 24.60; 0.80;2.0; 2; 1001; 23.76; 0.86;

Ve hi cl e Record

VISSIM Output

Distance Traveled 0.55 miFuel Use17.0585 (grams/mile)CO2 = 17.8462 (grams)CO = 7.5225 (grams)HC = 0.0994 (grams)NOx = 0.0931 (grams)

Control File: veh-Activity File: veh-

CMEM Output

VISSIM-CMEM-VISGAOST Interface

CO2 1.3416976E7 CO 1459375.8 HC 25465.648 NOx 29045.52 Fuel 3147279.0 Dist 20699.896

Summed Estimations

Distance Traveled 0.55 miFuel Use17.0585 (grams/mil e)CO2 = 17.8462 (grams)CO = 7.5225 (grams)HC = 0.0994 (grams)NOx = 0.0931 (grams)

Control Fi le: veh-Activity Fi le: veh-

Di stance Traveled 0.15 miFuel Use10.0585 (grams/mile)CO2 = 17.8462 (grams)CO = 7.5225 (grams)HC = 0.0994 (grams)NOx = 0.0931 (grams)

Control File: veh-Activi ty Fi le: veh-

Distance Traveled 0.35 miFuel Use13.1246 (grams/mile)CO2 = 22 .22 (grams)CO = 2.33 5 (grams)HC = 0.07 91 (grams)NOx = 0.0592 (grams)


Distance Traveled 0.35 miFuel Use13.1246 (grams/mile)CO2 = 22.22 (grams)CO = 2.335 (grams)HC = 0.0791 (grams)NOx = 0.0592 (grams)


Type CMEM Description MEM Code Percent

ULEV 51 0.08

PZEV 52 0.08

Tier 1 < 50k, low ratio 10 0.09

Tier 1 < 50k, high ratio 11

8

9

0.09

Tier 1 > 50k, low ratio 0.20

Tier 1 > 50k, high ratio 0.20

3-way catalyst, FI, > 50k miles low 4 0.13

3-way catalyst, FI, > 50k miles high 5 0.13

HDDV 1999-2002, 4 stroke, Elect 47 1.00

LDGV

Lookup Table

OR

LDV

HDD

CMEM

FIGURE 1 VISSIM–CMEM–VISGAOST integration.

It was also assumed that LDGVs can be represented in CMEM bytwo Tier 2 vehicle categories, ultra-low-emitting vehicles (ULEVs)and partial-zero-emitting vehicles (PZEVs); four Tier 1 vehicle cat-egories; and two categories of old vehicles (Table 1). The LDGVcategory was matched to these eight CMEM categories according tothe vehicle age distribution from MOBILE 6.2 (30) and the Tier 2phase-in schedule (31). CMEM does not include HDDVs after


2002, and thus the CMEM category of 1998–2002 HDDVs was cho-sen instead. Trucks manufactured before 1998 or after 2002 werenot considered in the study.

Many vehicle types can be defined in VISSIM; however, the ini-tial experiments were constrained to two: passenger cars and heavyvehicles (trucks). Depending on the VISSIM vehicle type, the Javaprogram utilizes either the CMEM LDV model or the CMEM HDDmodel. A CMEM model (LDV or HDD) computes fuel consumptionand vehicular emissions for each vehicle in the simulation outputs.The Java program summarizes individual vehicles’ fuel consumptionand emissions (CO, HC, NOx, and CO2) to obtain the total values forthe entire road network.

CASE STUDY

Study Network, Park City, Utah

To optimize signal timings for minimal fuel consumption, the ParkCity road network, located in Utah near Salt Lake City, was chosen.The network consists of two suburban arterials, SR-224 and SR-248,and many crossroads. The network, shown in Figure 2, has 14 sig-nalized intersections and average annual daily traffic of 32,000 and20,000 on SR-224 and SR-248, respectively.

0 0.5 1 km

FIGURE 2 VISSIM model of road network in Park City, Utah.

TABLE 1 Mapping of Vehicle Categories in MOBILE 6.2 and CMEM

CMEMType CMEM Description Code Percent

LDGV ULEV 51 0.08PZEV 52 0.08Tier 1 < 50k, low ratio 10 0.09Tier 1 < 50k, high ratio 11 0.09Tier 1 > 50k, low ratio 8 0.20Tier 1 > 50k, high ratio 9 0.203-way catalyst, FI, > 50k miles low 4 0.133-way catalyst, FI, > 50k miles high 5 0.13

HDDV 1999–2002, 4-stroke, electric 47 1.00

NOTE: FI = fuel injected.

VISSIM Model of Park City Network

Building, calibrating, and validating the VISSIM model requiredextensive field data collection and data reduction efforts. A team of10 students was employed and trained to collect various traffic dataduring three weeks in August 2005. All data used in this study werecollected between 4:00 and 6:00 p.m. on workdays under fair weatherand dry pavement conditions. The following data were collected:turning-movement counts, saturation flow rates, stopped delay at theintersections, spot speed data, corridor vehicle classification counts,and corridor travel times. On the basis of the collected data the fol-lowing parameters were adjusted to calibrate the VISSIM model: traf-fic inputs and routing decisions, the two car-following parametersin the Wiedemann 74 VISSIM model, control delays, desired speeddecisions, and vehicle compositions. Validation of the model wasdone with the corridor travel times. All of the segment travel timesfrom the field and VISSIM were close, but four of them (two in eachdirection) were still statistically different (two-tailed t-test was per-formed, with a = 0.05 and n = 15). A detailed description of the data,calibration process, and validation results was given elsewhere (9).

Field Signal Timings

The field signal timings were implemented under mixed actuated-coordinated and actuated-uncoordinated control. Intersections ofBonanza Drive and SR-248 and SR-248 and Comstock Drive areactuated-uncoordinated, and all others are coordinated. The first threeintersections in the Kimball Junction area were run on 128-s cycles.The other intersections all ran on 106-s cycles with exception ofDeer Valley and Bonanza Drive, which used double cycling. Thesignal timings in the field were monitored regularly, but there wereno recent major updates. Traffic engineers maintained the signaltimings to achieve good progression between intersections, whichwas reflected in the initial signal timings in the optimization.

VISGAOST Optimizations

There were two major objectives for the VISGAOST optimization ofsignal timings. The first objective was to compare estimates of thefuel consumption from CMEM LDV and HDD models with thosecomputed by VISSIM (node evaluation) based on a formula widelyused by major traffic signal optimization tools (TRANSYT-7F andSYNCHRO) (8). The formula used by VISSIM, TRANSYT-7F, andSYNCHRO reads as follows:

where

k1 = 0.075283 − 0.0015892 � Speed + 0.000015066 �speed2,

k2 = 0.7329,k3 = 0.0000061411 � speed2,F = fuel consumed (gal),

speed = cruise speed (mph),total travel = vehicle miles traveled (veh mi),total delay = total signal delay (h), and

stops = total stops (veh/h).

The second objective was to show that in order to minimize fuelconsumption or vehicular emissions, fuel consumption or particu-

F k k k= + +total travel total delay stops� � �1 2 3


lar vehicular emissions should be used as an objective functionwhen signal timings are optimized. In other words, when delaysor stops are used in the objective function, minimal delays or stopsare obtained but not necessarily the minimal fuel consumption orlowest vehicular emissions. Because of the time-consuming opti-mization process, these optimization experiments were limited tominimize fuel consumption and CO2 emissions. CO2 does not rep-resent a criterion pollutant, but because of the threat of global warm-ing, controlling this gas has become more important than ever. Othervehicular emissions (CO, HC, NOx) can also be optimized by theproposed VISSIM–CMEM–VISGAOST approach.

Seven optimization experiments were conducted. In total morethan 100 control variables for all intersections in the Park City networkwere optimized to reduce total delay (for the entire network, in hoursper hour), stops, throughput (total number of vehicles that completedtheir trips in the network), PI (PI = total delay + 10 * stops/h), CMEMfuel consumption, VISSIM fuel consumption, and CMEM CO2 emis-sions. Multiple optimizations, with various objective functions, wereperformed to show the difference in the lowest fuel consumption andCO2 achieved by each optimization method.

Each optimization started with the same initial signal timings fromthe field. Each optimization was based on evaluations of traffic andemissions performances accumulated during 60 min of simulationtime with an additional 10 min for warm-up. Simulation warm-up isnecessary to achieve steady-state traffic conditions in the network.Each optimization had 12,000 evaluations of various signal timingplans; 20 signal timing plans were operated through GA proceduresfor each of 600 generations. Previous experiments showed that thiscombination of GA population and generations yields the best results(9). In addition, each signal timing plan was evaluated for five ran-domly seeded simulation runs to account for variability of traffic flows.The optimizations were performed on 20 dual-processor computers.Overall, it took around 20 days of continuous simulation run time tocomplete the optimizations.

RESULTS AND DISCUSSION

Evaluation Results

Optimization of CMEM-estimated fuel consumption is shown inFigure 3, which demonstrates how the best fuel consumption andaverage fuel consumption vary over 600 generations. Spikes observ-able in Figure 3 reflect use of partial optimizations of signal timings(9). Similar trends were observed for six other optimization runs. Mostof the final signal timings, which reduce fuel consumption and CO2

emissions, seem to favor major-street operations and yield more delayto the side-street traffic. These signal timings exhibit higher cyclelengths and better progression on major streets. Once the optimiza-tions were finished, each of the seven best signal timing plans wasevaluated through 40 randomly seeded VISSIM simulations. VISSIMperformance measures were recorded and average statistics werecomputed. The 40 VISSIM runs were also linked with CMEM.CMEM’s estimates of fuel consumption and CO2 were recorded andaveraged. Mean values from these statistics are presented in Table 2for all seven objective functions.

Discussion of Results

Almost all of the optimization experiments found signal timingsthat reduce CMEM fuel consumption when compared with the initial

signal timings. However, not all of the objective functions workedeffectively. Although delay, throughput, and PI could be used as objec-tive functions without additional constraints, the other performancemeasures were not as effective in this role.

The major problem is represented by the way that an objectivefunction reports performance of the system if there is excessive delayin the network. In such a situation delay and PI significantly increase,whereas throughput significantly decreases. So if the GA proceduresuggests a signal timing plan that causes a traffic jam, these three per-formance measures, when used as objective functions in the GA, willdetect the problem and such a signal timing plan will be discarded.

However, when other performance measures are used as objectivefunctions, they may not necessarily recognize poor traffic conditions.For example, if traffic is jammed, vehicles move less and hence theystop less frequently (as recorded by VISSIM). So if the stops areminimized in the optimization, the traffic jam will be perceived as afavorable outcome. A similar situation occurs with emissions-relatedmeasures. If a vehicle is stopped and idling, it consumes less fuel than


one that runs at 40 mph. So although its fuel-per-mile consumptionis higher when the vehicle is idling, its fuel-per-second consumptionand emissions are lower. For this reason sometimes both VISSIM’sinternal fuel calculation procedure and CMEM report low fuel con-sumption and emissions associated with poor traffic conditions (whichare detected by the other performance measures). To illustrate theproblem, such an example is provided in Table 2, where VISSIM-reported fuel consumption is used as an objective function (note theminimal VISSIM-reported fuel consumption and the huge increasesin number of stops and delay).

Similar problems were observed when three other performancemeasures were used as objective functions in these experiments (stops,CMEM fuel use, and CMEM CO2). The ultimate solution to this prob-lem might be either selection of the reliable objective function or useof various metrics (e.g., stops and delay) as constraints in the optimiza-tion rather than in the objective function. To test this integration, thesecond method was used. The GA optimizations were constrained insuch a way that poor solutions, which minimize one objective function

650

655

660

665

670

675

680

685

0 50 100 150 200 250 300 350 400 450 500 550 600

Number of Generations

Fue

l Con

sum

ptio

n [g

al]

Average Fuel Consumption

Initial Fuel Consumption

Best Fuel Consumption

FIGURE 3 Optimization of CMEM fuel consumption.

TABLE 2 Measures of Effectiveness from 40-Run Tests

Mean Values

CMEM VISSIM CMEM CO2 ThroughputMOE Optimized Fuel Use [gal] Fuel Use [gal] [kg] Delay [h] Stops [veh] PI

Initial (no optimization) 685.1 780.7 10,349.6 185.1 24,317.1 8,020.1 252.7

Delay [h] 667.2 784.9 10,410.3 163.0 22,501.1 8,087.7 225.5

Stopsa 660.8 786.4 10,251.1 164.1 20,046.3 8,093.7 219.8

Throughput [veh] 674.2 792.8 10,514.4 177.3 26,897.4 8,094.3 252.0

PI 667.2 786.5 10,399.0 164.2 22,895.0 8,115.0 227.8

CMEM fuel use [gal]a 658.8 789.5 10,221.4 170.0 20,679.6 8,120.3 227.4

VISSIM fuel use [gal] 2,011.4 629.9 13,055.2 1,046.0 158,819.7 5,813.9 1,487.2

CMEM CO2 [kg]a 661.6 787.4 10,206.8 168.3 21,252.4 8,103.8 227.4

aConstrained optimization results.

but negatively affect all others, are discarded. Table 2 shows theresults of such an approach for the three objective functions whoseoptimizations were constrained (denoted by an asterisk).

The results from Table 2 show that almost every optimizationreduces CMEM-reported fuel consumption and CO2 emissions. How-ever, optimizations in which these two performance measures wereminimized generated minimal results. At the same time, VISSIM’sfuel consumption does not seem to be consistent even when properperformance measures are used as objective functions.

Overall, VISSIM fuel consumption does not represent an accu-rate value when compared with CMEM fuel consumption. The fuelconsumption reported by VISSIM should be lower than that fromCMEM because VISSIM reports fuel consumption only on the linkswithin node boundaries (user-defined areas around the intersections),whereas CMEM reports total fuel consumed on all links in the net-work. However, results from Table 2 show the opposite trend, becauseVISSIM does not calculate fuel consumption properly. There aremultiple reasons for this inaccuracy in VISSIM’s fuel consumptioncalculations: the formula used by VISSIM is based on aggregatedmeasures (speed, stops, delay, etc.) and cannot provide a level ofaccuracy achieved by CMEM calculations; VISSIM does not reportfuel consumption for those vehicles that are still within the nodeboundaries; VISSIM’s formula might be based on outdated emissionscharacteristics of an average vehicular fleet. Further research isneeded to investigate the inaccuracy of the fuel consumption reportedby VISSIM.

When CMEM fuel consumption is used as an objective functionto optimize signal timings, the optimal signal timings generate sig-nificantly lower fuel consumption than if delay or PI is used as anobjective function. The savings in fuel consumption when comparedwith the delay or the PI are around 1.5%. Although such savings maynot be seen as important, it is interesting to see that after 12,000 eval-uations the results do not show that there is a significant differencein fuel consumption between signal timings optimized for minimaldelay and minimal PI. In the past, the difference in fuel consumptionbetween signal timings that optimize delay and PI was estimatedto be between 1% and 3% (2). The current findings confirm thosefrom Smith et al., who directly minimized fuel consumption (withinSCOOT) and reported similar benefits over PI optimization of around2% (22).

CONCLUSIONS

The goal of this study was to present a new integration of existingtraffic operation, emissions estimation, and signal optimization mod-els. The study describes the integration of VISSIM, CMEM, andVISGAOST to optimize signal timings in such a way as to achieveminimal fuel consumption and vehicular emissions. The followingconclusions were reached:

1. Number of stops, fuel consumption, and CO2 emissions do notseem to be reliable objective functions in the optimization of signaltimings. Instead, they should be combined with other traffic perfor-mance measures, or additional constraints need to be introduced inthe optimization process. Further research is needed to investigatewhat the best objective function is to minimize fuel consumptionand emissions.

2. The VISSIM formula for fuel consumption is heavily influ-enced by number of stops and does not seem to be a reliable objec-tive function to minimize fuel consumption or emissions. VISSIM’s


method of estimating fuel consumption seems to significantly over-estimate total fuel consumption when it is compared with CMEMfuel consumption.

3. If fuel consumption is used as an objective function in a con-strained optimization of signal timings, the optimal signal timingswill generate fuel consumption 1% to 1.5% lower than that obtainedthrough a minimization of delay or PI. Although these results mayseem insignificant, they have the same order of magnitude as theresults obtained from the experiments that first investigated fuelminimization through signal timings (2).

4. Lengthy computation times make application of this researchimpractical for everyday optimization of signal timings. For this rea-son it is necessary to investigate which combination of delay, stops,and other potential performance measures would lead to minimalfuel consumption.

5. Future research should address additional optimization exper-iments with a variety of traffic networks and scenarios to develop ageneral strategy of how optimization of certain traffic metrics affectsfuel consumption and vehicular emissions. Eventually, results fromthe simulation testing should be validated by field measurements.

REFERENCES

1. Barth, M. J., and K. Boriboonsomsin. Real-World Carbon Dioxide Impactsof Traffic Congestion. In Transportation Research Record: Journal of theTransportation Research Board, No. 2058, Transportation Research Boardof the National Academies, Washington, D.C., 2008, pp. 163–171.

2. Robertson, D. I., C. F. Lucas, and R. T. Baker. Coordinating Traffic Sig-nals to Reduce Fuel Consumption. TRL Report LR934. TransportResearch Laboratory, Crowthorne, Berkshire, United Kingdom, 1980.

3. VISSIM 4.30 User Manual. Planung Transport Verkehr AG, Karlsruhe,Germany, 2008.

4. Scora, G., and M. Barth. Comprehensive Modal Emission Model(CMEM) Version 3.01 User’s Guide. University of California, Riverside,2006.

5. Nam, E. K., C. A. Gierczak, and J. W. Butler. Comparison of Real-Worldand Modeled Emissions Under Conditions of Variable Driver Aggres-siveness. Presented at 82nd Annual Meeting of the TransportationResearch Board, Washington, D.C., 2003.

6. Stathopoulos, F. G., and R. B. Noland. Induced Travel and Emissionsfrom Traffic Flow Improvement Projects. In Transportation ResearchRecord: Journal of the Transportation Research Board, No. 1842, Trans-portation Research Board of the National Academies, Washington, D.C.,2003, pp. 57–63.

7. Chen, K., and L. Yu. Microscopic Traffic-Emission Simulation and CaseStudy for Evaluation of Traffic Control Strategies. Journal of Trans-portation Systems Engineering and Information Technology, Vol. 7,No. 1, 2007, pp. 93–100.

8. Hale, D. Traffic Network Study Tool, TRANSYT-7F, U.S. Version T7F10.McTrans, University of Florida, Gainesville, 2006.

9. Stevanovic, A., P. T. Martin, and J. Stevanovic. VISSIM-Based GeneticAlgorithm Optimization of Signal Timings. In Transportation ResearchRecord: Journal of the Transportation Research Board, No. 2035,Transportation Research Board of the National Academies, Washington,D.C., 2007, pp. 59–68.

10. Stevanovic, J., A. Stevanovic, P. T. Martin, and T. Bauer. StochasticOptimization of Traffic Control and Transit Priority Settings in VISSIM.Transportation Research, Vol. 16C, No. 3, 2008, pp. 332–349.

11. Hallmark, S. L., and R. Guensler. Comparison of Speed and Accelera-tion Profiles from Field Data with NETSIM Output for Modal Air Qual-ity Analysis of Signalized Intersections. In Transportation ResearchRecord: Journal of the Transportation Research Board, No. 1664, TRB,National Research Council, Washington, D.C., 1999, pp. 40–46.

12. Rouphail, N. M., C. H. Frey, J. D. Colyar, and A. Unal. Vehicle Emis-sions and Traffic Measures: Exploratory Analysis of Field Observa-tions at Signalized Arterials. Presented at 80th Annual Meeting of theTransportation Research Board, Washington, D.C., 2001.

13. Marsden, G., M. Bell, and S. Reynolds. Towards a Real-Time MicroscopicEmission Model. Transportation Research, Vol. 6D, 2001, pp. 37–60.

14. Rakha, H. A., and Y. Ding. Impact of Vehicle Stops on Vehicle Energyand Emissions. Presented at 80th Annual Meeting of the TransportationResearch Board, Washington, D.C., 2001.

15. Unal, A., N. M. Rouphail, and H. C. Frey. Effect of Arterial Signaliza-tion and Level of Service on Measured Vehicle Emissions. In Transporta-tion Research Record: Journal of the Transportation Research Board,No. 1842, Transportation Research Board of the National Academies,Washington, D.C., 2003, pp. 47–56.

16. Midenet, S., F. Boillot, and J. C. Pierrelee. Signalized Intersection withReal-Time Adaptive Control: On-Field Assessment of CO2 and PollutantEmission Reduction. Transportation Research, Vol. 9D, 2004, pp. 29–47.

17. Li, X., G. Li, S. Pang, X. Yang, and J. Tian. Signal Timing of Intersec-tions Using Integrated Optimization of Traffic Quality, Emissions, andFuel Consumption: A Note. Transportation Research, Vol. 9D, 2004,pp. 401–407.

18. Coelho, M. C., T. L. Farias, and N. M. Rouphail. Impact of Speed Con-trol Traffic Signals on Pollutant Emissions. Transportation Research,Vol. 10D, 2005, pp. 323–340.

19. Park, J. Y., R. B. Noland, and J. W. Polak. Microscopic Model of AirPollutant Concentrations: Comparison of Simulated Results with Mea-sured and Macroscopic Estimates. In Transportation Research Record:Journal of the Transportation Research Board, No. 1750, TRB, NationalResearch Council, Washington, D.C., 2001, pp. 64–73.

20. Oda, T., M. Kuwahara, and S. Niikura. Traffic Signal Control for Reduc-ing Vehicle Carbon Dioxide Emissions on an Urban Road Network. InProceedings of the 11th World Congress on ITS, Nagoya, Japan, 2004.

21. Qu, T., L. R. Rilett, and J. Zietsman. Estimating Impact of Freeway SpeedLimits on Automobile Emissions. Presented at 82nd Annual Meeting ofthe Transportation Research Board, Washington, D.C., 2003.

22. Smith, K., K. Wood, and A. Ash. Managing Emissions from Vehiclesin Urban Systems. Presented at the AET European TransportationConference, Homerton College, Cambridge, United Kingdom, 2001.


23. SCOOT Traffic Handbook, Issue A 31-Dec-2001. Transport ResearchLaboratory Ltd., Crowthorne House, Wokingham, Berkshire, UnitedKingdom, 2001.

24. Goldberg, D. E. Genetic Algorithms in Search, Optimization and MachineLearning. Addison-Wesley Publishing Co. Inc., Reading, Mass., 1989.

25. Fellendorf, M. Integrated Simulation of Traffic Demand, Traffic Flow,Traffic Emissions and Air Quality. Proc., 8th International Sympo-sium on Transport and Air Pollution, Graz, Austria, 1999. www.english.ptv.de/download/traffic/library/1999%20ISTAP%20Symposium%20VISSIM.pdf.

26. Niittymaki, J., and M. Granberg. Development of Integrated Air Pol-lution Modeling Systems for Urban Planning: DIANA Project. Pre-sented at 80th Annual Meeting of the Transportation Research Board,Washington, D.C., 2001.

27. Malcom, C., G. Scora, and M. Barth. Validating a Micro-Scale Trans-portation/Emissions Model with Tunnel Study Data. Proc., 11th CRCOn-Road Vehicle Emissions Workshop, San Diego, Calif., 2001.

28. Chevallier, E. Microscopic Modeling Framework for Estimating Emis-sions from Traffic Management Policies. MS thesis. University of London,2005, pp. 19–24.

29. Kim, B. Y., R. L. Wayson, and G. Fleming. Development of the Traf-fic Air Quality Simulation Model. In Transportation Research Record:Journal of the Transportation Research Board, No. 1987, TransportationResearch Board of the National Academies, Washington, D.C., 2006,pp. 73–81.

30. User’s Guide to MOBILE6.1 and MOBILE6.2. U.S. Environmen-tal Protection Agency, 2003. www.epa.gov/otaq/models/mobile6/420r03010.pdf.

31. U.S. EPA Finalizes Tier 2 Standards and Limits on Gasoline Sulfur.U.S. Environmental Protection Agency, 2000. www.meca.org/galleries/default-file/mob21.pdf.

The Traffic Signal Systems Committee sponsored publication of this paper.

optimizing traffic control to reduce fuel consumption and vehicular emissions

Documents