Probabilistic Modeling of Acceleration in Traffic Networks as a Function of Speed and Road Type

Maya Abou-Zeid, Ismail Chabini, Edward K. Nam, and Alessandra Cappiello Probabilistic Modeling of Acceleration in Traffic Networks as a Function of Speed and Road Type Maya Abou-Zeid Massachusetts Institute of Technology 77 Massachusetts Aenue, Room 5- Cambridge, MA 39, USA Email: mga@mit.edu Tel: (67) 53-7338 Fax: (67) 58-873 Ismail Chabini (corresponding author) Massachusetts Institute of Technology 77 Massachusetts Aenue, Room -63 Cambridge, MA 39, USA Email: chabini@mit.edu Tel: (67) 53-464 Fax: (67) 58-873 Edward K. Nam Ford Motor Company Village Road, MD 383/ SRL Dearborn, MI 48 Email: enam@ford.com Tel: (33) 48-5833 Alessandra Cappiello Massachusetts Institute of Technology 77 Massachusetts Aenue, Room 5-4 Cambridge, MA 39, USA Email: alec@mit.edu Tel: (67) 53-79 Fax: (67) 58-873 TRB 3 Annual Meeting CD-ROM Paper reised from original submittal.

Maya Abou-Zeid, Ismail Chabini, Edward K. Nam, and Alessandra Cappiello Abstract: We propose a probabilistic approach to estimate the ariation of accelerations and decelerations in traffic networks as a function of speed and road type. We model acceleration as a random ariable and apply the probabilistic approach to a data set, for which we deelop a probabilistic acceleration distribution for eery speed range and road type. In almost all the cases, the goodness-of-fit of the modeled distribution function is acceptable, supporting the alidity of the probabilistic approach. It is seen that road type has little effect on the ariation of the distributions. Moreoe the standard deiation of the distributions decreases as the speed range increases. The deeloped model has a number of applications where only speed data are aailable, for example from field measurements or from a non-microscopic traffic model, and accelerations are sought. In this context, the acceleration model allows for more accurate emission modeling through the use of instantaneous emission models that require both speed and acceleration as input. The acceleration model can also be used in conjunction with speed data to model emissions as random ariables and generate their distributions for arious speed ranges and road types. TRB 3 Annual Meeting CD-ROM Paper reised from original submittal.

Maya Abou-Zeid, Ismail Chabini, Edward K. Nam, and Alessandra Cappiello 3 INTRODUCTION Characterizing trael behaior and ehicle actiity has been an important research topic that has numerous applications in traffic and emission modeling. Current trael demand models gie aerage speeds as outputs, which are not sufficient for the increasing data needs of emissions modelers. Recent emission research recognizes that the engine mode of operation will be a significant ariable in new modal emission models (,, 3, 4). When modal actiity exceeds specific thresholds of ariables such as powe positie kinetic energy, acceleration, or idle mode, emission leels can rise significantly. Thus, there is a need to understand how driing behaior and dynamic ehicular actiity affect the proportion of driing spent in different modes (idling, cruising, acceleration, deceleration, etc.). Sierra Research (5) deeloped representatie driing cycles for different facility types and congestion leels after analyzing instrumented and chase car data. While this research was a aluable addition to the literature, it does not present a statistical methodology for generating speed and acceleration distributions, which may be necessary for applications where specific driing cycles are an insufficient tool. Moreoe drier interaction with the ehicle in terms of acceleration and (to a lesser extent) deceleration patterns is important for emission modeling. For example, high-speed, high-acceleration, and heay braking actiities are typically exhibited by young male driers, and this in turn increases emission, while older driers might drie more conseratiely than younger driers (6, 7). Driing patterns might also be dependent on the particular city and the nature of its transportation network, i.e. whether it is mostly dominated by local streets or freeways (8), and on the traffic and control conditions (9). For these reasons, it is important to study those aspects of driing behaior and road characteristics that may influence the statistics of acceleration and deceleration eents for a gien speed. In this pape we deelop a noel approach for the quantification of acceleration and deceleration eents in a traffic network. The approach models acceleration as a random ariable whose distribution aries as a function of speed and road type. The basic motiation of this research is to integrate nonmicroscopic traffic models and instantaneous emission models, though there are other applications as well. Non-microscopic dynamic traffic assignment models are fast, applicable on a regional scale, and generally easier to calibrate than their microscopic counterparts. Their basic limitation is that they describe network conditions in terms of aerage link speeds, but do not proide accelerations. Load-based emission models, howee require both speed and acceleration as input. Therefore, a probabilistic acceleration model is an efficient method of oercoming the shortcomings of non-microscopic traffic models and proiding the necessary link to emission models. Een in the absence of a traffic model, the acceleration model is also useful when only speed data are aailable in a gien city from field measurements. This leads, for instance, to more accurate emission modeling using instantaneous emission models rather than aerage speed-based emission models (such as EPA s MOBILE6), which in principle do not properly quantify emissions from ehicles under dynamic conditions () and are thus an approximation at best. Seeral approaches for the quantification of acceleration and deceleration eents can be found in the literature. TRANSIMS, a traffic simulator based on cellular automata modeling, uses aggregate realworld frequencies of power factor (*speed*acceleration) to model the accelerations. The power factor is then used to estimate emissions (). Microscopic traffic models assign accelerations to indiidual ehicles based on the interaction among ehicles and the traffic regimes. Examples of these regimes are carfollowing and free-flowing. Therefore, in these models, acceleration is a function of parameters such as headway distribution, the relatie difference in speed between adjacent ehicles, and drier reaction time (). In this pape howee the focus is to deelop a probabilistic approach to estimate acceleration and deceleration actiity from the mere knowledge of speed, without modeling ehicular interactions at the microscopic leel. The analysis is conducted using data for four main road types: interstate highways, state highways, arterials, and collectors. It is well known that the potential for accelerating or decelerating decreases as speed increases because of power and traction limitations of the ehicle. Howee these limitations are usually insufficient to describe the dynamics of ehicles. It is also necessary to determine the statistical distribution of accelerations and decelerations within a gien speed range on a link, which is defined by the state ariables of density and flow (or aerage speed). TRB 3 Annual Meeting CD-ROM Paper reised from original submittal.

Maya Abou-Zeid, Ismail Chabini, Edward K. Nam, and Alessandra Cappiello 4 This paper is organized as follows. In the next section, we describe the data for which we deelop an acceleration model. This is followed by a presentation of the approach that we deeloped to calibrate the model. An analysis of the results is also gien. We then describe an application of the model. Finally we conclude the paper and gie future research directions. DATA The data sets of this research are obtained in conjunction with an intelligent cruise control study sponsored by the US Department of Transportation and conducted in South Eastern Michigan from July 996 to September 997 by the Uniersity of Michigan Transportation Research Institute (7). The model deeloped in this paper is built on real-world driing data collected during the first week of the study, during which the intelligent cruise control was not functional. 8 randomly-chosen driers from eight counties of South Eastern Michigan were selected to drie in metropolitan and rural areas of the state, using the same type of ehicle: an instrumented 996 Chrysler Concorde. Driers were classified into fie categories according to their driing behaior: Ultraconseratie: means an unusual tendency towards far ( far means large gap between leading and following ehicle) and/or slow driing. Planner: means an unusual tendency towards far and/or fast driing. Hunter/tailgater: means an unusual tendency towards fast and/or close driing. Extremist: means that the drier satisfies more than one of the aboe tendencies. Flow conformist: means that the drier satisfies none of the aboe tendencies. A flow conformist tends to trael at the same speed as other cars and at approximately the median headway time-gap. A sub-sample of eighteen driers, each conducting to 6 trips, was used to deelop the model presented in this paper. Most trip durations were less than 3 minutes. The eighteen driers belong to the following categories: two planners, three extremists, fie hunters, four ultraconseraties, and four flow conformists. The model does not capture differences in drier aggressieness because the intent is to isolate road type as the only independent ariable. Roads were classified into the following types: high-speed ramp, interstate highway, state highway, arterial, collecto light duty, alley or unpaed, unknown, and low-speed ramp. Only four road types (interstate highways, state highways, arterials, and collectors) are considered in the present study because of data aailability. Moreoe these road types coer most road types in a transportation network (except for on-ramps and off-ramps). The distribution of acceleration and deceleration data points, aggregated from all driers on eery road type, is shown in Table. These obserations correspond to second-by-second combinations of elocity and acceleration. CALIBRATION In this section, we describe the approach that we followed to deelop an acceleration model corresponding to the data described aboe. The same procedure can be used in general to derie statistical acceleration distributions from any gien acceleration data set though the fitted distributions might ary from one data set to another. The first fie trips hae been eliminated from eery drier s record to remoe some of the "noelty factor" bias that might be present at the beginning of the test since driers might not be accustomed to their new ehicles. The remaining data are diided into four subsets, one for each considered road type, to see whether acceleration and deceleration distributions are dependent on road type. The data in each subset are further diided into two groups: acceleration data (strictly positie alues) and TRB 3 Annual Meeting CD-ROM Paper reised from original submittal.

Maya Abou-Zeid, Ismail Chabini, Edward K. Nam, and Alessandra Cappiello 5 deceleration data (zero or negatie alues). Acceleration and deceleration are considered separately since in general they may not be similarly distributed. In the case of arterials, we hae diided the data into two subsets, S for calibration and C S for alidation. For the other road types, we did not alidate the model V because of the lack of a sufficient number of obserations for those road types. Consider one data group, for example, accelerations (strictly positie alues) on road type r. The statistical distribution analysis is conducted for km/h speed ranges. For each speed range, we would like to find a probability distribution that fits the acceleration distribution obtained from the sample for that particular speed range. Plotting the sample acceleration alues suggests a distribution similar (with a scale factor) to the density of a half-normal distribution with zero mean ( µ = ), and a standard deiation to be determined (see Fig. ). Let N, be the total number of acceleration obserations aggregated from all driers driing on a r road type r in a speed range. Let Q r, be the sample probability of obsering a certain alue of acceleration a, where a belongs to an acceleration interal of length. m/s ( e.g. (.,.4 ], (.4,.6 ],...) a belongs to an acceleration interal [ a, a ], we define T r, as the total number of acceleration obserations that are in the interal [ a, a ].Then Q r, is gien by: Q T = N. () The sample standard deiation is gien by the expression: ( a µ ) Q = ( a ) Q = a Q σ () = a A a A where A consists of all the acceleration alues in the data set, taken at. m/s interals. The half-normal probability density function, fitted to the acceleration data, is gien by: a A. If f a µ exp.5 σ = a exp. σ =. 5 πσ πσ (3) where σ is the standard deiation obtained from the acceleration sample, as gien by expression (). The half-normal distribution is truncated at some maximum acceleration alue a, so as to aoid predicting unrealistic high accelerations. The truncated distribution is then normalized so that the area under the resulting density function is one. The resulting density function is gien by: f normalized, = a f f da An error term ε, defined as the sum of the squares of the difference between the cumulatie sample probability F, r, F and the cumulatie truncated and normalized half-normal density function, r, is used to test the goodness-of-fit of the obtained distribution: [ F, F, ] ε = (5) a A (4) TRB 3 Annual Meeting CD-ROM Paper reised from original submittal.

Maya Abou-Zeid, Ismail Chabini, Edward K. Nam, and Alessandra Cappiello 6 Low alues of the error term ε indicate a good fit, while high alues suggest that the proposed distribution does not explain well the ariation in acceleration. This procedure was applied for eery road type and speed range in the acceleration data. The maximum acceleration alues at which the distributions are truncated are approximately equal to the maximum experimental alues of acceleration and deceleration for eery speed range, and they are gien by: a = 5 m/s for speed ranges from (-) to (7-8) km/h, a =. 5 m/s for speed range (8-9) km/h, a = m/s for speed range (9-) km/h, a =. 75 m/s for speed range (-) km/h, and a =.5 m/s for speed range (-) km/h. The steps described aboe to estimate accelerations are applicable to the deceleration data for eery road type and speed range. * The density function f of the combined distribution of acceleration and deceleration corresponding to road type r and speed range is obtained by weighing the truncated and normalized distribution with the probability of occurrence of acceleration and deceleration, respectiely: f * Pr,. = Pr,. a a f f f f da da for a for a > (6) where P, and P, are the probabilities of a deceleration realization and an acceleration realization, r r respectiely, obtained from the sample data for road type r and speed range. The other terms in expression (6) are as preiously defined. In the remainder of this pape we refer to the truncated normalized distribution as the halfnormal distribution. ANALYSIS OF RESULTS Comparison of Obsered Distributions to Fitted Half-Normal Distributions The error terms ε obtained upon fitting half-normal distributions (with zero mean and the standard deiations obtained from the sample) to the obsered alues of acceleration and deceleration are shown in Table (part a) for the four road types. These error alues are satisfactorily low. They support the alidity of the hypothesis that the accelerations and decelerations are probabilistically distributed, with the fitted distribution in this case being half-normal with zero mean and a standard deiation that decreases as the speed increases. Note that the error term alues of deceleration distributions are in most cases lower than those of acceleration distributions because of the aailability of more deceleration data and the absence of any acceleration alue in the interal (,.]. Howee error term alues are similar among different road types although there is an uneen distribution of data points among the road types. Fig. shows the cumulatie sample probability F and the cumulatie modeled distribution function F for accelerations (part a) and decelerations (part b), respectiely, on arterials in subset S C TRB 3 Annual Meeting CD-ROM Paper reised from original submittal.

Maya Abou-Zeid, Ismail Chabini, Edward K. Nam, and Alessandra Cappiello 7 for the speed range ( ) km/h. The goodness-of-fit is better for the deceleration data than for the,., as stated aboe. acceleration data due to lack of a sufficient number of obserations in the interal ( ] Validation Validation of the half-normal distributions was done only for the data of arterials because it contained enough obserations to allow for both calibration and alidation. For eery speed range, the same halfnormal distribution that was fitted to the acceleration data in subset S was compared to the acceleration C data of subset S to test whether the distributions deeloped from a certain sample can be applied to V another sample for the same road type. Validation was also done for the deceleration data on arterials. In both cases, the error terms obtained were acceptable, supporting the adoption of probabilistic models to estimate accelerations and decelerations. Fig. 3 shows the cumulatie sample probability of the acceleration data (part a) and deceleration data (part b) in subset S and the cumulatie distribution corresponding to V the modeled density function (which was deried from the calibration data set S ) on arterials for the C speed range (-) km/h. The error terms for all speed ranges are shown in Table (part b). Variation of Standard Deiation of the Distributions with Speed Range and Road Type Fig. 4 shows the ariation of standard deiation of acceleration (part a) and deceleration (part b) distributions among speed ranges for the four road types. Beyond a certain speed threshold, the standard deiations decrease with speed because at lower speeds, there is a higher probability of achieing high alues of acceleration and deceleration than at larger speeds. This diersity of acceleration and deceleration at lower speeds is an important phenomenon in estimating emissions related to engine load or power enrichment. This phenomenon is iolated for the first speed range, where the standard deiation increases when moing from ( ) km/h to ( ) km/h. The reason for this phenomenon could be that the maximum power aailable to ehicles at ery low speeds is lower than what driers desire as acceleration. This obseration might also be due to the common "lurch" (sudden positie ariation in acceleration) that follows a light change, or stop and go traffic, and might be made pronounced by the inexperience of the driers with a new ehicle throttle. This effect was also obsered in (8). Road type is seen to hae little effect on the ariation of acceleration and deceleration distributions, as shown in Fig. 4. While it is expected that stop and go conditions, characteristic of collectors and arterials, might lead to higher acceleration and deceleration alues, the results actually indicate that highways hae similar standard deiations to those of arterials and collectors, and in some cases hae een higher ariations. Note that higher speeds can be reached on highways than on arterials and collectors. MODEL APPLICATION In this section, we describe a general procedure for a possible application of the acceleration model. Then we depict one instance where the procedure has been applied. This application has been motiated by the integration of a non-microscopic traffic model and an instantaneous emission model. General Procedure The probabilistic nature of the acceleration model leads to a noel approach for emission modeling. Since accelerations are modeled as random ariables, emission factors which are functions of speed and acceleration will themseles be random ariables. Therefore, an instantaneous emission model combined with a probabilistic acceleration model can generate, for eery road type and speed range, a probabilistic emission distribution from which one can obtain multiple moments of emissions (expected alue, standard deiation, etc.). The approach summarized in the preious paragraph is documented in more details in (3). For a gien emission species i, ehicle category c, speed range, and road type r, an expected emission factor TRB 3 Annual Meeting CD-ROM Paper reised from original submittal.

Maya Abou-Zeid, Ismail Chabini, Edward K. Nam, and Alessandra Cappiello 8 e is calculated based on the probability of occurrence of eery acceleration and deceleration, and is i, c, r, gien by: a [ e, a) ] = e (, a) * i, c, r, = E i, c (, r, [, ] a i c f a a a e i c, a e da (7) In expression (7), ( ) is the emission facto obtained from any instantaneous emission model, for, emission species i, ehicle category c, speed, and acceleration a. a and a are the highest deceleration and acceleration realizations, respectiely, in speed range, as obtained from the sample data. * f is gien by expression (6). The expected emission factor is obtained by discretizing acceleration and deceleration alues in a, a. Its expression is: the interal [ ] = ( a) e i, c, ei, c. a S a, π (8) In the latter expression, { a h, a h /,...,. a 3h /, a / } S a = is the discretization interal, / h a h / and h can be set to any desired alue. Here it is set to. m/s. * π = f ( x). probability that the acceleration belongs to the interal ( a h /, a h / ) a h / dx, which is the This general procedure can be employed in two types of applications. First, it is useful for the integration of non-microscopic traffic models and emission models. In this case, the expected emission factors can be applied to the aerage speeds which are output by the traffic model in order to predict emissions. A specific application of this type is shown below. Second, the procedure can be used to enhance the accuracy of emission models predictions in cases where speed is obtained from field measurements, for example through loop detectors, and used as input to the acceleration model which would generate acceleration distributions for a gien road type. Any instantaneous emission model would then be able to predict emission distributions (or moments of emissions), gien the speed and acceleration (as well as other ehicle and roadway-related factors). Therefore, the acceleration model allows the deployment of more refined and detailed emission models in practice, as it allows the determination of acceleration, which is a quantity not measured in practice, ia the measurement of speed only. Application Example Below we describe a specific application where expected emission factors hae been generated based on speed data obtained from field measurements. The acceleration model has been used in (3) in conjunction with EMIT (EMIssions from Traffic), a recently deeloped emission and fuel consumption model. We proide a brief description of EMIT, show results of expected emission factors deried from EMIT, and describe the integration of EMIT with a non-microscopic traffic model. EMIT is a simple statistical model for instantaneous tailpipe emissions ( CO, CO, HC, and NO ) and fuel consumption of light-duty composite ehicles. In order to realistically reproduce the x behaior of the emissions, the explanatory ariables in EMIT hae been deried from the load-based approach, using some simplifying assumptions. The model is calibrated for a set of ehicles drien on standard as well as aggressie driing cycles, and is alidated on another driing cycle in order to assess its estimation capabilities. The preliminary results indicate that the model gies reasonable results compared to actual measurements as well as to results obtained with CMEM, a well-known load-based emission model (see Fig. 5). The goodness-of-fit of EMIT aries with different emission species (see Table 3), but the model has in general a reasonable predictie accuracy. Furthermore, the model, due to its simple structure, is relatiely easy to calibrate and requires less computational time than detailed load-based models. A detailed description of EMIT can be found in (3, 4). Expected emission factors hae been calculated in adance (off-line), according to the general procedure outlined aboe. The speed data used to compute expected emission factors are obtained from the data set described in this paper. Fig. 6 and Fig. 7 show the calculated expected emission factors of CO, TRB 3 Annual Meeting CD-ROM Paper reised from original submittal.

Maya Abou-Zeid, Ismail Chabini, Edward K. Nam, and Alessandra Cappiello 9 CO, HC, and NO as well as fuel rates for ehicle category 9 (defined in 3) as a function of speed on x arterials and highways, respectiely. In general, the expected emission factor (g/s) increases with speed because of the increase in fuel consumption rate. The expected emission factors are also compared with the facility-specific emission rates from MOBILE6. Note also that expected emission factors would in general be different for different ehicle categories. An integration component is designed to apply the expected emission factors to the output (i.e. time-dependent link speeds) of a mesoscopic traffic model, deeloped in (5), to predict total emissions as well as their spatial and temporal ariations. The combined model allows the ealuation of arious traffic management strategies and their effectieness in reducing traffic congestion, air pollution, and fuel consumption. For instance, in (3) arious scenarios of traffic conditions (with and without an incident) are considered to assess the impact of dynamic route guidance (an Intelligent Transportation Systems traffic management method) on trael time, emissions, and fuel consumption. CONCLUSION In this pape a probabilistic approach that models acceleration actiity as a random ariable has been deeloped. Statistical acceleration and deceleration distributions hae been deeloped as a function of realworld data of ehicle speeds and road types. As the speed range increases, the standard deiation of the acceleration and deceleration distributions decreases because at higher speeds only a limited range of accelerations and decelerations can be achieed due to power and traction limitations. This obseration was consistent among all road types. Moreoe the standard deiations are similar among road types, which might suggest that road type has little effect, if any, on acceleration and deceleration ariation. Howee this effect should be studied further with more data from other cities, since there is reason to beliee that driing behaior differs from city to city, especially those that hae more hills (8). For eery road type and speed range, and for both acceleration and deceleration, a half-normal distribution haing the same mean and standard deiation as the original data was fitted to the obserations. The fitted distribution was truncated at some maximum acceleration alue in order to consider only physically feasible accelerations. In almost all cases, the fit was ery close as indicated by low error term alues. This implies that the half-normal distribution well approximates the acceleration and deceleration actiity distributions for the gien data. The specific parameters of the distribution might hae to be calibrated separately for each city since there might be other factors, not captured by our model, that affect these distributions. Moreoe the distribution that would fit other acceleration and deceleration data from different regions might not be half-normal. Howee the same methodology deeloped in this paper can be used to deelop other acceleration probability distributions. A general procedure was gien to illustrate an application of the probabilistic modeling approach. Then specific results were proided where the acceleration model was used in conjunction with EMIT (3, 4), an instantaneous emission model, to generate expected emission factors for the purpose of integration with a non-microscopic traffic model. For further research, it would be useful to apply the methodology deeloped in this paper to other data sets (namely the Sierra chase car data) to inestigate further the nature of the fitted distributions as well as the effect of road type on these distributions. It would also be important to quantify the actiity from freeway ramps. Moreoe it would be interesting to disaggregate this model to assess the impact of drier aggressieness and ehicle type on the ariation of acceleration and deceleration distributions. TRB 3 Annual Meeting CD-ROM Paper reised from original submittal.

Maya Abou-Zeid, Ismail Chabini, Edward K. Nam, and Alessandra Cappiello ACKNOWLEDGEMENTS We would like to express our gratitude for Daid Chock from the Ford Scientific Research Laboratory for his wise technical counsel. This research benefited from financial support from the Ford Motor Company through the Ford-MIT Alliance. REFERENCES. Barth, M. Integrating a Modal Emissions Model into Various Transportation Modeling Frameworks. Transportation Planning and Air Quality III, Emerging Strategies and Working Solutions (S. Washington, ed.), ASCE, New York, 998, pp. 38-5.. Guensle R. S., S. Washington, and W. Bachman. Oeriew of the MEASURE Modeling Framework. Transportation Planning and Air Quality III, Emerging Strategies and Working Solutions (S. Washington, ed.), ASCE, New York, 998, pp. 5-7. 3. Washington, S., J. D. Leonard II, C. A. Roberts, T. Young, D. Sperling, and J. Botha. Forecasting Vehicle Modes of Operation Needed as Input to Modal Emissions Models. International Journal of Vehicle Design, Vol., Nos. -4 (Special Issue), 998, pp. 35-359. 4. Roberts, C. A., S. Washington, and J. D. Leonard II. Forecasting Dynamic Vehicular Actiity on Freeways: Bridging the Gap Between Trael Demand and Emerging Emissions Models. In Transportation Research Record 664, TRB, National Research Council, Washington, D. C., 999, pp. 3-39. 5. Carlson, T.R., and T. C. Austin. Deelopment of Speed Correction Cycles. Draft report prepared for the Enironmental Protection Agency, 997. 6. Wolf, J., R. Guensle S. Washington, W. Sarasua, C. Grant, S. Hallmark, M. Olieira, M. Koutsak, R. Thittai, R. Funk, and J. Hsu. Deelopment of a Comprehensie Vehicle Instrumentation Package for Monitoring Indiidual Tripmaking Behaior. Georgia Institute of Technology, Final Report GTI-R -995, April 999. 7. Fanche P. et al. Intelligent Cruise Control Field Operational Test. Final Report. U.S. Department of Transportation, May 998. 8. LeBlanc, D. C., F. M. Saunders, M. D. Meye and R. Guensler. Driing Pattern Variability and Impacts on Vehicle Carbon Monoxide Emissions. In Transportation Research Record 47, TRB, National Research Council, Washington, D. C., 995, pp. 45-5. 9. Special Report 9: Highway Capacity Manual, 3 rd ed. TRB, National Research Council, Washington, D.C., 998.. National Research Council Committee to Reiew EPA s Mobile Source Emissions Factor (MOBILE) Model. Modeling Mobile-Source Emissions. National Academy Press,.. Williams, M. D., G. Thaye M. J. Barth, and L. L. Smith. The TRANSIMS Approach to Emissions Estimation. Los Alamos National Laboratory, 999.. Ahmed, K. I. Modeling Driers Acceleration and Lane Changing Behaior. Ph.D. Thesis, Massachusetts Institute of Technology, Cambridge, MA, 999. 3. Cappiello, A. Modeling Traffic Flow Emissions. M.S. Thesis, Massachusetts Institute of Technology, Cambridge, MA,. TRB 3 Annual Meeting CD-ROM Paper reised from original submittal.

Maya Abou-Zeid, Ismail Chabini, Edward K. Nam, and Alessandra Cappiello 4. Cappiello, A., I. Chabini, E. K. Nam, A. Luè, and M. Abou-Zeid. A Statistical Model of Vehicle Emissions and Fuel Consumption. Proc. of the IEEE 5 th International Conference on Intelligent Transportation Systems, Singapore, Sept., pp. 8-89. 5. Bottom, J. A. Consistent Anticipatory Route Guidance. Ph.D. Thesis, Massachusetts Institute of Technology, Cambridge, MA,. TRB 3 Annual Meeting CD-ROM Paper reised from original submittal.

Maya Abou-Zeid, Ismail Chabini, Edward K. Nam, and Alessandra Cappiello LIST OF TABLES TABLE : Number of Obserations by Road Type. TABLE : Error Terms of Half-Normal Distributions Fitted to the Sample Acceleration and Deceleration Distributions, Obtained from Calibration on All Road Types (Part a) and Validation on Arterials (Part b). TABLE 3. R-square (R ) Between the Measured and the Predicted Emission (or Fuel Consumption) Rates from EMIT. Part a: Results for Calibration. Part b: Results for Validation. (from (3)). LIST OF FIGURES FIGURE : Acceleration Distribution for the Calibration Data Set on Arterials for the Speed Ranges -3 km/h (Part a), 5-6 km/h (Part b), and 8-9 km/h (Part c). FIGURE : Cumulatie Sample and Half-Normal Distribution Functions for the Acceleration Data (Part a) and the Deceleration Data (Part b) Used for Calibration on Arterials for the Speed Range - km/h. FIGURE 3: Cumulatie Sample and Half-Normal Distribution Functions for the Acceleration Data (Part a) and the Deceleration Data (Part b) Used for Validation on Arterials for the Speed Range - km/h. FIGURE 4: Variation of Standard Deiation of Acceleration Distributions (Part a) and Deceleration Distributions (Part b) Among Different Speed Ranges and Road Types. FIGURE 5: Category 9 - FTP bag. Second-by-Second Engine-Out (EO) and Tailpipe (TP) Emission Rates of CO and CO: Measurements (Calibration Data), EMIT Predictions, and CMEM Predictions. (from (3)). FIGURE 6. Expected Emission Rates in g/s and in g/km for Road Type Arterial and Vehicle Category 9. (from (3)). FIGURE 7. Expected Emission Rates in g/s and in g/km for Road Type Highway and Vehicle Category 9. (from (3)). TRB 3 Annual Meeting CD-ROM Paper reised from original submittal.

Maya Abou-Zeid, Ismail Chabini, Edward K. Nam, and Alessandra Cappiello 3 TABLE. Number of Obserations by Road Type. Road Type Interstate Highway State Highway Arterial Collector Acceleration 5,597 3,633,58 5,84 Deceleration,569 6,74 3,586,43 TRB 3 Annual Meeting CD-ROM Paper reised from original submittal.

Maya Abou-Zeid, Ismail Chabini, Edward K. Nam, and Alessandra Cappiello 4 TABLE *. Error Terms of Half-Normal Distributions Fitted to the Sample Acceleration and Deceleration Distributions, Obtained from Calibration on All Road Types (Part a) and Validation on Arterials (Part b). Part a Speed Acceleration Deceleration Range Interstate State Arterial Collector Interstate State Arterial Collector (km/h) Highway Highway Highway Highway -.895.89.6.63.46..6.87 -.588.844.77.549.56.69.43.8-3.374.49.43.6.7.97..5 3-4.5.68.57.834.35.934.448.337 4-5.6.659.853.68.494.4996.37.697 5-6.55.58.469.396.93.495.98.3373 6-7.346.368.444.489.863.8.34.95 7-8.4769.59.463.5967.79.333.39.35 8-9.63.795.739.8689.64.36.9.35 9-.7.8.985.967.64.3.34 N/A -.3349 N/A N/A N/A.5 N/A N/A N/A -.556 N/A N/A N/A.47 N/A N/A N/A * N/A indicates that not enough data were aailable to calibrate a model for the corresponding speed range and road type. Part b Speed Range (km/h) Acceleration Deceleration -.93.94 -.8.65-3.333.97 3-4.487.496 4-5.5469.97 5-6.55.333 6-7.5648.599 7-8.7844.384 8-9.668.65 TRB 3 Annual Meeting CD-ROM Paper reised from original submittal.

Maya Abou-Zeid, Ismail Chabini, Edward K. Nam, and Alessandra Cappiello 5 Table 3. R-square (R ) Between the Measured and the Predicted Emission (or Fuel Consumption) Rates from EMIT. Part a: Results for Calibration. Part b: Results for Validation. (from (3)). Part a CO CO HC NO X FR Engine-out module, category 7.98.87.58.86.97 Tailpipe module, category 7.98.84.53.79 Engine-out module, category 9.97.9.63.87.97 Tailpipe module, category 9.97.88.58.67 Part b CO CO HC NO X FR Engine-out module, category 7.96.46.5.83.94 Tailpipe module, category 7.96.36..63 Engine-out module, category 9.95.5..83.95 Tailpipe module, category 9.95.43.3.53 TRB 3 Annual Meeting CD-ROM Paper reised from original submittal.

..6.4.8..6 3.4 3.8 4. 4.6 Maya Abou-Zeid, Ismail Chabini, Edward K. Nam, and Alessandra Cappiello 6 Number of Obserations 5 5 5..6.4.8..6 3 3.4 3.8 4. 4.6 5 Part a Acceleration (m/s ) Number of Obserations 8 6 4..6.4.8..6 3 3.4 3.8 4. 4.6 5 Part b Acceleration (m/s ) 5 Number of Obserations 5 5 3 5 Acceleration (m/s ) Part c FIGURE : Acceleration Distribution for the Calibration Data Set on Arterials for the Speed Ranges -3 km/h (Part a), 5-6 km/h (Part b), and 8-9 km/h (Part c). TRB 3 Annual Meeting CD-ROM Paper reised from original submittal.

Cumulatie probability Maya Abou-Zeid, Ismail Chabini, Edward K. Nam, and Alessandra Cappiello 7 Cumulatie probability.8.6.4..5.5.5 3 3.5 4 4.5 5 Acceleration (m/s ) Cumulatie sample probability Cumulatie modeled distribution function Part a Cumulatie sample probability Cumulatie modeled distribution function.8.6.4. -5-4.5-4 -3.5-3 -.5 - -.5 - -.5 Deceleration (m/s ) Part b FIGURE : Cumulatie Sample and Half-Normal Distribution Functions for the Acceleration Data (Part a) and the Deceleration Data (Part b) Used for Calibration on Arterials for the Speed Range - km/h. TRB 3 Annual Meeting CD-ROM Paper reised from original submittal.

Cumulatie probability Maya Abou-Zeid, Ismail Chabini, Edward K. Nam, and Alessandra Cappiello 8 Cumulatie probability.8.6.4..5.5.5 3 3.5 4 4.5 5 Acceleration (m/s ) Cumulatie sample probability Cumulatie modeled distribution function Part a Part a Cumulatie sample probability Cumulatie modeled distribution function.8.6.4. -5-4.5-4 -3.5-3 -.5 - -.5 - -.5 Deceleration (m/s ) Part b FIGURE 3: Cumulatie Sample and Half-Normal Distribution Functions for the Acceleration Data (Part a) and the Deceleration Data (Part b) Used for Validation on Arterials for the Speed Range - km/h. TRB 3 Annual Meeting CD-ROM Paper reised from original submittal.

Maya Abou-Zeid, Ismail Chabini, Edward K. Nam, and Alessandra Cappiello 9.6.4 Part a Standard Deiation (m/s )..8.6.4. (-) (-) (-3) (3-4) (4-5) (5-6) (6-7) Velocity (km/h) (7-8) (8-9) (9-) (-) (-) Interstate highway State highway Arterial Collector.6 Standard Deiation (m/s ).4..8.6.4. Interstate highway State highway Arterial Collector (-) (-) (-3) (3-4) (4-5) (5-6) (6-7) (7-8) (8-9) Velocity (km/h) (9-) (-) (-) Part b FIGURE 4: Variation of Standard Deiation of Acceleration Distributions (Part a) and Deceleration Distributions (Part b) Among Different Speed Ranges and Road Types. TRB 3 Annual Meeting CD-ROM Paper reised from original submittal.

Maya Abou-Zeid, Ismail Chabini, Edward K. Nam, and Alessandra Cappiello EOCO (g/s) TPCO (g/s) EOCO (g/s) TPCO (g/s) 6 4 5..5. 5..5..4.7.... 5 6 7 8 9 3 5 6 7 8 9 3 5 6 7 8 9 3 5 6 7 8 9 3 5 6 7 8 9 3 Time (s) FIGURE 5. Category 9 - FTP bag. Second-by-Second Engine-Out (EO) and Tailpipe (TP) Emission Rates of CO and CO. Thick light line: measurements (calibration data); dark line: EMIT predictions; thin line: CMEM predictions. The top plot represents the speed trace. (from (3)) TRB 3 Annual Meeting CD-ROM Paper reised from original submittal.

Maya Abou-Zeid, Ismail Chabini, Edward K. Nam, and Alessandra Cappiello 3 FR (g/s).5.5 FR (g/km) 4 6 8 4 6 8 6 8 CO (g/s) 4 CO (g/km). 6 4 4 6 8 4 6 8.6 3 CO (g/s).4. CO (g/km) 4 6 8 4 6 8.9.8 HC (g/s).6.3 HC (g/km).6.4. 4 6 8 4 6 8.6.3 NOx (g/s).4. NOx (g/km)... 4 6 8 4 6 8 FIGURE 6. Expected Emission Rates in g/s (on the left) and in g/km (on the right) for Road Type Arterial and Vehicle Category 9. The expected emission rates in g/km of CO, HC, and NO X are compared with the facility-specific emission rates from MOBILE6 (thin line). (from (3)). TRB 3 Annual Meeting CD-ROM Paper reised from original submittal.

Maya Abou-Zeid, Ismail Chabini, Edward K. Nam, and Alessandra Cappiello FR (g/s).5.5 6 4 6 8 FR (g/km) 3 8 4 6 8 CO (g/s) 4 CO (g/km). 6 4 4 6 8 4 6 8.6 3 CO (g/s).4. CO (g/km) 4 6 8 4 6 8.9.8 HC (g/s).6.3 HC (g/km).6.4. 4 6 8 4 6 8.6.3 NOx (g/s).4. NOx (g/km)... 4 6 8 4 6 8 FIGURE 7. Expected Emission Rates in g/s (on the left) and in g/km (on the right) for Road Type Highway and Vehicle Category 9. The expected emission rates in g/km of CO, HC, and NO X are compared with the facility-specific emission rates from MOBILE6 (thin line). (from (3)). TRB 3 Annual Meeting CD-ROM Paper reised from original submittal.