ANALYSIS AND PREDICTION OF INDIVIDUAL VEHICLE ACTIVITY FOR MICROSCOPIC TRAFFIC MODELING. A Thesis Presented to The Academic Faculty

Transcription

1 ANALYSIS AND PREDICTION OF INDIVIDUAL VEHICLE ACTIVITY FOR MICROSCOPIC TRAFFIC MODELING A Thesis Presented to The Academic Faculty By Shauna L. Hallmark In Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy in Civil and Environmental Engineering Georgia Institute of Technology December 1999

2 TABLE OF CONTENTS THESIS APPROVAL ii TABLE OF CONTENTS......iii LIST OF FIGURES ix LIST OF TABLES...xi GLOSSARY / ACRONYMS.. xiv SUMMARY...xvii 1. INTRODUCTION BACKGROUND Automobile Exhaust Emission Ozone Carbon Monoxide Oxides of Nitrogen Pm Hydrocarbons Drawbacks to Traditional Emission Modeling Vehicle Activity Speed Estimates Volume Estimates Emission Rates TOWARDS A MODAL APPROACH FOR TRANSPORTATION- RELATED EMISSION MODELING Evidence of a Mode Specific Emission Relationship Tunnel Studies Activity Outside the FTP 25 iii

3 3.1.3 Enrichment.. 26 iv

4 Acceleration Grade Air Conditioner Use Rapid Load Reduction Towards a Modal Approach Improved Emission Factor Estimates Improved Vehicle Activity Estimates On-Road Vehicle Activity Modeling Simulation MEASURE Fundamentals of Vehicle Activity in Traffic Engineering Acceleration Performance of Passenger Cars Acceleration Performance of Heavy Trucks Deceleration Performance Discussion RESEARCH APPROACH Statement of Problem Hypothesis to be Tested Objectives Scope of Work Statistical Modeling Chi-Square Test Kolmogorv-Smirnov Two-Sample Linear Regression Hierachiacal Based Regression Tree Analysis Description of Test Applicability of Test to Research v

5 4.6 Research Scope and Presentation of Statistical Approach Response Variables Carbon Monoxide Model Hydrocarbon Model Oxides of Nitrogen Final Response Variables Independent Variables for Vehicle Activity Data Driver Variables Trip Purpose Demographics Vehicle Variables Roadway Variables Horizontal and Vertical Curvature Grade Distance Between Adjacent Intersections Number of Lanes Lane Width Speed Limit Environmental Factors Pavement Condition Weather Other Factors Pedestrian Activity Location Along Segment Physical Location of Site Queue Position Operational Characteristics Level of Service Volume to Capacity Volume Density Fleet Mix DATA PROTOCOLS. 103 vi

6 5.1 Data Collection Selection of Sampling Locations Advantage Laser Rangefinder JAMAR Boards Vehicle Attribute Data Site Attributes Data Collection Protocol Data Handling Laser Rangefinder RANGE.C Program ATTACH.C Stopline Distances Volume Calculations Percent Heavy Vehicle Calculations LOS and V/C Ratio Data Collection Sites PRESENTATION OF DATA Data Preparation Data Analysis Identification of Microscopic Activity Distribution Dependent Variables Identification of Microscopic Activity Distribution Independent Variables Results of Statistical Analysis for Passenger Cars Activity for Queue Vehicles From Stopping Point to 200 Feet Downstream ACCEL Model Percent Activity >= 6.0 mph/s (ACCEL.6) Percent Activity >= 3.0 mph/s (ACC.3) Percent Activity <= -2.0 mph/s vii

7 (DEC.2) Average Vehicle Speed (AVGSPD) Inertial Power Surrogate >= 120 mph 2 /s (IPS120) Summarization of Results for ACCEL Final Predictor Model for ACCEL Model Validation for ACCEL Final Model for Queued Vehicles for ACCEL Activity for Queued Vehicles From 200 to 400 Feet Downstream of Initial Stopping Point (ACCELPLUS Activity for Queued Vehicles From 400 to 600 Feet Downstream of Initial Stopping Point (ACCELPLUS400) Activity for Queued Vehicles From 600 to 1,000 Feet Downstream of Initial Stopping Point (ACCELPLUS600 and ACCELPLUS800) Activity for Queued Vehicles From Initial Stopping Point Upstream 200 Feet (DECEL) Activity for Queued Vehicles From 200 Feet Upstream of the Initial Stopping Point to a Point 400 Feet Upstream (DECELNEG200) Activity for Queued Vehicles From 400 Feet Upstream of the Initial Stopping Point to 600 Feet Upstream (DECELNEG400) Activity for THRU Vehicles at all Locations Heavy Trucks Activity for Queued Vehicles From Initial Stopping Point to 200 Feet Downstream (ACCEL) Activity for Queued Vehicles From 200 Feet Downstream to 800 Feet Downstream (ACCELPLUS200 to ACCELPLUS600) Activity for Queued Vehicles From 200 Feet Upstream to Stopping Point (DECEL) Activity for Queued Vehicles From 600 to 200 Feet viii

8 Upstream of Initial Stopping Point Activity for THRU Vehicles Comparison of Research to Existing Simulation Modeling Ranges of Field Data Comparison of Research to Existing Simulation Modeling Comparison of Research to Traffic Engineering Rates Comparison of Research to NCHRP Comparison of Research to FTP Range of Activity DISCUSSION AND CONCLUSIONS Model Limitations Future Research Needs Conclusions REFERENCES APPENDIX A1 204 APPENDIX A2 219 ix

9 LIST OF FIGURES Figure 2-1, Traditional Emission Modeling. 14 Figure 2-2, MOBILE Emissions Versus Speed Range for Carbon Monoxide 17 Figure 3-1, Modal Elements of a Vehicle Trip 34 Figure 3-2, Linear Speed-Acceleration Curve Figure 3-3, Maximum Acceleration on Upgrades for Passenger Cars by Speed Figure 4-1, Sample Vehicle Trace 62 Figure 4-2, Joint Acceleration-Speed Probability Density Function Figure 4-3, Comparison of Empirical cdfs for Acceleration on a 9% Grade (x) and -9% Grade (z). 69 Figure 4-5, Graduated Relationship Between Percent Hard Accelerations and Queue Position. 72 Figure 5-1, Data Collection and Reduction Methodology.111 Figure 5-2, LRF Geometry Accounted for in RANGE70.C..115 Figure 6-1, Schematic of Data Partions.132 Figure 6-2, Correlation Between V/C and Upstream Per Lane Volume (R 2 = 0.64) 142 Figure 6-3, Original Untrimmed Regression Tree Model for ACC6.146 Figure 6-4, Reduction in Deviance with the Addition of Nodes x

10 Figure 6-5, Normal Probability Plot of the Residuals for the Original Untrimmed Tree..147 xi

11 Figure 6-6, Trimmed ACC6 Model 148 Figure 6-7, Trimmed ACC3 Model Figure 6-8, Trimmed DECEL2 Model Figure 6-9, Trimmed AVG_SPD Model 153 Figure 6-10, Trimmed IPS120 Model 154 Figure 6-11, Comparison of CDFs for Dataset Out1 and Out Figure 6-12, Acceleration Distributions (mph/s) by Speed Ranges (mph) 173 Figure 6-13, Comparison of Time Spent in Each Acceleration Range for Field Data and NETSIM (-250 to 250 feet from the stopbar) 176 Figure 6-14, Comparison of Time Spent in Each Speed Range for Field Data and NETSIM (-250 to 250 feet from the stopbar) 176 Figure 6-15, Comparison of Time Spent in Each Acceleration Range for Field Data and NETSIM (-250 to 250 feet from the stopbar) (Midblock) 178 Figure 6-16, Comparison of Time Spent in Each Speed Range for Field Data and NETSIM (-250 to 250 feet from the stopbar) (Midblock) 178 Figure 6-17, Comparison of Field Data for First Vehicle in Queue with Linear Speed-Acceleration Relationship 182 Figure 6-18, Acceleration Distribution (mph/s) by Speed Ranges (mph).178 xii

12 LIST OF TABLES Table 3-1, Maximum Acceleration From Rest by Vehicle Type and Weight-to- Power Ratio..50 Table 3-2, Maximum Acceleration by Speed Range by Vehicle Type and Weight-to-Power Ratio 50 Table 3-3, Maximum Acceleration on Upgrades by Speed Range 51 Table 4-1, Joint Acceleration-Speed Probability Density Function Table 4-2, Modal Predictor Variables for Emission Rate Analysis for Passenger Cars.88 Table 4-3, Operational and Geometric Factors Hypothesized to Affect Modal Activity.90 Table 5-1, Example Data Collection Attribute Sheet 109 Table 5-2, Example Output from RANGE Table 5-3, Final Dataset Format 118 Table 5-4, Data Collection Sites Table 6-1, Data Partioning 131 Table 6-2, Full Untrimmed Regression Tree Results for ACC Table 6-3, Trimmed ACC6 Model Results Table 6-4, Trimmed ACC3 Model Results Table 6-5, Trimmed DECEL2 Model Results..150 xiii

13 Table 6-6, Trimmed AVG_SPD Model Results xiv

14 Table 6-7, Trimmed IPS120 Model Results.153 Table 6-8, Breakpoints for Data Stratification from Initial Queue Position Downstream 200 Feet Table 6-9, K-S Test Statistic for Comparison of Datasets 1 and 10 for Acceleration Distributions. 157 Table 6-10, K-S Test Statistic for Comparison of Datasets 1 and 10 for Speed Distributions 157 Table 6-11, Breakpoints for Data Stratification From the Initial Queue Position Downstream 200 Feet. 159 Table 6-12, Breakpoints for Data Stratification From 200 to 400 Feet Downstream of the Initial Queue Position 159 Table 6-13, Breakpoints for Data Stratification from 400 to 600 Feet Downstream of the Initial Queue Position 160 Table 6-14, Breakpoints for Data Stratification from 600 to 1000 Feet Downstream of the Initial Queue Position 162 Table 6-15, Breakpoints for Data Stratification from Initial Stopping Point Upstream 200 Feet 163 Table 6-16, Breakpoints for Data Stratification from 200 to 400 Feet Upstream of the Initial Queue Position 163 Table 6-17, Breakpoint for Data Stratification for THRU Vehicles for All Distances Upstream and Downstream of the Data Collection Site 165 Table 6-18, Breakpoints for Queued Heavy Vehicles from Initial Queue Position Downstream 200 Feet Table 6-19, Breakpoints for Queued Heavy Vehicles from 200 to 600 Feet Downstream of the Initial Queue Position xv

15 Table 6-20, Breakpoints for Queued Heavy Vehicles from the Initial Queue Position Upstream 200 Feet 168 Table 6-21, Breakpoints for Queued Heavy Vehicles from the Initial Queue Position from Upstream 200 to 600 Feet 168 Table 6-22, Breakpoints for THRU Heavy Vehicles for All Distances Upstream and Downstream of the Data Collection Site 169 Table 6-23, Field Data Acceleration Observations by Speed Range 171 Table 6-24, Comparison of Field Data and Traffic Engineering Handbook Maximum Acceleration by Speed Range 183 Table 6-25, Percent of Activity by Speed-Acceleration Ranges Outside the FTP 187 Table 7-1, Limits of Prediction for Independent Variables xvi

16 ACRONYMS CAAA: Clean Air Act Amendments CARB: California Air Resource Board CART: Classification and Regression Tree Analysis CBD: Central Business District CO: Carbon Monoxide CO 2 : Carbon Dioxide DMI: Distance Measuring Devices FHWA: Federal Highway Administration FTP: Federal Test Procedure GIS: Geographic Information System HC: Hydrocarbons HCS: Highway Capacity Software HPMS: Highway Performance Monitoring System HTBR: Hierchiacal Based Regression Tree ITS: Intelligent Transportation Systems JASPROD: Joint Acceleration-Speed Probability Density Function K/S: Kolmogorv-Smirnov xvii

17 LDV: Light Duty Vehicle LOS: Level of Service LRF: Laser Rangefinders MEASURE: Mobile Emission Assessment System for Urban and Regional Evaluation NAAQS: National Ambient Air Quality Standards NCHRP: National Highway Cooperative Program NO: Nitrogen Oxide NOx: Oxides of Nitrogen NO 2 : Nitrogen Dioxide O 3 : Ozone PPM: Parts Per Million Pb: Lead RMD: Residual Mean Deviance ROG: Reactive Organic Gas SO 2 : Sulfur Dioxide TCM: Transportation Control Measure TSP: Total Suspended Particulate VOC: Volatile Organic Compounds VMT: Vehicle Miles Traveled xviii

18 V/C: Volume to Capacity USEPA: United States Environmental Protection Agency UTPS: Urban Transportation Planning Software xix

19 SUMMARY Current research suggests that vehicle emission rates are highly correlated with modal vehicle activity and that specific instances of load induced enrichment may contribute a disproportionate share of motor vehicle emissions. Consequently, a modal approach to transportation-related air quality modeling is becoming widely accepted as more accurate in making realistic estimates of mobile source contribution to local and regional air quality. New vehicle modal emission rate models will assess emissions as a function of specific operating mode or engine load surrogates. These new models require that vehicle activity be input by fraction of time spent in different operating modes. However, the ability to realistically model microscopic on-road modal vehicle activity currently limits the implementation of these models. To provide better estimates of microscopic vehicle activity, field studies using laser rangefinding devices were undertaken to quantify actual vehicle behavior along signalized arterials and at signal-controlled intersections in Atlanta, Georgia. Data were analyzed to determine the fractions of vehicle activity spent in different operating modes, especially those that may lead to high engine load and elevated emissions. Statistical analysis of the data yields a model for prediction of microscopic vehicle xx

20 activity based on geometric and operational characteristics of the roadway. Research results will provide the ability to estimate microscopic vehicle activity as input to both local and regional transportation-related air quality models. Findings may also enhance current methods for estimating capacity and modeling traffic flow and may have applications for intelligent transportation systems (ITS). xxi

21 CHAPTER I 1. INTRODUCTION For at least ten years, the technical, scientific, and administrative community has expressed concerns about the current certification cycle for automotive emissions being representative of actual driving behavior (Cicero-Fernandez and Long, 1994). A major shortcoming of current emission modeling is the aggregated representation of on-road vehicle activity, which inaccurately characterizes on-road driving behavior. The current modeling philosophy is built on the assumption that drivers behave similarly, rather than being based on individual or actual driver behavior. Average behavior assumes that all drivers engage in driving patterns similar to those over which vehicle emissions have been tested, such as the Federal Test Procedure (FTP) Certification Cycle. Likewise, corresponding emission factors were developed from procedures based on the assumption that vehicles pollute similarly under an average range of speeds and vehicle miles traveled (VMT). This traditional approach neglects variations in driving behavior, especially extremes such as hard accelerations or stopand-go driving under congested conditions. 1

22 A large body of evidence suggests that under most on-road operating conditions, actual vehicle emissions can differ dramatically from those predicted by current mobile source emission models (Pierson et al., 1990; LeBlanc, 1994; Barth et al., 1997). Current research indicates that vehicle emissions rates are highly correlated with engine operating mode. In particular, vehicle operation leading to engine loading and elevated emissions are hard accelerations, air conditioner use, vehicle operation on a grade, and hard decelerations. Because recent research has indicated that various shortcomings exist in the data input, modeling, and output of traditional mobile source air quality models, current research activities are focusing on a modal approach to mobile source emission modeling. Modal or activity-specific models attempt to estimate emissions as a function of specific operating mode or engine load surrogates. To implement modal models, statistical distributions of vehicle activity corresponding to the amount of time that vehicles spend in different ranges of speeds and corresponding accelerations must be developed. Once vehicle activity is disaggregated into speed and acceleration distributions, activity-specific emission rates may be applied to estimate emissions. Modal emissions modeling is becoming widely accepted as a more theoretically accurate approach that will provide more realistic estimates of mobile source emissions contributions to local and regional air quality analysis. 2

23 Although a modal approach to emissions modeling offers promising benefits in terms of accuracy, a weak link is the ability to realistically model on-road modal vehicle activity. Currently, little data exists relative to how vehicles operate in a real world setting. Various activity estimation methods are in-use or proposed, such as simulation models. None of these methods have been validated as to whether the output realistically models the wide range of vehicle activity encountered on the roadway. Additionally, the ability does not exist to relate activity to external variables such as roadway grade or traffic volumes. This research was conducted as part of a study underway at Georgia Institute of Technology. Research was conducted under a cooperative grant from the U.S. Environmental Protection Agency (USEPA) and the Federal Highway Administration (FHWA). The principal goal of this research was to develop a model that can predict modal vehicle activity at signalized intersections and along signalized segments. Individual vehicle traces were collected on-road at signalized intersections with laser rangefinders (LRF) in the Atlanta, Georgia metropolitan area. With collection and analysis of field data, statistical distributions of vehicle activity were generated and tested using regression tree analysis to relate speed-acceleration profiles of vehicles to roadway characteristics such as grade, location along the study link, queue position, or 3

24 volume of roadway to physical capacity. Data were analyzed with Hierachical Based Regression Tree (HBTR) analysis and relevant predictor variables identified. The final model predicts microscopic vehicle activity based on those operational and geometric characteristics of the roadway, which were shown to influence vehicle activity such as grade, location along link, queue position, or volume of roadway to physical capacity. Model development ensured that final distributions of vehicle activity can be linked with the modal emission rates from Georgia Tech s MEASURE model to provide input to both regional and microscale air quality models. Chapter 2 of this work provides a background on air quality in general and discusses some of the inadequacies of traditional transportation-related air quality models. Chapter 3 overviews research that evidences a relationship between rate of emission output and engine operating mode and provides explanatory information on current research efforts for modal modeling. In Chapter 4, the various statistical models considered for data analysis are discussed and the final statistical model presented. The response variables based on emission factor models from MEASURE are explained and a list of all independent variables hypothesized to influence vehicle activity is presented. Chapter 5 covers data collection and handling. The methodology for calculating variables, such as level of service, is also covered. In Chapter 6, research results are presented along with the 4

25 final microscopic activity prediction models. A comparison of field data with various traffic engineering relationships and with simulation modeling is also provided. Finally, Chapter 7 presents a discussion and conclusion on research results. The significance of this research work is development of a model capable of predicting microscopic vehicle activity at signalized intersection based on roadway or operational characteristics that influence behavior. Because the results have described microscopic vehicle activity, research findings may also enhance current methods for estimating capacity and modeling traffic flow and may have applications for intelligent transportation systems (ITS). 5

26 CHAPTER II 2. BACKGROUND Degraded air quality continues to be a major concern in most major cities in the United States. Unhealthy levels of air pollution continue; posing health concerns, choking economic development, and threatening federal transportation dollars, despite advances in emissions control for both mobile and stationary sources. A significant share of blame for urban air problems can be directly attributed to increasing development and urban sprawl, which has resulted in a rapid increase VMT with the resulting emissions. The Clean Air Act Amendments of 1990 (CAAA) were issued as a legislative mandate to improve air quality in designated metropolitan areas. To regulate air pollution, National Ambient Air Quality Standards (NAAQS) were established, setting acceptable levels for specific airborne pollutants, including particulate matter, carbon monoxide (CO), oxides of nitrogen (NOx), sulfur dioxide (S0 2 ), ozone (O 3 ), and lead (Pb). NAAQs set the maximum air pollution concentrations allowable in any one area based on the minimum dose of the pollutant required to cause adverse health effects in the most sensitive members of the population (Kaliski, 1991). 6

27 Air pollution comes from a variety of sources, which can be divided into three main categories: Stationary sources: factories, power plants, smelters, etc.; Mobile sources: automobiles, trucks, buses, trains, and planes; and Natural sources: pollution from wildfires, windblown, dust, volcanic eruptions, etc. (USEPA, 1995a). This chapter provides background information on transportation-related air pollutants including the National Ambient Air Quality Standards. An introduction on traditional transportation-related air quality modeling along with a discussion on the drawbacks of the modeling process in terms of both emission factors and vehicle activity are presented. 2.1 Automobile Exhaust Emissions The transportation sector is directly responsible for a significant proportion of harmful ambient emissions (Anderson et al., 1996). Estimates for the amount of pollutants produced by motor vehicles vary from 33 to 50% of NOx, 33 to 97% of CO, 40 to 50% of HC, 50% of ozone precursors, and at least one-fourth of volatile organic compounds (VOC ) (Mullen et al., 1997; SCAQMD, 1996; EPA, 1995a; USDOT, 1993; CARB 1994; Chatterjee et al., 1997). Although not included in the NAAQs, particulate matter with aerodynamic size less than or equal to 10 microns (PM-10) is also released by motor vehicles from diesel 7

28 engines and tire wear. Hydrocarbons (HC) are also not included in the NAAQs but are included in mobile source emission modeling since they contribute to ozone formation. CO, NO x, and HC are a by-product of combustion and are found directly in automobile exhaust. Fuel evaporation also contributes emissions of VOC Ozone The NAAQS for ozone are 0.23 parts per million (ppm) for a one-hour period. This standard may not be exceeded more than 3 times over a continuous 3-year period (Chatterjee et al., 1997). This pollutant is a highly reactive form of oxygen. It is a colorless gas, characterized by a sharp odor. Ozone occurs naturally in the stratosphere but normally only in low doses (0.03 to 0.05 ppm) near the surface of the earth. Ozone is not emitted directly from mobile sources; rather it is produced by a complicated series of chemical and photochemical reactions between reactive organic compounds, oxides of nitrogen, and naturally occurring oxygen. Photochemical reactions require solar radiation to act as a catalyst; consequently peak concentrations of ozone are found around the middle of the day and climax during the summer months. The associated health effects of ozone include decreased breathing capacity, increased airway resistance, impaired host defenses, acute inflammation of the lung tissue, and respiratory cell damage. A correlation is hypothesized between an increasing number of 8

29 hospital admissions for all respiratory causes, including asthma, and an increase in ambient ozone, sulfates, or sulfur dioxide levels (SCAQMD, 1996; Mullholland, 1998) Carbon Monoxide The pollutant, carbon monoxide is a colorless, odorless, relatively inert gas introduced by both human and natural sources, such as forest fires. In urban areas, the primary source of CO is incomplete combustion of carbon-containing fuels, mostly gasoline. During optimum combustion, each carbon atom has affinity to bond with two oxygen atoms forming carbon dioxide (CO 2 ). When an oxygen deficiency is present in the engine, some carbon atoms are only able to bond with a single oxygen atom and the result is CO. Consequently, an overly rich air-fuel ratio is the primary cause of CO formation (King, 1995). Colder temperatures are more conducive to the formation of CO, consequently CO exceedances are more common in the winter (CARB 1995). Ambient concentration of CO are spatially and temporally correlated to the rate at which CO is emitted and prevailing meteorological conditions, with peak concentrations occurring in the fall and winter months. Because automobile exhaust is the major source of CO, high concentrations can result in urban areas with heavy traffic congestion (EPA, 1995a). NAAQS for CO are 35 ppm for a one-hour averaging period and 9 ppm for an 8- hour period, which is not to be exceeded more than once per year. 9

30 Carbon monoxide enters the bloodstream and displaces oxygen, binding with hemoglobin in the blood. This reduces the blood s ability to carry oxygen to the body s organs and tissues. Therefore, most of the toxic effects of CO are caused by reduced oxygen supply. Those at the highest risk from carbon monoxide are heart patients, smokers, and people who engage in heavy exercise (SCAQMD, 1996). Other health effects due to exposure to elevated CO levels include visual impairment, reduced work capacity, reduced manual dexterity, poor learning ability, and difficulty in performing complex tasks (EPA, 1995a) Oxides of Nitrogen The nitrogen content of both gasoline and diesel fuels is negligible. Oxides of nitrogen, a colorless gas, is actually formed from the destruction of atmospheric nitrogen (N 2 ), which makes up 80% of air, during the combustion process. Although ambient nitrogen and oxygen do not normally react, in the presence of sufficiently high temperatures, a chemical reaction occurs catalyzing oxygen and nitrogen to form nitrogen oxide (NO) (King, 1995). Formation of oxides of nitrogen is exacerbated by high temperature and high concentrations of oxygen. Once formed, oxides of nitrogen quickly react with oxygen and form nitrogen dioxide (NO 2 ), a reddish-brown gas with a bleach-like odor, which is primarily responsible for the brownish tinge characteristic of polluted air. Nitrogen dioxide also plays a major role in atmospheric reactions, which produce ground-level ozone (EPA, 1995). Critical engine variables that determine the amount of oxides of nitrogen produced 10

31 are the fuel/air equivalence ratio (Ø), the burned gas fraction of the in-cylinder unburned mixture, and spark timing. Once oxides of nitrogen are released into the atmosphere, a reaction occurs with reactive organic gas (ROG) to form ozone. This reaction is catalyzed by sunlight and therefore occurs more often in the summer corresponding to higher temperatures and greater amounts of sunlight (CARB, 1995). Besides ozone formation, nitrogen oxides in the air are a potentially significant contributor to a number of environmental effects such as acid rain and eutrophication in coastal waters. Eutrophication is an increase in nutrients that reduce the amount of oxygen in a body of water creating an environment that is destructive to fish and other animal life (EPA, 1995a). Children and adults with respiratory illnesses are most susceptible to oxides of nitrogen, which is a respiratory irritant and reduces resistance to respiratory infection such as influenza (SCAQMD, 1996) PM 10 PM 10 is a category of total suspended particulate (TSP), which is made up of a complex mixture of solid material suspended in the atmosphere. Finer fractions of TSP have greater effects on health and visibility than coarse fractions. PM 10 is particulate matter with diameter less than approximately 10 micrometers. The mobile source contribution to 11

32 particulate matter is a product of combustion, machinery, and tire wear (Chatterjee et al., 1997). The main health effects associated with PM 10 include increased mortality, exacerbation of preexisting respiratory and cardiovascular disease, changes in lung function and structure, altered defense mechanisms, and increased risks of developing cancer (SCAQMD, 1996). Children, the elderly, and persons with chronic lung disease, influenza, or asthma tend to be especially sensitive to the effects of particulate matter. In an acid form, PM-10 is destructive to manmade materials and is a major cause of reduced visibility in many parts of the United States (EPA, 1995a) Hydrocarbons Hydrocarbons are one of the three commonly modeled transportation-related pollutants (CO, NO x, and HC). Hydrocarbon emissions result from incomplete combustion of hydrocarbon-based fuel, such as gasoline. Fuel composition can significantly affect the types and amounts of hydrocarbons released. Hydrocarbons are released as part of the combustion process and from piston blow-by gases. Unburned hydrocarbons are released during fuel evaporation and through vents in the fuel tank and carburetor after engine shutdown (Heywood, 1988). Gasoline itself is a hydrocarbon compound and when burned properly, the hydrogen and carbon atoms split apart and then bond with oxygen to form water, 12

33 (H 2 0), or carbon dioxide, CO 2 (King, 1995). 2.2 Drawbacks to Traditional Emissions Modeling In order to meet CAAA goals and demonstrate progress towards conformity, the traditional mobile source emission modeling approach was developed. A flowchart of the general process is presented in Figure 2-1. In its most basic form, this approach simply multiplies an estimate of vehicle activity by an emission factor to determine the total quantity of pollution released by a roadway, group of roadways, or region. The most common emission factor models are the MOBILE family of models, widely used throughout the United States, and the EMFAC models used in California. Although emission estimates play an important role in determining a region s progress towards meeting air quality goals and influence transportation investment decisions, numerous inadequacies exist in both the data input and modeling methodology of traditional mobile source emission modeling. These inadequacies extend to both the vehicle activity side of the equation as well as to emission factor estimates. One of the main flaws is that the data used to support modeling often come from sources not intended to support air quality analysis and consequently may produce inaccuracy in the air quality analysis and the conclusions they present. 13

34 14

35 14 Figure 2-1: Traditional Emission Modeling (source: Roberts, 1999) 15

36 Another major flaw is the aggregate representation of on-road vehicle activity to estimate emissions, resulting in inaccurate characterization of actual driving behavior. The current modeling philosophy assumes that all drivers engage in driving patterns similar to those over which vehicle emissions have been tested. Likewise, corresponding emission factors were developed from aggregate representations of vehicles based on the assumption that vehicles pollute similarly under an average range of speeds and vehicle miles traveled (Guensler and Sperling, 1994) Vehicle Activity One of the main shortcomings of the traditional modeling approach is that it is unable to capture actual on-road vehicle behavior. Instead, activity estimates are frequently based on output from regional transportation modeling systems, which were developed to forecast the need for new highway facilities, rather than air quality modeling. Consequently, these models are not sensitive to the inputs and parameters required for air quality modeling, such as accurate estimates of vehicle speeds. Inaccuracies in vehicle estimates are related to data input, mathematical algorithms, and the calibration procedures of regional transportation models. Data input to regional models are often based on databases that are incomplete or outdated. Information about the number of trips, geographic distribution of those trips, and timing of tripmaking is often estimated from household surveys done years previously (USDOT, 1993). 16

37 Speed Estimates One of the major shortcomings of vehicle activity estimates is the use of average speed estimates derived from regional models. Historically, the MOBILE or EMFAC series of motor vehicle emission rate models estimated emissions as a function of average speed, consequently the modeled relationship between emissions and vehicle activity is highly speed dependent. Emission rates vary greatly across different speed ranges as shown in Figure 2-2 for CO in grams per mile. For CO, MOBILE5A emission rates are highest in the lower speed ranges and then reach their lowest rates in the middle speed ranges from 30 to 45 mph. Emission rates increase again after 55 mph. Locations on the emission curve where the slope is the steepest, indicate areas where emission rates are the most sensitive to changes in speeds. Inspection of Figure 2-2 indicates that an increase in average speed, from approximately 3 to 20 mph, reduces the emission rate from 130 to 20 g/mile. Logically, areas on the chart where emission rates are the most sensitive to changes in average speed are also locations where errors in estimating average speed would have the greatest impact to over or underestimate emissions. Around 60 mph, an error of only 1.2 mph will cause a 10% error in the CO emissions factor. At 20 mph, an error of 2.3 mph in average speed would be required to create that same 10% error in CO emissions (Chatterjee, et al., 1997). Although emissions are speed dependent in traditional modeling, the main data sources for both traffic volume and speed data are output from the traffic assignment 17

38 17 Figure 2-2: MOBILE Emissions Versus Speed Range for Carbon Monoxide 18

39 stage of travel demand modeling, whose main purpose is to forecast roadway volumes, not accurately replicate link speeds. Link speeds are used to calibrate the model for realistic volume output (Chatterjee et al 1997). Speeds input to four-step modeling are often the posted speed limit or default values, such as 45 mph for arterials, 35 mph for collectors, etc., instead of observed freeflow speeds. In some cases the use of speed limits or defaults may lead to underestimation of actual speeds since motorists frequently exceed speed limits. The use of average speeds also fails to accurately describe the wide ranges of vehicle activity actually found in normal driving. A group of vehicles at high speed coupled with a group of low speed vehicles, both of whom are operating in the higher emission factor ranges, could average out somewhere in the mid-speed ranges where emission factors are lowest for HC and CO (DeCarlo-Souza et al., 1995). Comparisons of modeled versus actual speeds have demonstrated that speeds modeled by travel demand forecasting models may exceed onroad speeds by 35%. Discrepancies in modeled and actual speeds may result from the fact that the capacity-restraint formula often used in travel demand models does not degrade for speed appropriately with considerable congestion. Speeds are used to calibrate travel demand models so that when modeled volumes replicate actual volumes, such as a screen line counts, match within a reasonable range of accuracy (commonly 10%), the model is considered to be calibrated. Rarely are model speeds compared against on-road speeds. Volume and speed estimates from regional models are also likely to become increasing 18

40 unreliable under congested conditions, which are a common occurrence in urban areas (USDOT, 1993) Volume Estimates A regional estimate for VMT, usually from the regional travel demand forecasting models, is often multiplied by emission factors in grams per mile to calculate total emissions produced. However, VMT output from regional models has several inherent inaccuracies. First, the road network used in travel demand models is not detailed enough for air quality modeling. Travel demand models use stick representation of the surface street system, which typically include only major roads such as arterials, freeways, and collectors. Consequently, VMT is available only for representative roadways in the network. Volume data is not available from travel demand modeling for all links in the network. Second, local street systems are not adequately accounted for in regional modeling (Chatterjee et al 1997). Local streets, themselves are not highly significant in the four-step modeling process since their purpose is only to provide access to the major street network. Consequently they are usually represented as centroid connectors. The lack of available data is particularly a problem as no accepted technique exists for VMT calculations for local roads. Local roads are typically low volume facilities, however they may make up a significant proportion of total miles of roadway in urban areas (Chatterjee et al., 1997). Consequently, lack of representation of both VMT and speeds on local roads presents a major deficiency in urban activity modeling. 19

41 2.2.2 Emission Rates A major shortcoming in traditional emission factor modeling is that a complete range of vehicle activity is not represented in the Federal Test Procedure on which MOBILE is based. The main algorithms in the MOBILE model were developed for the following default assumptions: average vehicle speed of 19.6 mph; ambient temperature of 75 F; and start mode fractions of 20.6% cold and 27.3% hot starts. Emission factor are based on the default assumptions and then adjusted by dimensionless correction factors to represent region-specific conditions, such as average speed, ambient temperature, percent cold starts, gasoline volatility, implementation of inspection/maintenance programs, and use of oxygenated fuels (Keenan and Escarpeta, 1995). In the development of the FTP, acceleration rates were artificially reduced to accommodate testing equipment capabilities. Additionally, the original objective of the FTP was to capture average driving not variations in speed, consequently more aggressive driving behavior such as high speed and high accelerations are not captured (USEPA, 1995b). The Federal Test Procedure, on which MOBILE is based, was established over two decades ago and was intended to replicate the operation of a typical in-use urban vehicle. 20

42 The FTP uses the average driving conditions, which are embodied in a pre-determined driving cycle, to determine emission factors (Barth et al., 1996). The FTP was developed to represent a typical driving pattern in primarily urban areas and was created to simulate a trip route in Los Angeles representative of a typical home based work trip. The original route was selected to match the engine operating mode distribution obtained in central Los Angeles using a variety of drivers and routes. The driving cycle is a particular pattern of idle, acceleration, cruise, and deceleration over which a vehicle is tested. Then emission factors specific to each cycle are produced (CARB, 1995). Emission factors were also developed using a small sample of vehicles, which may not be representative of the actual on-road fleet. This is particularly significant since emissions rates may vary even between vehicles of the same type based on miles accumulated on the vehicle, driving behavior, inspection and maintenance history, etc. (USDOT, 1993). Another drawback to current practice is that vehicle activity and speed estimates are link-based while emission factor models such as MOBILE are trip based. MOBILE estimates emissions over an entire trip, about 20 minutes, rather than for a particular link. Travel demand forecasting models are based on a street network represented by nodes (intersections) and links and as a result link-specific speed and traffic volumes are 21

43 generated. This trip-based emission factor modeling is therefore inconsistent with linkbased modeling (DeCorla-Souza et al., 1995). Inaccuracies in vehicle emissions models can also occur from errors in basic emission rates, as well as in correction factors, such as speed correction factors, which are used to adjust basic emissions rates. Basic emission rates are measured from a simulated pattern intended to be representative of "typical" city driving. This approach does not accurately reflect differing road facilities, vehicle types, and operational activity common to urban driving. Additionally, the overall average speed of 19.6 mph does not reflect actual speeds on urban collectors, arterials or freeways (USDOT, 1993). 22

44 CHAPTER III 3. TOWARDS A MODAL APPROACH FOR TRANSPORTATION- RELATED AIR QUALITY MODELING As explained in Chapter 2, various shortcomings exist in the traditional mobile source emission-modeling regime. The most significant drawback for both activity and emission factor modeling is the inability to model actual on-road vehicle behavior, especially activity outside the range of the FTP, and to correlate emission production specifically to operating mode. This chapter first outlines current research indicating that operating mode is related to emission output rates. Next, an overview is provided of research efforts focused on mode specific activity and emission factor estimates. Finally, an overview of current representations of speed/acceleration relationships common to traffic engineering are presented, to provide the reader with necessary background information. These relationships are often the basis for simulation models and other methods in use to create modal activity estimates. Later, in the data analysis chapter, Chapter 6, field data are compared with these traffic-engineering relationships. An overview of vehicle dynamics as they relate to speed and acceleration is also provided in Appendix A for more background information. 23

45 24

46 3.1 Evidence of a Mode Specific Emission Relationship An overview of contemporary research which has demonstrated inadequacies in the current average speed based approach and has indicated that emissions are related to engine operating mode are presented in the following sections Tunnel Studies Initial evidence that traditional modeling may not adequately represent actual on-road emissions was evidenced by studies in several traffic tunnels. An initial study was conducted in the Van Nuys Tunnel in California in 1987, which measured vehicle emissions with a mass flow study. The pollutant levels collected from the tunnel were three times higher for CO and four times higher for HC than predicted by EMFAC7C (Pierson et al., 1990). Additional studies were conducted in other tunnels. Emissions from motor vehicles for CO, NO, NOx, gas-phase speciated nonmethane hydrocarbons, and carbon compounds were measured in 1992 in the Fort McHenry Tunnel under Baltimore Harbor and the Tuscararoa Mountain Tunnel of the Pennsylvania Turnpike. The tunnels were characterized by high speeds with little acceleration. The vehicle fleet for both tunnels was relatively new with the median vehicle age less than 4 years old. Consequently, cleaner vehicles under steady speed conditions dominated the study. Results indicated that MOBILE4.1 and MOBILE5 only gave predictions within +-50% of observation with the MOBILE models tending to overpredict emissions (Pierson et al 1996). 25

47 3.1.2 Activity Outside the FTP Various studies show that a significant amount of on-road driving activity occurs outside the range of activity represented in the Federal Test Procedure (velocity >= 57 mph and acceleration >= 3.3 mph/s). The USEPA supported instrumentation of instrumented approximately 350 vehicles in Spokane, Washington; Baltimore, Maryland; and Atlanta, Georgia and recorded vehicle speed, engine speed, acceleration, and manifold absolute pressure (LeBlanc et al., 1995). Statistically significant differences were noted in vehicle speeds and acceleration characteristics across these cities. The three-city instrumented vehicle study also found accelerations ranging from a minimum of mph/s to a maximum of mph/s. Although these values, contrast sharply with the maximum acceleration in the FTP of 3.3 mph/s (USEPA, 1995b) they do appear to be extremely high. Trip lengths were also recorded for the three-city study and an average trip length of 4.9 miles in length discovered as compared to the 7.5 mile average trip used for the FTP, this suggests that that actual trip lengths may be much shorter than those modeled in the FTP (Enns et al., 1994). Another study using 1,100 miles of driving data from the Los Angeles area were used to develop seven cycles of vehicle activity. The researchers found differences between 26

48 freeway and arterial driving. Cycles representing freeway activity were much smoother in terms of speed and ranges of accelerations and decelerations as compared to arterial flow, which was much rougher. The authors indicated that 18.7% of driving time was spent in accelerations greater than 3 mph/s for arterials versus 2.3% for freeways and 32.9% of arterial and 7.6% of freeway activity was spent in decelerations less than -3mph/s (Effa and Larsen, 1994). This indicates that differences in modal activity may occur across roadway types Enrichment One of the extremes in vehicle activity that has been demonstrated to contribute a disproportionate share of emissions is commanded enrichment. Commanded enrichment is an engine-operating mode where the engine management feedback control system (which ensures stoichiometric operation) is overriden to increase the fuel:air ratio (LeBlanc et al., 1994; Ganesan, 1994). Commanded enrichment provides increased engine power output to enhance performance and also reduces peak exhaust gas temperatures protecting engine components and exhaust after-treatment systems from the high exhaust gas temperatures that would result under high load conditions. Commanded enrichment is typically called for whenever the engine operates under high load conditions, such as undergoing a hard acceleration, action against a grade, or pulling a load. The air-fuel ratios for commanded enrichment can be as rich as 11.7:1 27

49 compared to the normal stoichiometric air-fuel level of approximately 14.7:1 (Heywood, 1988). Engine-out CO emissions increase as the air-fuel ratio is enriched from stoichiometric levels. Emissions increase due to the lack of oxygen available to complete the combustion process, which normally results in conversion of hydrocarbons to CO 2 and water. Hydrocarbon emissions also increase under fuel-rich conditions, since less fuel is burned. The catalyst conversion efficiency levels for both HC and CO emissions are very sensitive to air-fuel ratios. In a fuel-rich combustion environment, the lack of oxygen causes the normal oxidation process that converts HC and CO into CO 2 and water vapor to drop off very quickly causing a reduction in catalyst conversion efficiency (Heywood, 1988). St. Denis and Winer (1994) used an instrumented Ford Taurus (1991 model year) to collect on-road driving data in California. The researcher found relevant differences between actual collected emissions and the amount of emissions calculated using the FTP procedure. The differences for CO were attributed to enrichment, which was estimated to be responsible for two-thirds of the difference between modeled and FTP calculated emissions. Study results also found that pollutants released during enrichment were two to three times higher than for stoichiometric operation indicating that engine operating mode is an important variable in determining emission output. 28

50 Other tests on individual vehicles reported that moderate to heavy engine loads lead to enrichment conditions that can increase gram/second emission rates for carbon monoxide by 2500 times and hydrocarbon emissions by 40 times compared to normal stoichiometric operation (LeBlanc, 1994; Barth et al, 1996) Acceleration Commanded enrichment is caused by engine loading. One factor leading to engine loading is "hard" accelerations. Research has indicated that a single "hard" acceleration event (enrichment event) may cause as much pollution as the remainder of the trip (Guensler, 1993). Emissions tests conducted at the California Air Resources Board, primarily on carbureted vehicles, showed a large increase of HC and CO during hard acceleration events. Later studies indicate that a single hard acceleration (> 6mph/s) could increase the total trip emission for carbon monoxide (CO) by a factor of two. Cicero-Fernandez and Long (1994) evaluated ten current technology vehicles over four testing cycles. One of cycles was a specially designed acceleration cycle, which included various acceleration events. Comparison of the acceleration specific cycle emissions to those from a comparable FTP cycle indicated that HC emissions were 3 times higher for the acceleration cycle than the FTP cycle. For CO, it was 19 times higher than the FTP cycle and for NO x the acceleration cycle was similar to the FTP cycle indicating a stable release of NOx even under enrichment. They also indicated that accelerations at low 29

51 to medium speeds had less pronounced emission increases than accelerations at higher speeds. Le Blanc et al. (1995) also conducted research that estimated emissions as a function of speed/acceleration ranges. The researchers found a correlation between increased CO output in g/s and the following: high speeds with speeds greater than 57 mph and acceleration rates less 1 mph/s; high accelerations greater than 3.3 mph/s with speeds less than 57 mph; and high speeds/high accelerations with speeds greater than 57 mph and accelerations greater than 1 mph/s. Research by CARB also found that hard accelerations triggered increased emissions, especially for CO. An increase in HC was also found to be significant under hard accelerations. Accelerations at mid to high speeds emitted more emissions than low to midrange speeds. For NO x, speed was found to be a more influential variable in emission rates than accelerations (CARB, 1997). Finally, Yu (1998) also found a correlation between speed, acceleration, and emission rates using remote sensing studies in Houston at five locations. From the data collected, a model was developed which correlated on-road vehicle exhaust emission rates 30

52 with the vehicle's instantaneous speed profile. Study results were compared with existing emission models. Results indicated that both the MOBILE and EMFAC emission models underestimated emissions for all vehicle types as compared to on-road estimates Grade Roadway gradient has been investigated as a geometric effect that may increase emissions. Acceleration against a grade results in additional load on the engine beyond that which is associated with normal driving. The higher mass flow associated with increased engine load is expected to produce higher emissions for all three pollutants (CO, NO x, and HC). It may also increase the frequency or extend the duration of enrichment which impacts CO emissions (USEPA, 1995b). Cicero-Fernandez et al. (1997) studied the effect of road grade and found that for each 1% increase in grade, the HC emission rate increased by 0.04 g/mile and the CO emission rate increased by 3.0 g/m. The study consisted of controlled runs with speeds between 35 and 55 mph and a maximum acceleration of 3.3 mph/s. Runs were conducted on both flat terrain and road segments with grades from 0 to 7% in Los Angeles, California. Both freeways and arterials were included in the study. Enns et al. (1994) tested 9 vehicles for grade influences and found increases in CO of 3.2 grams/mile with much smaller increases for HC and NO x. 31

53 The Fort McHenry tunnel study, discussed in section collaborates the correlation between grade and elevated emissions. The tunnel had upgrade and downgrade. The vehicle fleet was composed of relatively newer vehicles and vehicle activity in the tunnel was composed of smooth flow. Comparison of emissions from the upgrade and downgrade yielded an increase in emissions per mile by a factor of two. The authors concluded that the effect of grade was significant and that is should be included in transportation-related air quality modeling (Pierson et al., 1996) Air Conditioner Use Evidence has shown that air conditioner use in a vehicle results in elevated emissions. The effects of air conditioning use detailed in the previously described study by Cicero-Fernandez et al. (1997), who conducted controlled test vehicle runs with speeds between 35 and 55 mph and low acceleration on road segments with grades from 0 to 7%. With the air conditioner running at full setting, emissions increased by 0.07 g/mile for HC and 31.9 g/mile for CO. Enns et al. (1994) also described the impact of air conditioning use as increasing NO x emissions by 0.21 grams/mile. Another study found that air conditioning use (on max) in combination with roadway grades of up to 6.7% increased HC emissions up to 57% and CO up to 268% (CARB, 1997) Rapid Load Reduction Most research identified loading events, such as work against a grade or hard accelerations, as the culprits for elevated emissions in otherwise normally emitting 32

54 vehicles. Elevated HC emissions have also been associated with rapid load reduction and long deceleration events. During stoichiometric driving, the quantity of condensed fuel on the intake manifold walls is in rough equilibrium, dependent to some degree on the recent history of fuel injection (power level). With negative engine power, which often occurs in coastdown and braking, air flow continues with little or no fuel injection causing an extremely low ratio of fuel to air, which inhibits combustion. This allows the condensed fuel to be removed by evaporation over a period of several seconds resulting in elevated unburned hydrocarbons levels. A study by An et al. (1998a) tested 200 vehicles and found that this phenomenon contributed 10 to 20% of the overall HC emissions under various test cycles. 3.2 Towards A Modal Approach For at least ten years, the technical, scientific, and regulatory community has expressed concerns about the current certification cycle for automotive emissions being representative of actual driving behavior (Cicero-Fernandez and Long, 1994). Because recent research has indicated that various shortcomings exist in the data input, modeling, and output of traditional mobile source air quality models, current research activities are focusing on a modal approach to mobile source emission modeling. To address shortcomings in current transportation-related air quality models and provide agencies with enhanced tools for vehicle emission estimates, various modal modeling approaches have been suggested. Modal models attempt to estimate pollutants as a function of specific operating mode or engine load surrogates. To implement modal models, 33

55 statistical distributions of vehicle activity corresponding to the amount of time that vehicles spend in different ranges of speeds and corresponding accelerations must be developed. Once vehicle activity is disaggregated into speed and acceleration distributions, activityspecific emission rates may be applied. Modal emission modeling is becoming widely accepted as a more theoretically accurate approach that will provide more realistic estimates of mobile source contributions to local and regional air quality (Guensler, 1993; Barth et al, 1996; Washington, 1996). Figure 3-1 shows a schematic of the modal elements of a hypothetical trip, including idling, acceleration, cruise, and deceleration. This figure presents a schematic of a typical trip when broken down into its modal elements. Currently various research groups, both nationally and internationally, are working on various facets of modal modeling. Research efforts include development of activity specific emission factors, improved prediction of fleet mix, improved estimation of cold and hot start fractions, development of methods to better predict on-road vehicle modal activity, etc. Following is a summation of the major research efforts underway for emission factor and vehicle activity modeling. 34

56 Figure 3-1: Modal Elements of a Vehicle Trip Improved Emission Factor Estimates One of the major flaws in traditional modeling is the use of emission factors, which average emissions over a cycle, such as the FTP, and neglect extremes in vehicle activity. To improve emission factor estimates, various research efforts have focused on developing methods to relate emission output to specific vehicle activity such as a vehicle's instantaneous speed and corresponding acceleration. Post et al. (1985) tested 177 Australian light duty vehicles on a dynamometer and created averaged vehicle maps of emissions. Data were mapped into matrix format showing emission rates for specific speed/acceleration bins. The model was developed to make fuel 35

57 consumption predictions for any type of vehicle, engine capacity, and driving pattern. Emission rates for carbon monoxide, carbon dioxide, oxides of nitrogen, and hydrocarbons were output during the dynamometer testing and emission rates correlated to instantaneous velocity and acceleration. The three-parameters (speed, acceleration, and emissions) were used to develop a two-parameter (power and emissions) emission model. The power term is the product of speed and acceleration. Elevated emissions were noted for higher accelerations independent of the corresponding speed for both CO and HC. Increasing speeds were also correlated to elevated emissions for the two pollutants. Sierra Research has created driving cycles, based on chase car and instrumented vehicle data from Baltimore, Spokane and Los Angeles, which was collected during the FTP Revision Project. Cycles were constructed to match observed speed-acceleration and specific power frequency distribution of chase car driving data. Facility-specific cycles were constructed using randomly selected microtrips to match the speed-acceleration frequency distribution of all vehicle operation occurring under conditions of interest such as a particular facility type or level of service (LOS). Vehicle activity was collected using a specially designed instrumented vehicle with a grill mounted laser rangefinder and distance measuring instrument. This type of data collection allowed calculation of the instantaneous speed and acceleration of the "followed" vehicle. Other variables were collected such as LOS. Vehicle activity was categorized into six driving cycles defined by LOS of the freeway segments. 36

58 The level of service calculated, however, was a rough estimate based on data collector s perception of activity around the chase vehicle rather than a formal calculation (USEPA, 1997). An et al. (1997) outlined on-going development of a comprehensive modal emissions model capable of predicting emissions for a wide variety of light-duty cars and trucks based on engine operating mode. At the highest resolution, the model will predict second by second vehicle trajectories (location, speed, acceleration). Approximately 320 in-use vehicles were recruited and tested on a dynamometer over three different driving cycles. For each cycle, second-by-second emissions for CO 2, CO, NO x, and HC emissions were collected and analyzed for different driving conditions. Ultimately, the model will be able to predict emissions for a variety of light duty vehicles (LDVs) in different maintenance states (properly functioning, deteriorated, malfunctioning, etc.). The model will predict emissions and fuel consumption second by second (Barth et al. 1999). Another study described the use of SMOG DOG, a remote sensing technology that is able to simultaneously measure emission concentration for CO, HC, NO x and CO 2 in the dispersing exhaust cloud of vehicles as well as the instantaneous speed and acceleration of 37

59 the vehicle. Analysis of the data derived a relationship between pollutant concentrations and a vehicle's instantaneous speed profile. Data were collected for five highway locations for parts per million (ppm) of CO, HC, and NO x. Emission factors were calculated using regression equations for six independent variables. Of the six variables, speed, speed squared, acceleration squared, ambient temperature, and humidity proved to be relevant in predicting emission rates (Yu, 1999) Improved Vehicle Activity Estimates Accurate modal modeling requires two primary components; activity specific emission factors and accurate estimates of on-road vehicle activity. Currently, few models exist that accurately represent the range of activity that vehicles undergo as part of normal driving operation on any type of roadway. To address this, a number of efforts have been undertaken to more accurately model on-road vehicle activity. Various attempts have focused on collection of actual on-road data. However, collection, analysis, and model development of field data is extremely resource intensive. Consequently, many research efforts have utilized some manner of simulation model to obtain vehicle profiles to be coupled with mode specific emission production or fuel consumption factors. However, as will be discussed later in this work, there are several inherent inaccuracies in simulation models in terms of vehicle activity estimates, especially at signalized intersections where vehicle activity is especially complicated. The following sections describe the various research efforts for vehicle activity modeling. 38

60 On-Road Vehicle Activity Modeling Recent efforts undertaken by Grant (1998) were aimed at statistically relating observed speed/acceleration characteristics on freeways as a function of vehicle class, traffic flow, and geometric highway parameters using laser rangefinders. This new approach used aggregate measures of flow and roadway geometry to predict the important load-related measures of flow. While Grant s methods work for freeway segments in Atlanta, the general methods have yet to be applied to the more complex traffic flow conditions that occur on non-freeway roads. In particular, Grant derived a regression model that related the amount of activity where accelerations exceeded 3 mph/s. Final analysis indicated that percent of acceleration activity beyond the threshold of 3 mph/s for light duty vehicles was correlated with density of vehicles on the roadway, horizontal curvature, and the percent of heavy-duty trucks in the traffic stream. For decelerations less than 2 mph/s, a correlation was found between roadway density and curvature. Roberts et al. (1999) described development of a freeway modal activity model. Activity data were collected on California freeways using chase vehicles with SnapOn Scanners, Distance Measuring Devices (DMI), or Laser-tracking coupled with DMI measurements. Percentage of significant activity (acceleration greater than 3mph/s, PKE > 60, etc.) were analyzed with regression tree analysis and significant variables determined. 39

61 Significant explanatory variables included density, flow, and fraction of mainline volume merging or diverging in a weaving section Simulation Simulation offers an attractive method to easily create vehicle activity profiles. Many packages model vehicles on a microscale so that second by second parameters such as location, speed, or acceleration can be output. Following is a discussion of various research projects that have or are using simulation for microscopic vehicle activity. However, there are several inherent problems with each of the approaches that will affect activity output and ultimately air quality estimates based on simulated vehicle activity. In Chapter 6, results of this research are compared with simulation model output and the various models presented here are critiqued. As early as 1988, Al-Omishy and Al-Samarrai (1988) had developed a road traffic simulation model that predicted emissions based on vehicle type, location along the roadway, speed, and acceleration. The FORTRAN based traffic simulation model predicted both HC and NO x and could be used for evaluation of various traffic and pollutant control strategies. The simulation model estimated vehicle activity and location based on car-following theory. Matzoros (1990) reported the development of a modal emission simulation model. The model predicted air pollution concentrations for vehicles in urban areas. Modal activity simulated by the model included the formation and dissipation of queues as well as cruise, 40

62 idle, acceleration, and deceleration at different positions along a street link. The model included emission rates disaggregated by operating mode. Queue lengths were specifically modeled. Emission factors were provided for cruise, idling, acceleration, deceleration, and creeping. The highest emission rates for all pollutants (CO, HC, NO x, and lead) corresponded to acceleration. The modal approach was based on earlier work, which had observed that pollution concentrations were the highest near intersections, tailing off midblock. Model results were compared with data from two actual locations and it was found that, with the exception of NOx, an overall agreement exists between observed and modeled values. An analytical model to estimate intersection fuel consumption was created to investigate the effects of signal timing on fuel consumption by Liao and Machemehl (1998). The model attempted to identify inter-relationships between traffic characteristics, signal control strategies, and roadway geometric conditions based on consideration of vehicle operating conditions. The model used mathematical relationships to derive fuel consumption estimates rather than using simulation. A major research effort is currently in development by Los Alamos National Laboratory to develop the TRANSIMS model, which is a simulation system for the analysis of transportation options in metropolitan areas. The base of the system is a cellular automata 41

63 microsimulation model producing second-by-second vehicle positions defined by 7.5 meter cell locations. TRANSIMS is a set of integrated analytical and simulation models and supporting databases dealing with prediction and simulation of trips for individual households, residents, and vehicles as well as movement of individual freight loads (Williams et al. 1999). Many traffic simulation and optimization models such as TRANSYT-7F, INTEGRATION, FREQ, NETSIM, and INTRAS also have incorporated modules for estimating emissions. These modules are structured to be sensitive to modal model output but none of the models were developed based on on-road emission or vehicle activity data (Yu, 1999). Rakha et al. (1999) likewise is in the process of developing a modal microscopic simulation model, which combines a traffic simulation model with an emission module. The model uses car-following and lane changing logic to simulate vehicle activity. Speeds and headway are updated and calculated each tenth of a second. Acceleration is modeled as speeds are updated every deci-second based on the distance headway and speed differential between the subject vehicle and the vehicle immediately ahead of it, which can result in unrealistically high accelerations. To compensate, the model uses a linear acceleration decay function that decreases the vehicle's acceleration as a function of its speed, resulting in a linear speed/acceleration relationship. 42

64 Fuel consumption and emissions were calculated from data collected on a dynamometer at Oak Ridge National Lab. This provided fuel consumption and emission rates for a range of speeds from 0 to 75 mph km/h and for a range of accelerations from m/s 2 to 8.0 m/sec 2. The model was developed to represent typical driving conditions including idling, acceleration, and deceleration. Data were collected for eight vehicles and an emission model for a composite vehicle created. The simulation model models acceleration as a by-product of car-following logic and does not actually replicate realistic on-road behavior. In another study, the microsimulation model TRAF-NETSIM, used for urban roadways, and the microsimulation model INTRAS, used for modeling freeways, were used to develop a modeling framework for prediction of vehicle activity in regional areas for improved emission estimates. Output from the simulation modules was used to develop relationships between basic link characteristics and the time spent in each operating mode. Field data using instrumented vehicles for a freeway segment were used to validate model results. The relationships developed were then incorporated into a post processor from the Urban Transportation Planning Software (UTPS) four-step planning model so that regionwide estimates of vehicle activity can be applied with the existing state-of-practice in regional modeling (Skabardonis, 1997). 43

65 3.2.3 MEASURE To address the various problems with existing transportation-related emission models and shortcomings in more contemporary vehicle activity modeling such as simulation, a research-grade motor vehicle emissions model based in a geographic information system (GIS) platform is underway at Georgia Institute of Technology. The Mobile Emission Assessment System for Urban and Regional Evaluation (MEASURE) predicts emissions based on operating mode. Operating modes represented include cruise, idle, deceleration, acceleration, and other modes where power demand leads to enrichment. The model applies both vehicle characteristics, such as model year, engine size, etc, and speed/acceleration profiles to predict emissions. Because MEASURE resides in a GIS, it is able to capture both spatial and temporal characteristics of vehicle fleet and modal activity. It is also able to incorporate both existing and real-time datasets such as Highway Performance Monitoring System (HPMS) traffic counts. Many agencies already maintain traffic information in some form of a spatial database so the GIS platform allows integration of various datasets. The model also allows an enhanced approach to modeling emissions spatially (Guensler et al. 1998). The research model is based in a GIS package, which allow storage of all spatial and temporal attributes of the modeling regime and integration of a wide variety of data sources, spatial attributes, 44

66 and temporal distributions for use by external programs to estimate emissions. The GIS actually contains the physical transportation network with accompanying topology and attributes such as link length, number of lanes, grade, capacity, etc. Locations where enrichment is more likely to occur such as freeway on-ramps or signalized intersections can also be identified. The model takes in spatial data (road segment and census block) about vehicle activity and technology and outputs estimated, gridded, mobile exhaust emissions. It employs modal emission rates developed in house as well as MOBILE based emission rates. It relies on modal vehicle activity measures; starts, idle, cruise, acceleration, and deceleration. Vehicle technology characteristics (model year, engine size, etc) and operating conditions (road grade, traffic flow, etc) were also developed at a large scale (small zones and road segments). The scope of MEASURE is currently being expanded to include microscale air quality estimation and estimates of all mobile source pollutants (CO, HC, NO x, CO 2, and toxics). As designed, MEASURE will be able to read spatial databases and estimates of modal activity and then provide facility-level estimates as well as gridded estimates of the various pollutants. From a research perspective, MEASURE will be able to make comparisons of standard speed correction factor approaches versus activity-specific 45

67 approaches since both types will be embedded. From a practical perspective, it will provide planners and engineers a toolkit to generate regulatory level reporting as well as a provide flexible means of gaming various TCMs. Both of these capabilities make MEASURE an attractive tool for the research discussed in this paper. One of the important facets of the model are the modal emission rates generated during model development. These emission rates provide the ability to explore the emission impacts of changes in acceleration profiles or idling fractions. Some of the major features of MEASURE are: the model includes modal emission rates as well as MOBILE emission rates; user-defined grid cells; an improved spatial aggregation technique; and the inclusion of local road emissions. Modal emission rates are designed to estimate emissions for specific vehicle activities (idle, cruise, acceleration, and deceleration) and vehicle technology combinations (cold and warm engine starts, hot-stabilized, and enrichment). Over a region, engine start emissions are estimated for census blocks, and hot-stabilized and enrichment emissions are estimated for road segments (intersection to intersection). Emissions from the zones and lines are aggregated into user defined grid cells directly by completing polygon-on-polygon and line- 46

68 on-polygon spatial summarization. The main advantage of MEASURE is that it allows users to model a wide-range of strategies that may have an effect on emissions (i.e. signal timing and high-occupancy vehicle lanes) (Sarasua et al. 1999). 3.3 Fundamentals of Vehicle Activity in Traffic Engineering In the above sections, the evidence for moving towards a modal approach to analyzing transportation-related air quality is presented as well as an overview of other research efforts into modal modeling. Many research efforts focusing on the activity side of the emission equation are using either simulation modeling or other activity estimation based on speed/acceleration relationships commonly used in traffic engineering. Following is a synopsis of the common representations of vehicle activity used in various traffic engineering applications as well as description of several research efforts that attempt to better describe vehicle activity relationships. In the data analysis chapter, Chapter 6, field data are compared to the common traffic engineering relationships described below Acceleration Performance of Passenger Cars The most common comprehensive early study on vehicle acceleration/speed profiles was a research effort by St. John and Kobett (1978), reported in National Highway Cooperative Research Program (NCHRP) 185. The report presents analysis of speed and acceleration data points from a single passenger car (1970 Chevrolet Impala sedan) driven over a test course. The study used an on-board light beam oscillograph recorder to record a 47

69 time trace of speed and acceleration. Test results indicated a linear relationship existed between speed and acceleration for passenger cars. This relationship, for passenger cars on zero grade, is given by: a = a o [1-(V/V m )] (3-1) Where: a = acceleration capability at speed V; a o = maximum acceleration for speeds 0; V = vehicle speed; V m = a pseudo maximum speed indicated by the linear relation between acceleration and speed when data are fitted in the normal operating range. Maximum acceleration was calculated from an earlier study and was not explained in detail. The mathematical representation of maximum acceleration (a o ) is given by: Ao = [131.2/(W/bhp)] (3-2) Maximum acceleration is achieved at zero speed and linearly decreases as speed increases. Maximum acceleration is a linear function of the inverse of the weight/horsepower ratio. Major drawbacks to the test study were that data were limited to one older model year vehicle on a fixed test route and the relationship describes the upper bound of the 48

70 speed/acceleration curve rather than a distribution of speeds and accelerations expected under normal vehicle operation. This upper-bound linear speed-acceleration relationship is presented in Figure 3-2. A correlation between acceleration performance and grade was noted in the St. John and Kobett study as well. The relationship is described in Equation (3-3). a GV = a LV - Rg (3-3) 49

71 Figure 3-2: Linear Speed-Acceleration Curve where: a GV = acceleration capability at speed V on grade; a LV = acceleration capability at speed V on level terrain; g = acceleration due to gravity; and R = percent grade expressed as a decimal (for a 4% upgrade, R = 0.04). The acceleration capability of vehicles has primarily been used for road design. Acceleration capability influences passing zone lengths and acceleration lane lengths. Acceleration capability is also a factor in signal timing design, calculation of fuel economy and travel time values, and in estimating the return to normal traffic operation after a breakdown in traffic flow patterns (ITE, 1992). Microscopic simulation models often employ maximum on-road acceleration to generate individual vehicle activity profiles. The Traffic Engineering Handbook (ITE, 1994) lists maximum acceleration rates for both passenger cars and heavy trucks based on the vehicle's weight-to-power ratio. The weight/power ratio is a measure of the vehicle's ability to accelerate and maintain speed on upgrades. Weight-to-power ratio is the gross weight of the vehicle in pounds divided by the power in horsepower. Weight is a rough indicator of resistance to motion so the higher the 50

72 weight/power ratio the lower the acceleration performance, while a low weight/power ratio reflects higher performance capabilities (ITE, 1994). Details are shown in Table 3-1. Acceleration rates for different speed ranges are given in Table 3-2. Table 3-1: Maximum Acceleration from Rest by Vehicle Type and Weight-to- Power Ratio (source: Traffic Engineering Handbook 4th Edition (ITE, 1994)) Vehicle Type Passenger Car Tractor- Semitrailer Weightto-Power Typical Maximum Acceleration Rate on Level Road (mph/s) Ratio 0 to 10 0 to 20 0 to 30 0 to 40 0 to 50 (lb/hp) mph mph mph mph mph Table 3-2: Maximum Acceleration by Speed Range by Vehicle Type and Weightto-Power Ratio (source: Traffic Engineering Handbook 4th Edition (ITE, 1994)) Vehicle Type Passenger Car Weight-to- Power Ratio Typical Maximum Acceleration Rate on Level Road (mph/s) (lb/hp) 20 to to to to 60 mph mph mph mph

73 Tractor- Semitrailer Grade affects the maximum acceleration that can be achieved and according to the Traffic Engineering Handbook (ITE, 1994) has the following relationship, which is similar to NCHRP 185: a GV = a LV - Gg/100 (3-4) where: a GV = maximum acceleration rate at speed V on grade (ft/sec 2 ); a LV = maximum acceleration rate at Speed V in level terrain (ft/sec 2 ); G = Gradient (%); and g = acceleration of gravity (32.2 ft/sec 2 ). Maximum acceleration on upgrades by speed range is provided in Table 3-3 for passenger cars and heavy trucks. A graphical representation of maximum speed versus grade is shown in Figure 3-3. Table 3-3: Maximum Acceleration on Upgrades by Speed Range (source: Traffic Engineering Handbook 4th Edition (ITE, 1994)) Speed Passenger Car (30 lb/hp) Tractor-Semitrailer (200lb/hp) Change Level 2% 4% 6% 10% Level 2% 4% 6% 10% 0 to 20 mph * * 52

74 20 to 30 mph * * * 30 to 40 mph * * * 40 to 50 mph * * * * 50 to 60 mph * * * * *Truck unable to accelerate or maintain speed on grade Figure 3-3: Maximum Acceleration on Upgrades for Passenger Cars by Speed Acceleration Performance of Heavy Trucks NCHRP 185 also describes acceleration performance of heavy trucks. A heavy vehicle acceleration model was developed using a computer model to simulate vehicle 53

75 movement. Speed-acceleration traces were output and compared with heavy vehicle data from the Western Highway Institute and the Road Research Laboratory. The following formula was derived to describe the acceleration capabilities of large trucks (St. John and Kobett, 1978): A e = [?V/(?V + S p t s (A p - A c )]A p V > V 1 (3-5) Where: A e = effective acceleration (ft/sec 2 );? = a parameter that depends on the range of engine speeds; typical values range from 0.33 to 0.43 (0.4 is recommended); V = vehicle speed (ft/sec); S p = one time the sign of A p (which can be either + or -); t s = actual time required to shift gears (sec); A p = power-limited acceleration (with the engine employed and vehicle at speed (V) uses the average available net horsepower (ft/sec 2 ); A c = acceleration in coasting at vehicle speed V (ft/sec2); uses an average gear ratio for the coasting chassis losses; and V 1 = the maximum speed in lowest gear ratio (ft/sec). 54

76 The research acknowledged that the effect of grade was underestimated by applying the simple addition of a gravity component to zero-grade performance. When a heavy truck starts and accelerates on zero grade, the time to shift gears is about 1.5 seconds, consequently a large portion of the time is spent coasting without power applied. If a truck starts out on a positive grade, the acceleration in each gear ratio is lower and the time in each ratio is longer so the engine is usefully employed a larger percent of the time. In equation 3-5 above, the terms A c and A p included the direct added effect of grade (St. John and Kobett, 1978) Deceleration Performance Deceleration occurs either when the accelerator pedal is released due to the retarding effects of constant resistance to motion, an increase in resistance to motion, or when vehicle brakes are used. Without brakes, deceleration rates are greater at high running speeds since because of resistance to motion. At 70 mph, release of the accelerator pedal results in a deceleration rate of 2.2 mph/s. Around horizontal curves or on a gradient, the resistance to motion will increase and deceleration will occur without a corresponding increase on the accelerator pedal. Maximum deceleration occurs with braking and is determined by the retardation forces developed in brake drums or discs negating slip between the pavement and tire. Roadway and tire friction also affect maximum deceleration rates. Maximum deceleration is typically only applied in emergency situations. For traffic 55

77 engineering applications, such as determining vehicle clearance intervals at traffic signals, a common deceleration rate of 6.8 mph/s is used (ITE, 1994). 3.4 Discussion An overview of current literature which indicates a relationship between engine operating mode and emissions and literature that outlines other efforts to develop more realistic emission rates and activity estimates was presented in this chapter. The intent was to provide a background on previous and ongoing research into modal emissions modeling and to provide a sampling of the evidence suggesting that a modal approach is more accurate. It was beyond the scope of this paper to critique the validity or accuracy of each work. Several works employed questionable methods such as the level of service estimation used by Sierra Research or the unusually high acceleration rates reported in the three-city instrumented vehicle study. However, even with flaws, the bulk of research does point towards a relationship between engine mode and emission output. 56

78 CHAPTER IV 4. RESEARCH APPROACH This Chapter presents the research framework for this dissertation work. First the problem is defined, followed by presentation of the research hypothesis and research objectives. Statistical techniques considered for analysis of the data are discussed. A final statistical model is presented including an overview of the response variables. Following, the various predictor or independent variables hypothesized to affect vehicle activity, which were considered as part of experiment design, are outlined. Although presentation of the data collection protocol follows this chapter, a note is made of whether each independent variable could be and was actually included in the data collection phase of this research. 4.1 Statement of Problem The previous chapters explained the need for a modal approach to transportation-related air quality modeling. For a modal emission model to accurately predict emissions, both activity-specific emission rates and accurate estimates of vehicle activity are required. Much research activity has focused on the emission rate side of the modal-based emission prediction equation to develop methods that relate emissions specifically to a particular speed-acceleration combination or on identifying specific vehicle operating modes where emissions are disproportionate compared to 56

79

80 other modes. Efforts to model on-road vehicle activity have been undertaken to a much lesser extent. As discussed in Section 3.2.2, many emerging models are basing vehicle activity on a limited sample of vehicle profiles or on algorithms that attempt to simulate queuing, acceleration, and deceleration. None of these methods have been validated as to whether the output realistically models the wide range of vehicle activity encountered on the roadway. Additionally, current modeling efforts have not yet been shown to relate vehicle activity to on-road conditions such as grade or traffic volumes. Without accurate vehicle activity estimates, modal emission models are handicapped in their ability to successfully relate activity-specific emission rates with accurate estimates of vehicle activity. Consequently, the activity prediction side of the equation may be the limiting factor in accurate deployment of modal models. No modal model can be complete unless it addresses the air quality impacts of signalized intersections. By their nature, signalized intersections encompass much of the modal activity experienced by motor vehicles in an urban area. A significant amount of modal vehicle activity occurs within a relatively short distance of the intersection depending on queue length. Vehicles decelerate to a stop, idle, and then accelerate from rest. Even for vehicles not stopped or slowed by the signal, a large number of interactions with other vehicles occur leading to "rough" traffic flow. Significant modal activity may also occur at other locations along signalized links experiencing heavy congestion resulting in over-capacity stop and go conditions. Other locations of significant modal vehicle activity include freeway ramps and along 58

81 freeway segments where vehicles undergo braking due to interference with other vehicles and rapid accelerations when merging with existing traffic. Although, freeway segments may be the source of most hard accelerations at high speed, nonfreeway roadway links make up the bulk of existing roadways and the majority of vehicle activity. Consequently, the modal emissions impact of signalized intersections is highly relevant. 4.2 Hypothesis to be Tested This purpose of this work was to create a methodology to predict microscopic vehicle profiles at signalized intersections that can be used as input to regional or microscale transportation-related air quality models that use an activity-specific (modal) approach. The research hypothesis can be encapsulated by the following: Research Hypothesis: Modal activity on signalized roadways can be forecasted as a function of macroscopic activity (traffic volume, percent heavy trucks, etc.) and facility geometric properties (roadway grade, number of lanes, distance to downstream intersection, etc.). 4.3 Objectives To model activity at signalized intersections, this research had three main objectives, which are detailed in the following sections. 59

82 Objective 1: Develop a method to sample representative modal activity on signalized roadways to represent the widest range of geometric and operation conditions possible. The goal of this research was to develop a model that can predict modal vehicle activity at signalized intersections and along signalized roadway segments using actual field data. The model ideally should be able to predict vehicle activity based on those operational and geometric characteristics of the roadway, shown to influence vehicle activity. Consequently, data collection sites were selected to represent as broad a range of different characteristics as was economically viable. To forecast vehicle activity based on significant roadway and operational conditions, all the variables that may affect vehicle operation were identified and then those that could realistically be included were selected for the final experimental design. Objective 2: Develop a robust and repeatable methodology for forecasting modal activity on signalized roadways. This objective entails development of a statically valid methodology to analyze the data. Included in this portion of the research was identification of the "best" statistical procedure for analyzing the collected data. Various statistical methods that were considered are covered in section 4-4. An overview of the final statistical method is 60

83 provided in section 4-5. Selection of relevant response and potential predictor variables are discussed later. Objective 3: Develop a model that will integrate with MEASURE for output of vehicle activity given specific roadway, fleet mix, and operational characteristics. The final objective was to develop the framework for integration of a prediction model with the Georgia Tech MEASURE model discussed in Section Scope of Work To meet the objectives outlined for this work, various tasks were undertaken. First, an in-depth literature review was completed with relevant background information to this work as presented in Chapters 2 and 3. Next, the experiment was designed including a data collection plan and selection of appropriate statistical analysis procedures. The data collection is discussed in Chapter 5 and the selection of the appropriate statistical analysis procedure, including identification of response variables and final selection of predictor variables is the subject of this chapter. The defining feature of this research work was development of a method capable of generating complete vehicle profiles along the path of a signalized link through the intersection and onto the following link. An individual vehicle profile is illustrated in Figure 4-1. All individual vehicle traces are summed to reflect the number of vehicles on the link. The statistical analysis determined which of the 61

84 independent operational or geometric variables of the study locations were relevant in influencing vehicle activity. The final output of the statistical model is a Joint Acceleration-Speed Probability Density Function (JASPROD), which is a threedimensional (tri-variable) function of speed, acceleration, and the joint probability for a given speed-acceleration bin. An empirical JASPROD is created by sampling the Figure 4-1: Sample Vehicle Trace 62

85 simultaneous speed and acceleration trace of a vehicle along a specified path (or run), such as a vehicle's trajectory from the point of queuing to some point downstream. For the final model, data were divided by homogeneous zones of activity (distance from the stopping point or from the intersection stopbar) and by homogenous predictor variables determined by statistical analysis, as discussed in Chapter 6. Data were collected in one-second intervals so the resulting JASPROD are for one second intervals. JASPRODs are created by dividing vehicle traces into a matrix of speed and associated accelerations bins. Each bin has a unique speed and acceleration range. A JASPROD is shown in Figure 4-2 and Table 4-1. Once data are binned, the probability of any bin can be calculated by dividing its frequency by the sum of the frequencies of all bins. For each given geometric and operational condition that is investigated, the frequency of activity in a specific speed-acceleration bin is the number of seconds of operation in a given bin divided by the total number of seconds of activity. The sum of all frequencies for the vehicle trace will equal 1. The emission rate models, which will be described in Section 4.6, only require the fraction of activity for the specific modal variable which was shown to be correlated to emission rates (i.e. the percent of activity where acceleration >= 6.0 mph/s). Data could have been analyzed in this manner, however a method that allowed a distribution of data as output was desirable, since response variables may change in the future depending on results of on-going emission rate modeling. With output in the form of a distribution of data, the model output can be used with any 63

86 emission rate model that identifies critical modal variables. For example, if a 3- dimensional activity distribution is available and future research identifies acceleration greater than 5 mph/s as significant, the total fraction of activity that falls within this range can be selected from the JASPROD as shown in the shaded section in Table 4-1. This research model ultimately will be used as input to the Georgia Tech MEASURE model for regional air quality modeling and final data output designed for Figure 4-2: Joint Acceleration-Speed Probability Density Function (graphical form) 64

87 Table 4-1: Joint Acceleration-Speed Probability Density Function (matrix form) (Range of Activity Where Acceleration >= 5.0 mph/s is Shaded) Speed Acceleration (mph/s) (mph) compatibility with the model. The research model may also be used for microscopic air quality modeling, evaluating the effectiveness of transportation control measures (TCMs), or intelligent transportation system (ITS) alternatives. 4.5 Statistical Modeling The purpose of the statistical modeling was to determine which predictor variables influence vehicle activity behavior so that the data can be stratified by those variables and 3-dimensional matrices of speed and acceleration created. To determine which operational and geometric variables influence how vehicles operate, statistical analysis can be used to compare whether two distributions of data, which were disaggregated by the various variables, differ statistically. Unfortunately, no common 65

88 test existed, that allowed comparison of complex 3-dimensional distributions. The speed-acceleration matrix produced from data preparation was a three dimensional distribution (speed x acceleration x frequency). The three dimensional distribution may have been reduced to two dimensions if a distribution of speed versus frequency or acceleration versus frequency were used singularly or if a product of the two, a surrogate for power (speed x acceleration) versus frequency were used. There are various methods that may be used to test differences across two distributions. Goodness-of-fit tests examine two random samples to test the hypothesis that two unknown distributions are identical. Most methods are limited to 2-dimensional or simple 3-dimensional distributions. Regression uses a single response variable regressed against one or more predictor variables. For example, the percent of activity where acceleration are >= 3.0 mph/s can be regressed against a number of variables such as per lane volume, percent trucks, lane width, etc. A discussion of the various methods tested and the final methodology used follows Chi-Square Tests The chi-square test is the oldest and best known goodness-of-fit test. The test assumes that the observations are independent and that the sample size is reasonably large. This method can be used to test whether a sample fits a known distribution, or whether two unknown distributions from different samples are the same. The test 66

89 assumptions are that the sample is random and that the measurement scale is at least ordinal (Conover, 1980). The chi-square provides a relatively easy to apply approach for analyzing two dimensions of data (acceleration and frequency or speed and frequency). The main problem with using the chi-square is the orders of magnitudes of separate tests that would have to be conducted to test all possible combinations of variables in the datasets. For example, to just test 5 queue positions, grades from -9 to 9, and level of service, a total of 570 datasets would result, assuming data existed for each mutually exclusive group of variables. To compare the distributions to determine where they differ, a large share of the 570 datasets would have to be compared to the all the others. This quickly becomes logistically infeasible. Additionally, since the test only allows comparison across 2-dimensions, only distribution of accelerations from one dataset could be compared to another or the speed distribution of two datasets could be compared at a time. It does not allow comparison of the relationship between speed and acceleration except as a two-dimensional product. Another problem is applying the chi-square is that stratification of the data often resulted in cells with 0 observations, which presents difficulty for the chi-square test since it cannot handle cells with no observations. 67

90 4.5.2 Kolmogorv-Smirnov Two-Sample The Kolmogorov-Smirnov (K/S) two-sample test compares the empirical distibution functions of two samples, F 1 and F 2. The Kolmogorov-Smirnov test is a non-parametric test, which can be used to test whether two or more samples are governed by the same distribution by comparing their empirical distribution functions. The Kolmogorov-Smirnov two-sample test is illustrated in Figure 4-3 for a sample dataset using data for the first vehicle in the queue for an intersection with a positive 9% grade and data for the first vehicle in the queue for an intersection with a negative 9% grade. Calculations and graphs were made in SPLUS statistical software. In the figure the cumulative distribution function (cdfs) for each distribution is plotted. If the distributions are similar, the cdfs would also be similar. The wide variation between the two indicates that the two datasets were from quite different distributions. The test also provides a numerical solution as well as a visual plot. The Kolmogorov-Smirnov two-sample test provides an improved methodology over the chi-squared test since data does not have to be assigned arbitrarily to bins. Further, it is a non-parametric test so a distribution does not have to be assumed. However, the main disadvantage to the K/S is similar to the chi-square in that the orders of magnitudes of separate tests that would have to be conducted to test 68

91 Figure 4-3: Comparison of Empirical cdfs for Accelerations on a 9% Grade (x) and -9% Grade (z) all the possible combinations of variables in the datasets become logistically infeasible. The K/S also only allows comparison across 2-dimensions. The distribution of accelerations from one dataset could be compared to another or the speed distribution of two datasets could be compared. It does not allow comparison of the relationship between speed and acceleration except as a two-dimensional product. 69

92 4.5.3 Linear Regression Regression analysis is a statistical method used to explain a dependent variable as a mathematical function of one or more independent variables (Studenmund, 1985). For example, regression may be used to estimate the number of accidents along roadway segments as a function of pavement condition. Linear regression is a commonly used and easily understood statistical method. Linear regression explores relationships that can be described by straight lines or their generalization to many dimensions. Regression allows a single response variable to be described by one or more predictor variables. Linear regression was a viable analysis tool for the research data. This type of analysis would regress a single response variable based on relevant emission producing modes of activity against various predictor variables. The response variable would have to be a variable such as % of time spent in accelerations greater than 6.0 mph/s, which would be regressed against related variables such as LOS, distance to downstream intersection, per-lane volume, etc. This method would necessitate development and validation of a separate model for each response variable. The largest problem with this statistical approach is that a linear relationship, or a transformation form of a linear relationship, must exist between the response and relevant predictor variables. Relationships may exist between vehicle activity and certain ranges of a predictor variable but not others. For example, volume-to-capacity 70

93 may not influence vehicle behavior in the lower ranges, (i.e 0 to 0.7) but may be relevant at higher ranges (0.7 to 1.0+). Additionally, even at the higher range, a linear relationship may not exist (i.e. V/C from 0.7 to 1.0+ may affect vehicle operation but the effect is constant rather than linearly increasing or decreasing). Even with transformations on the data, such as taking the log, a relationship may not show up or a relationship may be forced between non-relevant ranges of the variable. Figure 4-5 shows this type of relationship between percent activity greater than or equal to 6 mph/s and queue position. As shown, an exact linear relationship does not exist between queue position and hard accelerations. An ordinary least squares regression of hard accelerations against queue positions only yields an R 2 value of 0.27, indicating that the model only explains 27% of the deviation. Another disadvantage to this method is that only a single response variable can be used. The relationship between speed and acceleration bins cannot be linked to geometric and operational characteristics. Also, an individual model must be developed for each response variable, rather than being able to derive a 3-dimensional distribution of speed and acceleration activity. 71

94 Figure 4-5: Graduated Non-Linear Relationship Between Percent Hard Accelerations and Queue Position Hierachiacal Based Regression Tree Analysis Binary recursive partioning, more commonly referred to as hierachiacial treebased regression, is similar to forward stepwise variable selection methods. It is also commonly referred to as classification and regression tree analysis (CART). One of the advantages to HTBR or CART is that it assists in detecting the underlying structure in data (Breiman et al., 1984). This technique generates a "tree" structure by 72

95 dividing the sample data recursively into a number of groups. The groups are selected to maximize some measure of difference in the response variable in the resulting groups. One of the advantages of regression tree analysis over traditional regression analysis is that it is a non-parametric method which by definition does not require any distribution assumptions and is more resistant to the effects of outliers (Roberts, 1999). The term CART is often used since it allows analysis of both classification and regression analysis. Classification analysis are those where the endpoints (or terminal nodes) are factors which are non-numeric. An example of a classification tree is rulebased method for determining the chances of survival or non-survival for a heart attack patient based on monitored variables, such as blood pressure, during the first 24 hours following the attack. Regression trees are those where the model endpoints end in predicted numerical values. Tree-based modeling is an exploratory technique that is increasingly being used for: devising prediction rules that can be rapidly and repeatedly evaluated; screening variables; assessing the adequacy of linear model; and summarizing large multivariate datasets (Mathsoft, 1997). 73

96 Description of Test The model uses a set of classification or predictor variables (x), and a single response variable (y). Regression tree rules are determined by a procedure known as recursive partioning. The regression tree methodology proceeds by iteratively asking and answering via a numerical search process: 1) Which variable of all of the predictor variables offered should be selected to produce the maximum reduction in variability of the response? and 2) Which value of the selected variable (discrete or continuous) results in the maximum reduction in variability of the response? The binary partioning algorithm recursively splits the data into increasingly homogenous regions. The splitting continues until either a desirable end condition is met with a homogeneous end node or too few observations exist to proceed further. Node homogeneity is determined by deviance where a deviance of zero indicates a perfectly homogenous node (Wolf et al 1997). formulas: The partioning process can be explained mathematically by the following three D s = Σ m=1 to M (Y ms - µ s ) 2 (4-1) (all x) = D s - (D l + D r ) (4-2) (all x) = Σ m=1 to M (Y m -µ s ) 2 - (Σ p=1 to P (Y pr - µ r ) 2 + Σ q=1 to Q (Y qr - µ r ) 2 ) (4-3) 74

97 D s in equation 4-1 is the deviance of node S, which is to be split into two new nodes (designated left and right). Each of the subnodes, left and right contain a portion of the sample points in s. D s is the sum of squared error (SSE) at node s, which is summed over all observations m in node s. The squared error term at node s is calculated by the difference of the mth observation of the dependent variable y and the mean µ of M observations in node s (Roberts, 1999). A split occurs on a "parent" node on a particular value of one of the independent variables specified. The deviance reduction function as shown in Equation 4-2 evaluates deviance over all independent variables where D l and D r are the residual mean deviances of the left and right nodes. An optimal split is selected from among all possible independent variables, X. Tree-based models have various advantages over linear and additive models. One of the main strengths of regression tree analysis is that independent variables do not have to be specified in advance. A regression tree picks only the most important predictor variables that result in the maximum reduction in deviance. Another advantage is that results are invariant with respect to monotone transformations of the independent variables so that the "right" transformation does not have to be sought. Regression trees are a nonparametric procedure, meaning that a functional form does not have to be specified. They are able to work well with data that have multiple structures rather than uncover a single dominant structure in data as many parametric models do. Regression trees are also robust to the effect of outliers since splits usually occur at non-outlier values (Roberts, 1999). 75

98 Applicability of Test to Research Regression tree analysis appears to offer the most feasible and appropriate approach to testing for differences in vehicle activity based on geometric or operational characteristics. The single largest advantage to regression tree is the ability to model non-linear relationships. The model divides responses by ranges of predictor variables in the data where a relationship is shown to exist and does not require a relationship to be derived between all ranges. For example, if grade greater than 5% is relevant, regression tree analysis can split the data at this point without forcing a relationship for other ranges of grade. Regression tree analysis also inherently handles correlation between predictor variables. For example, level of service may only be relevant if the intersections are less than 1000 feet apart. Regression tree analysis can model this type of relationship easily while regression analysis is unable to pick up this type of relationship. Chisquare and K/S can also model this type of relationship. However, it would require an immense number of modeling iterations to model all subsets of data. The main disadvantage to regression tree analysis is that only a single response variable can be used, as for linear regression. The relationship between speed and acceleration bins cannot be linked to geometric and operational characteristics. Also, an individual model must be developed for each response variable, rather than being able to derive a 3-dimensional distribution of speed and acceleration activity. 76

99 4.6 Research Scope and Presentation of Statistical Approach Initially, this research was funded by the United States Environmental Protection Agency and the Federal Highway Administration. The scope of research extends existing research one step further and provides a methodology to predict vehicle activity that can be reasonably applied. The research is not comprehensive nor applicable to all signalized roadways under all operating and geometric conditions. The research scope was limited by both resources and practical constraints. Ideally, given enough time and an unlimited budget, data for all combinations of geometric and operational conditions could be collected and analyzed. However this would be an enormous undertaking. Consequently, the main constraint was resources. Practical constraints also limited the research. Data were only observational and could only be collected by observing situations over which the researcher had no control. It was impossible to set up control groups and vary variables. Further, it was expected other factors, such as trip purpose or engine horsepower, may significantly affect vehicle activity. Although it was impossible to account for these types of factor, other research efforts are underway to accomplish just that (Wolf et al, 1999). The statistical approach used for data analysis for this research work involves a two-step process. First, HTBR was used to reduce the create a best "fit" model for each response variable which identified the predictor variables that most influence vehicle activity ranges determined to be relevant by emission rate models. However, since a representative range of response variables were used (deceleration, average 77

100 speed, medium acceleration, and hard acceleration activity), it may be assumed that these variables will affect all ranges of modal activity. For example, if grade influences both accelerations >= 3 mph/s and >= 6 mph/s, the inference would be drawn that it will also impact accelerations >= 5 mph/s. The second step is to validate the results of regression tree model using the K/S test to compare distributions of stratified data. 4.7 Response Variables Part of the MEASURE model research was to create new emission rates that are more representative of real-world modal activity. Emission rates are being developed for hydrocarbons, oxides of nitrogen, and carbon monoxide. The production of carbon monoxide at signalized intersections is of particular concern because of the immediate health effects. Therefore, CO is usually analyzed on a microscale, whereas HC and NO x are most often analyzed on a regional scale. Frequency of specific modes of operation were the response variable in the prediction process. The actual response variable were percent of activity within a given joint range of speeds and accelerations. In order to compare different locations with a differing number of observations for each speed-acceleration "cell", the cell aggregation was based on the results of emission rate modeling, discussed in the next section. Consequently, the response variable was prediction of frequency of activity 78

101 for a given range of vehicle activity, such as percent of activity where acceleration >= 3.0 mph/s. In the following sections, the development of emission rate models for carbon monoxide, hydrocarbons, and oxides of nitrogen are presented. Each model predicts pollutants based on relevant vehicle or operating mode variables. The relevant operating mode variables from the emission rate models will serve as the response variables for the statistical model. A major part of the MEASURE model research effort has been development of more accurate emission rate models, which reflect the influence of modal activity. A detailed description of how the carbon monoxide model was derived is provided in Section Since a similar process was used to derive the final model for hydrocarbons and oxides of nitrogen, they are also presented but without an in-depth description of the process. It should be noted that emission rate modeling will always be in a state of flux. Ongoing revisions to the model (addition of new test data, identification of previouslyundefined relationships due to improved explanatory power of the model as new variables are added, will result in revisions to the emissions-related variables of concern. Consequently, this research attempted to identify variables that influenced overall vehicle activity, not just predict a relationship between a single response variable such as accelerations >= 6.0 mph/s and predictor variables as explained in Section

102 4.7.1 Carbon Monoxide Model The CO model for passenger cars was developed by analyzing a data set of more than 13,000 hot-stabilized laboratory treadmill tests on 19 driving cycles (specific speed versus time testing conditions), and 114 variables describing vehicle, engine and test cycle characteristics (Fomung, 1999). The data set represents almost two decades of in-use driving tests conducted by the EPA and CARB and compiled by the EPA s Office of Mobile Sources for use in developing the MOBILE model. The emission rate model for CO, presented here, was estimated with a response variable as the logarithm of the emission rate ratio for carbon monoxide. The ratio is the vehicle emission rate (in grams/second) driven on a given cycle (or across a speed/acceleration matrix) divided by that vehicle s emission rate while driving on the FTP Bag 2. The model predicts the ratio of g/second emission rates for each vehicle technology group. The following sequence of equations shows the method of calculating the predicted emission rate for CO in units of either g/second or g/mile: Ψ CO (g/sec) = Ψ CO (g/mile) * S / t (4-4) Ψ COBag2 (g/sec) = Ψ COBag2 (g/mile) * 3.91/866 (4-5) R CO (rate ratio) = P CO (g/sec) / Ψ COBag2 (g/sec) (4-6) 80

103 Where Ψ CO is the measured or observed CO, P CO is the predicted CO, Ψ COBag2 is the FTP Bag2 rate of CO for a given vehicle, S is the driving cycle distance in miles, t is the cycle duration in seconds, 3.91 is the hot stabilized FTP Bag 2 sub-cycle distance in miles, and 866 is the FTP Bag2 sub-cycle duration in seconds. On a vehicle by vehicle basis this implies that after calculating R CO from the response variable, the predicted rate in g/second can be obtained by: P CO (g/sec) = R CO * Ψ COBag2 (4-7) Note that Equation 5-13 is similar in form to the embedded algorithm in MOBILE, which gives emission rates as BER x Correction Factors. Where BER stands for base emission rate, akin to Ψ COBag2 ; R CO is a composite representation of several variables and can be thought of as speed, load, and technology correction factors. Equation 4-7 can be easily converted to g/mile by using; P CO (g/mile) = R CO * Ψ COBag2 * 1/AVGSPD (4-8) Where AVGSPD is the average speed of the speed - acceleration profile of the driving schedule. 81

104 The CO model is presented in both an estimation form, and a prediction form. The estimation form is the regression equation 4-9: LogR CO = *AVGSPD *ACC *IPS *ips45sar *ips90tran *tran3idle *tran5mi *finj3sar *cat3tran *sar3tran *flagco (Fomung, 1999) (4-9) Where: AVGSPD is the average speed of the driving cycle in mph; ACC.3 is the proportion of the driving cycle on acceleration greater than 4.8 kph/s (3mph/sec); IPS.X is the proportion of the driving cycle on inertial power surrogate (IPS) (speed x acceleration) greater than X mph 2 /sec (Washington et al., 1994). Thus IPS.60 implies IPS greater than 60 mph 2 /sec; ips45sar2 is an interaction between IPS.45 (IPS >= 45 mph 2 /sec) and a vehicle with no air injection; ips90tran1 is an interaction variable for a vehicle with automatic transmission on IPS.90 IPS >= 90 mph 2 /sec; cat3idle is an interaction variable for a 3-speed manual transmission at idle; 82

105 tran5mi1 is an interaction variable for a 5-speed manual transmission vehicle with mileage <= 25k miles; finj3sar3 is an interaction variable for a vehicle that has throttle body fuel injection and pump air injection; cat3tran1 is an interaction variable for a vehicle with automatic transmission and TWC; sar3tran4 is an interaction variable for a vehicle with 4-speed manual transmission and pump air injection; and flagco is a flag used to tag a high emitting vehicle under CO emissions. and is given by: The prediction format is a more intuitive presentation for prediction purposes P CO (g/sec) = * FTP Bag2 * antilog{ *AVGSPD *ACC *IPS *ips45sar *ips90tr *tran3idle *tran *finj3sar *cat3tran *sar3tran *flagco} (4-10) The variables from the emission rate models were used as the response variables for the data model presented in this work. Only the emission rate models that related to vehicle activity and apply to the entire fleet were considered. Variables 83

106 that related to subfleet vehicles were not used even if they applied to vehicle activity such as the variable, ips45sar2 which is the percent of activity with IPS >= 45 mph 2 /sec for the subfleet of vehicles with no air injection. This type of variable was not included since data collection could not relate vehicle activity to subfleet characteristics. The model variables indicate that the significant modal predictor variables for carbon monoxide are average speed (AVGSPD), the percent of vehicle activity where acceleration exceeds 3.0 mph/s (ACC.3 ) and percent of activity where the inertia power surrogate is greater than or equal to 60 mph 2 /s (IPS.60). Only AVGSPD and ACC.3 were included as response variables since the final CO model was developed after the data analysis was initiated and IPS.60 could not be incorporated in a timely manner. Three other variables are related to both the vehicle's modal activity and a specific characteristic of the fleet. Percent of time spent idling was significant for 3- way catalyst equipped vehicles (cat3idle). IPS was significant for vehicles with automatic transmissions when IPS was greater than or equal to 90 mph 2 /s (ips90tr1). For vehicles with no excess air injection, IPS >= 45 mph 2 /s (ips45sar2) was a relevant variable. 84

107 To determine the emissions at an intersection with the modal emissions model, the technology group needs to be determined and then the vehicle speed/acceleration profiles estimated. Once the fleet distribution is known for an intersection, changes in operational characteristics will only affect the vehicle activity side so that impacts can be measured by the changes in relevant modal activity. These modal variables were used to predict emission ratios and combined with the corresponding FTP bag2 emission rate from look-up tables are used to derive emission rates, which are then fed into MEASURE HC Model The hydrocarbon emission rate model was derived similar to the CO model, which was described in detail in the above sections. The final emission rate model for HC is (Fomunug, 1999): LogR HC = *my *my *AVGSPD *finj2tran *cat2sar *cat3sar *cat3sar *sar3tran *sar1tran *cid *sar3kml *finj2km *flaghc *acc1finj *acc3cat *ips90sar *dps8finj2 (4-11) Where: my79 = model year < 79; my83 = 79 < model year < 83; 85

108 AVGSPD = average vehicle speed (mph); finj2tran4 = interaction variable for a 4-speed manual transmission vehicle with a carburetor; cat2sar1 = pre 1981 model year vehicle with "oxidation only" catalyst and unknown air injection type; cat3sar1 = pre 1981 model year vehicle with a TWC and unknown air injection type; cat3sar2 = vehicle with TWC and no air injection; sar3tran1 = automatic transmission vehicle with pump air injection; sar1tran5 = pre-1981 model year, 5-speed manual transmission vehicle of unknown air injection type; cid = cubic inches displacement; sar3km1 = vehicle with pump air injection and mileage <=25k miles; finj2km3 = vehicle with pump air injection and 50k < mileage <= 100k miles; flaghc = high emitting vehicle flag under HC emissions; acc1finj2 = carburetor-equipped vehicle operating with acceleration greater than 1 mph/s; acc3cat2 = oxidation only catalyst vehicle with acceleration greater than equal to 3.0 mph/s; ips90sar3 = vehicle with air pump and inertial power surrogate greater than or equal to 90 mph 2 /s; and dps8finj2 = proportion of drag power surrogate (DPS) speed x speed x 86

109 acceleration) greater than 8 mph 3 /s. Average vehicle speed is the single predictor modal activity variable that applies to the entire fleet and is the only variable that will be included in the research model specific to HC Oxides of Nitrogen Model The oxides of nitrogen emission rate model was derived similar to the CO model, which was described in detail in the above sections. The final emission rate model for NO x is (Fomunung, 1999): LogR nox = AVGSPD *IPS *ACC *DEC *finj2km finj2km *cat2km *cat3km *cat3km *finj1km3flagnox *finj3km3flagnox (4-12) Where: IPS.120 = proportion of activity where IPS >= 120 mph 2 /sec; ACC.6 = proportion of activity where acceleration >= 6.0 mph/s; DEC.2 = proportion of deceleration <= -2.0 mph/s; finj2km1 = carburetor equipped vehicle with mileage < 25k miles; finj2km2 = carburetor equipped vehicle with 25K, mileage <= 50k miles; cat2km3 = "oxidation only" catalyst vehicle with 50k < mileage <= 100k 87

110 miles; cat3km2 = TWC vehicle with 25K mileage <= 50k miles; cat3km3 = TWC vehicle with 50K < mileage <= 100k miles; finj1km3flagnox = second order interaction variable for a high emitting vehicle with port fuel injection and 50k < mileage <= 100k miles; and finj3km3flagnox = second order interaction variable for a high emitting vehicle with throttle body fuel injection and 50K < mileage <= 100k miles. The three modal activity predictor variables for the NOx emission rate model are proportion of activity where inertial power surrogate is greater than or equal to 120 mph 2 /sec, proportion of activity where acceleration is greater than or equal to 6.0 mph/s, and proportion of activity where acceleration is less than or equal to 2.0 mph/s Final Response Variables The final model included five response variables which where identified during the emission rate phase of MEASURE. These five variables are presented in Table 4-2. Table 4-2: Modal Predictor Variables for Emission Rate Analysis for Passenger Cars Predictor Description Variable AVGSPD Average speed ACC.3 Proportion of activity where acceleration >= 3.0 mph/s IPS.120 Proportion of activity where IPS >= 120 mph 2 /s ACC.6 Proportion of activity where acceleration >= 6.0 mph/s DEC.2 Proportion of activity where acceleration <= -2.0 mph/s 88

111 4.8 Independent Variables for Vehicle Activity Data Collection The goal of this research was to investigate the influence of different geometric and operational characteristics of signalized roadways that will affect modal frequencies of vehicle activity. It is expected that vehicles behave according to physical constraints of the vehicle, individual driver behavior, characteristics of the surrounding traffic stream, and physical characteristics of the roadway. Various factors may affect both driver behavior and vehicle operation. The Highway Capacity Manual (TRB, 1994), Traffic Engineering Handbook (ITE, 1994), and other resources identify several variables commonly shown to affect traffic operation. Factors may be broken down into categories of driver, vehicle, roadway, traffic, and environmental (ITE, 1994). The same variables that affect traffic operation are also theorized to influence modal activity. A list of these variables is shown in Table 4-3. The end result of the research will be a statistical analysis using the data to derive a relationship between the dependent variable, modal activity, and independent variables that may be used as predictors of modal activity. As a result, the factors that may be relevant independent variables were considered in the data collection process and sites were selected to represent as diverse a group of variables as was feasible. However, many variables could not realistically be collected and were not represented. Although the data collection is presented in Chapter 5, a note is made of whether each variable presented below was collected in the field or calculated if appropriate. 89

112 Table 4-3: Operational and Geometric Factors Hypothesized to Affect Modal Activity Factor Type Collected or Calculated Horizontal Curvature Roadway No Vertical Curvature Roadway only as part of grade Grade Roadway Yes Number of Lanes Roadway Yes Lane Width Roadway Yes Distance Between Intersections Roadway Yes Geographic Location (CBD, Roadway Yes Suburban, etc) Speed Limit Roadway Yes Vehicle Mix Traffic Yes Roadway Capacity Traffic Yes Level of Service Traffic Yes Density Traffic No Vehicle's Lane Position Traffic Yes On-Street Parking Traffic No Queue Position Traffic Yes Weather Conditions Environmental Yes Pavement Conditions Environmental No Type of Vehicle Vehicle Yes Trip Distance Driver No Driver Population Characteristics Driver No Trip Purpose Driver No Driver Variables The following sections describe variables listed in the Highway Capacity Manual (TRB, 1994), Traffic Engineering Handbook (ITE, 1994), or other source which may influence vehicle operation. 90

113 Trip Purpose The purpose for which a driver makes a trip may influence individual vehicle operation. A driver may behave differently for a morning trip to work than a trip to the store. This information could not be realistically collected and is not included as a variable. However, the majority of data collection took place either during the AM or PM peak period. Consequently it is likely that a high percentage of work trips are represented Demographics Drivers may behave differently depending on age, socio-economic status, drug or alcohol use, driving experience, psychological factors, driver familiarity with the roadway, stress, etc. (ITE, 1994). These factors may be significant but could not be collected Vehicle Variables A number of individual vehicle variables will affect vehicle activity profiles. The vehicle's weight, engine size, vehicle make, aerodynamic characteristics, etc. will determine the amount of fuel used and the physical operating constraints of each vehicle. The vehicle class was recorded but others specifics could not be collected given the available labor resources (an additional data collector is necessary for each sampling location to collect license plate) Roadway Variables The physical geometry of the roadway including factors such as horizontal and vertical curvature may affect driver behavior. Drivers slow around sharp horizontal curves and adjust speed to compensate for limited line of sight on both horizontal and vertical curves. Physically, vehicles experience different longitudinal and lateral 91

114 forces on curves than on straight sections and are likely to exhibit different speed profiles than vehicles on a straight stretch of roadway. Additionally, other geometric factors such as grade or lane width may affect driver and vehicle behavior Horizontal and Vertical Curvature Both the horizontal and vertical alignment of a particular roadway segment will have an impact on vehicle operation. Horizontal roadway curvature causes lateral acceleration on the vehicle causing additional load which may translate to either increased engine load and/or changes in frequencies in speed and acceleration as measured on-road. Klaubert and Jongedyk (1985) found that at a mean speed of 73 km/h on a 300-foot radius horizontal curve, road load torque increase by 124 N.m. for the automobiles tested. Given that lateral acceleration may increase engine loading leading to a possible increase in emission rates, horizontal curvature may be an important factor. However, the laser rangefinding equipment used for data collection use geometric equations to calculate the distance from the rangerfinder to the targeted vehicle and depend on a maintaining a constant line of sight. Consequently, accurate vehicle activity can only be collected with the rangefinder and vehicle in the same vertical and horizontal plane. This data collection method does not accommodate significant changes in horizontal alignment, so curvature could not be included as a predictive variable. Vertical curvature also impacts vehicle activity. Vertical curvature per se were not considered in the study for the same reason as horizontal curvature. However, a 92

115 vertical curve is made up of two different grades along a roadway alignment and grade was collected Grade Roadway grade will impact engine load and emissions and may influence on-road vehicle activity. Emission rates specific to grade will be necessary to capture the full effect of grade on emissions. Grades affect the way drivers operate vehicles. Most drivers either increase throttle position to maintain a constant speed or allow the vehicle to decelerate while maintaining the same throttle position (Grant, 1998). In either case, the change in engine activity may or may not be manifested on the road. For example, a driver who increases throttle position will increase engine load but the vehicle will maintain constant speed. The effect of grade may be manifested in both vehicle activity profiles as well as emissions rates. The effect of grade on vehicle operation is described in the Traffic Engineering Handbook (ITE, 1994). The roadway gradient affects the maximum vehicle acceleration achievable and is given by the equation a GV = a LV - Gg/100 (4-13) where: a GV = maximum acceleration at speed V on grade (ft/sec 2 ); a LV = maximum acceleration at speed V in level terrain (ft/sec 2 ); 93

116 G = Gradient (%); and g = acceleration of gravity (32.2 ft/sec 2 ). Not all grades may be significant. Grades in the range of a few percent may be negligible. The Traffic Engineering Handbook (ITE, 1994) states that grades above 3% begin to influence passenger car speeds. The length of the gradient may also affect modal activity. Grade was included as a variable Distance Between Adjacent Intersections Drivers are theorized to behave differently when they expect to stop frequently than when they are driving on long stretches of roadway. The greatest difference in speed and acceleration frequencies between segments with varying distances between signals is expected to occur midblock since higher freeflow speeds can be achieved on longer segments. Differences may also occur at the stopbar if drivers accelerate differently when presented with shorter distances between possible stops. Distance between intersections was collected Number of Lanes The number of lanes along a street segment will affect roadway capacity. Number of lanes may also affect vehicle activity. The presence of multiple lanes may encourage higher speeds. As the number of lanes increases, additional opportunities for conflict between adjacent vehicles may also increase, causing drivers to react and behave differently. A related variable is the vehicle's lane position. Drivers may behave differently depending on whether they are positioned in a lane opposing on-coming traffic, in a lane adjacent to the curb, or 94

117 sandwiched between other lanes in the lane group. The number of lanes was recorded for the study locations Lane Width The width of the traffic lane affects traffic operation. Narrow widths adversely impact capacity. Additionally, drivers may drive more conservatively when narrow lane widths exist. Lane widths data were also collected Speed Limit The posted speed limit may influence the cruise speed a vehicle attains. Consequently, a vehicle's acceleration and deceleration patterns may be influenced by the ultimate speed that the driver is trying to attain. The speed limit may also reflect other characteristics of the roadway such as functional class. However, evidence exists that drivers travel at the speed they consider safe rather obeying posted speed limits. Posted speed limits are easy to obtain for a given segment and were collected for all data collection locations Environmental Factors This section discusses various environmental factors, which are hypothesized to affect modal activity Pavement Condition The condition of the roadway affects speed and acceleration. The coefficient of friction between the tires and roadway affects a vehicle's ability to accelerate and decelerate. The coefficient of friction for dry pavement depends on the type of material used for construction, wear on the pavement, etc. All of the study locations consisted of asphalt roadways. None of the locations exhibited excessive wear. Further classification of pavement condition was not practical for this study. 95

118 Weather Weather conditions also affect both driver behavior and vehicle operation. Drivers usually adjust speed, braking, following distance, etc. when operating on snow, ice, or wet pavement. Drivers may be expected to drive more cautiously during adverse weather. Weather may also affect drivers psychologically. In addition to the effect on the driver, weather conditions may affect the operation of the vehicle. The presence of snow, ice, or water on the roadway reduces the coefficient of friction. A variety of weather conditions could not be represented in the data collection effort. Atlanta, Georgia rarely experiences snow. Additionally, the laser range finders do not operate well with excessive moisture so data were not collected in the rain Other Factors Other factors exist which may influence vehicle or driver behavior and did not fit into any of the preceding classifications. These factors are listed below Pedestrian Activity Pedestrian activity should influence driver behavior. Whether pedestrians are crossing at the intersection itself, walking alongside lanes of traffic, or crossing at non-intersection locations, pedestrians may either physically interfere with traffic operation or influence the way drivers proceed along the roadway segment. However, this variable was not accounted for as part of the data collection effort since the overwhelming majority of the data were collected on arterials in locations where little pedestrian activity was noted. 96

119 Location Along Segment Along a street segment with traffic control, vehicles operations will be influenced by their location along the link in relation to the traffic signal. For example, vehicles midblock are expected to maintain cruise speed while at the intersection they are undergoing acceleration and deceleration. All of this activity is dependent on a vehicle s location along a segment. The majority of extreme modal activity is likely to occur at the intersection stopbar. Once a vehicle reaches cruise speed, they will typically only accelerate or decelerate only due to interactions with surrounding vehicles, to change position, or leave the roadway. While not recorded directly, with a known distance to the stopline and laser output, the location along the link could be calculated Physical Location of Site The geographic location of an intersection may influence traffic activity. The landuse characteristics surrounding a roadway segment may influence both interactions between vehicles and driver behavior. Segments along retail areas are more likely to have a large percentage of vehicles exiting or entering via driveways. This can lead to significant variations in speeds and consequently increased interactions between vehicles. Industrial locations have less driveway activity but may have more heavy vehicles that interfere with the traffic stream. The Central Business District (CBD) is a location classification that is commonly used in traffic engineering applications such as calculation of LOS and may be characterized by a combination of on-street parking, pedestrian activity, and 97

120 delivery trucks. Suburban areas are located away from the city center and are characterized by a minimal number of businesses. Data collection sites were identified given one of the following designations: CBD; Suburban; Commercial; and Industrial. The CBD category is for any sites located in the central business district area of Atlanta. Suburban describes areas outside the CBD where a minimal number of driveways or businesses were located. The commercial category indicates areas where a substantial number of businesses and driveways are located in the study location. Industrial areas were those located in areas with industrialized land uses Queue Position The position of the vehicle in the queue is hypothesized to affect vehicle activity profiles. With free-flow conditions downstream, the first vehicle in the queue is expected to accelerate unrestrained to the desired speed. The second and subsequent queue positions will, to various degrees, have their behavior constrained by the vehicle or vehicles ahead. Vehicle activity may vary between all queue positions. However, at some position in the queue, it is expected that vehicle interactions and behavior will become more uniform (i.e. the tenth vehicle in the queue may behave similar to the ninth vehicle). Queue position was noted. 98

121 4.8.6 Operational Characteristics Operational characteristics such as volume or level of service may influence vehicle activity Level of Service Level of service is characterized by the average stopped delay per vehicle over a 15 minute analysis period (TRB, 1994). Delay is a complex variable based on a number of factors including quality of progression, cycle length, green ratio, and the v/c ratio. Since delay plays a major role in determing LOS, significant delay at an intersection with even low volumes may result in degraded LOS. Hence LOS may only be correlated with vehicle activity if the amount of delay a vehicle experiences influences the way it accelerates or decelerates. Level of Service is an indication of the effectiveness used to describe how well a roadway is functioning. Level of service is easily understood and widely used and was calculated Volume to Capacity Ratio The volume to capacity ratio (v/c) is a measure of capacity sufficiency. It is the ratio of volume or rate of flow to capacity. A v/c ratio greater than 1 indicates that demand is exceeding the computed capacity of the roadway segment (McShane & Roess, 1990). Capacity at signalized intersections is calculated for each lane group and is the maximum rate of flow that can pass through the intersection under prevailing conditions. Factors that affect capacity include: volumes at other approaches; 99

122 turning movement distributions; bus activity; parking; pedestrian activity; number of lanes; grade; signal timing; percent heavy vehicles; and lane width. For signalized intersections, the capacity of a lane group is calculated by: c i = s i (g i /C) (4-14) Where: c i = capacity of lane group i; s i = saturation flow rate for lane group i (vehicles per hour of effective green); g i /C = effective green ratio for lane group i. Ideal saturation flow rate (s o ) is the maximum flow rate capable of passing thru an intersection for a given lane group given constant green and ideal conditions. The saturation flow rate used to calculate capacity reflects an adjustment to the ideal 100

123 saturation flow based on prevailing conditions and is impacted by lane width, percent heavy vehicles, grade, on-street parking, bus activity, area type (CBD versus non- CBD), and distribution of turning movements (TRB, 1994). Volume to capacity may impact vehicle activity. At higher v/c, more interactions between vehicles are expected affecting driving patterns. At lower v/c the number of interactions between vehicles on the roadway are expected to be decline. Volume to capacity was calculated Volume As more vehicles occupy the same amount of roadway, an increased number of interactions will occur. Additionally, a vehicle's ability to achieve and maintain it's desired freeflow will be significantly impacted. Volume is necessary to calculate v/c ratios and is highly correlated. Volume was collected in the field and is described in Chapter Density Density is the number of vehicles occupying a given section of roadway at a particular instant. It is a function of rate of flow and average speed given by: D = v/s (4-15) where: D = density; v = rate of flow; and 101

124 S = average travel speed (mph) (McShane & Roess, 1990). Density may provide a more realistic measure of vehicle interactions than volume to capacity. Volume to capacity at the intersection only indicates the relationship between volume of the roadway and the number of vehicles that can be discharged from the intersection. Because average speed was not collected for the entire intersection, density could not be calculated Fleet Mix Fleet mix is the proportion and types of vehicles occupying a given roadway segment. The vehicle type describes the physical limits of activity that an individual vehicle is capable of achieving. The types of vehicles in the traffic stream will influence the ability of other vehicles to operate. For example, heavy vehicles are physically incapable of achieving the same acceleration rates at a given velocity than passenger vehicles. The percent of heavy vehicles in a traffic stream will therefore affect the speed and acceleration of passenger cars around them. Fleet mix was collected via vehicle counts. 102

125 CHAPTER V 5. DATA PROTOCOLS The primary research goal was to develop representative distributions of vehicle activity at signalized intersections as a function of vehicle attributes, physical roadway characteristics, or roadway operating characteristics. Modal data were collected through empirical measurement of speeds and accelerations of vehicles with laser rangefinders (LRF). Data collection, data preparation, data handling, and attribute integration protocols used in this research are detailed in the following sections. The data collection procedure is described followed by an overview of how attribute data were calculated and matched to field datapoints. A total of 26 locations representing a range of geometric and operation characteristics were studied. Several locations were studied on more than one date. 5.1 Data Collection The data collection methodology is presented in this section, including a technical description of the hardware used. The selection of sampling locations is given in Section 5.1.1, the equipment used for data collection is described in Sections 103

126 5.1.2 and 5.1.3, and data collection protocol and a description of site attributes that were collected are presented in Sections and Selection of Sampling Locations Sampling procedure is a critical component of experimental design. An effort was made to collect as much data as could logistically be collected, reflecting as many of the independent variables as possible. However, time and resource constraints limited the actual amount of data that could be collected in the field. Study sites were chosen based on three criteria. First, candidate locations were selected to represent as many of the independent variables covered in Chapter 4 as possible. The following is a list of the minimum variables that were considered in the site selection process: grade; level of service; volume to capacity; location; distance between signalized intersections; vehicle type; and percent heavy-duty vehicles. Physical constraints of the data collection process served as the second criteria for selecting study locations. Because data collection occurred alongside roadways, sites 104

127 had to be chosen to minimize interference with surrounding objects, such as trees, or adverse geometry, abrupt changes in grade or horizontal alignment. Most data collection locations were selected so that a consistent grade existed throughout the intersection approach. The third consideration in site selection was to minimize influence on interaction with the traffic stream. Areas with sidewalks or wide shoulders were selected so that data collection personnel could be safely located away from the traffic stream. Additionally, locations were chosen and equipment set up so that data collection was as unobtrusive as possible to minimize distraction to drivers or influence driver behavior. Ideally, specific information about each vehicle "tracked", such as model, make, year, engine size, etc., would be recorded and speed-acceleration activity related to individual vehicle characteristics. Initially an attempt was made to record each vehicle's license plate so that information for each vehicle could be extracted from the Georgia vehicle registration database. This approach was abandoned early in the data collection process for several reasons. First, the number of data collection personnel available was limited. It was difficult for the person track the vehicle to also record the license plate. Second, a number of vehicles had no plate, unreadable plates, or were from outside the state. Discarding otherwise "good" datapoints because the license plate wasn t available would have seriously compromised the final 105

128 sample size. Third, even if data could be disaggregated by individual vehicle type, given all the variables that would be part of the statistical analysis, it would become almost impossible to derive relationships on as detailed a level as individual vehicle parameters. Additionally, even if the data could be disaggregated to the level of the individual vehicle, it would be difficult for an agency to provide fleet mix at this level of detail. In many cases, fleet mix is only defined by percent passenger cars and percent heavy trucks rather than by separate technology groups. However, different vehicles do exhibit different operational parameters. Vehicles were divided into several categories including passenger cars, passenger trucks, vans, buses, and heavy trucks. During data collection, each vehicle was assigned a corresponding vehicle category. Passenger trucks include light duty trucks, Jeeps, sport utility vehicles, etc. Vans include minivans and other vehicles reasonably classified as passenger vehicles. Vans larger than normal passenger vehicles, such as transit vehicles, were classified as buses. Passenger cars were designated as passenger vehicle not falling into one of the two preceding categories. Buses included any type of bus or large commercial van. Heavy trucks were designated when possible, with their official classification (2A6, 3AD, etc.) and included all trucks with six or more wheels Advantage Laser Rangefinder Individual vehicle activity profiles were collected in the field using hand-held laser rangefinding devices, also called laser guns. The equipment used were 106

129 Advantage Laser Rangefinders manufactured by Laser Atlanta Optics. The LRF are portable, handheld devices capable of measuring the distance to an object at a high sampling frequency (238.4 distance measurements per second) with a manufacturer s accuracy specification of 0.1 feet (rms) over 2,500 feet (Laser Atlanta, 1997). No minimum effective range for the LRF exists. The maximum effective range is 2,500 feet. Actual range is governed by practical considerations such as the type of vehicle, sight constraints, and interference between the vehicle "tracked" and surrounding vehicles and in all cases the actual range was less than the maximum range of 2,500 feet. Readings from the laser gun can be stored by either outputting the datafile to a computer via serial port interface or by storing data on a SRAM PCMCIA card, which inserts into the rear of the gun. For data collection, SRAM cards were used. Data streamed to the output port are stored in null data files that were created on the card. Each time the LRF trigger is pulled, all subsequent readings are stored to the first available null data file on the SRAM card. Consequently, a unique file is stored on the SRAM card for each vehicle observed (Grant, 1997). For a more in-depth discussion on laser-range finding technology, the reader is referred to Grant, JAMAR Boards JAMAR boards are industry standard data collection devices commonly used in traffic engineering studies. They are used for traffic engineering data collection including volume counts, vehicle classification studies, and intersection turning movement counts. JAMAR boards have the capability of recording up to three directional movements for a total of four intersection approaches. They also have the 107

130 ability to simultaneously bin volume counts into one of three individual bins so that a vehicle mix can be monitored concurrently with volume counts. JAMAR boards were used for turning movement counts for the study sites in question. Turning movement counts were collected as well as classification of heavy vehicles. Using the classification buttons, vehicles were assigned to one of three classes: passenger vehicle : includes passenger cars, light duty trucks, and vans; heavy trucks: defined as any vehicle with more than 4 wheels; or bus: includes buses of all sizes. Turning movement counts were downloaded from the JAMAR boards to a PC using the JAMAR technologies proprietary software, PETRA. From PETRA, counts were output to a text file. Each file contained the time interval and volume counts by movement for each of the three bins. Final output from the JAMAR boards is a tally of vehicle volumes by lane group for each approach in 1-minute intervals. The 1- minute intervals were later aggregated to 15-minute periods Vehicle Attribute Data Concurrent with laser gun data collection, attribute information was recorded for each vehicle "tracked." Since the laser gun did not have a time stamp, the only method to attach one, when using the SRAM cards, was to manually record the time and later attach this to the file. A time stamp was only necessary to match volume, 108

131 LOS, and V/C to the data. The time was recorded manually on the data collection sheet every few minutes and corresponded to individual vehicles. Other attributes recorded for each vehicle, including the type of vehicle, lane the vehicle was occupying, queue position, and the unique number from the LRF, described above were recorded for each vehicle. An example of an attribute sheet is shown in Table Site Attributes General information such as the weather conditions, location, date, etc. were recorded for the study session. Information about each location, including grade, distance to the nearest upstream and downstream intersection, lane width, number of lanes, posted speed limit were recorded for each session. Table 5-1: Example Data Collection Attribute Sheet Jimmy Carter at Live Oak 11-May-97 Weather: hot, sunny Distance to stopline: 104 feet Time Vehicle Type Queue Position Lane P=xxxxx 7:22 Car :24 2A :25 Car Thru :25 Car Thru :28 Van :30 Car

132 5.1.6 Data Collection Protocol The signal timing for each intersection studied was also collected in the field. Signal timing is necessary for calculation of volume to capacity ratios as well as level of service calculations. For optimal data collection, sites were selected and set up to be unobtrusive as possible. Personnel and equipment were located either on the sidewalk or in the right of way, as far away from the traffic stream as feasible without compromising line of sight. LRFs were mounted on tripods to allow for continuous and uninterrupted vehicle tracking. Data collection consisted of the data collector "locking" the laser gun onto a selected vehicle and then following that vehicle until loss of lock occurred. Data collectors attempt to "lock" onto a location on the vehicle, such as the license plate, and then maintain lock on that position. Data for each vehicle are downloaded from the LRF and stored as a unique file on a data card. 5.2 Data Handling Once data were collected in the field, they were later downloaded from the data storage cards (SRAM cards) and reduced to usable datasets. A flowchart detailing the data collection and reduction procedure is provided in Figure 5-2. A description of the data reduction process follows. 110

133 Figure 5-1: Data Collection and Reduction Methodology 111

134 5.2.1 Laser Rangefinder During each data collection session, data were stored on 2 MB SRAM cards. The LRF output for each observation was streamed at the rate of readings per second to the next available null file on the card. Each card has the ability to hold the lesser of either 100 files or 2 megabytes of data. At the end of the session, data were downloaded to a PC via a MS-DOS batch file (CAPTURE.C), which copies each file off the PCMCIA card and then systematically creates null files on the PCMCIA card. Downloaded data files were named following the naming convention DATA.000, DATA.001,, DATA.099. As a result, data files from each data collection session had the same file names, so once data were downloaded they were zipped and stored under a specific zip filename and directory. Once a vehicle has been "tracked" and the trigger released, the LRF displays a unique number in the visual display (P=xxxxx). Each file will also occupy the number of bytes corresponding to the recorded P=xxxx value. For example, if the value P=8891 was observed, once the file was downloaded it would occupy 8,891 bytes. Because it is statistically unlikely that any two records contain the same number of bytes, given the LRF sampling frequency, this number was unique. The number was recorded and later correlated with other manually recorded data collected for the vehicle such as type of vehicle or queue position. 112

135 5.2.2 RANGE.C Program After data were downloaded and stored in unique directories to prevent overwriting data from one session with another, a program written in C language was used to calculate speed and acceleration from the distance information. RANGE.C was written by Chris Grant of Georgia Tech as part of his dissertation work (Grant, 1998). The program expects datafiles as input with the naming convention DATA.xxx in order, starting with DATA.000, which are located in the same directory. Each datafile is read consecutively until the last data file in the directory is reached. For each directory, RANGE70.C reads each datafile and then records results to a single output file for the directory. For each second of data, time, distance from the laser gun, speed, and acceleration followed by the vehicle number are reported. RANGE70.C calculates speed and acceleration using a smoothing algorithm. It also attempts to throw out erroneous readings. Table 5-2 provides an example of RANGE70.C output. Additionally, RANGE70.C requires the offset distance between the LRF and the data collection as input and with this value, calculates actual Euclidean distances between successive movements of the vehicle. Because data collection takes place at the side of the road, the laser gun is not able to take a straightline reading to the vehicle. The readings actually report the hypotenuse of the distance between the vehicle and the LRF. RANGE70.C accounts for this and computes the straight-line distance to the vehicle for speed and acceleration calculations and distance output. This is illustrated in Figure 5-2. Because different offsets will affect 113

136 final output, each data file was run using the distance from the laser gun to the center of each lane where data collection took place. For example, if data collection occurred six feet from the edge of pavement and two twelve foot lanes were sampled, RANGE70.C would have been run twice. First an offset distance to the center of the first lane of 12 feet (6 + 12/2) would be used. Next, RANGE would be rerun using an offset of 24 feet for the distance to the center of the second lane ( /2). Often during the data collection process, other vehicles interfered with "tracking" a targeted vehicle. This usually resulted in no data output for the vehicle in question or a series of speeds and accelerations near 0 as output. An example of this is found in the data output for vehicle 3 in Table 5-2. These data were manually removed so that "bad" data did not skew analytical results. Additionally, each vehicle in queue was tracked from rest if possible so that several to many seconds of idling were recorded. Idling time was removed from the record sets because total delay could not be captured for each vehicle and was not the subject of this research. Delay can be calculated using a number of programs including the Highway Capacity Software (HCS) and can be represented as seconds of activity in zero acceleration and zero speed. 114

137 115 Figure 5-2: LRF Geometry Accounted for in RANGE70.C 115

138 Table 5-2: Example Output from RANGE Laser Offset : 12.0 Time=17.65, Dist= 33.9, Speed= 0.0, Accel=-0.1 Time=18.66, Dist= 37.0, Speed= 2.1, Accel= 2.0 Time=19.66, Dist= 43.2, Speed= 4.2, Accel= 2.1 Time=20.67, Dist= 54.1, Speed= 7.4, Accel= 3.1 Time=21.68, Dist= 69.5, Speed= 10.5, Accel= 3.1 Time=22.68, Dist= 89.4, Speed= 13.5, Accel= 3.0 Time=23.69, Dist= 110.4, Speed= 14.2, Accel= 0.7 Time=24.70, Dist= 131.3, Speed= 14.1, Accel=-0.1 Time=25.70, Dist= 152.9, Speed= 14.6, Accel= 0.5 Time=26.71, Dist= 177.5, Speed= 16.7, Accel= 2.0 Time=27.72, Dist= 205.1, Speed= 18.7, Accel= 2.0 Time=28.72, Dist= 234.9, Speed= 20.2, Accel= 1.5 Time=29.73, Dist= 267.2, Speed= 21.8, Accel= 1.6 Time=30.74, Dist= 301.0, Speed= 22.9, Accel= 1.1 Time=31.74, Dist= 335.0, Speed= 23.0, Accel= 0.1 # 1 Vehicle Time= 4.56, Dist= 292.7, Speed= 28.1, Accel=-2.5 Time= 5.57, Dist= 330.3, Speed= 25.5, Accel=-2.6 Time= 6.58, Dist= 366.0, Speed= 24.2, Accel=-1.3 Time= 7.58, Dist= 400.8, Speed= 23.6, Accel=-0.6 Time= 8.59, Dist= 436.2, Speed= 24.0, Accel= 0.4 Time= 9.60, Dist= 473.1, Speed= 25.0, Accel= 1.0 Time=10.60, Dist= 510.4, Speed= 25.2, Accel= 0.2 # 2 Vehicle Time= 2.55, Dist= 268.1, Speed= 0, Accel=0 Time= 3.56, Dist= 311.2, Speed= 0, Accel=0 Time= 4.56, Dist= 311.2, Speed= 0, Accel=0 Time= 5.56, Dist= 311.2, Speed= 0, Accel=0 Time= 6.56, Dist= 311.2, Speed= 0, Accel=0 # 3 Vehicle Time=13.62, Dist= 26.1, Speed= -0.0, Accel=-0.0 Time=14.63, Dist= 28.1, Speed= 1.4, Accel= 1.4 Time=15.64, Dist= 33.0, Speed= 3.4, Accel= 2.0 Time=16.64, Dist= 43.3, Speed= 6.9, Accel= 3.6 Time=17.65, Dist= 59.1, Speed= 10.7, Accel= 3.7 # 4 Vehicle 116

139 5.2.3 ATTACH.C Attribute data for each vehicle were manually collected during the data collection process as described earlier and later matched with output from the laser rangefinders so that observations of modal activity for individual vehicles could be sorted by lane, queue position, vehicle type, etc. Attributes including a vehicle identification number, data collection time, queue position, lane, grade, distance to upstream and downstream intersections, and speed limit were recorded for each vehicle and manually entered in a spreadsheet after data collection. Data were later exported to a column delimited text file. A program written in C, ATTACH.C, was used to match attribute output with actual vehicle data output from RANGE70. ATTACH.C outputs comma delimited data that can be easily be imported to a spreadsheet or database file. A final dataset for each card used during each data collection session was created. An example of a final dataset is shown in Table 5-3. The following sections describe the attribute data Stopline Distance Using the known distance from the where the LRF was positioned to the intersection stopline and the location output as part of RANGE.c, the vehicle's instantaneous location from the intersection stopline was calculated and attached as an attribute to each record which represented on second of vehicle activity. 117

140 Table 5-3: Final Dataset Format Deering ID Time Type Queue Lane Stop VC LOS Speed Accel UP Down % FILE Up Down Gr ade Speed Lane s Lane Locatioition Cond- Dist (mph) (mph/s) Vol Vol HV Dist Dist. Limit Width 28 10:10 2A A DEER IND DRY 28 10:10 2A A DEER IND DRY 28 10:10 2A A DEER IND DRY 28 10:10 2A A DEER IND DRY 28 10:10 2A A DEER IND DRY 28 10:10 2A A DEER IND DRY 28 10:10 2A A DEER IND DRY 28 10:10 2A A DEER IND DRY 28 10:10 2A A DEER IND DRY 28 10:10 2A A DEER IND DRY 28 10:10 2A A DEER IND DRY 28 10:10 2A A DEER IND DRY 28 10:10 2A A DEER IND DRY 28 10:10 2A A DEER IND DRY 28 10:10 2A A DEER IND DRY 28 10:10 2A A DEER IND DRY 118

141 5.2.5 Volume Calculations Volume counts for each study location were taken in one-minute intervals using JAMAR boards. Counts were output from the JAMAR boards using JAMAR Technologies software, PETRA and turning movement volumes calculated by 15- minute intervals for the duration of the study period by vehicle type. Final volumes reflect merging of passenger car, heavy truck, and bus bins for each interval. Upstream volumes were determined by adding all turning movement volumes for the upstream approach for the 15-minute interval. Downstream volumes were calculated by summing the through movements for the study approach plus the volume of vehicles from other approaches turning either left or right into the downstream link of the approach Percent Heavy Vehicles Calculations Heavy vehicle percents were calculated using the following equation: P hv = (H + B)/(H+B+C) (5-1) Where: Phv = percent heavy vehicles for the 15 minute period; H = total number of heavy trucks for the 15 minute period; B = total number of buses for the 15 minute period; C = total number of passenger cars for the 15 minute period. 119

142 5.2.7 LOS and V/C Ratio Later V/C and level of service were calculated and manually added to the spreadsheet files. V/C and LOS were calculated for 15-minute intervals using the Highway Capacity Software and were manually related to the data by the corresponding time period. Once data have been reduced and attributes attached, data can be binned by desired groupings such as activity on specific grade or under a particular LOS. Data can be sorted by location along a link so that critical locations for modal activity and possible enrichment are identified. 5.3 DATA COLLECTION SITES A total of 26 locations were studied in the Atlanta, Georgia metropolitan area resulting in a total of 95 datafiles. Each datafile represents data collected on a single card during the data collection process. Each datafile represents between 200 and 500 seconds of vehicle activity. A summary of data collection activity is shown in Table 5-4. Final datasets represent LOS ranging from A to F with levels A, B, and C being the most represented. V/c ranges from 0.2 to 1.2. Per lane volumes vary from a minimum of 143 to a maximum of The following grades are represented: -9%, -8%, -5%, -4%, -3%, -2%, -1%, 0, +1%, +2%, +3%, +4%, +%5, 8% and +9%. From 120

143 2 to 5 lanes were represented. A two or three lane roadway was the most common configuration with lane widths varying from 9 feet to 12 feet. Posted speed limits were 30, 35, 40, and 45 mph. Downstream distances varied from 756 to 4,118 feet and upstream distances varied from 300 to 5,544 feet. The most sampled queue positions were the first or second in the queue or a "thru" vehicle since they were the easiest to sample and were also the most common. However, this did not skew test results since vehicles with different queue positions were analyzed separately. 121

144

145 Table 5-4: Data Collection Sites 122 Date Day Filename Intersection Location Starting Approach Lanes Lane Grade Speed Upstream Downstream Time Width Limit Distance Distance 25-Oct-96 Friday Chrisatt 10th & West Peachtree Midblock 3:45 PM WB % Oct-96 Friday Dharmnet 10th & Acceleration 3:45 PM NB % Peachtree 7-Nov-96 Thurs Nov7cd1 Peachtree & Acceleration 4:30 PM NB % th 7-Nov-96 Thurs Nov7cd2 Peachtree & 10 th Acceleration 4:30 PM NB % Feb-97 Friday Will1 Jimmy Carter Deceleration 8:33 AM WB % & Williams 21-Feb-97 Friday Will2 Jimmy Carter Deceleration 9:18 AM WB % & Williams 21-Feb-97 Friday Will3 Jimmy Carter Acceleration 8:37 AM WB % & Williams 21-Feb-97 Friday Will4 Jimmy Carter & Williams Acceleration 9:23 AM WB % Mar-97 Friday Rockbd1 Rockbridge midblock 9:05 AM WB % Mar-97 Friday Rockbd2 Rockbridge midblock 9:18 AM WB % Mar-97 Friday Rockbd3 Rockbridge acceleration 8:25 AM WB % Mar-97 Thurs Deer1cd1 Northside & Deering 20-Mar-97 Thurs Deer1cd2 Northside & Deering 20-Mar-97 Thurs Deer1cd3 Northside & Deering 20-Mar-97 Thurs Deer1wht Northside & Deering 21-Mar-97 Friday Deer1cd1 Northside & Deering 21-Mar-97 Friday Deercd2 Northside & Deering 4-Apr-97 Friday Deer3cd1 Northside & Deering acceleration 5:03 PM NB % acceleration 5:51 PM NB % deceleration 5:03 PM NB % deceleration 5:41 PM NB % acceleration 8:15 AM SB % acceleration 9:01 AM SB % acceleration 8:13 AM SB %

146 Table 5-4: Data Collection Sites (Cont.) 123 Date Day Filename Intersection Location Start Time Approach Lanes Lane Grade Speed Upstream Downstream Width Limit Distance Distance 4-Apr-97 Friday Deer3cd3 Northside & deceleration 8:09 AM SB % Deering 4-Apr-97 Friday Deer3wht Northside & deceleration 8:40 AM SB % Deering 9-Apr-97 Wed Deer4cd1 Northside & Deering acceleration 5:17 PM NB % Apr-97 Wed Deer5cd1 Northside & acceleration 5:04 PM NB % Deering 16-Apr-97 Wed Deer5cd2 Northside & acceleration 5:57 PM NB % Deering 16-Apr-97 Wed Deer5cd3 Northside & Deering deceleration 5:05 PM NB % Apr-97 Monday Wp&15wht West Peachtree acceleration 8:25 AM NB 5 9-3% & 15th 21-Apr-97 Monday Wp&15cd1 West Peachtree acceleration 9:15 AM NB 5 9-3% & 15th 22-Apr-97 Tuesday Ev&jccd1 Everest & deceleration 8:43 AM WB % Jimmy Carter 22-Apr-97 Tuesday Ev&jccd3 Everest & acceleration 8:49 AM WB % Jimmy Carter 22-Apr-97 Tuesday Ev&jcwht Everest & Jimmy Carter Acceleration 9:30 AM WB % May-97 Thurs Ndruid2a North Druid Hills & LaVista Deceleration 8:23 AM WB May-97 Thurs Ndruid2b North Druid Deceleration 8:56 AM WB Hills & LaVista 1-May-97 Thurs Ndruidwht North Druid Hills & LaVista acceleration 8:16 AM WB May-97 Friday Piedcd2 Piedmont acceleration 8:00 AM EB May-97 Friday Card3out Piedmont deceleration 7:58 AM EB May-97 Friday Whiteout Piedmont deceleration 8:43 AM EB May-97 Friday 16-May-97 Friday Nt&10wh1 Northside & 10 th midblock 8:13 AM SB % Nt&10wh2 Northside & midblock 8:53 AM SB % th 124

147 Table 5-4: Data Collection Sites (Cont.) 124 Date Day Filename Intersection Location Time Approach Lanes Lane Width Grade Speed Limit Upstream Distance Downstream Distance 16-May-97 Friday nt&10cd3 Northside & 10 th 16-May-97 Friday nt&10cd2 Northside & 10 th acceleration 8:07 AM SB % May-97 Wed Mari1cd2 Marietta & acceleration 4:53 PM WB % Chatahochee 21-May-97 Wed Mari1cd3 Marietta & Chatahochee acceleration 5:24 PM WB % May-97 Friday Mari2cd2 Marietta & acceleration 7:47 AM EB % Chatahochee 23-May-97 Friday Mari2wht Marietta & Chatahochee acceleration 8:21 AM EB % May-97 Wed Mari3cd1 Marietta & deceleration 5:35 PM WB % Chatahochee 28-May-97 Wed Mari3wht Marietta & Chatahochee deceleration 6:17 PM WB % Jun-97 Friday Pc&lkcda Peachtree & deceleration 8:03 AM SB 3 9-3% Lakeview 6-Jun-97 Friday Pc&lkcdb Peachtree & Lakeview deceleration 8:45 AM SB 3 9-3% Jul-97 Wed Pc&jowht Peachtree acceleration 5:35 PM NB % Industrial & Johnson Ferry 9-Jul-97 Wed Pc&jocd2 Peachtree both 5:40 PM NB % Industrial & Johnson Ferry 15-Aug-97 Wed 1hwy29bk Hwy 29 & deceleration 8:52 AM SB % North Druid Hills 15-Aug-97 Wed 1hwy29wh Hwy 29 & North Druid Hills deceleration 8:08 AM SB % Aug-97 Friday Sp&16cd1 Spring &16th deceleration 8:38 AM SB %

148 Table 5-4: Data Collection Sites (Cont.) 125 Date Day Filename Intersection Location Time Approach Lanes Lane Grade Speed Downstream Upstream Width Limit Distance Distance 29-Aug-97 Friday Sp&16cd2 Spring &16th acceleration 9:16 AM SB % Aug-97 Friday Sp&16wht Spring &16th deceleration 8:00 AM SB % Sep-97 Friday Carrolc2 Marietta & deceleration 9:18 AM EB % Carrol 5-Sep-97 Friday Carrolc1 Marietta & deceleration 8:43 AM EB % Carrol 5-Sep-97 Friday Carrolwh Marietta & Carrol deceleration 8:05 AM EB % Sep-97 Friday Pch&key1 Peachtree acceleration 8:55 AM SB % Industrial & Cross Key 12-Sep-97 Friday Pch&key2 Peachtree acceleration 9:36 AM SB % Industrial & Cross Key 12-Sep-97 Friday Pch&keyw Peachtree Industrial & Cross Key acceleration 8:20 AM SB % Oct-97 Wed Pc&ky2c1 Peachtree & acceleration 9:12 AM SB % Cross Key 1-Oct-97 Wed Pc&ky2c2 Peachtree & acceleration 8:09 AM SB % Cross Key 1-Oct-97 Wed Pc&key2w h Peachtree & Cross Key acceleration 8:45 AM SB % Oct-97 Monday Phill1cd1 Pleasant Hill & deceleration 7:34 AM EB % Satellite 6-Oct-97 Monday Phill1cd2 Pleasant Hill & deceleration 8:15 AM EB % Satellite 6-Oct-97 Monday Phill1wht Pleasant Hill & acceleration 9:00 AM EB % Satellite 13-Oct-97 Monday Phill2cd2 Pleasant Hill & both 8:09 AM EB % Satellite 126

149 Table 5-4: Data Collection Sites (Cont.) 126 Date Day Filename Intersection Location Time Approach Lanes Lane Grade Speed Upstream Downstream Width Limit Distance Distance 13-Oct-97 Monday Phill2cd1 Pleasant Hill & both 7:31 AM EB % Satellite 13-Oct-97 Monday Phill2wht Pleasant Hill & both 8:44 AM EB % Satellite 27-Oct-97 Monday 2hwy29c1 Highway 29 & deceleration 7:40 AM SB % North Druid Hills 27-Oct-97 Monday 2hwy29c2 Highway 29 & deceleration 8:13 AM SB % North Druid Hills 3-Nov-97 Monday 2sp&16c2 Spring &16th both 8:50 AM SB % Nov-97 Monday 2sp&16c3 Spring &16th both 8:12 AM SB % Nov-97 Monday 2sp&16wh Spring &16th both 7:29 AM SB % Dec-97 Wed Mr-mdwb Marietta & midblock 11:43 AM WB % Chatahochee 3-Dec-97 Wed Mari-mid Marietta & Chatahochee midblock 12:26 PM EB % Dec-97 Friday Law-mid Lawrenceville midblock 11:25 AM EB % Hwy 19-Dec-97 Friday Ind-mdc1 Indian Trails midblock 1:38 PM WB % by I_85 19-Dec-97 Friday Ind-mdc2 Indian Trails midblock 1:14 PM WB % by I_85 19-Dec-97 Friday Ind-mdc3 Indian Trails by I_85 midblock 12:46 PM WB % Apr-98 Monday Ch420c1 Marietta & both 7:58 AM EB % Chatahochee 20-Apr-98 Monday Ch420c2 Marietta & deceleration 7:14 AM EB % Chatahochee 11-May-98 Monday j&ok1cd1 Jimmy Carter & Live Oak acceleration 7:15 AM WB % May-98 Monday j&ok1cd2 Jimmy Carter & Live Oak acceleration 7:50 AM WB %

150 Table 5-4: Data Collection Sites (Cont.) 127 Date Day Filename Intersection Location Time Approach Lanes Lane Grade Speed Upstream Downstream Width Limit Distance Distance 11-May-98 Monday j&ok1wht Jimmy Carter & Live Oak acceleration 8:27 AM W % Jun-98 Thursd d0604cd1 Northside & acceleration 11:04 AM SB % ay Deering 4-Jun-98 Thursd ay d0604cd2 Northside & Deering acceleration 12:10 PM SB % Jun-98 Monday Deer0608 Northside & Deering acceleration 10:09 AM SB % Jun-98 Monday j&ok2cd1 Jimmy Carter both 7:36 AM WB % & Live Oak 8-Jun-98 Monday j&ok2cd2 Jimmy Carter both 8:09 AM WB % & Live Oak 8-Jun-98 Monday j&ok2wht Jimmy Carter both 8:45 AM WB % & Live Oak 22-Jun-98 Monday j&ok3cd2 Jimmy Carter both 7:34 AM WB % & Live Oak 22-Jun-98 Monday j&ok3wht Jimmy Carter & Live Oak both 8:09 AM WB % May-97 Wed e&jc2c2 Everest & acceleration 4:59 PM EB % Jimmy Carter 7-May-97 Wed e&jc2c3 Everest & Jimmy Carter acceleration 5:44 PM EB %

151 CHAPTER VI 6. PRESENTATION OF DATA This chapter presents the data analysis segment for this research work. A total of 26 locations in the Atlanta, Georgia metropolitan area were sampled in the data collection process. For the 26 sites, at total of 95 data files were downloaded from the PCMCIA cards. Over all sites surveyed, a total of 4,097 passenger vehicles and 326 heavy vehicles were sampled. A total of 26,941 seconds of passenger vehicle activity and 3,830 seconds of heavy vehicle activity were recorded and were available for data analysis. 6.1 Data Preparation After data processing as described in Chapter 5, a unique spreadsheet was created for each datafile (one per SRAM for each site). The spreadsheet contained a second by second profile for each vehicle with the related attributes such as queue position or grade. Prior to statistical analysis, each spreadsheet was converted to a common file format readable by S-PLUS statistical software. Data were separated into two vehicle type categories. The "passenger vehicle" designation included vehicles such as cars, light duty trucks with 4 wheels, passenger 128

152 vans, and sport utility vehicles. "Heavy vehicles" included trucks with 6 or more wheels. Various vehicle activity profiles were observed for buses. However, the MEASURE model does not currently include parameters for buses so this data were not used in the analysis. Once data were separated by vehicle type, they were disaggregated into two hundred-foot incremental distances according to a vehicle s position from the point of queuing. Data were disaggregated in this manner for two reasons. First, when collecting data it was almost impossible to sample a complete vehicle trace as was shown in Figure 4-1. A complete vehicle trace would follow a vehicle from rest to a distance downstream, such as 1000 feet, without interruption. In reality complete vehicle traces could not be collected because of interference in tracking the vehicle. Interference came from a number of sources such as surrounding vehicles, vegetation, the tracked vehicle changing lanes, etc. Consequently, by analyzing data in specific segments, incomplete vehicle traces can be used without compromising the integrity of the data. Second, it was expected that activity would differ by location from the intersection stopline. For example, vehicles behave differently when starting from rest at the intersection than when they are cruising midblock. However, at some point along a link it is expected that activity will become homogenous and can be grouped. Additionally, although activity for the first and tenth vehicle in a queue is 129

153 expected to be dissimilar as the vehicles accelerate off the stopline, at some point downstream, both vehicles should have reached their cruising speed and may have similar vehicle profiles. Partitioning the data allows locations where the data act more similarly to be identified. Data were partitioned according to the grouping conventions listed in Table 6-1 and an example of this disaggregation is shown in Figure 6-1. To determine the fraction of activity in each response variable category, each of the 95 datafiles were prepared according to the following criteria: 1) Data were separated by queue position, level of service, volume to capacity, percent heavy vehicles, upstream per lane volume, and downstream per lane volume. In most cases upstream and downstream distances, grade, number of lanes, width, etc. were consistent for the entire data collection site. 2) Data were then divided by 200-foot increments from the queuing point according to the convention described above. 3) Seconds of activity for each response variable were calculated (i.e. seconds of activity for the group where acceleration >= 6 mph/s). 4) Disaggregated data were summed by total seconds of activity and total seconds of activity in each response category were calculated. 5) Percent of activity in each response category, such as % activity with acceleration >= 3.0 mph/s, was calculated by: 130

154 % Activity = Seconds of response activity total seconds of activity. Table 6-1: Data Partitioning Name ACCEL ACCELPLUS200 ACCELPLUS400 ACCELPLUS600 ACCELPLUS800 ACCELPLUS1000 THRU DECEL DECELNEG200 DECELNEG400 DECELNEG600 DECELNEG800 DECELNEG1000 Description Vehicle activity from the stopping point downstream 200 feet for vehicles stopped by the traffic signal. Activity from 200 to 400 feet downstream of vehicle's initial stopping point Activity from 400 to 600 feet downstream downstream of vehicle's initial stopping point Activity from 600 to 800 feet downstream downstream of vehicle's initial stopping point Activity from 800 to 1000 feet downstream downstream of vehicle's initial stopping point Activity from 1000 to 1200 feet downstream downstream of vehicle's initial stopping point Activity for vehicles not stopped by the traffic signal and vehicles captured during "midblock" data collection. Data were divided by 200 foot increments before and after the intersection stopbar Vehicle activity from 200 feet upstream of the vehicle's stopping point to the stopping point for vehicles stopped by the traffic signal. Vehicle activity from 400 to 200 feet upstream of the vehicle's stopping point. Vehicle activity from 600 to 400 feet upstream of the vehicle's stopping point. Vehicle activity from 800 to 600 feet upstream of the vehicle's stopping point. Vehicle activity from 1000 to 800 feet upstream of the vehicle's stopping point. Vehicle activity from 1200 to 1000 feet upstream of the vehicle's stopping point. 131

155 132

156 132 Figure 6-1: Schematic of Distance Partions 133

157 6.2 Data Analysis Data were analyzed using hierachachial tree based regression analysis and then validated using the Kolmorgorov-Smirnov two sample test in S-PLUS statistical software version 4.5 from Mathsoft (Mathsoft, 1997). This analysis technique generates a "tree" structure by dividing the sample data recursively into a number of groups. The groups are selected to maximize some measure of difference in the response variable in the resulting groups. One of the advantages of regression tree analysis over traditional regression analysis is that it is a non-parametric method, which by definition does not require any distribution assumptions and is more resistant to the effects of outliers (Roberts, 1999). In growing a regression tree, the binary partitioning algorithm recursively splits the data in each node until the node is homogenous or the node contains too few observations. If left unconstrained, a regression tree model can "grow" until it results in a complex model with a single observation at each terminal node that explains all the deviance. However, for application purposes, it is desirable to create an end product that balances the model's ability to explain the maximum amount of deviation with a simpler model that is easy to interpret and apply. The software allows the user to interact with the data in the following manner to select variables and help simplify the final model: 134

158 Response variable: the response variable is selected by the user from a list of fields from the data set; Predictor variables: one or more independent variables can be selected by the user from a list of fields associated with the dataset; Minimum number of observations allowed in a single split: sets the minimum number of observations that must be present before a split is allowed (default is 5); Minimum node size: sets the allowed sample size at each node (default is 10); Minimum node deviance: the deviance allowed at each node (default is 0.01). Tree size is not limited and the resulting model may be more complex than necessary. To simplify the model, several methods can be used. First, the minimum number of observations, minimum node size, and minimum node deviance can be increased or decreased either singly or in combination. Three other functions can be used to simplify the tree without sacrificing goodness-of-fit. Pruning reduces the nodes on a tree by successively snipping off the least important splits. The importance of a subtree is measured by a cost-complexity measure defined by: D k (T') = D(T') + k. size(t') (6-1) where: D k (T') = deviance of the subtree T'; 135

159 k = cost-complexity parameter; and size(t') = number of terminal nodes of T'. Cost complexity pruning determines the subtree T' that minimizes D k (T') over all subtrees. The larger the value for k, the fewer subnodes that will result (Mathsoft, 1997). The second function that can be used to simplify the model is shrinking. Shrinking reduces that number of effective nodes by shrinking the fitted value of each node towards its parent node. The shrunken fitted values are computed according the following algorithm: y(node) = k.? (node) + (1 - k). y(parent) (6-2) where: k = shrinking parameter, may be either a scalar of vector (0<k<1);? (node) = the usual fitted value for a node; and y(parent) = the shrunken fitted value for the node's parent. Snipping (snip.tree) allows the user to interactively remove nodes and try various modifications to the original model. Implications of using any of the procedures (prune, shrink, snip, modifying minimum number of observations, modifying minimum node size, or modifying minimum node deviance) can be evaluated by observing normal probability plots of the residuals for the tree object, comparing residual mean deviance for different models, 136

160 or inspecting a plot of the reduction in deviance with the addition of nodes. Breiman et al. (1984) indicate that too large a tree will have a higher true misclassification rate than the right sized tree, while too small a tree will not use some of the classification information available. Breiman et al. (1984) suggest starting with an initial large tree model and then pruning back to the right root node. The residual mean deviance (RMD) is an indicator of regression tree "fit". It is the mean deviance of the data samples in the terminal nodes of an estimated tree model. RMD is calculated by summing of the deviance of all the data samples for all the terminal nodes. The summed deviance is then divided by the number of terminal nodes. A lower value for RMD indicates a "better" fit (Roberts, 1999). Under a normal (Gaussian) assumption, terms in the residual mean deviance are the squared differences between the observations and the predicted values (Mathsoft, 1997). A plot of the model in S-PLUS may also be used to estimate the relative importance of splits on a particular variable. When using the parameter of non-uniform spacing to plot the regression tree model, the software plots the tree legs in approximation to the importance of the split. Consequently, longer tree legs indicate that the variable explained more variation than a shorter tree leg (Mathsoft, 1997). 137

161 6.2.1 Identification of Microscopic Activity Distribution Dependent Variables As discussed in Section 4.7, various microscopic activity variables have been identified which may be highly relevant to emission producing activity. An in-depth overview of the dependent variables was provided in that section. In short the following dependent or response variables used in the statistical modeling are: 1) Acceleration >= 3.0 mph/s (ACC.3): The proportion of activity for the segment where instantaneous acceleration rates are greater than or equal to 3.0 mph/s. 2) Acceleration >= 6.0 mph/s (ACC.6): The proportion of activity for the segment where instantaneous acceleration rates are greater than or equal to 6.0 mph/s. 3) Deceleration <= -2.0 mph/s (DEC.2): The proportion of activity for the segment where instantaneous acceleration rates are less than or equal to -2.0 mph/s. 4) Average Speed (AVGSPD): The average speed for the segment (mph). 5) Inertial Power Surrogate >= mph 2 /s (IPS120): The proportion of activity for the segment where inertial power surrogate (approximated by the product of velocity and acceleration) is greater than or equal to mph 2 /s Identification of Microscopic Activity Distribution Independent Variables 138

162 Section 4.8 detailed 21 variables that were hypothesized to influence microscopic activity. There may be additional variables, which were not considered and may have contributed to model error. It is theorized that the single most relevant variable in predicting microscopic vehicle activity is driver behavior. Variables that can serve as surrogate variables for individual driver characteristics include trip purpose (work commute, shopping, recreation, education) and driver characteristics (age, income, occupation, etc.). Unfortunately, with the type of study performed, it was impossible to collect any of the variables related to individual drivers. Of the 21 original variables considered, the final data model was only able to realistically include 13 variables. A more in-depth discussion of how each variable was calculated is provided in Chapter 5. Following is the final list of the predictor variables used in the statistical analysis with their designation name in the database in parentheses: Vehicle queue position (QUEUE): represents the position of the vehicle in queue at the stopbar of the intersection (1, 2, 3, 4.). Vehicles not stopped at the intersection were designated as THRU vehicles. Volume to capacity (VC): the volume to capacity ratio for the segment calculated using HCS in 15-minute intervals. Level of Service (LOS): level of service for the segment, calculated using HCS in 15- minute intervals. 139

163 Number of lanes (NO_LANES): the number of lanes in the direction of travel for the segment studied. Lane width (WIDTH): the average lane width for the segment. Upstream volume (UPSTREAM): volume by 15-minute intervals for the upstream segment of the intersection studied adjusted to hourly volume Volume was divided by the number of lanes yielding per lane volume. Downstream volume (DOWNSTREAM): volume by 15-minute intervals for the downstream segment of the intersection studied adjusted to hourly volume. Volume was divided by the number of lanes yielding per lane volume. Upstream distance (UPDIST): distance from the intersection studied to the nearest upstream signalized intersection. Downstream distance (DOWNDIST): distance from the intersection studied to the nearest downstream-signalized intersection. Grade (GRADE): grade for the segment. Percent heavy vehicles (PER_HV): percent heavy vehicles for the link. Speed limit (SPEEDLIMIT): the posted speed limit for the link studied. Location (LOCATION): a categorical variable that indicates the most typical land use surrounding the link being studied. The designations include a) Industrial, b) Commercial, c) Suburban, and d) Central Business District (CBD). 140

164 Link length (LINKDIST): this variable was used for "thru" vehicles and was the length of the street segment from signalized intersection to signalized intersection where data collection took place. This was used in place of upstream link distance and downstream link distance since it was difficult to interpret what upstream and downstream distances were for thru vehicles since they occupied positions both before and after the intersection stopbar. Link volume (VOLUME): this variable was the per lane volume of the street segment where data collection took place and was used for "thru" vehicles only instead of upstream and downstream volumes. Several of the variables, which were considered and collected, were not included in the final statistical analysis. Density was identified as a variable that may be influential in affecting vehicle activity. Although, it is relatively easy to calculate, it requires the average speed for the segment. Speeds were available by queue categories, such as average speed for "THRU" vehicles for the segment or average speed for the first vehicle in the queue. However, an average speed representative of all activity on the segment could not be calculated. Consequently, density was not included. Pavement condition (wet, dry, icy) was also dropped as an independent variable. As discussed previously, most data collection took place under dry pavement conditions. 141

165 Before proceeding with the statistical analysis, the various predictor variables were investigated to determine whether they were correlated. Correlation between variables may result in false partitioning of data. Correlation was found to exist between the following variables: LOS and upstream per lane volume; LOS and downstream per lane volume; Volume to capacity and upstream per lane volume; Volume to capacity and downstream per lane volume; and Volume to capacity and LOS. Figure 6-2 demonstrates the degree of correlation between volume to capacity and upstream per lane volume. Correlation indicates that two or more of the independent variables had a high level of linear relationship between them. Because of the strong correlation, only one of the correlated variables was tested at a time and the variable yielding the best model was selected for the final analysis. For example, volume to capacity would first be included as a predictor variable along with the other non-correlated variables such as grade. Variables correlated with V/C would be excluded from the analysis (LOS, up and downstream volume). Second, LOS would be tested without volume to capacity or upstream or downstream volume. Third, upstream and downstream volumes were be used as separate independent variables along with the non-correlated variables. The best of the three models would then be selected. 142

166 Figure 6-2: Correlation Between V/C and Upstream Per Lane Volume (R 2 = 0.64) Variables such as lane width, grade, percent heavy vehicles, and CBD vs. non-cbd were used in the calculation of volume to capacity and level of service. However, a strong correlation was not detected between any of these variables and no further action was taken. 143

167 6.3 Results of Statistical Analysis for Passenger Cars Final results of regression tree results and model validation for each data partion unit as listed in Table 6-1 are presented below for both passenger cars and heavy vehicles. An in-depth discussion of the statistical analysis, assumptions, final model selection, and model validation for each response variable for the data partion ACCEL is presented below. The data partion, ACCEL, represents queued passenger cars from the initial queuing position downstream 200 feet. Since the analysis procedure is similar, the final model for each subsequent data partion is provided in the following sections without a detailed description of interim analysis steps and final model selection protocol Activity for Queue Vehicles From Stopping Point to 200 Feet Downstream ACCEL Model Described below are the five models (one for each of the response variables) for passenger cars stopped at the traffic signal. Data were analyzed for a distance of 200 feet downstream of the vehicle's initial queuing position. Next, model validation is discussed and the final model is presented Percent Activity >= 6.0 mph/s (ACC.6) To arrive at the "best" initial model, various regression tree models were created. Since several of the variables were highly correlated, a test run was made with different combinations of correlated variables as 144

168 described above. The initial model with the lowest deviation or best fit was used. Next, to simplify the model, various combinations of: 1) increasing allowed deviance at the nodes, 2) increasing or decreasing the minimum number of observations required before a split occurs, and 3) increasing or decreasing the minimum number of nodes were tested to simplify and improve model simplicity and "fit". The initial model was created by allowing the tree to grow unconstrained for the first cut. Once an initial model was created, the "snip.tree" function in S-PLUS was used to simplify the model by removing the lower branches of the "tree" that explained the least deviance. Each resulting "tree" was examined to ensure that the model's predictive ability wasn't compromised by allowing the overall amount of deviance to increase significantly. Figure 6-3 illustrates the initial tree model used for ACC6 (percent of activity >= 6.0 mph/s) for data from queue vehicles from the stopping point to a distance 200 feet downstream. Results for the initial model are given in Table 6-2. As noted, the tree grew into a complex model with a considerable number of branches and 13 terminal nodes. To simplify the model, various combinations of the prune, snip, and shrink functions were experimented with. The "snip.tree" function ended up being the most useful tool in simplifying trees. As explained previously, the first split in the regression tree explains the most deviation with following split subsequently explaining less of the deviation. Figure 6-4 illustrates the 145

169 amount of deviance explained corresponding to the number of terminal nodes. As shown, the first 13 nodes (not terminal nodes) explain 92% of the deviance. The additional 26 nodes combined only, explain 8% of the deviance. Figure 6-5 illustrates a normal probability plot of the residuals for the original untrimmed tree. Table 6-2: Full Untrimmed Regression Tree Results for ACC6 for Passenger Cars From Stopping Point to 200 Feet Downstream Summary(acc6accel.tree) Regression tree: Tree(formula = ACC6 ~ QUEUE + UPSTREAM + DOWNSTREAM + PER.TRUCKS + DOWNDIST + GRADE + LOCATION + NO.LANES + SPEEDLIMIT, data = CarsAccelClean, na.action = na.omit, mincut = 5, minsize = 10, mindev = 0.1) Variables actually used in tree construction: [1] "QUEUE" "GRADE" "DOWNSTREAM" "DOWNDIST" "UPSTREAM" "LOCATION" Number of terminal nodes: 13 Residual mean deviance: = / 386 Distribution of residuals: Min. 1 st Qu. Median Mean 3 rd Qu. Max e

170 Figure 6-3: Original Untrimmed Regression Tree Model for ACC6 for Passenger Cars From Stopping Point to 200 Feet Downstream Figure 6-4: Reduction in Deviance with the Addition of Nodes 147

171 Figure 6-5: Normal Probability Plot of the Residuals for the Original Untrimmed Tree A simplified model was derived which ends in six terminal nodes as compared to the 13 terminal nodes in the initial model. The residual mean deviance only increased from to and yielded a much cleaner model that the initial one. Results are shown in Table 6-3 and Figure 6-6. As noted the independent model variables are queue position, roadway grade, and downstream per lane volume. 148

172 Table: 6-3: Trimmed ACC6 Model Results for Queued Passenger Cars From Stopping Point to 200 Feet Downstream Regression tree: tree(formula = ACC6 ~ QUEUE + UPSTREAM + DOWNSTREAM + PER.TRUCKS + DOWNDIST + GRADE + LOCATION + NO.LANES + SPEEDLIMIT, data = CarsAccelClean, na.action = na.omit, mincut = 5, minsize = 10, mindev = 0.1) snip.tree(tree = a6accel.snip3, nodes = 5) Variables actually used in tree construction: [1] "QUEUE" "GRADE" "DOWNSTREAM" Number of terminal nodes: 5 Residual mean deviance: = / 394 Distribution of residuals: Min. 1st Qu. Median Mean 3 rd Qu. Max e

173 Figure 6-6: Trimmed ACC6 Model for Queued Passenger Cars From Stopping Point to 200 Feet Downstream Percent Activity >= 3.0 mph/s (ACC.3) This section describes the final regression tree models for the response variable ACC.3 (percent of activity >= 3.0 mph/s). Table 6-4 provides model results and Figure 6-7 shows the final regression tree model. In the final model, queue position and grade were the most significant variables. The final model had a rather poor fit with a RMD of Percent Activity Where Acceleration <= -2.0 mph/s (DEC.2) Regression tree results for the response variable DEC.2 (percent of vehicle activity for the indicated position where deceleration was less than or equal to -2.0 mph/s) are given in 150

174 Table 6-5 and Figure 6-8. Note that downstream per lane volume with a single split on per lane volume of 862 was the only variable for the final regression tree model. Table: 6-4: Trimmed ACC.3 Model Results for Queued Passenger Cars From Stopping Point to 200 Feet Downstream Regression tree: Tree(formula = ACC3 ~ QUEUE + UPSTREAM + DOWNSTREAM + PER.TRUCKS + DOWNDIST + GRADE + WIDTH + LOCATION + NO.LANES + SPEEDLIMIT, Data = CarsAccelClean, na.action = na.omit, mincut = 5, Minsize = 10, mindev = 0.1) Snip.tree(tree = a3accel.snip3, nodes = 3) Variables actually used in tree construction: [1] "QUEUE" "GRADE" Number of terminal nodes: 3 Residual mean deviance: = / 396 Distribution of residuals: Min. 1st Qu. Median Mean 3rd Qu. Max e

175 Figure 6-7: Trimmed ACC.3 Model for Queued Passenger Cars From Stopping Point to 200 Feet Downstream Table: 6-5: Trimmed DEC.2 Model Results for Queued Passenger Cars From Stopping Point to 200 Feet Downstream Regression tree: tree(formula = Decel2 ~ QUEUE + UPSTREAM + DOWNSTREAM + PER.TRUCKS + UPDIST + DOWNDIST + GRADE + WIDTH + LOCATION + NO.LANES + SPEEDLIMIT, data = CarsAccelClean, na.action = na.omit, mincut = 5, minsize = 10, mindev = 0.1) snip.tree(tree = last.tree, nodes = 2) Variables actually used in tree construction: [1] "DOWNSTREAM" Number of terminal nodes: 2 Residual mean deviance: = / 397 Distribution of residuals: Min. 1st Qu. Median Mean 3rd Qu. Max e

176 Figure 6-8: Trimmed DEC.2 Model for Queued Passenger Cars From Stopping Point to 200 Feet Downstream Average Vehicle Speed (AVG-SPD) The next response variable was average speed for the indicated position in mph. For the data partion from 0 to 200 feet from the point of queue, the single predictor variable for average speed was queue position with a residual mean deviance of Table 6-6 provides model 153

177 results and Figure 6-9 shows the final regression tree model. The analysis showed that queue positions 1, 2, and 3 were similar and queue positions 4 and higher were similar Inertial Power Surrogate >= 120 mph 2 /s (IPS120) The response variable is inertial power surrogate (IPS120--the product of speed and acceleration) that equaled or exceeded 120 mph/s 2 for the indicated position. Table 6-7 provides model results and Figure 6-10 shows the final regression tree model. The final variables included queue position and roadway grade, with the 1 st queue position in one split and all other queue positions in the other. For the first queue position, grade was divided into values < and values >= Grade did not apply to higher queue positions. Table: 6-6: Trimmed AVG_SPD Model Results for Queued Passenger Cars From Stopping Point to 200 Feet Downstream Regression tree: tree(formula = SPEED ~ QUEUE + UPSTREAM + DOWNSTREAM + PER.TRUCKS + DOWNDIST + GRADE + WIDTH + LOCATION + NO.LANES + SPEEDLIMIT, data = CarsAccelClean, na.action = na.omit, mincut = 5, minsize = 10, mindev = 0.1) snip.tree(tree = spdaccel.snip3, nodes = c(3, 2)) Variables actually used in tree construction: [1] "QUEUE" Number of terminal nodes: 2 Residual mean deviance: = 6331 / 397 Distribution of residuals: Min. 1st Qu. Median Mean 3rd Qu. Max e

178 Figure 6-9: Trimmed AVG_SPD Model for Queued Passenger Cars From Stopping Point to 200 Feet Downstream Table: 6-7: Trimmed IPS120 Model Results for Queued Passenger Cars From Stopping Point to 200 Feet Downstream Regression tree: Tree(formula = PKE120 ~ QUEUE + UPSTREAM + DOWNSTREAM + PER.TRUCKS + UPDIST + DOWNDIST + GRADE + WIDTH + NO.LANES + SPEEDLIMIT, Data = CarsAccelClean, na.action = na.omit, mincut = 5, minsize = 10, mindev = 0.1) Snip.tree(tree = last.tree, nodes = c(3, 4)) Variables actually used in tree construction: [1] "QUEUE" "GRADE" Number of terminal nodes: 3 Residual mean deviance: = / 396 Distribution of residuals: Min. 1st Qu. Median Mean 3rd Qu. Max e

179 Figure 6-10: Trimmed IPS120 Model for Queued Passenger Cars From Stopping Point to 200 Feet Downstream Summarization of Results for ACCEL As described in section 4.6, the statistical approach used for data analysis involved a two-step process. Hierarchical based regression tree analysis was first used, as described in the preceding sections, to identify the predictor variables with the greatest power to explain the most variation in each of the five response variables. Next, the predictor variables were used to stratify the original datasets, into three-dimensional matrices in the form of a Joint Acceleration-Speed Probability Density Function that can be used as input to MEASURE. Because data were originally collected in 156

180 one-second intervals, JASPROD are in one-second "bins". JASPRODs are created by dividing vehicle traces into a matrix of speed and associated accelerations bins according to the operational or geometric characteristics, which were shown to be statistically significant Final Predictor Model for ACCEL The variables that were shown to be the most relevant from regression tree analysis in influencing activity traces for passenger cars from the initial point of queuing at the signalized intersection to a point 200 feet downstream, for all the response variables, include roadway grade, queue position, and downstream per lane volume. Relevant queue positions include the first vehicle in queue, second and third vehicles in queue combined, and the fourth vehicle in queue and higher combined. Grade is significant for the first, second, and third queue positions. According to the data analysis, vehicle activity for this segment should be stratified by queue position, and then other variables as shown in Table 6-8. Table 6-8: Breakpoints for Data Stratification From the Initial Queue Position Downstream 200 Feet 1 st in Queue 2 nd and 3 rd in Queue 4 th in Queue and Grade < -4.5 Grade < -4.5 Down per lane < 862 Down per lane < 862 Down per lane < 862 Down per lane >= 862 Down per lane >= 862 Down per lane >= > Grade > = > Grade >= > Grade >= -1.5 Grade >= -1.5 Grade >= -0.5 Down per lane < 862 Down per lane < 862 Down per lane >= 862 Down per lane >=

181 Model Validation for ACCEL Model validation was difficult since the dataset was not large enough to reserve a subset of sufficient size for validation. Additionally, resources did not allow additional data collection to provide a "control" data sample. However, the methodology can be validated internally. Following is description of data validation to demonstrate the process of internal validation. Results are presented for passenger cars for the data segment from the initial stopping point downstream 200 feet. To validate model results, initial raw field data for the indicated segment were divided by the factors listed in Table 6-9. Distributions of speed and acceleration for data subsets were each compared using the non-parametric Kolmogorov-Smirnov goodness of fit test for two independent samples using S-PLUS 4.5. The Kolmorgorov-Smirnov test was described in more detail in section A description follows for a comparison of two datasets. Results of the K-S test for the first dataset (queue position = 1, grade < -4.5%, downstream per lane volume < 862 {out1}) versus the second dataset, where queue position and downstream per lane volume were held constant and the grade changed (queue position = 1, grade >= -0.5, and downstream per lane volume < 862 {out10}) are provided in Tables 6-9 and The K-S was used to compare both speed and acceleration. As shown, the null hypothesis that the distributions are the same was rejected for both speed and acceleration. This indicated that the data should indeed be divided by these parameters 158

182 from the regression tree analysis. Figure 6-11 compares the cumulative distributions for the two datasets. Table 6-9: K-S Test Statistic for Comparison of Datasets 1 and 10 for Acceleration Distributions Ks.gof(out1accel,out10accel) Two-Sample Kolmogorov-Smirnov Test Data: out1 and out10accel Ks = , p-value = 0 Alternative hypothesis: Cdf of out1 does not equal the cdf of out10accel for at least one sample point. Table 6-10: K-S Test Statistic for Comparison of Datasets 1 and 10 for Speed Distributions Two-Sample Kolmogorov-Smirnov Test Data: out1speed and out10speed Ks = , p-value = Alternative hypothesis: cdf of out1speed does not equal the cdf of out10speed for at least one sample point. 159

183 Figure 6-11: Comparison of CDFs for Dataset Out1 and Out10 If the K-S test indicates that there is no statistical difference between the acceleration and speed distributions of two data subsets, a strong case can be made for aggregating the data up a level Final Model for Queued Vehicles for ACCEL After the K-S tests were performed to validate the results of the regression tree analysis, a final model was selected which reflected any changes to the data divisions indicated by the K-S tests. If the K-S test indicated that the distributions were similar, the data from the two distributions were combined. The final model, which governed how the final datasets were disaggregated for MEASURE is presented in Table As noted the only difference between the original divisions of data suggested by the regression tree analysis, Table 6-8, and the final model was in how the downstream per lane volume variable was ultimately divided. The original three divisions for downstream per lane volume were: 1) downstream < 862; 2) 862 <= downstream < 902; and 3) downstream > =902. The three were collapsed into two divisions after using the K-S test: A) downstream < 862 and B) downstream >=

184 Table 6-11: Breakpoints for Data Stratification From the Initial Queue Position Downstream 200 Feet 1 st in Queue 2 nd and 3 rd in Queue 4 th in Queue and Higher Grade < -4.5 Grade < -4.5 Down per lane < 862 Down per lane < 862 Down per lane < 862 Down per lane >= 862 Down per lane >= 862 Down per lane >= > Grade > = > Grade >= > Grade >= -1.5 Grade >= -1.5 Grade >= -0.5 Down per lane < 862 Down per lane < 862 Down per lane >= 862 Down per lane >= 862 since division 2 (862<= downstream < 902) was shown to have the same distribution as division 3 ( downstream >= 902) Activity for Queued Vehicles From 200 to 400 Feet Downstream of Initial Stopping Point (ACCELPLUS200) The next data partion modeled was passenger vehicle activity that encompassed a distance from 200 feet downstream of the queued vehicle's initial position to a point 400 feet downstream. The final regression tree model results are presented in Appendix B. Regression tree analysis and the K-S test were used and indicated that queue position and grade were the most relevant variables in explaining variation. Table 6-12 illustrates the final data breakdown by queue position. Table 6-12 Breakpoints for Data Stratification From the 200 to 400 Feet Downstream of the Initial Queue Position 161

185 1 st in Queue 2 nd and 3 rd in Queue 4 th and Higher in Queue Grade < -6.5 Grade < -4.5 Grade < > Grade >= > Grade >= > Grade >= > Grade >= -4.5 Grade >= -1.5 Grade >= -1.5 Grade >= Activity for Queue Vehicles From 400 to 600 Feet Downstream of Initial Stopping Point (ACCELPLUS400) The next data partion was activity from 400 feet downstream of the queued vehicle's initial queuing point to a point 600 feet from the initial queuing point. The final regression tree model results are presented in Appendix B. For this data segment, queue position, downstream per lane volume, distance to the nearest downstream intersection and percent heavy vehicles were indicated as being relevant. According to model results, the first and second vehicles in queue should be combined and 3rd and higher queue positions combined. For the 1st and 2nd queue positions, distance to the nearest downstream intersection was relevant. For 3 rd and higher queue positions, percent heavy vehicles in the traffic stream was significant. Final operational and geometric predictor variables after model validation are demonstrated in Table Table 6-13: Breakpoints for Data Stratification From the 200 to 400 Feet Downstream of the Initial Queue Position 1 st or 2 nd in Queue 3 rd or Higher in Queue Down per lane < 451 Down per lane < 878 Downdist < 803 Percent trucks < 5.5 Downdist >= 803 Percent trucks >= > Down per lane >= 451 Down per lane >= 878 Down per lane >=

186 163

187 6.3.4 Activity for Queue Vehicles From 600 to 1,000 Feet Downstream of Initial Stopping Point (ACCELPLUS600 and ACCELPLUS800) The next data partion was activity that covered distances from a point 600 feet downstream of the queued vehicle's initial queuing point to a point 1000 feet from the initial queuing point. Data were initially divided by 200 feet increments. However, data were combined from two segments, ACCELPLUS600 and ACCELPLUS800, since fewer data were collected at increasing distances from the data collection location. Additionally, at some point along a signalized link, it is expected that vehicle activity will become more homogenous. The distance segment was included as a variable to test whether it was in important factor in influencing vehicle activity (i.e were data from ACCELPLUS600 measurably different from ACCELPLUS800). The final regression tree model results are given in Appendix B. Statistical analysis indicated that posted speed limit, grade, and downstream per lane volume are the most relevant variables that influence vehicle activity for this data segment. Data should be divided first by the posted link speed limit with speeds less than 45 mph in one set and posted speeds of 45 mph and higher in another and then further divided by grade and downstream per lane volume. The final model is shown in Table

188 Table 6-14: Breakpoints for Data Stratification From the 200 to 400 Feet Downstream of the Initial Queue Position Speedlimit < 45 mph Speedlimit >= 45 mph Down per lane < 491 Down per lane < 491 Down per lane >= 491 Grade < -1.5 Grade >= -1.5 Down per lane >= 491 Grade < -1.5 Grade >= Activity for Queue Vehicles From Initial Stopping Point Upstream 200 Feet (DECEL) After data collected from the stopping point forward for queue vehicles were analyzed for various distances, deceleration activity that occurred prior to the vehicle's queuing position was analyzed. The first deceleration data partion was activity from the vehicle's queuing position upstream 200 feet. The final regression tree model results are found in Appendix B. The final combination of geometric and operational variables that influence vehicle activity for this data segment include distance to the nearest upstream signalized intersection, upstream per lane volume, and queue position. The data should first be stratified by data collected at locations where the upstream distance is less than 1,168 and then data collected in locations with a distance to the nearest upstream intersection is greater than 1,168 and less than 3,432 feet. The next set of data should be divided by segments where the nearest upstream intersection is greater than 3,432 feet. The final rules 165

189 for division of data according to regression tree and K-S test results for this data segment are provided in Table Table 6-15: Breakpoints for Data Stratification From the Initial Queue Position Upstream 200 Feet Updist < > Updist >= 1168 Updist > st & 2 nd in Queue 1 st & 2 nd in Queue 1 st & 2 nd in Queue Up per lane < 613 Up per lane < 613 Up per lane < 613 Up per lane >= 613 Up per lane >= 613 Up per lane >= rd and higher queue positions 3 rd thru 8 th position in queue 3 rd and higher queue positions Up per lane < 613 Up per lane < 613 Up per lane < 613 Up per lane >= 613 Up per lane >= 613 Up per lane >= th in queue and higher Up per lane < 613 Up per lane >= Activity for Queue Vehicles From 200 Feet Upstream of the Initial Stopping Point to a 400 Feet Upstream (DECELNEG200) The second deceleration data partion was activity from 200 feet to 400 feet upstream of the vehicle's queuing position. The final regression tree model results are presented in Appendix B. The final combination of variables that were shown to be the most relevant in influencing activity traces according were grade, upstream per lane volume, and queue position. According to the data analysis, vehicle activity for this segment should be stratified by queue position and then other variables as shown in Table

190 Table 6-16: Breakpoints for Data Stratification From 200 to 400 Feet Upstream of the Initial Queue Position 1 st and 2 nd in Queue 3 rd in Queue 4 th in Queue and Higher Grade < -6.5 Grade < -6.5 Grade < > Grade >= -6.5 Grade >= > Grade >= -6.5 Grade >= 0 Up per lane < 447 Grade >= 0 Up per lane >= Activity for Queued Vehicles From 400 Feet Upstream of the Initial Stopping Point to a 600 Feet Upstream (DECELNEG400) For the data segment from 400 to 600 feet upstream of the initial queuing position, only a single variable influenced vehicle activity. No activity was observed for the response variables of acceleration >= 6.0 mph/s, acceleration >= 3.0 mph/s, or IPS >= 120. The single predictor variable which explained the most deviation in both average speed and percent of activity where acceleration <= -2.0 mph/s is upstream per lane volume with splits on upstream < 601 vehicles per lane per hour and upstream >= 601 vehicles per lane per hour "THRU" Vehicles at All locations Vehicles not stopped at the intersection were analyzed separately from stopped vehicles stopped since their vehicle activity traces are expected to be much different in the vicinity of the intersection. Data were partitioned into 200-foot segments as for queued vehicles. However all data partions were included in a single analysis for "THRU" vehicles 167

191 and distance was included as a variables to test whether the location from the stopline affects vehicle activity. The variables downstream and upstream volume were replaced by the variable VOLUME as described in Section The variable LINKDIST was also included which reflected the length of the link where data collection was taking place and is explained in section Including midblock data, the distances for data collection ranged from 2,000 feet upstream of the intersection stopbar to 1,200 feet downstream of the intersection stopbar. Regression tree analysis indicated that midblock data were statistically different from upstream and downstream data positions. The designation for THRU data are those from a distance 1000 feet upstream of the intersection stopline to 1200 feet downstream. All other locations may be considered MIDBLOCK. Additionally, the posted link speed limit, link volume, and link distance were shown to influence THRU vehicle activity. The final data divisions are shown in Table Table 6-17: Breakpoints for Data Stratification For Thru Vehicles for All Distances Upstream and Downstream of the Data Collection Intersection Midblock Link distance < 3004 Link distance >= 3004 At intersection stopline 168

192 Speed limit = 30 Speed limit = 35 or 40 Speed limit > 40 Link vol. < 543 Link vol. < > Link vol. >= 543 Link length < > Link vol. >= 777 Link length >= 3004 Link Vol. >= > Link Vol. >= 543 Link length < 3004 Link length >= 3004 Link vol >=

193 6.4 Heavy Trucks The various regression tree models for heavy vehicles were much easier to run. In many cases most of the initial regression tree models were simple enough that further trimming was not warranted. This is likely due to the fact that heavy vehicle activity has much less variation to begin with than passenger car activity since vehicle operation may be constrained by vehicle rather than driver constraints. Presented below are the final models from regression tree analysis and K-S validation for each data segment position for heavy trucks Heavy Vehicle Activity for Queue Vehicles From Stopping Point to 200 Feet Downstream (ACCEL) Model This model provides results for heavy vehicles that were stopped at the traffic signal and includes data for a distance from the vehicle s initial queuing position downstream 200 feet. The variables that were shown to be the most relevant are roadway grade and queue position as shown in Table Table 6-18: Breakpoints for Data Stratification From the Initial Queue Position Downstream 200 Feet 1 st and 2 nd Queue Positions Grade < > Grade >= -4.5 Grade > rd thru 6 th Queue 7 th and Higher Queue Positions Positions No further division necessary No further division necessary 170

194 6.4.2 Heavy Vehicle Activity From 200 feet From Stopping Point to 800 Feet Downstream (ACCELPLUS200 to ACCELPLUS600) This model provides results for heavy vehicles that were stopped at the traffic signal and include data from a point 200 feet downstream of the vehicle's initial queuing position to a position 800 feet from the initial stopping point. Two data partions were combined, so distance from the initial stopping point was also included as an independent variable. The variables shown to be the most relevant in influencing activity traces for this data segment for heavy trucks include speed limit, grade, and percent trucks. According to the final data analysis, vehicle activity for this segment should be stratified by the variables as shown in Table Heavy Vehicle Activity From 200 feet Upstream to Stopping Point (DECEL) The final combination of geometric and operational variables that influence vehicle activity from the initial point of queue to a position 200 feet upstream include queue position and grade. According to the regression tree analysis and K-S, vehicle Table 6-19: Breakpoints for Data Stratification From 200 to 600 Feet Downstream of the Initial Queue Position Speed limit < 36 Speed limit >= 36 Grade < -4.5 Grade < -4.5 Grade >= -4.5 Grade >= -4.5 Percent trucks < 2.5 Percent trucks <

195 Percent trucks < 2.5 Percent trucks < 3.5 Percent trucks >= 2.5 Percent trucks >= 3.5 activity for the first, second, and third queue positions behaved similarly and should be combined. Then data for the 4th and higher queue positions should be combined. The final rules for division of data for this data segment is provided in Table Heavy Vehicle Activity for Queue Vehicles From 200 feet up to All Prior Upstream Positions (DECELNEG200 to DECELNEG400) This data segment was analyzed by including datasets for activity from a point 200 feet above the initial queuing location to any point upstream of that position. Data include activity from 200 feet to 600 feet upstream. Distance from the initial queuing position was also included as an independent variable to test if the distance from the signal was relevant. The only variables shown to be relevant in influencing activity traces for heavy trucks was upstream per lane volume as shown in Table Table 6-20: Breakpoints for Data Stratification From the Initial Queue Position Upstream 200 Feet 1 st, 2 nd and 3 rd in Queue 4 th and Higher in Queue Grade < 3.5 Grade < 3.5 Grade >= 3.5 Grade >=

196 Table 6-21: Breakpoints for Data Stratification From 200 to 600 Feet Upstream of the Initial Queue Position All queue positions Upstream < 605 Upstream >= Heavy Vehicle Activity for "THRU" Vehicles for All Positions This model provides results for heavy vehicles that were not stopped at the traffic signal and includes data for all distances before and after the stopbar including midblock. Data for all THRU vehicles were combined and was composed of midblock data, data collected immediately upstream of the intersection, and data collected immediately downstream of the intersection. Link per lane volume (VOLUME) and link distance (LINKDIST) were also included as variables as explained for the passenger vehicle THRU data segment. Regression tree analysis indicated that midblock data were not statistically different from upstream and downstream data positions. After regression tree and K-S analysis the only relevant variable was percent grade of the segment being studied as shown in Table Comparison of Data to Existing Relationships In this section, field data are compared against other relationships that describe speed and acceleration activity. A presentation of the ranges of acceleration found by speed range is presented. Data are also compared against the values from the Traffic Engineering Handbook (ITE, 1992), simulation models, and NCHRP 185. Activity 173

197 Table 6-22: Breakpoints for Data Stratification For Thru Vehicles for All Distances Upstream and Downstream of the Data Collection Intersection All midblock and intersection activity Grade < -4.5 Grade >= -4.5 collected in the field that fell outside the range of activity in the FTP was also included Ranges of Field Data An overview of the data collected is given in Table 6-23, which provides a bin count by speed and acceleration range for all recorded values for all speed ranges for passenger cars. Accelerations greater than and equal to 11.5 mph/s were combined into the 12 mph/s bin. Accelerations less than and equal to mph/s were combined into the -12 mph/s bin. Figure 6-12 illustrates a graph of acceleration ranges by speed category (each speed category sums to 1). This shows the variation in acceleration activity by speed range. As shown, a significant variety exists in the data for accelerations across all speed ranges. This data was intended to show that relationships that model acceleration as an inversely proportional linear relationship to speed, do not provide a statistical distribution of actual onroad vehicle activity. As noted, the most variation in acceleration ranges occurs at the lower speed ranges from 0 to 35 mph speed bins. 174

198

199 171 Table 6-23: Field Data Acceleration Observations by Speed Range Acceleration Velocity (mph) (mph/s) Plus Plus Column Total

200 173 Figure 6-12: Acceleration Distribution (mph/s) by Speed Ranges (mph) 177

201 6.5.2 Comparison of Research to Existing Simulation Modeling The use of simulation models by various research groups to output individual activity profiles was discussed in Chapter 3. Simulation models offer attractive advantages for modal activity modeling. They are readily available and often allow differing levels of analysis with both simple and detailed data input. A major advantage to simulation modeling is the ability to make multiple runs and compare different scenarios, such as comparing the effect of different traffic timing plans on individual vehicle delay. The use of simulation models for signalized intersections is especially promising because intersections are locations of significant modal activity. Along signalized links, vehicle activity is particularly impacted by intersection characteristics such as cycle length, which can easily be modeled by simulation. However, simulation models often employ theoretical profiles of vehicle acceleration and speed relationships. The algorithms were intended to model gross measures of traffic activity, such as changes in cycle length or the effect of an incident. The models have been validated under these conditions and perform well for the applications for which they were developed. Internal algorithms, however, remain unvalidated for predicting individual vehicle activity. Additionally, most models are incapable of integrating temporal and spatial characteristics of traffic and roadways. 178

202 To explore whether simulation models can be used to output realistic estimates of individual vehicle activity and to identify drawbacks in their use, a companion study to this research work (Hallmark and Guensler, 1999) compared individual activity output from a simulation model with the field-collected vehicle profiles at signalized intersections that was part of this work. For the comparison, a single study intersection was modeled using simulation runs from NETSIM, the non-freeway, urban traffic simulation module of the TRAF (CORSIM) traffic simulation model family. Instantaneous speed/acceleration output from NETSIM for the study intersection was compared with the field data. A brief overview of the results are discussed below, for a more detailed explanation, the reader is referred to Hallmark and Guensler (1999). Comparison of NETSIM and field data for the same sample intersection demonstrated significant differences. Figure 6-13 and Figure 6-14 shows frequency of activity by acceleration range for each model and frequency of activity by speed range for a 500-foot segment. Note that NETSIM underpredicts higher acceleration ranges (3 to 8 mph/s) for the study intersection. As shown, NETSIM also underpredicts vehicle activity in the higher speed ranges (45 to 65 mph). 179

203 Figure 6-13: Comparison of Percent Time Spent in Each Acceleration Range for Field Data and NETSIM (-250 to 250 feet from the stopbar) Figure 6-14: Comparison of Percent Time Spent in Each Speed Range for 180

204 Field Data and NETSIM (-250 to 250 feet from the stopbar) The subset of midblock activity data was also analyzed. Figures 6-15 and 6-16 show frequency plots by speed and acceleration for a 500-foot segment of vehicle activity 2000 feet downstream of the study intersection. Unlike on-road vehicle activity, NETSIM predicts few midblock acceleration events. Once a vehicle achieves it s desired speed, modeled acceleration activity remains fairly static. NETSIM also has a narrow range of midblock speeds, ranging from 25 to 55 mph. Field data speeds range from 0 to 65 mph. Field data show much greater acceleration variations and wider speed ranges. As demonstrated in Figure 6-15, the field data midblock shows accelerations ranging from - 6 mph/s to 7 mph/s. The simulation model data only show activity for the acceleration ranges from -4 to 3 mph/s. As shown in Figure 6-16, NETSIM has much narrower speed ranges than demonstrated by the field data. No downstream queuing or significant driveway interactions were noted, which would influence variations in speed and acceleration in the field data. 181

205 Frequency Field NETSIM Acceleration (mph/s) Figure 6-15: Comparison of Percent Time Spent in Each Acceleration Range for Field Data and NETSIM (midblock) 18% 16% 14% Frequency 12% 10% 8% 6% Field Data NETSIM 4% 2% 0% Speed (mph) 182

206 Figure 6-16: Comparison of Percent Time Spent in Each Speed Range for Field Data and NETSIM (midblock) Results of the study indicate that even though the NETSIM model may be calibrated correctly to predict aggregate flows or speeds, it is not necessarily calibrated to provide accurate speed/acceleration profiles. If NETSIM or similar simulation models do not predict speed/acceleration profiles correctly, the ultimate impact is largely dependent on the emission factors that are applied to the data. Emissions predicted from modal emission rate models, which predict significantly higher emissions at higher engine loads, will be adversely affected by errors in predicted speed/acceleration profiles, especially in the extreme speed/acceleration bins. When modal emission factors indicate that average speeds are a highly significant variable (as they are for oxides of nitrogen), NETSIM outputs are likely to underestimate modal emissions. When high accelerations at low to medium speed ranges are more significant, NETSIM has the potential to over-represent emissions (Hallmark and Guensler 1999). One of the main reasons simulation models are unable to realistically model vehicle activity are the underlying assumptions about vehicle behavior used in the model. In models, such as NETSIM, a desired speed is assigned to each vehicle, which then attempts to reach that target speed. The actual speeds attained are a function of interference with traffic 183

207 control devices and interference with surrounding vehicles. The accelerations corresponding to each instantaneously generated speed are constrained by car-following logic and an upper bound maximum acceleration, which is a function of speed. The maximum acceleration at any given speed is determined by a linear speed-acceleration relationship with maximum acceleration occurring at zero velocity and zero acceleration at the maximum velocity. The relationship is similar to that reported in NCHRP 185 (11 from TRB1999). TRAF version 5.0, used in the analyses reported here, allows users to define the maximum acceleration for zero speed on dry level roads for a specified vehicle type (USDOT, 1995). A later version of the program allows user defined maximum acceleration rates for specified speed ranges (FHWA, 1995). The problem with activity modeling that uses this linear speed-acceleration where maximum acceleration is constrained by upper bound depending on the particular speed, is that a vehicle can select any acceleration range up to that upper bound. No statistical distribution of actual speeds and corresponding acceleration is actually incorporated into the model. A plot of speed versus acceleration for data 0 to 250 feet from the stopbar for the first vehicle is shown in Figure for the first vehicle in the queue from the NETSIM dataset described above, with the field data set. Data were extracted from NETSIM for 212 "first in the queue" vehicles and field data provided 37 "first in the queue" vehicles. Even 184

208 though data were available for roughly three times as many vehicles, the NETSIM simulation data show much less variation Figure 6-17: Comparison of Field and NETSIM Data for the First Vehicle in the Queue (stopbar to 250 feet downstream) than the field data. Additionally, acceleration peaks are noted in the 7 to 22 mph speed range for the field data and from 0 to 10 mph for the simulation data. Although the NETSIM microsimulation model has been presented, other simulation models also have potential pitfalls that affect their ability to accurately model microscopic 185

209 vehicle activity. The TRANSIMS simulation model based vehicle activity and position on car-following theory rather than field studies of vehicle activity. One main drawback to the model is that the cellular automata model describes vehicle position in units of cells, velocity in units of cells per second and acceleration in units of cells per second per second. Since the typical cell size is 7.5 meters, speed is modeled in 16 mph increments, which is too aggregated for direct use in modal emissions modeling. Another drawback is that it does not accurately represent acceleration events, which are a major variable in emissions (Williams et al., 1999). Other traffic simulation and optimization models such as TRANSYT-7F, INTEGRATION, FREQ, NETSIM, and INTRAS calculate emissions but base output on existing logic which is not expected to realistically model microscopic vehicle activity since none of the models were developed based on on-road emission or vehicle activity data (Yu, 1999). The simulation model proposed by Rakha et al. (1999) bases vehicle activity on car following logic constrained by a linear acceleration decay function and may be characterized by unrealistically high accelerations. To compensate, the model uses a linear acceleration decay function that decreases. However, the application of the linear decay function has not been validated with field studies. 186

210 6.5.3 Comparison of Research to Traffic Engineering Rates Evaluation of field data indicated that measured on-road maximum acceleration exceed the published values from the Traffic Engineering Handbook (ITE, 1994) as listed in Tables 3-1 and 3-2 in Chapter 3. A comparison of field data with the Traffic Engineering Handbook values is provided in Table Values in the Traffic Engineering Handbook were listed by weight to power ratio. Since this value was not available for the field data, the maximum value for any weight to power ratio from the Traffic Engineering Handbook were compared to the field collected values. All data are for level roadways (-1% to 1% grades). As noted, all field values exceeded the maximum published values for both passenger cars and heavy trucks, indicating that commonly used acceleration rates may not adequately represent on-road acceleration Comparison of Data to NCHRP 185 Some of the earlier speed acceleration relationships in traffic engineering were based on NCHRP 185, which derived a linear speed-acceleration relationship as described in Section A comparison of this relationship with field data for the Table 6-24: Comparison of Field Data and Traffic Engineering Handbook Maximum Acceleration by Speed Range (mph/s) Vehicle Type 0 to 10 mph 10 to 20 mph 20 to 30 mph 30 to 40 mph 40 to 50 mph 50 to 60 mph Passenger Cars from Traffic Eng. Handbook

211 Passenger Cars from Field Data Heavy Trucks from Traffic Eng. Handbook Heavy Trucks from Field Data first vehicle in queue for acceleration off the stopbar is shown in Figure Only data for the first vehicle in queue are presented since they are the only vehicles in the traffic stream that enjoy unconstrained movement. Only data collection sites with no downstream backup were included so that vehicle activity represents unconstrained acceleration. As shown, vehicle activity does not follow a linear relationship. At low speeds, the vehicle is unable to achieve high on-road acceleration. Acceleration ability increases with increasing speed until approximately the 10 to 25 mph speed 188

212 Figure 6-18: Comparison of Field Data for First Vehicle in Queue with Linear Speed-Acceleration Relationship range when on-road acceleration decreases. Also demonstrated by this figure, is that onroad vehicles undergo a wide distribution of vehicle activity at any given speed range beyond 0-5 mph. This indicates that the linear speed-acceleration relationship is too simplistic to adequately model on-road vehicle activity for specialized applications such as air quality modeling Comparison of Data to FTP Range of Activity 189