Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
135 5 SUMMARY This chapter summarizes the steps involved in developing a travel time reliability monitoring system (TTRMS), starting with the various potential data sources and ending with metrics that can be used to assess performance. The chapter draws on the methods presented in Chapters 2 and 3 and identifi es case studies from Ch apter 4 that illustrate real-world applications of the methods. The overall process can be described using the diagram shown in Figure 5.1; each of the steps described in this chapter refers to this diagram. The process is illustrated as linear, but in reality it can be recursive and nonlinear due to a number of factors: ⢠Types of source data available, both recurrent traffi c-related data and nonrecur- rent event data (e.g., weather, incidents); and ⢠Use cases of interest (e.g., real-time versus historical evaluation). The statistical methods necessary to conduct the analyses tend to be sensitive to the data being analyzed. The methodological details in Chapter 3 and its support- ing reports demonstrated the potential complexities involved, and the case studies in Chapter 4 illustrated how a customized travel time reliability analysis needs to be fi t to a particular location. As a result, generalizations on the methods for analysis in this chapter are only illustrative of the actual process that a specifi c analyst will follow in a given situation and are not necessarily the one that will be applicable to every system in every instance.
136 GUIDE TO ESTABLISHING MONITORING PROGRAMS FOR TRAVEL TIME RELIABILITY STEP 1. COLLECT AND MANAGE TRAFFIC DATA The first step in the analytical process, illustrated in Figure 5.2, is to collect and store all types of traffic data. These traffic data should be both mode neutral (e.g., freeway lane data) and mode specific (e.g., transit data). These sources can be divided into two major groups: infrastructure-based sources and vehicle-based sources. Infrastructure-Based Sources Infrastructure-based detectors collect single-point count and occupancy data or double-point speed data anonymously (i.e., without specifically identifying a particular vehicle). Typical applications include the following types of infrastructure-based sensors: ⢠Loop detectors, illustrated in the San Diego, California, and Northern Virginia case studies in Chapter 4; ⢠Video image processors, illustrated in the Atlanta, Georgia, case study in Chapter 4; ⢠Wireless magnetometer detectors; and ⢠Radar detectors. Figure 5.1. Information flow in a TTRMS. Measure Data Computation Engine Characterize Identify Understand Decision Sources of Congestion Travel times without incidents Travel times during incidents Travel time during an incident Low travel times High travel times Travel time
137 GUIDE TO ESTABLISHING MONITORING PROGRAMS FOR TRAVEL TIME RELIABILITY Management of the data from each of these detector sources follows the general methods outlined in Chapter 2. Vehicle-Based Sources Vehicle-based detectors collect data about specific vehicles, either when they pass a fixed point, as in the case of automated vehicle identification (AVI), or as they travel along a path for automated vehicle location (AVL). Automated Vehicle Identification Data AVI data collection sources, which include Bluetooth readers, electronic toll tag readers, and license plate readers, detect passing vehicles at sensor locations. The data are postprocessed to identify matches for specific vehicles passing successive sensors, allow ing the vehicleâs trip time between two points to be directly computed. Chapter 4 presents an application of AVI data in the form of Bluetooth data in the Sacramentoâ Lake Tahoe, California, case study. Figure 5.2. The role of data. Measure Data Computation Engine Characterize Identify Understand Decision Sources of Congestion Travel times without incidents Travel times during incidents Travel time during an incident Low travel times High travel times Travel time
138 GUIDE TO ESTABLISHING MONITORING PROGRAMS FOR TRAVEL TIME RELIABILITY Automated Vehicle Location Data AVL data collection sources provide perhaps the richest source of data in that they identify traceable paths for individual vehicles through a system. Chapter 4 presented the application of AVL data for travel time reliability analysis for two case studies: 1. In the San Diego case study, transit travel time reliability was studied using data collected from AVL and automated passenger count data from San Diegoâs transit system. 2. In the New York/New Jersey case study, data from global positioning system (GPS)âbased technology from ALK Technologies, Inc., were used to study travel times for vehicle trips in the New York City metropolitan area. STEP 2. MEASURE TRAVEL TIMES The second step in the analytical process, illustrated in Figure 5.3, is to measure the travel times, ultimately at the route level. To get to this final product, the process first requires imputing missing data (Step 2A), computing segment travel times (Step 2B), and finally computing route travel times (Step 2C). Figure 5.3. Measuring travel times. Measure Data Computation Engine Characterize Identify Understand Decision Sources of Congestion Travel times without incidents Travel times during incidents Travel time during an incident Low travel times High travel times Travel time
139 GUIDE TO ESTABLISHING MONITORING PROGRAMS FOR TRAVEL TIME RELIABILITY Step 2A. Impute Missing Values A major challenge in producing useful travel time reliability performance measures is the reality that the data sources are often incomplete for a variety of reasons. Sensor coverage can often be incomplete, either due to gaps in system coverage or individual sensor malfunctioning. Infrastructure-Based Sensors Chapter 3 presented a series of techniques for identifying malfunctioning detectors, discarding the poor data, and filling in the missing data with imputed values. These techniques include a series of regression methods and the use of temporal or cluster medians. Case studies in Chapter 4 illustrate methods for imputing missing infrastructure- based data: ⢠The case studies for San Diego and Northern Virginia illustrate the use of imputa- tion for loop detector data. The degree to which imputation is necessary directly affects the quality of the data used for analysis. For San Diego, 5-minute travel times were discarded if they were computed from data in which more than 20% was imputed. When imputation was necessary, the robust methods described in Chapter 3 were employed. For the Northern Virginia case study, the majority of detectors on selected corridors were reporting no data or bad data, thus requiring the use of less robust imputation algorithms. ⢠The case study for Atlanta illustrates the use of imputation for video-based detec- tion systems. Vehicle-Based Sensors Vehicle-based sensors also require imputation to correct poor or missing data. An illustration of collection and imputation of AVI data (specifically, Bluetooth data) is provided in Chapter 4 in the case study for SacramentoâLake Tahoe. Similarly, an il- lustration of collection and imputation of AVL data (specifically, GPS data) is provided in Chapter 4 in the case study for New York/New Jersey. Both case studies illustrate the challenges of matching the AVI or AVL data to the network and filtering the data to that which is relevant to the segments under study. Because both AVI and AVL data are associated with individual vehicles that can enter and leave the network between measurement points, these outliers need to be identi- fied and removed. The methods for doing this are included in the segment travel time calculations in Step 3, as they require analysis and processing of the entire data set to identify which data points should be considered assigned to the link and which are likely outliers. Step 2B. Compute Segment Travel Times The methods used to calculate segment travel times, including their distribution, depend on the type of sensors used as the data source.
140 GUIDE TO ESTABLISHING MONITORING PROGRAMS FOR TRAVEL TIME RELIABILITY Infrastructure-Based Sensors In this case, segment travel times are computed from either the measured or estimated speeds at each sensor. For double-loop detectors, the speeds are directly measured. However, for single-loop detectors, the speeds are estimated using tools like a g-factor, which combines an assumed vehicle length with the flow and occupancy data from the single-loop detector. This method is described in Chapter 3. While single-loop and double-loop detectors can be used to measure or estimate average speeds, they do not directly report individual vehicle travel times. Chapter 3 contains an example of how to synthesize such individual vehicle times from infra- structure-based sensor data and an estimate of the distribution of vehicle speeds. Vehicle-Based Sensors: Automated Vehicle Identification Data Unlike infrastructure-based sensors, vehicle-based sensors (both AVI and AVL) can make available observations of individual vehicle travel times from one location to another. Thus direct measures of segment travel times can be provided. The challenge in determining vehicle segment travel times is matching the AVI or AVL data to the network. The analytical techniques for estimating segment-level travel times and rates are presented in Chapter 3. A methodology for filtering AVI data (Bluetooth data) is given in Chapter 3, and an example of the conversion of AVI data into segment travel times is provided in the SacramentoâLake Tahoe case study presented in Chapter 4. An important challenge in dealing with AVI data is identifying the most appropriate observations from the mul- tiple responses received by the AVI reader as the vehicle passes. As described in detail in the case study, three processes are recommended: 1. Identification of passage times for vehicles at an individual AVI reader. This in- cludes aggregating device observations into visits and selecting the appropriate time to represent the vehicleâs visit at that device. 2. Generation of passage time pairs. Three methods have been illustrated to accom- plish this: maximum origin and destination permutations, use of all origin visits, and aggregation of visits. 3. Generation of segment travel time histograms. This process involves filtering out- liers across days and across time intervals, removing intervals with few observa- tions, and removing the data from highly variable intervals. Vehicle-Based Sensors: Automated Vehicle Location Data An example of the conversion of AVL data (GPS data) into segment travel times is pro- vided in the New York/New Jersey case study presented in Chapter 4. Illustrated here are techniques for matching GPS data to the network and identifying the GPS pings that are closest to the monuments that define the measurement points for the segments.
141 GUIDE TO ESTABLISHING MONITORING PROGRAMS FOR TRAVEL TIME RELIABILITY Step 2C. Compute Route Travel Times The construction of route travel times, including their distribution from a set of seg- ment-level observations, is significantly more complicated than simply adding travel times together (as can be done carefully when looking only at averages). This compli- cation occurs because correlation exists between adjacent segments. Chapter 3 pre- sented methods to estimate route travel times and their distributions. Infrastructure-Based Data The calculation of average route travel times from infrastructure-based data is rel- atively straightforward as long as attention is paid to both location and time. The process is described in Chapter 3. However, the determination of route travel time distributions from infrastructure-based sensors is difficult due to the thinness of infor- mation about specific vehicles on the network. Two methods are illustrated: ⢠Monte Carlo simulation is used in conjunction with correlation matrices known as link incidence matrices. This method is described in Chapter 3. ⢠Monte Carlo simulation is used in conjunction with a prediction technique using probe vehicles and point queue estimations. This method is illustrated in Chapter 4 using Berkeley Highway Laboratory data. Vehicle-Based Data The construction of route travel time distributions from vehicle-based data is signifi- cantly simpler and can be done directly if enough data are present. The Sacramentoâ Lake Tahoe and New York/New Jersey case studies in Chapter 4 illustrate ways to do this based on data from AVI and AVL systems, respectively. In the absence of sufficient data to construct route-level travel times and distribu- tions directly, the same techniques described for infrastructure-based sensors can be employed. The significant difference is that the segment-level travel times and distribu- tions are based on synthesized values rather than direct observations from AVI or AVL data. STEP 3. CHARACTERIZE OBSERVED TRAVEL TIMES As shown in Figure 5.4, characterization involves labeling each travel time observation based on the operating conditions that were extant when the travel time was observed. An example might be a travel time that was observed late at night under uncongested conditions when there was no nonrecurring event. Another example might be a time that was observed during the heavily congested afternoon peak when an incident had just occurred. A characterization that is particularly effective uses a two-dimensional categoriza- tion based on the nominal congestion level that was occurring at the time the observa- tion was obtained and the nonrecurring event (including none) that was taking place. These pairwise combinations are called regimes. Chapter 3 defined this idea, Chapter 4 provided illustrations in the use cases, and Appendix D supplies additional informa- tion. Using this characterization puts observations into bins under which the operating
142 GUIDE TO ESTABLISHING MONITORING PROGRAMS FOR TRAVEL TIME RELIABILITY Figure 5.4. Characterizing the observed travel times. Measure Data Computation Engine Characterize Identify Understand Decision Sources of Congestion Travel times without incidents Travel times during incidents Travel time during an incident Low travel times High travel times Travel time conditions were very similar. The amount of traffic using the facility is similar because of the level of congestion (essentially a surrogate for the demand-to-capacity ratio), and the nonrecurring event label puts it into a bin with other observations that were obtained under similar conditions. STEP 4. COLLECT, MANAGE, AND IMPUTE NONRECURRING EVENT DATA The power of a TTRMS lies in its ability to connect nonrecurring event data to the travel times observed. These situations often act as sources of congestion, as illustrated in Figure 5.5. Nonrecurring event data come in a variety of forms, as discussed in Chapter 2. The Guide has many examples of placing nonrecurring event data into a TTRMS. In some systems the nonrecurring data are integrated into a single database. An example of this type of system is the Georgia DOTâs Navigator system, presented in Chapter 4 for the Atlanta case study. In other systems, multiple data sources are post- processed to integrate various nonrecurring event sources into a single database. An example of this type of system is the performance measurement system (PeMS) in San
143 GUIDE TO ESTABLISHING MONITORING PROGRAMS FOR TRAVEL TIME RELIABILITY Diego, which integrates incident data from the statewide Traffic Accident and Surveil- lance Analysis System, lane closure data from Caltrans, and weather data from the Automated Weather Observing System reported for San Diego International Airport. These examples are presented in Chapter 4 for the San Diego case study. The integration of nonrecurring event data typically sometimes involves manual intervention to properly integrate and assign event data to the correct time periods. Particular challenges include the actual assignment of incident data to time periods, because the impact of the incident on traffic may extend well beyond the time period during which the actual incident occurs, is reacted to, assessed, and cleared. Similar challenges are involved with weather data. Nonrecurring event data may require imputation to develop consistent reporting of events across the time periods being analyzed. A common example includes the notation of incident events and achieving consistency in the way that incidents are recorded. The Atlanta case study illustrates a more detailed method for coding inci- dents to increase their potential utility for analysis. Figure 5.5. The role of sources of congestion. Measure Data Computation Engine Characterize Identify Understand Decision Sources of Congestion Travel times without incidents Travel times during incidents Travel time during an incident Low travel times High travel times Travel time
144 GUIDE TO ESTABLISHING MONITORING PROGRAMS FOR TRAVEL TIME RELIABILITY STEP 5. IDENTIFY SOURCES OF CONGESTION AND UNRELIABILITY Once route travel time calculations or distributions have been assembled, they can be analyzed in conjunction with nonrecurring event data to identify sources of unreliabil- ity, as illustrated in Figure 5.6. Chapter 3 described two methods of source identifica- tion: a tagging approach and a statistical approach. Tagging Approach The tagging approach involves matching nonrecurring event data with travel time data as the data are collected. Effectively, nonrecurrent event data are captured in real time as the events occur so they can be archived into databases tailored to each event type and then matched to travel time data for analysis and categorization purposes. Recur- ring congestion levels are identified and tagged to the travel time observations by time of day. This allows each observation (e.g., each 5-minute data point) to be tagged with a regimeâa combination of congestion level and nonrecurring event condition. This method is described in Chapter 3 and demonstrated in Chapter 4 for the San Diego and Atlanta case studies. Once the data points have been tagged, distributions are generated for each combination of nonrecurring event and recurring congestion level. Figure 5.6. Identifying sources of unreliability. Measure Data Computation Engine Characterize Identify Understand Decision Sources of Congestion Travel times without incidents Travel times during incidents Travel time during an incident Low travel times High travel times Travel time
145 GUIDE TO ESTABLISHING MONITORING PROGRAMS FOR TRAVEL TIME RELIABILITY Statistical Approach The statistical approach involves two steps. The first step involves identifying outly- ing travel times for which explanations should be sought for the discordant values observed. The second step involves seeking evidence of nonrecurring events that are likely to have caused the unusual travel times. The strength of the method is that in- formation is sought for every event that involves outlying values. Its weakness is that causal events can be missed because they do not produce significant impacts; more- over, it may be difficult to find such information ex post facto because the explanatory data have been lost. Applications of this technique can be seen in Chapter 4 in the San Diego, Northern Virginia, and SacramentoâLake Tahoe case studies to reveal explana- tions for unusual travel times and rates. STEP 6. UNDERSTAND THE IMPACT OF THE SOURCES OF UNRELIABILITY The impact of the sources of unreliability can be analyzed using the distributions de- veloped in Step 5 for specific facilities or routes. Figure 5.7 illustrates this step. Different metrics can be used to quantify the impact of the sources of unreliability. The methodology presented in Chapter 3 uses semivariance measures to identify the reliability impacts of congestion. This methodology is demonstrated using data from Figure 5.7. Understanding recurrent and nonrecurrent causes of congestion. Measure Data Computation Engine Characterize Identify Understand Decision Sources of Congestion Travel times without incidents Travel times during incidents Travel time during an incident Low travel times High travel times Travel time
146 GUIDE TO ESTABLISHING MONITORING PROGRAMS FOR TRAVEL TIME RELIABILITY the San Diego case study. The Atlanta case study uses multistate models to inform on the reliability impacts of nonrecurrent congestion. Both cases studies were presented in Chapter 4. These metrics can be used for various purposes by different types of transportation system users. The use cases in Chapter 4 demonstrated how the metrics can be used to prioritize facilities based on relative impacts and to make planning and programming decisions. Chapter 4 also showed how agency administrators can view the travel time reliability performance of a subarea, how agency planners can determine when a route is unreliable, and how roadway system managers can be alerted when travel time vari- ability exceeds a threshold. STEP 7. MAKE DECISIONS The final step in the analytical process, illustrated in Figure 5.8, is to use the un- derstanding gained from the analysis to make decisions. These decisions vary widely based on the user cases under consideration. A variety of use cases are presented in Chapter 4 to illustrate these potential applications. It was not the purpose of Project L02 to determine what decisions would be most appropriate for specific situations, Figure 5.8. Making decisions. Measure Data Computation Engine Characterize Identify Understand Decision Sources of Congestion Travel times without incidents Travel times during incidents Travel time during an incident Low travel times High travel times Travel time
147 GUIDE TO ESTABLISHING MONITORING PROGRAMS FOR TRAVEL TIME RELIABILITY but it was the intent of the project to ensure that the data provided by the reliability monitoring system can be used for such purposes. CONCLUSION A TTRMS can be a powerful asset in the decision-making process for managing trans- portation facilities. Many key questions can be answered with such a system, including ⢠What is the distribution of travel times in the system? ⢠How is the distribution affected by recurrent congestion and nonrecurring events? ⢠How are freeways and arterials performing relative to performance targets set by the agency? ⢠Are capacity investments and other improvements really necessary given the cur- rent distribution of travel times? ⢠Are operational improvement actions and capacity investments improving the travel times and their reliability?
