Using Archived AVL-APC Data to Improve Transit Performance and Management (2006)

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.

79 Archived AVL-APC indeed holds great potential for improv- ing transit management and performance. This report has reviewed the history and state of the practice in AVL-APC data collection and analysis. System design affects in many ways the type and quality of data captured, which in turn affects the types of analyses that the data can support. To develop guid- ance for system design, existing and potential analyses and tools that use AVL-APC data to improve management and per- formance are reviewed and their data needs analyzed. Analy- sis tools for running time, waiting time, and crowding were developed in the course of this project. AVL systems have traditionally been designed primarily for real-time applications. For investigating specific incidents, archives of almost any AVL data stream can be used in play- back mode. However, for analysis of historical data combining multiple days of observations, data requirements substantially exceed what will suffice for real-time monitoring. The follow- ing are the main conclusions regarding how AVL-APC systems should be designed to provide a valuable data archive: • Storing data on board frees the system from the capacity restrictions of transmitting data records over the air. • Time-at-location data (i.e., stop and timepoint records) is needed for analyses that aggregate over multiple days of observation, such as running time and schedule adherence analysis. Location-at-time (i.e., polling) data is not suitable. • Stop-level records permit running time analysis and sched- ule making at greater geographic detail than timepoint records, better serving passenger information needs and supporting better operational control. • Integrating on-board devices adds information to the data stream that can be valuable in its own right as well as aid in matching captured data to the base map and schedule. The most valuable devices to integrate are door sensors, odome- ter (transmission), and radio control head. • Designers should pay attention to a data collection and processing system’s ability to accurately determine arrival and departure times at stops, identify holding, and deal with multiple apparent stops and starts at bus stops and in terminal areas. Having data from both door sensors and odometers is particularly valuable in this respect. • Integrating AVL with the fare collection system offers a potentially powerful means of measuring ridership pat- terns, because matching fare media serial numbers offers a means of observing linked trips. Automatic data collection can revolutionize schedule plan- ning and operations quality monitoring as agencies shift from methods constrained by data scarcity to methods that take advantage of data abundance. The large sample sizes afforded by automatic data collection allow analyses that focus on extreme values, which matter for schedule planning (e.g., how much running time and recovery time are needed, what head- way is needed to prevent overloads) and service quality mon- itoring (e.g., how long must passengers budget for waiting, how often do they experience overcrowding). Stop-level data recording provides a basis for stop-level scheduling, a practice with potential for improved customer information and better operational control. With AVL-APC data, trends can be found that might otherwise be hidden, such as operator-specific ten- dencies and sources of delay en route. Regularly analyzing AVL data gives a transit agency a tool for taking greater control of its running times by offering a means of detecting causes of delay and evaluating the effectiveness of countermeasures. Two sets of analysis tools were developed as part of this project. One uses running time data to suggest periods of homogeneous running times, analyze user-selected running time periods and scheduled running times, and create stop- or timepoint-level schedules. It includes a valuable “what-if” tool that allows schedule planners to propose a scheduled running time period and running time, and immediately see how that running time would have performed based on the historical data. The tool offered for segment-level run- ning times uses a statistical approach that, if combined with C H A P T E R 1 3 Conclusions

operational control in the form of holding early trips, has great potential to improve on-time performance. These run- ning time tools are part of the software package TriTAPT, developed at the Delft University of Technology, and are avail- able with no license fee to U.S. and Canadian transit agencies through the end of 2009. New tools were also developed on a spreadsheet platform to evaluate waiting time and crowding from the customer’s perspective using AVL and APC data. Unlike traditional meth- ods, they focus on the extreme events (e.g., very early and late buses, very long headways, very crowded buses) that most affect customer satisfaction. A whole new framework was developed for evaluating passenger waiting time, one that gives attention to the time that passengers have to budget for waiting, not just the time they actually spend waiting. Three new measures of waiting time are proposed: budgeted waiting time, potential waiting time, and equivalent waiting time, the latter being a comprehensive summary of passengers’ waiting cost. This framework is superior to traditional measures of waiting time because it accounts for the impact of service unreliability on passenger waiting time. The new crowding measures developed are defined from the passenger rather than vehicle perspective. With APC data, the percentage of passengers at the maximum load point who sit and who stand can be inferred, and they can be further divided into those who do and do not sit next to an unoccu- pied seat, and those standing at various levels of crowding. The result is a distribution of passengers by crowding experi- ence, something that should correlate well with passenger complaints and with service objectives. Routine use of automatic passenger counts poses a special challenge because imperfect counting accuracy requires trip- level parsing and balancing both for consistency and to avoid drift errors. Accuracy measures are described and analyzed. The researchers show, both from empirical data and from log- ical considerations, that accuracy in measured load (and, by extension, in passenger-miles) can be significantly worse than accuracy in on-off counts; therefore, load accuracy should be an important consideration in system specification, design, and testing. Shortcomings in several existing methods of load balancing are pointed out; some of them can bias load and passenger- miles estimates upwards. A method of trip-level balancing is presented that prevents not only negative departing load, but also negative through load, a stronger feasibility criterion. Approaches and data structures are described for dealing with end-of-line passenger attribution, especially on routes ending in loops and on interlined routes in which passengers can be inherited from one trip to the next. On routes ending with short loops (short enough that it can be assumed that no passenger trips both begin and end in the loop), a parsing and balancing method was developed that makes the loop effec- tively serve as a zero-load point. On routes lacking a natural zero-load point, such as downtown circulators or routes with loops on both ends, APC systems may need operator input to fix the load at a key point on each round trip. The typical fleet penetration rate (10% to 15%) for APCs is shown to be adequate for all passenger count applications except for monitoring extreme crowding. The sample sizes typically afforded by APCs are shown to be sufficient to sat- isfy NTD requirements for passenger-miles reporting. Statisti- cal requirements on systematic error in load or passenger-miles measurements to meet NTD reporting requirements are also given. Analysis procedures using AVL-APC data should account for variable sampling rates by aggregating and weighting observations based on the schedule. Simply aggregating over all the stop or timepoint records for a chosen route, period, and date range can bias results in favor of trips that were measured more frequently. While some analyses involve simple aggregation over a selected set of records, others require that analysis follow a particular sequence of stops. For such an analysis to involve data from multiple patterns operating on a common trunk, a special data structure is required to align the patterns. Such a data structure, called the “virtual route,”was developed in Tri- TAPT; it permits analysts to see, for example, headway analy- sis and passenger load profiles on trunks shared by multiple routes and patterns. The development of AVL-APC data collection and analysis capability poses numerous organization challenges. Perhaps the greatest is raising awareness within the organization of the value of archived AVL-APC data in order to ensure that AVL systems, which are often procured to serve real-time applica- tions, have a design and the database support needed to achieve their potential for archived data analysis. 80