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.

66 One of the hard lessons learned is that off-line analysis has different data needs than real-time monitoring and that, therefore, AVL systems designed for real-time monitoring may not deliver the type and quality of data needed for off- line analysis. Considering the ways off-line data can be used to improve operations and management, and considering the way AVL system design affects what data is captured as well as its quality, this chapter presents findings related to AVL system design. The emphasis of this chapter is on AVL rather than APC system design because APCs have always been designed for off-line analysis. To a large extent, this chapter summarizes Furth et al. (6). 10.1 Off-Vehicle versus On-Vehicle Data Recording There are two options for recording data with AVL systems: on-vehicle (in an on-board computer) or off-vehicle (by sending data messages via radio to a central computer). As discussed in Section 2.3, the radio channel capacity limits an AVL system’s ability to record data over the air. On-vehicle data storage is clearly superior in that it presents no effective limit on data recording. Where radio-based systems are still contemplated, buyers must learn the capacity of any proposed system to collect timepoint and/or stop data, along with random event data. Capacity depends on the number of radio channels available, the number of buses instrumented, message length, and specifics of the technology used. The number of radio chan- nels available to a transit system is strictly limited and varies by location, because radio channels are allocated by govern- ment. Radio-based systems have been successfully configured to record useful data for off-line analysis; for example, Metro Transit’s system makes timepoint records from all buses and stop records from about 15% of the fleet by sending messages over the air to a central computer. 10.2 Level of Spatial Detail As discussed in Sections 2.3.1 and 4.2.1, the choices in spatial detail of basic AVL records are polling records (occur- ring at arbitrary locations, when the bus is polled), time- point records, and stop records. Collecting data at a finer level is also possible. 10.2.1 Time-at-Location and Location- at-Time Data Polling data can be characterized as location-at-time data, giving bus location at an arbitrary time; stop and timepoint records, on the other hand, have time-at-location data, giving the time at which a bus arrives or departs from a specific loca- tion. Most off-line analyses, including analyses of running time and schedule adherence, require knowledge of departure time from standard locations. Therefore, stop and timepoint records are inherently better suited to off-line analysis of AVL data. Theoretically, time-at-location could be estimated from polling data by interpolation. This method introduces inter- polation errors, whose magnitude can be almost as large as the polling interval on segments in which the plausible bus speed has a wide range (e.g., because of intermittent traffic congestion). During periods of traffic congestion, it can be difficult with polling data to determine whether a bus report- ing coordinates close to a stop is in a queue waiting to reach the stop, is at the stop, or has already left the stop and is wait- ing in a traffic queue. The project survey did not find a single case of a transit agency routinely using polling data for off-line analysis except for playback to investigate incidents. Researchers used such a data stream from Ann Arbor for some operational analyses (41), but the process of going from raw poll messages to trajectories matched to route and schedule was too involved to become routine. The three case study agencies with round- robin polling data do not use it off line except for incident C H A P T E R 1 0 Designing AVL Systems for Archived Data Analysis

investigation using playback. Also, the survey indicates that all of the traditional AVL suppliers, even if they still use polling to support real-time applications, include timepoint records in their data streams as well. However, a relatively new entry to the market, stand-alone “next arrival” systems, uses only polling data to track bus loca- tion. Like earlier AVL systems, they are designed around a real- time application and, according to the interviewed vendor, use polling data to minimize the amount of equipment installed in the vehicles, making such systems less expensive. This vendor claims to have obtained good test results using the data from its system for off-line analysis of on-time performance. The next arrival system’s data stream includes predicted arrival time at stops (based on proprietary algorithms); as buses get close to a stop, predicted arrival time should become a rather accurate measure of actual arrival time, especially if the polling cycle is short, and therefore might be used as an approximation. A drawback of next arrival systems is that, while their application focuses on arrival time, most running time and schedule adherence analyses are concerned with departure time. 10.2.2 Timepoint versus Stop-Level Data Recording Given that obtaining time-at-location data is important, what location detail is needed: stop level or timepoint level? Of course, stop-level data is needed for passenger counts; but, for operations analysis, what is the incremental value from getting data at all stops as well as at timepoints? Because scheduling practice in the United States is based on timepoints, timepoint data is all that is needed for traditional running time and schedule adherence analyses. Metro Transit’s AVL-APC system design emphasizes this distinction: on buses with APCs, stop records are created; while, on buses with only AVL, only timepoint records are created. Timepoint data tends to be favored by systems that rely primarily on radio transmis- sion for data recording, because timepoint messages do not consume much radio channel capacity–timepoint messages are not very frequent and tend to be rather short, including only timepoint ID, time and location stamp, and identifiers. (Inter- estingly, this issue does not arise in the Netherlands, because almost every stop is a timepoint there. Also, stop spacing in the Netherlands tends to be about 60% greater than in the United States, resulting in fewer stops.) However, stop-level detail offers advantages to a transit agency willing to go beyond traditional scheduling and oper- ations analyses. Those advantages stem from (1) finer geo- graphic detail for operations analysis and planning, and (2) the fact that stops are where customers meet the system, making stops a natural unit for customer-oriented schedul- ing and service quality analyses. In Section 4.4.5, several advantages of stop-level scheduling were cited, including bet- ter customer information (both off line, as part of trip plan- ning, and in real time, for predicting arrival time), finer con- trol, and ability to apply conditional (schedule-based) signal priority. Stop-level analysis helps enable operations analysis to identify points of delay, determine the impact of changes to stop location or traffic control, and analyze bunching. Stop-level data also permits more customer-oriented service quality analysis, such as enabling determination of passenger waiting time at stops. With stop-level data, bus arrival and departure times are easier to determine at the end of the line. If the last segment’s data is unreliable, what is lost becomes much smaller with stop-level data. The practice of making timepoint, but not stop, records appears to be partly a relic of past practice, partly a limitation of radio-based data communication, and partly a simplifica- tion (for example, an agency with timepoint records only has to make sure its timepoints are mapped correctly, not all its stops). In today’s technology age, with on-board data storage possible at relatively little cost, there seems little reason to set- tle for less than stop-level data. 10.2.3 Interstop Data Automatically collected data on what happens between stops is not nearly as important as data about stops. How- ever, there is nearly no marginal cost to making interstop records, which can support some useful applications. Exam- ples include monitoring maximum speed (both as a check for speeding and as a measure of quality of traffic flow), moni- toring time spent below crawl speed (as a measure of delay), and treating the bus as a GPS probe for mapping bus paths. Another possibly valuable use, mentioned earlier, is to analyze operations at terminals to help better determine actual arrival and departure times. One possible configuration, applied at NJ Transit, permits records at regular, user-set intervals; for mapping a bus’s path, the interval can be made quite small. Eindhoven’s configura- tion, in which a record is made whenever speed crosses a crawl-speed threshold, can be generalized. By using a few dif- ferent thresholds, users could estimate not only delay (time spent below crawl speed), but also a speed profile, which might be used to characterize traffic quality or to monitor speed in different speed zones. Tri-Met’s configuration, in which only maximum speed between stops is recorded, is partly a concession to limited on-board data storage (which was an important factor in the mid-1990s). Given the current availability of low-cost on-board data storage, frequent inter- stop records can easily be accommodated. However, until now, the value of much interstop detail was not yet proven. Frequent interstop records detail can aid in matching. For instance, speed records may help resolve situations such as 67

when a bus stops twice at the same stop, is jockeying around at a layover, or holds (to avoid running early) away from a stop. NJ Transit is interested in using frequent interstop records for improving maintenance management by correlat- ing operations measures with maintenance needs, particu- larly if future generations of its data collection system can integrate data from the vehicle drivetrain system. 10.2.4 Arrival/Departure Time Accuracy AVL systems vary widely in the data captured with respect to arrival and departure time at stops and timepoints. Some off-line analyses need arrival time, some departure time, and some both. Therefore, an AVL system will be more valuable if it detects and records both arrivals and departures. Door sen- sors and the recording of door open and close events help improve the accuracy of arrival and departure detection as well. For example, suppose a bus stops two or more times in the neighborhood of a stop. Was the first the stop and the sec- ond simply traffic delay, or was it the reverse? Or did the bus open it doors both times, so that arrival time should be taken from the first stop and departure time from the second? Without a door sensor, arrival is frequently detected by a vehicle entering a 10-m radius zone around a stop. Around major stops and terminals, the zone can be quite a bit larger, which can distort arrival time if a bus faces congestion getting to the stop (e.g., because a traffic queue or another bus is block- ing the stop). Likewise, zonal detection can distort departure time if departing buses encounter congestion before leaving the stop zone, which can happen at near-side stops, at stops where buses have to await a gap or yield to crossing pedestrians before entering the traffic stream, and at terminals. Knowing both when doors close and when the bus actually departs is valuable for detecting holding, which is important for running time analysis. Because of the possibility of hold- ing, door close events are not sufficient to determine depar- ture time. Therefore, while door sensors are valuable, they are not sufficient. (In many Latin American cities, where buses routinely operate with doors open, they are almost useless.) 10.2.5 Route Endpoint Identification As mentioned earlier, many AVL systems are weak in deter- mining when a bus arrives and departs a route terminal. For running time and schedule adherence analyses, these data items are critical, and system features that make their correct identification easier are valuable. Such system features include door open and close records, frequent interstop records in ter- minal regions, odometer-based records to supplement GPS- based records in terminal regions, and better algorithms for interpreting bus movements in terminal areas in order to bet- ter distinguish genuine departures from movements within a terminal area. When route terminals are located in zones with poor GPS reception (downtown or a covered terminal), data from supplemental devices and logic to interpret it are espe- cially valuable. 10.3 Devices to Include Integrating other devices in an AVL or APC system can add value either because the data those devices provide is inher- ently valuable or because of synergies that make the new data helpful for interpreting other AVL data. Door sensors have already been mentioned for their value to help match location, determine arrival and departure times, and identify control time. APC systems virtually always include door sensors; their inclusion in AVL systems would be a benefit as well. Odometer (transmission) data is helpful for determining bus speed, which can be valuable in its own right, and can be used to determine when a bus departs from a stop and when/where buses are delayed in traffic. Most AVL systems have odometer connections as a backup to GPS or to determine distance traveled between signposts. Gyroscopes add richness to the data provided by odome- ters, allowing vehicles to be tracked off route and permitting matching based on turning locations. Recording fare transactions in the AVL data stream is a means of getting location-stamped boarding data, which can be especially valuable to a transit system lacking APCs. When the payment medium is electronic and therefore offers an ID unique to the passenger, location-stamped fare transaction data offers the further opportunity for inferring link-trip and transfer information. However, there has been little experi- ence to date with location-stamped fare records; this area has considerable opportunity for research and development. Integrating the radio system’s control head offers the opportunity to capture records of sign-in data, valuable for matching, and of operator-observed or -initiated events including pass-ups, special passengers (e.g., wheelchair and bicycle users, fare evaders), and traffic events (e.g., draw- bridge up). These event records provide data that is valuable in its own right for direct analysis. Such records can also add detail and accuracy matching to running time and service analyses, for example, by helping to confirm and perhaps explain long delays or indicating when the “everybody boards the first bus” assumption behind waiting time calcu- lations is violated. While radio-based systems always benefit from this connection, it would also benefit passenger count- ing and event recording systems. Also of potential value, but not yet applied (to the researchers’ knowledge), would be coded records of standard radio messages initiated by the control center, such as instructions to hold for a connecting passenger. 68

Integration with the wheelchair lift (or lift sensors) would provide a more accurate and automated record of lift use than relying on operators to initiate a message. Integration with the destination sign might prove useful to help with matching. Integration with a stop announcement system or next arrival system does not bring in any new data; however, it cre- ates an incentive for stop matching to be accurate and thereby benefits AVL-APC data analysis. AVL systems have long attempted to use data from the vehi- cle’s mechanical system (in addition to that from the odome- ter), such as oil temperature and air pressure. In real time, alarms from these systems have delivered so many false posi- tives that they tend to be ignored. Whether the recording and off-line analysis of mechanical data integrated with location data can deliver new insights on mechanical performance is a rich area for further exploration. 10.4 Fleet Penetration and Sampling AVL systems, when installed, are usually installed on the entire fleet. APCs have traditionally been installed on about 10% to 15% of the fleet. Chapter 9 discusses how fleet penetration affects sample size, and what sample size need is for passenger count–related data items such as load, boardings, and passenger-miles. The general principle is that if the data is used only to determine mean values, small samples are sufficient; however, when extreme values are important, a complete or at least large sample is preferred. For passenger count data, 10% penetra- tion is more than enough for boardings and passenger-miles data, for which only mean values are needed. For load on crowded bus routes, a near-complete sample is desirable so that extreme values can be observed. With a small fraction of the fleet instrumented, large sample sizes can still be obtained from crowded routes if the instrumented buses are dispro- portionately allocated to crowded routes. On the side of operations data, only near-100% penetra- tion will provide the large sample sizes needed to determine extreme values used in statistically based running time analy- sis and design. With a small fraction of the fleet equipped (as in a traditional APC system), large sample sizes can be obtained by aggregating over time, but at the risk of some of the data being out of date. Headway analysis requires 100% instrumentation on a route at a given time, which can only be achieved with either 100% penetration or careful allocation of the instrumented sub-fleet. Another benefit of full coverage is the ability to investigate complaints. As many complaints arise from extreme events (long waits, overcrowding), full coverage would be most help- ful in such investigations. Many transit agencies report that managing the allocation of an instrumented sub-fleet can be a large headache and that concerns other than data collection (e.g., who gets the new buses) often control the allocation, frustrating data col- lection plans. A major motivation for instrumenting only a fraction of the fleet with APCs has been their cost. Tri-Met has shown that by integrating APCs with an AVL system, the incremental cost of an APC can be reduced to the $1,000 to $3,000 range; Tri-Met now treats them as standard equipment included in all new bus purchases. Note that Tri-Met uses a rather simple-technology APC and that more complex and (presumably) accurate APCs may not offer such an attractive incremental cost. 10.5 Exception Reporting versus Exception Recording Exception reporting is certainly a valuable management tool, available for use with any AVL-APC data archive. It should be distinguished, however, from “exception record- ing,” the practice of only recording a bus’s location if it is off schedule or off route. This protocol of exception recording does not permit analysis of normal operations and should therefore be avoided. 69