Cover Image

Not for Sale



View/Hide Left Panel
Click for next page ( 30


The National Academies | 500 Fifth St. N.W. | Washington, D.C. 20001
Copyright © National Academy of Sciences. All rights reserved.
Terms of Use and Privacy Statement



Below are the first 10 and last 10 pages of uncorrected machine-read text (when available) of this chapter, followed by the top 30 algorithmically extracted key phrases from the chapter as a whole.
Intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text on the opening pages of each chapter. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

Do not use for reproduction, copying, pasting, or reading; exclusively for search engines.

OCR for page 29
29 4.2 Key Dimensions of Data Needs misbehavior and customer complaints, incidents, or accidents. These cases can be investigated using playback (a capability Along with listing uses of AVL-APC data, this chapter seeks of many AVL systems oriented to real-time applications) in to identify the particular data needs for each use, so that peo- which data from the subject trip is viewed as if it were hap- ple involved in AVL-APC system design can better determine pening in real time. If stop or timepoint records are stored in what features are needed to support various analyses. As indi- a database, cases can be investigated more efficiently using cated in Table 4, the needs of various analysis tools can be general query capability. examined along three dimensions: basic record type, data For investigating incidents, it is often helpful simply to deter- detail, and sampling rate. mine whether the bus in question was really there. Several tran- sit agencies report having refuted accident claims by referring 4.2.1 Basic Record Type to archived AVL data. GPS-based location systems can indicate whether a bus was off route. Being able to relate the AVL data to Four basic record types, giving vehicle location at regular operator-initiated event records and to control messages sent intervals, are considered. (For more detail on record types, see by dispatchers can be helpful for some types of investigations. Sections 2.3.1 and 2.3.2.) Polling records are not suited to stop With interstop records, speed and possibly acceleration or timepoint matching and therefore are only suitable for man- information can be extracted from an AVL database; however, ual investigations involving playback. Timepoint records and to date, no agencies are known to sample so frequently as to stop records differ chiefly in level of geographic detail. In addi- give their AVL system the "black box"function of trip recorders tion, stop records, which may or may not include passenger that have become common for accident investigations in avia- counts, are assumed to include the time at which doors opened tion and trucking. and closed, or time of passing a stop if a bus does not stop. For investigating customer complaints, simple playback Timepoint records are assumed only to have either arrival or (or better yet, a database query) can identify whether a bus was departure times. Interstop records are records of speed between very early or late, or (with GPS data) off route. However, stops; the term can also refer to summary data about what because being early is often a matter of only 1 or 2 min, time- occurred between stops that may be part of a stop record. point records are better than polling records for verifying com- plaints about early buses. Also, because an early arrival does not 4.2.2 Data Detail necessarily mean an early departure, the location data should indicate departure time. This second dimension indicates what additional data are Some agencies use their AVL data to investigate operator needed, either as additional items in the basic record type or overtime claims. AVL data has shown, for example, that an captured in infrequent event records. Data detail is mainly operator who reported late to the garage actually finished his affected by what devices are integrated into the AVL system. last trip on time, suggesting that the late pull-in was inten- tional. Speed data from interstop records can be used to mon- 4.2.3 Sampling Rate itor speeding; however, experience has shown that speed data can be quite unreliable. As mentioned earlier, analyses involving estimation of mean values require a relatively small sample, while esti- mating extreme values or proportions requires large sample 4.4 Running Time sizes. Therefore, it makes sense, and is consistent with prac- Analyzing and scheduling running time is one of the rich- tice, to treat this dimension as binary: either 100% (all vehi- est application areas for archived AVL-APC data. Without AVL cles equipped) or 10% (a sample of the fleet equipped), data, agencies must set running times based on small manual noting however that some uses that demand large sample samples, which simply cannot account for the running time sizes on a subset of routes can be accommodated without variability that comes with traffic congestion. instrumenting the entire fleet, if the instrumented percent- Buses are scheduled at the timepoint level; therefore, sched- age is managed carefully. uling demands timepoint data. Because schedules sometimes refer to arrivals as well as departures, it is helpful if timepoint records include both arrival and departure times. 4.3 Targeted Investigations Running time analyses that require only estimation of mean Analysis tools may be used in targeted investigations, which values, or that involve only occasional studies (e.g., delay and may be conducted to support legal, payroll, operations, main- dwell time analysis), can be conducted with only a sample of the tenance, and other functions. For this category, transit agen- fleet equipped with AVL. However, routine scheduling applica- cies use archived AVL data to investigate suspected operator tions based on extreme values need the entire fleet equipped.

OCR for page 29
30 As part of this project, a set of improved analysis tools was such a schedule, a high percentage of trips depart almost developed for analyzing running time. They are mentioned in exactly on schedule, and the low percentage of trips that run this section, but illustrated and further described in Chapter 5. late are not far behind schedule. The amount of slack put into a schedule is often a simple fraction of mean running time, with ad hoc adjustments 4.4.1 Allowed Time, Half-Cycle Time, and based on experience. A more scientific, data-driven approach Recovery Time is to use a percentile value, or "feasibility criterion." To illus- A common analysis examines the distribution of observed trate, a feasibility criterion of 85% means setting allowed running time for scheduled trips across the day compared with time equal to 85th-percentile observed running time; such scheduled running time, also called "allowed time."An example a schedule can be completed on time 85% of the time. This is given in Figure 2, where the vertical bars show mean observed approach is used in Chapter 5, with both the viewpoint of running time, the short lines show 85th-percentile values, and design ("tell me the 85th-percentile running time") and the arrows indicate maximum observed running times. Heavy analysis ("tell me what feasibility I'd get if I added 1 minute horizontal lines show scheduled running time. This figure also to allowed time"). distinguishes net and gross trip time, as explained in Section Analysis of running time is also pertinent for determining 4.4.4. When the number of observations is not too large, a scat- how much recovery time to schedule at the end of the line. terplot showing every observation can be useful. The time from a bus's departure at one terminal to its next Based on the observed distribution of running time for departure in the reverse direction has been called the "half- either a single scheduled trip or a set of contiguous trips in a cycle time"; it is the sum of running time and recovery time. period that will be scheduled as a group, schedule makers can Because the purpose of recovery time is to limit the likeli- choose a value for allowed time according to their preferred hood that delays encountered in one trip will propagate to scheduling philosophy. Some schedule makers prefer to base the next, half-cycle time is based logically on a high-percentile schedules on mean running time. An alternative approach, value of running time. Tri-Met has begun to systematically aimed at improving schedule adherence, is to intentionally revise its half-cycle times, basing them on a 95% feasibility put slack into the schedule; this approach has to be coupled criterion so that there will be only a 5% chance that a bus will with an operating practice of holding at timepoints. With arrive so late that it starts the next trip late. For this applica- Gross and net route section times, mean, 85% and max values Company: Hermes Departure times Dates: 2000/05/01 until 2000/05/26 Trips scheduled: 520 (Count) Line: 1 From: Station NS From: 07:00 Mon Tue Wed Thu Fri Sat Sun Total Trips used: 427 (82%) Route: 1 To: Castilielaan Until: 12:00 4 4 4 4 4 0 0 20 Trips excluded: 2 ( 0%) 14 15 19 17 16 17 16 12 18 18 18 18 16 Count 19 16 15 17 18 15 18 9 18 16 17 18 17 Tritapt 1.0 (b82) license holder is Peter Knoppers, Technische Universiteit Delft. Copyright 1997-2006 TU Delft 25 20 route section time [m] 15 10 5 0 08:00 10:00 12:00 07:00 09:00 11:00 time [hh:mm] Source: Hermes (Eindhoven), generated by TriTAPT Figure 2. Observed running time by scheduled trip.

OCR for page 29
31 tion, scheduled recovery time is set to be the difference between delays and running time variability occur. Second, it tends to 95th-percentile running time and allowed time. put less slack in the early part of the route and more in the later The chief engineer for operations analysis at Brussels' tran- part of the route, which is a better way to distribute slack than sit agency has recently been working with a scheduling software simply applying it proportionately throughout. It results in vendor to develop reports to support statistically based sched- holding buses less in the early part of the route because they ules. Their approach uses three parameters, as explained in this may need that time later in the route. example: If the three parameters are 95%, 80%, and 5 min, then Steve Callas, the manager of Tri-Met's AVL data analysis half-cycle time is set equal to 95th-percentile running time, and program, has suggested that if operating practice is such that allowed time is set equal to 5 min less than 80th-percentile run- buses do not hold at timepoints, it may be better to use a low ning time. That way, there will be an 80% chance that a trip fin- feasibility criterion because running early is more harmful to ishes no more than 5 min late, and a 95% chance that the next passengers than running late. For example, a statistically trip can start on time. based approach for segment-level scheduling might be to base departure time at a timepoint on 40th-percentile observed cumulative running time (i.e., from the start of the line to a 4.4.2 Segment Running Time point). That way, there will be only a 40% chance of a bus Scheduling running time on segments, or equivalently set- departing early, and those that depart early should not be very ting departure (or arrival) times at timepoints relative to the far ahead of schedule. trip start time, can either precede or follow scheduling route times. One approach, common in U.S. practice, is to first 4.4.3 Choosing Homogeneous determine segment running time based on mean observed Running Time Periods running time by segment, perhaps adding a certain percent- age for slack, and then constructing route time as simply the Another problem in running time analysis is choosing the aggregation over the segments. boundaries of running time periods within which allowed time Simply aggregating over segments will not work with a is constant. Establishing periods of homogeneous running feasibility-based approach, because the sum of the parts will time involves a trade-off between short periods within which not yield a valid measure for the whole. For example, the scheduled running times match the data well versus longer sum of 85th-percentile segment running times does not periods of constant allowed time but greater variability. A com- equal a route's 85th-percentile running time; the sum will be mon logic for resolving this trade-off is first to determine, for far greater, in fact. each scheduled trip, an ideal allowed time (e.g., mean running The Delft University of Technology has developed the Pass- time, or 85th-percentile running time, depending on the ing Moments method of extending statistically based schedul- desired feasibility criterion) and then to make running time ing to the segment level (31). Applied at several Dutch transit periods as long as possible subject to the restriction that no agencies, the Passing Moments method bases timepoint sched- more than a certain percentage of the scheduled trips in that ules on f-percentile completion time from each timepoint to period have an ideal running time that deviates by more than the end of the line, where completion time is running time given tolerance from the suggested running time for that period. from a point to the end of the line, and f is the feasibility cri- terion (e.g., 85%). Segment running times are determined by 4.4.4 Excluding Holding (Control) Time working backwards from the end of the line, without ever explicitly analyzing observed running time on timepoint-to- Ideally, scheduling tools should use net running time, which timepoint segments. This approach was designed to overcome excludes holding time, also called "control time." Identifying operators' resistance to holding, which is the key to good sched- what part of observed running time is holding time can be ule adherence. If a schedule is written based on mean running tricky, requiring greater data detail, and is done by only a few time, operators know that, if they hold at a timepoint, they will agencies with AVL-APC data. Ideally, stop records should indi- have a 50% chance of finishing the trip late and thereby getting cate both when doors open and close, and when the wheels a shortened break; therefore, they are reluctant to hold. With start to roll, something provided by at least some APC vendors. 85th-percentile allowed times between each timepoint and the If the bus is ahead of schedule, any unusual gap between door end of the line, operators know that even if they hold, they have close time and departure time can be interpreted as holding. a high chance of finishing on time. The running time analysis shown earlier in Figure 2 distin- The Passing Moments method is one of the scheduling tools guishes gross from net running time. The gray bars show net described in Chapter 5. The Passing Moments method has two running time; their black tops are control time, making the advantages compared to setting slack time simply proportional combined height equal to gross running time. Of the short to mean running time. First, it is sensitive to where on the route horizontal lines representing 85th-percentile values, those

OCR for page 29
32 extending to the left of a scheduled start time are gross running Signal priority is another application that needs stop-level times, and those extending to the right are net running times. schedules, if priority is conditional on buses being late. Con- Because in good weather holding may occur with the doors ditional priority is a form of operational control, pushing open, NJ Transit's system is being upgraded to provide a record late buses ahead while holding early buses back, and to be of how many passengers boarded and alighted every few sec- effective, it needs a finely tuned schedule at every signalized onds (the main on-board computer frequently queries the on- intersection. This system has been used very effectively in board APC analyzer while doors are open and writes records), Eindhoven (32). Conditional priority without fine-tuned allowing NJ Transit to recognize periods of inactivity even schedules can easily devolve into either unconditional prior- while the doors are open. Eindhoven's algorithm looks for ity (because buses always arrive late) or no priority (because unusually long dwell times when the bus is ahead of schedule. buses arrive early). To have a good control margin, schedules Event codes for lift use or a fare dispute can also help iden- have to be written so that the probability of arriving late is far tify holding time by explaining the cause for long dwell times. from the extremes of 0% and 100%. Another version of holding is still harder to detect--"killing time" en route to avoid being early. Data and algorithms that 4.4.6 Speed and Delay Analysis would help detect "killing time" would be valuable and are being developed at NJ Transit. Speed, delay, and dwell time studies are analyses that help support a transit agency's efforts to improve commercial speed, something that benefits both operations and passengers. 4.4.5 Stop-Level Scheduling "Speed" in this context is average speed over a segment, not In many European countries including the Netherlands, instantaneous or peak speed. A display such as given in Fig- schedules are written at the stop level. In fact, on many routes, ure 3 showing delay by segment (or, alternatively, average speed every stop is a timepoint. Making (almost) every stop a time- by segment) helps a transit agency to identify problem loca- point has the advantage of replacing occasional large holding tions, to monitor the impacts of actions that affect speed, and actions with frequent small ones, which are less obvious and to monitor and document historic trends in operating speed. irritating to passengers. In that figure, the thin horizontal lines are individual observa- Stop-level schedules fit well with the trend of giving cus- tions of delay by segment; the box height is the 85th-percentile tomers better information. Stops are where customers meet the delay, and the bar inside the box indicates mean delay. Analysts system, and where they need to know scheduled departure will be interested not only in average delay, but also in how vari- times. Internet-based trip planners need stop-level schedules, able it is, and in the likelihood of extreme values. as do real-time next-arrival systems (to know stop-to-stop A report showing delays or speeds between stops offers a expected running time). In current practice, agencies with cus- richer, more geographically detailed view than one using tomer information systems like trip planners and real-time timepoint segments. Another reason to prefer stop records as information systems estimate stop-level departure and run- the basis of delay analysis is that it allows dwell time and con- ning time by interpolation between timepoints; using stop- trol time (which almost always occur at stops) to be removed, level AVL data to develop stop-level schedules offers an obvious which puts a clearer focus on the effects of the roadway and improvement. Not providing the public with stop-level sched- traffic on bus speed and delay. ules is a good example of practice being driven by the historic "Delay" can be defined in several ways. Two definitions of lack of data--practice that should change as advanced tech- delay are (1) the travel time between stops minus the average nology is deployed in transit. travel time measured during non-congested periods such as Stop-level schedules are valuable for control even when early morning or late evening and (2) the amount of time every stop is not a timepoint. Some AVL systems display spent at speeds below 5 km/h minus the time spent at stops. schedule deviation to the operator, who can use this infor- (Eindhoven uses this second definition.) To support this def- mation all along the route to try to adjust bus speed. For inition of delay, an AVL system needs records of when speed example, in Eindhoven a small display for operators shows thresholds are crossed. schedule deviation in units of 10 s. Such a system can be effective, however, only if it is based on a realistic, finely 4.4.7 Dwell Time Analysis tuned schedule. Simply interpolating between timepoints is too approximate if speed between timepoints is not uni- Transit agencies also try to improve commercial speed by form, such as when the route passes major intersections. For reducing dwell time, using such measures as low-floor buses operators to have confidence in, and therefore use and ben- or changes to fare collection equipment and practices. Stop efit from, a schedule deviation display, they need data-based records with door open and close times allow agencies to ana- stop-level schedules. lyze dwell time to determine impacts and trends.