Webinar: Making use of automated data collection to improve transit effectiveness
1. Making Use of Automated Data Collection
to Improve Transit Effectiveness
Presented by
Nigel Wilson
Professor of Civil and Environmental Engineering
Massachusetts Institute of Technology
Bus Rapid Transit Centre of Excellence
Webinar
January 25, 2013
2. OUTLINE
• Key Automated Data Collection Systems (ADCS)
• Key Transit Agency/Operator Functions
• Impact of ADCS on Functions
• Traditional Relationships Between Functions
• State of Research/Knowledge
• Examples of Recent Research with ADCS in London
• OD Matrix Estimation
• Service Reliability Metrics
• Train Load Inference
Nigel Wilson MIT 2
January 25, 2013
3. Transit Agencies are at
a Critical Transition
in Data Collection Technology
Manual Automatic
• low capital cost • high capital cost
• high marginal cost • low marginal cost
• small sample sizes • large sample sizes
• aggregate • more detailed, disaggregate
• unreliable • errors and biases can be estimated
and corrected
• limited spatially and temporally • ubiquitous
• not immediately available • available in real-time or quasi real-
time
Nigel Wilson MIT 3
January 25, 2013
4. Automated Data Collection Systems
• Automatic Vehicle Location Systems (AVL)
• bus location based on GPS
• train tracking based on track circuit occupancy
• available in real time
• Automatic Passenger Counting Systems (APC)
• bus systems based on sensors in doors with channelized passenger
movements
• passenger boarding (alighting) counts for stops/stations with fare barriers
• train weighing systems can be used to estimate number of passengers on
board
• traditionally not available in real-time
• Automatic Fare Collection Systems (AFC)
• increasingly based on contactless smart cards with unique ID
• provides entry (exit) information (spatially and temporally) for individual
passengers
• traditionally not available in real-time
Nigel Wilson MIT 4
January 25, 2013
5. ADCS - Potential and Reality
Potential
• Integrated ADCS database
• Models and software to support many agency decisions using ADCS
database
• Providing insight into normal operations, special events, unusual weather,
etc.
• Provide large, long-time series disaggregate panel data for better
understanding of travel behavior
Reality
• Most ADCS systems are implemented independently
• Data collection is ancillary to primary ADC function
• AVL - emergency notification, stop announcements
• AFC - fare collection and revenue protection
• Many problems to overcome:
• not easy to integrate data
• requires substantial resources
Nigel Wilson MIT 5
January 25, 2013
6. Key Transit Agency/Operator Functions
A. Off-Line Functions
• Service and Operations Planning (SOP)
• Network and route design
• Frequency setting and timetable development
• Vehicle and crew scheduling
• Performance Measurement (PM)
• Measures of operator performance against SOP
• Measures of service from customer viewpoint
Nigel Wilson MIT 6
January 25, 2013
7. Key Transit Agency/Operator Functions
B. Real-Time Functions
• Service and Operations Control and Management (SOCM)
• Dealing with deviations from SOP, both minor and major
• Dealing with unexpected changes in demand
• Customer Information (CI)
• Information on routes, trip times, vehicle arrival times, etc.
• Both static (based on SOP) and dynamic (based on SOP and
SOCM)
Nigel Wilson MIT 7
January 25, 2013
8. Impact of ADCS on Functions
IMPACT ON SOP
• AVL: detailed characterization of route segment running times
• APC: detailed characterization of stop activity (boardings, alightings, and
dwell time at each stop)
• AFC: detailed characterization of fare transactions for individuals over
time, supports better characterization of traveler behavior
IMPACT ON PM
• AVL: supports on-time performance assessment
• AFC: supports passenger-oriented measures of travel time and reliability
IMPACT ON SOCM
• AVL: identifies current position of all vehicles, deviations from SOP
IMPACT ON CI
• AVL: supports dynamic CI
• AFC: permits characterization of normal trip-making at the individual level,
supports active dynamic CI function
Nigel Wilson MIT 8
January 25, 2013
9. Relationships Between Functions
• Real-time functions (SOCM and CI) based on
• SOP
• AVI data
• Reasonable as long as SOP is sound and deviations from it
are not very large
• Fundamentally a static model in an increasingly dynamic
world
Nigel Wilson MIT 9
January 25, 2013
10. State of Research/Knowledge in SOCM
• Advances in train control systems help minimize impacts of
routine events
• Major disruptions still handled in individual manner based
on judgement and experience of the controller
• Little effective decision support for controllers
• Models are often deterministic formulations of highly
stochastic systems
• Simplistic view of objectives and constraints in model
formulation
• Substantial opportunities remain for better decision support
Nigel Wilson MIT 10
January 25, 2013
11. State of Research/Knowledge in CI
• Next vehicle arrival times at stops/stations well developed
and increasingly widely deployed
• Pre-trip journey planner systems widely deployed but with
limited functionality in terms of recognizing individual
preferences
• Strongly reliant on veracity of SOP
• Ineffective in dealing with major disruptions
Nigel Wilson MIT 11
January 25, 2013
12. Evolution of Customer Information
• Operator view --> customer view
• route-based --> OD-based
• Static --> dynamic
• based on SOP --> based on SOP modified by current system
state and control actions
• Pre-trip and at stop/station en route
• Generic customer specific customer
• Request-based systems anticipatory systems
• Agency/operator developed systems "app"
developers using real time data feeds from agency
Nigel Wilson MIT 12
January 25, 2013
13. Key Functions
Service and Operations
Planning (SOP)
Off-line Functions
Performance
Measurement (PM)
ADCS System Monitoring, ADCS
Supply Analysis, and Prediction Demand
Service
Management Customer
(SOCM)
Real-time Functions Information (CI)
Nigel Wilson MIT 13
January 25, 2013
14. Real Time Functions
Vehicle Locations ADCS Loads
Incidents/Events
Monitoring
CONTROL CENTER
Dynamic Estimation of current Information
Supply rescheduling conditions - travel times
Demand
- paths
Prediction
Nigel Wilson MIT 14
January 25, 2013
15. Examples of Recent Research
with ADCS in London
• OD Matrix Estimation
• Reliability metrics
• Individual train load inference
Nigel Wilson MIT 15
January 25, 2013
16. Public Transport OD Matrix Estimation
Function: Service and Operations Planning
Objective: Estimate passenger journey OD matrix at network
level
Network attributes:
• multi-modal rail and bus systems
• entry fare control only and/or entry+exit fare control
Source:
"Intermodal Passenger Flows on London’s Public Transport Network: Automated Inference of Full Passenger
Journeys Using Fare-Transaction and Vehicle-Location Data. Jason Gordon, MST Thesis, MIT (September 2012).
Nigel Wilson MIT 16
January 25, 2013
17. Trip Chaining: Basic Idea
Each AFC record includes:
• AFC card ID
• transaction type
• transaction time
• transaction location: rail station or bus route (time-matching with
AVL data)
B (locB, timeB) D (locD, timeD)
A (locA, timeA) C (locC, timeC)
The destination of many trip segments (TS) is close to the origin of the
following trip segment.
Nigel Wilson MIT 17
January 25, 2013
18. Trip-Chaining Method for OD Inference
Key assumptions for destination inference to be correct:
• No intermediate private transportation mode trip segment
• Passengers will not walk a long distance
• Last trip of a day ends at the origin of the first trip of the day
Nigel Wilson MIT 18
January 25, 2013
19. Trip-Chaining Method for OD Inference
• Infer start and end of each trip segment for individual AFC
cards
• Link trip segments into complete (one-way) journeys
• Integrate individual journeys to form seed OD matrix by
time period
• Expand to full OD matrix using available control totals
• station entries and/or exits for rail
• passenger entries and/or exits by stop, trip, or period for bus
• train load weight data
Nigel Wilson MIT 19
January 25, 2013
20. Summary Information on London Application
• Off-line tool developed
• Oyster fare transactions/day:
• Rail (Underground, Overground, National Rail): 6 million (entry & exit)
• Bus: 6 million (entry only)
• For bus:
• Origin inference rate: 94%
• Destination inference rate: 74%
• For full public transport network:
• 73% of all fare transactions are included in the seed matrix
• Computation time for full London OD Matrix (including both seed
matrix and scaling):
• 30 mins on 2.8 GHz Intel 7 machine with 8 GB of RAM
Nigel Wilson MIT 20
January 25, 2013
21. Reliability Metrics
Function: Performance Measurement
Objective: Characterize transit service reliability from
passenger's perspective, consistently across modes
Applications:
• London rail services
• Underground and Overground
• entry and exit fare transactions
• train tracking data
• London bus services
• typically high frequency
• entry fare transactions only
• AVL data
Nigel Wilson MIT 21
January 25, 2013
22. Excess Journey Time (EJT): Overground Example
Nigel Wilson MIT 22
January 25, 2013
23. Example: Reliability Metrics - Rail
High Frequency Service
• use tap-in and tap-out times to measure actual station-station journey
times
• characterize journey time distributions measures such as Reliability
Buffer Time, RBT (at O-D level):
% of Journeys
RBT
Travel Time
50th perc. 95th perc.
RBT = Additional time a passenger must budget to arrive on time for most of their
trips (≈ 95% of the time)
Nigel Wilson MIT 23
January 25, 2013
24. Line Level ERBT: Underground Example
12.00
Excess RBT
10.00
Baseline RBT
Travel Time (min)
8.00 5.22
1.86
6.00
1.56 2.36
4.00
4.18 5.52 4.18 5.52
2.00
0.00
NB SB NB SB
(5.74) (10.74) (6.54) (7.38)
February November
Period-Direction
Victoria Line, AM Peak, 2007
Nigel Wilson MIT 24
January 25, 2013
25. Reliability Metrics: Bus
Challenge to measure passenger journey time because:
• no tap-off, just tap-on
• tap-on occurs after wait at stop, but wait is an important part of
journey time
Strategy to use:
• trip-chaining to infer destination for all possible boardings
• AVL to estimate:
• average passenger wait time (based on assumed passenger arrival
process)
• actual in-vehicle time
Nigel Wilson MIT 25
January 25, 2013
26. Individual Train Load Inference
Function: Off-line (SOP and PM)
Objective: Estimate the passenger load for each train at each
station
Application: London Underground
Data:
• Oyster: entry station and time, exit station and time for each
passenger
• NetMIS: train arrival and departure times at stations
Nigel Wilson MIT 26
January 25, 2013
27. Individual Train Load Inference
Passenger Load-based Measures
Passenger loading on a vehicle can affect perceptions of
time:
• seated vs standing
• density for standees
• ability/willingness to board a vehicle
But not generally available from most ADCS systems.
Potentially valuable addition to Customer Information and
Operations Control Functions, if load estimation is feasible in
real-time
Nigel Wilson MIT 27
January 25, 2013
28. Individual Train Load Inference
Challenges:
• Oyster Times are truncated to the minute, so entry and
exit times have errors of up to 59 seconds
• Train tracking (NetMIS) data is incomplete
• In-station access, egress, and interchange times are
variable
• across stations, individuals, and instances
• Capacity constraints can be binding and are variable
• Many OD pairs have multiple paths available and actual
path chosen is unknown
Nigel Wilson MIT 28
January 25, 2013
30. Individual Train Load Inference
Definition: "Itinerary" refers to a feasible (sequence of)
train(s) for a specific passenger journey.
Proposed Methodology:
1. Categorize each observed passenger journey (OD pair, entry,
and exit times) as:
• Single path or multiple paths (based on OD pair and path choice
models)
• Single itinerary or multiple itineraries (based on NETMIS data and
Oyster trip times)
Nigel Wilson MIT 30
January 25, 2013
31. Proposed Methodology (cont'd)
2. Select most likely path for each multiple-paths case
3. Select most likely itinerary for each multiple itinerary case
4. Assign passengers
5. Compare assigned flows with train capacities. If
assignment is infeasible, return to Step 2, otherwise
terminate
Nigel Wilson MIT 31
January 25, 2013
32. Model Exploration
• Apply to Victoria and Jubilee Line OD pairs only – those
with complete train tracking data
• Examine the following passenger OD pair groups:
A • no interchange C • interchange
• capacity not • capacity not binding
(33% of (1%)
binding • single path
passengers)
• single path
B • no interchange D • interchange
• capacity may be • capacity not binding
(3%) (1%)
binding • multiple paths
• single path
Nigel Wilson MIT 32
January 25, 2013
33. Conclusions on LU Train
Load Inference Application
• Inferring train loads on LU not yet feasible (off-line):
• Missing NetMIS data (only 3 of 11 LU lines have complete train-
tracking data)
• Oyster Timestamp Truncation
• Clock Misalignment
• Basic groundwork for tackling the off-line problem has
been laid
• Generation of itineraries works well
• Keys to improving results are:
• improved modeling of access and egress times
• improved path choice models
Nigel Wilson MIT 33
January 25, 2013
34. Future Research
• Develop better models to support SOCM and CI functions
• descriptive
• predictive
• normative
• Evaluate marginal costs and benefits for additional data
availability in real time
• AFC data
• APC data
Nigel Wilson MIT 34
January 25, 2013
35. February Webinar
Regulatory Organization and Contractual Relations Between Agents
Professor Rosário M. R. Macário
Departamento de Engenharia Civil, Arquitectura e Georrecursos
Instituto Superior Técnico, Lisbon Portugal
Thursday, February 21st, 2013 at 1300 CLST (UTC-3)
The regulatory setting of urban transport is the background scene where economic and
institution relations occur. These are in turn regulated through contractual relation. The
last decades provided a strong development in this field establishing a clear cause-effect
relation between regulatory setting – contractual form – performance of operators and
system.
No universal solution exists and for each reality one must chose the type of contract that
provides the best fitting and the highest performance incentive to all intervening agents. In
the case of introduction of a new mode (e.g. BRT) these aspects gain a critical importance
since change is introduced in the system and agents behaviours do adjust to the new
circumstances.