MobileOD: travel patterns from large scale mobile phone data
- 1. Deriving 24/7 Operational OD Matrices From AirSage Mobile Phone Data Sacramento Pilot Study and Beyond October 2011 Jingtao Ma, PhD, PE, Mygistics, Inc.
- 2. Agenda Brief overview of OD derivation methodology and techniques AirSage data processing MobileOD pilot for Sacramento, CA Pre-processing: sample trips Projection based on CTPP survey data Hourly Vehicular OD (path flow) refinement based on static traffic assignment Vehicular path flow estimation based on observed path choice Path matching Path flow aggregation OD estimation (TFlowFuzzy) from path flows 2
- 3. Traditional Methods for Operational OD Derivation Travel demand model: Calculated, not observed and thus only as good as the model itself Only a fixed point snapshot of the mobility pattern Active probing: Automated number plate recognition (ANPR) or Bluetooth MAC matching Potentially more accurate, but usually case by case on a small scale Relatively slow turnaround Very expensive Passive probing: GPS based navigation devices Small samples Biased towards fleets and are thus not representative of a communitys travel patterns 3
- 4. OD Derivation Methods: Why Mobile OD? Mobile OD: travel pattern inference from mobile phone traces also a passive probing method In general: Sprint High device penetration: >85% conservatively estimated (285M devices/308M population in US) Wide overage Ubiquitous usage Travel patterns could be Verizon Weekday versus weekend Seasonal variation, special events Work trips/non work trips Continuous OD at fine grain spatial/temporal resolutions What is offered to clients Off-the-shelf 24/7 operational OD Add-on survey tool for household surveys as alternative to traditional GPS tracking Long-distance, inter-regional, external-external travel data 4
- 5. How AirSage Technology Works AirSage patented WiSETM platform transforms normal operational signaling data from wireless carriers into real-time and historical location and movement data. CDMA network techonology: Sprint & Verizon Currently 35 million Sprint devices in US; 90 million Verizon devices to be added
- 6. Operational 24/7 MobileOD Workflow AirSage Public NAVTEQ Various Sources Mobile Sightings Socio-economics Navigation Net Traffic Detectors Trips Block groups Model Traffic Paths Travel survey network counts Projected Mobile based OD Path flow Mygistics/PTV Operational proprietary 24/7 MobileOD
- 7. Sacramento Pilot: Project Background Customer Fehr & Peers Associates I-80/CA-65 Interchange improvement project Study period: 6-10AM, and 3-7PM A lengthy process was originally proposed for demand estimation Initial discussion at TRB 2011 7
- 8. Sacramento Pilot: Mobile Phone Data Encrypted Sprint subscribers data from one mobile switch coverage area for October 2010 Total mobile sightings: 256 million (255,828,842) Filtered and analyzed: 98 million Subscribers: more than128 thousand 400,000 sightings from 600 randomly selected subscribers 8
- 9. Snowball Trip Identification and Analysis System (STIAS) An Expert System Rule-based knowledge base Inference engine 20+ rules, one inference engine Mygistics proprietary
- 10. Trip Identification: The Mygistics Difference 14 randomly selected subscribers from the Sacramento dataset Regression Analysis: Eyes vs Myg-alg 0.4.1 Trips from three methods 80 70 60 Eyes Myg-alg0.4.1 AirSage 22 18 7 50 16 8 15 Eyes 40 68 54 6 Predicted Y 8 6 7 30 22 20 10 20 13 10 2 25 41 22 46 17 26 10 0 R2 = 0.89 9 9 3 0 20 40 60 21 25 5 Mygi-Alg 0.4.1 6 4 2 10 13 1 Regression analysis: Eyes vs AirSage 28 18 7 80 38 27 5 70 327 280 113 100% 85.6% 34.6% 60 80 Improvement 50 Eyes 60 factor of 2.5 40 30 Eyes 40 20 Alg0.4.1 20 10 R2 = 0.11 AirSage 0 0 0 5 10 15 20 25 30 AirSage 1 3 5 7 9 11 13 10
- 11. STIAS: Benchmark & Validation Do these numbers apply to the entire dataset? For these samples: 280 versus 113 (MYG alg 0.4.1 vs. AirSage Known Trips) Factor of 2.47 For the entire Sacramento dataset: 2.20 million vs. 1.04 million Factor of 2.12 The sample benchmarking favored Myg-alg 0.4.1 a little, but not too much Mygistics currently working on version 0.5, hopefully to get to the point of 90+% of trips identifiable by human eyes Which will bring to the same level of factor 2.5 11
- 12. OD Matrices from STIAS Identified trips mapped to TAZs Hourly aggregate over all weekdays of October 2010 288 thousand (non-zero) active O-D pairs 1070 active TAZ 1.14 million OD pairs 12
- 13. Path Matching (Trajectories) Path search & enumeration from VISUM For Sacramento, 65 million paths stored for query GIS functions in PostGIS assisted in path matching Shortest distance from via points to candidate paths Selected the most likely one(s) Using observed paths for OD refinement improves accuracy and requires fewer counts 13
- 14. Sacramento Pilot: Results Sample OD from identified trips mapped to TAZs OD projection based on CTPP survey to generate better seed matrix TFlowFuzzy (OD refinement in VISUM) (8x1h) Traffic assignment and matrix verification R^2 RMSE(%) 6AM 0.92 42 7AM 0.94 26 8AM 0.91 26 9AM 0.91 28 3PM 0.87 30 4PM 0.86 30 5PM 0.86 29 6PM 0.86 30 (Link/turn counts vs. model volume after matrix refinement) 14
- 15. Market Response to Date Ongoing projects, proposals, request for information Positive feedback for the Sacramento pilot project Active discussion on social media (LinkedIn groups, ITS America, etc.) Inquiries for new proposals and projects Interest from researchers, consultants and government agencies 15
- 16. The beginning of the more research and applications Ongoing projects, proposals, request for information 24/7 hourly OD matrices 16
- 17. The beginning of the more research and applications Ongoing projects, proposals, request for information 24/7 hourly OD matrices 17
- 18. OD Matrices Analysis Identified trips mapped to TAZs Hourly aggregate over all weekdays of October 2010 597,529 for Mobile OD (block group level for two months data) (non-zero) active O-D pairs 308,988 for weekdays 102,571 for weekends 158,617 for event days Active OD Pairs Sample Size Internal + Paths/Active OD External=Num of Pair (Internal/ Paths External) Weekdays 289,059+1992 51.7% 41 days 270,661+245,851=5 1.95 (0.93/12.3) 9=308,988 16,512 Weekends 82,642+19,929 17.2% 16 days 27,771+84,075=111 1.85 (0.34/4.2) =102,571 ,846 Event Days 138,688+19,92 26.5% 4 days 21,222+80,795=102 1.92 (0.15/4.1) 9=158,617 ,017 18
- 19. The beginning of the more research and applications Ongoing projects, proposals, request for information Trip mode inference Activity chain and tour imputation 19
- 20. The beginning of the more research and applications Ongoing projects, proposals, request for information Travel behavior change from continuous observations and more yet to explore 20
- 21. Mygistics MobileOD Full OD trip tables, not OD samples 24 hourly matrices for 7 days a week Census block group resolution (custom zone structure possible) Internal, external/internal and external/external trips Survey add-on tools (on-board survey, household survey) 21
- 22. Contact Jingtao Ma jma@mygistics.com 503-575-2191 ext 2802 22