Implementation of CMAP’s Activity-Based Model with EMME

1. Implementation of CMAP’s Activity-Based Model with EMME Model City 2011: 22nd International Emme Users’ Conference

2. Presentation Outline • Project Need • CT-RAMP ABM • Road Pricing • EMME Implementation • Pricing Sensitivity Tests • Next Steps • Questions

3. CMAP Region • Population: 10.5 million • Modeling Region • 21 counties in 3 states • Neighboring MPOs • SE Wisconsin • NW Indiana • 1,944 TAZs • Road Network • 15.0K nodes • 44.3K links • Rail Network • 6.6K nodes • 19.5K links

4. Current Model System • Four-step trip based model • Fortran • Trip generation • Trip distribution • Mode choice • EMME • Time-of-day factoring (8 time-of-day periods) • Assignments and skimming • External, truck, and airport trips • Fixed trip tables • SAS for pre- and post-processing • Auto: SOV and HOV2+ • Trucks: heavy, medium, light, and b-plate (<4 tons) • Transit • 7 line haul modes • CTA: metro, local bus, express bus • Metra: commuter rail • PACE: express, regional, and local buses • 12 auxiliary modes

5. Policy Environment • GO TO 2040 • Regional comprehensive plan adopted in 2010 • Recommendations • Implement congestion pricing • Implement parking pricing • Increase commitment to transit • Need improved tools for testing pricing policies: ABM

6. Project • Develop pricing (demonstration) ABM • Borrow ABM from other MPOs (Atlanta, San Francisco Bay Area, San Diego, etc) • Develop base year synthetic population • Integrate with CMAP highway and transit networks • Re-estimate/calibrate key components • Destination choice • Mode choice • Prove usefulness of ABM; develop full ABM later

7. CT-RAMP Family of Models • Coordinated Travel Regional Activity-based Modeling Platform • Key features: • Explicit intra-household interactions and Coordinated Daily Activity Patterns (CDAP) • Near - continuous temporal dimension (30 minutes) • Java-based package for ABM construction

8. Members of CT-RAMP Family • 1st generation: • Columbus, OH (MORPC) – in practice since 2004 • Lake Tahoe, NV (TMPO) – in practice since 2006 • 2nd generation: • Atlanta, GA (ARC) – in practice since 2009 • San-Francisco Bay Area, CA (MTC) – in practice since 2010 • 3rd generation: • San-Diego, CA (SANDAG) – started in 2008 • Phoenix, AZ (MAG) – started in 2009 • Jerusalem, Israel (JTMT) – started in 2009 • Chicago (CMAP) – started in 2010 • Every model has many unique features

9. CT-RAMP Person Types

10. CT-RAMP Activity Types

11. CT-RAMPModel Structure • Auto ownership model • Destination choice models • Time-of-day choice models • Mode choice models Model Re-estimated for CMAP Pricing ABM

12. Distributed Modeling System • Main Machine • Manages model run system • Stores in-memory households, persons, matrices • Skimming and assignment for two time periods • 2 Six-Core Intel Xeon 2.66 GHz, 144 GB RAM, $10K • 3 Worker Machines • Solves model components (for bundles of households) • Skimming and assignment for two time periods • 2 Six-Core Intel Xeon 2.66 GHz, 144 GB RAM, $10K • Uses Java JPPF to run worker node processes and Microsoft PsExec to run EMME processes on workers

13. Distributed Model System

14. Road Pricing Essentials • Variation in Value of Time: • ABM operates with a continuous VOT distribution • EMME requires discrete classes (High VOT, Low VOT) • Vehicle occupancy: • ABM and EMME operates with 3 discrete classes (SOV, HOV2, HOV3) • Route type choice: • ABM and EMME explicitly treat toll and non-toll users for each segment

15. Advanced VOT Techniques in ABM • Basic VOT estimated for each travel purpose and person type • Situational variation of VOT applied for each person based on lognormal distribution • Car occupancy accounted by cost sharing: • VOT for HOV2 is 1.6 of highest participant VOT • VOT for HOV3+ is 2.3 of highest participant VOT • For static assignments VOT has to be aggregated across individuals into discrete vehicle classes

16. Example of VOT Distribution

17. Initial Value-of-Time Segmentation

18. Route Type Choice • Currently implemented as binary choice (toll vs. non-toll); can be extended to distinguish between managed lanes (toll vs. non-toll) and general purpose lanes (toll vs. non-toll) • Explicit modeling and analysis of toll users at OD level • Accounts for (negative) toll bias • Allows for VOT variation / segmentation beyond 12 assignable classes

19. Applied Segmentation Rules • Assignable trip tables are segmented by 44 classes: • 12 core auto components are generated by ABM • 8 truck components are handled by route type choice model implemented in EMME • 12 external components are handled by route type choice model implemented in EMME • 12 airport travel components are handled by route type choice model implemented in EMME

20. Desired Multi-Class Assignment Classes

21. EMME Implementation Constraints • Currently multi-class-assignment is limited to 12 classes (will be extended soon to 30) • It will be beneficial to consider more than 2 VOT classes, for example (Low, Medium, High) • Possible implementation scheme: • Pre-assign heavy and (possibly) medium trucks since they follow planned routes (4 classes) • Assign the rest of classes with heavy and medium trucks preloaded (16 classes)

22. Current Multi-Class Assignment Classes

23. Assignment and Skimming Macro

24. Equilibration Details • The model system requires 3-4 global iterations to reach a reasonable level of convergence • Assignment and skimming macro is run before each global iteration (to generate LOS for ABM) and after the last iteration (to assign the final results) • Assignment and skimming macro requires 4 internal iterations to equilibrate core and non-core components in route type choice • Smart schemes are applied w.r.t highway assignment accuracy at early internal iterations and ABM accuracy at early global iterations

25. EMME Integration • Eight databanks stored in the project folder on main machine • PsExec copies two banks to each remote worker machine • PsExec runs EMME macros remotely • PsExec copies the banks back to the main machine • Java-based ABM reads skims directly from the databank • ABM is run (with sampling) • ABM writes demand matrices directly to the databank

26. Run Times • EMME Skimming and Assignment • 8 databanks, 4 machines (12 threads each) • Module 5.21: 6 hours • 1 thread / databank • Module 5.22: 1 hour 20 minutes • 12 threads / databank • CT-RAMP ABM • 20% population: 4 hours • 100% population: 17 hours • Total Run Time for 1 iteration • 5 hours 20 minute (with 20% sample) • Will be reduced with additional machines (which is planned) 5.22 saves 78% on skimming and assignment time!

27. Pricing Sensitivity • Trips To/From the CBD • Scenario: Global pricing, 5X all toll costs

28. Pricing Sensitivity • Trips To/From the CBD • Scenario: Congestion Pricing, 5X peak toll costs

29. Next Steps • Additional scenario testing, including corridor specific tests and cordon pricing • Demonstrate usefulness of pricing ABM to policymakers • Full ABM implementation, including revised transit modeling procedures • Improve implementation with: • Three additional worker machines • Potentially EMME Modeller for data I/O, overall model running, automated creation of inputs, etc

30. Questions? Matt Stratton, mstratton@cmap.illinois.gov Kermit Wies, kwies@cmap.illinois.gov Ben Stabler, stabler@pbworld.com Peter Vovsha, vovsha@pbworld.com Surabhi Gupta, guptas@pbworld.com