1. Tennessee Statewide Model Integration with the National Long Distance Passenger Model and Calibration to AirSage Data
Vince Bernardin, PhD, RSG Hadi Sadrsadat, PhD, RSG Steven Trevino, RSG Chin-Cheng Chen, RTCSNV
May 16, 2017

2. Background

3. TDOT and Tennessee MAJOR INTERSTATE CORRIDORS
Tennessee crossed by I-40, I-75, I-24, and I-65
Long distance truck and passenger traffic fight for capacity with local travelers especially in Knoxville and Chattanooga where mountainous topography limits alternative through routes
Major studies examining needs for these existing major corridors
Possible extension of I-69

4. Tennessee Statewide Travel Model - History
ORIGINAL 2005 TSTM
Fixed trip tables, limited coverage
VERSION 2 TSTM (2014)
3-step, pivot-point model, 3x network & zonal detail
Accelerated schedule to support state’s long range plan
TDOT FIRST EXPERIENCE WITH BIG DATA IN TSTM2
8 week dataset from American Transportation Research Institute (ATRI) to support the Version 2 TSTM
Over 234,000 individual trucks
Over 6.5 million truck trips
Represents roughly 11% of the trucks on the road for 56 days

5. Memphis 1,000 Truck Sample

6. Same 1,000 Trucks After 24 Hours

7. Same 1,000 Trucks After 48 Hours

8. Same 1,000 Trucks After 72 Hours

9. Same 1,000 Trucks After 5 Days

10. Same 1,000 Trucks After 7 Days

11. Number of Count Stations Correlation Coefficient
TSTM2 Success
SUBSTANTIAL IMPROVEMENT IN MODEL PERFORMANCE
Access to TN State University System
Volume Range
Number of Count Stations
Percent Error
Correlation Coefficient
MAPE
%RMSE
Model
Standard 1
5,000
7,288
10.2
0.69
66.8
91.1
101.4
10,000
1,943
5.5
0.61
31.1
39.6
56.3
20,000
1,700
0.8
0.79
21.0
27.8
51.4
30,000
747
-2.1
0.85
15.9
21.5
35.7
40,000
318
0.6
12.0
18.1
32.0 +
661
-0.1
0.94
11.5
15.6
21.6
Total
12,645
2.1
0.97
47.9
37.0
60.0

12. TSTM Version 3 BUILD ON SUCCESSES AND CAPITALIZE ON OPPORTUNITIES
New Commodity Flow Freight Model
Make ATRI-based trucks sensitive
Advanced trip-based model to replace 3-step
Mode & Destination choice models, Linked NHB trips
New AirSage+ATRI-based pivot point
Incorporate New National Long Distance Model
Calibrate to AirSage
Successful use in Chattanooga
Successful use for intercity corridors

13. Model Design & Implementation

14. Tennessee Statewide Travel Model
LONG DISTANCE PASSENGER DEMAND
INPUTS
Networks
SE Data
1. Network Skimming
2. Synthetic Population Expansion
DEMAND MODELS
Freight Demand
Short Distance Passenger Demand
Long Distance Passenger Demand
3. rJourney National Long Distance Model
4. Matrix Aggregation / Disaggregation
ASSIGNMENT

15. 1. Network Skimming INTEGRATION
Statewide model network in TN integrated with National Highway Planning Network outside TN
National Zone to National Zone (NUMA to NUMA) skim
NUMA to TAZ correspondence for centroids

16. 2. Synthetic Population / Socioeconomics
EXPANSION
Expand (or sample) base population rather than re-synthesize
Based on TAZ households within TN
Separate county household totals file outside TN
Detailed future demographic not needed outside TN
Faster than re-synthesizing population for whole nation
SIZE VARIABLES
Similarly, socioeconomic size variables for destination choice are scaled
Based on TAZ employment, etc., within TN
Separate county employment totals file outside TN

17. 3. National Long Distance Model
COMPONENTS
Household level disaggregate simulation like ‘simplified’ activity-based model
Tour generation-scheduling-duration, and party size by business and leisure
Destination and mode choice
INPUTS/OPTIONS
Networks/travel times and costs
Autos (from TSTM)
Bus (from TSTM)
Rail (exogenous input)
Air (exogenous input)
Possibility for many policy scenarios/alternative futures
5 LD Tour Types in rJourney
Visit friends & family
Leisure
Personal Business
Commute
Employer’s Business

18. 4. Matrix Aggregation/Disaggregation
RECONCILING ZONE SYSTEMS
rJourney national model produce NUMA to NUMA trip table
Within TN (and immediate halo) NUMAs disaggregated to TAZ
Beyond, NUMAs aggregated to larger zones (e.g., states)
Final trip table has manageable number of zones  faster assignment
Same network uses different zones/centroids for skimming and for assignment

19. Data Development

20. Big Data Fusion – AirSage & ATRI
FILTERED OUT “INTERMEDIATE” STOPS ON LONG DISTANCE TRUCK TRIPS IN ATRI
Meant to make ATRI trips comparable to both AirSage and commodity flows (FAF/Transearch)
Used similar but slightly different algorithm than AirSage – compared distances, if AB + BC ≈ AC then drop B B A C B A C

21. AirSage – ATRI Integration
SUBTRACTED ATRI FROM AIRSAGE
Before filtering, more negatives:
11% of cells
0.20% of trips
After filtering, less negatives:
1.3% of cells
0.09% of trips
Still not perfect, but filtering ATRI reduced negatives by 87%
Remaining negatives indicate remaining issue with intermediate stops, or perhaps coverage drops along some Interstates
BEFORE
AFTER

22. AirSage Expansion Problem
AIRSAGE DEFAULT CARRIER MARKET SHARE METHOD
Standard practice
AirSage does preliminary expansion based on carrier market share by census geography
analysts scale for “auto occupancy” (actually, vehicle trips/cell trips)
Previously worked reasonably well for both urban areas and intercity corridors
But in TN statewide context
significant urban under-loading (e.g., -10%)
and rural/intercity over-loading (e.g.,+15%)

23. AirSage Expansion Adjustments
TWO-STAGE ADJUSTMENT
How best to expand to traffic counts?
Parametric – fit distance-based expansion factor curves for residents and non-residents
Non-parametric – used ODME for residual adjustments
Avoid massive ODME adjustment, provide explanation/understanding of bias and correction
Trip length bias hypothesis from previous finding with other passive data (ATRI) and general explanation of observed pattern
Under-representation of shorter distance/duration trips confirmed by both parametric and non-parametric methods

24. Parametric Adjustment
Statistic
Before
After
Urban Percentage of Error
-10%
-2%
Rural Percentage of Error
+15%
+5%
TRIP LENGTH CORRECTION
Resident Expansion Factor = *Exp(-0.051*Length)
Visitor Expansion Factor = *Exp(-0.020*Length)
Visitors are already long distance travelers – may be more likely to have cell phones / higher auto occupancy
100 mi trip is 12 times as likely to be detected as a 10 mi trip
Didn’t we rescale to 0% error? If so, we should update all of this.

25. Non-Parametric Adjustment (ODME)
CONTROLS
Minimum factor 0.5
Maximum factor 5.0
10 iterations
RESULTS
Matrix MAPE 1.29
RMSE vs. counts from 66.3% to 36.1%
Modest additional increase in short trips
Didn’t we rescale to 0% error? If so, we should update all of this.

26. Calibration Results

27. TN Internal Districts
DISTRICT SCHEME (INTERNALS)

28. Long Distance Trip Generation / Frequency
MODELS VS. AIRSAGE
No purpose in AirSage, different purposes in v2 vs v3, so only compare total long distance trip rates
TSTM3 / rJourney generally reproduce AirSage well
Old model based on NCHRP 735 was 17% low – but offset by lower vehicle occupancy (3.0 vs 3.5)

29. Trip Distribution – State Border Crossings
INBOUND/OUTBOUND VS WITHIN STATE
AirSage shows much more balanced within-state and inbound/outbound long distance trips
Gravity models a la NCHRP 735 cannot reproduce this pattern
rJouney reproduces the pattern after the addition of a psychological boundary term to its destination choice models
Trip Type
AirSage
TSTM v2
TSTM v3
I-I Trips
110,779
11,597
112,741
I-E and E-I Trips
88,422
177,468
78,790
Total
199,201
189,065
191,531

30. Trip Length Distribution
WITHIN STATE TRIPS
Generally good, except too many <75 mi trips

31. Trip Length Distribution
WITHIN STATE + INBOUND/OUTBOUND
TSTM3 / rJourney generally tracks AirSage
Model expected to be lower at extreme long distances b/c AirSage data districts do not cover whole country

32. TN Internal Districts
DISTRICT SCHEME (INTERNALS)

33. TN External Districts
DISTRICT SCHEME (EXTERNALS)

34. Within TN Trip Distribution
DISTRICT-TO-DISTRICT COMPARISON
Importance of comparing actual OD pattern, not just TLFD
Very good agreement, overall
District level origins & destinations all within 1.5%
District level ODs all within 2% except within Nashville district
Relative Percentage Difference (Model Version 3 vs AirSage) I-I Trips

35. To/From TN Trip Distribution
DISTRICT-TO-DISTRICT COMPARISON
Generally good agreement
District level origins & destinations all within 10%, most within 3%
Smoky Mtns not attracting enough to/from Knoxville
District level ODs all within 4% except within Nashville - Northcentral
Relative Percentage Difference (Model Version 3 vs AirSage) I-E & E-I Trips

36. Conclusions

37. Conclusions NATIONAL LONG DISTANCE MODEL
Capable of integration with statewide models
Feasible implementation structures, runtimes (~3-5 hrs)
Capable of calibration to new data
Calibration to regional data important
May need new psychological boundaries, etc.
AIRSAGE LONG DISTANCE DATA
Valuable new data capable of supporting long distance modeling
Data reveals patterns significantly different than NCHRP 735 national defaults
Possible trip length bias makes scaling/expansion challenging – and important

38. DIRECTOR OF TRAVEL FORECASTING
Contact
Vince Bernardin, Jr, PhD
DIRECTOR OF TRAVEL FORECASTING