2. About this presentation
Requirements for transport modelling
Three case studies
o Trip generation
o Trip distribution
o Trip distribution and mode choice
Summary
Page 2
4. Requirements for Transport Modelling
Replicating base year situations: In practice, it is used as
key measurement of model quality and is often the focus
of model development
Forecasting future year implications: In practice, it
receives less attention although forecasting is the ultimate
use of models
6. Case Study – Trip Generation
Alternative B: Introduce retail job density as an additional
explanatory factor for the number of attractions of
shopping trips
Presentation Title Page 6August 17, 2016
Modelling trip attraction to major shopping centres:
HBS trip attraction = a * number of retail jobs +
b * Ln(retail job density, in jobs / hectare)
Alternative A: Flags a number of shopping centers as
‘special generators’ and apply adjustment factors derived
from observed data
HBS trip attraction = [a * number of retail jobs] * adjustment factors
7. Validation Performance
Presentation Title Page 7August 17, 2016
The Alternative B produces considerably worse validation results
for the largest centres (A and B) but better outcomes for relatively
smaller shopping centres (C to H).
8. Forecasting Performance
Page 8
Alternative A: growth in jobs and growth in trip attractions are similar at
all the shopping centres
Alternative B, faster growth rate for trip attractions at the shopping
centres with relatively lower density, and slower growth rate at D where
the density reaches a very high level
10. Case Study – Trip Distribution
August 17, 2016
Calibrate gravity model for HBW trip distribution:
Alternative A:
- Use observed trip productions and attractions as inputs for
model calibration, even though it is modelled trip ends are
that input to the gravity model for model application
- Only include the OD movements with observed trips > 0,
excluding the ODs with zero trips, even though some of
the zeros may reflect travellers’ choice of not travelling
11. Case Study – Trip Distribution
Presentation Title Page 11August 17, 2016
Calibrating gravity model for HBW trip distribution:
Alternative B:
- Use modelled trip productions and attractions as inputs for
calibrating the gravity model parameters
- Include all the o-d pairs including those with zero trips.
This requires aggregating the movements to sector (SA2)
level to remove the zeros that are due to the small sample
size in HTS
12. Base Year Performance – for model calibration
Page 12August 17, 2016
Observed number of trip productions and attractions are
used to produce the trip distribution curves
13. Base Year Performance – for model validation
Presentation Title Page 13August 17, 2016
Modelled number of trip productions and attractions are
used to produce the trip distribution curves
16. Case Study – Trip Distribution and Mode Choice
Presentation Title Page 16August 17, 2016
Alternative A: mode choice occurs after the trip
distribution (or destination choice)
This ordering is computationally convenient: OD trip
tables from distribution provide inputs to the mode
choice
Arguably, this order also fits in a modelers’ experience
of real life in so far as on many occasions travelers’
decide on their destinations first and then choose
between transport modes
Develop destination and mode choice model for HBW trips:
17. Case Study – Trip Distribution and Mode Choice
Presentation Title Page 17August 17, 2016
Develop destination and mode choice model for HBW trips:
Alternative B: mode choice occurs before the trip
distribution (or destination choice).
Less computationally straightforward: the generalized
cost input for mode choice must be aggregated at trip
production level, instead of at trip (origin-destination)
level.
Due to the aggregation of cost inputs, the resultant
base year validation outcomes tend to be less desirable
than the destination choice first approach.
18. Validation Performance - Overall mode shares
Presentation Title Page 18August 17, 2016
Similar performance between the two alternatives
19. Validation Performance – Mode share for SA2 to SA2
movements
Presentation Title Page 19August 17, 2016
Alternative A:
20. Validation Performance – Mode share for SA2 to SA2
movements
Presentation Title Page 20August 17, 2016
Alternative B – similar performance to Alternative A
21. Forecasting performance
August 17, 2016 Page 21
Car GC = 30
mins
PT GC = 70
min
Car GC = 30
mins
PT GC = 70
min
D1
D2
Base case assumptions:
1000 trips from origin to
destination D1, D2, via
Car or PT
Destination and Mode
choice results in base case
(for both alternative):
D1, Car = 478 trips
D1, PT = 22 trips
D2, Car = 478 trips
D2, PT = 22 trips
22. Forecasting performance
August 17, 2016 Page 22
Car GC = 40
mins
PT GC = 70
min
Car GC = 30
mins
PT GC = 70
min
D1
D2
Scenario testing:
Car travel cost to D1
increases from 30 to 40
minutes
Destination and Mode choice
results , Alternative A
(double constrained)
D1, Car = 432 trips
(decrease from 478 in Base)
D1, PT = 68 trips (increase
from 22 trips in Base)
D2, Car = 478 trips
(unchanged from Base)
D2, PT = 22 trips
(unchanged from Base)
23. Forecasting performance
August 17, 2016 Page 23
Car GC = 40
mins
PT GC = 70
min
Car GC = 30
mins
PT GC = 70
min
D1
D2
Scenario testing:
Car travel cost to D1
increases from 30 to 40
minutes
Destination and Mode choice
results , Alternative B
D1, Car = 394 trips (more
significant decrease)
D1, PT = 106 trips (more
significant increase)
D2, Car = 480 trips (nearly
unchanged)
D2, PT = 20 trips (nearly
unchanged)
25. Implications
Overfitting: Use synthetic measures for fitting base year
outcomes to observed data, which may affect forecasting
performance adversely.
Underperforming: Accommodate underperformance
(compared with overfitting option) in base year validation,
in exchange of improved forecasting performance
Another option:
Common practice?