A Calibration Procedure for Microscopic
Traffic Simulation
Lianyu Chu, University of California, Irvine
Henry Liu, Utah State University
Jun-Seok Oh, Western Michigan University
Will Recker, University of California, Irvine
Outline
• Introduction
• Data preparation
• Calibration
• Evaluation of the overall model
• Discussion
• Conclusion
Introduction to Microscopic simulation
• Micro-simulation models / simulators
– AIMSUN, CORSIM, MITSIM, PARAMICS, VISSIM…
– model traffic system in fine details
• Models inside a simulator
– physical components
– roadway network, traffic control systems, driver-vehicle units, etc
– associated behavioral models
– driving behavior models, route choice models
• To build a micro-simulation model:
– complex data requirements and numerous model
parameters
– based on data input guidelines and default model
parameters
Objective
• Specific network, specific applications
• Calibration:
– adjusting model parameters
– until getting reasonable correspondence between model
and observed data
– trial-and-error, gradient approach and GA
• Current calibration efforts: incomplete process
– driving behavior models, linear freeway network
• Objective:
– a practical, systematic procedure to calibrate a
network-level simulation model
Study network
Data inputs
• Simulator: Paramics
• Basic data
– network geometry
– Driver Vehicle Unit (DVU)
– driver behavior (aggressiveness and awareness factors)
– Vehicle performance and characteristics data
–
–
–
–
vehicle mix by type
traffic detection / control systems
transportation analysis zones (from OCTAM)
travel demands, etc.
• Data for model calibration
– arterial traffic volume data
– travel time data
– freeway traffic data (mainline, on and off ramps)
Freeway traffic data reduction
• Why
– too many freeway data, showing real-world traffic variations
– calibrated model should reflect the typical traffic condition
of the target network
– find a typical day, use its loop data
• How to find a typical day
– vol(i): traffic volume of peak hour (7-8 AM)
– ave_vol: average of volumes of peak hour
– investigating 35 selected loop stations
– 85% of GEH at 35 loop stations > 5
GEH 
Vol (i)  ave _ Vol 2
(Vol (i )  ave _ Vol ) / 2
Calibration procedure
Basic data input / Network coding
Calibration of driving behavior models
Calibration of routing behavior model
Model Fine-tuning
Reference OD from planning model
Route choice adjustment
Total OD estimation
Reconstruction of timedependent OD demands
N
Volume, Travel
time match?
Y
Overall model
validation / evaluation
Determining number of runs
 2
N  (t / 2 
)
 
• μ, δ:
– mean and std of
MOE based on the
already conducted
simulation runs
• ε: allowable error
• 1-α: confidence
interval
Start
Original nine runs
Calculating the mean and its std
of each performance measure
Calculating the required # of runs
for each performance measure
Additional one
simulation run
N
Is current # of
runs enough?
Y
End
Step 1/2: Calibration of
driving behavior / route behavior models
• Calibration of driving behavior models:
– car-following (or acceleration) , and lane-changing
– sub-network level
– based on previous studies
– mean target headway: 0.7-1.0
– driver reaction time: 0.6-1.0
• Calibration of route behavior model
– on a network-wide level.
– using either aggregated data or individual data
– stochastic route choice model
– perturbation: 5%, familiarity: 95%
Step 3: OD Estimation
• Objective: time-dependent OD
• Method:
– first, static OD estimation
– then, dynamic OD
• Procedure:
– Reference OD matrix
– Modifying and balancing the reference OD demand
– Estimation of the total OD matrix
– Reconstruction of time-dependent OD demands
Reference OD matrix
• Reference OD matrix
– from the planning model, OCTAM
• Modifying and balancing the reference OD demand
– problems with the OD from planning model
– limited to the nearest decennial census year
– sub-extracted OD matrix based on four-step model
– morning peak hours from 6 to 9; congestion is not cleared at 9 AM
– balancing the OD table: FURNESS technique
– 15-minute counts at cordon points (inbound and outbound)
– total generations as the total
Estimation of the total OD matrix
• A static OD estimation problem
– least square
– tools, e.g. TransCAD, QueensOD, Estimator of Paramcis
• Our method:
– simulation loading the adjusted OD matrix evenly
– 52 measurement locations (13 mainline, 29 ramp, 10 arterial)
– quality of estimation: GEH
M (n)  M (n)2
– GEH at 85% of measurement locations < 5
GEH 
obs
( M obs (n)  M sim (n)) / 2
– modification of route choices
– OD adjustment algorithm: proportional assignment
– assuming the link volumes are proportional to the OD flows
• Result:
– 96% of all measurement locations < 5
sim
Reconstruction of time-dependent OD
• A dynamic OD demand estimation problem
– research level, no effective method
– a fictitious network or a simple network
– practical method:
– FREQ: freeway network
– QueensOD, Estimator of PARAMICS, etc.
• Profile-based method:
– profile: temporal traffic demand pattern
– based on the total OD demand matrix
– assign total OD to a series of consecutive time slices
Finding OD profiles
• Find the profile of each OD pair
• General case (from local to local):
– profile(i, j) = profile(i) , for any origin zone, j =1 to N,
– profile(I): vehicle generation pattern from an origin zone
• Special cases:
– local to freeway
– estimated by traffic count profile at a corresponding on-ramp location
– freeway to local
– estimated by traffic count profile at a corresponding off-ramp location
– freeway to freeway*
– roughly estimated by traffic count profile at a loop station placed on
upstream of freeway mainline
– needs to be fine-tuned
• volume constraint at each time slice
Examples of OD profiles
Origin
1
1
2
3
4
2
Destination
3
4
profile(i)
(known)
Fine-tuning OD profiles
• Optimization objectives
– Min (Generalized Least Square of traffic counts
between observed and simulated counts over all
points and time slices)
– step 1:minimizing deviation of peak hour (7-8 AM)
– criteria: more than 85% of the GEH values < 5
– step 2: minimizing deviation of whole study period
at five-minute interval
– together with next step
– 52 measurement points
• Result:
– step 1: 87.5% of all measurement locations
Step 4: overall model fine-tuning
• Objectives:
– check/match local characteristics: capacity, volumeoccupancy curve
– further validate driving behavior models locally
– reflect network-level congestion effects
• Calibration can start from this step if:
– network has been coded and roughly calibrated.
– driving behavior models have been roughly calibrated and
validated based on previous studies on the same network.
– one of the route choice models in the simulator can be
accepted.
– OD demand matrices have been given.
Model fine-tuning method
• Parameters:
– Link specific parameters
– signposting setting
– target headway of links, etc
– Parameters for car-following and lane-changing models
– mean target headway
– driver reaction time
– Demand profiles from freeway to freeway
• Objective functions:
– min (observed travel time, simulated travel time)
– min (Generalized Least Square of traffic counts over all
points and periods)
• Trial-and-error method
Some calibrated OD profiles
Percentage of total demand
12.0%
10.0%
8.0%
6.0%
4.0%
2.0%
0.0%
6:00 6:15 6:30 6:45 7:00 7:15 7:30 7:45 8:00 8:15 8:30 8:45 9:00 9:15 9:30 9:45
Time of day
a freeway zone to a freeway zone
an arterial zone to an industrial zone
a freeway zone to an arterial zone
an artertial zone to a freeway zone
volume-occupancy curve
Loop station @ 2.99
Simulation
120
100
80
60
40
20
0
30-sec Volum e
30-sec Volum e
Real world
0
20
40
60
Percent occupancy
80
120
100
80
60
40
20
0
0
20
40
60
Percent occupancy
80
Evaluation of Calibration (I)
• Measure for goodness of fit:
– Mean Abstract Percentage Error (MAPE)
1 T
MAPE   (( M obs (t )  M sim (t )) / M obs (t ))
T t 1
600
Travel time (sec)
Travel time (sec)
300
200
100
3.1% (SB)
400
200
8.5% (NB)
0
0
6:00
6:30
7:00
7:30
8:00
simulation
8:30
9:00
observation
9:30
10:00
6:00
6:30
7:00
7:30
8:00
simulation
Comparison of observed and simulated
travel time of SB / NB I-405
8:30
9:00
observation
9:30
10:00
Evaluation of Calibration (II)
700
600
500
400
300
200
250
1000
200
800
150
600
400
100
200
50
100
0
0
6:05 6:30 6:55 7:20 7:45 8:10 8:35 9:00 9:25 9:50
405N0.93ml-sim
6:05 6:30 6:55 7:20 7:45 8:10 8:35 9:00 9:25 9:50
405N0.93ml-real
405N1.93ff-sim
6:
05
6:
30
6:
55
7:
20
7:
45
8:
10
8:
35
9:
00
9:
25
9:
50
0
405N1.93ff-real
405S3.31ml-sim
1000
1000
200
800
800
150
600
600
400
400
200
200
0
0
405S3.31ml-real
405N3.04ml-sim
405N3.04ml-real
50
405N3.86ml-sim
405N3.86ml-real
6:
05
6:
30
6:
55
7:
20
7:
45
8:
10
8:
35
9:
00
9:
25
9:
50
0
6:
05
6:
30
6:
55
7:
20
7:
45
8:
10
8:
35
9:
00
9:
25
9:
50
6:
05
6:
30
6:
55
7:
20
7:
45
8:
10
8:
35
9:
00
9:
25
9:
50
100
133s9.37ml-sim
5-min traffic count calibration at major freeway measurement locations
(Mean Abstract Percentage Error: 5.8% to 8.7%)
133s9.37ml-real
Discussion
• Completeness and quality of the observed data
– Especially important for calibration result
– Quality of the observed data
– Calibration errors might have been derived from
problems in observed data
– Probe vehicle data with about 15-20 minute intervals
cannot provide a good variation of the travel time
– Quantity / Availability of observed data
– cover every part of the network
– some parts of the network were still un-calibrated
because of unavailability of data
Conclusion
• Conclusion
– a calibration procedure for a network-level simulation model
– responding to the extended use of microscopic simulation
– the calibrated model:
– reasonably replicates the observed traffic flow condition
– potentially applied to other micro-simulators
• Future work:
– inter-relationship between route choice and OD estimation
– an automated and systematic tool for microscopic
simulation model calibration/validation