Traffic Flow Modeling with Realtime Data for Online Network Traffic Estimation and Prediction

1
Traffic Flow Modeling with Real-time Data for
On-line Network Traffic Estimation and Prediction

• Hani S. Mahmassani
• Based on Doctoral Dissertation Research of
• Xiao Qin

2
Traffic Flow Modeling with Real-time Data for
On-line Network Traffic Estimation and Prediction

• Modern transportation network management depends
  on the availability of timely and accurate
  estimates of prevailing and incipient traffic
  conditions.

3
Traffic Flow Modeling with Real-time Data for
On-line Network Traffic Estimation and Prediction

• Modern transportation networks depend for their
  success on the availability of timely and
  accurate estimates of prevailing and emerging
  traffic conditions.
• This calls for a real-time traffic estimation
  and prediction system that provides predictive
  traffic conditions as a basis for traffic control
  and information supply for online support of
  decision-making by traffic management operators
  and travelers.

4
Traffic Flow Modeling with Real-time Data for
On-line Network Traffic Estimation and Prediction

• The system uses advanced traffic network models
  (i.e., dynamic traffic assignment type models) to
  combine historical data as well as real-time
  traffic data from different sources.

5
Traffic Flow Modeling with Real-time Data for
On-line Network Traffic Estimation and Prediction

• 
  
  
• Require reliable and robust traffic flow models,
  capable of representing the dynamic evolution of
  traffic over space and time and ensure online
  operational capability in the traffic estimation
  and prediction system.

6
Descriptive conditions Prediction Normative
guidance
Traffic Management Center
Real-time Traffic Estimation / Prediction System
Guidance (VMS), Signal control,
Network
Real-time Traffic Flow Models
Advanced Traffic Models
Fundamental core
Surveillance system
Real-time traffic data
7
Research Problem

• Given
• Real-time traffic sensor data stream (on a
  subset of links in a network)
• Goal
• Formulate a dynamic traffic flow model driven by
  real-world observations, which would be suitable
  for mesoscopic simulation within DTA model
  system.
• Develop efficient algorithms for model
  calibration and application
• Develop an effective framework for model
  integration within an online traffic estimation
  and prediction system to support traffic
  operations management and information supply.

8
Traffic flow modeling

• Needed when the traffic process has to be
  simulated for a number of different scenarios and
  conditions.
• Cause-effect models
• Univariate models (e.g. Smoothing methods,
  Artificial neural networks(ANN), ARIMA models)
• lack representation of any underlying structural
  relations that drive traffic dynamics
• For a large scale network, macroscopic approach
  towards traffic flow modeling is a preferred
  choice.
• Robust (collective effects at a certain level
  of aggregation)
• Efficient (less computational cost)

9

• Transfer function method
• Used for study of dynamic response occurring in a
  dynamic input-output system
• deals with
• time-lagged relationship between output and input
• time series pattern (autocorrelation) in the
  residuals

, , , polynomial
functions of backward shift operator B
Transfer Function Model for Dynamic System with
Noise
10
Macroscopic Modeling

• Kinematic (hydrodynamic) traffic theories
• First-order continuum model (LWR model)
• Based on equilibrium speed-density relation
  (i.e., Modified Greenshields model)
• More suitable for long term offline operational
  planning
• Higher-order continuum model (PW model)
• More complicated, but potentially more accurate
• Various specifications and estimation methods
  inconclusive.

Modified Greenshields model
Higher-order continuum model
?
?
?
11
Density Anticipation
Speed Relaxation
Speed Convection
traffic flow
i-1
i1
i
12
Macroscopic Modeling

• Kinematic (hydrodynamic) traffic theories
• First-order continuum model (LWR model)
• Based on equilibrium speed-density relation
  (i.e., Modified Greenshields model)
• More suitable for long term offline operational
  planning
• Higher-order continuum model (PW model)
• More complicated, but potentially more accurate
• Various specifications and estimation methods
  inconclusive.

13
Real-time Traffic Flow Model

• Based on availability of real-time traffic sensor
  data streams (speed and density)
• Embeds the physical concepts of the higher-order
  continuum model within the structural formulation
  of
• dynamic response occurring in a dynamic
  cause-effect system
• time-lagged relationship between effect and
  cause
• autocorrelated residuals

transfer function models
, , , , ,
polynomial functions of backward shift operator
B white noise
14
Real-time Traffic Flow Model (contd)

• to model a dynamic speed-density relation

15
Real-time Traffic Flow Model (contd)

• Applied to mesoscopic type assignment-simulation-b
  ased traffic models (DYNASMART)

Link movement
Static model
Link Movement
Dynamic model
DYNASMART Model Structure
16
Real-Time Dynamic Traffic Assignment System
Architecture
DYNASMART-X
Traffic Network Surveillance (e.g. PeMS) and
Information Distribution (TMC)
Traffic State Estimation (DTA Simulator)
Consistency Checking
OD Estimation / Prediction
Traffic State Prediction (DTA Simulator)
Route Guidance
17
Model Estimation

• A least-squares based method (Liu and
  Hanssens,1982) to estimate multiple input
  transfer functions
• An iterative procedure to attain consistency and
  efficiency of estimates.

18
Speed prediction

• minimum mean square error prediction
• Conditional expectation at present time t.

forecasting
t
tl
current time
19
Adaptive model calibration and speed prediction

• Rolling horizon scheme
• Use most recent traffic data

• Advantages
• Constant computation cost
• Constrained prediction variance
• Faster response to actual traffic dynamics

20
Short term correction
Traffic Estimation

• Error propagation
• A tendency to deviate from actual situation
• Need an adjustment process to reduce
  inconsistency.
• Identify discrepancies -gt adjustment on speeds
• PI control in feedback control theory
• Corrective control action is dependent on present
  and past errors
• Deviations to be considered in control action
• Speeds
• Speeds densities
• Parameters
• Pre-determined
• Adaptively estimated

Traffic Prediction
21
Numerical Experiments

• Given real-time data of speed and density on a
  subset of links in a network
• Objectives evaluate model performance in
  predicting consistent dynamic link speeds in real
  time

• Irvine, CA
• Loop detector data
• 30-second sampling interval
• From 400 am to 1000 am, 5 days
• 13 study links

22
Numerical Experiments (contd)

• Stand-alone (independent) experiments
• Individual link level evaluation
• Do not require traffic simulation
• Densities are known values
• Integrated experiments
• Network-wide evaluation
• Used as the traffic flow model within DTA-type
  traffic simulation
• Densities are simulated values

23
Numerical experiments (contd)

• Four model variants (to accommodate the data
  availability in real world)
• Type I relaxation
• Type II relaxation convection
• Type III relaxation anticipation
• Type IV relaxation convection anticipation

24
Numerical Experiments (contd)

• Measure of effectiveness
• Weighted root mean squared error (RMSE) of
  estimates against observations
• Equilibrium relation
• Modified Greenshields model
• Input
• Benchmark

25
Stand-alone Experiments

• Performance evaluation across the four model
  variants with different static models
• Robust performance is insensitive to the quality
  of the static relation.
• Effective all variants are effective the model
  Type III (with relaxation and anticipation
  effects) is most effective for speed estimation.
• Sensitivity analysis of rolling-horizon scheme
• roll period 2.5 min, 5 min
• calibration horizon 50 min, 60 min

26
Integrated Experiments

• Partial observations
• 13 study links
• Static model-applied links (SM links)
• Dynamic model-applied links (DM links)
• Performance analysis
• Model compatibility
• Model transferability
• Model estimation capability with short term
  correction
• Model prediction capability
• Impact of temporal scale (of observation
  interval)
• Impact of spatial scale (of link segment)

27
Integrated Experiments (contd) ? model
compatibility

• Vary the of DM links that are randomly selected
  from the study links over the network

28
Integrated Experiments (contd) ? model
compatibility

• Errors reduced for DM links and on average
• Trend for SM links

Error reduction with of DM links over network
29
Integrated Experiments (contd) ? model
compatibility

• Lowers the estimation errors
• Significant improvement during rush hours

Errors for static model vs. dynamic model (on
all the 13 study links)
30
Integrated Experiments (contd) ? model
compatibility

• Shift from high-error regions to low-error regions

Frequency histogram of overall errors from
static model vs. dynamic model for the study
links
31
Integrated Experiments (contd) ? model
compatibility

• Vary the of DM links that are randomly selected
  from the study links along a corridor

32
Integrated Experiments (contd) ? model
compatibility

• Errors reduced for DM links and on average
• Trends for SM links on the impacted and
  non-impacted corridors

Error reduction with of DM links on I-405N
33
Integrated Experiments (contd) ? model
compatibility

• Include additional higher-order dynamic factors
  (convection and/or anticipation) when observed
  links are predominantly connected.

Errors from SM, DM I, DM II, DM III, and DM IV
34
Integrated Experiments (contd) ? model
compatibility

• To decide on an appropriate traffic flow model
  strategy for a specific network
• Detector configuration in the network
• Amount (dense)
• Spatial distribution (balanced)
• Interconnectedness

35
Integrated Experiments (contd) ? model
transferability

• Transfer the dynamic model (adaptively
  calibrated)
• to other links (study links, or non-study
  links)

36
Integrated Experiments (contd) ? model
transferability

• The value of even a small of sensors can be
  increased by judiciously transferring the dynamic
  model properties to adjacent links

Error reduction with different model transfer
configuration
37
Integrated Experiments (contd) ? short term
correction

• Useful to reduce the simulation errors
• More apparent impact for static model than for
  dynamic model

Errors for static model vs. dynamic model
(without or with short term correction)
38
Integrated Experiments (contd) ? short term
correction

• Shift from high-error regions to low-error regions

Frequency histogram of overall errors from
static model vs. dynamic model for the study
links
39
Integrated Experiments (contd) ? short term
correction

• Estimated time series

40
Summary and Conclusion

• Introduces a new perspective to the
  specification, calibration and application of
  ITS-oriented traffic flow models.
• Takes advantage of massive real-time traffic
  sensor data and view them as a mine of
  information to uncover dynamic traffic patterns.
• Explores the effects of higher-order dynamic
  factors for triggering traffic dynamics.
• Formulates the model as a general dynamic system
  through techniques adapted from the transfer
  function model.
• Flexible and effective with adaptive model
  estimation and prediction based on
  rolling-horizon framework.
• Establishes mechanism for short term correction
  to improve consistency of speed prediction.

41
Summary and Conclusion (contd)

• Stand-alone application
• Reliable and robust
• Combined impact of relaxation and anticipation is
  most effective.
• Rolling-horizon scheme
• Integrated application
• Importance of sensor amount, distribution and
  interconnectedness
• Transferability among neighboring links of the
  same type
• Functionality of short term correction
• Enhancement to traffic prediction capability
• Values of real-time data with different
  observation intervals

42
Future Extensions

• Include seasonal patterns in the model
  formulation.
• Explore other forms of the higher order continuum
  traffic flow models.
• Use the proposed dynamic model as a mechanism for
  short term correction.
• Evaluate and extend the model with a larger
  amount and more diverse types of data (other than
  loop detector data).

43
LIGHT AT THE END OF THE TUNNEL?
Thank you Q A