Predicting for the adaptive transport system
Francisco Pereira
camara@dtu.dk
The prediction -optimization pipeline (cycle)
What is preached…
…what we risk getting into
…what we risk getting into
The prediction -optimization pipeline (cycle)
Wrong balancing Wrong pricing
In non -recurrent conditions…
Wrong routing Wrong scheduling
In non -recurrent conditions…
Attention to stress scenarios
Large events Incidents
System breakdowns
…
Focus of this talk
Non -recurrent scenarios
Demand Supply
Expectable
Unexpectable
Non -recurrent scenarios
Demand Supply
special events, demonstrations,
holidays Expectable
Unexpectable
Non -recurrent scenarios
Demand Supply
special events, demonstrations,
holidays
road works, closures, mega
events Expectable
Unexpectable
Non -recurrent scenarios
Demand Supply
special events, demonstrations,
holidays
road works, closures, mega
events Expectable
crisis situations, unknown high
fluctuations Unexpectable
Non -recurrent scenarios
Demand Supply
special events, demonstrations,
holidays
road works, closures, mega
events Expectable
crisis situations, unknown high
fluctuations
incidents, weather, crisis situations
Unexpectable
Non -recurrent scenarios
Demand Supply
special events, demonstrations,
holidays
road works, closures, mega
events Expectable
crisis situations, unknown high
fluctuations
incidents, weather, crisis situations
Unexpectable
Expectable demand
Event information is usually online
● Event homepages
● Event listings
● Social media (e.g. twitter, facebook )
● News feeds
…
Lots of potential sources, but a lot in free form text
Topic
m ode ling Machine -
inte rpre table fe ature s
Accounting for event information
Data preparation process
Data preparation process
● Features used:
- Topics (K=24) - Google hits
- Facebook likes
- Time to event start, time since event start - Location
Bayesian additive model
Gaussian process that models routine component
Gaussian process that models the effect of events
Probabilistic
Graphical Model
Public transport arrivals in Singapore
Data:
● 5 months of smartcard data (bus, metro, light rail)
● Sources: singaporeexpo.com.sg , upcoming.org , last.fm, timeoutsingapore.com
● 2 study areas
○ Singapore Indoor Stadium
○ Singapore Expo
large e ve nt
Observed arrivals
Model WITHOUT events information Model WITH events information
An example
Public transport arrivals in Singapore
RAE: relative absolute error CorrCoef: correlation coefficient R2: coefficient of determination Unit: # passengers /15 min
Exploiting the additive structure
Non -recurrent scenarios
Demand Supply
special events, demonstrations,
holidays
road works, closures, mega
events Expectable
crisis situations, unknown high
fluctuations
incidents, weather, crisis situations
Unexpectable
Unexpectable supply
Model degradation
Simulation + machine learning
Test
network
- Hillerød Motorvej
Offline calibration
Multiple scenarios
Simulation + machine learning
Just -in-time model updating
Results summary
Predicting for the adaptive transport system
Francisco Pereira
http:// mlsm.man.dtu.dk
camara@dtu.dk
