1. Predicting Traffic
Dmitriy Bespalov

2. Imagine… You are driving on a highway Congestion Ahead!
GPS
Congestion Ahead!
Faster route found!
Take next exit!

3. Imagine… You are sitting in your cubicle
Receive a regular alert from Traffic.com:
If you leave now, you will spend 30 min driving
If you leave in 30 minutes: 45 minutes
If you leave in 1 hour: 60 minutes
Better go now, dude!

4. ? Predicting Traffic Sensor Network History of road conditions
real-time data
history data ?
Predicting
Server
predictions

5. Predicting Traffic 37 min Data Locations + + How long?
Belmont
How long? +
12 min
37 min
I-476
Need to predict travel times on smallest granularity roads’ segments – Data Locations

6. Predicting Traffic For single road’s segment (Data Location):
Record delay values for one day:
Perform same process for multiple days:
Delay (min)
Time of day (hours)
6 am
12 pm
6 pm
12 am
Historical Data

7. Last 20 minutes’ travel-times
Predicting Traffic
For single road’s segment (Data Location):
Historical Data
Find similar
pattern
Last 20 minutes’ travel-times
Prediction
Length
Predicted
Delay
Match!

8. How Predictions Are Made
Storing historical data takes a lot of space:
One reading every minute: 1440 per day
90 days of readings for historical data set
And this is for one segment!
Philadelphia has about 430 such segments
Total: 55,728,000 readings to store
And simply storing data is not enough:
Need to query historical data efficiently
Otherwise no use for real-time system

9. How Predictions Are Made
Compression required to reduce data size
Used piece-wise linear data fitting
Points were replaced with line segments
While closely fitting original data
14 points for 1 day of readings (1440 before)

10. How Predictions Are Made

11. How Predictions Are Made

12. How Predictions Are Made
delay
Historical Data
time
Find similar
pattern
Last 20 minutes’ travel-times

13. How Predictions Are Made
Historical Data
Last 20 minutes
Data
Location 2
Better Match!
Data
Location 1
Data
Location 3
time

14. How Predictions Are Made
For every match, assign distance measure
Sort all matches based on distance
Extract k-closest matches
For each of k matches produce prediction
Combine k predictions for final answer

15. Prediction Accuracy
The predictions are nearly perfect…

16. Prediction Accuracy
The predictions are nearly perfect…

17. Prediction Accuracy
The predictions are nearly perfect…

18. Prediction Accuracy
The predictions are nearly perfect…

19. Prediction Accuracy
The predictions are nearly perfect…

20. Prediction Accuracy
The predictions are nearly perfect…

21. Prediction Accuracy
… yeah, right…

22. Prediction Accuracy
… yeah, right…

23. Prediction Accuracy
… yeah, right…

24. Prediction Accuracy Average prediction error: 0.17 minutes
Standard deviation: 0.67 minutes
Average relative accuracy: 75%

25. Demo

26. How to Improve Accuracy?
Use better similarity measure for patterns
Classify historical data:
Traffic on Mondays is usually similar
Traffic on Sundays is usually not the same as on Mondays
Don’t need to search Sundays if today is Monday
Learn accuracy adjusting functions
Function will attempt to correct predicted values
Learn these functions in the background of Predicting Engine:
Using real and predicted pairs
Learn separate functions for various classes of incidents:
No incidents, rain, snow, accident, road work, etc

27. Other R&D Efforts Predicting model can be used:
Effective traffic data-mining tool
Compression of traffic data
Querying/matching capabilities
“Probe” data integration project

28. Other R&D Efforts “Probes” are actual moving vehicles
GPS data: latitude, longitude, heading, speed
Proposed new algorithm for travel-times
Calculus-based
Can handle constantly moving sensors
Coverage by “probes” is very sparse
e.g. average 6 “probes” / 20 min on entire I-76

29. Other R&D Efforts Try to build “general trend” function
Assume traffic is periodic
delay
time
Monday #4
Monday #3
Monday #2
Monday #1

30. Q&A

31. Are These Patterns Similar?
time
delay
time
delay
(x4,y4)
(x2,y2)
(x1,y1)
(x3,y3)

32. Are These Patterns Similar?
time
delay
time
delay
time
delay
distance = area between two pattern plots

33. Data & CPU Requirements
For Philadelphia:
Predicting time:
Using 3 months of historical data, prediction update takes less than 5 seconds to compute
System Memory:
Predicting engine with 3 months of history can fit in less than 100 Mb
DB Storage:
3 months of history needs ~ 50 Mb of DB storage

34. Predicting Engine’s Architecture
Traffic Data
History
Collector
Prediction
Models
Database
Traffic
Server
Metro map with
predicted values
Save
Prediction Model
Real-time
traffic data
Predicting Engine
Predicting Server (Philadelphia)
Traffic Conditions
History
Prediction Model
Pre-saved
Predicting Server
(Philadelphia)
Predicting Server
(New York)
Prediction Model Interface
Predicting Server
(Los Angeles)
Prediction Model
Implementation

35. Mondays of August 2005

36. Mondays of August 2005

37. Mondays of August 2005