Data-driven traffic management combining analytical models & machine learning Dr. Ing. Lorenzo Meschini, Vice President Product Management & Engineering Real-Time MSc.-Ing. Nora Szabo, Senior Sales Manager
About PTV Group Active customers >50,000 use PTV Software (>2,500 Cities) Global presence
“Empowering transportation and mobility – for a cleaner and smarter future.”
with 28 locations on all continents Founded
1979 as a spin-off of the of
Transportation Karlsruhe Institute Technology 45% Mobility Colleagues 38% ~900 worldwide
Headquarters in Karlsruhe, Germany
Real-time traffic management: Data-driven simulation approach
Network models
Real-time data
Real-time traffic & mobility simulation
Benefits: •
Space propagation: Estimating the current traffic situation not only where data are available, but across the whole network
•
Time propagation: Forecasting the future traffic situation which will arise in the absence of remedial actions
•
Proactive decision-making and management: Evaluation of future traffic situation as it would be influenced by remedial actions
PTV methodology for real-time mobility management TM Functionality
Traffic Estimation
Traffic Forecast
Scenario Evaluation & Decision Support
“What is going on?“
“What is going to happen?“
“What would happen if?“ “What should we do?“
Observed data approach
Maybe (with extensive measures)
X
X
Statistical approach
✓
“Usual“ conditions only
X
Simulation Approach
✓
✓
✓
Real-time traffic management installations European cities
European countries
Non-European cities
•
Paris
•
Slovenia
•
Abu Dhabi, UAE
•
Strasbourg
•
Czech Republic
•
Sydney, Australia
•
Vienna
•
Romania
•
Taichung, Taiwan
•
Salzburg
•
Erfurt
•
Ingolstadt
•
Regensburg
•
Gdansk
•
Krakow
•
Torino
•
Moscow
•
York
Best practice: Regional Traffic Supervisor Piemonte Region
•
Coverage: 25,000 sq. Km, 5M inhabitants, 34,000 km of roads
• Dynamic model • 2,000 zones • hourly OD matrices for 3 demand segments and 6 day types • Supply: 76,000 links • +0’, +15’, +30’, +45’, +60’ forecasts produced every 5 minutes
Best practice: Regional Traffic Supervisor Piemonte Region
• Real-time traffic monitoring • Real-time traffic estimation and short-term prediction • Generate alarms in presence of traffic disruptions (congestion, accidents, …) • Traffic control and supervision • Info mobility
Best practice: Taichung PTV Optima (Short-term traffic forecast) + PTV Balance (Adaptive singal control) = Proactive Adaptive network signal control PTV Optima for Taichung
-9,4%
+7,6%
+8,4%
Total travel time
Total Flows
Average speed
PTV Balance for Taichung
PTV methodology for real-time mobility management – can we do better? TM Functionality
Observed data approach Statistical approach Simulation Approach … What‘s next?
Traffic Estimation
Traffic Forecast
Scenario Evaluation & Decision Support
“What is going on?“
“What is going to happen?“
“What would happen if?“ “What should we do?“
X
X
✓
“Usual“ conditions only
X
✓ ?
✓ ?
✓ ?
Maybe (with extensive measures)
Data-driven modelling INPUT
OUTPUT
• Graph • Zones • Floating Car Data
• Model
DATA FUSION
Data-driven outputs
Trip Origins and destinations
Distribution of Paths & Flow
Speed Profiles and Patterns
Improved forecast of traffic conditions using machine learning (AI) Problem: •
Up to now, forecast of traffic conditions is usually based on static approach.
•
Current traffic conditions are not taken into account.
Solution & added value: •
Artificial Intelligence/ Machine Learning uses historical data to automatically detect typical daily time profiles in an unsupervised, fully automatic approach.
•
This data is clustered and provides the fundament for a forecast of up to 60 minutes in advance.
?
Technical Implementation: Machine Learning for short-term forecast of traffic conditions 1/3
Used Data: • Data coming in time profiles, e.g. traffic flow at count locations, speeds at count locations, occupancy of parking lots • The AI algorithms need historic time profiles for training and real-time data to forecast under current conditions Data provider:
• Owner of count locations, e.g. urban / regional road authorities • Their data providers (e.g. Materna)
Historical Data 15’ Intervals
Forecast: Identified time profiles (typical days) as input for the short-term forecast
Algorithms of Machine Learning
Normalisation
Technical Implementation: Machine Learning for short-term forecast of traffic conditions 2/3 AUTOMATIC DAY TYPES CLUSTERISATION
Generic Weekdays (as Monday, Tuesday, Wednesday) are recognised as similar
Unsupervised learning
Technical Implementation: Machine Learning for short-term forecast of traffic conditions 3/3 AUTOMATIC DAY TYPES CLUSTERISATION Unsupervised learning
Some days are special. For example: Pre- & post-holidays in Christmas period
Machine learning application – Traffic forecast service for the German Mobility Data Space (MDS) developed by Acatech End client: Acatech, a thinktank advising the German government about technological trends
Production year: 2021 Real-time Traffic Data from Materna •
3,234 Count Locations on the NRW Trunk Road Network (mainly on Motorways)
•
Live traffic speed
Challenge: Improve traffic using the “Mobility Data Space” (a platform for sharing data between different transportation industry players) in the following way: - Read data. - Process data. - Put back added-value data. Solution: PTV Optima Machine Learning Forecast uses real-time data to automatically detect typical daily time profiles in an unsupervised, fully automatic approach. Impact: - Easy KPIs monitoring. - Provide solution as fast as possible. - Decrease operational and services costs. - Self-learning approach.
Use case demonstration
Combination of analytical + ML forecast •
104 streets covered by data
•
~105 total streets
•
Forecast up to 60’ ahead in 1’
•
Error on flows reduced by 23% (GEH 6.9 → 5.3)
•
Error on speeds reduced by 55% (MAPE 16.2 → 7.3)
L
PTV methodology for real-time mobility management TM Functionality
Traffic Estimation
Traffic Forecast
Scenario Evaluation & Decision Support
“What is going on?“
“What is going to happen?“
“What would happen if?“ “What should we do?“
X
X
X
Simulation Approach
✓ ✓✓
“Usual“ conditions only
Data + Models + ML
✓✓
✓✓
Observed data approach Statistical approach
Maybe (with extensive measures)
✓
✓ ✓✓
Non-scientific methodology for real-time mobility management
Machine learning statistical approach
Data Statistical forecast
Model
Models Data
Data-driven models
Forecasting techniques combination
Model-based forecast
Conclusion
• Data-driven methodologies allows to streamline definition and calibration of tarditional transport models. • Clear advantages in using combine analytical and machine learning approaches for traffic and mobility forecast.
• Availability of significant, clean, normalized and labeled datasets in the mobility field is still an open issue.
Questions