Schiphol is Europa’s best connected airport en verwerkt op piekdagen tot 235.000 passagiers. Om deze soepel door de processen te leiden is een betrouwbare prognose van de drukte noodzakelijk. Schiphol laat zien hoe zij datatoepassingen ontwikkelt om het aantal reizigers zo accuraat mogelijk te voorspellen en hiermee processen in te richten.
2. Agenda
2
Introduction
• About Schiphol
• Traffic Analysis & Forecasting team
Implementation of machine learning
• Operational forecasting
• Tough ride to get it implemented
• What did/do we do to make it a success?
Questions
4. Traffic Analysis & Forecasting Team
4
Reporting &
Analysis
• Network analysis
(connectivity)
• Benchmarking
• Opportunities /
threats
• Regional Airports
Capacity
• Forecasting
Short term
=> hiring staff
Mid/Long term
=> infra planning
Long term
=> spatial planning
• Limitations: airport slots,
airport & airspace capacity,
environment, etc.
Commercial
• Airline business
cases: Route /
business
development
• Cargo cases
• Forecasts to set
airport charges
€
€
5. Agenda
5
Introduction
• About Schiphol
• Traffic Analysis & Forecasting team
Implementation of machine learning
• Operational forecasting
• Tough ride to get it implemented
• What did/do we do to make it a success?
Questions
8. Confidential 8
Lazy / Lean…
Why would an Airport work so hard to predict, while Airlines…
• Have insights in sales & searches
• ‘manage their yields’ (start promotion / discounts)
have best insights on the expected number of passengers per flight
Increase the share of
airline input
9. Confidential 9
… but stay in control…
Avoid full dependency of airlines
Validate Airlines’ input (no strategic behaviour?)
Strong
tooling
10. Confidencelevel
Time
10
Not user friendly and a
Lack of faith in
‘machine’
Results were strong, now
get organized
professionally
Make the
best & full
blown
Looks great!
Now make it
robust
Building a GUI
is not our
strength
The ML engine
is OK, though
btw, legacy tool
isn’t perfect
either
That last
5% is
tough
Promising
results
Forget about
bureaucracy
& go!
… hence we developed a tool using
machine learning
12. Confidential
Predict
Train
• Predict the past
• Make it realistic: do not use future information
Feedback: measure how well it performed
How to build a robust model?
13. Confidential
• What machine learning
algorithm to use
• What data to include and
what not
• How to model holdiays
• Importance weighting of
more recent data/holidays
What did we tune?
13
14. Confidential
Dealing with the Dutch holidays
14
• Holidays appear in a lot of different
configurations
• Employ time-series analysis and
disentangle holidays per region
• These predictions: additional
inputs to machine learning
algorithm
• Final tweak: weights, more
importance to holiday periods
Central +20%
North + 15%
South + 5%
15. Adding weights
• Holiday periods
• Aircraft size
• Historic years
• Winter/Summer season trends
Weights for Summer 2017
training
16. Measure whether it improved
16
Sum of all flights
departing in the
same 15 minute
bracket
ML prediction too highML prediction too low
Test data
17. When is it ‘good enough to go’?
17
Daily totals
Grouped by airline
Grouped by destination
Single flight level
18. Confidential
Reliable training data doesn’t
always appear accurate
Don’t rely on
historic data to be
fully accurate.
Things we revealed in our historic data
• 100 seats and…. 160 passengers
• -/- 61 passengers to Mumbai
• Freight shipped from Amsterdam to…
Amsterdam!
21. See where you can improve
Schedule
Load factor
Transfer
share
# seats
Test data
22. ML Roller coaster learnings
22
Got organized seriously
Define “when is it good enough to go?”
Hired skilled professionals
New tools = new user skills
Compare & validate
Data accuracy
23. Agenda
23
Introduction
• About Schiphol
• Traffic Analysis & Forecasting team
Implementation of machine learning
• Operational forecasting
• Tough ride to get it implemented
• What did/do we do to make it a success?
Questions