Additional capacity with Machine Learning ART meeting Bretigny Bob Graham & Floris Herrema
Runway Throughput Challenge 16 to 28 congested airports 1.5+ million flights not catered for 160+ million people not able to fly Runway throughput solutions exist promising significant benefit: Wake Re-categorisation Time based separation supporting multiple concepts Enhanced approaches (glideslope, displaced thresholds) The runway itself remains a frontier restraining the solutions and holding capacity secrets Machine learning offers a way to unlocking these secrets
Will we ever automate the tasks of an ATCO? ATC position 70s ATC position today 2 Air traffic controllers Analogic display Paper strips Phone coordination VHF radio clearance National radar feed 2 Air traffic controllers Digital display Paper or electronic strips Phone coordination VHF radio clearance National radar feed Multiple support systems in situation assessment, less in decision making
Final algorithm How to identify and analyse operational, cost, efficiency, runway-throughput and safety patterns and precursors at final approach and the runway in order to model and support tactical and strategical decision making and alerting solutions? Arrival Runway Occupancy Time (AROT) Taxi Out Time (TXOT) Prediction time in minutes Post processing AROT TXOT T2F & TAS Optimised separation Nrex 0.5 15 60 Time to Fly (T2F) True Airspeed (TAS) Optimised separation Runway exit prediction (Nrex)
User interface example AROT = 71 seconds
The Runway opportunity From a data/machine learning and operational perspective Aircraft type Weather Controller Runway configuration Airline policy Geography Pilot Data Day/ season
How the opportunities were addressed Rules Answers Data mining Answers Machine Learning Rules Decision trees Data Data Descriptive analytics Clean large historical data sets Feature selection Learn from historical data sets Predictive analytics Deep understanding of contributing factors Automatic and precise prediction of aircraft behaviour impacting runway performance
Prototypes What solutions were tested and technical approach used Post-processing predictions based on historical observations Time to fly (T2F) True Airspeed (TAS) Arrival Runway Occupancy Time (AROT) Taxi-Out Time (TXOT) Optimised separation Real time predictions based on historical observation Runway exit utilised (Nrex) Real time prototype
Normalized feature importance Runway exit utilised - features 1 0
LiDAR and ADS-B measurements Real time wind and ADS-B data Remote access
Runway Exit utilised (Nrex) Runway 06 exits RWEP: Runway Exit Point ALDT: Actual Landing Time
2 mile spacing
My position My position
Data capture 10NM EZY475A M 060 160 A319 IBE652R H 060 160 A343 Aircraft_ICAO_Cat Height_10NM IAS_10NM WindSpeed_3000ft WindDirection_3000ft 2700ft < Height < 3300ft 10NM 14
Data capture 5NM EZY475A M 060 160 A319 Height_5NM IAS_5NM WindSpeed_1500ft WindDirection_1500ft IBE652R H 060 150 A343 1200ft < Height < 1800ft 5NM 15
Data capture 2NM EZY475A M 060 160 A319 Height_2NM IAS_2NM WindSpeed_600ft WindDirection_600ft IBE652R H 060 140 A343 400ft < Height < 800ft 2NM 16
Approach & Methodology Consider (real-time) ADS-B and LiDAR data 13 variables 25 days of historical data 930 arrival flights Retrain the model every day with 2 hours of data Assess Nrex for Runway 06 arrivals and per Aircraft ICAO category For instance Air France A320 Runway 06 Start prediction 10:30 Technique used Random forest Output txt file (E1, E2, E3, E4)
Example Height_10NM IAS_10NM WindDirection_3000ft WindSpeed_3000ft Aircraft_ICAO_Cat Height_5NM IAS_5NM WindDirection_1500ft WindSpeed_1500ft Height_2NM IAS_2NM WindDirection_600ft WindSpeed_600ft 10NM 5NM 2NM
Real-time and historical prediction results Real time ADS-B Reliability (%) E3 or E4 Number of flights Light 61 280 Medium 64 400 Heavy 59 250 Real time ADS-B + LiDAR data Reliability (%) E3 or E4 Number of flights Light 74 280 Medium 78 400 Heavy 73 250 Real time ADS-B + LiDAR data Reliability (%) E3/E4 Number of flights Medium 81 400
Conclusion Machine learning techniques seem to be very effective in terms of Nrex performance prediction. Predictions can be computed in just a few seconds after the tree has been extracted. The tree supports quick and intuitive decisions. Such predictions are useful and can be used to support: In-depth analysis of runway efficiency and understanding of procedural improvements Development of a Tactical tool to alert ATC on runway exit utilised and extending AROT and later additional features e.g. predicted separation at exit of landed aircraft compared to following aircraft etc. Development of a Strategic tool to support ATC supervisor on; Coordination of the runway configuration and changing of the sequence algorithm Input to AMAN/DMAN or TBS to refine sequence development and time based separation Improved acceptance rate based on traffic mix, weather and runway configuration
Recommendations Include more real time ADS-B and LiDAR measurements. The methodology can easily be transposed to any other airport using real time ADS-B and LiDAR measurements. Live trial in 2018 To validate the accuracy of the Nrex for a couple of days. To test the prediction speed in real time.
Thank you With this study a better prediction will be established of the Nrex patterns and precursors, this research will stimulate further Dynamic Pair-wise Separation studies.