Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a...
Transcript of Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a...
![Page 1: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/1.jpg)
Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system
Herbert Naessens, 11 October 2018
![Page 2: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/2.jpg)
In this presentation
Who is EUROCONTROL Maastricht UAC
ATC to ATM, and the problem of predictability
Rationale for a deep neural network
Details of the implementation
Integration in the operational system
Real-life results
![Page 3: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/3.jpg)
EUROCONTROL Maastricht Upper Area Control Centre
Cross-border ATC
Upper area
> 1.8 million flights(5700 on peak day)
Highest controller productivity
Driven by innovation
![Page 4: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/4.jpg)
EUROCONTROL Maastricht Upper Area Control Centre
https://www.youtube.com/watch?v=gBwwik4F2Og
![Page 5: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/5.jpg)
Air Traffic Control to Air Traffic Management(ATC to ATM)
Amount of traffic an air traffic controller can handle safely has a limit
Traditional approach of splitting sectors in smaller pieces has reached limits
Delays have been increasing last couple of years
Vision: avoid peaks in individual sectors by proactive traffic measures (= Air Traffic Flow and Capacity Management)
Sector workload prediction 3h-30min horizon from ‘now time’
Detection of upcoming traffic clusters 30-10min horizon from ‘now time’
But predictability degrades quickly when look-ahead is 10min or longer ...
![Page 6: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/6.jpg)
Challenges to predict traffic for MUAC
1. Flights not conforming to the route in the filed flight plan because air traffic controllers give permission to fly shorter routes (local & upstream)
2. Uncertainty of departure times at airports in the vicinity
3. Rate of climb/descent, ground speed
![Page 7: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/7.jpg)
The problem
Flight plan
Flown trajectory
Is it possible to predict controller instructions ?And, by extension, the
flown trajectory ?
![Page 8: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/8.jpg)
Machine Learning
Key enabler is availability of historical data
Supervised learning:
Training DataLabels
Machine Learning Algorithm
Predictive ModelNew Data Predicted Label
online
offline
![Page 9: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/9.jpg)
Machine Learning Algorithms
Several machine learning algorithms have been evaluated
Decision Trees
Random Forests
Kernel Support Vector Machines
K-Nearest Neighbours
Neural Networks
Random forest with adequate pruning offered the best results out of the box.
With lots of tuning, a deep neural network could surpass the results by a small margin.
![Page 10: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/10.jpg)
Rationale for a deep neural network
Random forest required disproportionate more computing resources if amount of training data and number of predictors increased
The serialised model was much smaller with a neural network important for scalability: training is done offline; the serialised model is deployed as adaptation data to the production environment
Off-the-shelf libraries (TensorFlow)
high degree of customisability, e.g. custom cost functions
API integration with existing application codesolution had to be integrated in EUROCAE ED-153 Software Assurance Level 4 (SWAL4) environment written in Java
offloading computations to GPU cards speeds up training
![Page 11: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/11.jpg)
Target data to be predicted
Intersection observed trajectory and the MUAC Area of Responsibility (AoR)
Simplified to 4 points by iteratively applying the Douglas-Peuckeralgorithm For 99.6% of the flights, the lateral deviation does not exceed 5NM at any point along the trajectory. For 89%, the lateral deviation does not exceed 1NM
Makes sense because flown route is typically result from ‘direct-to’ and ‘heading’ instructions, and controllers like to minimize the amount of instructions
![Page 12: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/12.jpg)
Target data to be predicted
Spacing if colinear
![Page 13: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/13.jpg)
Transformation of target data to be predicted
New X–axis
New Y–axis
![Page 14: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/14.jpg)
Transformation of target data to be predicted
The x/y coordinates of the 4 points are rotated and scaled.
NCOP-BPXCOP axis from filed plan data known prior to prediction !
Coordinates on the new X-axis are scaled by 0.5
normalization for the target data
scaling along the new X-axis allows for a more optimal cost function
generic sanity checking on the output data
![Page 15: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/15.jpg)
Predictors : flight plan data
NCOP
XCOPBPXCOP
ADES
bearing to ADEP
+ day of week
+ expected 30min entry interval
![Page 16: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/16.jpg)
Predictors : military areas
15NMx15NM grid cells, indicating upper reserved FL re-usability of old training data if names or geographic definitions change
Based on theoretical time flight could reach cell
![Page 17: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/17.jpg)
Predictors : military areas
NFL = 250
RFL = 350XFL = 350FL = 450
FL = 300
![Page 18: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/18.jpg)
Neural network
3 hidden layers of 170 units with ELU activationdropout for regularisation
NCOPXCOPBPXCOPADEPBearing to ADESDay of weekEntry time intervalNFLRFLXFL 8 coordinates
+ noise
![Page 19: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/19.jpg)
Cost function
Most correct cost function would be lateral distance between the position on the predicted route and the position on the real route at equivalent progression times. difficult from a practical perspective
Pragmatic:
Distances at entry and exit
Area of polygon / L
![Page 20: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/20.jpg)
Training data
Flow from UK to south / south-east
~10% of all traffic, suffers heavily from route deviations
15 January 2015 – 20 March 2018 (more than 362.000 flights)
Incremental training with 2.600.000 batches of 1000 random samples
Neural network has also been trained on all flows (> 3.5 million flights), but is not yet in operational use due to integration issue legacy system
![Page 21: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/21.jpg)
Example prediction
Figure 2: prediction (red) for flight of figure 1 (blue) Figure 3: prediction for flight with active military areas
![Page 22: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/22.jpg)
Integration in operational system
iFMP FDP
Operational ATM Flow & Capacity management System
machine learning
neural net
Offline
historical data
TSA mgr
predict module
neural net
In operational use since January 2018
Flight data
Schedule military reservations
Route What-If
What-If trajectory
![Page 23: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/23.jpg)
Real life results
![Page 24: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/24.jpg)
Real life video
<iFMP video>
![Page 25: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/25.jpg)
Real life video
<iFMP video>
![Page 26: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/26.jpg)
Real life use case
CFE53TK EGLC to LIML
27’ delay because included in regulation OLNO sector
Neural network predicted that flight would not fly through OLNO but LUX
Flight was excluded by FMP operator from regulation : no delay
under study: exclude flights via NM B2B interface, enabling automation of this workflow
![Page 27: Predicting flight routes with a deep neural networkScience+in... · Predicting flight routes with a deep neural network in the Air Traffic Flow and Capacity Management system Herbert](https://reader031.fdocuments.in/reader031/viewer/2022031423/5c743bbf09d3f22e5a8bb45c/html5/thumbnails/27.jpg)