Trajectory prediction accuracy and its improvement
Created by Gaumont Noé supervised by Hadjaz Areski and Marceau-Caron Gaetan
Study of temporal predictions on characteristic trajectory points, such as the top of climb. In order to do so, real tracks and predictions are used in two ways:
Rate the predictions made by the FMS or with BADA
Develop innovative methods which hopefully make better predictions
Several methods have been tested during the intership:
Using aircraft previous positions in order to extrapolate futur positions, ie fit a curve to the flight profile.
It's a complex method with an expensive computation time. This method needs a lot of past positions to achieve an accurate prediction.
Use existing predictions and then improve them. Learn an offset to apply from predictions made by the FMS. The computed offset depends on the remaining time before the estimated TOC.
Corrections are pretty good for long term prediction, for shorter prediction corrections are poorer. There is still a lot of uncertainty even for short term predictions. If the FMS makes erratic prediction, this method will be affected.
Just like for BADA, compute the rate of climb for each layer. The ROC are learned from a dataset. The main difference is that an entire distribution is stored for each layer instead of a single value for the BADA. The flight history will also be taken into account by storing the rank in the distribution for each ROC recorded. Future ROC are estimated according to this history
Predictions are more accurate. There is less uncertainty in prediction. This method use only trustworthy data, currently collected by ADS-B