Two important drawbacks of state-of-the-art methods for predicting aircraft trajectories are that (a) they are limited to single trajectory predictions, and (b) their prediction horizon is a short time one. Consequently, the network effect resulting from the interactions of multiple trajectories is not considered at all, which may lead to huge prediction inaccuracies due to several reasons. This is due to the complex nature of the ATM system, which impacts the trajectory predictions in many different ways.
The design of the new data-driven methods for single trajectory prediction are focusing on increasing predictability and improving scalability, in order to be able to handle extended airspaces and large volumes of data simultaneously.
Multiple algorithms implemented for single trajectory prediction in DART are exploring different directions: Either by ingesting raw data, or enriched surveillance datasets with additional variables, or derived datasets such as enriched trajectories and AIDL datasets. In doing so, DART provides a comprehensive evaluation of state of the art machine learning algorithms for single trajectory prediction, also combined with clustering and model-based approaches, exploiting varying features concerning 4D trajectories.
Algorithms benchmarking activities included comparison between data-driven predictions versus flown trajectories, and data driven predictions versus Eurocontrol Network Manager pre-flight prediction. Results show that data-driven methods can achieve high accuracy in predicting trajectories, also when they exploit information about flight plans.
Validation results help to understand that the combination of models and data-driven approaches is the correct way to evolve current operational systems towards the implementation of Trajectory Based Operations (TBO).
Towards delivering an understanding on the suitability of applying agent-based modelling techniques taking into account multiple trajectories and towards assessing the impact of traffic to individual trajectory predictions, the focus is on the DCB problem in Air Traffic Management, whose solution takes place at the pre-tactical stage: our objective is to predict delays that are applied to the flights.
To this end, DART makes the following contributions:
DART provided two formulations of the demand-capacity balance (DCB) problem using multi-agent Markov Decision Processes (MDP), modelling flights as agents whose decisions range in the space of their preferred/allowed delays: A “flat” and a hierarchical model.
DART designed and devised four multi-agent reinforcement learning methods towards assessing the impact of traffic to individual trajectories for resolving the DCB problems at the planning phase.
These methods provide a shift-of-paradigm for regulating flights: Each agent, corresponding to a trajectory decides its own delay w.r.t to operational constraints, own constraints on delays, and of course, according to the cost of strategic delay imposed.
Evaluation results for all methods show that the proposed methods are capable to resolve hotspots, while keeping the average delay for flights at low levels, also compared to CFMU regulations, with fairness. All methods can incorporate stakeholders’ constraints on flight delays.
