The overall objective of this project is to develop a novel approach to storm-avoidance aircraft trajectory optimization that combines optimal control techniques with imitation learning. The hypothesis it that the ideal manoeuvres to avoid the cells of a storm can be learned from demonstrations, which in this context, are storm-avoidance trajectories predicted by analyzing real storm-avoidance trajectories, rather than determined using pure aircraft trajectory optimization techniques. Both the prediction of the aircraft trajectory to avoid storms and its optimisation will be carried out based on the predicted evolution of storms using artificial intelligence techniques. The problem will be solved in the 3D space, i.e., the optimized storm-avoidance trajectories will have three spatial components plus time.

STORM-ATS

Subproject 1 (PID2021-122323OB-C31)

The storm-avoidance aircraft trajectory optimization method developed in Subproject 1 combines imitation learning with optimal control and, besides the weather forecast, it considers multiple aircraft described by precise dynamical models, the operational constraints, and all the uncertainties of the problem. It additionally allows a probabilistic upper bound on the risk of encountering a cell of a storm to be included in the formulation to take into account the risk aversion of the pilots. It is also endowed with decision-making capabilities to solve possible contingencies.

Subproject 2 (PID2021-122323OB-C32)

Two different storm-avoidance aircraft trajectory prediction methods will be developed in Subproject 2, one for the short-term time horizon (0-2 hours) and another one for the long-term time horizon (beyond 2 hours). Several factors are considered in the short-term aircraft trajectory prediction method, such as the three-dimensionality of the weather phenomena, the inherent uncertainty of the weather forecasts -which are in this case ensemble nowcasts-, the uncertain storm-avoidance strategies, and the uncertainty in the take-off time of those aircraft that are on-ground at the time of the prediction. The same factors are considered in the long-term aircraft trajectory prediction method, except the weather forecasts, which in this case are ensemble prediction systems. One important aspect in this case is to establish the probabilistic relationship between the expected trajectory deviations and the convection indicators provided by the ensemble prediction systems.

Subproject 3 (PID2021-122323OB-C33)

In Subproject 3, numerical weather predictions and machine learning techniques will be used to generate different predictions of the evolution of the storms, one for the short-term time horizon (1-2 hours), and another one for the long-term time horizon (up to 36 hours). The hypothesis is that there is a correlation between numerical weather prediction parameters and convective weather phenomena, which will be revealed using advanced neural network architectures.

The result of this project will represent a step forward in effectively and efficiently preventing the disruptive effects of adverse weather conditions on the air transport system, while maintaining high safety standards.