Synthetic Generation of Air Traffic Trajectories Using Deep Learning

This thesis explores the application of deep learning models to generate synthetic air trajectories, represented as sequences of sectors and time spent within each sector. The study compares different neural network architectures, including LSTM and Transformer, trained and evaluated on operational datasets. Additionally, the research investigates the effect of conditioning models with external variables, such as weather conditions, to simulate more realistic flight scenarios. Given the computational cost of training, parameter optimization was conducted using a progressive approach. While the study is exploratory, the results offer a foundation for future developments in air traffic simulation and management, aligning with the objectives of the European SESAR 3 program.