Aircraft trajectory prediction using deep learning

With the continued growth of global air traffic, accurate trajectory prediction has become increasingly important for ensuring safety and improving efficiency, particularly within congested terminal airspaces. This thesis presents a data-driven, neural network-based approach for predicting the trajectories of arrival aircraft in the Istanbul Terminal Maneuvering Area, focusing on the spatiotemporal dynamics of this complex environment. The model is trained using 2023 Automatic Dependent Surveillance–Broadcast data, which includes flight parameters such as position, altitude, and speed. A comprehensive preprocessing pipeline is applied, including noise filtering, interpolation of missing values, temporal segmentation, and grouping by callsign to generate time-series samples. Each sample is formatted in JSONL, with input vectors containing both aircraft-specific features and slot-based attributes of surrounding aircraft. This structure allows the model to incorporate local traffic complexity and proximity interactions. A Multilayer Perceptron architecture is adopted to process the structured inputs. The model is tested on a scenario entirely excluded from training: 23 June 2024, the busiest traffic day of the year at Istanbul Airport, with 1,612 recorded flights. In this scenario, trajectories of six aircraft present in the TMA are predicted. Performance is assessed using Dynamic Time Warping and Mean Absolute Percentage Error. The model yields an average horizontal deviation of 2.34 NM and a vertical error of 369 feet per step. These results suggest that the model is capable of generating reasonably accurate trajectory predictions under high-density traffic conditions, while considering the presence and dynamics of nearby aircraft.