Relative replacement window function in flight trajectory prediction

This paper presents a data-driven approach for short-term flight trajectory prediction within the Terminal Maneuvering Area (TMA) using deep learning models. Specifically, a relative displacement formulation is proposed to improve the learning efficiency and generalization capability of the predictive models. Three neural network architectures are evaluated: a GRU Encoder–Decoder model, a GRU Encoder–Decoder model with attention mechanism, and the FlightBERT transformer-based model. The relative displacement concept enables the models to focus on the sequential motion behavior of aircraft, minimizing the influence of geographic dependencies and reducing error accumulation over extended prediction horizons. Experimental evaluations are conducted using real-world ADS-B data collected from Hong Kong International Airport (HKIA) arrivals. Quantitative analysis and residual error distributions demonstrate that the GRU Encoder–Decoder model consistently achieves superior predictive performance across multiple target sequence lengths, outperforming both the attention-enhanced GRU model and the FlightBERT model. The results highlight the effectiveness of relative displacement modeling in enhancing trajectory prediction accuracy, contributing to safer and more efficient air traffic control operations.