Accurate and fast-converging trajectory prediction based on Long Short-Term Memory neural networks

Abstract

With the emergence of Urban Air Mobility (UAM), the safety and efficiency of airspace operations will largely depend on the necessary evolutions of conventional Air Traffic Management (ATM) decision support systems. In this context, trajectory predictionwill be one of the most critical functions in future air traffic deconfliction services, and suitable algorithms will have to be implemented in both ground-based and airborne systems. These algorithms must be far more accurate, efficient and flexible than in present-day ATM. This study introduces a Long Short-Term Memory (LSTM)-based adjustable interpolation algorithm, which can be incorporated into future UAM decision support system architectures. In the absence of UAM operational data, the verification of the proposed algorithm focuses on a series of scenarios encompassing both airliner and helicopter flight trajectories. Results demonstrate that the proposed method reduces computation time by half without significantly sacrificing prediction accuracy compared to conventional linear interpolation methods. Furthermore, accuracy improvements of at least 50% are achieved compared to raw data, with no substantial increase in computational time. Additionally, the algorithm complexity is evaluated via big O notation analysis, showing that our proposed approach allows to train accurate prediction models efficiently even when a large amount of training iterations is required. With further developments, this algorithm shows high potential as the foundation trajectory prediction for UAM services in dense urban airspace, enhancing conflict detection and resolution capabilities and mitigating risks.