Multi-aircraft attention-based model for perceptive arrival transit time prediction

The states and trajectories of other aircraft are crucial in predicting arrival transit time; yet, current research predominantly concentrates on individual aircraft prediction and inadequately considers other aircraft within the airspace. The oversimplification of existing models raises concerns regarding their relevance and real-time applicability. Indeed, to effectively assist decision-making processes in air traffic management, we need solutions that are accurate, computationally efficient, and consistent with air traffic controller operations. To this end, we leverage the attention mechanism—which has demonstrated success in natural language processing—to appropriately consider all aircraft in the airspace in deriving a perceptive multi-aircraft transit time prediction. To achieve this, we propose a modified attention layer that can realistically mimic aircraft’s paying attention to others in a dynamic environment. The introduced model demonstrates a notable reduction in absolute prediction error by approximately 25% compared to state-of-the-art approaches. The functionality and effectiveness of the proposed attention layer are rigorously validated through extensive evaluation during the model’s learning process. Additionally, we introduce a model detachment technique in the feature importance analysis to determine the features that influence the attention decision of one flight with respect to another. The promising results highlight the potential of employing the customized attention mechanism in multi-agent systems both within and beyond air transportation research. • Comprehensive attention-based ML model for aircraft’s arrival time forecasts. • New attention layer to better capture arrival traffic flow dynamics. • Comparison against state-of-the-art approaches using a Hong Kong case study. • New detachment technique to unveil model’s interpretability and feature importance.