A Deep Neural Network Approach for Prediction of Aircraft Top of Descent

抄録

Abstract— An arrival flight starts to transit from the cruise phase to the descent phase at the top of descent (TOD). Pilots get to know the TOD locations via onboard devices, while controllers can estimate the TOD locations with the help of radar surveillance and simple rules. In order to help controllers to get a better situation awareness of the traffic surrounding an aerodrome, it is of great operational importance to get an accurate prediction of the TOD locations for arrival flights. In this paper, we propose to apply deep learning for TOD location prediction for arrival flights. To do so, a TOD-specific feature engineering is suggested and applied to historical flight trajectories. Then the simple yet effective multilayer perceptron neural network model is adopted for TOD prediction. A case study on the arrival flights to Singapore Changi airport with respect to one-month historical trajectory data is carried out. Experiments demonstrate that the adopted deep learning method is effective for TOD location prediction. When compared against several typical machine learning models for regression, the adopted model yields a mean square error of 0.0039, which is smaller than the error achieved by the comparison models. Meanwhile, the adopted deep learning model yields TOD location prediction errors of 0.29 nautical miles (NM) on average with a standard deviation of 46.88 NM.