An air traffic prediction model based on kernel density estimation

Abstract

This paper revisits a link transmission model that is designed for nationwide air traffic prediction. The prediction accuracy relies on the estimate of traversal time of each link, which is obtained through statistical analysis of historical trajectories. As the most straightforward approach, the average traversal time is often used in the model implementation. But the outliers inherent in the data samples can easily distort the estimate. To address this issue, this paper proposes to use the mode of the traversal times which corresponds to the value reaching the peak of the probability density function of data samples. The continuous probability density function is estimated using a non-parametric approach, kernel density estimation. As the mode is resistant to the outliers, using the mode to parameterize the link transmission model is a more robust approach. Simulations based on historical traffic data of three months show that, in comparison with the conventional mean approach, use of the kernel density estimation in the sector count prediction leads to a 6% reduction in modeling errors.