Airport aircraft taxi time prediction modeling using machine learning approaches

Abstract

Various planning tools for air traffic control support require a model that can predict an aircraft taxi time for optimized scheduling under limited resources. However, because the existing aircraft taxi time prediction models were not developed for use in these planning tools, there are several limitations in directly applying the existing models to the planning tools. In addition, accurate aircraft taxi time prediction is required in various fields related to air traffic management, such as flight delay estimation. Therefore, this dissertation discusses how to develop an aircraft taxi time prediction model that can be directly applied to various fields related to air traffic management such as planning tools and flight delay estimation. First, a method was proposed to develop a model for predicting the unimpeded taxi time, which is the time required when an aircraft moves at a nominal speed under no congestion on the airport surface. The unimpeded taxi time is used in the planning tools to calculate the earliest possible time for an aircraft to arrive at a specific location. The factor that has the greatest influence on the unimpeded taxi time is the ground travel distance. A model is proposed that can predict the unimpeded taxi time by accurately modeling the ground travel distance using the node-link model and classification tree. In addition, in this dissertation, a model development method is proposed to predict the taxi-out time of the departure flights in the presence of congestion on the airport surface. Existing taxi-out time prediction models that consider congestion could predict the taxi-out time just before or near the departure time of the aircraft. However, planning tools require the predicted taxi-out time to consider congestion from a point much earlier than the departure time of the individual flight. Therefore, to solve this problem, the proposed model was implemented so that a taxi-out time that considers congestion could be predicted from 2 hours before the departure time using Long Short-Term Memory, a time series prediction technique based on deep learning.