Data-driven runway occupancy time prediction using explainable machine learning model

• Main Author: Chow, Hong Wei
• Other Authors: Sameer Alam
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Mechanical engineering
• Online Access: https://hdl.handle.net/10356/150660
• Institution: Nanyang Technological University

With an increasing amount of flights, the demand for runways at airports increases as well. Innovative mechanisms are required to maximise the use of a runway such that the demand can be met. Such mechanisms include the prediction of Runway Occupancy Time (ROT), so that the Air Traffic Controllers (ATCs) are able to gauge how much time a particular flight needs on the runway. This allows them to prepare the next flight for the runway and effectively reduce the buffer times between flights, thus increasing the overall efficiency of the runway. The objective of this paper is to develop an explainable machine learning model to predict Runway Occupancy Time. The Decision Tree Regressor was chosen for this study and its performance was compared to other more complicated machine learning models. The Decision Tree Regressor, unlike the other machine learning algorithms, provides explicit rules on how the predictions of the ROT is derived. An example of a generated rule for runway 02L of Singapore Changi Airport is that if an aircraft is a medium aircraft from airline XXX, arriving between 2100 and 2159 hours UTC, with an approach speed of more than 83.344 m/s at the final approach fix, and with the trailing aircraft traveling slower, the predicted ROT will be 42.6 seconds. Results show that the Decision Tree Regressor has the least runtime out of all the models at 0.28 minutes during training and its prediction capabilities are also on par with the rest of the machine learning models. The Root Mean Square Error for the Decision Tree Regressor is 5.96 seconds, which is only 0.20 seconds away from the best performing machine learning model. This, coupled with the rules that the Decision Tree Regressor can provide, makes it easier for end-users to to accept the prediction results without compromising on the accuracy. Permutation importance was also applied to the decision tree, providing an insight into what affects the ROT the most.