Date of Award

Fall 2022

Access Type

Thesis - Open Access

Degree Name

Master of Science in Aeronautics

Department

College of Aviation

Committee Chair

Dahai Liu, Ph.D.

First Committee Member

Hong Liu, Ph.D.

Second Committee Member

Ahmed Abdelghany, Ph.D.

College Dean

Alan J. Stolzer, Ph.D.

Abstract

Growth in air traffic demand in the United States has led to an increase in ground delays at major airports in the nation. Ground delays, including taxi time delays, directly impacts the block time and block fuel for flights which affects the airlines operationally and financially. Additionally, runway configuration selection at an airport significantly impacts the airport capacity, throughput, and delays as it is vital in directing the flow of air traffic in and out of an airport. Runway configuration selection is based on interrelated factors, including weather variables such as wind and visibility, airport facilities such as instrument approach procedures for runways, noise abatement procedures, arrival and departure demand, and coordination of ATC with neighboring airport facilities. The research problem of this study investigated whether runway configuration selection and taxi out times at airports can be predicted with hourly surface weather observations. This study utilized two sequence-to-sequence Deep Learning architectures, LSTM encoderdecoder and Transformer, to predict taxi out times and runway configuration selection for airports in MCO and JFK. An input sequence of 12 hours was used, which included surface weather data and hourly departures and arrivals. The output sequence was set to 6 hours, consisting of taxi out times for the regression models and runway configuration selection for the classification models. For the taxi out times models, the LSTM encoder-decoder model performed better than the Transformer model with the best MSE for output Sequence 2 of 41.26 for MCO and 45.82 for JFK. The SHAP analysis demonstrated that the Departure and Arrival variables had the most significant contribution to the predictions of the model.

For the runway configuration prediction tasks, the LSTM encoder-decoder model performed better than the Transformer model for the binary classification task at MCO. The LSTM encoder-decoder and Transformer models demonstrated comparable performance for the multiclass classification task at JFK. Out of the six output sequences, Sequence 3 demonstrated the best performance with an accuracy of 80.24 and precision of 0.70 for MCO and an accuracy of 77.26 and precision of 0.76 for JFK. The SHAP analysis demonstrated that the Departure, Dew Point, and Wind Direction variables had the most significant contribution to the predictions of the model.

Download

Share

COinS