Degree Type

Dissertation

Date of Award

2019

Degree Name

Doctor of Philosophy

Department

Aerospace Engineering

Major

Aerospace Engineering

First Advisor

Peng . Wei

Abstract

Air transportation direct share is the ratio of direct passengers to total passengers on a certain origin and destination (O&D) pair. It is an essential factor in air transportation planning, airline market strategy making, and airport operations scheduling. Better understanding and more accurate forecasting of direct share can benefit air transportation planners, airlines, and airports in multiple ways. However, in most of the previous research and practice, direct share is usually assumed as a fixed ratio, which is not hold for the air transportation practice, especially for long term forecasting. This research aims to analyze the characteristics of O&D direct share from multiple perspectives and develop accurate O&D direct share forecasting model, which can be a promising and reliable replacement for the model used by the Federal Aviation Administration (FAA) Terminal Area Forecast (TAF).

To analyze the characteristics of O&D direct share, a database containing rich information about O&D features, air travelling features, and socio-economic features is developed based on data mining on the Airline Origin and Destination Survey (DB1B) database, Air Carrier Statistics (T100) database, and IHS Global Insight Economic database. Based on the analysis of important features in the developed parametric regression models, the features that have significant impacts on O\&D direct share are identified.

To develop accurate O&D direct share forecasting models, the problem is studied under both panel data context and time series context. Under the panel data context, the observations of direct share on different O&Ds are assumed from the same population. Both parametric and non-parametric models are investigated. Based on the comparison, the developed nonparametric models can provide better forecasting performance compared to the model used by FAA TAF. The Gradient Boosting Machine (GBM) model can outperform other models explored in this research. To further improve the forecasting performance, a novel Category-based learning method is proposed, which can further improve the forecasting performance by efficiently categorizing the data based into different categories. The problem is studied under the time series context as well, for which the direct share time series are generated for each O&D pair. To exploit the modeling capabilities of different models, both classic time series models and artificial neural networks are explored. Based on the modeling and forecasting performance analysis and comparison, it is shown that, direct share time series is O&D specific, which means for different O&D pairs, the characteristics of the direct share time series and the best model which can describe and forecast the direct share time series are different. Based on this fact, we proposed a novel modeling method which combines the ideas of time series modeling and supervised learning based on feature engineering. The forecasting performance is further improved by the newly proposed model, especially for O\&D pairs with major market changes in history. The model developed under time series context can provide better forecasting performance comparing to the models developed under the panel data context, especially when generating longer term forecasting. To automatically select the best direct share time series forecasting model for different O&D pairs, a hybrid framework is proposed in this research for practical implementation.

Copyright Owner

Xufang Zheng

Language

en

File Format

application/pdf

File Size

103 pages

Available for download on Monday, November 30, 2020

Share

COinS