Data-Driven Departure Flight Time Prediction Based on Feature Construction and Ensemble Learning

Abstract

Temporal–spatial resource optimization within the terminal maneuvering area has become an important research direction to meet the growing demand for air traffic. Accurate departure flight time prediction from taking off to the metering fixes is critical for departure management, connecting the surface operations, and overhead stream insertion. This paper employs ensemble learning methods (including bagging, boosting, and stacking) to predict departure flight times via different metering fixes based on four feature categories: initial states, operating situation, traffic demand, and wind velocity. The stacking method employs a linear regressor, a support vector regressor, and a tree-based ensemble regressor as base learners. The Guangzhou Baiyun International Airport case study shows that the stacking method proposed in this work outperforms other methods and could achieve satisfactory performance in departure flight time prediction, with a high prediction accuracy of up to 89% within a 1 min absolute error and 98% within a 2 min absolute error. Besides, the affecting factors analysis indicates that the operation direction, flight distance, and traffic demand in different areas significantly improve prediction accuracy.