A Decision-Tree Based Continuous Learning Framework for Real-Time Prediction of Runway Capacities

Abstract

Accurate and up-to-date runway capacity estimates are vital for efficient management of airport operations. They are also a pre-requisite for the strategic and tactical measures used to mitigate traffic congestion and optimize air traffic operations at airports. Due to increasing availability of surveillance data, Machine Learning (ML) methods are used for capacity prediction. However, such ML models are static in nature and cannot be easily refitted/re-trained to learn the dynamic changes runways continually experience. These ML methods either fail to consider key factors that might affect runway capacity or try to incorporate every possible factor and suffer from the curse of dimensionality. This paper proposes a machine learning based capacity prediction model that utilizes innovative feature engineering methods to approximate a set of variables that better explain the dynamics of a runway system and can learn incrementally. The capacity prediction model utilizes the Adaptive Random Forest, an online machine learning model, which is built with decision trees that can learn incrementally. The capacity prediction model is then trained and tested on data from Philadelphia International Airport (PHL). Results demonstrate that the ML models can perform runway capacity prediction every hour for six hours rolling window horizon, for both arrivals and departures, which on an average, deviate by less than 3 flights from the actual count. In terms of prediction accuracy, models achieve a Mean Average Percentage Error of 12.05% for arrivals and 13.16% for departures.