A multi-level predictive methodology for terminal area air traffic flow

Abstract

Over the past few decades, the air transportation system has grown significantly. In particular, the number of passengers using air transportation has greatly increased. As the demand for air travel expands, airport departure/arrival demand almost reaches its capacity. In consequence, the level of delays increases since the system capacity cannot manage the increased demand. With this trend, the national airspace system (NAS) will be saturated, and the congestion at the airport will become even more severe. As a result of congestion, a considerable number of flights experience delays. According to the Bureau of Transportation Statistics (BTS), over 1 million flights are operated in a year, and about twenty percent of all scheduled commercial flights are delayed more than 15 minutes. These delays cost billions of dollars annually for airlines, passengers, and the US economy. Therefore, this study seeks to find out why the delays occur and to analyze patterns in which the delays occurred. Analysis of airport operations generally falls into a macro or micro perspective. At the macro point of view, very few details are considered, and delays are aggregated at the airport level. Especially, shortfalls in airport capacity and a capacity-demand imbalance are the primary causes of delays in this respect. In the micro perspective, each aircraft is modeled individually, and the causes of delays are reproduced as precisely as possible. Micro reasons for air traffic delays include inclement weather, mechanics problems, operation issues. In this regard, this research proposes a methodology that can efficiently and practically predict macro and micro-level air traffic flow in the terminal area. For a macro-level analysis of delays, artificial neural networks models are proposed to predict the hourly airport capacity. Multi-layer perceptron (MLP), recurrent neural network (RNN), and long short-term memory (LSTM) are trained with historical weather and airport capacity data of Hartsfield-Jackson Atlanta airport (ATL). In the performance evaluation, the models have presented decent predictive performance and successfully predicted the test data as well as the training data. On the other hand, Random Forests and AdaBoost are implemented in the micro-level modeling of the air traffic. The micro-level models trained with on-time flight performance data and corresponding weather data focus on a classification of the individual flight delays. The model provides interpretability and imbalanced data handling while the accuracy is as good as the existing methods. Lastly, the predictive model for individual flight delays is refined using the cost-proportionate rejection sampling (costing) method. Along with the integration of the costing method, general machine learning algorithms have been converted to cost-sensitive classifiers. The cost-sensitive classifiers were able to account for asymmetric misclassification costs without losing their diagnostic functionality as binary classifiers. This study presents a data-driven approach to air traffic flow management that can effectively utilize air traffic data accumulated over decades. Through data analysis from the macro and micro perspective, an integrated methodology for terminal air traffic flow prediction is provided. An accurate prediction of the airport capacity and individual flight delays will assist stakeholders in taking more informed decisions.