A comparative analysis of models for predicting delays in air traffic networks

In this paper, we compare the performance of different approaches to predicting delays in air traffic networks. We consider three classes of models: A recently-developed aggregate model of the delay network dynamics, which we will refer to as the Markov Jump Linear System (MJLS), classical machine learning techniques like Classification and Regression Trees (CART), and three candidate Artificial Neural Network (ANN) architectures. We show that prediction performance can vary significantly depending on the choice of model/algorithm, and the type of prediction (for example, classification vs. regression). We also discuss the importance of selecting the right predictor variables, or features, in order to improve the performance of these algorithms. The models are evaluated using operational data from the National Airspace System (NAS) of the United States. The ANN is shown to be a good algorithm for the classification problem, where it attains an average accuracy of nearly 94% in predicting whether or not delays on the 100 most-delayed links will exceed 60 min, looking two hours into the future. The MJLS model, however, is better at predicting the actual delay levels on different links, and has a mean prediction error of 4.7 min for the regression problem, for a 2 hr horizon. MJLS is also better at predicting outbound delays at the 30 major airports, with a mean error of 6.8 min, for a 2 hr prediction horizon. The effect of temporal factors, and the spatial distribution of current delays, in predicting future delays are also compared. The MJLS model, which is specifically designed to capture aggregate air traffic dynamics, leverages on these factors and outperforms the ANN in predicting the future spatial distribution of delays. In this manner, a tradeoff between model simplicity and prediction accuracy is revealed. Keywordsdelay prediction; network delays; machine learning; artificial neural networks; data mining