Supervised learning in air traffic management : a comparative study of algorithms for flight delay prediction

Abstract

A series of delays can have a strong effect on aviation, frustrates people traveling and affects the industry’s finances. In this work, two models were used to precisely predict when delays might occur: the PyTorch based Multi-layer Perceptron (MLP) and CatBoostClassifier. The work involves a dataset of domestic flights in the U.S.(August 2013 to August 2023) and checks if each flight was late or on time by looking at the delay in minutes. A variety of techniques have been added to ensure that the results are fair and can be repeated. The performance of accuracy, precision, recall and F1-score indicates that CatBoostClassifier achieves higher scores than MLP, with 98.6% accuracy and an F1-score of 99.1%, compared to 91.8% accuracy and an F1-score of 94.8% from MLP. Looking at the data shows that most delays are reported in June, July and December, and specific airports and carriers are often involved. From these findings, CatBoostClassifier shows it can handle tabular data and advises on the best ways airlines and air traffic management can plan and work together efficiently in aviation analytics.