Using predictive machine learning models to analyse weather resilience and robustness in U.S Airp

Eduardo Filipe Montenegro Salvador Leitão da Fonseca (Student)

Student thesis: Master's Thesis

Abstract

Flight delays pose a major challenge to the aviation sector, disrupting operations, reducing efficiency, and undermining passenger experience. This thesis examines airport resilience through the lens of robustness, defined as the ability to sustain operational performance under adverse weather conditions. A dataset combining meteorological information from Meteostat/NOAA and airline performance metrics from Sage Data was constructed for six major U.S. airports (ATL, DEN, DFW, JFK, LAX, ORD) covering 2010–2025. Machine learning models, including linear regression, random forest, and gradient boosting, were applied to predict on-time flight performance and quantify the impact of meteorological variables such as precipitation, wind speed, and temperature extremes. Tree-based models outperformed linear models, with random forest achieving the highest predictive accuracy. Crucially, results reveal substantial variation in robustness across airports: Denver (DEN) and Atlanta (ATL) exhibited the highest resilience, maintaining more stable on-time performance despite weather disruptions, whereas New York (JFK) and Chicago (ORD) were most sensitive, with delays strongly linked to adverse conditions. Dallas/Fort Worth (DFW) and Los Angeles (LAX) showed intermediate robustness. By integrating predictive modelling with the conceptual framework of resilience, this thesis advances both academic understanding and practical assessment of airport performance. The comparative insights provide an evidence base for targeted strategies to mitigate vulnerabilities and strengthen resilience within the U.S. aviation system.