UC Berkeley

Experiencing exceptionally high delays constitutes a “bad day” for the National Airspace System (NAS). This study applies machine-learning methods to model system-wide delay and predict high-delay days in the NAS during the 2010s. The analysis considers a broad range of potential contributors, including queuing delays, terminal conditions, convective weather, wind, traffic volume, and special events. Penalized regression, kernelized support vector regression, and ensemble regression models are trained to relate system delay to these spatially and temporally defined features. The selected model’s learned weights are used to examine the spatial distribution and temporal consistency of feature importance. Queuing delays, convective weather, and wind emerge as the most influential contributors to system delay. Model-predicted delays indicate an increasing frequency of high-delay days over the decade. Counterfactual analyses further suggest worsening convective-weather conditions after 2014 and a demand surge in 2013 that was subsequently offset by increased system capacity.