Machine learning and time series decomposition approaches for predicting airline crew non-availability

Airline crews serve as the foundational component of effective flying operations. It is becoming increasingly critical to have accurate assessments of how much of an entire airline crew is available for duty. In collaboration with Sabre Airline Solutions, we develop and analytical framework for aircraft crew non-availability forecasting and scheduling prediction using several machine learning models. We focus on non-availability counts for multiple different crew ranking categories and ground activities. The data provided comes from several major airlines and takes the form of assignments that render crew members unavailable for flying from the past 10 years, and includes timestamps, base locations, and designations for activity type, crew type, and crew category. We attempt to formulate predictions for the number of unique crew members rendered unavailable for flight assignments on a daily, weekly, and monthly basis. Our model accounts for the temporal and seasonal aspects of crew non-availability, as well as the effects from assignment features. A dashboard application is provided that allows planners to choose what subsets of the overall crew population to analyze, vary the parameters of the prediction model, and visualize the non-availability counts.