Learning and Optimization for Reliable Airline Operations

Abstract

The aviation industry has experienced significant volatility and complexity in recent years, exacerbated by unpredictable factors such as adverse weather, airport congestion, and mechanical problems. This research aims to answer two key questions for improving airline operations: First, how can we create realistic schedules? To address this, we need to reliably predict flight segments. Second, how can we reduce costs while operating under uncertainty? One effective approach is by building robust schedules for crews. By tackling these questions, this thesis addresses two critical problems in airline operations: block time prediction and reliable crew pairing optimization. These problems are particularly challenging due to the long-term nature of prediction and planning, where limited knowledge about future conditions necessitates dealing with significant levels of uncertainty.

We first look at flight segment prediction through the lens of block-time - the total time from when an aircraft leaves the gate at the departure airport to when it arrives at the gate at the destination airport. Here we introduce an embedding-based neural network model to improve the long-term predictive accuracy of block time, which is essential for optimizing resource utilization and effective planning. Traditional methods in the literature often fail to capture the complex latent relationships between factors affecting block time and struggle with high-cardinality categorical data and temporal variations. Our large-scale study on the 2018 Bureau of Transportation Statistics (BTS) dataset incorporates high cardinality categorical features through embedding-based representation learning. The model's efficacy is demonstrated through rigorous experimental evaluations against ten baselines, showing significant improvements in long-term forecast accuracy for airline planning.

While flight segment prediction provides extensive knowledge about delays, it cannot be directly used to determine if a crew will successfully make a connection. This is because planning needs information about the probability that a crew can make a connection. Therefore, in the second study we introduce SurvCG, which integrates survival analysis into the crew pairing problem to quantify and manage uncertainties in flight schedules. Traditional approaches often assume deterministic scenarios or rely on predefined uncertainty sets that do not adequately reflect the true variability and unpredictability inherent in airline operations. SurvCG overcomes these limitations by incorporating survival probabilities into the optimization process, thus introducing a dynamic reliability score into the crew pairing cost function. This adjustment allows for the prioritization of crew schedules that are less likely to be disrupted by uncertainties, leading to more reliable and cost-effective operations. Additionally, this study also introduces a new metric to assess the predictive power of survival models in correctly predicting reliability scores. The efficacy of SurvCG is demonstrated through extensive simulations of different scenarios with varying disruptions, which show a significant reduction in propagated delay and disruptions.