Root cause analysis of ATC delays: A case study on KLM flights at Schiphol Airport

Abstract

Due to the continuous growth of air traffic up to the year 2020, the air transportation network has become more complex, and the airports and airspace busier. However, capacities have not grown at the same rate as air traffic, making Air Traffic Control one of the most encountered primary delays. A data driven approach is taken in order to expose the drivers of the ATC delays for KLM Royal Dutch Airlines flights at Schiphol Airport. The used data consists of public and proprietary data, and contains information related to the weather, KLM flight operations, operational data, and Airport Collaborative Decision Making. To perform the analysis, two causal methods were used, association rule mining as a baseline method and a Bayesian network as the state-of-the-art model. Both methods were able to identify various conditions that trigger and/or prevent ATC delay occurrence, and agreed on the majority of the identified influential factors of the ATC delay. It was found that the main influences of ATC delay are the average startup delay of flights in the 20 minute time interval of the flight's departure, as well as the received pure ATFM delay and the assigned regulation delay key. Additionally, other influential parameters on the ATC delay related both to the amount of traffic volume and congestion at the airport, as well as individual variables of the flight, such as the propagation of arrival delay, the number of updates in the CDM process and the delay in the closure of the doors. The main discrepancies in the results could be attributed to the limitations of both methods. In general, it was found that both methods are suitable to diagnose direct causes or influencing factors on a target variable. The Bayesian network method was found to be more suitable to better understand a system and the dynamics between a large number of variables, as the conditional dependencies can be observed from the learned structure, and are not hidden in a large number of frequent patterns. However, first diagnoses of influential variables can also be done using association rule mining, which could find more indirect effects on the target variable compared to the Bayesian network, in which indirect relations might be lost in the structure learning process.