On any given day, constraints in the National Airspace System, for instance weather, necessitate the implementation of Traffic Flow Management initiatives, such as Ground Delay Programs. The parameters associated with these initiatives, for example the location, scope, duration, etc., are typically left to human decision makers, who must rely on intuition, past experience, and weather and traffic forecasts. Although the decisions of these traffic flow specialists are recorded on a daily basis, few studies have attempted to apply data mining techniques to these archives in an attempt to identify patterns and past decisions that could ultimately be used to influence future decision-making. The goal of this study is to take a preliminary step towards informing future decision-making by proposing a technique for identifying similar days in the National Airspace System in terms of the Ground Delay Programs that were operationally implemented. Hence an airport perspective is being taken to identify these similar days, as opposed to considering possible airspace features. A modified k-means clustering algorithm is applied to all days in 2011, resulting in the identification of 18 clusters that represent unique combinations of Ground Delay Program that were historically implemented. A given day was described in terms of the presence or absence of 33 features that were a combination of Ground Delay Program locations and causes. By far the largest cluster that was identified consisted of 73 days on which low ceiling related Ground Delay Programs impacted San Francisco International Airport. In an attempt to verify the stated cause of the Ground Delay Programs, an Expectation Maximization clustering algorithm was applied to the 8,760 hourly Meteorological Aerodrome Reports, scheduled arrival rate and Ground Delay Program start and end time records for 2011. In general, clusters were identified that corroborated the stated causes of the Ground Delay Programs. However, these clusters often contained a significant number of members for which a Ground Delay Program did not occur. Findings from this initial study indicate that it is possible to identify similar days under which the National Airspace System operates, and clustering techniques appear to be promising methods for identifying the major causes of Ground Delay Programs.
展开▼