Characterizing and classifying historical days based on weather and air traffic

This article describes our identification of sets of similar days where similarity is defined in terms of the conditions relevant to the planning of an air traffic flow management initiative. The work here represents a first step toward the construction of a decision support tool for dispatchers at airline operations centers and officials at the Federal Aviation Administration Air Traffic Control System Command Center. A side product of our work is an taxonomy of reasonable approaches for categorizing calendar days in aviation systems research. Terminal Area / Aerodrome Forecast and Aviation Routine Weather Report data describe forecast and observed weather at airports, respectively. Aviation System Performance Metrics data summarize airport operations. Reasonable methods for defining features within available data include: applying expert judgment, using Principal Component Analysis to capture variance among days, and summarizing observations of specific weather variables via weighted averages where weights reflect levels of air traffic. Our preferred features, based on expert judgment, include counts of scheduled arrivals and departures, the minimum forecast visibility, the maximum forecast runway crosswinds, and forecasts of snow, thunderstorm, and rain at the three busiest airports (or, in the case of crosswinds, at four key runways) in the New York area during key blocks of time. Many efforts in aviation systems use records on the presence or absence of traffic flow management initiatives at various time points as label data. We instead use a Partitioning Around Medoids algorithm for clustering as we do not wish to model current decision making or use a Euclidean distance metric but do wish to assign all days to a cluster. Our results indicate weak structure in our feature data; days are not arranged into a handful of clusters of days that each contain days which are strongly similar to one another.