UC Berkeley

Airport arrival capacity, referred to here as the airport acceptance rate (AAR), is strongly influenced by the weather in the vicinity of the airport and thus AAR prediction necessitates an airport-specific weather forecast. Weather forecasts, however, are seldom accurate in predicting the actual weather conditions. Strategic decisions, for example arrival rates in a ground delay program (GDP), must be made ahead of time, usually more than two hours, when there is an uncertainty about the future capacity. This research uses probabilistic capacity scenarios to represent the uncertainty in the future arrival capacity. A probabilistic capacity scenario is defined as a time series of AAR values with which a certain probability of realization is associated. A set of probabilistic capacity scenarios may be used to represent the uncertainty in arrival capacity at an airport over the course of the day.

There has been considerable research in developing GDP models that determine efficient ground delay decisions and require probabilistic capacity scenarios as inputs. It is assumed that the capacity scenarios can be developed from weather forecasts or can be obtained from the expertise of the air traffic managers. There is, however, considerably less literature on the development of specific day-of-operation probabilistic capacity scenarios from weather forecasts. This limits the use of these GDP models in real- world application. This thesis fills that gap and presents methodologies to generate probabilistic capacity scenarios from weather forecasts.

In this thesis we develop methodologies for generating probabilistic capacity scenarios using a widely available airport-specific weather forecast called the Terminal Aerodrome Forecast (TAF). These methodologies require the issued TAF forecast and the realized capacity for days in the past. We apply and assess the performance of these methodologies on four US airports: San Francisco International Airport, Boston Logan International Airport, Chicago O'Hare International Airport and Los Angeles International Airports. Though we have focused on these airports as case studies, the TAF-based scenario generation techniques can be applied to any airport.

In the first methodology, TAF Clustering, the scenarios are representative capacity profiles for days having similar TAFs. Groups of similar TAFs are found using K-means clustering and the number is verified using Silhouette value. In the second methodology, Dynamic Time Warping (DTW) Scenarios, the scenarios are the actual realized capacity profiles for days that have similar TAFs. The similarity between TAFs is determined using a statistical technique for comparing multidimensional time series called DTW.

DTW Scenarios uses three airport specific input parameters. These parameters control the numbers and the probabilities of the scenarios. We determine the values of the parameters through optimization to maximize the performance of the scenarios through minimizing average delay costs. The optimal values are determined through a specialized algorithm designed for situations where evaluating the objective function is computationally expensive.

For San Francisco International Airport we also use another forecast: the San Francisco Marine Initiative forecast (STRATUS) to develop the scenarios. In this methodology called, Fog burn-off clustering, the scenarios are representative capacity profiles for days that have the fog burn-off time in the same quarter hour.

We measure the efficacy of the various scenario generation methodologies in a real world setting based on 45 historic days for each of the four case-study airports. For each day, the generated scenarios are provided as inputs to a static stochastic ground delay model (SSGDM) that determines the series of planned arrival rates that minimize the sum of ground delay costs and expected air delay costs, assuming that the plan is not adjusted to evolving information. The ground delay is determined directly from the SSGDM whereas the realized air delay is determined from a queuing diagram based on the planned arrival rate and the realized arrival capacity. The realized delay costs are averaged over 45 days for each airport, and is the metric used to compare the different scenario generation methodologies. Employing this approach, we compare the different methods for capacity scenario generation against each other and against two other reference cases. Under the first reference case, Naïve Clustering, the scenarios are developed from historical capacity data without the use of the weather forecast. Groups of similar arrival profiles are determined though K-means clustering. In the second reference case, Perfect Information, we assume that the GDP is planned based on perfect information about the future arrival capacity.

Our results show that, on average, scenarios generated using the TAF-based DTW method results in the lowest delay cost amongst all scenario based methodologies. It is shown that capacity scenarios generated using day-of-operation weather forecasts can reduce the cost of delays by 5%-30% compared to scenarios that do not make use of weather forecast. The benefit of the TAF based approach is more pronounced on days that have a greater capacity-demand imbalance when compared to Naïve Clustering.