The worldwide increase in demand for air traffic over the last several years has coincided with increased delays in the National Airspace System (NAS). According to Department of Transportation estimates, the annual cost of these delays is over $3 billion in costs to airlines alone. What's more, 76.9% of the delays in 2007 were weather-related, an increase over previous years. This situation will only get worse, as demand for airspace is projected to climb further in the coming years, and it will become increasingly important to operate the airspace system efficiently even in the face of storms.

Researchers have been working on the challenge of minimizing air traffic delays for several decades, mostly in the field of Air Traffic Flow Management (ATFM). Historically, ATFM models have taken expected capacity to be a known input, and have only focused on optimizing routing decisions based on this input. Under clear weather conditions, these expected capacities are fairly stable and tend to reflect reality. However, under stormy weather conditions, capacity is highly variable, and the use of estimated capacity for planning may be unreliable. More recently researchers have developed stochastic models for ATFM, which take as input a probabilistic weather forecast or a probability distribution for capacity.

The missing piece has been that a realistic probabilistic model of airspace capacity under convective weather conditions does not exist. Such a model is critical for decreasing weather-related delays, yet it has been a challenging problem due to the chaotic nature of weather and to the unique requirements of ATFM forecasts. More concretely, forecasts for ATFM must be reasonably precise and fine-grained, in both spatial and temporal scales. Knowing that there is a 30% chance of rain in the Boston area this afternoon, for instance, does not help to determine if there will be a route open from the east into Boston Logan Airport at 5PM, or if flights should incur delay on the ground and avoid entrance into Boston between 5 and 9PM.

One approach toward determining the capacity of the terminal-area airspace is by computing the bottlenecks in the weather-impacted region, to determine aircraft routes that maximize the throughput of the airspace. This research draws on the research of Mitchell, Polishchuk and Krozel (2006, 2007) in estimating airspace capacity under convective weather. We compare the airspace capacities and bottlenecks predicted by the weather forecasts with those of the observed weather data. This gives us one potential metric by which weather forecasts could be evaluated. We discuss the advantages and disadvantages of using the airspace capacities predicted in this manner to make routing decisions for arrival traffic flows.

We then present the first steps towards creating a stochastic model of terminal-area capacity. We consider a scenario in which short-term routing decisions are made based on MIT Lincoln Laboratory's (deterministic) 0-2 hour convective weather forecast product, and the predictions of the Vertically Integrated Liquid (VIL) in the airspace surrounding the airport. Specifically, we consider the following question: if aircraft are routed along trajectories that are clear according to the t-minute weather forecast, what is the probability that that these trajectories will actually be viable in the observed weather? The rationale behind this question is that by measuring forecast accuracy at the route-based level rather than the pixel-based level, we will be able to draw conclusions that are more suited for ATFM purposes. For instance, a 10km x 10km storm cell forecast might be displaced by 10km to the west when observed, resulting in no correct pixels (in a pixel-by-pixel comparison). For air traffic planning purposes, however, this forecast is potentially quite good because moving an aircraft's planned trajectory 10km east might be acceptable.

To answer the question raised, we use a data-driven approach to develop a model of route robustness. We synthesize hundreds of routes in the forecast grid, and correlate features of these routes with blockage in the observed grid, where a route is considered to be blocked if there does not exist a viable route in a small neighborhood of the original route. We use these correlations to determine the probability of a route being blocked, given the features of that route in the forecast data. Features of interest include the maximum, average and standard deviations of the VIL forecast along the route, and the closest distance of the route to forecast Level 3 weather. In our analysis, we use Corridor Integrated Weather System (CIWS) forecasts and observed weather data for the Hartsfield-Jackson Atlanta International Airport (ATL) terminal area from the summer of 2007. Our results show that it is possible to synthesize routes that end up being open in the true weather grid with high probability, even for the 60-minute forecast horizon. This suggests that subject to minor adjustments in routes, planning arrival trajectories at a 10-, 30-, and 60-minute time horizon is quite reasonable.

In the paper, we will present the details of this route robustness model, and show how it can be used to create a stochastic forecast of terminal airspace capacity. We will also discuss how this model can be integrated with Air Traffic Management decision-support tools. Although this work focuses on the terminal-area, and arrivals in particular, the methodology developed can be modified to model departures as well as en-route airspace.