11.2 Advanced Forecasts, Constraint Translations, and Decision Models for Improved Air Traffic Management Given Weather Uncertainty

Thursday, 16 January 2020: 10:45 AM
206A (Boston Convention and Exhibition Center)
Michael Robinson, The MITRE Corporation, McLean, VA; and T. Niznik, J. K. Williams, and C. P. Taylor

Current models and procedures for air traffic management (ATM) center upon deterministic and experience-driven (anecdotes) solutions for highly uncertain operational disruptions. What is needed instead is a practical, data-driven process and decision-making approach that supports incremental ATM informed by probabilistic guidance that explicitly considers uncertainty in balance with the costs and risks for decision options. Such a capability will make ATM during impactful weather conditions more consistent and efficient, and promises to reduce avoidable delay, operating costs, and passenger disruptions.

To address these challenges, the MITRE Corporation, American Airlines (AA), and The Weather Company, an IBM Business (TWC) partnered together to define and test an alternative ATM framework to support a repeatable, data-driven process for cost-effective, proactive decisions in the face of uncertainty. Targeting an initial domain (New York LaGuardia [LGA] airport), weather constraints (reduced capacity or airport acceptance rate [AAR] given surface winds and ceilings / visibility), and ATM response (Ground Delay Program [GDP]), this framework focuses on predicting a range of impacts and response likelihoods to support improved , incremental decisions most beneficial to the flight operator.

As part of this framework, a novel approach and model were developed for stochastic predictions of LGA AAR using high-resolution forecast data. This approach accounts for the uncertainty in the expected outcome given a specific weather constraint, represents the inertia present in the system in that rapidly changing conditions do not necessarily lead to rapid changes in response, and accurately predicts AARs influenced by transient constraints. With this AAR predictor as a foundation, an additional model was developed to predict the likelihood (and associated characteristics) of a LGA GDP given anticipated capacity reductions relative to scheduled demand and associated, modeled delay resulting from subsequent demand-capacity imbalances. This information may then be used by traffic coordinators at American Airlines to make informed proactive decisions to adapt its flight schedule.

The AAR and GDP prediction models are driven by TWC’s Probabilistic Forecasts on Demand (PFoD) which provide hourly probability distribution function (PDF) and related descriptions of forecast uncertainty for 10 weather variables at any global location through an API. Of particular value are PFoD’s calibrated ensemble “prototype” forecasts designed to provide a set of equally-likely multivariate weather trajectories that span the range of possible outcomes. Applying the AAR and GDP predictors to 100 prototype forecasts for wind speed, direction, gusts, ceiling height, and visibility produces 100 predicted trajectories for airport constraint and potential GDP outcomes. In our collaborative efforts, these forecasts, updating hourly, extended to 24-hour lead time in one hour forecast increments.

As part of this presentation, we will describe the prototype weather forecasts and impact translation models and demonstrate performance for predicting AAR reduction and GDP at LGA airport. We will also present initial results from testing a real-time prototype developed for AA operational decision-makers to consider utility and outstanding challenges for ‘day-of’ and ‘next-day’ airport impact planning given a range of predictions derived from ensemble-based, translated weather constraints. We will then introduce an Adaptive Planning Framework (APF) that seeks to cluster common constraint / response scenarios, as defined by individual trajectories of AAR / GDP predictions and derived from TWC ensemble forecasts, in order to identify not only the initial recommendations for actions, but subsequent contingency plans for each scenario represented. With APF, the sequence of actions under multiple scenarios is optimized simultaneously, allowing for both the cost of incorrect actions and the risks of delayed decisions to be explicitly assessed through cost function evaluations. In this way, APF permits incremental, risk-managed, data-driven decision-making. Finally, we will discuss the applicability and potential benefits of this collective research to an airline operation.

Approved for Public Release; Distribution Unlimited. Public Release Case Number 19-2527

© 2019 The MITRE Corporation. ALL RIGHTS RESERVED

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner