Enhancements to Convective Weather Avoidance Modeling Using Probabilistic Machine Learning

Decision support tools that can translate weather hazards into anticipated impact on air traffic operations are critical for safe and effective management. The Convective Weather Avoidance Model (CWAM) was developed to address this need. CWAM correlates observable weather parameters with pilot behaviors to produce a Weather Avoidance Field (WAF). This field provides the probability of pilot avoidance in the presence of convective weather, and it can inform air traffic controllers of the likelihood that a pilot will avoid particular weather events. CWAM has been used to great effect in decision aids such as the Traffic Flow Management System Route Availability Planning Tool, NASA’s Dynamic Weather Rerouting, and the Arrival Route Status and Impact; however, CWAM suffers from several shortcomings. First, CWAM is based upon a relatively small number (order several thousand) of human-selected cases of pilot avoidance, often involving a hard threshold for avoidance/non-avoidance determination. Second, CWAM is focused on tactical pilot avoidances only and does not estimate the impact of convective weather for more strategic Traffic Flow Management planning needs, such as Ground Delay Programs and Playbook Reroutes. Third, CWAM considers only two weather parameters as hazards along the pilot’s path: the spatial extent of intense precipitation and height of radar echo tops.

This paper will present solutions to these shortcomings and analyze the effectiveness these solutions have on the current model. To address the issues of human avoidance detection as well as the bias of the current CWAM toward tactical avoidance impacts, a probabilistic method of automated avoidance detection will be presented. This method leverages large quantities of traffic information and is used to construct a database (order tens of thousands) of avoidance cases without any human supervision. Additionally, these avoidance cases are not solely focused on tactical pilot deviations, but rather contain all types of avoidance (e.g., strategic reroutes). This allows for the expansion of CWAM to cover strategic Traffic Flow Management Initiatives as well. The method by which these data are used to improve CWAM will also be presented: a machine learning method that is trained to couple the weather hazard data and the avoidance data in order to learn avoidance behaviors. This machine learning method addresses the third shortcoming in that additional weather hazard data can be easily incorporated into the model. These additional data, such as rapidly-developing storms in NEXRAD radar data, introduce more hazards that affect pilot behavior. Finally, an evaluation between the automated avoidance and machine learning method with the prior method will be shown.

DISTRIBUTION STATEMENT A. Approved for public release. Distribution is unlimited. This material is based upon work supported by the Federal Aviation Administration under Air Force Contract No. FA8702-15-D-0001. Any opinions, findings, conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the Federal Aviation Administration.