Improvements to Convective Weather Avoidance Modeling Using Supervised Learning

Decision support tools that can translate weather hazards into anticipated impact on air traffic are critical for safe and effective operations. 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 to 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’s Route Availability Planning Tool, NASA’s Dynamic Weather Routes system, the Arrival Route Status and Impact prototype, and the NextGen Weather Processor Traffic Flow Impact (TFI) tool; 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 avoidance only and does not estimate the impact of convective weather for strategic traffic management planning needs, such as Ground Delay Programs or 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, a probabilistic latent variable model is used to quickly make these attributions. This method leverages large quantities of traffic information and is used to construct a database of avoidance cases without requiring human supervision. These data are then used to train a convolutional neural network that couples the weather hazard data and the avoidance data in order to learn avoidance behaviors. Two versions of this model are trained: one trained with just the radar derived precipitation intensity and echo top height and another trained with additional weather parameters. These additional data, such as rapidly‐developing storms in NEXRAD radar data, expand the set of hazards that affects pilot behavior. A validation approach will be presented that compares the current CWAM model and the machine learning CWAM models. Finally, the effect these models have on airspace permeability and traffic flow will also 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.

- Indicates paper has been withdrawn from meeting

- Indicates an Award Winner

See more of: Advances in the Use of Artificial Intelligence Techniques in Support of Aviation, Range, and Aerospace Meteorology
See more of: 19th Conference on Artificial Intelligence for Environmental Science

<< Previous Abstract | Next Abstract >>