Learning With Confidence the Likelihood of Flight Diversion Due to Adverse Weather at Destination

Abstract

When an aircraft is unable to land at its original destination airport, it is diverted to an alternate airport. This event has economic and operational implications for airspace users and airports, as well as negative environmental consequences. Diversions are triggered by many reasons, including adverse weather (e.g., low visibility) or medical emergencies. While the latter are certainly unpredictable, the former could be learned from observed diversions, given the weather at the estimated time of arrival and the landing capabilities of the aircraft and airport. Unfortunately, only airspace users are aware of the reason for the observed diversions. This implies that some (unknown) diversion in the historical data should not be learned by the model because they were triggered by reasons different from adverse weather. This paper presents a gradient boosted decision trees model that learned the likelihood of diversions due to adverse weather from noisy labels using confident learning. Results indicate that this method is effective in pruning of diversions caused by reasons other than adverse weather, and that diversions could be predicted with high precision and moderate recall.