Minneapolis–Saint Paul International Airport (MSP), like the rest of the airport community has been using customary methods to acquire and apply weather and runway friction knowledge from various sources in the runway closure decision process. Since MSP experiences frequent winter storms, they needed an automated system - Runway Friction and Closure Prediction System (RFCPS) than can establish relationship between weather variables, friction observations and runway closure times. The RFCPS developed by National Center for Atmospheric Research (NCAR) uses a backend forecast engine combined with a machine learning module and rules of practice to predict runway friction and the onset and duration of runway closures for Minneapolis-Saint Paul International Airport.
This paper will discuss methods used to obtain and organize real-time and archived runway friction data and related meteorological observation data from MSP. It will discuss methods for quality control (QC) algorithms that were applied to get rid of erroneous data and assemble predictor and target data sets more suitable for various machine learning techniques. It will cover the challenges faced in deciding the best predictor and target variables for friction machine learning models. It will also go over how real time observations were used to forward error correct the friction forecast after they have been created by the ML models. Lastly, the paper will illustrate how the machine learning results for runway friction were translated to runway closure start and stop times. It will discuss the rules of practice used to come up with the runway closure guidance alerts generated by the RFCPS.