A machine learning approach to wind estimation and uncertainty using remote dropsondes and deterministic forecasts

Abstract

The U.S. Air force currently has a need for high altitude, unguided airdrops without making two passes over a drop zone (DZ). During conventional high altitude drops, aircrews fly over a DZ, release a dropsonde, compute a payload release point, loop back to the DZ, and release a payload. This work proposes a machine learning method that enables a single pass airdrop mission where a dropsonde is released en route to a DZ, the dropsonde measurement is merged with a weather forecast using machine learning methods, and the aircrew releases the payload when they reach the drop zone. Machine learning models are trained to use a deterministic weather forecast and a dropsonde measurement to predict the winds over a DZ. The uncertainty in the DZ wind prediction is inferred using quantile regression. The uncertainty estimate is nonstatic meaning it is unique for each airdrop mission, and the uncertainty estimate is derived from data that is already available to aircrews. The quantile regression uncertainty estimate replaces the single pass mission's potential need for ensemble forecasts. The developed models are evaluated using data near Yuma, AZ, with later evaluation of several other locations in the US. The machine learning models are shown to improve the accuracy of the wind prediction at the DZ from a remote location up to 117 km away by up to 43% over other methods. To generalize findings, we develop models at several US locations and demonstrate the machine learning methodology is successful at other geographic locations. Models trained on data from a set of DZs are then shown to be transferrable to DZs unseen by models during training. This moves the wind prediction methodology closer to a global solution. The inferred prediction uncertainty is found to reliably reflect the accuracy in the wind prediction. The dynamic wind uncertainty estimate allows for the assessment of mission risks as a function of day-of-drop conditions. For nominal drop parameters, single pass airdrop missions were simulated around the Yuma DZ, and the machine learning methodology is shown to be approximately 20% more accurate than other methods.