Estimating Wind Fields Using Physically Inspired Neural Networks With Aircraft Surveillance Data

Abstract

Wind estimation plays an important role in many aspects of our world, both for nowcasting and forecasting. This is also true for aircraft flight routes; where wind field knowledge can contribute to flight planning and safety. Aircraft can relay flight information back to ground-based receivers, from which local wind measurements can be derived. Such measurements can be used in interpolation or a meteo-particle model to produce estimates elsewhere; yet those means do not take into account wind dynamics. A physically inspired neural network approach is applied in this work. Flight paths are derived from the aircraft measurements, and used to simulate training data from ECMWF ERA5 reanalysis data. Next, actual aircraft measurements serve as input, with the goal of predicting an entire wind field over the Netherlands and surrounding areas at cruise altitudes. The network can be leveraged by introducing physical losses, which are found to smooth the predicted flows. The network is able to predict flow fields on both simulated and real measurements, given sufficient input data. It improves over the existing methods of relying on 6-hour forecasts or using the meteo-particle model. Magnitude error of wind is reduced to 2.85 m/s, coming from 3.88 m/s for 6-hour GFS forecasts and 4.78 m/s for the Meteo-Particle model. Directional, physically inspired networks have an error of 11.2 degrees, compared to 14.4 and 17.3 degrees for 6-hour GFS forecasts and Meteo-Particle model, respectively. However, both error metrics fluctuate significantly depending on whether flow is uniform or of divergent nature at a particular day. The network is fit for nowcasting and future work can be done for longer-term forecasting.