Wind estimation by multirotor dynamic state measurement and machine learning models

We present multiple models for the problem of wind estimation by a multirotor drone in low altitude atmospheric turbulence. Data is collected by test flights with an instrumented drone in proximity to an anemometer for training, validation, and testing. Three machine-learning models are developed: a long-shortterm-memory (LSTM) neural-network, an artificial neural-network, and a Gaussian-process-regression. These models are developed with variations in the inputs, considering the addition of drag estimations by known equations of motion and motor speeds, both of which improved performance. The LSTM model performed the best reaching 0.34 m/s root-mean-square-error in validation. The similarity between all model's performance without optimization indicates that we may be approaching the limits of accuracy of the experimental "ground truth"-a single anemometer that cannot be exactly co-located with the drone. The performance demonstrated so far suggests that the method may be useful in pollutant plume characterization, preliminary wind surveys, and other applications.